content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
In this paper we deal with the well-posedness of the Cauchy
problem for a linear hyperbolic operator whose coefficients depend only on
time. Namely, we consider the equation
\begin{equation}\label{eq1}
u_{tt}-\sum_{i,j=1}^n a_{ij}(t)u_{x_ix_j}=0
\end{equation}
in $[0,T]\times\mathbb R^n$, with initial data
\begin{equation}\label{eq2}
u(0,x)=u_0(x),\quad u_t(0,x)=u_1(x)
\end{equation}
in $\mathbb R^n$. The matrix $(a_{ij})_{i,j}$ is supposed to be real and symmetric. Setting
\begin{equation}\label{eq3}
a(t,\xi):=\sum_{i,j=1}^n a_{ij}(t)\xi_i\xi_j/|\xi|^2,\quad(t,\xi)\in[0,T]
\times(\mathbb R^n\setminus \{0\}),
\end{equation}
we assume throughout that $a(\cdot,\xi)\in L^\infty(0,T)$ for all $\xi\in\mathbb R^n\setminus\{0\}$.
Moreover, we suppose that the equation (\ref{eq1}) is strictly hyperbolic, i.e.
\begin{equation}\label{eq4}
\Lambda_0\geq a(t,\xi)\geq \lambda_0> 0
\end{equation}
for all $(t,\xi)\in[0,T]\times(\mathbb R^n\setminus \{0\})$.
It is a classical result that if the coefficients $a_{ij}(t)$'s
are real integrable functions, then the Cauchy problem (\ref{eq1}), (\ref{eq2})
is well posed in ${\mathcal A}'(\mathbb R^n)$, the space of real analytic functionals; moreover, if
the initial data vanish in a ball, then the solution vanishes in a cone, whose
slope depends on the coefficients $a_{ij}(t)$'s (see \cite[Theorems 1 and 3.a]{CDGS}).
On this basis, various well-posedness results can be proved by mean of the Paley-Wiener theorem (in the version of
\cite[p. 517]{CDGS}, to which we refer here and throughout) and some energy estimates. If the
coefficients $a_{ij}(t)$'s are Lipschitz-continuous then the Cauchy problem
(\ref{eq1}), (\ref{eq2}) is well posed in Sobolev spaces. Relaxing this regulatity assumption, one has that if the $a_{ij}(t)$'s
are Log-Lipschitz-continuous or H\"older-continuous of index $\alpha$, then
(\ref{eq1}), (\ref{eq2}) is well posed in $C^\infty$ or in the Gevrey space $\gamma^{(s)}$ for
$s<{1\over 1-\alpha}$ respectively (see \cite[Theorem 3.b,c]{CDGS}).
Suitable counterexamples show that in each case the regularity assumption on the $a_{ij}(t)$'s is sharp for the well posedness of (\ref{eq1}), (\ref{eq2}) in the corresponding function space.
It is a remarkable fact that in the above mentioned counterexamples the coefficients $a_{ij}(t)$'s are in fact $C^\infty$ for $t\not=0$, and each time the specific regularity fails only at $t=0$. In \cite{CDSK1} the authors showed that a control on the rate of the loss of Lipschitz regularity of the $a_{ij}(t)$'s as $t\to 0$ allows to recover well-posedness of (\ref{eq1}), (\ref{eq2}) in suitable function spaces. To be more specific, if the $a_{ij}(t)$'s are of class $C^1$ in $]0,T]$ and $|a_{ij}'(t)|\leq Ct^{-p}$, then (\ref{eq1}), (\ref{eq2}) is well posed in $C^\infty$ when $p=1$, and in the Gevrey space $\gamma^{(s)}$ for $s<{p\over p-1}$ when $p>1$. Concerning $C^\infty$ well-posedness, it was proved in \cite{CDSR} that a control on the second derivative of the $a_{ij}$'s as $t\to0$ allows to relax slightly the growth assumption on the first derivative up to $|a_{ij}'(t)|\leq Ct^{-1}|\log t|$.
In \cite{KR} some of the above results were extended to the case in which the coefficients $a_{ij}$'s depend also on the $x$ variable in $C^{\infty}$ fashion.
In this paper we consider non Lipschitz coefficients whose regularity is ruled by a modulus of continuity $\mu$, with a constant which blows up as $t\to 0$. More precisely, we assume that
\begin{equation}
|a_{ij}(t+\tau)-a_{ij}(t)|\leq \frac C{\nu(t)}\mu(\tau),\quad 0\leq\tau\leq\tau_0,\quad t,t+\tau\in\,]0,T],
\end{equation}
where $\nu(t)^{-1}$ is possibly non integrable at $t=0$ and where $\mu$-continuity is possibly strictly weaker than Lipschitz continuity. We investigate how the interaction between $\nu$ and $\mu$ affects the well-posedness of (\ref{eq1}), (\ref{eq2}).
In Section 2 we prove a technical regularization result for the coefficients $a_{ij}$'s.
In Section 3 we consider locally H\"older continuous coefficients satisfying
\begin{equation}|a_{ij}(t+\tau)-a_{ij}(t)|\leq \frac C{t^p}\tau^\alpha,\quad 0\leq\tau,\quad t,t+\tau\in\,]0,T]\end{equation}
with $0<\alpha<1$ and $p>1$, and we obtain well-posedness in the Gevrey space $\gamma^{(\sigma)}$ for $\sigma<\frac{p}{p-\alpha}$, a condition which fits perfectly with the ones of \cite{CDGS} and \cite{CDSK1}.
In Section 4 we consider the problem of $C^\infty$ well-posedness and we identify a precise relation between $\mu$ and $\nu$ which guarantees the latter. In particular we obtain well-posedness for coefficients satisfying
\begin{equation}|a_{ij}(t+\tau)-a_{ij}(t)|\leq \frac{C}{t|\log t|}\,\frac{\tau|\log \tau|}{\log|\log\tau|},\quad 0\leq\tau,\quad t,t+\tau\in\,]0,T],\end{equation}
where one can easily see that $\nu(t)^{-1}$ is non integrable and $\mu$-continuity is strictly weaker than Lipschitz continuity. Also in this situation the results fits with the ones contained in \cite{CDGS} and \cite{CDSK1} and contain them as particular cases.
\section{Approximation}
We begin by recalling the notion of {\it modulus of continuity}.
\begin{defn} Let $\tau_0>0$.
A function $\mu: [0,\, \tau_0]\to [0,\,+\infty[$ is a {\rm modulus of continuity} if it is continuous, concave, strictly increasing and $\mu(0)=0$.
\end{defn}
Let $\mu$ be a modulus of continuity and let $a\colon [0,T]\to\Bbb R$ be a bounded function. Without loss of generality we can assume that $\tau_0\leq T$. We assume that
\begin{equation}|a(t+\tau)-a(t)|\leq \frac C{\nu(t)}\mu(\tau),\quad 0\leq\tau\leq\tau_0,\quad t,t+\tau\in\,]0,T],\end{equation}
where $\nu\colon\,]0,T]\to\,]0,+\infty[$ is a non-decreasing continuous function such that, for some $\kappa>0$,
\begin{equation}\nu(t/2)\geq \kappa \nu(t),\quad t\in\,]0,T].\label{doubling}\end{equation}
\begin{rem}
Condition (\ref{doubling}) is satisfied whenever $\nu$ is concave. Moreover, it is satisfied by $\nu(t)=t^p$ for every real exponent $p>0$. On the other hand, it is not satisfied if $\nu(t)$ tends to $0$ too fast as $t\to 0$, e.g. by $\nu(t)=e^{-1/t}$.
\end{rem}
Now let $0<\epsilon\leq\tau_0\leq T$ and define
\begin{equation}
\tilde a_\epsilon(t):=\begin{cases}a(\epsilon)&\text{for $t\leq\epsilon$},\\a(t)&\text{for $\epsilon\leq t\leq T$},\\a(T)&\text{for $T\leq t$}.\end{cases}\label{reg1}
\end{equation}
Let $\rho\in C^\infty(\Bbb R)$ with ${\rm supp}\,\rho\subset[-1,1]$, $\rho(s)\geq 0$, $\int_{\Bbb R}\rho(s)\,ds=1$, set $\rho_\epsilon(s):=\frac1\epsilon\rho(\frac s\epsilon)$, and define
\begin{equation}a_\epsilon(t):=\int_{-\epsilon}^{\epsilon}\rho_\epsilon(s)\tilde a_\epsilon(t-s)\,ds, \quad t\in\Bbb R.\label{reg2}\end{equation}
We have the following
\begin{prop}\label{prop1}
Under the above hypotheses, there exist constants $C'$ and $C''>0$ such that, for $0<\epsilon\leq\tau_0$,
\begin{equation}
|a_\epsilon(t)-\tilde a_\epsilon(t)|\leq {C'}\min\left\{1,\frac1{\nu(t)}\mu(\epsilon)\right\},\quad t\in\,]0,T]\label{appr1}
\end{equation}
and
\begin{equation}
|a_\epsilon'(t)|\leq \frac{C''}\epsilon\min\left\{1,\frac1{\nu(t)}\mu(\epsilon)\right\},\quad t\in\,]0,T].\label{appr2}
\end{equation}
The constants $C'$ and $C''$ depend only on $C$, $\rho$, $\kappa$ and $\|a\|_{\infty}$.
\end{prop}
\begin{proof}
We have
\begin{multline*}
|a_\epsilon(t)-\tilde a_\epsilon(t)|=\left|\int_{t-\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)(\tilde a_\epsilon(s)-\tilde a_\epsilon(t))\,ds\right|\\
\leq \int_{t-\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)|\tilde a_\epsilon(s)-\tilde a_\epsilon(t)|\,ds.
\end{multline*}
If $t\geq2\epsilon$, then $t-\epsilon\geq t/2\geq\epsilon$, so $\nu(t-\epsilon)\geq\nu(t/2)\geq\kappa\nu(t)$. Therefore, we have
\begin{multline*}
|a_\epsilon(t)-\tilde a_\epsilon(t)|\leq \int_{t-\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)| a(s)- a(t)|\,ds\\
\leq \int_{t-\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)\frac C{\nu(t-\epsilon)}\mu(|s-t|)\,ds\\
\leq \int_{t-\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)\frac C{\nu(t/2)}\mu(|s-t|)\,ds\\
\leq \int_{t-\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)\frac{C/\kappa}{\nu(t)}\mu(|s-t|)\,ds\leq \frac{C/\kappa}{\nu(t)}\mu(\epsilon).
\end{multline*}
If $0<t\leq\epsilon$, then $\tilde a_\epsilon(s)=\tilde a_\epsilon(t)=a(\epsilon)$ for $s\leq\epsilon$, and therefore,
\begin{multline*}
|a_\epsilon(t)-\tilde a_\epsilon(t)|\leq \int_{\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)| a(s)- a(\epsilon)|\,ds\\
\leq \int_{\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)\frac C{\nu(\epsilon)}\mu(|s-\epsilon|)\,ds\\
\leq \int_{\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)\frac C{\nu(t)}\mu(\epsilon)\,ds\leq \frac{C}{\nu(t)}\mu(\epsilon).
\end{multline*}
If $\epsilon\leq t\leq 2\epsilon$, then $0\leq t-\epsilon\leq \epsilon\leq t$. Therefore, we have
\begin{multline*}
|a_\epsilon(t)-\tilde a_\epsilon(t)|\leq \int_{t-\epsilon}^{\epsilon}\rho_\epsilon(t-s)| a(\epsilon)- a(t)|\,ds+\int_{\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)| a(s)- a(t)|\,ds\\
\leq \int_{t-\epsilon}^{\epsilon}\rho_\epsilon(t-s)\frac C{\nu(\epsilon)}\mu(|t-\epsilon|)\,ds+\int_{\epsilon}^{t+\epsilon}\rho_\epsilon(t-s)\frac C{\nu(\epsilon)}\mu(|t-s|)\,ds\\
\leq \frac{C}{\nu(\epsilon)}\mu(\epsilon)\leq \frac{C}{\nu(t/2)}\mu(\epsilon)\leq \frac{C/\kappa}{\nu(t)}\mu(\epsilon).
\end{multline*}
The thesis follows setting $C':=\max\{C, C/\kappa, 2\|a\|_\infty\}$.
In order to estimate $a'_\epsilon$, we observe that
\begin{multline*}
|a'_\epsilon(t)|=\left|\int_{t-\epsilon}^{t+\epsilon}\rho'_\epsilon(t-s)\tilde a_\epsilon(s)\,ds\right|=\left|\int_{t-\epsilon}^{t+\epsilon}\rho'_\epsilon(t-s)(\tilde a_\epsilon(s)-\tilde a_\epsilon(t))\,ds\right|\\
\leq \int_{t-\epsilon}^{t+\epsilon}|\rho'_\epsilon(t-s)||\tilde a_\epsilon(s)-\tilde a_\epsilon(t)|\,ds.
\end{multline*}
Then we procede as above, noticing that $\rho'_\epsilon(t)=\frac1{\epsilon^2}\rho'(\frac t\epsilon)$, and hence $$\int_{t-\epsilon}^{t+\epsilon}|\rho'_\epsilon(t-s)|\,ds=\frac{\|\rho'\|_{L^1}}\epsilon.$$
The thesis follows setting $C'':=\|\rho'\|_{L^1}\max\{C, C/\kappa, \|a\|_\infty\}$.
\end{proof}
\section{Well posedness in Gevrey spaces}
In this section we shall prove that if the coefficients $a_{ij}$'s are locally H\"older continuous of exponent $\alpha$, with a H\"older constant which grows like $t^{-p}$ as $t\to0$, then the Cauchy problem (\ref{eq1}), (\ref{eq2}) is well posed in a suitable Gevrey space $\gamma^{(\sigma)}$, where $\sigma$ depends on $\alpha$ and $p$.
As we pointed out in the Introduction, since the coefficients $a_{ij}$'s
are real integrable functions, the Cauchy problem (\ref{eq1}), (\ref{eq2})
is well posed in
${\mathcal A}'(\mathbb R^n)$, the space of real analytic functionals (which have by definition compact support). Moreover, if
the
initial data vanish in a ball, then the solution vanishes in a cone, whose basis is the same ball and whose slope depends on the coefficients $a_{ij}$'s. Therefore, it will be sufficient to
show that if $u_0$ and
$u_1$ belong to a a suitable Gevrey space $\gamma^{(\sigma)}$ and have compact support, then the corresponding solution
$u$ is not only in $W^{2,1}([0,T],{\mathcal A}'(\mathbb R^n))$, but it belongs
to the same Gevrey space in the $x$ variable for all $t\in[0,T$]. The result for initial data which do not have compact support follows by an exhaustion argument. Our main tools in the proof
will be the Paley-Wiener theorem and energy estimates.
\begin{thm}\label{th1}
Let $p> 1$ and $0<\alpha<1$, and assume that there exists a constant $C>0$ such that the function $a=a(t,\xi)$ defined by (\ref{eq3}) satisfies
\begin{equation}|a(t+\tau,\xi)-a(t,\xi)|\leq \frac C{t^p}\tau^\alpha,\quad 0\leq\tau,\quad t,t+\tau\in\,]0,T]\label{control1}\end{equation}
for all $\xi\in \Bbb R^n\setminus\{0\}$.Then the Cauchy problem
(\ref{eq1}), (\ref{eq2}) is $\gamma^{(\sigma)}$-well-posed for
$1\leq\sigma<
\frac{p}{p-\alpha}$.
\end{thm}
\begin{rem}
For a fixed $p>1$, passing to the limit as $\alpha\to1$ we regain the result of \cite{CDSK1}. In the same way, for a fixed $\alpha<1$, passing to the limit as $p\to 1$ we extend to $p=1$ the result of \cite{CDGS} which was valid only for $p<1$. The case $\alpha=1$, $p=1$ was considered in \cite{CDSK1} and will be reconsidered here in a more general context: in this case one has well posedness in $C^\infty$.
\end{rem}
\begin{rem}
The result in Theorem \ref{th1} can be considered sharp in the following sense. Let $p_0> 1$ and $0<\alpha_0<1$. It is possible to construct a positive function $a\in C^\infty(]0,T])\cap C([0,T])$ such that
$$|a(t+\tau)-a(t)|\leq \frac C{t^{p_0}}\tau^{\alpha_0},\quad 0\leq\tau,\quad t,t+\tau\in\,]0,T],$$
and it is possible to construct two functions $u_0,\ u_1\in\gamma^{(s)}(\mathbb R)$, for all $s>\frac{p_0}{p_0-\alpha_0}$ such that
the Cauchy problem
$$
\left\{
\begin{array}{ll}
u_{tt}-a(t)u_{xx}=0\\[0.2cm]
u(0,x)=u_0(x),\ u_t(0,x)=u_1(x)
\end{array}\right.
$$
has no solution in $C^1([0,r[\,;\, D'^{(s)})$, for all $s>\frac{p_0}{p_0-\alpha_0}$ and for all $r>0$ (here $D'^{(s)} $ denotes the set of Gevrey-ultradistributions of index $s$). The construction of such a counterexample is exactly the same as that contained in Theorem 5 in \cite{CDSK1}.
\end{rem}
\begin{rem} A result analogous to that of Theorem \ref{th1} can be proved if the singularity of the $a_{ij}$'s is located at $t=T$, with only minor obvious changes in the proof. As a consequence, the result is still valid if the coefficients have a finite number of singularities, where the loss of regularity is controlled as in (\ref{control1}).
\end{rem}
\begin{proof}[Proof of Theorem \ref{th1}]
We take the Fourier transform of $u$ with respect to $x$, and we denote it by $\hat u$. Equation
(\ref{eq1}) then transforms to
\begin{equation}\label{eq2.1}
\hat u_{tt}(t,\xi)+a(t,\xi)|\xi|^2\hat u(t,\xi)=0.
\end{equation}
Let $\epsilon$ be a positive parameter and for each
$\epsilon$ let $a_\varepsilon\colon[0,T]\times(\mathbb R^n\setminus\{0\})\to\mathbb R$ be
defined according to (\ref{reg1})-(\ref{reg2}).
We define the {\em approximate energy} of $\hat u$ by
\begin{equation}\label{eq2.2}
E_\varepsilon(t,\xi):=a_\varepsilon(t,\xi)|\xi|^2|\hat u(t,\xi)|^2+|\hat u_t(t,\xi)|^2,\quad (t,\xi)\in
[0,T]\times(\mathbb R^n\setminus\{0\}).
\end{equation}
Differentiating $E_\varepsilon$ with respect to $t$ and using (\ref{eq2.1}) we get
\begin{multline*}
E'_\varepsilon(t,\xi)=a'_\varepsilon(t,\xi)|\xi|^2|\hat u(t,\xi)|^2+
2a_\varepsilon(t,\xi)|\xi|^2{\rm Re}(\hat u_t(t,\xi)\bar {\hat u}(t,\xi))\\
+2{\rm Re} (\hat u_{tt}(t,\xi)\bar {\hat u}_t(t,\xi))\\
\leq \left(\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}+
\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}|\xi|\right)E_\varepsilon(t,\xi).
\end{multline*}
By Gronwall's lemma we obtain
\begin{equation}\label{eq2.3}
E_\varepsilon(t,\xi)\leq E_\varepsilon(0,\xi)
\exp\left(\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+
|\xi|\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t\right)
\end{equation}
for all $t\in[0,T]$ and for all $\xi\in\mathbb R^n$, $|\xi|\geq1$.
By Proposition \ref{prop1} with $\mu(\tau)=\tau^\alpha$ and $\nu(t)=t^p$ we have
\begin{multline*}
\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+
|\xi|\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t\\
\leq \int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+
|\xi|\int_0^T\left(\frac{|a(t,\xi)-\tilde a_\epsilon(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}+
\frac{|a_\varepsilon(t,\xi)-\tilde a_\epsilon(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\right)\,d t\\
\leq\int_0^{\epsilon^{\alpha/p}}\frac{C''}{\lambda_0\epsilon}\,dt+\int_{\epsilon^{\alpha/p}}^T\frac{C''}{\lambda_0\epsilon}
t^{-p}\epsilon^\alpha\,dt+\frac{2\Lambda_0}{\lambda_0^{1/2}}|\xi|\epsilon \\+ |\xi|\left(\int_0^{\epsilon^{\alpha/p}}\frac{C'}{\lambda_0^{1/2}}\,dt+\int_{\epsilon^{\alpha/p}}^T\frac{C'}{\lambda_0^{1/2}}
t^{-p}\epsilon^\alpha\,dt\right)\\
\leq M|\xi|\epsilon+ M\left(|\xi|+\frac1\epsilon\right)\left(\epsilon^{\alpha/p}+(\epsilon^{\alpha/p})^{1-p}\epsilon^\alpha\right)\\= M|\xi|\epsilon+ 2M\left(|\xi|+\frac1\epsilon\right)\epsilon^{\alpha/p},
\end{multline*}
where $M$ depends on $C'$, $C''$, $\lambda_0$, $\Lambda_0$, $\alpha$ and $p$.
Choosing $\epsilon=|\xi|^{-1}$ we get
\begin{equation}\label{mart1}
\left[\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+
|\xi|\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t\right]_{\epsilon=|\xi|^{-1}}\leq M+4M |\xi|^{\frac{p-\alpha}{p}}.
\end{equation}
Putting together (\ref{eq2.3}) and (\ref{mart1}) we get
\begin{equation}
E_{1/|\xi|}(t,\xi)\leq e^Me^{4M|\xi|^{\frac{p-\alpha}{p}}}E_{1/|\xi|}(0,\xi)\end{equation}
and, finally,
\begin{equation}
|\hat u_t(t,\xi)|^2+|\xi|^2|\hat u(t,\xi)|^2\leq \frac{e^M\Lambda_o}{\lambda_0}e^{4M|\xi|^{\frac{p-\alpha}{p}}}\left(|\hat u_t(0,\xi)|^2+|\xi|^2|\hat u(0,\xi)|^2\right)\label{mart2}
\end{equation}
Now if $u_0$, $u_1\in\gamma^{(\sigma)}\cap C^\infty_0$, the Paley-Wiener
theorem ensures that there exist $K,\delta>0$ such that
\begin{equation}\label{eq2.4}
|\hat u(0,\xi)|^2+|\hat u_t(0,\xi)|^2\leq K\exp(-\delta|\xi|^{1/\sigma})
\end{equation}
for all $\xi\in\mathbb R^n$, $|\xi|\geq1$. It follows from (\ref{mart2}) that if $\sigma<p/(p-\alpha)$,
then there exist
$K',\delta'>0$ such that
\begin{equation}\label{eq2.5}
|\hat u(t,\xi)|^2+|\hat u_t(t,\xi)|^2\leq K'\exp(-\delta'|\xi|^{1/ \sigma})
\end{equation}
for all $t\in[0,T]$ and for all $\xi\in\mathbb R^n$, $|\xi|\geq1$ and, therefore, $u\in W^{2,1}([0,T],\gamma^{(\sigma)})$. The proof is complete.
\end{proof}
\section{Well posedness in $C^\infty$}
Let $\psi\colon [1,+\infty[\to\,]0,+\infty[$ be a strictly increasing continuous function, such that $\psi'$ is non-increasing and $e^r\psi'(r)$ is non-decreasing. Moreover, we assume that
\begin{enumerate}
\item $\lim_{r\to\infty}\psi(r)=\chi$, $0<\chi\leq+\infty$;
\item $\lim_{r\to\infty}\psi'(r)=\eta$, $0\leq\eta<+\infty$;
\end{enumerate}
We set
\begin{equation}
\nu(t):=\begin{cases}\frac{t}{\psi'(|\log t|)}&\text{for $0<t\leq e^{-1}$},\\\null&\null\\ \frac{e^{-1}}{\psi'(1)} &\text{for $e^{-1}\leq t$}.\end{cases}\label{blow}
\end{equation}
A direct computation shows that $\nu$ is a non-decreasing continuous function and that
$\nu(t/2)\geq (1/2) \nu(t)$ for $t\in\,]0,T]$.
We define
\begin{equation}
\mu(\tau):=\frac{\tau|\log \tau|}{\psi(|\log \tau|)}\label{mod}
\end{equation}
and we assume that $\mu$ is strictly increasing and concave in $]0,\tau_0]$ for a suitable $\tau_0>0$, so it is a modulus of continuity.
\begin{thm}\label{th2}
Let $\nu=\nu(t)$ and $\mu=\mu(\tau)$ be as above, and assume that there exists a constant $C>0$ such that the function $a=a(t,\xi)$ defined by (\ref{eq3}) satisfies
\begin{equation}|a(t+\tau,\xi)-a(t,\xi)|\leq \frac{C}{\nu(t)}\mu(\tau),\quad 0\leq\tau\leq\tau_0, \quad t,t+\tau\in\,]0,T],\label{control2}\end{equation}
for all $ \xi\in \Bbb R^n\setminus\{0\}$. Then the Cauchy problem
(\ref{eq1}), (\ref{eq2}) is well-posed in $C^\infty$.
\end{thm}
\begin{rem}Examples of functions satisfying all the above properties are $\psi(r)=1-e^{-\alpha r}$ with $0<\alpha\leq1$, $\psi(r)=1+\log r$ and $\psi(r)=r^\beta$ with $0<\beta\leq1$.
In particular, we have:
\begin{itemize}
\item if $\psi(r)=r$ we have $\eta=1$ and $\chi=+\infty$ and we get $\mu(\tau)=\tau$ and $\nu(t)=t$, that is the situation considered in \cite{CDSK1};
\item if $\psi(r)=1-e^{-\alpha r}$ we have $\eta=0$ and $\chi=1$ and we get $\mu(\tau)=\tau|\log\tau|/(1-\tau^\alpha)$, which is equivalent to $\mu(\tau)=\tau|\log\tau|$, and $\nu(t)=\alpha t^{1-\alpha}$, that is a situation covered by the result of \cite{CDGS}, since $\nu(t)^{-1}$ is integrable;
\item if $\psi(r)=1+\log r$ or $\psi(r)=r^\beta$ with $0<\beta<1$, we have $\eta=0$ and $\chi=+\infty$, and we get $\mu(\tau)=\tau|\log\tau|/(1+\log|\log\tau|)$ or $\mu(\tau)=\tau|\log\tau|^{1-\beta}$. In both cases $\mu$-continuity is weaker than Lipschitz continuity. Moreover we have $\nu(t)=t|\log t|$ or $\nu(t)=t|\log t|^{1-\beta}$, so in both cases $\nu(t)^{-1}$ is not integrable.
\end{itemize}
The case in which $\mu(\tau)=\tau|\log\tau|$ and $\nu(t)^{-1}$ is not integrable is not covered by Theorem \ref{th2}, and we were not able to find a counterexample to $C^\infty$ well posedness either, so the question remains open. On the other hand, when $\mu(\tau)=\tau|\log\tau|$ and $\nu(t)=t$ by Theorem \ref{th1} we get authomatically $\gamma^{(\infty)}$ well posedness.
\end{rem}
\begin{rem} A result analogous to that of Theorem \ref{th2} can be proved if the singularity of the $a_{ij}$'s is located at $t=T$, with only minor obvious changes in the proof. As a consequence, the result is still valid if the coefficients have a finite number of singularities, where the loss of regularity is controlled as in (\ref{control2}).
\end{rem}
\begin{proof}[Proof of Theorem \ref{th2}]
Like in the proof of Theorem \ref{th1}, we take the Fourier transform $\hat u$ of $u$. Equation
(\ref{eq1}) then transforms to
\begin{equation}\label{meq1}
\hat u_{tt}(t,\xi)+a(t,\xi)|\xi|^2\hat u(t,\xi)=0.
\end{equation}
For $0<\epsilon\leq\tau_1:=\min\{\tau_0, T,e^{-1}\}$ we define $a_\varepsilon\colon[0,T]\times(\mathbb R^n\setminus\{0\})\to\mathbb R$ according to (\ref{reg1})-(\ref{reg2}). Again, we define an {\em approximate energy} of $\hat u$ by
\begin{equation}\label{meq2}
E_\varepsilon(t,\xi):=a_\varepsilon(t,\xi)|\xi|^2|\hat u(t,\xi)|^2+|\hat u_{t}(t,\xi)|^2,\quad (t,\xi)\in
[0,T]\times(\mathbb R^n\setminus\{0\}).
\end{equation}
Differentiating $E_\varepsilon$ with respect to $t$ and using (\ref{meq1}) we get
\begin{multline*}
E'_\varepsilon(t,\xi)=a'_\varepsilon(t,\xi)|\xi|^2|\hat u(t,\xi)|^2+
2a_\varepsilon(t,\xi)|\xi|^2{\rm Re}(\hat u_{t}(t,\xi)\bar {\hat u}(t,\xi))\\
+2{\rm Re} (\hat u_{tt}(t,\xi)\bar {\hat u}_t(t,\xi))\\
\leq \left(\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}+
\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}|\xi|\right)E_\varepsilon(t,\xi).
\end{multline*}
By Gronwall's lemma we obtain
\begin{equation}\label{meq3}
E_\varepsilon(t,\xi)\leq E_\varepsilon(0,\xi)
\exp\left(\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+
|\xi|\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t\right)
\end{equation}
for all $t\in[0,T]$ and for all $\xi\in\mathbb R^n$, $|\xi|\geq1$.
By Proposition \ref{prop1} with $\mu(\tau)$ and $\nu(t)$ given by (\ref{mod}) and (\ref{blow}), we have
\begin{multline*}
\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t=\int_0^\epsilon\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+\int_\epsilon^{e^{-1}}\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+\int_{e^{-1}}^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t\\
\leq\frac{C''}{\lambda_0\epsilon}\left(\epsilon+\int_{\epsilon}^{e^{-1}}
\frac{\psi'(|\log t|)}{t}\frac{\epsilon|\log \epsilon|}{\psi(|\log \epsilon|)}\,dt+\int_{e^{-1}}^{T}\frac{\psi'(1)}{e^{-1}}\frac{\epsilon|\log \epsilon|}{\psi(|\log \epsilon|)}\,dt\right).\\
\end{multline*}
Since $$\frac{\psi'(|\log t|)}{t}=-\frac{d}{dt}\psi(|\log t|)$$ and $\psi(|\log\epsilon|)\geq\psi(|\log\tau_1|)$, we obtain
\begin{equation}
\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t\leq M''\left(1+|\log\epsilon|\right).
\end{equation}
On the other hand
\begin{multline*}
\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t
=\int_0^T\left(\frac{|a(t,\xi)-\tilde a_\epsilon(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}+
\frac{|a_\varepsilon(t,\xi)-\tilde a_\epsilon(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\right)\,d t\\
\leq\frac{2\Lambda_0}{\lambda_0^{1/2}}\epsilon + \frac{C'}{\lambda_0^{1/2}}\left(\epsilon+\int_{\epsilon}^{e^{-1}}
\frac{\psi'(|\log t|)}{t}\frac{\epsilon|\log \epsilon|}{\psi(|\log \epsilon|)}\,dt+\int_{e^{-1}}^{T}\frac{\psi'(1)}{e^{-1}}\frac{\epsilon|\log \epsilon|}{\psi(|\log \epsilon|)}\,dt\right).
\end{multline*}
Arguing as above we, get
\begin{equation}\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t\leq M'\epsilon\left(1+|\log\epsilon|\right).\end{equation}
Choosing $\epsilon=|\xi|^{-1}$ we get
\begin{equation}\label{meq4}
\left[\int_0^T\frac{|a'_\varepsilon(t,\xi|}{a_\varepsilon(t,\xi)}\,d t+
|\xi|\int_0^T\frac{|a_\varepsilon(t,\xi)-a(t,\xi)|}{a_\varepsilon(t,\xi)^{1/2}}\,d t\right]_{\epsilon=|\xi|^{-1}}\leq M(\left(1+\log|\xi|\right)
\end{equation}
for $|\xi|\geq\tau_1^{-1}$.
Putting together (\ref{meq3}) and (\ref{meq4}) we get
\begin{equation}
E_{1/|\xi|}(t,\xi)\leq e^M |\xi|^M E_{1/|\xi|}(0,\xi)\end{equation}
and, finally,
\begin{equation}
|\hat u_{t}(t,\xi)|^2+|\xi|^2|\hat u(t,\xi)|^2\leq \frac{e^M\Lambda_o}{\lambda_0}|\xi|^M\left(|\hat u_{t}(0,\xi)|^2+|\xi|^2|\hat u(0,\xi)|^2\right).\label{meq5}
\end{equation}
Now if $u_0$, $u_1\in C^\infty_0$, the Paley-Wiener
theorem ensures that for all $\zeta>0$ there exists $K_\zeta>0$ such that
\begin{equation}\label{meq6}
|\hat u(0,\xi)|^2+|\hat u_{t}(0,\xi)|^2\leq K_\zeta |\xi|^{-\zeta}
\end{equation}
for all $\xi\in\mathbb R^n$, $|\xi|\geq\tau_1^{-1}$. It follows from (\ref{meq5}) that for all $\theta>0$
there exist
$K'_\theta>0$ such that
\begin{equation}\label{meq7}
|\hat u(t,\xi)|^2+|\hat u_{t}(t,\xi)|^2\leq K'_\theta |\xi|^{-\theta}
\end{equation}
for all $t\in[0,T]$ and for all $\xi\in\mathbb R^n$, $|\xi|\geq\tau_1^{-1}$, and therefore, $u\in W^{2,1}([0,T],C^\infty_0)$. The proof is complete.
\end{proof}
|
\subsubsection*{Related literature}
This work is related to two research streams.
In the first one, a number of strategies have been employed to optimize the gear-shift control of (hybrid) EVs. In~\cite{SainiSinghEtAl2016} a genetic algorithm is used to optimize the shifting patterns for an EV, whilst Pontryagin's principle has been used in~\cite{RitzmannChristonEtAl2019} to find the optimal gear selection and power distribution for a hybrid electric vehicle (HEV). The combination of dynamic programming (DP) and Pontryagin's principle has been successfully applied to find the optimal control of the gearshift command in HEVs~\cite{NgoHofmanEtAl2012}. Similarly, a combination of convex optimization and DP was employed to find the optimal gearshift command and power split for hybrid electric passenger~\cite{NueeschElbertEtAl2014,RobuschiSalazarEtAl2020} and race vehicles~\cite{DuhrChristodoulouEtAl2020}.
However, none of these techniques involve the optimization of the design of the transmission.
The second research line is concerned with the design of (hybrid) electric powertrains. The components' size and the energy management strategy of an HEV has been optimized jointly, leveraging convex optimization~\cite{MurgovskiJohannessonEtAl2012} and particle swarm optimization~\cite{EbbesenDoenitzEtAl2012}, but without accounting for the transmission.
Focusing on EVs, the transmission design and control problem has been studied analytically~\cite{HofmanSalazar2020}, and solved simultaneously using derivative-free methods~\cite{HofmanJanssen2017} and convex optimization~\cite{VerbruggenSalazarEtAl2019}, which was also applied to e-racing in~\cite{BorsboomFahdzyanaEtAl2021,LocatelloKondaEtAl2020}. However, all these approaches only consider FGTs or CVTs.
One of the challenges of finding the optimal gear-ratios and gear-shifting strategy for an MGT is the combinatorial nature of the gear-shifting problem, and the fact that the problems of finding the optimal shifting patterns and the optimal ratios are intimately coupled.
In~\cite{LeiseSimonEtAl2019}, mixed-integer nonlinear programming was used to find the optimal gear-ratios for an MGT, jointly computing the optimal gear-shifting control strategy.
The methods proposed in~\cite{SorniottiSubramanyanEtAl2011} compute the optimal gear-ratios for a given gear-shift map, whilst the authors of~\cite{GaoLiangEtAl2015} leverage DP to find the optimal gear-shift trajectory for a set of possible gear-ratios. In both contributions, the missing part of the problem is solved by exhaustive search.
In~\cite{MorozovHumphriesEtAl2019} high-fidelity nonlinear simulations are used to compare the design of the EMs and gear-ratios, and to search for the most advantageous drivetrain layout.
Finally, the authors of~\cite{VerbruggenSilvasEtAl2020} solve the joint e-powertrain design and control problem via derivative-free optimization for different architectures.
Overall, whilst all these methods in some form contribute to solve the problem of finding the optimal gear-ratios and control strategy, they are based on nonlinear, derivative-free and/or combinatorial optimization methods, or exhaustive search, resulting in high computation times without providing global optimality guarantees.
In conclusion, to the best of the authors' knowledge, a method to jointly optimize the design and control of an MGT for an EV in a computationally efficient manner is not available yet.
\subsubsection*{Statement of Contributions}
Against this backdrop, this paper presents a computationally efficient framework to jointly optimize the design and the control of the transmission of a central-drive EV on a given drive cycle.
To this end, we devise a convex yet high-fidelity EM model (see Fig.~\ref{fig:MotorMap}), and leverage it to compute the optimal CVT control trajectories and FGT design in closed form.
To tackle the complexity of jointly optimizing the gear-ratios and gear-shifting patterns of an MGT, we combine our closed-form solution for the FGT design with DP in an iterative fashion.
Finally, we validate and benchmark our methods with nonlinear simulations and mixed-integer quadratic programming (MIQP), respectively, and compare the achievable performance of an FGT-, CVT- and MGT-equipped EV in terms of energy efficiency. Thereby, the computational efficiency of our method enables us to also optimize the EM size via exhaustive search.
\subsubsection*{Organization}
The remainder of this paper is structured as follows: Section~\ref{sec:methodology} presents a model of a central-drive EV, including a convex EM model, and states the minimum-energy transmission design and control problem. We present solution algorithms to find the optimal gear designs and strategies for a CVT, FGT and MGT in Section~\ref{sec:solutionalgorithms}, whilst in Section~\ref{sec:results}, we offer a numerical case study comparing the three types of transmissions and present validation results.
Finally, Section~\ref{sec:conclusion} draws the conclusions.
\subsection{Energy-optimal Solution for the CVT}
\label{sec:CVT}
For the CVT, Problem~\ref{prob:GeneralProblem} consists of a pure optimal control problem.
Arguably, minimizing $J$ in~\eqref{eq:totalEnergyLoss} is equivalent to minimizing $P_\mathrm{m,loss}(t)$ at every time-instant. Rewriting the losses as a function of the gear-ratio $\gamma(t)$ we get
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
P_\mathrm{m,loss}(t) = \frac{d_0(t)}{\gamma(t)} + d_1(t) + d_2(t) \cdot \gamma(t),
\label{eq:Plossmodelconversion}
\end{equation}
\end{small}%
\endgroup
where the coefficients $d_0(t)$, $d_1(t)$ and $d_2(t)$ are related to the coefficients from~\eqref{eq:Plossmodel} as $d_0(t) = \frac{p_0(P_\mathrm{m}(t))}{v(t)/r_\mathrm{w}}\geq0$, $d_1(t) = p_1(P_\mathrm{m}(t))$ and $d_2(t) = p_2(P_\mathrm{m}(t))\cdot \frac{v(t)}{r_\mathrm{w}}\geq0$. Crucially, since $P_\mathrm{m}(t)$ is known in advance, these coefficients' trajectory is also known in advance.
We rewrite constraints~\eqref{eq:maxmotorspeed}, \eqref{eq:maxtorque} and \eqref{eq:maxPower} for $\gamma$ as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{aligned}
\gamma(t) &\geq 0\\
\gamma(t) &\geq \frac{|T_\mathrm{m,w}(t)|}{T_\mathrm{m,max}}\\
\gamma(t) &\leq \,\omega_\mathrm{m,max}\cdot \frac{r_\mathrm{w}}{v(t)}.
\label{eq:gammaconstraintsCVT}
\end{aligned}
\end{equation}
\end{small}%
\endgroup
Therefore, the minimal and maximal values for $\gamma(t)$ are
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{aligned}
\gamma_\mathrm{min}(t) &= \max\left(0,\frac{|T_\mathrm{m,w}(t)|}{T_\mathrm{m,max}} ,\gamma_{\mathrm{cvt},\min}\right)\\
\gamma_\mathrm{max}(t) &= \mathrm{min}\left( \omega_\mathrm{m,max}\cdot \frac{r_\mathrm{w}}{v(t)}\right).
\end{aligned}
\label{eq:ymin}
\end{equation}
\end{small}%
\endgroup
Since we consider only the control of the CVT and constraint~\eqref{eq:SlopeStartConstraint} only affects transmission design, it is not taken into account here.
We first consider an unconstrained operation. Since $P_\mathrm{m,loss}$ is a convex function for $\gamma(t)\geq 0$, we set the derivative of \eqref{eq:Plossmodelconversion} equal to zero and we solve for $\gamma(t)$. This way, the optimal unconstrained gear-ratio is
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}\label{eq:ucoptimumCVT}
\gamma_\mathrm{uc}^\star(t) = \sqrt{\frac{d_0(t)}{d_2(t)}}.
\end{equation}
\end{small}%
\endgroup
Since~\eqref{eq:Plossmodelconversion} is convex for non-negative arguments and is a function of a single optimization variable, its global optimum $\gamma^\star(t)$ corresponds to the feasible value that is closest to its unconstrained minimizer $\gamma_\mathrm{uc}^\star(t)$.
Therefore, the optimal gear-ratio $\gamma^\star$ is found as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}\label{eq:optimumCVT}
\gamma^\star(t) = \mathrm{min}(\gamma_\mathrm{max}(t),\mathrm{max}(\gamma_\mathrm{min}(t),\gamma_\mathrm{uc}^\star(t))).
\end{equation}
\end{small}%
\endgroup
This way, the global optimum of Problem~\ref{prob:GeneralProblem} can be efficiently computed by simple matrix operations.
\subsubsection{Optimal Gear-shifting Trajectory}
\label{sec:MGTgeartrajectory}
We consider the set of gear ratios $\gamma_i$ to be given, for which we have to solve problem~\ref{prob:GeneralProblem} to find the optimal gear trajectory.
This way, the objective $J$ is solely influenced by the gear-shifting trajectory defined by the binary variable $b_i(t)$ introduced in Section~\ref{sec:problemstatement}.
If we neglect gear-shifting costs setting $c_\mathrm{shift}=0$, we can minimize the $J$ by determining the resulting $P_\mathrm{m,loss}(t)$ for each gear $\gamma_i$ and choose $b_i^\star(t) = 1$ if the corresponding gear $\gamma_i$ results in the lowest feasible power loss at time-step $t$. This approach entails choosing the minimum element in $T+1$ vectors of size $n_\mathrm{gears}$, which is parallelizable and can be solved extremely efficiently.
If gear-shifting costs are included ($c_\mathrm{shift}>0$), Problem~\ref{prob:GeneralProblem} can be rapidly solved with dynamic programming due to the presence of only one state and input variable constrained to sets with cardinality $n_\mathrm{gears}$.
\subsubsection{Optimal Gear-ratios}
\label{sec:MGTGearRatios}
We consider the gear-shifting trajectory defined by the binary variable $b_i(t)$ to be given.
Since the gear trajectory is given, $C_\mathrm{shift}$ is constant and, therefore, minimizing $J$ corresponds to minimizing the energy losses~\eqref{eq:Plossmodelconversion}, which we rewrite as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{split}
\sum_{t=0}^T P_\mathrm{m,loss}(t)\cdot\Delta t = \sum_{i=1}^{n_\mathrm{gears}} \frac{f_{0,i}}{\gamma_i} + f_{1,i} + f_{2,i} \cdot \gamma_i ,
\end{split}
\label{eq:PlossMGT}
\end{equation}
\end{small}%
\endgroup
where the constant coefficients
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation*}
\begin{aligned}
f_{0,i} &= \sum_{t=0}^T d_0(P_\mathrm{m}(t)) \cdot b_i(t)\geq 0\\
f_{1,i} &= \sum_{t=0}^T d_1(P_\mathrm{m}(t)) \cdot b_i(t)\\
f_{2,i} &= \sum_{t=0}^T d_2(P_\mathrm{m}(t))\cdot b_i(t)\geq 0
\end{aligned}
\end{equation*}\\
\end{small}%
\endgroup
can be computed in advance.
We adapt the bounds for each $\gamma_i$ from~\eqref{eq:minmaxFGT}, considering that the constraints must hold only for the selected gear:
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{aligned}
\gamma_{\mathrm{min},i} &= \max_{t:b_{i}(t) = 1}\left(0,\frac{|T_\mathrm{m,w}(t)|}{T_\mathrm{m,max}}\cdot \right)\;\forall i=\{2,...,n_\mathrm{gears}\}\\
\gamma_{\mathrm{max},i} &= \min_{t:b_{i}(t) = 1}\left(\omega_\mathrm{m,max}\cdot \frac{r_\mathrm{w}}{v(t)}\right) \forall i=\{1,...,n_\mathrm{gears}\}.
\end{aligned}
\label{eq:minmaxMGT}
\end{equation}
\end{small}%
\endgroup
Hereby, the towing constraint~\eqref{eq:SlopeStartConstraint} only holds in first gear:
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\gamma_{\mathrm{min},1} = \max_{t:b_{1}(t) = 1}\left(0, \frac{|T_\mathrm{m,w}(t)|}{T_\mathrm{m,max}} , \frac{m_\mathrm{v} \cdot g \cdot \sin{(\alpha_0)} \cdot r_\mathrm{w}} {T_\mathrm{m,max} \cdot \eta_\mathrm{mgt}}\right).
\label{eq:minMGT1}
\end{equation}
\end{small}%
\endgroup
The energy-optimal gear-ratios can be found in a similar fashion to Sections~\ref{sec:CVT} and \ref{sec:FGT}. Since \eqref{eq:PlossMGT} is convex for $\gamma_i\geq 0$, we set its derivative equal to zero to compute the unconstrained optimum as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{split}
\gamma_{\mathrm{uc},i}^\star = \sqrt{\frac{f_{0,i}}{f_{2,i}}} \quad\forall i=\{1,...,n_\mathrm{gears}\}. \\
\end{split}
\end{equation}
\end{small}%
\endgroup
Hereby, since~\eqref{eq:PlossMGT} corresponds to the sum of $n_\mathrm{gears}$-times one-dimensional problems that are convex for non-negative arguments, its constrained minimum is the point that is closest to the unconstrained optimum whilst still satisfying the constraints
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\gamma_i^\star = \mathrm{min}(\gamma_{\mathrm{max},i} ,\mathrm{max}(\gamma_{\mathrm{min},i},\gamma_{\mathrm{uc},i}^\star)) \quad\forall i=\{1,...,n_\mathrm{gears}\}.
\label{eq:constrainedoptimumMGT}
\end{equation}
\end{small}%
\endgroup
\subsubsection{Finding the Optimal Gear Strategy and Design}
We finally solve Problem~\ref{prob:GeneralProblem} for the MGT iterating on the methods presented in Section~\ref{sec:MGTgeartrajectory} and~\ref{sec:MGTGearRatios}, as shown in Algorithm~\ref{alg:Iterative}.
Thereby, we terminate the iteration when the relative difference in the cost function $J$ is smaller than a tolerance~$\epsilon$.
\begin{algorithm}[t!]
\caption{Iterative Algorithm}
\label{alg:Iterative}
\begin{algorithmic}
\State $\gamma_i^\star = \gamma_{i,\mathrm{start}} $
\While{ $ \| J - J_\mathrm{prev} \| \geq \epsilon\cdot\|J\|$}
\State $J_\mathrm{prev} = J$
\State Compute $b_i^\star(t)$ for $\gamma_i^\star$ as in Section~\ref{sec:MGTgeartrajectory}
\State Compute $\gamma_i^\star$ for $b_i^\star(t)$ as in~\eqref{eq:constrainedoptimumMGT}
\State Compute $J$ for $b_i^\star(t)$ and $\gamma_i^\star$ as in~\eqref{eq:totalEnergyLoss}
\EndWhile
\end{algorithmic}
\end{algorithm}
\subsection{Note on Optimality}
Whilst the methods devised to solve Problem~\ref{prob:GeneralProblem} for the FGT and CVT owe global optimality guarantees to the convex problem's structure, the iterative approach presented in Section~\ref{sec:MGT} to solve the combinatorial Problem~\ref{prob:GeneralProblem} for an MGT sacrifices global optimality for computational efficiency.
However, the optimality benchmark provided in Section~\ref{sec:validation} is very promising.
\subsubsection*{Discussion}
\subsection{Compact Car Case Study}
\label{sec:numericalresults}
\begin{figure}
\centering
\includegraphics[width = \linewidth]{Figures/Results_Bar_Chart.pdf}
\caption{Optimal total energy usage for the FGT, CVT and MGTs.}
\label{fig:ResultChart}
\end{figure}
Fig.~\ref{fig:ResultChart} summarizes the results obtained for a 2-speed to 5-speed MGT (with and without gear-shifting costs), an FGT and a CVT, whereby the EM size is jointly optimized. The resulting EM operating points are shown in Fig.~\ref{fig:OperatingPoints}.
In addition, the resulting power loss, vehicle mass and EM size can be found in
\ifextendedversion
Table~\ref{tab:Results} in Appendix~\ref{sec:ResultTables}.
\else
the extended version of this paper~\cite{HurkSalazar2021}.
\fi
\begin{figure}
\centering
\includegraphics[width = \columnwidth]{Figures/OperatingPoints.pdf}
\caption{Resulting operating points for the FGT, CVT and MGTs without gear-shifting costs plotted over the scaled original EM map.}
\label{fig:OperatingPoints}
\end{figure}
We observe the FGT yielding the worst performance. The main reason for this is that, owing to the towing constraint~\eqref{eq:SlopeStartConstraint}, the EM needs to be able to deliver a high torque and therefore has a larger size compared to the other transmissions. This effect, combined with the FGT inability of choosing the EM speed in turn results in a lower number of operating points lying in the EM high-efficiency region.
Interestingly, the CVT is also outperformed by every MGT, as the additional benefits of choosing any gear-ratio do not compensate for its higher weight and lower transmission efficiency.
For the MGTs, we observe the operating points converging to the areas of high efficiency as the number of gears is increased.
The benefits of adding gear ratios reach a point of diminishing returns between a 3-speed and a 4-speed MGT.
Whilst the energy losses decrease as the number of gears is increased, the increase in required mechanical energy due to the increased transmission mass causes an overall higher energy consumption.
Finally, also when including a small gear-shifting cost, the MGTs still outperform the CVT and the FGT. As a result of the decreased benefit of gear changes, the best performing transmission changes from the 3-speed MGT to the 2-speed MGT.
Further figures of the EM speed and the gear-ratio trajectories over the drive cycle can be found in
\ifextendedversion
Appendix~\ref{sec:speedandgear}.
\else
the extended version of this paper~\cite{HurkSalazar2021}.
\fi
\subsection{Validation}
\label{sec:validation}
To benchmark the optimality of our methods, we first compute the globally optimal solution for a slightly less accurate quadratic version of the EM model which can be solved to global optimality with mixed-integer quadratic programming (MIQP)---the details on the implementation can be found in
\ifextendedversion
Appendix~\ref{sec:MIQP}.
\else
extended version of this paper~\cite{HurkSalazar2021}.
\fi
Second, we apply the proposed methods to the same quadratic motor model and compare our results to the global optima computed via MIQP.
For the FGT and the CVT, the results obtained with quadratic programming and the results obtained with the proposed methods applied on the quadratic model are identical.
Solving the MIQP for a 2-speed MGT with Gurobi~\cite{GurobiOptimization2021} took \unit[15]{min} to converge for a fixed EM size, whilst no convergence was achieved for MGTs with 3 gear-ratios or more.
The gear ratios obtained with MIQP for a 2-speed MGT are $\gamma_1^\star = 6.14$ and $\gamma_2^\star = 14.40$, whilst our Algorithm~\ref{alg:Iterative} converged to $\gamma_1^\star = 6.09$ and $\gamma_2^\star = 14.44$.
The resulting total energy consumption differs less than 0.03\%: a difference which can be ascribed to numerical tolerances, indicating a good agreement between our solution and the global optimum.
\begin{figure}
\centering
\includegraphics[width = \linewidth]{Figures/SimulationResults_Bar_Chart.pdf}
\caption{Comparison of the results for the FGT, MGT and CVT for the fractional model and the quadratic model, and the corresponding simulations.}
\label{fig:SimResultChart}
\end{figure}
Finally, in order to validate the results stemming from different EM models, the results obtained with the fractional EM model~\eqref{eq:Plossmodel} and the quadratic model are simulated using the original nonlinear motor data. This final step on the one hand assesses the performance that would be achieved in practice, and on the other hand validates the accuracy of our models.
Fig.~\ref{fig:SimResultChart} summarizes the results obtained with our methods and their performance in simulation, showing that our proposed fractional EM model~\eqref{eq:Plossmodel} can predict the energy consumption of the EM very accurately, outperforming the quadratic model for all the transmission technologies.
Additional results including the optimal EM size and vehicle mass are provided in
\ifextendedversion
Table~\ref{tab:Results} in Appendix~\ref{sec:ResultTables},
\else
extended version of this paper~\cite{HurkSalazar2021},
\fi
where we observe that both the proposed EM model and the quadratic model always result in the same EM size.
\subsection{Discussion}
A few comments are in order. First, we assume the EM power as given. This assumption is acceptable for central-drive or symmetric in-wheels architectures without active torque vectoring, and enables to pre-compute the power demand on a drive cycle. Thereby, our method is readily applicable also in scenarios where only a predefined fraction of braking energy can be recuperated.
Second, we minimize the electric power provided to the EM, which is almost equivalent to the power provided by the battery, and approximate the shifting costs with a constant value, leaving a more careful analysis accounting for the battery efficiency and shifting dynamics to a journal extension of this paper.
\subsection{Vehicle and Transmission}
\label{sec:VehicleModel}
For a given discretized driving cycle consisting of a speed trajectory $v(t)$, acceleration trajectory $a(t)$ and road grade trajectory $\alpha (t)$, the required power at the wheels is
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{aligned}
P_{\mathrm{t}}(t)=& m_{\mathrm{v}} \cdot\left(c_{\mathrm{r}} \cdot g \cdot \cos (\alpha(t))+g \cdot \sin (\alpha(t))+a(t)\right) \cdot v(t) \\
&+\frac{1}{2} \cdot \rho \cdot c_{\mathrm{d}} \cdot A_{\mathrm{f}} \cdot v(t)^{3},
\end{aligned}
\label{eq:dcycle}
\end{equation}
\end{small}%
\endgroup
where $m_\mathrm{v}$ is the total mass of the vehicle, $c_\mathrm{r}$ is the rolling friction coefficient, $g$ is the gravitational acceleration, $\rho_\mathrm{a}$ is the air density, $c_\mathrm{d}$ is the air drag coefficient and $A_\mathrm{f}$ is the frontal area of the vehicle.
Similarly, the required torque at the wheels is
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\begin{aligned}
T_{\mathrm{t}}(t)=& m_{\mathrm{v}} \cdot\left(c_{\mathrm{r}} \cdot g \cdot \cos (\alpha(t))+g \cdot \sin (\alpha(t))+a(t)\right).
\end{aligned}
\label{eq:Tdcycle}
\end{equation}
\end{small}%
\endgroup
We assume the gearbox efficiency to have a different constant value for each transmission technology. This way, the EM power required to follow the drive cycle is
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
P_\mathrm{m}(t) = \left\{\begin{array}{ll}
P_\mathrm{t}(t) \cdot \eta_\mathrm{fgt}^{-\mathrm{sign}(P_\mathrm{t}(t))} & \text{ if FGT } \\
P_\mathrm{t}(t)\cdot \eta_\mathrm{mgt}^{-\mathrm{sign}(P_\mathrm{t}(t))} & \text{ if MGT } \\
P_\mathrm{t}(t) \cdot \eta_\mathrm{cvt}^{-\mathrm{sign}(P_\mathrm{t}(t))} & \text{ if CVT,}
\end{array}\right.
\label{eq:Pmotor}
\end{equation}
\end{small}%
\endgroup
where $\eta_\mathrm{fgt}$, $\eta_\mathrm{mgt}$ and $\eta_\mathrm{cvt}$ are the efficiencies of the FGT, MGT and CVT, respectively.
Similarly, the EM torque at the wheels (i.e., without including the gear-ratio) is
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
T_\mathrm{m,w}(t) = \left\{\begin{array}{ll}
T_\mathrm{t}(t) \cdot \eta_\mathrm{fgt}^{-\mathrm{sign}(T_\mathrm{t}(t))} & \text{ if FGT } \\
T_\mathrm{t}(t) \cdot \eta_\mathrm{mgt}^{-\mathrm{sign}(T_\mathrm{t}(t))} & \text{ if MGT } \\
T_\mathrm{t}(t) \cdot \eta_\mathrm{cvt}^{-\mathrm{sign}(T_\mathrm{t}(t))} & \text{ if CVT},
\end{array}\right.
\label{eq:Tmotorw}
\end{equation}
\end{small}%
\endgroup
Assuming the friction brakes to be used only when the EM is saturated at its minimum power or torque (in order to maximize regenerative braking), $P_\mathrm{m}(t)$ and $T_\mathrm{m,w}(t)$ can be known in advance and the operating point of the electric motor is solely determined by the transmission. Thereby, the rotational EM speed $\omega_\mathrm{m}(t)$ is related to the wheels' speed through the transmission ratio $\gamma(t)$ as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\omega_\mathrm{m}(t) = \gamma(t) \cdot \frac{v(t)}{r_\mathrm{w}} ,
\label{eq:motorspeed}
\end{equation}
\end{small}%
\endgroup
where $r_\mathrm{w}$ is the wheels' radius. The transmission ratio (including the final drive) is subject to optimization and is constrained depending on the transmission technology as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
\gamma(t)\left\{\begin{array}{ll}
=\gamma_{1} & \text{ if FGT } \\
\in \{ \gamma_1,... ,\gamma_{n_\mathrm{gears}} \} & \text{ if MGT } \\
\in\left[\gamma_{\mathrm{cvt},\min }, \gamma_{\mathrm{cvt},\max }\right] & \text{ if CVT,}
\end{array}\right.
\label{eq:transmission}
\end{equation}
\end{small}%
\endgroup
where $\gamma_1,...,\gamma_{n_\mathrm{gears}} $ are fixed gear-ratios of the FGT (for $n_\mathrm{gears}=1$) and the MGT and $n_\mathrm{gears}$ is the number of gears in the MGT, whilst $\gamma_\mathrm{min}$ and $\gamma_\mathrm{max}$ are the gear-ratio limits of the CVT which we consider as given.
For the sake of simplicity, as a performance requirement, we focus on the stand-still towing requirements on a slope of $\alpha_0$:
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
m_\mathrm{v} \cdot g \cdot \sin{(\alpha_0)} \cdot r_\mathrm{w} \leq T_\mathrm{m,max} \cdot \left\{\begin{array}{ll}
\eta_\mathrm{fgt} \cdot \gamma_1 & \text{ if FGT } \\
\eta_\mathrm{mgt} \cdot \gamma_1 & \text{ if MGT} \\
\eta_\mathrm{cvt} \cdot \gamma_\mathrm{max} & \text{ if CVT}.
\end{array}\right.
\label{eq:SlopeStartConstraint}
\end{equation}
\end{small}%
\endgroup
However, our framework readily accommodates additional requirements which we leave to an extended version of this paper.
\subsection{Vehicle Mass}
The mass of the vehicle depends on the transmission used.
The total mass of the vehicle $m_\mathrm{v}$ is the sum of a base weight $m_0$, the weight of the gearbox $m_\mathrm{g}$ and the weight of the motor $m_\mathrm{m}$, yielding
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
m_\mathrm{v} = m_0 + m_\mathrm{g} + m_\mathrm{m} .
\label{eq:vehiclemass}
\end{equation}
\end{small}%
\endgroup
Similarly to \cite{VerbruggenSalazarEtAl2019,BorsboomFahdzyanaEtAl2021}, the motor mass $m_\mathrm{m}$ is modeled linear in relation to the maximum motor power $P_\mathrm{m,max}$ as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
m_\mathrm{m} = \rho_{\mathrm{m}}\cdot P_\mathrm{m,max} ,
\end{equation}
\end{small}%
\endgroup
where $\rho_\mathrm{m}$ represents the specific mass of the motor.
The mass of the gearbox is modeled linearly with the number of gears as
\par\nobreak\vspace{-5pt}
\begingroup
\allowdisplaybreaks
\begin{small}
\begin{equation}
m_\mathrm{g} = \left\{\begin{array}{ll}
m_{\mathrm{g},0} + m_{\mathrm{g}} & \text{ if FGT } \\
m_{\mathrm{g},0} + m_{\mathrm{g}}\cdot n_\mathrm{gears} & \text{ if MGT} \\
m_{\mathrm{cvt}} & \text{ if CVT,}
\end{array}\right.
\label{eq:transmissionmass}
\end{equation}
\end{small}%
\endgroup
where $m_{\mathrm{g},0}$ is the base mass for an FGT and an MGT, $m_{\mathrm{g}}$ is the added mass per gear, whilst $m_{\mathrm{cvt}}$ is the mass of a CVT.
\section{Introduction}
\input{chapters/ProjectBackground}
\section{Methodology}
\label{sec:methodology}
\input{chapters/VehicleModel}
\subsection{Motor Model}
\label{sec:MotorModel}
\input{chapters/MotorModel}
\subsection{Problem Statement}
\label{sec:problemstatement}
\input{chapters/ProblemStatement}
\section{Solution Algorithms}
\label{sec:solutionalgorithms}
\input{chapters/CVT}
\subsection{Energy-optimal Solution for the FGT}
\label{sec:FGT}
\input{chapters/FGT}
\subsection{Energy-optimal Solution for the MGT}
\label{sec:MGT}
\input{chapters/MGT}
\section{Results}
\label{sec:results}
\input{chapters/Results}
\section{Conclusions}
\label{sec:conclusion}
\input{chapters/conclusions}
\section*{Acknowledgment}
We would like to thank Mr.\ Olaf Korzilius, Dr.\ Steven Wilkins and Dr.\ Pascal Etman for the fruitful discussions as well as Dr.\ Ilse New for proofreading this paper.
\bibliographystyle{IEEEtran}
|
\section*{A. Experimental setting}
\subsection*{1. Surgical procedures}
LFPs recordings are performed on Wistar rats, which are maintained under standard environmental conditions in the animal research facility of the Department of Biomedical Sciences of the University of Padova. All the procedures are approved by the local Animal Care Committee (O.P.B.A.) and the Italian Ministry of Health (authorization number 522/2018-PR). Young adult rats aged 36 to 43 days and weighting between 150 and 200 g are anesthetized with an intra-peritoneal induction mixture of tiletamine-xylazine (2 mg and 1.4 g/100 g body weight, respectively), followed by additional doses (0.5 mg and 0.5 g/100 g body weight) every hour. The anesthesia level is constantly monitored by testing the absence of eye and hind-limb reflexes and whiskers’ spontaneous movements. Each animal is positioned on a stereotaxic apparatus where the head is fixed by teeth- and ear-bars. To expose the cortical area of interest, an anterior-posterior opening in the skin is made in the center of the head and a window in the skull is drilled over the somatosensory barrel cortex at stereotaxic coordinates $-1 \divisionsymbol -4$ AP, $+4 \divisionsymbol +8$ ML referred to bregma \cite{swanson2003}. A slit in the meninges is then carefully made with fine forceps at coordinates $- 2.5$ AP, $+ 6$ ML for the subsequent insertion of the recording probe. As a reference, the depth is set at 0 µm when the electrode proximal to the chip tip touches the cortical surface. The neuronal activity is recorded from the entire barrel cortex (from 0 to $- 1750$ $\mu m$), which is constantly bathed in Krebs’ solution (in mM: NaCl 120, KCl 1.99, NaHCO\textsubscript{3} 25.56, KH\textsubscript{2}PO\textsubscript{4} 136.09, CaCl\textsubscript{2} 2, MgSO\textsubscript{4} 1.2, glucose 11). An Ag/AgCl electrode bathed in the extracellular solution in proximity of the probe is used as reference.
\subsection*{2. Recordings}
LFPs are recorded through a custom-made needle which integrates a high density array, whose electrodes are organized in a $64\times4$ matrix. The operation principle of the multi-electrode-arrays used to record LFPs is an extended CMOS based EOSFET (Electrolyte Oxide Semiconductor
Field Effect Transistor). The recording electrodes are 7.4$\mu$m in diameter size and the needle is 300 $\mu$m in width and 10$mm$ long. The x- and y-pitch (i.e. the distance between adjacent recording sites) are 32$\mu$m. The multiplexed signals are then digitized by a NI PXIe-6358 (National Instruments) up to 1.25MS/s at 16bit resolution and saved to disk by a custom LabVIEW acquisition software.
The LFP signal is sampled at 976.56 Hz and band-pass filtered (2-300 Hz). The dataset analyzed for this work consist in 20 trials of basal activity lasting 7.22 seconds, that are recorded from 4 rats.
\subsection*{3. Barrel cortex}
The barrel cortex is the region of the primary somatosensory cortex (S1) that encodes tactile sensory inputs from the rodents' whiskers. The barrel presents a high degree of segmentation into vertical columns and horizontal layers. The whisker-related barrel column is a cylindrical structure spanning vertically the six layers of the barrel cortex, although its border is defined exclusively by spatially aligned sub-cellular structures in layer 4 called barrels.
Each barrel-column is composed of an archetypal circuit that is repeated in each column. Importantly, the barrels are laid out in a pattern that is nearly identical to the whiskers on the rat's snout that, together with the underlying neural circuit, suggests a highly specific correspondence between whiskers and barrel columns.
\begin{figure}[t!]
\centering
\includegraphics[width = 0.9\textwidth]{figS1.png}
\caption{\footnotesize Comparison between the stationary distributions obtained in the main text and the results $10^3$ simulations. Semi-transparent lines represent different simulations. Filled areas of the plots represent one standard deviation from the mean distribution. (a) Probability distribution of $\mathcal{D}$. (b) Probability distribution of a single $v_i$.}
\label{fig:probability_single}
\end{figure}
\section*{B. The extrinsic model}
\noindent As explained in the main text, we introduce a model of $N$ variables $(v_1, \dots, v_N)$ that are conditionally independent given the state of the external input. In particular, let us assume that there are some parameters $\bm{\xi} = (\xi_1, \dots, \xi_M)$ controlling the extrinsic modulation, that is
\begin{equation}
\label{eqn:conditionally_independent}
p(v_i, v_j, t \,|\, \bm{\xi}) = p(v_i, t \,|\, \bm{\xi}) \, p(v_j, t \,|\, \bm{\xi}) \quad\quad \forall i, j.
\end{equation}
Since at this moment we assume no direct interaction is happening, we think about the probability distribution $p(v_i | \bm{\xi})$ as shaped by extrinsic activity alone. That is, once we specify $\bm{\xi}$, no interaction happens between the $i$-th variable and the $j$-th variable. However, we now suppose that we do not have access on the states of the external parameters. Hence, we can only hope of describing to the joint probability distribution
\begin{equation}
\label{eqn:joint_prob_model}
p(v_1, \dots, v_N, t) = \int d\bm{\xi} \prod_{i=1}^N p(v_i, t \,|\, \bm{\xi})p(\bm{\xi}, t)
\end{equation}
which is what we typically observe in an experimental setting. In general, due to the marginalization over the external parameters this probability distribution is not factorizable, and thus in this sense it is not trivial.
In general, we want to work in the stationary limit and thus there is a complication we need to take into account. In particular, since the integrand of Equation (\ref{eqn:joint_prob_model}) can be arbitrarily complicated, we would like to be able to perform the stationary limit before the marginalization, i.e,
\begin{align*}
\lim_{t\to\infty}p(v_1, \dots, v_N, t) = \int d\bm{\xi} \prod_{i=1}^N & \lim_{t\to\infty} p(v_i, t \,|\, \bm{\xi}) \cdot \lim_{t\to\infty} p(\bm{\xi}, t).
\end{align*}
This limit, of course, is not always true, but it does hold if we assume that the time-scales of the two processes - the process for $\bm{v}$ and for $\bm{\xi}$ - are separated. In particular, we are interested in the limit in which the timescale of $\bm{\xi}$ is slower than $\bm{v}$, so that $\bm{v}$ relaxes to its stationary state in a time-frame in which the external modulation can be considered constant.
\begin{figure}[t]
\centering
\includegraphics[width=14cm]{figS2.png}
\caption{\footnotesize Comparison between the analytical expressions of the joint probability distribution in Eq. (\ref{eqn:pstat_joint}) and its factorization $p(v_i, v_j)$. (a) The factorized distribution. (b) The joint distribution. (c) Comparison between the two in a log-plot. (d) Relative difference between the two with respect to the factorized distribution. (e-f) Comparison between the joint and the factorized distributions, with the same parameters except for $\mathcal{D}^* = 5$.}
\label{fig:probability_joint_factor}
\end{figure}
Let us now describe the particular choices of the main text, where
\begin{equation}
\label{eqn:OUprocess_vi}
\frac{dv_i(t)}{dt} =-\frac{1}{\gamma_i} v_i(t) + \sqrt{\mathcal{D}(t)} \eta(t)
\end{equation}
is the process that generates $\bm{v}$ and the single-parameter external modulation is
\begin{equation}
\mathcal{D}(t) = \begin{cases}
\mathcal{D}^* & \text{if} \quad D(t) \le \mathcal{D}^* \\
D(t) & \text{if} \quad D(t) > \mathcal{D}^*
\end{cases}
\end{equation}
where
\begin{equation}
\label{eqn:OUprocess_D}
\frac{dD(t)}{dt} =-\frac{1}{\gamma_{D}} D(t)+ \sqrt{\theta} \eta(t).
\end{equation}
For the above considerations to hold, we assume that $\gamma_{D} \gg \gamma_i$. This is simply a time-scale separation limit. In this limit, the process of $v_i$ reaches stationarity much faster than the process of $\mathcal{D}$, thus the overall stationary distribution is the stationary distribution $p(v_i | \mathcal{D})$ averaged over the stationary distribution $p(\mathcal{D})$. Hence we have
\begin{align}
\label{eqn:pstat_D}
p(\mathcal{D}) = \frac{1}{2}\left[1+ \text{Erf}\left(\frac{\mathcal{D}^*}{\sqrt{\theta \gamma_D}}\right)\right]\delta(\mathcal{D}-\mathcal{D}^*) + \frac{H(\mathcal{D}-\mathcal{D}^*)}{\sqrt{\pi \theta \gamma_D}}e^{-\frac{\mathcal{D}^2}{\theta \gamma_D}},
\end{align}
where $H$ is the Heaviside step function. Then, the computation of the stationary probability distributions is quite easy. The single-unit probability is given by
\begin{align}
\label{eqn:pstat_v}
p(v_i) & = \int d\mathcal{D} \, p(v_i | \mathcal{D}) \, p(\mathcal{D}) = \int d\mathcal{D} \, \frac{1}{\sqrt{\pi \gamma_i \mathcal{D}}}\, e^{-\frac{v_i^2}{\gamma_i \mathcal{D}}}\, p(\mathcal{D}) \nonumber\\
& = \frac{1+ \text{Erf}\left(\frac{\mathcal{D}^*}{\sqrt{\theta \gamma_D}}\right)}{2\sqrt{\pi \mathcal{D}^* \gamma_i}}e^{-\frac{v_i^2}{\mathcal{D}^*\gamma_{i}}} + \frac{1}{\sqrt{\pi^2 \theta \gamma_D\gamma_{i}}}\int_{\mathcal{D}^*}^\infty \frac{dD}{\sqrt{D}}e^{-\left[\frac{D^2}{\theta \gamma_D} + \frac{v_i^2}{D\gamma_{i}}\right]}
\end{align}
and, in principle, we can compute the probability distributions for any number of variables $\{v_i\}$ in the same way.
We note that although before the marginalization the two-point probability distribution is factorizable
\begin{equation*}
p(v_i, v_j, \mathcal{D}) = p(v_i, v_j | \mathcal{D}) p(\mathcal{D}) = p(v_i|\mathcal{D})p(v_j|\mathcal{D})p(\mathcal{D}),
\end{equation*}
the marginalization itself breaks the factorization, namely
\begin{equation*}
\int d\mathcal{D} \, p(v_i, v_j,\mathcal{D}) \ne \left[\int d\mathcal{D} \, p(v_i, \mathcal{D})\right]\left[\int d\mathcal{D} \, p(v_j, \mathcal{D})\right].
\end{equation*}
Hence, we can write down the two-unit joint probability as
\begin{align}
\label{eqn:pstat_joint}
p(v_i, v_j) = \frac{1+ \text{Erf}\left(\frac{\mathcal{D}^*}{\sqrt{\theta \gamma_D}}\right)}{2\pi \mathcal{D}^* \sqrt{\gamma_i\gamma_j}}e^{-\frac{1}{\mathcal{D}^*}\left(\frac{v_i^2}{\gamma_i}+\frac{v_j^2}{\gamma_i}\right)} + \frac{1}{\sqrt{\gamma_i\gamma_j\gamma_D \pi^3\theta}} \int_{\mathcal{D}^*}^\infty \frac{dD}{D} e^{-\frac{1}{D}\left(\frac{v_i^2}{\gamma_i}+\frac{v_j^2}{\gamma_i}\right)}e^{-\frac{D^2}{\theta \gamma_D}}
\end{align}
where the second term is not factorizable.
From these probabilities we can immediately show that
\begin{equation}
\label{eqn:uncorrelated}
\ev{v_i v_j} - \ev{v_i}\ev{v_j} = 0 \quad \forall i \ne j
\end{equation}
and, in general, all the expectation values where a variable $v_i$ appears an odd number of times vanish. {This follows from the simple fact that, for instance,}
\begin{align*}
\ev{v_i v_j} = \int dv_i \,dv_j\, p(v_i, v_j) v_i\, v_j = \int d\mathcal{D} \, p(\mathcal{D}) \left(\int dv_i \, v_i\,p(v_i|\mathcal{D})\right)\left(\int dv_j\, v_j \,p(v_j|\mathcal{D})\right) = 0 \quad \forall i \ne j
\end{align*}
{since $\int dv_i \,v_i \,p(v_i | \mathcal{D}) = 0$. In fact, let us note that this is not a general feature of these kind of models, but it is rather a property inherited from the choice of an OU process for $v_i$ given $\mathcal{D}$, for which all the odds moment vanish.} The diagonal entries of the covariance matrix, on the other hand, are given by
\begin{equation*}
\langle v_i^2 \rangle = \frac{\gamma_i \mathcal{D}^*}{2}\left[1+ \text{Erf}\left(\frac{\mathcal{D}^*}{\sqrt{\theta \gamma_D}}\right)\right] + \sqrt{\frac{\gamma_i^2 \theta \gamma_D}{16 \pi}}e^{-\frac{(\mathcal{D}^*)^2}{\theta \gamma_D}}
\end{equation*}
Thus, in this particular setting, the units are uncorrelated and have a variance that is independent on the other units. At the same time, however, it is clear that in general $p(v_i, v_j) \ne p(v_i)p(v_j)$ so while uncorrelated the variables are not independent.
\begin{figure}[t]
\centering
\begin{minipage}{1\textwidth}
\centering
\includegraphics[width = 10cm]{figS3.png}
\end{minipage}
\caption{\footnotesize Comparison between the analytical expressions of the joint probability distribution $p(v_i, v_j)$ given by Eq. (\ref{eqn:pstat_joint}) and the one obtained from with $10^3$ simulations. We also show the corresponding results for the factorized probabilities. The blue line corresponds to the analytical expression of $p(v_i, v_j)$. The corresponding dots represent the histogram of the distribution obtained from $10^3$ simulations, and the semitransparent filled areas represent one standard deviation from this estimate. Similarly, the gray dashed line represent the analytical expression of $p(v_i)p(v_j)$. (a) Section along the $v_j$ direction for small $v_j$, so that we are looking at the tails of the distribution. Even though the estimate along the tails is noisy, we clearly see that the estimate from the simulations lies along the analytical prediction. (b) As before, but for higher $v_j$. (c-d) As before, but with values of $v_j$ close to zero so we look at the bulk of the distribution near its peak. Even though joint probability and its factorization now are more similar, once more the estimate from the simulation match the analytical expression $p(v_i, v_j)$.}
\label{fig:probability_data_comparison}
\end{figure}
\section*{C. Testing the predictions of the extrinsic model}
\noindent We now show that simulations of the extrinsic model described by Equation \ref{eqn:OUprocess_vi} agree with the analytical results. Whenever not specified, we assume that the parameters of the model are given by $\mathcal{D^*} = 0.3$, $\theta = 1$, $\gamma_D = 10$ together with $\gamma_i = 0.1$, $\gamma_j = 0.5$. Thus, we are in the limit of timescale separation considered in the main text.
Let us begin with Figure \ref{fig:probability_single}. Albeit trivial, we first check in Figure \ref{fig:probability_single}a that the stationary distribution of $\mathcal{D}$ is indeed the one of Eq. (\ref{eqn:pstat_D}) and in Figure \ref{fig:probability_single}b that the stationary distribution of a single $v_i$ does correspond to the analytical expression of Eq. (\ref{eqn:pstat_v}). If we compare this distribution to a standard distribution of an Ornstein-Uhlenbeck process with a diffusion coefficient equal to the mean $\ev{\mathcal{D}(t)}$ \cite{Note1} we immediately see that the distribution of our model is considerably more peaked around zero and displays longer tails. Indeed, one expects that due to the fact that $\mathcal{D}^* < \ev{\mathcal{D}(t)}$ the system tends to wander more easily close to zero, especially in the time windows where the diffusion coefficient is constant and equal to $\mathcal{D}^*$. At the same time, the fact that $\mathcal{D}(t)$ can change in time favors the presence of values of $v$ that are larger in absolute value, which is the mechanism at the origin of the bursty behavior seen in the main text.
In Figure \ref{fig:probability_joint_factor} we look instead at the properties of the joint probability distribution $p(v_i, v_j)$. The most natural quantity to compare this distribution with is its factorization $p(v_i)p(v_j)$, which is equivalent to ignoring the feedback effects between $v_i$ and $v_j$ due to the shared extrinsic modulation of $\mathcal{D}(t)$. Since we are setting $\mathcal{D}^* = 0.5$, we expect that these effects are going to be particularly relevant for the dynamics of the model. In particular, in Figure \ref{fig:probability_joint_factor}c-d we see that the most important differences between the two occur in the tails of the two-dimensional distribution, with the joint distribution typically showing dramatically longer tails. This translates to the fact that far-from-zero values of the two variables can occur more easily at the same time.
The situation is completely reversed when we increase $\mathcal{D}^*$. In Figure \ref{fig:probability_joint_factor}e-f we see that for $\mathcal{D}^* = 8$ the joint probability distribution and the factorized distribution are almost indistinguishable. Hence, this example shows explicitly that if $\mathcal{D}^*$ is high enough the dependence induced by the extrinsic modulation vanishes.
Let us keep focusing on the case $\mathcal{D}^* = 0.5$ for the time being. In Figure \ref{fig:probability_data_comparison}a-d we compare the one-dimensional sections of the analytical expression of the joint probability distribution with the results of $10^3$ simulations of the model, together with the sections of the factorized distribution. The joint distribution estimated from the simulation matches particularly well the analytical prediction. Once more, and perhaps more clearly, in Figure \ref{fig:probability_data_comparison}a-b we see the stark difference that emerges along the tails between the joint probability distribution and its factorization. Interestingly, panel (c) and panel (d) show that the situation in the bulk of the distribution is reversed with respect to the tails and now the joint probability distribution is more peaked with respect to is factorization, albeit only slightly. That is, the modulation in the low $\mathcal{D}^*$ regime favors both large values of $v_i$ and $v_j$ and values very close to zero.
Overall, this brief analysis showed us how the bursty behavior that we see for small values of $\mathcal{D}^*$ emerges from the underlying probability distributions, which in turn emerge from the simple marginalization that occurs in Eq. (\ref{eqn:joint_prob_model}). Similar arguments, albeit impractical, could be carried out for the probability distributions beyond the two-point ones. In a sense, one could argue that the fundamental properties of the model are inherited from the fact that there are some unobserved physical quantities, and these are the quantities that drive the global response of the single variables.
For the specific case of a double Ornstein-Uhlenbeck process we chose, the net effect of the marginalization is the widening of the tails of both the one-point $p(v)$ and the two-point $p(v_i, v_j)$ probability distributions when $\mathcal{D}^*$ is small enough. As we increase $\mathcal{D}^*$, this effect becomes less and less important until it is completely negligible. In this sense, we can effectively think of $\mathcal{D}^*$ as a control parameter that changes the qualitative behavior of the system. Most importantly, the fact that the tails of the joint probability distribution are wider when $\mathcal{D}^*$ is small reflects dynamically in the emergence of a non trivial coordination between the variables, from which in turn power-law avalanches emerge.
\begin{figure}[t]
\centering
\includegraphics[width=16cm]{figS4.png}
\caption{\footnotesize Comparison between (a) the correlations of the data and (c) the correlations of the interacting model after solving the inverse problem. Panel (b) shows the inferred matrix $\tilde{A}$, which is independent on the parameters of the model.}
\label{fig:inverse_problem}
\end{figure}
\section*{D. The multivariate Ornstein-Uhlebeck process}
\noindent The process studied in the main text is a multivariate Ornstein-Uhlebeck process \cite{gardiner2004handbook} of the form
\begin{equation}
d \boldsymbol{v}(t)=-A \boldsymbol{v}(t) d t+ B(t)d \boldsymbol{W}(t),
\label{Lyapu_EQ}
\end{equation}
where in our case $B(t)$ is a diagonal matrix whose diagonal elements are given by $\sqrt{\mathcal{D}(t)}$. In the case of decoupled units, which we use to model the extrinsic activity, the matrix $A$ is again diagonal with entries $A_{ij} = \delta_{ij}/\gamma_i$.
The formal solution of this stochastic process is given by
\begin{align*}
\boldsymbol{v}(t) = & \exp \left(-A t \right) \boldsymbol{v}(0) +\int_{0}^{t}\exp \left[-A\left(t-t^{\prime}\right)\right] B\left(t^{\prime}\right) d \boldsymbol{W}\left(t^{\prime}\right).
\end{align*}
In particular, since in our case the matrix $B$ is on itself a stochastic variable, from now on we will focus on the original variables conditioned on $B$, which in turn we assume follows its own stationary distribution $p(B)$. We call these conditioned variables $\boldsymbol{v}(t,B)$.
If the eigenvalues of $A$ have all positive real parts, a stationary solution of Eq. \ref{Lyapu_EQ} exists and it is of the form \cite{Note2}
\begin{align*}
\boldsymbol{v}_{s}(t,B)=\int_{-\infty}^{t} \exp \left[-A\left(t-t^{\prime}\right)\right] B d \boldsymbol{W}(t^{\prime}).
\end{align*}
Hence, the stationary covariance matrix conditioned on $B$ reads as
\begin{align*}
\sigma(B) & = \left\langle\boldsymbol{v}_{s}(t,B), \boldsymbol{v}_{\mathbf{s}}^{\mathrm{T}}(t,B)\right\rangle \\
& = \int_{-\infty}^{t} d t^{\prime} \exp \left[- A (t- t^{\prime})\right] B B^{\mathrm{T}} \exp \left[ A^{\mathrm{T}} (t -t ^\prime) \right]
\end{align*}
which solves the algebraic equation
\begin{align*}
A \sigma(B)+\sigma(B) A^{\mathrm{T}} = M,
\end{align*}
where the matrix $M$ is given by
\begin{align*}
M = & \int_{-\infty}^{t} A \exp \left[-A\left(t-t^{\prime}\right)\right] B B^{\mathrm{T}} \exp \left[-A^{\mathrm{T}}\left(t-t^{\prime}\right)\right] d t^{\prime} +\int_{-\infty}^{t} \exp \left[-A\left(t-t^{\prime}\right)\right] B B^{\mathrm{T}} \exp \left[-A^{\mathrm{T}}\left(t-t^{\prime}\right)\right] A^{\mathrm{T}} d t^{\prime}.
\end{align*}
We can rewrite it as
\begin{align*}
M = \int_{-\infty}^{t} \frac{d}{d t^{\prime}}\left\{\exp \left[-A\left(t-t^{\prime}\right)\right] B B^{\mathrm{T}} \exp \left[-A\left(t-t^{\prime}\right)\right]\right\} d t^{\prime}
\end{align*}
Carrying out the integral, we find that the lower limit vanishes due to the assumed positivity of the eigenvalues of $A$ and hence only the upper limit remains, giving
\begin{align}
A \sigma(B)+\sigma(B) A^{\mathrm{T}}=B B^{\mathrm{T}},
\end{align}
which is a continuous Lyapunov equation \cite{Lyapunoveq} for the covariance matrix $\sigma(B)$ \cite{Gilson2016}. Then, we only need to marginalize over $B$ and we obtain the equation for the covariance matrix of our original variables $\boldsymbol{v}_s(t)$
\begin{equation}
\label{lyapunov}
A \sigma +\sigma A^{\mathrm{T}}= Q,
\end{equation}
where for the sake of brevity we call $\sigma$ the covariance matrix of the original variables and $Q$ is a diagonal matrix whose elements are given by
\begin{equation}
Q_{ij} = \delta_{ij}\int_{\mathcal{D}^*}^\infty \mathcal{D} \,p(\mathcal{D})\, d\mathcal{D} := \delta_{ij} f (\mathcal{D}^*, \gamma_D, \theta).
\label{Q_eq}
\end{equation}
Then, taking the transpose of Eq. (\ref{lyapunov}), and assuming that the matrix A is symmetric, we end up with a Lyapunov equation for the matrix A. Thus, the inverse problem of reconstructing the connectivity matrix from the data reduces to solving numerically $\sigma A + A \sigma = Q$, starting from the correlation matrix of the data and given a set of parameters for the model.
Hence we end up with a model
\begin{equation}
\dot{v}_i(t) = -\sum_j A_{ij}v_j(t) + \sqrt{\mathcal{D}(t)} \xi_i(t)
\end{equation}
where $A_{ij}$ depends on the the parameters of the stochastic modulation $(\mathcal{D}^*, \gamma_D, \theta)$. Notice that if we write $\tilde{A}_{ij} = A_{ij}/f(\mathcal{D}^*, \gamma_D, \theta)$ we need to solve the Lyapunov equation $\sigma \tilde A + \tilde A\sigma = \mathbb{1}$ that only depends on $\sigma$, the correlation matrix of the data. If we introduce $\tilde{\mathcal{D}} = \mathcal{D}/f$ and $\tilde{v}_i = v_i/\sqrt{f}$ we can write
\begin{equation*}
\dot{\tilde v}_i(t) = -\sum_j \tilde{A}_{ij} \tilde v_j(t) + \sqrt{\tilde{\mathcal{D}}(t)} \xi_i(t)
\end{equation*}
and clearly
\begin{equation*}
\ev{\tilde v_i \tilde v_j} - \ev{\tilde v_i}\ev{\tilde v_j} = \frac{\ev{ v_i v_j} - \ev{v_i}\ev{v_j}}{f}
\end{equation*}
so the correlation between $\tilde v_i$ and $\tilde v_j$ is proportional to the correlation between $v_i$ and $v_j$. This means that at different $(D^*, \theta, \gamma_D)$ we simply find a rescaled interaction matrix $A_{ij}$, but the scaling of correlation length does not change.
{The stationary probability distribution solution of the interacting model can be written, in general, as}
\begin{equation}
p(v_1, \dots, v_N) = \frac{1 + \mathrm{Erf}\left[\frac{\mathcal{D}^*}{\sqrt{\theta\gamma_D}}\right]}{2\sqrt{(\pi\mathcal{D}^*)^N \det \Sigma}} e^{-\frac{1}{\mathcal{D}^*}\vb{v}^T {\Sigma}^{-1}\vb{v}} + \frac{1}{\sqrt{(\gamma_D\theta)^N \pi^{N+1}\det \Sigma}}G_N\left(\frac{\mathcal{D}^*}{\sqrt{\theta \gamma_D}}, \frac{\vb{v}^T {\Sigma}^{-1}\vb{v}}{\sqrt{\theta \gamma_D}}\right)
\end{equation}
{where $\left({A}{\Sigma} + {\Sigma}{A}^T \right)/2 = \mathbb{1}$ and}
\begin{equation}
G_N(\alpha, \beta) = \int_{\alpha}^{\infty} \frac{dx}{x^{N/2}}e^{-\frac{\beta}{x}-x^2}.
\end{equation}
{In general we could define a multivariate information between these $N$ variables and the results of the main text would not change. In practice, however, it is very hard to perform the related numerical integration if $N$ is large. Therefore, in the main text we show an example of two interacting units that interact through the matrix element $\tilde{A}_{12} = 2$. The mutual information of this model receives a distinct (and constant) contribution from the interaction matrix and from the modulation induced by $\mathcal{D}^*$. In fact, if we take $A = \mathrm{diag}(\gamma_1, \dots, \gamma_N)$, i.e., we consider the extrinsic model, the only the contribution to the mutual information comes from the external modulation, whereas the addition of interactions simply shift the mutual information by a constant value at all $\mathcal{D}^*$.}
\section*{E. Scaling of the correlation length}
\noindent To study if the correlations of the system under study exhibit critical-like properties we determine the correlation length $\xi$ of the system at various system’s sizes. The correlation length can be defined as the average distance at which the correlations of the fluctuations around the mean crosses zero \cite{Cavagna2010}, and it is known to diverge at criticality in the thermodynamic limit \cite{binney1992}.
For finite systems, however, this behavior can be demonstrated by showing that the correlation length grows with system size. Thus, we first compute for each time series their fluctuations around the mean, namely
\begin{equation}
\Tilde{v}_i(t) = v_i(t) - \frac{\sum_{i= 1}^Nv_i(t)}{N}
\end{equation}
where $1/N \sum_{i= 1}^Nv_i(t)$ is the mean activity, i.e. the mean at time $t$ computed over the $N$ channels, and $v_i$ is the activity at channel $i$. By definition the mean of the fluctuations vanishes, i.e. $\frac{\sum_{i=1}^N \Tilde{v}_i(t)}{N} = 0, \forall t$. As said, different sizes (portions of the array) of the system are selected and, importantly, the mean activity is computed for each system size, considering the channels inside the observation window.
In particular, since the maximum system size (corresponding to $N$ channels) both in our experimental data and in our model is fixed, we investigate how $\xi$ changes with system sizes corresponding to different subsamples from the multi-array probe \cite{Ribeiro2020}. In fact, through simulations on control models that display a critical point \cite{martin2020}, subsampling has been shown to be practically equivalent to consider systems of different sizes. We assume that the units of our model have the same topology of our data, i.e. we assume that the units are placed as the channels in the $55 \times 4$ array of our experimental setup.
\begin{figure}[t]
\centering
\includegraphics[width=13cm]{figS5.png}
\caption{Avalanche statistics in four different rats. (1a-4a) The distribution of the avalanches' sizes is consistently a power-law, with an exponent that slightly depends on the single rat. (1b-4b) The avalanche durations are once more power-law distributed in all rats with some variability in the exponents, even though the range accessible with the experimental setup only covers two decades. (1c-4c) The crackling-noise relation, however, is consistently satisfied in each rat.}
\label{fig:resting_avalanches}
\end{figure}
Since in our case the array shape is rectangular, we consider the number of rows as the relevant dimension and build subsampled systems of size $L \times 4$, with $L$ that decreases from the maximum of $55$ channels up to $5$ channels.
Next, for each system's subset, we compute the average correlation function of the fluctuations between all pairs of channels separated by a distance $r$,
\begin{equation}
\ev{C(r)} = \ev{\frac{\ev{\left(\tilde{v}_{i}-\overline{\tilde{v}}_{i}\right)\left(\tilde{v}_{j}-\overline{\tilde{v}}_{j}\right)}_{t}}{\sigma_{\tilde{v}_{i}} \sigma_{\tilde{v}_{j}}}}_{i, j}
\end{equation}
where $\ev{\cdot}_t$ stands for the average over time, $\ev{\cdot}_{i,j}$ the average over all pairs of channels separated by a distance $r$ and
\begin{gather*}
\overline{\tilde{v}}_{i} =\frac{1}{T} \sum_{t=1}^{T} \tilde{v_i}\left(t\right) \\
\sigma_{\tilde{v}_{i}}^{2} =\frac{1}{T} \sum_{t=1}^{T}\left(\tilde{v_i}\left(t\right)-\overline{\tilde{v}}_{i}\right)^{2}
\end{gather*}
with $T$ is the length of the time series. Then $\xi$ is computed as the zero of the correlation function $C(r = \xi) = 0$. To reduce the noise effects, results were averaged across all possible sub-regions for any given size. Then the $\xi$ are plotted against the relative system size $L$ and the slope of the fit is obtained through a linear regression.
\section*{F. Avalanches statistics in LFPs data}
\subsection*{1. LFPS peaks detection and avalanche analysis}
{
For the detection of LFP events, the standard deviation (SD) and the mean of the signal was computed for each channel. In order to distinguish real events from noise, a three SD threshold was chosen basing on the distribution of the signal amplitudes which significantly deviated from a Gaussian best fit above that threshold. Both negative and positive LFPs (i.e., nLPFs adn pLFPs, respectively) were considered as events in accordance with previous works \cite{Shew2015}. Within our specific experimental settings, one reason is that across the depth of the cortex there are polarity changes of the LFP signal because of compensatory capacitive ionic currents particularly along dendrites of pyramidal cells \cite{Buzsaki}. Since in our experiments electrodes span multiple cortical layers, both nLFPs and pLFPs were found and detected. Moreover, alternatively, pLFPs can be related to activation of populations of inhibitory neurons leading to inhibitory outward postsynaptic currents, which also justifies their inclusion in the events count. For detection, each deflection was considered terminated only after it crossed the mean of the signal. After defining events of activity, the avalanche statistics are analyzed through the usual methods of the literature. In particular, the data are temporally binned, avalanches are defined as sequences of bins that present activity, and an avalanche ends once an empty bin is found - the temporal bin chosen is the average inter-event interval, as it is typically done in avalanche analysis \cite{beggs2003}. Then, the distribution $p(s)$ of the avalanches sizes - the number of events in each avalanche - and of the avalanche duration $p(T)$ are computed and fitted using a corrected maximum likelihood method \cite{Gerlach2019, Mariani2021b}. In particular, the power law distribution is tested following the method proposed in \cite{Clauset, Gerlach2019}. Avalanche sizes and lifetimes are fitted with discrete power laws $p(y; \alpha) = \frac{y^{-\alpha}}{\sum_{x = x_{min}}^{x = x_{max}}x^{-\alpha}}$. The parameter $x_{max}$ is set to the maximum observed size or duration. $x_{min}$ is selected as the one that minimizes the Kolmogorov-Smirnov distance (KS) between the cumulative distribution function (CDF) of the data and the CDF of the theoretical distribution fitted with the parameter that best fits the data for $y \geq x_{min}$ \cite{Clauset}.
To assess goodness-of-fit we compared the experimental data against 1000 surrogate datasets drawn from the best-fit power law distribution with the same number of samples as the experimental dataset. The p-value of the power-law fit was defined as the fraction of the KS statistics of the surrogates which were greater than the KS statistic for the experimental data. The data were considered power law distributed if the the p-value was greater than 0.1.
We also take into account the fact that while maximum likelihood methods rely on the independence assumption, actual data often display correlations, and this may lead to false rejection of the statistical laws \cite{Gerlach2019}. As the author of \cite{Gerlach2019} suggest, before performing the fit and assessing p-values, we undersample the data in order to decorrelate them, by estimating the time $\tau^*$ after which two observations (e.g., the avalanche sizes) are independent from each other, as done in \cite{Gerlach2019, Mariani2021b}.
}
\subsection*{2. Avalanches statistics}
\noindent Here we report the avalanches statistics from the other rats that we analyzed. The avalanches statistics is computed by considering all available the $20$ trials of basal activity for each rat, that are $7.22 \mathrm{s}$ long. Results are reported in Figure \ref{fig:resting_avalanches}. Inter-rat variability is present with respect to avalanche exponents, and is expected as found in previous experiments \cite{Fontenele2019, Shew2015, bansal2020}. Moreover, a theoretical explanation for the difference in avalanche exponents has been also recently proposed as signature of quasi-criticality \cite{Fosque2021}. However, the fundamental point is that the crackling-noise relation is always verified compatibly with the experimental errors, a feature that is usually considered a hallmark of criticality \cite{Fontenele2019}.
\begin{figure}[t]
\centering
\includegraphics[width=10cm]{figS6.png}
\caption{Collapse of the average profile of avalanches of varying duration in the extrinsic model, for the low $\mathcal{D}^*$ regime. (a) Profile of the avalanches before the rescaling. (b) If we rescale with an exponent $\delta \approx 1.33$, which is remarkably close to the one found in the main text through the crackling-noise relation, we obtain an optimal collapse onto the same scaling function.}
\label{fig:collapse}
\end{figure}
\section*{G. Additional avalanches statistics in the model}
\noindent Another signature of criticality is the collapse of the average profile of avalanches of widely varying duration onto a single scaling function. For avalanches of duration $T$ we can write down the average number of firing at time $t$ as $s(t,T) = T^{\delta-1} F(t/T)$ where F is a universal scaling function that determines the shape of the average temporal profile. $\ev{S(T)}$ and $s(t,T)$ are related by $\ev{S(T)} = \int_0^Ts(t,T)dt$. At the critical point we expect that plots of $t/T$ versus $s(t,T)T^{1-\delta}$ for different $T$ will collapse onto the same universal scaling function \cite{Sethna}.
Thus, finding the exponent for which the goodness of the collapse is higher provides another way to estimate $\delta$. For testing the avalanche shape collapse, we used the methodology introduced in \cite{Timme2016}. To determine the quality of the collapse, the averaged and rescaled avalanche profiles of different lifetimes $F(t/T) = T^{1-\delta}s(t/T,T) $ are first linearly interpolated at $1000$ points along the scaled duration. The variance across the different $F(t/T)$ is calculated at each interpolated point, and the shape collapse error $\epsilon(\delta)$ is then defined as the mean variance divided by the squared span of the avalanche shapes, where the span equals the maximum minus the minimum value of all rescaled avalanche profiles. In the presented analysis, avalanche shapes of $T>10$ bins with at least 10 samples were used.
The collapse has been tested on the extrinsic model in the case of low $\mathcal{D}^*$, and the results are plotted in Figure \ref{fig:collapse}. We find that the exponent that minimizes $\epsilon(\gamma)$ turns out to be $\approx 1.33$, close to the estimates of $\delta $ found through the linear fit of average size given duration and through the prediction of the crackling-noise relation in the main text. Again, it is also close to the apparently super-universal exponent found in \cite{Fontenele2019} and in \cite{Friedman2012}.
\section*{H. The crackling noise relation and its link with criticality}
\label{sec:crackling}
{The so called ``crackling noise" relation among avalanches' distributions' exponents \cite{Sethna} is usually considered a strong sign of criticality \cite{Fontenele2019}, however it is possible to show that it can be valid under general conditions that do not necessarily require criticality. Indeed, following arguments similar to the one proposed in \cite{Scarpetta}, if we express the size $S$ of an avalanche as a function of its duration $T$ by assuming that $S \sim T^\delta$, i. e. we assume that fluctuations in the size fixed the duration can be disregarded, the following equivalence holds:}
\[p(S(T)) \left|\frac{dS}{dT}\right|dT = p(T)dT\]
{Then, if we assume that both $p(S(T))$ and $p(T)$ are described by power laws, with exponents respectively $\tau$ and $\tau_t$, we see that the above equation reduces to}
\[(T^\delta)^{-\tau} \delta T^{\delta-1} = T^{-\tau_t}\] {that leads ideed to the relation} \[\delta = \frac{\tau_t -1}{\tau -1}.\]
{Thus, to derive the above relation we had to assume that $p(S)$ and $p(T)$ are well described by power laws, and that fluctuations in the size given the duration are negligible ($S \sim T^\delta$). These features are certainly proper of critical points (where moreover the exponent $\delta$ can be expressed as a combination of other critical exponents \cite{Sethna}), but can be found to be valid also away from criticality, as our modeling framework shows.}
\section*{I. Alternative models}
{
The multivariate Ornstein-Uhlenbeck (MOU) process has been previously used in the literature \cite{Saggio2016, Arbabyazd2021, Gilson2019, Gilson2016} to model neural activity at large scale (e.g. fMRI signals). However, it is a paradigmatic model which, albeit simple enough to allow for an analytical treatment, does not account for many biological aspects. Thus, in order to show that our results generalize to more complex situations, here we introduce a more biologically sound model that can describe the population-level neural activity measured by our LFPs data, the Wilson-Cowan model. It includes both excitatory and inhibitory synapses and non-linearities in the transfer function. We extend the framework of the extrinsic model presented in the main text, and we also provide a possible biological generative mechanism for the external time varying modulation.}
\begin{figure}
\centering
\includegraphics[scale = 0.45]{figS9.png}
\caption{{(a) External input $h = \frac{E^{(h)} + I^{(h)}}{2}$ in a balanced state ($\omega^{(h)}_{E} = 50.05$ and $\omega^{(h)}_{I} = 49.05$) with high level of noise amplitude $\sigma^{(h)}$ (in orange), and low level of noise amplitude (in blue). When the noise level is high enough bursty behavior is present, that is reflected also on the firing rate $\frac{E_i + I_i}{2}$ of the units that receive $h$ as input (Figure (b), in orange). Instead, if the noise level is low, $h$ converges to the up state predicted by the model in the absence of noise (Fig. (a), blue), hence also the firing rate of the units (Figure (b), blue) doesn't display transient behaviors. } }
\label{fig:avawilson}
\end{figure}
{The model considered is a stochastic Wilson-Cowan model \cite{diSanto2018, Benayoun}, which describes the dynamics of the activity, i.e., the density of active neurons, for the two subpopulations of excitatory (E) and inhibitory (I) neurons. Each unit $i$, i.e., each neural population, is now represented by a subpopulation of excitatory neurons and a subpopulation of inhibitory neurons, whose activity evolves according to the following set of stochastic equations:}
\begin{equation}
\label{eq:wilson}
\begin{cases}
\frac{dE_i}{dt} = -\alpha E_i + (1-E_i) f(\omega_{E} E_i - \omega_{I} I_i + h) + \sqrt{(\alpha E_i + (1-E_i) f(\omega_{E} E_i - \omega_{I} I_i + h))}\eta_{E_i}\\
\frac{dI_i}{dt} = -\alpha I_i + (1-I_i) f(\omega_{E} E_i - \omega_{I} I_i + h) + \sqrt{(\alpha I_i + (1-I_i) f(\omega_{E} E_i - \omega_{I} I_i + h))}\eta_{I_i}
\end{cases}
\end{equation}
{where $\alpha$ is the rate of spontaneous activity decay, $\omega_{E, I}$ are the synaptic efficacies, $s = \omega_{E} E_i - \omega_{I} I_i + h$ is the averaged incoming current, $\eta_{E,I}$ are uncorrelated Gaussian white noises, with network-size dependent amplitude $\sigma \sim \frac{1}{\sqrt{\mathcal{N}}}$ \cite{Benayoun}, interpreted in the Ito sense. The incoming current is modeled as the sum of all synaptic inputs plus an external input current $h$, filtered by a sigmoid response function}
\begin{equation}
\begin{cases}
f(s) = \tanh(s) \hspace{0.22cm} s \geq 0\\
f(s) = 0 \hspace{1.2cm} s < 0.
\end{cases}
\end{equation}
{We consider a case in which the units are inhibition dominated, i.e. when $\omega_{I} > \omega_{E}$ ($\omega_{I} = 7$, $\omega_{E} = 6.8$). In this situation the noise level is not going to produce significant changes in the dynamics, and the noise amplitude of $\eta_{E_i}$, $\eta_{I_i}$ is set to a low value, $\sigma = 10^{-4} \, \forall i$. Similarly, the value of $\alpha$ is not going to affect the inhibition dominated dynamics, and it is set to $\alpha = 1$. For each unit, we consider the quantity $\Sigma_i = (E_i + I_i)/2$, that describes the firing rate of the neural population $i$.}
{As in the main text, we consider an external, time-varying input that enters the model through $h$. As a potential candidate for a biological realization of this external, stochastic driving, we consider the effective input that comes from other, yet unobserved, neural populations. Hence, we model $h$ as the firing rate $h = (E^{(h)} + I^{(h)})/2$ of another Wilson-Cowan model. In particular, we take it to be in the balanced state $0 \ll \omega^{(h)}_0 \ll \omega^{(h)}_{\rm sum} = \omega^{(h)}_E + \omega^{(h)}_I$ and with a sufficiently high level of noise, so that the resulting firing rate spends most of the time close to the down state, while showing frequent bursts of activity \cite{Benayoun}. Note that, since the balanced state requires that $\omega_0 \ll \omega_S$, it does not coincide with the ``critical" state $\omega_0 = 0$ - being balanced is not a sufficient condition and neither necessary for being critical.}
{We end up with a time-varying modulation}
\begin{equation}
\label{eq:wilsoninput}
\frac{dh}{dt} = \frac{d}{dt}\left[\frac{E^{(h)} + I^{(h)}}{2}\right]
\end{equation}
{that is qualitatively similar to the one considered in the main text, alternating periods of silence and bursts, but which is more realistic from a biological point of view. In particular, we choose $\omega^{(h)}_{E} = 50.05$, $\omega^{(h)}_{I} = 49.95$, $h^{(h)} = 10^{-3}$, $\alpha^{(h)} = 0.1$, and, importantly, $\sigma^{(h)}$, the amplitude of the noise for both the excitatory and the inhibitory populations, is increased to a relatively high value, $1.2 \times 10^{-3}$ (i e. corresponding to a finite but still large network size), so that the up state can be destabilized by the noise. Intuitively, we are considering a case in which $i = 1, \dots, N$ populations of neurons evolve according to the WC model in the inhibition dominated phase and are all receiving the same input by another population of neurons in a balanced state.}
{From each firing rate $\Sigma_i$ we generate spike trains and analyze avalanches in the usual way, by temporal binning the events with the average inter-event interval (see the main text). Again, we find that avalanche sizes and durations are power law distributed, satisfying the crackling noise relation (Fig. \ref{fig:avawilson}). We note that we do not expect to find the exponents of avalanches' sizes and durations found in \cite{diSanto2018}, even though $h$ itself is a WC model. In fact, in \cite{diSanto2018} avalanches were found by considering the excursions of $h$ over a small threshold, in contrast to the study of spike trains carried out here. Moreover, and most importantly, the external input $h$ that we are considering feeds the units $i$ passing through the non linearity of $f(s)$, hence $h$ and $\Sigma_i$ do not coincide.}
{Thus, we have shown that the results presented in the main text, and in particular the ones of the extrinsic model, extend well beyond the paradigmatic example of a multivariate Ornstein-Uhlenbeck process with a shared external input. With respect to the approach presented in this section, the MOU has the significant advantage of being amenable to analytical treatments and to the inclusion of interactions reconstructed from the data. However, the WC model has a deeper biological relevance and proves that the mechanism proposed here for avalanche generation appears to be quite general. Finally, we note that $h = (E^{(h)} + I^{(h)})/2$ is only one of the possible choices that could be considered for the external modulation. Other candidates include a simple Ornstein Uhlenbeck process, that is often used to model noisy neural activity \cite{ricciardi1979ornstein} or the Brunel's model in the Synchronous Irregular phase \cite{Brunel}, which is known to be able to reproduce avalanching behavior \cite{Touboul2017}.}
\begin{figure}
\centering
\includegraphics[scale = 0.3]{figS7.png}
\caption{Avalanches with Wilson Cowan units.}
\label{fig:avawilson}
\end{figure}
\end{document}
|
\section{Introduction}
String theory (see \cite{string} for a review) is the most hopeful scheme for complete quantum gravity, which is supposed to be valid up to the Planck scale. String length is a fundamental parameter in string theory, which is to be determined by experiment \cite{witten}. It has an inverse relation with the string coupling constant $g_s$, and the effective string length in physical world of four dimensions has a relation with the extra dimension volume in string theory. In this letter, we propose a possible way to determine the effective string length through experiment.
Our work in based on string gas cosmology (SGC) \cite{sgc}, which is a model of the early universe with minimal and crucial string inputs: a gas of closed strings \cite{blocks} as building blocks coupled to the space; the string oscillatory with winding modes as freedom; and T-duality as symmetry. SGC model can be embedded into string theory. The topics of extra dimension \cite{extra}, moduli fitting \cite{moduli}, elements of open string and D-brane \cite{brane}, the ``trans-Planckian'' problem \cite{trans-1}, and the ``swampland'' problem \cite{swampland} are discussed in related literature. As explained in \cite{sgc}, T-duality implies the energy spectrum of string states is invariant when winding and momentum quantum numbers are interchanged with $ (n,m) \rightarrow (m,n)$, and the consequence of T-duality resolves cosmological singularities immediately as the temperature in SGC has a maximum, the Hagedorn temperature $T_H$ \cite{hagedorn}.
SGC is a thermal dynamics system of strings. With assuming the matter is in thermal equilibrium, it predicts an almost scale-invariant spectrum of cosmological scalar perturbations in the Hagedorn phase, and thus it provides an alternative scenario to slow-roll inflation for the origins of the large-scale structure. In fact, the mechanism for the generation of the primordial perturbations is intrinsically stringy, as it predicts a spectrum with red tilt, unlike a large and phenomenological blue tilt in spectrum from particle thermodynamic fluctuations.
Since the first gravitational wave signal from a binary black hole merger on 2015 September 14th detected by LIGO \cite{ligo}, tens of BBH mergers and one binary neutron star (BNS) signal have been identified by the LIGO-VIRGO network (see \cite{ligo-virgo}). In spired of that, the detection of stochastic gravitational-wave background (SGWB) becomes a hot topic recently, though that has been discussed in a series of previous works (e.g. \cite{allen,2008PhLB..663...17F,2012PhRvD..85j4024W,2011ApJ...729...59Z,2011ApJ...739...86Z,2014PhRvD..89h4046R,2019arXiv191109745M,2018PhRvD..98d4020F,fan}). Current network of ground-based GWDs have set directional limits on the SGWB \cite{ligo03}.
SGWB is also predicted from SGC, and a rough scale invariant power spectrum with a blue tilt is shown in (\cite{bluetilt-1, bluetilt-2}). Recently, probing quantum gravity effect through stochastic gravitational wave background in SGC is studied in \cite{calcagni}. We revisit this topic in this work. With the help of Lambert W function, an exact energy density spectrum of gravitational waves background with the parameter of scalar-to-tensor is derived, and new feathers of the spectrum are then found.
\section{The SGWB from SGC}
We begin with the ansatz with a space-time metric containing both the linear scalar metric fluctuations $\phi$ and the gravitational wave fluctuations $h_{ij}$:
\begin{equation}
\label{fun-space}
ds^2 = a^2(\eta)((1 + 2\phi(x, \eta))d\eta^2 - [(1-2\phi) \delta_{ij} + h_{ij} ] dx_{i}dx_{j} ) ~,
\end{equation}
where $\eta$ is conformal time, $a(\eta)$ is the scale factor describing the background cosmology. We have chosen a gauge in which the metric corresponding to the cosmological perturbations is diagonal, and assumed that there is no anisotropic stress.
The inhomogeneities in SGC are not vacuum fluctuations, but rather thermal ones. Both of the scalar and gravitational waves (tensor metric) fluctuations are determined by the Einstein constraint equations, in terms of the fluctuations of the energy-momentum tensor. Specifically the gravitational fluctuations can be expressed as (see \cite{bluetilt-1})
\begin{eqnarray}
\label{fun-1}
&\Delta_{h}(k)& = <|h(k)^2|> = \frac{16 \pi^2 G^2}{k^{4}} \left<\delta T^{i}_{j}(k) \delta T^{i}_{j}(k)\right> ~, (i \ne j)
\end{eqnarray}
where $G$ is Newtons gravitational constant, and $h(k)$ is the amplitude of each of the two polarization modes of gravitational waves. As explained in \cite{bluetilt-2}, this is the gravitational fluctuations at present time generated from the usual theory of cosmological fluctuations.
The correlation of Eq. (\ref{fun-1}) is calculated in \cite{bluetilt-1} in a fixed-coordinate. Here we express that in co-moving coordinate in term of the frequency $f$ as
\begin{eqnarray}
\label{fun-2}
&\Delta_{h}(f)& = 16 \pi^2 G^2 \times {T \over {l_s^3}} \times (1 - \theta)\times \ln^2{\left[\frac{(1 -\theta)}{4\pi^2 l_s^2 f^2/c^2}\right]} ~,
\end{eqnarray}
where c is the light speed. We have set $\theta=T/T_{H}$, where $T=T(f_*)$ is the temperature when the mode with the frequency $f_*$ exits the Hubble radius. The power spectrum is independent on $a(\eta)$ and thus it is scale-invariant. As explained in \cite{bluetilt-2}, the temperature $T$ is independent on $f$ to the first approximation, and thus in this case, $T$, $\theta$ and $l_s$ are free parameters. In hagedorn phase with temperature close to the Hagedorn value, the power spectrum reduces to
\begin{eqnarray}
\label{fun-3}
&\Delta_{h}(f)& = (\frac{l_{pl}}{l_s})^{4} \times \theta \times (1 - \theta) \times \ln^2{\left[\frac{(1 -\theta)}{4\pi^2 l_s^2 f^2/c^2}\right]} ~,
\end{eqnarray}
where $l_s$ and $\theta$ are the free parameters, with $ 0 < \theta <1$. A suspected blue tilt of the tensor model is predicted in (\cite{bluetilt-1, bluetilt-2}) from the logarithmic term in Eq. (\ref{fun-3}), however, we will show that this blue tilt is prohibited by the current B-polarization observation.
The energy intensity spectrum can be characterized by the dimensionless quantity as (see \cite{neil})
\begin{eqnarray}
\label{fun-4}
&\Omega_{0}(f)& = \frac{1}{12} \frac{2\pi f}{H_{0}} \times \Delta_{h} ~,
\end{eqnarray}
where the Hubble constant at present is $H_0=3.2\times 10^{-18} h_{100}/sec$, and we use $h_{100}=0.65$.
The tensor-to-scalar $r$ in both Hagedorn phase and non-Hagedorn phase in SCG is defined as (see \cite{trans-1})
\begin{eqnarray}
\label{lambert1}
r(f) &=& (1 -\theta)^2 \times \ln^2{\left[\frac{(1 -\theta)}{4\pi^2 l_s^2 f^2/c^2}\right]} .
\end{eqnarray}
With rewriting it as
\begin{eqnarray}
\label{lambert2}
\frac{r^{1/2}}{4\pi^2 l_s^2 f^2} &=& \frac{(1 -\theta)}{4\pi^2 l_s^2 f^2} \times \ln{\left[\frac{(1 -\theta)}{4\pi^2 l_s^2 f^2/c^2}\right]} ~,
\end{eqnarray}
we find the $\theta$ can be solved inversely, and given by the so-called Lambert W function. The prototype Lambert W function is the inverse function of $x=y \cdot e^{y}$, and the solution is $y=W(x)$, however for Eq. (\ref{lambert2}), which is of the form $x=y\cdot \ln(y)$, the inverse function is given as $y=x/W(x)$. We write that as below immediately
\begin{equation}
\label{lambert3}
\theta(f) = 1- \frac{r^{1/2}}{W(\frac{r^{1/2}}{4\pi^2 l_s^2 f^2/c^2})} ~~.
\end{equation}
The prototype Lambert W function has two branches, $W_{0}$ and $W_{-1}$, but for Eq. (\ref{lambert2}), the function only has one branch $W_{0}$ sine the term in the left of Eq. (\ref{lambert3}) is always larger than 0. Now, we have derived the energy density $\Omega_0$ in term of $r$, in both Hagedorn phase and non-Hagedorn phase.
The tensor-to-scalar parameter r can be determined by the primordial CMB B-mode polarization experiment. In 2018, the Planck team announced the result, a tightest tensor-to-scalar ratio of $r \leq 0.06$ is given in \cite{b-mode}, combining with the BICEP2/Keck Array BK14 data.
Constraining the SGWB through CMB observation is studied in \cite{trans-1} and \cite{calcagni}, however, as they ignored the logarithmic term in Eq. (\ref{lambert1}), the constraining is loose. We give our analysis below, with using the exact spectrum derived above.
\section{Rule out the non-Hagedorn phase}
The non-Hagedorn phase is difficult to handle in SGC (e.g. see \cite{calcagni}), as the form of $T(k)$ is unknown. However, in this section, we show that the non-Hagedorn phase can be ruled out by the CMB data. It is reasonable to assume the string length $l_s \subset (l_{pl},l_{electron})$, where $l_{pl}$ is the Planck scale and $l_{electron}$ is the radius of electron, and with a measurable frequency $ f \subset (10^{-10}Hz, 10^{10}Hz)$, $\theta$ from Eq. (\ref{lambert3}) can be illustrated in Fig. \ref{ratio}. Two observations: first, a steep drop, which is caused by the logarithmic term in (\ref{lambert1}), is observed within $r\subset (10^3,10^5$). However, the steep drop is in fact prohibited as the recent B-polarization experiment \cite{b-mode} has given $r \leq 0.06$. Second, a constraint $\theta \geq 0.995$ is found in the region of $r \leq 0.06$. This implies only the mechanics of the mode exiting from the Hubble radius from Hagedorn phase with $T \geq 0.995 T_{H}$, is allowed, while the non-Hagedorn phase mechanics is ruled out.
\begin{figure}[hbt!]
\centering
\includegraphics[width=100mm]{ratio.png}
\caption{\label{ratio}The dependence of $\theta$ with r, with the measurable frequency of $f \subset (10^{-10}Hz, 10^{10}Hz)$, and the string length $l_s \subset (l_{pl},l_{electron})$. The dash one represents the upper boundary while the solid line represents the lower boundary.}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=100mm]{theta.png}
\caption{\label{theta}The dependence of $\Omega_0/(l_{pl}/l_s)^4$ with r. We do that with a measurable frequency of $ f \subset (10^{-10}Hz, 10^{10}Hz)$, and a string length of $l_s \subset (l_{pl},l_{electron})$. The dash one represents the upper boundary while the solid line represents the lower boundary. }
\end{figure}
As the measurable frequency and the string length assuming above are quite loose, our conclusion is quite general. With the same assumption, we find a similar steep drop in the spectrum of $\Omega_0/(l_{pl}/l_s)^4$ in dependent of $r$ in Fig. \ref{theta}, which is obviously prohibited by CMB observations with the same reason. Thus, only the red region is the figure is left, from the aspect of CMB. \\
\section{SGWB detectability from the Hagedorn phase}
The spectrum of gravitational waves from SCG can be simply written in the form of $\Omega \sim f\times ln^2[(1-\theta)/f^2]$. It is clear that such a spectrum with a special logarithmic term is unique. However, we don't know how significant the logarithmic term makes in the measurable frequency range ($10^{-4}Hz$, $10^{4}Hz$) of the current detector like aLIGO, and the upcoming detectors such as ET, DECIGO and LISA.
\begin{figure}[hbt!]
\centering
\includegraphics[width=120mm]{spectrum.png}
\caption{\label{spectrum}The spectrum region of the energy density $\Omega_{0}$ of gravitational waves at present time from SGC with setting $l_s=10^{6} l_{pl}$ and $r \subset (10^{-6},0.06)$. The cyan region is the the case with the logarithmic term in Eq.(\ref{lambert1}), while the red region is the case without. The dash one represents the upper boundary while the solid line represents the lower boundary. }
\end{figure}
The simplest way to see the uniqueness is comparing the spectrum with the one without such a term, which is of the power-law form $\Omega \sim f$. This is illustrated in Fig. \ref{spectrum}, with setting $l_s=10^{6} l_{pl}$ and $r \subset (10^{-6},0.06)$. The cyan region is the case with the logarithmic term, while the red region is the case without. It is clear that the spectrum region with the logarithmic term is higher than the one without, around 3 order.
Thus, we see that the logarithmic term makes significant change in the spectrum in the measurable range. Thus, is fair to say that, in case such a spectrum is found, the origin of gravitational waves from SCG is confirmed.
\section{The string length constraining}
The relation $\Omega_{0} \sim (l_{pl}/{l_s})^{4}$ in Eq. (\ref{fun-4}) implies the energy density spectrum is sensitive to the string length, and the smaller the string length, the easier the gravitational waves to be detected. The detectability of GWDs for SGWBs radiation \cite{allen} could be quantitatively characterized by the ratio of `signal'~(S) to `noise'~(N) which is given by an integral over frequency $f$ after correlating signals for duration $T$:
\begin{equation}
\label{SN1}
\left( \frac { S } { N } \right) ^ { 2 } = \frac { 9 H _ { 0 } ^ { 4 } } { 50 \pi ^ { 4 } } T \int _ { 0 } ^ { \infty } d f \frac { \gamma ^ { 2 } ( f ) \Omega _ { \mathrm { 0 } } ^ { 2 } ( f ) } { f ^ { 6 } P _ { 1 } ( f ) P _ { 2 } ( f ) } ~,
\end{equation}
where the Hubble constant $H _ { 0 } = 3.2 \times 10 ^ { - 18 } h _ { 100 } / \rm sec $ is the rate at which our universe is currently expanding and $h_{100}$ is a dimensionless parameter for Hubble constant and is assumed to be 0.65 in this paper. $P_i(f)$ is the one-side noise power spectral density which describes the instrument noise for GWDs in the frequency domain. $\gamma(f)$ denotes the overlap reduction function between two GWDs which encodes the relative positions and orientations of a pair of GWDs \cite{flanagan}.
Overall, we can infer from Eq.~\eqref{SN1} that the detectability of a network of GWDs to the SGWs are determined by the noise power spectral density together with the overlap reduction function of GWDs. The detailed information and description of overlap reduction function of multiple detector pairs together with their sensitivity curves we used in later analysis could be found in \cite{fan}. In order to detect SGWBs with $5\%$ false alarm and $95\%$ detection rate, the total signal-to-noise ratio (S/N) threshold $S/{N_\mathrm{opt}}$ in Eq.~\eqref{SN1} should be 3.29 \cite{allen}.
Given a fixed SNR threshold of 3.29, we could derive the relation between $l_s$ and $\theta$ in Eq (\ref{fun-2}) by Eq (\ref{fun-3}) as presented in Fig.\ref{constrain} for GWD networks by correlating their signals for 3 months $(T = 10^7 s \sim $3 months). By illustration, we simulated three GWD networks. They are respectively, the second generation GWD network, L(LIGO livingston)-H(LIGO Hanford); the third generation GWD network, ET(Einstein Telescope)- CE(Cosmic Explorer), and Wuhan(kHz)-Aus(kHz), which assumed 2 detectors with kilo-hertz detector sensitivity respectively located in Wuhan of China and Australia.
\begin{figure}[hbt!]
\centering
\includegraphics[width=120mm]{SNR329_theta_ls_c2_log.png}
\caption{\label{constrain}The string length $l_s$ versus the tensor-to-scalar $r$ given that network S/N of 3.29. `L' is for `LIGO Livingston', `H' is `LIGO Hanford'. 'ET' represents the third generation detector Einstein Telescope, while 'CE' is Cosmic Explorer. Wuhan(kHz) means a kilo-Hertz detector located in Wuhan of China. Aus(kHz) represents the same kilo-Hertz detector located in Australia. The red vertical line represents the $l_s/l_{pl}$ at position of $\theta$ with $r=0.06$. }
\end{figure}
The relation $\Omega_{0} \sim \theta \times (1 - \theta)\times \ln^2{[(1 -\theta)]}$ gives the dependence of the sensitivity and $\theta$. As shown in the figure, the constraining capability for $\theta$ is almost independent on the sensitivity of GWDs. To be specific, lower right region, namely larger theta and smaller $l_s$, is easier to be tested as sensitivity of second generation detector network is enough to detect with higher SNR. On the other hand, upper left region, which corresponds to smaller theta and larger $l_s$, is harder to be explored, because GWDs with extremely high sensitivity are also limited to detect this region.
The relation $\Omega_{0} \sim (l_{pl}/{l_s})^{4}$ in Eq. (\ref{fun-2}) implies the energy density is sensitive to the string length, and the smaller the string length, the easier the gravitational waves to be detected. This could be seen from this figure that given the same $\theta$, the third generation detector network with higher sensitivity could detect larger string length $l_s$ with the same network SNR compared to the seconds generation detector network, which means detector networks with higher sensitivity has the ability to explore larger parameter space. And the constraining capability for $l_s$ is mainly determined by the GWD sensitivity.
With $r<0.06$, we find from Fig.\ref{constrain} that the string length $l_s$ is confined to be lower than $8\sim$ orders of the Planck scale from the third generation of GWD, while from the second generation of GWD, it is confined to be lower than $7\sim$ orders of the Planck scale. It is estimated in \cite{witten} that the string length might be of $15 \sim$ order of the Planck scale. However, if that is right, finding the gravitational waves from SCG through the current the upcoming detectors is completely impossible.
\section{Conclusion}
In conclusion, SGC is a thermal dynamics system of strings, and its prediction can be used to test string theory. The scalar perturbation from SCG remains it as one of the origins of large-scale structure.
SGWB from SCG is studied in this work. With using the Lambert W function, we derived the exact energy density spectrum of SGWB in terms of the measurable tensor-to-scalar. With that spectrum, we found the non-Hagedorn phase in SCG is ruled out by the current B-polarization experiment data . As the logarithmic term plays a critical role, the spectrum from Hagedorn phase was found to be unique in the measurable frequency range. The most important, we found the string length can be confined to be lower than $7\sim $ orders of the Planck scale with the current B-polarization experiment data. \\
\acknowledgments
X. F. thanks R. Brandenberger for pointing out the references on string gas cosmology scenario. X. F. is supported by the National Natural Science Foundation of China(under Grants No.11922303) and Hubei province Natural Science Fund for the Distinguished Young Scholars (2019CFA052).
|
\section*{Affiliation notes}
$^{\rm I}$ Deceased\\
$^{\rm II}$ Also at: Italian National Agency for New Technologies, Energy and Sustainable Economic Development (ENEA), Bologna, Italy\\
$^{\rm III}$ Also at: Dipartimento DET del Politecnico di Torino, Turin, Italy\\
$^{\rm IV}$ Also at: M.V. Lomonosov Moscow State University, D.V. Skobeltsyn Institute of Nuclear, Physics, Moscow, Russia\\
$^{\rm V}$ Also at: Institute of Theoretical Physics, University of Wroclaw, Poland\\
\section*{Collaboration Institutes}
$^{1}$ A.I. Alikhanyan National Science Laboratory (Yerevan Physics Institute) Foundation, Yerevan, Armenia\\
$^{2}$ AGH University of Science and Technology, Cracow, Poland\\
$^{3}$ Bogolyubov Institute for Theoretical Physics, National Academy of Sciences of Ukraine, Kiev, Ukraine\\
$^{4}$ Bose Institute, Department of Physics and Centre for Astroparticle Physics and Space Science (CAPSS), Kolkata, India\\
$^{5}$ Budker Institute for Nuclear Physics, Novosibirsk, Russia\\
$^{6}$ California Polytechnic State University, San Luis Obispo, California, United States\\
$^{7}$ Central China Normal University, Wuhan, China\\
$^{8}$ Centro de Aplicaciones Tecnol\'{o}gicas y Desarrollo Nuclear (CEADEN), Havana, Cuba\\
$^{9}$ Centro de Investigaci\'{o}n y de Estudios Avanzados (CINVESTAV), Mexico City and M\'{e}rida, Mexico\\
$^{10}$ Chicago State University, Chicago, Illinois, United States\\
$^{11}$ China Institute of Atomic Energy, Beijing, China\\
$^{12}$ Chungbuk National University, Cheongju, Republic of Korea\\
$^{13}$ Comenius University Bratislava, Faculty of Mathematics, Physics and Informatics, Bratislava, Slovakia\\
$^{14}$ COMSATS University Islamabad, Islamabad, Pakistan\\
$^{15}$ Creighton University, Omaha, Nebraska, United States\\
$^{16}$ Department of Physics, Aligarh Muslim University, Aligarh, India\\
$^{17}$ Department of Physics, Pusan National University, Pusan, Republic of Korea\\
$^{18}$ Department of Physics, Sejong University, Seoul, Republic of Korea\\
$^{19}$ Department of Physics, University of California, Berkeley, California, United States\\
$^{20}$ Department of Physics, University of Oslo, Oslo, Norway\\
$^{21}$ Department of Physics and Technology, University of Bergen, Bergen, Norway\\
$^{22}$ Dipartimento di Fisica dell'Universit\`{a} 'La Sapienza' and Sezione INFN, Rome, Italy\\
$^{23}$ Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Cagliari, Italy\\
$^{24}$ Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Trieste, Italy\\
$^{25}$ Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Turin, Italy\\
$^{26}$ Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Bologna, Italy\\
$^{27}$ Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Catania, Italy\\
$^{28}$ Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Padova, Italy\\
$^{29}$ Dipartimento di Fisica e Nucleare e Teorica, Universit\`{a} di Pavia, Pavia, Italy\\
$^{30}$ Dipartimento di Fisica `E.R.~Caianiello' dell'Universit\`{a} and Gruppo Collegato INFN, Salerno, Italy\\
$^{31}$ Dipartimento DISAT del Politecnico and Sezione INFN, Turin, Italy\\
$^{32}$ Dipartimento di Scienze e Innovazione Tecnologica dell'Universit\`{a} del Piemonte Orientale and INFN Sezione di Torino, Alessandria, Italy\\
$^{33}$ Dipartimento di Scienze MIFT, Universit\`{a} di Messina, Messina, Italy\\
$^{34}$ Dipartimento Interateneo di Fisica `M.~Merlin' and Sezione INFN, Bari, Italy\\
$^{35}$ European Organization for Nuclear Research (CERN), Geneva, Switzerland\\
$^{36}$ Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture, University of Split, Split, Croatia\\
$^{37}$ Faculty of Engineering and Science, Western Norway University of Applied Sciences, Bergen, Norway\\
$^{38}$ Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Prague, Czech Republic\\
$^{39}$ Faculty of Science, P.J.~\v{S}af\'{a}rik University, Ko\v{s}ice, Slovakia\\
$^{40}$ Frankfurt Institute for Advanced Studies, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany\\
$^{41}$ Fudan University, Shanghai, China\\
$^{42}$ Gangneung-Wonju National University, Gangneung, Republic of Korea\\
$^{43}$ Gauhati University, Department of Physics, Guwahati, India\\
$^{44}$ Helmholtz-Institut f\"{u}r Strahlen- und Kernphysik, Rheinische Friedrich-Wilhelms-Universit\"{a}t Bonn, Bonn, Germany\\
$^{45}$ Helsinki Institute of Physics (HIP), Helsinki, Finland\\
$^{46}$ High Energy Physics Group, Universidad Aut\'{o}noma de Puebla, Puebla, Mexico\\
$^{47}$ Hiroshima University, Hiroshima, Japan\\
$^{48}$ Hochschule Worms, Zentrum f\"{u}r Technologietransfer und Telekommunikation (ZTT), Worms, Germany\\
$^{49}$ Horia Hulubei National Institute of Physics and Nuclear Engineering, Bucharest, Romania\\
$^{50}$ Indian Institute of Technology Bombay (IIT), Mumbai, India\\
$^{51}$ Indian Institute of Technology Indore, Indore, India\\
$^{52}$ Indonesian Institute of Sciences, Jakarta, Indonesia\\
$^{53}$ INFN, Laboratori Nazionali di Frascati, Frascati, Italy\\
$^{54}$ INFN, Sezione di Bari, Bari, Italy\\
$^{55}$ INFN, Sezione di Bologna, Bologna, Italy\\
$^{56}$ INFN, Sezione di Cagliari, Cagliari, Italy\\
$^{57}$ INFN, Sezione di Catania, Catania, Italy\\
$^{58}$ INFN, Sezione di Padova, Padova, Italy\\
$^{59}$ INFN, Sezione di Pavia, Pavia, Italy\\
$^{60}$ INFN, Sezione di Roma, Rome, Italy\\
$^{61}$ INFN, Sezione di Torino, Turin, Italy\\
$^{62}$ INFN, Sezione di Trieste, Trieste, Italy\\
$^{63}$ Inha University, Incheon, Republic of Korea\\
$^{64}$ Institute for Gravitational and Subatomic Physics (GRASP), Utrecht University/Nikhef, Utrecht, Netherlands\\
$^{65}$ Institute for Nuclear Research, Academy of Sciences, Moscow, Russia\\
$^{66}$ Institute of Experimental Physics, Slovak Academy of Sciences, Ko\v{s}ice, Slovakia\\
$^{67}$ Institute of Physics, Homi Bhabha National Institute, Bhubaneswar, India\\
$^{68}$ Institute of Physics of the Czech Academy of Sciences, Prague, Czech Republic\\
$^{69}$ Institute of Space Science (ISS), Bucharest, Romania\\
$^{70}$ Institut f\"{u}r Kernphysik, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany\\
$^{71}$ Instituto de Ciencias Nucleares, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico\\
$^{72}$ Instituto de F\'{i}sica, Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil\\
$^{73}$ Instituto de F\'{\i}sica, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico\\
$^{74}$ iThemba LABS, National Research Foundation, Somerset West, South Africa\\
$^{75}$ Jeonbuk National University, Jeonju, Republic of Korea\\
$^{76}$ Johann-Wolfgang-Goethe Universit\"{a}t Frankfurt Institut f\"{u}r Informatik, Fachbereich Informatik und Mathematik, Frankfurt, Germany\\
$^{77}$ Joint Institute for Nuclear Research (JINR), Dubna, Russia\\
$^{78}$ Korea Institute of Science and Technology Information, Daejeon, Republic of Korea\\
$^{79}$ KTO Karatay University, Konya, Turkey\\
$^{80}$ Laboratoire de Physique des 2 Infinis, Ir\`{e}ne Joliot-Curie, Orsay, France\\
$^{81}$ Laboratoire de Physique Subatomique et de Cosmologie, Universit\'{e} Grenoble-Alpes, CNRS-IN2P3, Grenoble, France\\
$^{82}$ Lawrence Berkeley National Laboratory, Berkeley, California, United States\\
$^{83}$ Lund University Department of Physics, Division of Particle Physics, Lund, Sweden\\
$^{84}$ Moscow Institute for Physics and Technology, Moscow, Russia\\
$^{85}$ Nagasaki Institute of Applied Science, Nagasaki, Japan\\
$^{86}$ Nara Women{'}s University (NWU), Nara, Japan\\
$^{87}$ National and Kapodistrian University of Athens, School of Science, Department of Physics , Athens, Greece\\
$^{88}$ National Centre for Nuclear Research, Warsaw, Poland\\
$^{89}$ National Institute of Science Education and Research, Homi Bhabha National Institute, Jatni, India\\
$^{90}$ National Nuclear Research Center, Baku, Azerbaijan\\
$^{91}$ National Research Centre Kurchatov Institute, Moscow, Russia\\
$^{92}$ Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark\\
$^{93}$ Nikhef, National institute for subatomic physics, Amsterdam, Netherlands\\
$^{94}$ NRC Kurchatov Institute IHEP, Protvino, Russia\\
$^{95}$ NRC \guillemotleft Kurchatov\guillemotright Institute - ITEP, Moscow, Russia\\
$^{96}$ NRNU Moscow Engineering Physics Institute, Moscow, Russia\\
$^{97}$ Nuclear Physics Group, STFC Daresbury Laboratory, Daresbury, United Kingdom\\
$^{98}$ Nuclear Physics Institute of the Czech Academy of Sciences, \v{R}e\v{z} u Prahy, Czech Republic\\
$^{99}$ Oak Ridge National Laboratory, Oak Ridge, Tennessee, United States\\
$^{100}$ Ohio State University, Columbus, Ohio, United States\\
$^{101}$ Petersburg Nuclear Physics Institute, Gatchina, Russia\\
$^{102}$ Physics department, Faculty of science, University of Zagreb, Zagreb, Croatia\\
$^{103}$ Physics Department, Panjab University, Chandigarh, India\\
$^{104}$ Physics Department, University of Jammu, Jammu, India\\
$^{105}$ Physics Department, University of Rajasthan, Jaipur, India\\
$^{106}$ Physikalisches Institut, Eberhard-Karls-Universit\"{a}t T\"{u}bingen, T\"{u}bingen, Germany\\
$^{107}$ Physikalisches Institut, Ruprecht-Karls-Universit\"{a}t Heidelberg, Heidelberg, Germany\\
$^{108}$ Physik Department, Technische Universit\"{a}t M\"{u}nchen, Munich, Germany\\
$^{109}$ Politecnico di Bari and Sezione INFN, Bari, Italy\\
$^{110}$ Research Division and ExtreMe Matter Institute EMMI, GSI Helmholtzzentrum f\"ur Schwerionenforschung GmbH, Darmstadt, Germany\\
$^{111}$ Russian Federal Nuclear Center (VNIIEF), Sarov, Russia\\
$^{112}$ Saha Institute of Nuclear Physics, Homi Bhabha National Institute, Kolkata, India\\
$^{113}$ School of Physics and Astronomy, University of Birmingham, Birmingham, United Kingdom\\
$^{114}$ Secci\'{o}n F\'{\i}sica, Departamento de Ciencias, Pontificia Universidad Cat\'{o}lica del Per\'{u}, Lima, Peru\\
$^{115}$ St. Petersburg State University, St. Petersburg, Russia\\
$^{116}$ Stefan Meyer Institut f\"{u}r Subatomare Physik (SMI), Vienna, Austria\\
$^{117}$ SUBATECH, IMT Atlantique, Universit\'{e} de Nantes, CNRS-IN2P3, Nantes, France\\
$^{118}$ Suranaree University of Technology, Nakhon Ratchasima, Thailand\\
$^{119}$ Technical University of Ko\v{s}ice, Ko\v{s}ice, Slovakia\\
$^{120}$ The Henryk Niewodniczanski Institute of Nuclear Physics, Polish Academy of Sciences, Cracow, Poland\\
$^{121}$ The University of Texas at Austin, Austin, Texas, United States\\
$^{122}$ Universidad Aut\'{o}noma de Sinaloa, Culiac\'{a}n, Mexico\\
$^{123}$ Universidade de S\~{a}o Paulo (USP), S\~{a}o Paulo, Brazil\\
$^{124}$ Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil\\
$^{125}$ Universidade Federal do ABC, Santo Andre, Brazil\\
$^{126}$ University of Cape Town, Cape Town, South Africa\\
$^{127}$ University of Houston, Houston, Texas, United States\\
$^{128}$ University of Jyv\"{a}skyl\"{a}, Jyv\"{a}skyl\"{a}, Finland\\
$^{129}$ University of Kansas, Lawrence, Kansas, United States\\
$^{130}$ University of Liverpool, Liverpool, United Kingdom\\
$^{131}$ University of Science and Technology of China, Hefei, China\\
$^{132}$ University of South-Eastern Norway, Tonsberg, Norway\\
$^{133}$ University of Tennessee, Knoxville, Tennessee, United States\\
$^{134}$ University of the Witwatersrand, Johannesburg, South Africa\\
$^{135}$ University of Tokyo, Tokyo, Japan\\
$^{136}$ University of Tsukuba, Tsukuba, Japan\\
$^{137}$ Universit\'{e} Clermont Auvergne, CNRS/IN2P3, LPC, Clermont-Ferrand, France\\
$^{138}$ Universit\'{e} de Lyon, CNRS/IN2P3, Institut de Physique des 2 Infinis de Lyon , Lyon, France\\
$^{139}$ Universit\'{e} de Strasbourg, CNRS, IPHC UMR 7178, F-67000 Strasbourg, France, Strasbourg, France\\
$^{140}$ Universit\'{e} Paris-Saclay Centre d'Etudes de Saclay (CEA), IRFU, D\'{e}partment de Physique Nucl\'{e}aire (DPhN), Saclay, France\\
$^{141}$ Universit\`{a} degli Studi di Foggia, Foggia, Italy\\
$^{142}$ Universit\`{a} di Brescia, Brescia, Italy\\
$^{143}$ Variable Energy Cyclotron Centre, Homi Bhabha National Institute, Kolkata, India\\
$^{144}$ Warsaw University of Technology, Warsaw, Poland\\
$^{145}$ Wayne State University, Detroit, Michigan, United States\\
$^{146}$ Westf\"{a}lische Wilhelms-Universit\"{a}t M\"{u}nster, Institut f\"{u}r Kernphysik, M\"{u}nster, Germany\\
$^{147}$ Wigner Research Centre for Physics, Budapest, Hungary\\
$^{148}$ Yale University, New Haven, Connecticut, United States\\
$^{149}$ Yonsei University, Seoul, Republic of Korea\\
\bigskip
\end{flushleft}
\endgroup
\section*{Acknowledgements}
\input{fa_2021-04-27.tex}
\end{acknowledgement}
\bibliographystyle{utphys}
|
\section{Introduction}
The Elsevier cas-sc class is based on the
standard article class and supports almost all of the functionality of
that class. In addition, it features commands and options to format the
\begin{itemize} \item document style \item baselineskip \item front
matter \item keywords and MSC codes \item theorems, definitions and
proofs \item lables of enumerations \item citation style and labeling.
\end{itemize}
This class depends on the following packages
for its proper functioning:
\begin{enumerate}
\itemsep=0pt
\item {natbib.sty} for citation processing;
\item {geometry.sty} for margin settings;
\item {fleqn.clo} for left aligned equations;
\item {graphicx.sty} for graphics inclusion;
\item {hyperref.sty} optional packages if hyperlinking is
required in the document;
\end{enumerate}
All the above packages are part of any
standard \LaTeX{} installation.
Therefore, the users need not be
bothered about downloading any extra packages.
\section{Installation}
The package is available at author resources page at Elsevier
(\url{http://www.elsevier.com/locate/latex}).
The class may be moved or copied to a place, usually,
\verb+$TEXMF/tex/latex/elsevier/+,
or a folder which will be read
by \LaTeX{} during document compilation. The \TeX{} file
database needs updation after moving/copying class file. Usually,
we use commands like \verb+mktexlsr+ or \verb+texhash+ depending
upon the distribution and operating system.
\section{Front matter}
The author names and affiliations could be formatted in two ways:
\begin{enumerate}[(1)]
\item Group the authors per affiliation.
\item Use footnotes to indicate the affiliations.
\end{enumerate}
See the front matter of this document for examples.
You are recommended to conform your choice to the journal you
are submitting to.
\section{Bibliography styles}
There are various bibliography styles available. You can select the
style of your choice in the preamble of this document. These styles are
Elsevier styles based on standard styles like Harvard and Vancouver.
Please use Bib\TeX\ to generate your bibliography and include DOIs
whenever available.
Here are two sample references:
See \citet{Fortunato2010}. Also refer \citet{Fortunato2010,NewmanGirvan2004}.
More citations are here \citep{Fortunato2010,Vehlowetal2013}.
\section{Floats}
{Figures} may be included using the command, \verb+\includegraphics+ in
combination with or without its several options to further control
graphic. \verb+\includegraphics+ is provided by {graphic[s,x].sty}
which is part of any standard \LaTeX{} distribution.
{graphicx.sty} is loaded by default. \LaTeX{} accepts figures in
the postscript format while pdf\LaTeX{} accepts {*.pdf},
{*.mps} (metapost), {*.jpg} and {*.png} formats.
pdf\LaTeX{} does not accept graphic files in the postscript format.
\begin{figure}
\centering
\includegraphics[scale=.75]{figs/Fig1.pdf}
\caption{The evanescent light - $1S$ quadrupole coupling
($g_{1,l}$) scaled to the bulk exciton-photon coupling
($g_{1,2}$). The size parameter $kr_{0}$ is denoted as $x$ and
the \PMS is placed directly on the cuprous oxide sample ($\delta
r=0$, See also Table \protect\ref{tbl1}).}
\label{FIG:1}
\end{figure}
The \verb+table+ environment is handy for marking up tabular
material. If users want to use {multirow.sty},
{array.sty}, etc., to fine control/enhance the tables, they
are welcome to load any package of their choice and
{cas-sc.cls} will work in combination with all loaded
packages.
\begin{table}[width=.9\linewidth,cols=4,pos=h]
\caption{This is a test caption. This is a test caption. This is a test
caption. This is a test caption.}\label{tbl1}
\begin{tabular*}{\tblwidth}{@{} LLLL@{} }
\toprule
Col 1 & Col 2 & Col 3 & Col4\\
\midrule
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
12345 & 12345 & 123 & 12345 \\
\bottomrule
\end{tabular*}
\end{table}
\section[Theorem and ...]{Theorem and theorem like environments}
{cas-sc.cls} provides a few shortcuts to format theorems and
theorem-like environments with ease. In all commands the options that
are used with the \verb+\newtheorem+ command will work exactly in the same
manner. {cas-sc.cls} provides three commands to format theorem or
theorem-like environments:
\begin{verbatim}
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newdefinition{rmk}{Remark}
\newproof{pf}{Proof}
\newproof{pot}{Proof of Theorem \ref{thm2}}
\end{verbatim}
The \verb+\newtheorem+ command formats a
theorem in \LaTeX's default style with italicized font, bold font
for theorem heading and theorem number at the right hand side of the
theorem heading. It also optionally accepts an argument which
will be printed as an extra heading in parentheses.
\begin{verbatim}
\begin{theorem}
For system (8), consensus can be achieved with
$\|T_{\omega z}$ ...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{theorem}
\end{verbatim}
\newtheorem{theorem}{Theorem}
\begin{theorem}
For system (8), consensus can be achieved with
$\|T_{\omega z}$ ...
\begin{eqnarray}\label{10}
....
\end{eqnarray}
\end{theorem}
The \verb+\newdefinition+ command is the same in
all respects as its \verb+\newtheorem+ counterpart except that
the font shape is roman instead of italic. Both
\verb+\newdefinition+ and \verb+\newtheorem+ commands
automatically define counters for the environments defined.
The \verb+\newproof+ command defines proof environments with
upright font shape. No counters are defined.
\section[Enumerated ...]{Enumerated and Itemized Lists}
{cas-sc.cls} provides an extended list processing macros
which makes the usage a bit more user friendly than the default
\LaTeX{} list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list counter
type and its attributes.
\begin{verbatim}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.', so that the item counter will be suffixed
by a period.
\item You can use `a)' for alphabetical counter and '(i)' for
roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\item One more item before we start another.
\item One more item before we start another.
\item One more item before we start another.
\end{verbatim}
Further, the enhanced list environment allows one to prefix a
string like `step' to all the item numbers.
\begin{verbatim}
\begin{enumerate}[Step 1.]
\item This is the first step of the example list.
\item Obviously this is the second step.
\item The final step to wind up this example.
\end{enumerate}
\end{verbatim}
\section{Cross-references}
In electronic publications, articles may be internally
hyperlinked. Hyperlinks are generated from proper
cross-references in the article. For example, the words
\textcolor{black!80}{Fig.~1} will never be more than simple text,
whereas the proper cross-reference \verb+\ref{tiger}+ may be
turned into a hyperlink to the figure itself:
\textcolor{blue}{Fig.~1}. In the same way,
the words \textcolor{blue}{Ref.~[1]} will fail to turn into a
hyperlink; the proper cross-reference is \verb+\cite{Knuth96}+.
Cross-referencing is possible in \LaTeX{} for sections,
subsections, formulae, figures, tables, and literature
references.
\section{Bibliography}
Two bibliographic style files (\verb+*.bst+) are provided ---
{model1-num-names.bst} and {model2-names.bst} --- the first one can be
used for the numbered scheme. This can also be used for the numbered
with new options of {natbib.sty}. The second one is for the author year
scheme. When you use model2-names.bst, the citation commands will be
like \verb+\citep+, \verb+\citet+, \verb+\citealt+ etc. However when
you use model1-num-names.bst, you may use only \verb+\cite+ command.
\verb+thebibliography+ environment. Each reference is a
\verb+\bibitem+ and each \verb+\bibitem+ is identified by a label,
by which it can be cited in the text:
\noindent In connection with cross-referencing and
possible future hyperlinking it is not a good idea to collect
more that one literature item in one \verb+\bibitem+. The
so-called Harvard or author-year style of referencing is enabled
by the \LaTeX{} package {natbib}. With this package the
literature can be cited as follows:
\begin{enumerate}[\textbullet]
\item Parenthetical: \verb+\citep{WB96}+ produces (Wettig \& Brown, 1996).
\item Textual: \verb+\citet{ESG96}+ produces Elson et al. (1996).
\item An affix and part of a reference:
\verb+\citep[e.g.][Ch. 2]{Gea97}+ produces (e.g. Governato et
al., 1997, Ch. 2).
\end{enumerate}
In the numbered scheme of citation, \verb+\cite{<label>}+ is used,
since \verb+\citep+ or \verb+\citet+ has no relevance in the numbered
scheme. {natbib} package is loaded by {cas-sc} with
\verb+numbers+ as default option. You can change this to author-year
or harvard scheme by adding option \verb+authoryear+ in the class
loading command. If you want to use more options of the {natbib}
package, you can do so with the \verb+\biboptions+ command. For
details of various options of the {natbib} package, please take a
look at the {natbib} documentation, which is part of any standard
\LaTeX{} installation.
\section{Introduction}
Two classfiles namely \file{cas-sc.cls} and \file{cas-dc.cls} were
written for typesetting articles submitted in journals of Elsevier's
Complex Article Service (CAS) workflow.
\subsection{Usage}
\begin{enumerate}
\item \file{cas-sc.cls} for single column journals.
\begin{vquote}
\documentclass[<options>]{cas-sc}
\end{vquote}
\item \file{cas-dc.cls} for single column journals.
\begin{vquote}
\documentclass[<options>]{cas-dc}
\end{vquote}
\end{enumerate}
and have an option longmktitle to handle long front matter.
\section{Front matter}
\begin{vquote}
\title [mode = title]{This is a specimen $a_b$ title}
\tnotemark[1,2]
\tnotetext[1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[2]{The second title footnote which is a longer text
matter to fill through the whole text width and overflow into
another line in the footnotes area of the first page.}
\author[1,3]{CV Radhakrishnan}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-7511-2910]
\cormark[1]
\fnmark[1]
\ead{<EMAIL>}
\ead[url]{www.cvr.cc, <EMAIL>}
\end{vquote}
\begin{vquote}
\credit{Conceptualization of this study, Methodology,
Software}
\address[1]{Elsevier B.V., Radarweg 29, 1043 NX Amsterdam,
The Netherlands}
\author[2,4]{Han Theh Thanh}[style=chinese]
\author[2,3]{CV Rajagopal}[%
role=Co-ordinator,
suffix=Jr,
]
\fnmark[2]
\ead{<EMAIL>}
\ead[URL]{www.sayahna.org}
\credit{Data curation, Writing - Original draft preparation}
\address[2]{Sayahna Foundation, Jagathy, Trivandrum 695014,
India}
\author[1,3]{Rishi T.}
\cormark[2]
\fnmark[1,3]
\ead{<EMAIL>}
\ead[URL]{www.stmdocs.in}
\address[3]{STM Document Engineering Pvt Ltd., Mepukada,
Malayinkil, Trivandrum 695571, India}
\cortext[cor1]{Corresponding author}
\cortext[cor2]{Principal corresponding author}
\fntext[fn1]{This is the first author footnote. but is common
to third author as well.}
\fntext[fn2]{Another author footnote, this is a very long
footnote and it should be a really long footnote. But this
footnote is not yet sufficiently long enough to make two lines
of footnote text.}
\end{vquote}
\begin{vquote}
\nonumnote{This note has no numbers. In this work we
demonstrate $a_b$ the formation Y\_1 of a new type of
polariton on the interface between a cuprous oxide slab
and a polystyrene micro-sphere placed on the slab.
}
\begin{abstract}[S U M M A R Y]
This template helps you to create a properly formatted
\LaTeX\ manuscript.
\noindent\texttt{\textbackslash begin{abstract}} \dots
\texttt{\textbackslash end{abstract}} and
\verb+\begin{keyword}+ \verb+...+ \verb+\end{keyword}+
which contain the abstract and keywords respectively.
Each keyword shall be separated by a \verb+\sep+ command.
\end{abstract}
\begin{keywords}
quadrupole exciton \sep polariton \sep \WGM \sep \BEC
\end{keywords}
\maketitle
\end{vquote}
\begin{figure}
\includegraphics[width=\textwidth]{sc-sample.pdf}
\caption{Single column output (classfile: cas-sc.cls).}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{dc-sample.pdf}
\caption{Double column output (classfile: cas-dc.cls).}
\end{figure}
\subsection{Title}
\verb+\title+ command have the below options:
\begin{enumerate}
\item \verb+title:+ Document title
\item \verb+alt:+ Alternate title
\item \verb+sub:+ Sub title
\item \verb+trans:+ Translated title
\item \verb+transsub:+ Translated sub title
\end{enumerate}
\begin{vquote}
\title[mode=title]{This is a title}
\title[mode=alt]{This is a alternate title}
\title[mode=sub]{This is a sub title}
\title[mode=trans]{This is a translated title}
\title[mode=transsub]{This is a translated sub title}
\end{vquote}
\subsection{Author}
\verb+\author+ command have the below options:
\begin{enumerate}
\item \verb+auid:+ Author id
\item \verb+bioid:+ Biography id
\item \verb+alt:+ Alternate author
\item \verb+style:+ Style of author name chinese
\item \verb+prefix:+ Prefix Sir
\item \verb+suffix:+ Suffix
\item \verb+degree:+ Degree
\item \verb+role:+ Role
\item \verb+orcid:+ ORCID
\item \verb+collab:+ Collaboration
\item \verb+anon:+ Anonymous author
\item \verb+deceased:+ Deceased author
\item \verb+twitter:+ Twitter account
\item \verb+facebook:+ Facebook account
\item \verb+linkedin:+ LinkedIn account
\item \verb+plus:+ Google plus account
\item \verb+gplus:+ Google plus account
\end{enumerate}
\begin{vquote}
\author[1,3]{Author Name}[type=editor,
auid=000,bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-7511-2910,
facebook=<facebook id>,
twitter=<twitter id>,
linkedin=<linkedin id>,
gplus=<gplus id>]
\end{vquote}
\subsection{Various Marks in the Front Matter}
The front matter becomes complicated due to various kinds
of notes and marks to the title and author names. Marks in
the title will be denoted by a star ($\star$) mark;
footnotes are denoted by super scripted Arabic numerals,
corresponding author by of an Conformal asterisk (*) mark.
\subsubsection{Title marks}
Title mark can be entered by the command, \verb+\tnotemark[<num>]+
and the corresponding text can be entered with the command
\verb+\tnotetext[<num>]+ \verb+{<text>}+. An example will be:
\begin{vquote}
\title[mode=title]{Leveraging social media news to predict
stock index movement using RNN-boost}
\tnotemark[1,2]
\tnotetext[1]{This document is the results of the research
project funded by the National Science Foundation.}
\tnotetext[2]{The second title footnote which is a longer
text matter to fill through the whole text width and
overflow into another line in the footnotes area of
the first page.}
\end{vquote}
\verb+\tnotetext+ and \verb+\tnotemark+ can be anywhere in
the front matter, but shall be before \verb+\maketitle+ command.
\subsubsection{Author marks}
Author names can have many kinds of marks and notes:
\begin{vquote}
footnote mark : \fnmark[<num>]
footnote text : \fntext[<num>]{<text>}
affiliation mark : \author[<num>]
email : \ead{<emailid>}
url : \ead[url]{<url>}
corresponding author mark : \cormark[<num>]
corresponding author text : \cortext[<num>]{<text>}
\end{vquote}
\subsubsection{Other marks}
At times, authors want footnotes which leave no marks in
the author names. The note text shall be listed as part of
the front matter notes. Class files provides
\verb+\nonumnote+ for this purpose. The usage
\begin{vquote}
\nonumnote{<text>}
\end{vquote}
\noindent and should be entered anywhere before the \verb+\maketitle+
command for this to take effect.
\subsection{Abstract and Keywords}
Abstract shall be entered in an environment that starts
with \verb+\begin{abstract}+ and ends with
\verb+\end{abstract}+. Longer abstracts spanning more than
one page is also possible in Class file even in double
column mode. We need to invoke longmktitle option in the
class loading line for this to happen smoothly.
The key words are enclosed in a \verb+{keyword}+
environment.
\begin{vquote}
\begin{abstract}
This is a abstract. \lipsum[3]
\end{abstract}
\begin{keywords}
First keyword \sep Second keyword \sep Third
keyword \sep Fourth keyword
\end{keywords}
\end{vquote}
\section{Main Matter}
\subsection{Tables}
\subsubsection{Normal tables}
\begin{vquote}
\begin{table}
\caption{This is a test caption.}
\begin{tabular*}{\tblwidth}{@{} LLLL@{} }
\toprule
Col 1 & Col 2\\
\midrule
12345 & 12345\\
12345 & 12345\\
12345 & 12345\\
\bottomrule
\end{tabular*}
\end{table}
\end{vquote}
\subsubsection{Span tables}
\begin{vquote}
\begin{table*}[width=.9\textwidth,cols=4,pos=h]
\caption{This is a test caption.}
\begin{tabular*}{\tblwidth}{@{} LLLLLL@{} }
\toprule
Col 1 & Col 2 & Col 3 & Col4 & Col5 & Col6 & Col7\\
\midrule
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
12345 & 12345 & 123 & 12345 & 123 & 12345 & 123 \\
\bottomrule
\end{tabular*}
\end{table*}
\end{vquote}
\subsection{Figures}
\subsubsection{Normal figures}
\begin{vquote}
\begin{figure}
\centering
\includegraphics[scale=.75]{Fig1.pdf}
\caption{The evanescent light - $1S$ quadrupole coupling
($g_{1,l}$) scaled to the bulk exciton-photon coupling
($g_{1,2}$). The size parameter $kr_{0}$ is denoted as $x$ and
the \PMS is placed directly on the cuprous oxide sample ($\delta
r=0$, See also Fig. \protect\ref{FIG:2}).}
\label{FIG:1}
\end{figure}
\end{vquote}
\subsubsection{Span figures}
\begin{vquote}
\begin{figure*}
\centering
\includegraphics[width=\textwidth,height=2in]{Fig2.pdf}
\caption{Schematic of formation of the evanescent polariton on
linear chain of \PMS. The actual dispersion is determined by
the ratio of two coupling parameters such as exciton-\WGM
coupling and \WGM-\WGM coupling between the microspheres.}
\label{FIG:2}
\end{figure*}\end{vquote}
\subsection{Theorem and theorem like environments}
CAS class file provides a few hooks to format theorems and
theorem like environments with ease. All commands the
options that are used with \verb+\newtheorem+ command will work
exactly in the same manner. Class file provides three
commands to format theorem or theorem like environments:
\begin{enumerate}
\item \verb+\newtheorem+ command formats a theorem in
\LaTeX's default style with italicized font for theorem
statement, bold weight for theorem heading and theorem
number typeset at the right of theorem heading. It also
optionally accepts an argument which will be printed as an
extra heading in parentheses. Here is an example coding and
output:
\begin{vquote}
\newtheorem{theorem}{Theorem}
\begin{theorem}\label{thm}
The \WGM evanescent field penetration depth into the
cuprous oxide adjacent crystal is much larger than the
\QE radius:
\begin{equation*}
\lambda_{1S}/2 \pi \left({\epsilon_{Cu2O}-1}
\right)^{1/2} = 414 \mbox{ \AA} \gg a_B = 4.6
\mbox{ \AA}
\end{equation*}
\end{theorem}
\end{vquote}
\item \verb+\newdefinition+ command does exactly the same
thing as with except that the body font is up-shape instead
of italic. See the example below:
\begin{vquote}
\newdefinition{definition}{Definition}
\begin{definition}
The bulk and evanescent polaritons in cuprous oxide
are formed through the quadrupole part of the light-matter
interaction:
\begin{equation*}
H_{int} = \frac{i e }{m \omega_{1S}} {\bf E}_{i,s}
\cdot {\bf p}
\end{equation*}
\end{definition}
\end{vquote}
\item \verb+\newproof+ command helps to define proof and
custom proof environments without counters as provided in
the example code. Given below is an example of proof of
theorem kind.
\begin{vquote}
\newproof{pot}{Proof of Theorem \ref{thm}}
\begin{pot}
The photon part of the polariton trapped inside the \PMS
moves as it would move in a micro-cavity of the effective
modal volume $V \ll 4 \pi r_{0}^{3} /3$. Consequently, it
can escape through the evanescent field. This evanescent
field essentially has a quantum origin and is due to
tunneling through the potential caused by dielectric
mismatch on the \PMS surface. Therefore, we define the
\emph{evanescent} polariton (\EP) as an evanescent light -
\QE coherent superposition.
\end{pot}
\end{vquote}
\end{enumerate}
\subsection{Enumerated and Itemized Lists}
CAS class files provides an extended list processing macros
which makes the usage a bit more user friendly than the
default LaTeX list macros. With an optional argument to the
\verb+\begin{enumerate}+ command, you can change the list
counter type and its attributes. You can see the coding and
typeset copy.
\begin{vquote}
\begin{enumerate}[1.]
\item The enumerate environment starts with an optional
argument `1.' so that the item counter will be suffixed
by a period as in the optional argument.
\item If you provide a closing parenthesis to the number in the
optional argument, the output will have closing
parenthesis for all the item counters.
\item You can use `(a)' for alphabetical counter and `(i)' for
roman counter.
\begin{enumerate}[a)]
\item Another level of list with alphabetical counter.
\item One more item before we start another.
\begin{enumerate}[(i)]
\item This item has roman numeral counter.
\end{vquote}
\begin{vquote}
\item Another one before we close the third level.
\end{enumerate}
\item Third item in second level.
\end{enumerate}
\item All list items conclude with this step.
\end{enumerate}
\section{Biography}
\verb+\bio+ command have the below options:
\begin{enumerate}
\item \verb+width:+ Width of the author photo (default is 1in).
\item \verb+pos:+ Position of author photo.
\end{enumerate}
\begin{vquote}
\bio[width=10mm,pos=l]{tuglogo.jpg}
\textbf{Another Biography:}
Recent experimental \cite{HARA:2005} and theoretical
\cite{DEYCH:2006} studies have shown that the \WGM can travel
along the chain as "heavy photons". Therefore the \WGM
acquires the spatial dispersion, and the evanescent
quadrupole polariton has the form (See Fig.\ref{FIG:3}):
\endbio
\end{vquote}
\section[CRediT...]{CRediT authorship contribution statement}
Give the authorship contribution after each author as
\begin{vquote}
\credit{Conceptualization of this study, Methodology,
Software}
\end{vquote}
To print the details use \verb+\printcredits+
\begin{vquote}
\author[1,3]{V. {{\=A}}nand Rawat}[auid=000,
bioid=1,
prefix=Sir,
role=Researcher,
orcid=0000-0001-7511-2910]
\end{vquote}
\begin{vquote}
\cormark[1]
\fnmark[1]
\ead{<EMAIL>}
\ead[url]{www.cvr.cc, www.tug.org.in}
\credit{Conceptualization of this study, Methodology,
Software}
\address[1]{Indian \TeX{} Users Group, Trivandrum 695014,
India}
\author[2,4]{Han Theh Thanh}[style=chinese]
\author[2,3]{T. Rishi Nair}[role=Co-ordinator,
suffix=Jr]
\fnmark[2]
\ead{<EMAIL>}
\ead[URL]{www.sayahna.org}
\credit{Data curation, Writing - Original draft preparation}
. . .
. . .
. . .
\printcredits
\end{vquote}
\section{Bibliography}
For CAS categories, two reference models are recommended.
They are \file{model1-num-names.bst} and \file{model2-names.bst}.
Former will format the reference list and their citations according to
numbered scheme whereas the latter will format according name-date or
author-year style. Authors are requested to choose any one of these
according to the journal style. You may download these from
The above bsts are available in the following location for you to
download:
\url{https://support.stmdocs.in/wiki/index.php?title=Model-wise_bibliographic_style_files}
\hfill $\Box$
\end{document}
\section{Introduction}
Ocean oil spills have devastating impacts on marine eco\-systems and human society in the surrounding coastal areas. The 2010 Deep Water Horizon spill in the Gulf of Mexico lasted 87 days and was estimated to have released over 3 million barrels of oil. It impacted over 1,600 miles of coastline, killed over 8,000 marine animals/seabirds and caused direct economic loss from fishing and tour industries estimated at tens of billions of dollars \cite{graham:bp2011}.
There are certain areas that are highly sensitive to small amount of oil pollutants, such as shrimp farms. The toxic effect of hydrocarbons can lead to mortality, if the levels exceed the threshold concentration despite the fact that these areas can be geographically far from the center of oil spills \cite{bodkin_esler_rice_matkin_ballachey:2014, Gracia:2020}. Understanding and being able to predict these far-field impacts of ocean oil spill is critical for EPA first response teams and shrimp farmers to allocate resource to prevent loss and mitigate risk \cite{IPIECA:2004}.
How much oil (measured as oil concentration level) and how long it takes for surface oil to propagate to the highly sensitive areas are challenging problems. The transport of oil spilled into the ocean is a complex process that depends in a critical way on the current, wind, temperature and chemical composition of the oil and seawater \cite{mishra:oilspillwethering}. We focus our study on investigating the feasibility and effectiveness of using the Lattice Boltzmann Method for solving the Advection-Dif\-fusion Equation (LBM-ADE) to model and simulate ocean oil spill far-field impacts to address the aforementioned challenge problems.
We developed a prototype model and simulation in limited scale, a sub area of the Gulf of Mexico, with assimilation of real data from the UWIN-CM ocean current model \cite{uwin-cm:ocean-model}. We conducted two simulation experiments in comparison of the LBM-ADE model vs GNOME (General NOAA Operational Modeling Environment), a modeling tool which the Office of Response and Restoration’s (ORR) Emergency Response Division has been using to predict the possible oil transports on an ocean surface \cite{noaa:gnome}. GNOME is an ADE-based tool with Lagrangian particles, which relies on the accuracy of the ocean surface current velocity data to produce quality results. Our study shows the kinetic theory based LBM model outperforms the stochastic particle based GNOME model in accuracy and computation time due to their fundamental difference in representation of oil pollution advection and diffusion mechanisms. We propose the LBM as a viable alternative to the Lagrangian particle calculation component of GNOME in modeling the far-field impacts to highly sensitive areas.
There are many publications regarding using LBM to model and solve ocean flow problems. Wolf-Gladrow’s work \cite{gladrow:lbm} used the LBM to solve the linearized Munk Problem \cite{munk:oceanmodell}. In another LBM application of ocean models, Nuraiman \cite{nuraiman:lbm-ocean} used a 1D Shallow Water Equation representation of the Navier Stokes Equations coupled to the 2D Navier Stokes Equations to form an LBM model using the Bhatnagar, Gross and Krook (BGK) kinetic theory \cite{cercignani:bgk}. But to our knowledge there have been only a few studies of oil spill tracking using the LBM. One of the most comprehensive studies was done by \cite{maslo:lbmoill} and showed good agreement between simulated results and satellite observations from an oil spill in the Gulf of Beirut on July 15, 2006. Their LBM model used a two relaxation parameter technique to facilitate numerical stability. In addition, a flux limiter computational technique was used to resolve sharp numerical boundaries, which led to negative densities. Further, an interpolation technique was used to permit a non-square lattice to resolve the flow along the elongated coastline studied. In addition, Ha and Ku \cite{haku:lbmoil} used an LBM model to simulate an advective-diffusion formulation of the spread of an oil slick on the sea surface and confirmed the functionality of their model. Further, Li et al. \cite{lmk:lbmcde} solve the 2D convection-diffusion equation using the LBM. Other advection-diffusion equation solutions are presented by Dedits et al. \cite{dpsv:lbm-ade}. While not specifically studying oil transport, Li and Huang \cite{lihuang:lbm-shallow-water} used a coupled LBM formulation of the Shallow Water Equation and Contamination Concentration Transport. Excellent agreement was obtained between numerical predictions and analytical solutions in the pure diffusion problem and convection–diffusion problem. Banda and Seaid \cite{banda:lbm-shallow-water} also developed an LBM model to solve shallow water equations as the depth-averaged incompressible Navier-Stokes equations with conservation of momentum under the assumption that the vertical scale is much smaller than any typical horizontal scale and the pressure is hydrostatic. Then they apply their shallow water model to simulate pollutant transport in the Strait of Gibraltar.
Our literature review shows most of ocean oil spill and contamination transport models are based on the ADE. In addition to the above cited research, GNOME is an ADE-based tool, which has been used to predict the possible route, or trajectory, a pollutant might follow on the surface of water. It relies on the accuracy of the ocean surface current velocity field to produce quality results. Also, GNOME is based on simulating an ensemble of stochastic particles such that the accuracy depends crucially on the number of particles in the simulation. In contrast, LBM computes the distribution function directly and is expected to be computationally more efficient in situations where GNOME would require an exceedingly large number of particles in order to provide an accurate solution.
During the first phase of our research, we developed an LBM-ADE model and simulation that is capable of providing numerical solutions as an LBM-ADE solver. To validate the model, we performed a benchmark test using a Gaussian Hill concentration with a simplified velocity field first. However, the ocean surface current constitutes a much more complex velocity field that is temporal-spatial dependent. We tested the LBM-ADE solver against a Finite Differential Method (FDM) ADE solver using a perturbation of the Taylor-Green velocity field. To the best of our knowledge, no such benchmark has been done in the past for an LBM-ADE model using a velocity field as complex as the perturbed Taylor-Green field. Our first phase study shows the LBM-ADE model achieves great results in comparisons vs analytical solution in Gaussian Hill case and vs the FDM-ADE solution in Taylor-Green case \cite{zzhang:ams2020}.
Afterward, we conducted a benchmark study of LBM-ADE and GNOME against an analytical ADE solution and a comparison study of the LBM-ADE model vs the GNOME model in prediction the surface transport of spilled oil in the same sub area of the Gulf of Mexico, with assimilation of real data from the UWIN-CM ocean current model.
We achieved accurate results using Gaussian Hill concentrations with two ocean current scenarios: linear ocean current and real ocean current from (UWIN-CM). These results suggest that the LBM-ADE is a promising model that is capable of predicting spilled oil transport on an ocean surface \cite{zzhang:wsc2020}.
\section{Methodology}
\subsection{GNOME Model}
GNOME \cite{noaa:gnome} is an ADE-based model using Lagrangian particles, referred as Lagrangian elements (LEs), to represent oil pollutants. All the LEs trajectories collectively represent the path of oil transport on the ocean surface. In a simplified ocean oil spill scenario without considering weathering process, each LE is moved independently by two forces, namely advection and diffusion. Furthermore GNOME assumes the advection and diffusion processes are independent of each other.
GNOME models the advection process using a forward Euler scheme. Assume at time step $t$, a LE is at the point $p(x,y,t)$. It calculates the LE position after one time step $t+\Delta t$ at the point $p(x+\Delta x, y+\Delta y, t+\Delta t)$ using the velocity field for each element.
Diffusion is modeled as stochastic processes where a set of LEs engage a 2-D random walk with a displacement probability such that the mean value remains zero, but the variance grows linearly with time. It has been shown that a long series of random steps will converge to a Gaussian distribution with variance growing linearly with time \cite{csanady:diffusion}. For a given input of diffusion coefficient, GNOME uses a particle random walk simulation, a stochastic process, to approximate the diffusion process.
\subsection{LBM-ADE Model} \label{sec:lbm-ade-model}
We use the LBM to model ocean oil pollutants as a set of particles with certain density and mass located on a virtual grid (lattice) that maps over an area of ocean with boundary conditions representing coastal lines or islands. This model makes it possible to track particle spatial positions and microscopic momenta from a continuum to just a handful and similarly discrete in distinct steps. Particle positions are confined to the nodes of the lattice. Variations in momentum that could have been due to a continuum of velocity directions and magnitudes and varying particle mass are reduced (in a simple 2D model) to 9 directions and a single particle mass \cite{sukop:lbm-geoscientist}. Figure \ref{fig:lbm-particles} shows the Cartesian lattice and the velocities $e_a$ (where a = 0, 1 … 8) is a direction index and $e_0 = 0$ denotes particles at rest. This model is known as D2Q9 as it is 2 dimensional and contains 9 velocities. It can be generalized to a 3 dimensional model as D3Q27 if we replace the lattice in D2Q9 with a cube with length, width and height are one lattice unit.
The next step is to incorporate the single-species distribution function f, which has only nine discrete ‘bins’ instead of being a continuous function. The distribution function can conveniently be thought of as a histogram representing a frequency of occurrence. For example the shaded area in Figure \ref{fig:lbm-particles} shows a likely oil pollutant propagation pattern after one time step.
\begin{figure}[htb]
{
\centering
\includegraphics[width=0.45\textwidth]{figs/LBM_ParticleDist.png}
\caption{The particle distribution function represents the percentage of particles in the velocity bins.\label{fig:lbm-particles}}
}
\end{figure}
Accordingly, the macroscopic fluid density is defined as:
\begin{equation} \label{eq:lbm-mass}
\rho = \sum_{a=0}^{8} f_a
\end{equation}
The macroscopic velocity $u$ is an average of the microscopic velocities $e_a$
weighted by the directional densities $f_a$ as defined as:
\begin{equation} \label{eq:lbm-velocity}
u=\frac{1}{\rho}\sum_{a=0}^{8} f_a e_a
\end{equation}
This simple equation allows us to pass from the discrete microscopic velocities that comprise the LBM back to a continuum of macroscopic velocities representing the fluid’s motion.
When incorporating external forces, such as wind, gravity and others, that interact with the ocean water, Equation \eqref{eq:lbm-force}
can be modified as:
\begin{equation} \label{eq:lbm-force}
u=\frac{1}{\rho}\sum_{a=0}^{8} f_a e_a + \frac{F\Delta t}{2\rho}
\end{equation}
where the first term is the velocity due to mass density redistribution with conservation of momentum and the second term is due to external forces \cite{kkkssv:lbm}.
Equation \eqref{eq:lbm-force} is a generalization of the LBM that is applicable to both NSE and ADE models. In the LBM-NSE model, Equation 3 provides a mathematical base for developing a velocity projection schema that integrates the ocean surface velocity into a LBM model as an external input, and then uses it to update local equilibrium distribution functions $f^{eq}$. On the other hand, in the LBM-ADE model, we ignore the first term in \eqref{eq:lbm-force} and only consider the second term as an advective velocity resulting from external forces since ADE only conserves mass, not momentum.
The next steps are streaming and collision of the particles via the distribution function. The simplest approach to approximate the collision can be defined as:
\begin{equation} \label{eq:lbm-collison}
f_a(x+e_a\Delta t, t+\Delta t)=f_a(x,t)- \frac{f_a(x,t)-f_a^{eq}(x,t)}{\tau}
\end{equation}
where $\tau$ is a relaxation time used in the BGK operator. Although they can be combined into a single statement as above, collision and streaming steps must be separated if solid boundaries are present because the bounce back boundary condition is a separate collision. Collision of the fluid particles is considered as a relaxation towards a local equilibrium. \hfill The parameter $\tau$ is a relaxation time to reach equilibrium. \\*
A D2Q9 equilibrium distribution function $f^{eq}$ is defined as:
\begin{equation} \label{eq:lbm-feq}
f_a^{eq}(x,t)=w_a\rho(x,t)\left(1+3\frac{e_au}{c^2}+\frac{2(e_au)^2}{9c^4}+\frac{3u^2}{2c^2}\right)
\end{equation}
where the weights $w_a=\left(\frac{4}{9},\frac{1}{9},\frac{1}{9},\frac{1}{9},\frac{1}{9},\frac{1}{36},\frac{1}{36},\frac{1}{36},\frac{1}{36}\right)$ and c is the velocity on the lattice, one lattice unit per time step $(lu/\Delta t)$ in the simplest implementation. Note that if the macroscopic velocity u = 0, the equilibrium $f_a^{eq}$ are simply the weights times the fluid density.
To implement the LBM model as a simulation program, \cite{bao:lbm-fluid} presented an algorithm outline that can be summarized as follows:
\begin{enumerate}
\item Initialize $\rho$, $\mu$, $f_a$ and $f_a^{eq}$;
\item Streaming step: move $f_a$ $\rightarrow$ $f_a^*$ in the direction of $e_a$, where $f_a^*$ holds intermediate values of density distribution after the streaming step.
\item Compute macroscopic $\rho$ and $\mu$ from $f_a^*$ using above equations \eqref{eq:lbm-mass} and
\eqref{eq:lbm-velocity};
\item Compute $f_a^{eq}$ using
equation \eqref{eq:lbm-force};
\item Collision step: calculate the updated distribution function using equation \eqref{eq:lbm-collison}:
$f_a=f_a^*-\frac{f_a^*- f_a^{eq}}{\tau}$;
\item Repeat steps 2 to 5.
\end{enumerate}
During the streaming and collision step, the boundary nodes require some special treatments for the distribution functions in order to satisfy the imposed macroscopic boundary conditions.
The LBM as described has been shown to be second order accurate in time and space to the 2D incompressible Navier Stokes Equations by \cite{kkkssv:lbm} and separately by \cite{gladrow:lbm}.
LBM is a kinetic theory based modeling technique that can provide numerical solutions for a range of flow problems whose underline physics are governed by NSE and/or ADE. When the NSE and the ADE are applied in near incompressible fluids, they can be expressed as equations \eqref{eq:lbm-nse-pde} and \eqref{eq:lbm-ade-pde}. The Navier Stokes Equations are given by
\begin{equation} \label{eq:lbm-nse-pde}
\frac{\partial u}{\partial t}+u\nabla u=-\frac{\nabla p}{\rho}+\nu\nabla ^2 u+F
\end{equation}
where $u$ is fluid velocity, $P$ is fluid pressure, $\rho$ is fluid density, $\nu$ is fluid kinematic viscosity, and $F$ is an external force. The Advection-Diffusion Equation is written as
\begin{equation} \label{eq:lbm-ade-pde}
\frac{\partial C}{\partial t}+u\nabla C=D\nabla ^2 C+q\,
\end{equation}
where $C$ is mass concentration, $D$ is diffusion coefficient (assume isotropic diffusion), $u$ is fluid velocity as an advection force; and $q$ is a source term.
The LBM-NSE model is defined by the set of equations \eqref{eq:lbm-mass} to \eqref{eq:lbm-feq}. While the LBM-NSE model conserves both mass and momentum, the LBM-ADE model only conserves mass (referred as concentration $C$). A LBM-ADE model is defined by the set of equations \eqref{eq:lbm-ade-concentration} to \eqref{eq:lbm-ade-geq} as below:
\begin{equation} \label{eq:lbm-ade-concentration}
C = \sum_{a=0}^{8} g_a
\end{equation}
\begin{equation} \label{eq:lbm-ade-collison}
g_a(x+e_a\Delta t, t+\Delta t)=g_a(x,t)- \frac{g_a(x,t)-g_a^{eq}(x,t)}{\tau_g}
\end{equation}
\begin{equation} \label{eq:lbm-ade-geq}
g_a ^{eq}(x,t)=w_aC(x,t)\left(1+3\frac{e_a u}{c^2}+\frac{2(e_a u)^2}{9c^4}+\frac{3u^2}{2c^2}\right)
\end{equation}
\begin{equation} \label{eq:lbm-ade-D}
D_l=c^2 \left( \tau_g - \frac{\Delta t}{2} \right)
\end{equation}
where the $g_a$ are the directional densities of concentration; $g_a^{eq}$ is the equilibrium density function; $\tau_g$ is the relaxation time and $u$ is a velocity vector due to advection forces, while $e_a$ and $w_a$ are the same as previously defined in \eqref{eq:lbm-velocity} and \eqref{eq:lbm-feq}.
Since the equations for the NSE \eqref{eq:lbm-collison} and equation for the ADE \eqref{eq:lbm-ade-collison} are the same, the algorithm outlined by \cite{bao:lbm-fluid} is also applicable to the ADE.
\section{Model Configurations and Simulations}
LBM and GNOME are both ADE based modeling tools that share the same governing physics laws that can be defined as a partial differential equation \eqref{eq:lbm-ade-pde} where we set $q=0$ except in the initial condition for a one-time oil spill release.
While the LBM model takes a kinetic approach and the GNOME model takes a stochastic approach, they both provide numerical solutions that approximate and converge to the analytical solution of equation \eqref{eq:lbm-ade-pde}.
But there are differences between them which present several challenges in the evaluation of their far-field impacts. To conduct a scientfically sound comparison between LBM and GNOME, we need to carefully design and configure our experiments taking into consideration the following issues: (i) LBM works with a Eulerian specification of flow field, while GNOME works with a Lagrangian specification of flow field; (ii) LBM uses units of measurement of space and time in lattice units, while simulations in GNOME uses units of measurement in meters and seconds; (iii) LBM uses a square lattice to represent a ocean surface area of $1^{\circ}$ by $1^{\circ}$ in longitude and latitude, while GNOME uses a non-square area of 97.904 km by 111.194 km due to the curvature of the earth's surface in longitude and latitude; and (iv) we need to specify the oil spill volumes and diffusion coefficients that are equivalent in both units of measurement in LBM and GNOME.
\subsection{Far-Field Effect Benchmark}
To assist in analyzing the low density effectiveness of GNOME and LBM-ADE we provide a benchmark example which calculates the elapsed time,$ T_{thresh}$, and location of the maximum Y extent, $Y_{thresh}$, of the minimum non-zero normalized oil concentration from a hypothetical point source spill with a constant linear current, $ (U_x,U_y)= (0.0,-1.0)$. For this particular case of a constant advection field, we can calculate the theoretical probability density values
based on the analytical solution of the corresponding ADE and relate these results to the Monte Carlo analysis of the LEs in GNOME. Details of this calculation are provided in the Appendix. In the case for the far-field effect, the stochasticity parameter $\sigma$ in the underlying probability density is given by $\sigma=1.414$. The non-dimensional scaling parameters are: length $(l=100.0)$, time $(t = 6.42)$, and velocity, $v=(0.0,-1.0)$.
GNOME is then run with this configuration and diffusion, $D_g= 297761.0 cm^2/sec$. The number of LEs at $T_{thresh}$ and $Y_{thresh}$ is simulated and the probability is calculated. To align the initial LE distribution with the initial Gaussian used in the LBM, GNOME is run for 32 time steps with a zero advective velocity field corresponding to setting a current=(0.0,0.0) .
Next, LBM-ADE is run with equivalent parameters and $\tau_g=0.760$ and the normalized concentration, $C_l$, at $T_{thresh}$ and $Y_{thresh}$ is calculated.
The resulting values of the probability and normalized concentrations are presented in Table~\ref{tab:benchmark-gnome-lbm}, where $T_{thresh}$ includes the additional time steps for the development of the initial Gaussian distribution. The results show good agreement among GNOME and the LBM-ADE.
Figure \ref{fig:far_field_benchmark} shows the results of the far-field effect, low density, benchmark at $T_{thresh}$ in our computational domain with constant southward velocity at 0.026778 m/s. The red dots are the locations of the 10,000 LEs The green dot represents the only LE north of $Y_{thresh}$ shown as the green dashed line.
The contours show the normalized concentration results from the LBM-ADE simulation.
These results show that LBM-ADE model can provide information of far-field effect with lower oil density to the area beyond the $Y_{thresh}$ while GNOME model can not, since there are no particle presents beyond $Y_{thresh}$.
\begin{figure}[htb]
{
\centering
\includegraphics[width=0.45\textwidth]{figs/far_field_benchmark.png}
\caption{The Far-Field Effect Benchmark for both GNOME and LBM.\label{fig:far_field_benchmark}}
}
\end{figure}
\begin{table*}[htb]
\centering
\caption{Benchmark Results Compares Results from GNOME and LBM-ADE\label{tab:benchmark-gnome-lbm}}
\begin{tabular}{llllll}
\hline
Method & Probability & Total LE Count & $Y_{thresh}$ & $T_{thresh}$ & LEs beyond Threshold\\ \hline
GNOME & $1.00x10^{-4}$ & 10,000 & $28.87^{\circ}$ & 306,000s & 1\\
LBM & $1.20x10^{-4}$ & NA & $1.14x10^{2}$ in lattice length & 296 in lattice time & NA\\
Analytic ADE & $1.22x10^{-4}$ & NA & 7.42 & 6.42 & NA \\
\hline
\end{tabular}
\end{table*}
\subsection{GNOME and LBM ADE Model Configurations and Results Analysis}
In this study, we compare the LBM-ADE solution with GNOME in the same configuration with a velocity field that comes from a real ocean data model (UWIN-CM) to evaluate their effectiveness of far-field impact of ocean oil spill.
UWIN-CM is a fully coupled atmosphere–wave–ocean system data model. We used a subset of data from the UWIN-CM ocean model to cover a 1 degree square area of the Gulf of Mexico centered at -88.4 longitude and 28.8 latitude over three days from Feb. 07, 2016 at 16:00:00 until Feb. 10, 2016 at 18:00:00.
We used bi-linear interpolation spatially to generate a velocity field for the LBM computational domain at each time step while $\Delta t=15$ minutes. The ocean surface velocity field is assimilated in the LBM-ADE model as an advection velocity at each time step. In GNOME the ocean current data is used in the same way except the velocity data and time steps are stored in a netCDF file, then loaded into the model during model configuration.
We introduce a one-time oil spill as a Gaussian hill in the center of the grid at time step $t=0$ as an initial oil spill in both the GNOME and the LBM-ADE. The volume of a Gaussian hill is a function of $C_0$ and $\sigma_0$ which are calculated to be the volume of oil specified in GNOME. GNOME uses a point source of 10,000 LEs to represent the oil volume. Then we let the models run 296 time steps, with $\Delta t=15$ minutes, a total of 74 hours.
Since in GNOME the initial oil spill is modeled as a point source release, all the LEs are located at the center point of the grid instead of a Gaussian hill distribution as in LBM-ADE. It is interesting to point out that, regardless this initial oil distribution difference, the GNOME concentration diffusion converges to LBM-ADE after a transient period of approximately 40 time steps.
We show the comparison results of both LBM-ADE and GNOME in snap shots at time steps 2, 50, 100 and 150 in Figure \ref{fig:lbm-gnome-ocean} with background of ocean current in the study domain of Gulf of Mexico, where the contours show oil density in LBM and the red color particles are LEs in GNOME. These times were chosen to omit the GNOME transient effect and the loss of concentration at the north eastern boundary.
Figure \ref{fig:lbm-gnome-ocean} clearly shows that the contours can reach far out from the center mass of oil concentration where there are no presents of particles. Therefore, the LBM-ADE model is more effective to model the far-field impact of ocean oil spill to certain highly sensitive areas in terms of predicting how far a small amount of oil can reach and how long it takes.
\begin{figure*}[htb]
{
\centering
\includegraphics[width=.45\textwidth]{figs/lbm_gnome_map1.jpg} \hfill
\includegraphics[width=.45\textwidth]{figs/lbm_gnome_map2.jpg} \hfill
\includegraphics[width=.45\textwidth]{figs/lbm_gnome_map3.jpg} \hfill
\includegraphics[width=.45\textwidth]{figs/lbm_gnome_map4.jpg} \hfill
\caption{LBM oil density and GNOME particles with ocean current.}
\label{fig:lbm-gnome-ocean}
}
\end{figure*}
\subsection{Discussion of GNOME Diffusion and LBM BGK Collision and Their Relation}
GNOME incorporates diffusion as horizontal eddy diffusivity in the water using the model of a random walk which theoretically corresponds to the solution of Advection-Dif\-fusion Equation (\ref{eq:lbm-ade-pde}) where $D$, referred to $D_g$, corresponds to an isotropic diffusion \cite{noaa:technotes}:
\begin{quote}
"Diffusion is a way to capture [...] small scale circulations that [are] not captured in the underlying circulation model. This is often known as subgrid scale circulation. These small eddies [...] serve to spread things out, or diffuse them."
\end{quote}
To effectively represent diffusion during a spill the value is calibrated using over flight data \cite{noaa:gnome}. The process is somewhat of an art which, constrained by the current grid scale and geographic complexity, strives to represent the spread without diluting the concentration. The default value used by GNOME is $1 \times 10^5 \ cm^2/sec$.
In our LBM simulation, diffusion is incorporated into the model through the relaxation parameter, $\tau_g$, based on Equation (\ref{eq:lbm-ade-D}). The physical scaling used in GNOME is used to calculate the equivalent $\tau_g$ from the value of $D$. To avoid numerical instability, $\tau_g>0.50$. While this limits the LBM for very small values of $D$, for modeling oil spills this does not present a significant limitation since for small $D$ the trajectory of the oil is driven primarily by the current . In this analysis we have used various values given in the Table~\ref{tab:diffusion-parameters}. It shows the LBM-ADE model and its numerical solutions are robust enough to handle a wider range of values of $\tau_g$.
Table~\ref{tab:diffusion-parameters} shows the values of the corresponding diffusion for a non-dimensional Monte Carlo model, GNOME, LBM-ADE and $\tau_g$. While $D$ and $D_g$ are independent of the numerical stability of their respective models, $D_l$ is directly related to $\tau_g$ which contributes to the numerical stability of the LBM-ADE.
\begin{table}[htb]
\centering
\caption{Diffusion Parameters in Monte Carlo, GNOME and LBM-ADE Models.For this example the non-dimensional scaling parameters are:
length $(l=10.0)$, time $(t=2.0)$, and velocity $(v=(0.0,-1.0)$.\label{tab:diffusion-parameters}}
\begin{tabular}{llll}
\hline
$D$ & $D_g(cm^2/s)$ & $D_l(l^2/lt)$ & $\tau_g(lt)$ \\ \hline
$0.28125$ & $2.61x10^6$ & $0.760134$ & $2.78040$ \\
$0.2112$5 & $1.96x10^6$ & $0.507945$ & $2.21284$ \\
$0.15125$ & $1.40x10^6$ & $0.408783$ & $1.72635$ \\
$0.10125$ & $9.04x10^5$ & $0.273648$ & $1.32095$ \\
$0.06125$ & $5.68x10^5$ & $0.165540$ & $0.99662$ \\
$0.03125$ & $2.90x10^5$ & $0.084459$ & $0.75338$ \\
$0.01125$ & $1.04x10^5$ & $0.030405$ & $0.59122$ \\
$0.00125$ & $1.16x10^4$ & $0.003378$ & $0.51014$ \\
\hline
\end{tabular}
\end{table}
\section{Complexity Analysis and Performance Evaluation}
In our study, both GNOME and LBM are ADE based models which provide numerical solutions that approximate the propagation of oil pollutants on ocean surface. In addition to evaluating their effectiveness of modeling ocean oil spill far-field impacts, we conducted analytical analyses and empirical tests to compare their performance in terms of accuracy and computation complexity.
\subsection{Accuracy of the LBM and GNOME} \label{lbm_gnome_accuracy}
Using an analysis based on the Chapman-Enskog expansion, Kruger \cite{kkkssv:lbm} shows that the Lattice Boltzmann Equation with the BGK collision operator (\ref{eq:lbm-ade-collison}) and quadratic equilibrium (\ref{eq:lbm-ade-geq}) recovers the ADE (\ref{eq:lbm-ade-pde}) up to $O(\Delta t ^2)$ and $O(\epsilon^2)$ where $\epsilon$ is the Knudsen number. Moreover, they show that the resulting error term is independent of the velocity. In \cite{Zhang-Shi:2012}, T. Zhang, et al. also analyze the convergence of this LBM model and present numerical experiments which demonstrate the $O(\epsilon^2)$ convergence in space. Above reference \cite{kkkssv:lbm} also notes that the error term can be eliminated by introducing an artificial source term as shown in \cite{Chai-Zhao:2013}. In the same work \cite{Chai-Zhao:2013}, Chai and Zhao show that an LBM model with an artificial source term is exactly $O(\epsilon^2)$ convergent in space and the authors present numerical examples showing that their revised LBM model is slightly more accurate than \cite{Zhang-Shi:2012} for the same benchmark problem.
For our LBM model using a $(lx,ly)$ lattice in a $1^{\circ} \times 1^{\circ}$ computational domain the theoretical accuracy of the normalized density is $O(\epsilon^2)$, $1/(1^{\circ} \times 1^{\circ})$ which for $lx=ly=200$ is $2.5 \times 10^{-5}$. In addition, smaller concentrations can be calculated in the far field limited by numerical round-off error, which may be smaller than the theoretical accuracy and provide valuable information.
The "mover" components of GNOME which will be replaced by LBM are the LE trajectories and diffusion. As far as the trajectories, GNOME uses a forward Euler method \cite{noaa:gnome} which is locally $O(\Delta t ^2)$ \cite{Isaacson:1966, stoer:Numerical-Analysis} and globally $O(\Delta t)$. As far as the diffusion, GNOME uses classical diffusion, see Equation \ref{eq:lbm-ade-pde2}, which is modeled by a stochastic differential equation using a Monte Carlo method based on Brownian Motion.
\begin{equation} \label{eq:lbm-ade-pde2}
\frac{\partial C}{\partial t}=D\nabla ^2 C
\end{equation}
The accuracy of the stochastic differential equation solved is $O(\sqrt\Delta t)$ \cite{Haugh:2017,Glasserman:2003,Gardiner:2009}, here the size of $\Delta t$ determines the degree of (non-)smoothness of the LE paths. For evaluating {\em rare events} for comparison with LBM, the density of the smallest detectable concentration is determined by the number of LEs rather than the smoothness of the path and time step.
For our GNOME example, the spatial resolution of the concentration of particles is $(\frac 1 N$ $\frac A {\Delta x \Delta y})$ where $A$ is the study area and $\Delta x$ and $\Delta y$ are the grid size of the bins used for accumulating particles for the purpose of calculating densities. We use $\Delta x=200, \Delta y=200$ and an area of $1^{\circ} \times 1^{\circ}$ centered at (28.8,-88.4). The smallest detectable non-zero probability of a particle in a bin is $1/N$ which is $1 \times 10^{-4}$. Throughout the far field where there are no particles the probability is constant equal to 0.
\subsection{GNOME and LBM-ADE Computation Complexity Comparison}
GNOME simulates the ocean oil spill surface propagation based on an ensemble of Lagrangian particle trajectories which are combination of particle advection movements and stochastic Brownian motions with specified diffusion coefficient, $D_g$. The basic computation unit is a Lagrangian element (LE) which consumes certain computation resources (CPU/GPU time and memory space) in each time step. The total computation complexity in a given GNOME simulation is proportional to the total number of LEs and simulation time steps, $O(N_{le}) \times N_s$, where $N_{le}$ is the number of LEs and $N_s$ is the number of simulation steps.
On the other hand, LBM-ADE uses a kinetic process to simulate ocean oil spill surface propagation based on oil concentration density and movements on a lattice grid which are governed by a stream-collision equation with the BGK operator \eqref{eq:lbm-ade-collison} and the equilibrium equation \eqref{eq:lbm-ade-geq}. The LBM-ADE numerical solution is computed using an algorithm presented in Bao and Meskas \cite{bao:lbm-fluid}. In this case, the basic computation unit is a lattice node. The total computation complexity in a given LBM-ADE simulation is proportional to the total number of lattice nodes and simulation time steps, $O(N_{l}^2) \times N_s$, where $N_{l}^2$ is the total number of lattice nodes in a 2D square lattice and $N_s$ is the number of simulation steps.
We conducted simulation tests to generate empirical data sets which are used to construct computation complexity profiles for both GNOME and LBM-ADE models. For each simulation run, we use the same number of simulation steps (297 steps) in both GNOME and LBM-ADE models, so that we only need to consider the computation times related to total number LEs in GNOME and total number lattice nodes in LBM-ADE respectively. We implemented a Python version of GNOME simulation program using pyGNOME API published by NOAA \cite{noaa:technotes} and we also implemented a Python version of the LBM-ADE simulation program. Both simulation programs run on a Dell Latitude E6430 laptop computer with Intel dual-core CPU at 2.90 GHz, 8 GB RAM and 64-bit Windows 10 operating system.
Figure \ref{fig:lbm_time} shows the LBM-ADE computation time profile as the lattice size ($N_l$) changes from 50 to 500 with incremental of 50. The elapsed time is averaged over 10 simulation runs for each lattice size. It shows the LBM-ADE computation time scaling agrees with the estimated scaling of $O(N_{l}^2)$.
\begin{figure}[htb]
{
\centering
\includegraphics[width=0.45\textwidth]{figs/lbm_time.png}
\caption{LBM-ADE computation time profile.\label{fig:lbm_time}}
}
\end{figure}
Figure \ref{fig:gnome_time} shows GNOME computation time profile as the number of LE changes from 10,000 to 200,000. The elapse time is averaged over 10 simulation runs for each total number of LEs It shows the GNOME computation time is linear related to the total number of LEs.
\begin{figure}[htb]
{
\centering
\includegraphics[width=0.45\textwidth]{figs/gnome_time.png}
\caption{GNOME computation time profile.\label{fig:gnome_time}}
}
\end{figure}
Referring to the example mentioned in section \ref{lbm_gnome_accuracy}, the LBM-ADE model with a (200 x 200) lattice can achieve accuracy of the normalized density at level of $2.5 \times 10^{-5}$ and it takes 6.91 seconds as Figure \ref{fig:lbm_time} shows. For GNOME to reach the same level of accuracy, it would need 40000 LEs and it will take 8.40 seconds as Figure \ref{fig:gnome_time} shows. we found that, in this case, the LBM-ADE and the GNOME are comparable with LBM-ADE slightly outperforms GNOME in computation times. As Figure \ref{fig:far_field_benchmark} shows if we were estimating extremely low oil density at $1 \times 10^{-6}$ level, LBM-ADE will be far more advantage than GNOME in computation time.
\subsection{Potential Performance Gains in Parallel Processing}
Both LBM and GNOME algorithms lend themselves to parallel computing
techniques. Since each time step of the LBM algorithm can be
structured so that all calculations on nodes are locally dependent,
either depending only on the node or at most depending on nodes one
lattice edge apart, the algorithm is favorable to parallelization.
The explicit time step of the LBM process can be divided into two sub-steps (see sec. \ref{sec:lbm-ade-model}):
\begin{enumerate}
\item The first is comprised of the relaxation, the equilibrium and
the macroscopic calculations which depend only on the node being
updated. Parallel domain partitioning techniques \cite{huelsemann:2006} enable simultaneous update of each
node.
\item The second is streaming, which depends on the node being updated
and the immediate neighboring nodes. Parallel domain partitioning
techniques and ghost cells for sub-domain boundaries enables
parallel processing of each sub-domain simultaneously.
\end{enumerate}
Good speed-ups have been reported for multiple PUs (processing units)
by Körner C., et al \cite{koerner-etal:2006} for 64 CPU and linear
scale-up (fixed CPU load by scaling the domain size). In addition,
effective massive parallel techniques using GPGPUs (General Purpose
Graphics Processing Units) and CUDA are described by Jan{\ss}en and
Krafczyk \cite{janssen-krafczyk:2011} with "one order of magnitude
faster than comparable CPU implementations". Parallel threads
techniques using OpenMP and GPUs where used for an LBM solution of the Shallow Water Equations by Kevin Tubbs where linear speed-up was
obtained for up to 6 threads for OpenMPI \cite{tubbs:2010,tubbs-tsai:2010}.
Parallelizing each time step of the particle based algorithms in
GNOME is straightforward because each particle moves independently of each other particle. These algorithms are often referred to as
{\em embarrassingly parallel}. Typical parallel techniques such as using
threads, multiple CPUs and message passing and GPU can be used with
various speed-up improvements \cite{anderson-etal:2013,lang-prehl:2017}.
\section{Conclusion}
We study and evaluate GNOME and the LBM-ADE model in their capability and effectiveness in terms of modeling ocean oil spill far-field impacts to certain sensitive areas. The results show that
GNOME is not necessarily the most appropriate tool to use for analyzing far-field impact with very small oil densities.
The LBM-ADE model, on the other hand, can provide numerical estimations to certain far-field area with very low oil density while GNOME cannot. In their performance evaluations, we find the LBM-ADE model performs better in accuracy and computation complexity. For these reasons, the LBM-ADE model is a viable alternative to GNOME in modeling the far-field impacts to highly sensitive areas.
For future research, We will explore the possibilities of using LBM-NSE as well as LBM-ADE to model multi species and multi phase flows. This will enhance our ability to model the ocean oil weathering process, such as mixed water and oil droplets in ocean water columns and their far-field impacts to marine lives beneath ocean surface.
\section{Appendix: Estimating Limits of Monte-Carlo Simulations}
In some special cases, it is possible to evaluate the limits of Monte-Carlo simulations exactly. The results of this analysis can be used to create benchmark situtations for the numerical algorithms. To illustrate one particular approach, we consider the simple case of a constant advection field in $y$-direction with $v_y<0$ such that the oil particles will be advected downward and no advection in the $x$-direction ($v_x=0$). In this case, we can solve the advection-diffusion equation explicitly. For a $\delta$-initial condition corresponding to a release of the oil at the origin, the probability density of finding a particle at a location $(x,y)$ after a time $T$ is given by
\begin{equation}
p(x,y,T) = \frac{1}{2\pi\sigma^2T}{\mathrm{e}}^{-(x^2+(y-v_yT)^2)/(2\sigma^2T)}
\end{equation}
From there, we can compute the probability of finding a particle above $y>L>0$. For typical parameter values, corresponds to a rare event as the advection is pulling the particle downwards and only a few particles will be able to reach a point with $y>L$. The exact probability $P$ is found from the probability density above as
\begin{eqnarray}
P(T) &=& \int_{-\infty}^{\infty} dx\int_L^{\infty}p(x,y,T),\,dy \nonumber \\
&=& 1-\Phi\left(\frac{L-v_yT}{\sigma\sqrt{T}}\right)
\end{eqnarray}
Here, $\Phi$ denotes the cumulative normal distribution. This result can be interpreted as follows: For very small times, it is unlikely for particles to reach the region $y>L>0$ simply because the particles need time to diffuse away from the source.
\begin{figure}[htb]
{
\centering
\includegraphics[width=0.35\textwidth]{figs/prob_T.png}
\caption{Example illustrating the probability $P(T)$. Here, $v_y = -1$, $\sigma = 1$, and $L = 1$. The time $T$ ranges from 0 to 10 .\label{fig:prob_T}}
}
\end{figure}
As the time increases, particles leave the source and they are being diffused and advected. Most particles are moved downward by the drift $v_y<0$, away from the region $y>L>0$. However, for some particles corresponding to a few rare events, the diffusion happens to act against the drift and there is a chance for those particles to reach this region. As time grows, however, it becomes again less and less likely for the particles to reach this region since the drift is constantly dragging them downward. Figure \ref{fig:prob_T} illustrates this dependence of the probability on the time $T$.
Assume now that we run a Monte-Carlo simulation with $N_{max}$ particles. Clearly, if we expect less than one particle to reach the area with $y>L$, we will not be able to rely on Monte-Carlo simulations to collect sufficient statistics about that region. Therefore, setting $P= 1/N_{max}$ gives us an estimate for the limit of the Monte-Carlo method. Using the formula above, we can compute the maximum distance $L$. Solving for $L$ we obtain
\begin{equation}
L = v_yT + \sigma\sqrt{T}\Phi^{-1}\left(1-\frac{1}{N_{max}}\right)
\end{equation}
where $\Phi^{-1}$ denotes the inverse cumulative normal distribution. Note that $L=L(T)$ is still a function of the time $T$. However, it is easy to see that, as a function of $T$, its maximum is given by
\begin{equation}
L_{max} = \frac{1}{4|v_y|}\sigma^2\left(\Phi^{-1}\left(1-\frac{1}{N_{max}}\right)\right)^2
\end{equation}
which allows us to estimate the limit of Monte-Carlo simulations in this particular case.
\section*{Acknowledgments}
Computational support was provided by The City University of New York High Performance Computing Center, which is operated by the College of Staten Island funded, in part, by the National Science Foundation grants CNS-0958379 and CNS-0855217, and DMS-2012548. This pro\-ject was also supported, in part, by the PSC-CUNY Research Award 63570-00 51.
We thank Robyn N. Conmy, PhD. US EPA Office of Research and Development, Remediation Pollution Control Division for her invaluable comments and guidance in understanding environmental transport and weathering of oil. We further thank Andrew Poje, PhD. Professor of Mathematics, College of Staten Island for his assistance with UWIN-CM.
\bibliographystyle{cas-model2-names}
|
\section{Introduction}
In the era of high-cadence, all-sky surveys, the processing and analysis of the vast amounts of resulting data is a major challenge. To date, a number of wide-field surveys have been commissioned to conduct large, repeated photometric surveys both in the optical (e.g., the Sloan Digital Sky Survey \linebreak (SDSS; \citealt{York2000}), the Palomar Transient Factory (PTF; \citealt{Law2009}, \citealt{Rau2009}) and Panoramic Survey Telescope and Rapid Response System (Pan-STARRS; \citealt{Hodapp2004}, \citealt{Chambers2016}) and the near-infrared (e.g., the Vista Variables in the Via Lactea survey conducted with the VISTA telescope; \citealt{Minniti10}). Repeated observations of the same parts of the sky enable time-domain studies, which are key for the identification and analysis of varying and transient astrophysical events. Along these lines, repeated surveys of large areas of sky are critical for statistical studies of varying and transient sources. Indeed, studies based on data from the aforementioned large surveys have shown that the properties that describe the level of variability of sources (such as variability timescales and amplitudes) correlate with other physical properties of the sources (e.g. \citealt{MacLeod2010}). Constraining such correlations will therefore enable us to use variability to infer other physical properties for large samples of astronomical sources. For example, measuring the variability of large numbers of AGN may, in the future, provide a further handle on the mass distribution of supermassive black holes (\citealt{Caplar2017},\citealt{Sanchez2018}). To achieve the full benefits of high-cadence, wide field surveys, it is critical that we are able to process large samples of photometric data efficiently and to a high level of precision.
In order to address the challenges presented by the data volume and rate delivered by wide-field, high cadence surveys, many of these projects have invested significant resources into the development of efficient data processing pipelines \linebreak (e.g., SDSS;\citealt{Lupton2001},Pan-STARRS; \citealt{Magnier2016}). The forthcoming Legacy Survey of Space and Time (LSST) to be conducted with the Vera C. Rubin Observatory (\citealt{Ivezic19}), represents a step-change in both data quantity and delivery rate. As a consequence, major efforts are currently being made by the LSST team to ensure that the data processing pipelines -- the LSST Science Pipelines (hereafter referred to as simply the ``LSST stack''; \citealt{Juric2017}) -- are capable of handling the data flow from the telescope.\footnote{The LSST stack software is available at https://github.com/lsst} The pipeline will deliver both of the main LSST data products i.e. those from the nightly processing and the annual releases.
Rather than being a single-purpose pipeline, however, the LSST stack has been designed to be adaptable for surveys conducted by facilities other than the Vera Rubin Observatory. Indeed, the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP;\citealt{Aihara2018a}), which is in the process of conducting a deep, multi-band imaging survey of selected fields, is also using a version of the LSST stack (\texttt{hscPipe}; \citealt{Bosch2018}) to process the data from that survey.
With this in mind, we have adapted the the LSST stack to process data from the Gravitational wave Optical Transient Observatory (GOTO; Steeghs et al. in prep.) -- a wide-field (currently $\sim$40 sq. deg), high-cadence survey telescope based on La Palma whose primary scientific objective is the identification of optical counterparts of gravitational wave events. The GOTO collaboration has developed their own in-house processing pipeline that has been optimised for the rapid follow-up of gravitational wave events. However, while GOTO's survey depth is shallower ($\sim20$ mag limit for a 3 minute exposure in dark time) than the LSST, both have a similar single-pointing field-of-views and cadences, meaning that the LSST stack is a viable alternative pipeline for non-primary science data products. In order to process GOTO data using the LSST stack, we have developed our own ``obs package'', \texttt{obs\_goto}, which is described in more detail in \cite{Mullaney20}.
Perhaps the most important data to come out of repeated, wide area surveys such as GOTO and the LSST is that of lightcurves; i.e., time-series data that describes how the flux of an astronomical object changes over time. In order to obtain lightcurves, flux measurements of the same object extracted from multiple observations must be associated with one another. One way of achieving this is via positional matching, in which a ``blind'' source detection (e.g., SExtractor; \citealt{BA1996}) algorithm is run on each incoming science exposure and common sources (i.e., those associated with the same physical object) are identified as those that lie within a given matching radius of each other in each observation. This method is, however, subject to a number of issues such non-detection in a survey because of low S/N in a given exposure (in which case, it is possible that a detected neighbouring object may be incorrectly matched, especially in a crowded field) and deblending failures. In an attempt to address these issues, the technique of ``forced photometry'' was developed. With this method, photometric measurements (e.g. flux) of sources are performed with the positions (and, if necessary, other parameters such as shape) fixed at those specified in a reference catalogue. Using this method we can, to a degree, mitigate the issues of non-detections or blended sources in a survey as the photometry will be measured for each position in the reference catalogue.
Motivated by the key role that forced photometry will play in the coming years with current and future multi-wavelength wide field surveys, we investigate the performance of the LSST stack's forced photometry task on wide-field survey data obtained by GOTO. Using this method we are also able to assess GOTO's photometric performance and compare against results obtained via ``blind source photometry'' i.e. photometry measurements of those sources identified in GOTO images via standard ``blind'' source detection, such as that performed by SExtractor (\citealt{BA1996}). Throughout this study, we use v18.01 \linebreak (released July 2019) of the LSST stack, which was the most up-to-date version when we began processing our data. It is important to note, however, that v18.01 of the LSST stack utilises the now near-obsolete “Generation 2” Butler to organise and retrieve data, which at the time of writing has largely been replaced by the ``Generation 3'' butler.
In this paper, we report on the forced photometry results we obtain by processing GOTO data using the LSST stack, with a particular emphasis on the quality of the photometry measurements. In the following section we provide a brief description of the GOTO survey, while in section 3 we give an overview of the forced photometry task and also how we filter bad data points from our light curves. In section 4, we present the results of various quality-assurance tests of the forced photometry measurements. Finally in section 5, we summarise our work and present our conclusions.
\section{The Gravitational-wave Optical Transient Observer}
The GOTO prototype is located on the summit of El Roque de los Muchachos (La Palma, Spain). It consists of an array of 40~cm-diameter astrographs (f/2.5) attached to the same mount, with each astrograph equipped with a 50M pixel detector with a field-of-view of roughly 5 sq. degrees (and a corresponding pixel scale of 1.24 arcsec). At the time of writing, GOTO consists of eight astrographs (hereafter, unit telescopes, or UTs) which is the full complement for a GOTO mount, although the data described in this work was obtained prior to the second set of four UTs being added i.e. during the GOTO prototype phase. A planned second dome located alongside the first will host an identical mount resulting in a total of 16 UTs and a total field-of-view of 80 sq. degrees, enabling repeat observations of the whole observable sky every few nights. A southern node is planned to be located in Australia which will provide full sky coverage for the GOTO survey. GOTO's default observing mode is the so-called ``survey mode'', in which the sky is repeatedly observed in a systematic way. This can be interrupted at any time to undertake prioritised observations to follow-up a transient event, such as a gravitational wave event (\citealt{Dyer2020}).
The four UTs used to obtain the data analysed in this study were aligned such that they deliver a contiguous field-of-view of roughly 20 sq. degrees per mount pointing. Each UT is equipped with a filter wheel consisting of standard Baader R, G, and B filters, plus a broad L-band filter which covers the optical passband between $\sim$400-700nm. The L-band is used as the primary filter for the survey as it maximises the amount of light reaching the detectors in a given exposure.
Every night the GOTO Telescope Control System (G-TeCS; \citealt{Dyer2018}, \citealt{Dyer2020}) decides whether to open the dome given various criteria, including the local weather conditions. The pilot controls the hardware and will stop the operations if conditions are not appropriate and it will close the dome. Should the conditions improve/deteriorate during the night, the full robotic system will automatically resume/pause operations. The observations are also controlled by this system, with a scheduler deciding, in real-time, the optimal observations to conduct to achieve the primary science objective (i.e., the detection and identification of the optical counterparts of gravitational wave events and other transient sources).
To identify the optical counterparts of gravitational wave events, GOTO repeatedly surveys the whole observable sky to ensure that up-to-date reference images exist to compare against incoming follow-up observations. This repeated survey provides the opportunity for science projects beyond identifying the optical counterparts of gravitational wave events. These include, for example, time-domain astrophysics and transient detections via forced photometry or image differencing on repeated observations of the sky.
Included in the LSST stack is a suite of software that is capable of conducting both image differencing and forced photometry. This includes the production of a set of reference images and catalogues (see \citealt{Mullaney20}) and, as we describe next, the tasks required to perform forced photometry based on the positions of sources in the aforementioned reference catalogue. The LSST stack also includes software capable of conducting image differencing, although this is beyond the scope of this study.
\section{Data and Pipeline products}
Typically, GOTO begins each night by obtaining a number of calibration frames; specifically bias, dark, and sky-flat frames. If the observing criteria are met (e.g., suitable weather conditions) then, after conducting a set of focusing exposures, it begins science observations. The data presented in this paper are the result of the nightly processing of raw images observed between the 24 February and the 25th October, 2019. Coadded images, from which reference catalogues are constructed, were produced by combining frames from dates spanning the 24 February to 12 March, 2019 (see \citealt{Mullaney20}). This selection resulted in reference images and catalogues spanning ${\rm 2~h \lesssim RA \lesssim 20~h}$ and ${\rm -20~deg \lesssim Dec. \lesssim 90~deg}$, which represents roughly 50\% of the sky observable from GOTO's location on La Palma and avoids the densest parts of the Galactic plane (see Fig. \ref{fig:coverage}). At each pointing, GOTO takes three back-to-back 60~s exposures which together are termed a ``visit'' . On coadding these three exposure we achieve an L-band $5\sigma$ magnitude limit of $\sim$20~mag (see section 4). While we describe the processing of GOTO images with the LSST software in detail in \cite{Mullaney20}, we feel it is important to highlight some of the key processing steps using the LSST stack here in order to provide some context.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/coverage}
\caption{The region of the sky covered by the GOTO data that we processed using the LSST stack. These data cover the region spanning roughly -30 to 90 deg in declination and 15 to 315 deg in right ascension. This represents roughly 50\% of the sky observable at the location of the GOTO prototype on La Palma, Spain.}
\label{fig:coverage}
\end{figure}
First, the raw data are ingested into a database using the header information and master calibration frames that are produced by combining the individual bias, dark, and flat frames. This task can be done nightly or calibration frames from different nights can be combined. These master calibration frames are then used to correct science exposures for so-called instrument signatures, after which the individual science frames undergo background subtraction, PSF-characterisation, and astrometric and photometric calibration using a number of selected sources as astrometric and photometric standards. While the LSST stack can be fed bad pixel masks for flagging purposes, we did not provide these since they weren't available during GOTO's prototype phase when our data were taken. However, on visual inspection of the CCD images, it is clear that bad pixels represent far fewer than one in ten thousand pixels, and so we are confident that the impact of not including bad pixel masks has a minimal impact on our results. In this paper, we present the results from forced photometry using, as references, catalogues generated by running the LSST stack's detection algorithm on the coadded frames, and adopting a $5\sigma$ detection threshold (\citealt{Mullaney20}). In the following subsection we present the method we used to perform forced photometry on GOTO images using the LSST stack.
\subsection{Forced Photometry}
\label{frcphot}
Forced photometry is a technique that was developed to deal with issues arising from cross-matching between sky surveys conducted at different wavelengths and/or different telescopes. Depending on the shapes of their (observed-frame) spectral energy distributions, different sources may or may not be formally detected in different surveys. In such cases, a simple positional match may wrongly associate a source detected in one survey with a different, nearby, source in the other when, in reality, it is not formally detected in the latter. Further, sources that are close -- but still resolved as separate -- in one survey may be blended together in another survey, meaning that a simple positional match will associate all the resolved sources in one survey with the single unresolved source in the other. Forced photometry attempts to solve both these issues by performing photometry on one survey based on the positions of detected sources in another (e.g. \citealt{Bovy2012}, \citealt{Lang2014}, \citealt{Nyland2017}). In this respect, it is similar to the ``list-driven'' photometry technique described in \cite{Aigrain15}, although in our case we utilise our own catalogue derived from the coadded frames (see \citealt{Mullaney20}), rather than an external catalogue. Meaningful upper limits can be obtained via forced photometry in cases where a source would be formally undetected in a given survey. Of course, unless mitigating steps are taken, forced photometry can still suffer from flux contamination due to varying PSFs between different science frames. A further major drawback of forced photometry is that a transient source that does not exist in the reference catalogue would not be measured (unless, of course, it is associated with an existing, detected source, such as a supernova within a detected galaxy). To solve this problem, other techniques could be used, such as image differencing, which is beyond the scope of this study. Similarly, high proper motion stars may also be missed by forced photometry if the reference catalogue does not include proper motion information. As we are using our own catalogue derived from coadded GOTO frames as a reference, we do not have this information to hand. Since high proper motion stars represent an extremely small number of all astronomical sources, especially in the region outside the Galactic plane covered by our reference catalogue, we do not attempt to account for such sources.
To perform forced photometry on our incoming science frames, we use the LSST's {\tt forcedPhotCcd.py} task. This task finds the sources within the reference catalogue that overlap with the incoming science frames, and performs various (user-specified) photometric measurements at the positions of those sources. This approach means that every measurement in the incoming science frame is associated with an object ID within the reference catalogue. This association makes extracting light curves for a given object straightforward, as the user simply needs to specify the object ID of the source they are interested in.
By default, {\tt forcedPhotCcd.py} performs forced photometry on every incoming science frame. However, GOTO takes multiple (usually three) back-to-back exposures for each pointing, which are grouped together according to their visit and CCD numbers. Each visit is identified via a unique identification number which associates it with a given pointing (see \citealt{Dyer2018} for further details). Since each exposure in a given visit is taken back-to-back, it is unlikely that there will be much change between exposures so, rather than perform forced photometry on every incoming exposure, we instead decided to coadd (using the LSST stack's {\tt snapCombine.py} task) the three back-to-back exposures to increase the depth of the forced photometry.\footnote{We choose to coadd, rather than take a median of, the three input exposures because taking a median can affect the PSF in non-trivial ways.} Prior to coaddition, however, each individual exposure requires instrument signature removal (ISR) and warping to a common WCS. We have therefore written our own wrapper for {\tt forcedPhotCcd.py} (named {\tt singleVisitDriver.py}) that processes and coadds individual frames prior to also performing forced photometry.
We have made some further modifications relating to how the uncertainties associated with the photometric zero-point are calculated for each coadded exposure. In v18.01 of the LSST stack, the zero-point uncertainty is calculated as $\sqrt{\Sigma(1/\sigma^2)}$, where $\sigma$ is the so-called ``instrumental error'' associated with each measured source arising from photon noise. While this would hold true if the {\it only} source of error was counting statistics, in GOTO's case there are other sources of errors (e.g., varying conditions across the CCD) that would not get captured by this method. To obtain a more appropriate estimate of the uncertainty in the zero-point of a given frame($\sigma\textsubscript{zp}$), we instead use the standard deviation of the absolute difference between the instrumental magnitudes and the calibrated magnitudes (i.e., $\Delta m$) of the $N$ stars used to obtain the zero-point, i.e.,
\begin{equation}
\sigma\textsubscript{zp} = \sqrt \frac{{\Sigma \left(\Delta m - \overline{\Delta m}\right)^2}}{N}.
\end{equation}
Within the LSST stack, the error on the zero-point is then added in quadrature to the instrumental error for each measured source. A histogram of zero-point values for all our frames is presented in fig. \ref{fig:zeropoint}.
We have measured the total time taken to undertake the entire forced photometry task (i.e. from the calibration of the three back-to-back exposure to their coaddition to the forced measurements performed on the coadded frame) and report that it takes an average of $\sim$20~s per coadded frame.\footnote{ We used a Dual Intel Xeon E5-2697v3 2.60~GHz CPU with 28 cores/56 threads with access to 256 GB of RAM to process all our data. The reported time is the average wall-clock time to process a single coadded frame on a single core.} We note finally that, to speed up processing, we only perform forced photometry on the nightly frames (i.e., we do not also perform blind detection and measurement on the nightly frame). At the position of every reference source within the boundaries of a given coadded exposure we measure aperture photometry (using aperture radii: 5.58, 7.44, 11.16, 14.88, 29.76, and 59.52 arcsec which correspond to 4.5, 6, 9, 12, 24, and 48 pixels, respectively) and PSF photometry (using a PSF modelled using Principal Component Analysis; see \cite{Mullaney20} for more details). Our LSST stack-processed data is photometrically-calibrated using Pan-STARRS PS1 (\citealt{Magnier2016}) g-band PSF photometry, adopting appropriate colour terms to convert to the L-band (see \citealt{Mullaney20}). We use Pan-STARRS PSF photometry, as that is what is recommended for point sources, which the vast majority of calibration sources are.
\subsection{Lightcurves}
\label{lightcurves}
The catalogues generated by {\tt forcedPhotDriver.py} contain the position of each reference source, aperture photometry measurements using various pre-defined aperture radii (see previous section) and their respective errors, and PSF photometry measurements and their errors. Forced photometry metadata are also generated containing information on the epoch, UT, filter, observed target or tile, visit number, zero-point and seeing of exposure. In this paper, we predominantly rely on aperture photometry for the results presented in section 4. We do, however, compare PSF photometry lightcurves against those extracted from aperture photometry as a way to estimate the quality of the PSF photometry.
Light curves are the main means by which the data from high cadence photometric surveys are analysed. As such, it is vitally important that the data used to construct lightcurves are reliable. There are various reasons, however, why this may not be the case. For example, data collected during nights of poor photometric quality, or instrumental or pipeline failures (which may or may not get flagged). It is therefore important to pre-process the light curves to ``clean'' the data of spurious photometric measurements which are not accounted-for by the reported uncertainties (e.g., in some cases, poor quality photometry is captured by the large uncertainties associated with the measured zero-point, but this is not always the case). In an attempt to remove poor quality data, we exclude any that arise from exposures with photometric zero-points that deviate by more than three standard deviations from the average. On further investigation these deviant zero-point values arose from frames that were affected by poor observing conditions such as thin cloud. The mean value of the photometric zero-point as measured by the LSST stack corrected for the exposure time for the data presented in this paper is 22.52 with standard deviation of 0.51 (see fig. \ref{fig:zeropoint}).
\begin{figure}
\includegraphics[width=\columnwidth]{figures/zeropoint}
\caption{The distribution of photometric zero-points for the processed frames. The average and standard deviation of this distribution of 22.52 and 0.51. We exclude from further analysis any frames whose zero-points deviate from the average by more than three standard deviations. In further investigation, it was found that these were typically affected by poor observing conditions, including thin cloud cover.}
\label{fig:zeropoint}
\end{figure}
Additionally, we exclude those sources that are flagged as interpolated, which, for the most part, arise due to the presence of saturated pixels in their footprint. In these cases the flux measurement fails because the masked pixels are not included when summing the flux within an aperture. If they were not excluded, such sources would have an underestimated measured flux which would not be reflected in the photometric uncertainty.
\section{Results}
In this section, we first compare the LSST software stack results for GOTO forced photometry against the blind source photometry from the Pan-STARRS DR1 catalogue as well as those measured by {\sc gotophoto} -- the in-house pipeline of the GOTO collaboration.\footnote{At the time of writing, {\sc gotophoto} does not conduct forced photometry.} We then assess whether the forced photometry results are self-consistent by testing the precision of the photometry across multiple nights. Using the photometric repeatability we obtain an estimate of the survey depth -- one that is independent of that estimated for the deeper reference images and catalogue described in \cite{Mullaney20}. In this section we also investigate the quality of our photometric uncertainties by assessing whether the reported uncertainties account for the scatter in the difference in measured photometry across multiple nights. Finally, we compare GOTO light curves generated by the LSST stack against those extracted from the Zwicky Transient Facility (ZTF; \citealt{Bellm2019}) and GAIA DR2 (\citealt{Clementini19}) databases for a number of known variable stars.
\subsection{Photometry}
\label{photoqual}
\begin{figure*}
\includegraphics[width=\columnwidth]{figures/frcAp_vs_PS}
\includegraphics[width=\columnwidth]{figures/frcPSF_vs_PS}
\centering
\caption{The difference between GOTO photometry measured using the LSST stack and Pan-STARRS PSF photometry, plotted as a function of L-band magnitude. The left-hand plot shows the magnitude difference arising from GOTO aperture photometry, whereas the right-hand plot shows the difference arising from GOTO PSF photometry. Both plots show data arising from all four UTs for a single pointing.}
\label{fig:LvsPS}
\end{figure*}
The first approach we take to estimate the quality of the photometry as measured by the LSST stack is to compare against the magnitudes reported in the Pan-STARRS DR1 catalog (\citealt{Magnier2016}). The Pan-STARRS photometry is calibrated using the \texttt{ubercal} method described in \cite{Schlafly2012}. Our choice of Pan-STARRS is motivated by the fact that it has a very similar sky coverage to the GOTO survey, but is significantly deeper than the GOTO survey, so all non-transient sources detected by GOTO {\it should} have a Pan-STARRS counterpart (the main exception being transient sources in GOTO). This means that we can obtain comparison statistics down to the detection limit of GOTO (i.e., we are not limited by the depth of Pan-STARRS). In Figure \ref{fig:LvsPS} we present plots showing the magnitude difference between the GOTO magnitudes measured by the LSST stack and the colour corrected Pan-STARRS g-band magnitude versus the GOTO magnitude. We present results from both aperture (11.16~arcsec; left) and PSF (right) photometry. This plot includes sources from all four CCDs for a single pointing, although we obtain similar results for all pointings (caveat those pointings filtered-out via the method outlined in \S\ref{lightcurves}). This comparison with Pan-STARRS suggests that the PSF photometry is more precise for this particular epochal pointing, although as we shall see from the repeatability test, aperture photometry results are in general more accurate, and especially for sources brighter than $\sim17^{\rm th}$ magnitude. Both PSF and aperture photometry suggest that, for sources fainter than $18^{\rm th}$ magnitude down to the detection limit (i.e., $\sim$19.5), GOTO photometry as measured by the LSST stack is within 0.2 magnitude RMS of the Pan-STARRS photometry. Between 16th and 18th magnitude it is within $\sim$0.03-0.06 mag and for sources brighter the 16th magnitudes, the RMS is $\sim$0.01-0.02 mag.
\begin{figure*}
\includegraphics[width=\columnwidth]{figures/frcAp_vs_gotoflow}
\includegraphics[width=\columnwidth]{figures/frcPSF_vs_gotoflow}
\centering
\caption{As for fig. \protect\ref{fig:LvsPS}, but instead showing the difference between LSST stack-measured photometry, and that measured by {\sc gotophoto}.}
\label{fig:Lvsgotoflow}
\end{figure*}
We also compare outputs from the LSST stack against those obtained with {\sc gotophoto}, the in-house photometric pipeline developed by the GOTO collaboration. {\sc gotophoto} uses SExtractor's (\citealt{BA1996}) {\tt MAG\_AUTO} photometry measurements for photometric calibration, which it compares against Pan-STARRS DR1 g-band PSF photometry to obtain photometric zero-points. Unlike the LSST stack-processed data, the version of {\sc gotophoto} used for our comparison does not apply colour terms to convert between Pan-STARRS g-band photometry measurements to GOTO's L-band (although there are some colour cuts on the stars chosen to calibrate and image to remove strong outliers, i.e., -0.5 < g-r < 1.0).\footnote{Colour terms will be implemented in future versions of {\sc gotophoto}} We note that here we are comparing forced photometry measurements (from the LSST stack) against measurements of sources obtained via blind detection (from {\sc gotophoto}). This caveat shouldn't be a concern for isolated sources, which form the vast majority, although it could mean that some sources are deblended in the blind catalogue, but not in the forced photometry catalogue (or vice versa). {\sc gotophoto}'s measurements are based on SExtractor's (\citealt{BA1996}) {\tt MAG\_AUTO} aperture photometry performed on the same (but median-combined, rather than coadded) back-to-back exposures as those we performed forced photometry on. In Figure \ref{fig:Lvsgotoflow} we present plots showing the magnitude difference between the GOTO magnitudes measured by the LSST stack and those measured by {\sc gotophoto} versus GOTO magnitude. From this comparison we note that the vast majority are within an RMS of 0.1~mag, even in the case of the faintest sources. There is, however, a systematic offset between the results of the two pipeline which is likely due to the application of colour terms when we process the data with the LSST stack (which are not applied to {\sc gotophoto} magnitudes). The PSF photometry appears once again to result in a smaller scatter than aperture photometry for this particular pointing, although we note that other pointings produce similar overall results. The accuracy of PSF photometry vs. aperture photometry is investigated using the photometric repeatability, which we consider next.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/repeat}
\centering
\caption{Photometric repeatability for 11.16~arcsec aperture photometry as measured with the LSST stack, plotted as a function of L-Band magnitude. Each point represents a single reference source within a 4 UT pointing. We use the RMS of the photometry of these sources measured across multiple nights as our measure of repeatability; see Section \ref{photoqual} for details. A photometric precision of 0.02 mag (shown as the red line) is achieved for bright (i.e., $m_L\lesssim15$) sources. }
\label{fig:repeatability}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{figures/psfVsApRMS}
\centering
\caption{As for fig. \protect\ref{fig:repeatability}, but now also showing the repeatability of PSF photometry for comparison. While PSF photometry is less precise for the brighter sources, it may be a better choice for fainter sources. We also use the inter-night RMS to estimate the survey depth by using 0.2 mags RMS as an estimate of the 5$\sigma$ detection threshold. This is shown as the horizontal red line at $RMS=0.2$~mag, which corresponds to L-band survey depths of 18.6 and 19.4 for 11.16 arcsec aperture photometry and PSF photometry, respectively.}
\label{fig:psfvsAp}
\end{figure}
With photometric repeatability tests, we are investigating the level of consistency between multiple photometric measurements of the same source across multiple nights. We estimate this quantity for both types of photometry measurements obtained using our implementation of the LSST stack, i.e., 11.16 arcsec aperture and PSF photometry. We investigate how the photometric repeatability changes as a function of magnitude. To do this, we use data obtained from observations of a region of sky that has been visited the most number of times by GOTO between 2019-03-01 and 2019-07-31 (inclusive) and calculate the RMS of the magnitude of all sources in one pointing (from all four UTs). To clean the photometric data of spurious measurements we use a similar method using pixel flags as that outlined in section 3.2. Further, we remove measurements from any frame whose zero-point deviates by more than three standard deviations from 22.52 (i.e., the mean zero-point reported in \S\ref{lightcurves}). Both of these cleaning steps are straightforward to do using the metadata provided by the LSST stack. We note that we do not filter for known variable sources. Such sources will increase the measured inter-night RMS, but they represent such a small proportion of sources that we do not expect them to have any measurable effect on our repeatability measurements.
In figure~\ref{fig:repeatability} we plot the inter-night RMS of aperture photometry measurements from the aforementioned observations. From this plot we see that for sources brighter than $15^{\rm th}$~mag the internal photometric precision, as measured by photometric repeatability, is typically below $\sim$0.02 mag (red line on the plot). The RMS increases with decreasing brightness due to the increase in size of the photometric uncertainties.
We compare the repeatability of PSF and aperture photometry in fig. \ref{fig:psfvsAp}. This plot shows that PSF photometry is less consistent between nights than aperture photometry for sources brighter than around $m_{\rm L} = 17$. This limitation is expected due to the various difficulties associated with performing PSF photometry on bright sources arising from, e.g., bright spikes or saturated pixels. PSF photometry, however, is found to perform better than aperture photometry for fainter sources. In the current GOTO prototype system, the PSF can vary over the field-of-view, especially at the edges. However, as shown in Fig. 5 of \cite{Mullaney20}, the LSST stack's PSF modelling software is able to account for this variation and, in general, does a good job of reproducing the PSF across the frame.
In the aforementioned analysis we exclusively used 11.16 arcsec aperture, i.e. $\sim$2.5 times the typical FWHM of the PSF size ($\sim$4.5~arcsec), for aperture photometry. However, as mentioned in \S\ref{frcphot}, our implementation of the LSST stack returns measurements obtained with multiple different sized apertures. In general, we find that smaller apertures reproduce the inter-night RMS of PSF photometry more faithfully than larger apertures (i.e., smaller apertures result in a larger systematic RMS at brighter magnitudes, but smaller RMS values at fainter magnitudes). Considering this, it may be beneficial in the future to attempt to adjust the aperture to match the size of the PSF, at least when measuring faint sources.
\subsection{Survey depth from repeated photometry}
In \cite{Mullaney20} we characterised the depth and the detection completeness of the coadded images based on the average magnitude of a $5\sigma$ detected sources. This was further verified using injected sources. However, another way to obtain an estimate of the magnitude limit for the GOTO survey is from photometric repeatability. The 5$\sigma$ detection corresponds to a S/N of 5 and so to a flux RMS of $\sim$ 0.2 mags (\citealt{Masci2019}; note that in this regime we are dominated by random, as opposed to systematic, errors). Using aperture photometry measurements from the same set of observations as used in \S\ref{photoqual} we calculate the median magnitude of sources with a S\/N of between 4.5 and 5.5. As shown in Figure \ref{fig:psfvsAp}, this corresponds to a magnitude limit for the L-band of 18.6~mag. This value is brighter than the L-band magnitude limit of 19.6~mag for the coadded references exposures presented in \cite{Mullaney20}. This discrepancy arises from the fact that the analysis described in this paper is based on forced photometry performed on single visit frames. While each of these visit frames is produced by mean-combining three back-to-back exposures, each coadd (or part thereof) may have been produced from the combination of more than three exposures.
One important issue relating to the estimation of the survey depth relates to the choice of the type of photometry used (i.e., aperture vs. PSF) and, in the case of aperture photometry, the size for aperture. This choice is especially relevant for the faint sources that we use to define the survey depth. In our case the 11.16~arcsec aperture will include more background flux than a smaller aperture, possibly leading to an overestimation of the brightness of the faint sources. We can see from Figure \ref{fig:repeatability} that if we use the PSF photometry for the fainter sources then we would obtain a limiting magnitude of 19.4~mag (note that PSF photometry was not performed by \citealt{Mullaney20}, so a comparison between forced and reference PSF photometry cannot be made).
\subsection{Observational photometric uncertainties}
\begin{figure}
\includegraphics[width=\columnwidth]{figures/errorplots}
\caption{Plots used to assess the quality of the uncertainties of our photometry measurements. The top panel shows the robust standard deviation of the quantity of Eqn. \ref{zij} as a function of the median magnitude for the GOTO light curves from forced photometry. The middle shows the intrinsic scatter of each light curve measured as the robust standard deviation of the magnitude as a function of the magnitude and the bottom panel shows the median error of each light curve as a function of the magnitude. The method evaluates the photometric uncertainties following the method outlined in \protect\cite{Suberlak2017}.}
\label{fig:errorplots}
\end{figure}
As well as assessing the quality of the absolute photometric measurements, it is important that we also assess their uncertainties. This quantity is particularly relevant for the analysis of variable sources, since we need to know whether differences in the measured photometry over multiple nights are physical in origin (i.e., genuine), or simply due to the statistical variances in our measurements. As such, we must carefully evaluate whether we are over or under-estimating our uncertainties.\footnote{An example of the effects of poor error estimation is presented in \citealt{Suberlak2017}, in which they find that the quasar variability levels observed in Catalina Real-time Transient Survey (CRTS) data (\citealt{Graham2014}) actually arises from underestimated errors.}
\begin{figure*}
\includegraphics[width=\linewidth]{figures/errorcorrepochs}
\caption{Standard deviation of the photometric scatter normalised for the reported photometric uncertainties (left) and intrinsic standard deviation (right) as function of magnitude and coloured by the number of epochs in the light curve. There is no obvious correlation for either of them with the number of epochs.}
\label{fig:errorepochs}
\end{figure*}
As described in \S\ref{frcphot}, the uncertainties reported by the LSST stack on each photometric measurement are obtained by combining, in quadrature, instrumental uncertainties (i.e., those principally arising from photon noise) with calibration uncertainties (i.e., those arising from uncertainties in the zero-point). To assess the quality of these uncertainties we use the method outlined in \cite{Suberlak2017} that they use to assess the quality of the uncertainties reported by the CRTS. We repeatedly measure the photometry of stars extracted from the \cite{Ivezic2007} catalogue of standard stars.\footnote{To create this catalogue, \cite{Ivezic2007} used repeat measurements of $\sim$1 million $m_r=14-22$ stars in SDSS Stripe82 to verify that they are non-variable.} For each standard star, $j$, covered by our repeat GOTO observations, $i$, we calculate $z_{ij}$:
\begin{equation}
z_{ij} = \frac{m_{ij}-\overline{m_{ij}}}{\epsilon_{ij}}
\label{zij}
\end{equation}
\noindent where $m_{ij}$ and $\epsilon_{ij}$ are the measured L-band photometries and associated uncertainties of source $j$ from observation $i$, and $\overline{m_{ij}}$ is the mean of all $m_{ij}$, averaged over $i$, weighted according to inverse uncertainty. We then take the standard deviation of these $z_{ij}$ values for each star, using the definition of standard deviation used in \cite{Suberlak2017}, which is less affected by outliers:
\begin{equation}
\label{sdev}
\sigma_j(z) = 0.741\times IQR(z_{ij})
\end{equation}
\noindent where $IQR(z_{ij})$ is the 25\%-75\% interquartile range of the $z_{ij}$ values over all observations, $i$, of each standard star, $j$. As such, there are $j$ $\sigma_j(z)$ values, i.e., one per star.
There are $\sim$13000 standard SDSS stars within the current GOTO reference catalogue used as the basis for our forced photometry. The r-band magnitude range of these stars is 14-20. In the top panel of Figure \ref{fig:errorplots} we plot $\sigma_j$ for each of these stars. For non-variable stars, it is expected that the $\sigma_j$ values would follow a distribution centred at unity and display no dependence on magnitude. However, as for CRTS, we find that this is not the case for GOTO photometry as measured by the LSST stack. Instead, we find that the photometric uncertainties are overestimated by a factor of $\approx$2 (i.e., $\sigma_j\approx0.5$) in the case of sources brighter than $m\sim15$, and underestimated by a factor of 1.2--1.3 (i.e., $\sigma_j\approx0.8$) for sources fainter than $m\sim17$ and brighter than $m\sim18.5$.
\begin{figure*}
\includegraphics[width=0.45\linewidth]{figures/stst1.png}
\includegraphics[width=0.45\linewidth]{figures/stst2.png}
\centering
\caption{Lightcurves of standard SDSS stars from aperture (upper panels) and PSF (lower panels) photometry. Aperture photometry gives reduced-$\chi^2$ closer to unity, which is what is expected for non-variable stars. Aperture photometry performs better at brighter (i.e., < 15 mag) magnitudes.}
\label{fig:LC1}
\end{figure*}
In the middle panel of Figure \ref{fig:errorplots} the standard deviation of the magnitude difference (again calculated using Eq. \ref{sdev}, but with $\Delta m_{ij}=m_{ij}-\overline{m_{ij}}$ in place of $z_{ij}$), which we denote $\sigma_j(\Delta m)$. Finally, in the bottom panel of Figure \ref{fig:errorplots} we plot the median error of each of our sources, again averaged over all observations, $i$. From these lower two panels, we find that the standard SDSS stars brighter than $m\sim15$ have a standard deviation in $\Delta_m$ that is less than $\approx$0.015 whereas the minimum uncertainty for these magnitudes is $\approx$0.032. Again, this implies that the uncertainty estimate is too large by a factor of $\approx$2 in this bright regime.
Since the aforementioned results are based on standard deviation measurements, it is important to ensure that they are not affected by artificial factors such as the number of observations (e.g., the standard deviation will only start to approximate to the size of the uncertainty after a large number of measurements). To test for this, we explore whether the trends seen in the top plot of Figure \ref{fig:errorplots} changes as a function of the number of epochs (see Figure \ref{fig:errorepochs}). However, while there is perhaps some evidence of larger uncertainties at fewer epochs, this effect is very weak and certainly not large enough to explain the trend seen in Figure \ref{fig:errorplots}.
Our results therefore support the application of correction factors to the error bars, such as those presented in \cite{Suberlak2017} for the CRTS data, prior to using the photometric uncertainties when studying source variability. In our case, we calculate the correction factor by fitting a 4\textsuperscript{th} degree polynomial to the median values of $\sigma_L$ shown in the top plot of Figure \ref{fig:errorplots}. The resulting polynomial has the following terms:
\begin{equation}
0.0074 x^4 - 0.5168 x^3 + 13.42 x^2 - 153.4 x + 651.5
\end{equation}
We then use this polynomial to correct our uncertainties. These corrected uncertainties are included in the following section where we present some examples of lightcurves measured using forced photometry on GOTO data by the LSST stack.
\subsection{Light curve analysis}
As a final approach to characterise the forced photometry results, we examine some light curves generated using the nightly catalogues generated from our forced photometry measurements. First, we compare the light curves generated using aperture photometry against those using PSF photometry. For standard SDSS stars (\citealt{Ivezic2007}) brighter than $\sim$17.5 we find that the aperture photometry light curves are better, in terms of reduced-$\chi^2$ after single flux fitting of the lightcurves, than the light curves obtained from PSF photometry, especially at the very bright end (< 15 mag; see Fig.\ref{fig:LC1}). We find, however, that for stars fainter than $\sim$17.5 both types of photometry give similar values of reduced-$\chi^2$, with the PSF photometry perhaps giving slightly better measurements, as implied by Fig. \ref{fig:psfvsAp}.
\begin{figure*}
\includegraphics[width=0.45\linewidth]{figures/cep6}
\includegraphics[width=0.45\linewidth]{figures/cep7}
\includegraphics[width=0.45\linewidth]{figures/rr4}
\includegraphics[width=0.45\linewidth]{figures/rr7}
\centering
\caption{Examples of GOTO L-band light curves from aperture photometry measured with the LSST stack, presented as phase plots, of periodic variable stars. Also included in these plots are ZTF g-band and GAIA G-band light curves, colour-corrected to L-band. These variable stars have period from 0.5 to 6 days and belong to two different classes (top row: Cepheid variables, bottom row: RR Lyraes, the id numbers refer to GAIA ids). GOTO light curves contain fewer data points since it has (a) been operating for a shorter period of time than the other surveys and (b) the ZTF in particular has a higher cadence than the GOTO prototype. The shapes of the GOTO, ZTF and GAIA phase plots are very similar, demonstrating that GOTO will be a valuable resource for time-domain studies of variable sources once it has its full complement of 16 UTs in both the Northern and Southern hemispheres.}
\label{fig:gotoztf}
\end{figure*}
Finally, to characterise the performance of the forced photometry performed by the LSST stack task on GOTO data, we compare our light curves against those for the same sources obtained by the ZTF and GAIA surveys. ZTF is a wide-field survey that uses the Palomar 48 inch Schmidt telescope and with a dedicated camera of 47 deg$^2$ field of view (\citealt{Graham2019}). Its large FOV allows ZTF to scan the northern sky with a 3-day cadence in g and r bands. It also scans the visible Galactic plane every night. The survey has a median limiting g-band and r-band magnitude of $\sim$20.8 and $\sim$20.4, respectively. The ZTF survey is similar to GOTO's high-cadence survey, although GOTO's cadence with 4 UTs is lower than that of ZTF. However, when GOTO is fully deployed (with 16 UTs in both the northern and southern hemispheres), its cadence will match that of ZTF. By contrast, GAIA is a space based mission whose primary goal was to measuring accurate positions, parallaxes, and proper motions of over 10 billion stars (\citealt{Gaia16}). To achieve this goal, it observed those stars multiple times during its nominal five-year mission and, in doing so, obtained multi-epoch photometry measurements.
We select four different variable stars (two Cepheids and two RR Lyrae stars) from the General Catalogue of Variable Stars (GCVS 5.1; \citealt{Samus2017}) with periods of 0.5 to 6 days (these periods are those reported by that same catalogue). The ZTF light curves are generated using data from their second data release (DR4; \citealt{Masci19}) and only include epochs from the public survey. We present the results in figure \ref{fig:gotoztf}. The GAIA lightcurves are generated from the second GAIA data release (DR2; \citealt{Gaia18}). We have colour corrected both the ZTF g-band and and GAIA G-band photometry to create synthetic ZTF and GAIA ``L''-band photometry. The light curves in this case are given as phase plots in which repeated observations of the same part of the cycle are effectively ``folded''. We calculate the phase of a given observation by subtracting a reference starting time (the same start time is used in the case of both GOTO, ZTF, and GAIA), dividing by the period of the source, then taking the remainder fraction of the period. We note that that after applying the correction for the errors as implied by Fig. \ref{fig:errorplots}, the uncertainties generated by our modified forced photometry task are larger than those on the ZTF and GAIA light curves.
For these comparison plots for the GOTO lightcurves we have used the LSST stack aperture photometry as produced by the forced photometry task whereas ZTF uses PSF photometry. The larger number of data points on the ZTF and GAIA light curves are partly due to both having operating for a longer period than we are considering in this study (i.e., ZTF DR4 covers the period between March, 2018 and June, 2020, and GAIA covers the period between 25 July 2014 and 23 May 2016 ), and especially ZTF's higher cadence over this time. Having said that, it is clear the GOTO lightcurves track the shape of their respective ZTF and GAIA light curves very closely. As such, this comparison make us confident that GOTO will be a valuable resource for measuring variable sources, particularly when it has its full complement of UTs.
\section{Summary}
In this paper we have used the LSST stack to perform forced photometry on images obtained with the GOTO prototype. This has involved the development of the \texttt{obs\_goto} package which works as the interface between the GOTO data and the LSST stack. The \texttt{obs\_goto} package is described in more detail in \cite{Mullaney20}, together with a description of the production of a series of reference images and subsequent reference catalogue for a large fraction (i.e., $\sim50\%$) of the GOTO-observable sky.
In this paper we have presented the nightly processing of the data through the forced photometry task. We used the reference catalogue described \cite{Mullaney20} to obtain the positions of sources for which we measure the flux within a number of apertures. We have also performed PSF photometry for these sources. From the nightly forced photometry catalogue we were then able to generate light curves for the sources. We then assessed the quality of the measured forced photometry by comparing our results to those obtained by {\sc gotophoto} -- GOTO's own processing pipeline, whose photometry measurement is based on SExtractor -- Pan-STARRS, ZTF, and GAIA.
After comparing against colour-corrected Pan-STARRS g-band photometry, we found that our L-band photometric measurements were consistent to within 0.01 mag (rms) for brighter sources ($\sim$14 mag) to 0.2 mag (RMS) for fainter sources ($\sim$ 18 mag). We also performed internal photometric tests by assessing the consistency of repeated measurements of standard stars. This assessment showed that the typical precision for bright (i.e. $<$ 15.5), unsaturated sources is 0.02 mag. This assessment also indicated that the GOTO aperture photometry from the LSST stack is more precise than PSF photometry which has a precision of 0.04 mag for bright sources. The survey depth of a GOTO pointing ($\sim$ 19.4 mag), again measured via the repeatability of aperture photometry, is found to be slightly brighter than that reported ($\sim$19.6 mag) for the reference catalogue described in \cite{Mullaney20}. Finally, by comparing the measured uncertainties to the standard deviation of repeat-observed sources, we found that the photometric errors associated with brighter sources are overestimated by a factor of $\sim$2, whereas they are underestimated by a factor of $\sim$1.3 for sources fainter than $\sim17$.
Our results demonstrate the feasibility of using the LSST stack to process and perform forced photometry measurements on GOTO data. In section 3.1 we have described the steps we took to adapt the LSST stack to process and mean-combine GOTO data from a single pointing and perform forced photometry on the resulting frame. This adaptation is included in our \texttt{obs\_goto} package. This highlights a particular advantage of using the LSST stack -- i.e., that the user can modify or even write tasks as part of the ``obs\_package’’ to process their own data in the way they wish, while exploiting the various modules that make up the LSST stack. Finally, the results from our various quality assurance tests demonstrate that the data obtained via the LSST stack's processing of GOTO frames can be used for the scientific analysis of light curves.
\section*{Acknowledgements}
We thank the referee for their constructive comments, which helped to clarify some key points. We also thank those members of the LSST Community Forum whose help and advice enabled us to conduct this study. The Gravitational-wave Optical Transient Observer (GOTO) project acknowledges the support of the Monash-Warwick Alliance; Warwick University; Monash University; Sheffield University; the University of Leicester; Armagh Observatory \& Planetarium; the National Astronomical Research Institute of Thailand (NARIT); the University of Turku; Portsmouth University; and the Instituto de Astrof\'{i}sica de Canarias (IAC). This paper makes use of software developed for the Large Synoptic Survey Telescope. We thank the LSST Project for making their code available as free software at http://dm.lsst.org. Based on observations obtained with the Samuel Oschin 48-inch Telescope at the Palomar Observatory as part of the Zwicky Transient Facility project. ZTF is supported by the National Science Foundation under Grant No. AST-1440341 and a collaboration including Caltech, IPAC, the Weizmann Institute for Science, the Oskar Klein Center at Stockholm University, the University of Maryland, the University of Washington, Deutsches Elektronen-Synchrotron and Humboldt University, Los Alamos National Laboratories, the TANGO Consortium of Taiwan, the University of Wisconsin at Milwaukee, and Lawrence Berkeley National Laboratories. Operations are conducted by COO, IPAC, and UW. R.P.B., M.R.K. and D.M-S. acknowledge support from the ERC under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 715051; Spiders).
\bibliographystyle{pasa-mnras}
|
\section{Introduction}
In the 1990s, the field of cosmology encountered serious breakthrough with the discovery of accelerated expansion of the Universe. This led to the current evolutionary status that is determined by the three contributions to the total density parameter ($\Omega$): the ordinary matter ($\Omega_m\approx5\%$), the dark matter ($\Omega_{DM}\approx 27\%$) and the dark energy ($\Omega_{DE} \approx 68\%$), where the latter contribution is responsible for the accelerated expansion phenomena \cite{riess1998}. Importantly, the nature of the {\it dark sector} in the energetic budget is also the reason for intense experimental and theoretical studies that exposed shortcomings of the general relativity (GR) \cite{huterer2017,copeland2006}. In this respect, one of the most promising proposals to deal with the mentioned issues is based on the extensions and modifications of the Einstein's framework of GR \cite{joyce2016}. Note that the proposals of this type are quite successful not only in the context of the aforementioned cosmic expansion but also in other issues of the modern cosmology \cite{nojrii2017}. Moreover, in terms of the quantum gravity, the standard GR cannot be renormalized but higher-curvature terms in the gravitational action may tackle this issue \cite{Julve1978,kiefer2006,Hamber2009}.
In the literature the GR has been extended in many ways, mainly by including additional curvature terms, scalar fields or coupling with matter. For a well-written reviews on the modified gravities we refer reader to \cite{capozziello2011,nojiri2011,clifton2012,joyce2016}. Among many interesting approaches, one of the most popular proposals is the so-called $f(R)$ gravity introduced by Buchdachl in 70's \cite{buchdachl1970,sotiriou2010,bohmer2008}, where standard GR action is generalized to the arbitrary function of the Ricci scalar ($R\rightarrow f(R)$). Further curvature invariants were included in the $f(\mathcal{G})$ gravity, where arbitrary function of the Gauss-Bonnet invariant is added to the Lagrangian \cite{nojiri2005}. In a wide class of the higher-curvature models, the $f(R,\mathcal{G})$, $f(R,R_{\mu\nu}R^{\mu\nu})$ or theories with the higher-order of derivatives have also been introduced and studied extensively \cite{elizalde2010,de_la_Cruz_Dombriz2012,sharif2016,lambiase2015,easson2004,sharif2016}. The authors of \cite{caroll2005} considered combinations of the Ricci scalar, the Ricci squared ($R_{\mu\nu}R^{\mu\nu}$) and the Riemann squared ($R_{\mu\nu\alpha\beta}R^{\mu\nu\alpha\beta}$) terms and shown that such models can provide alternative to the dark energy. Moreover, theories of this type were recently studied in the context of divergences and renormalization group frameworks and play important role in the perturbative approach to the quantum gravity \cite{Avramidy1985,Hamber2009,Ohta2018}. Note that the general relativity can be further extended by the scalar-geometry or scalar-matter couplings. In this case gravitational action includes terms with contributions not only from geometry but also various functions of the scalar field, matter Lagrangian or trace of the energy-momentum tensor. Representative example of the scalar-geometry coupling is the Jordan-Brans-Dicke theory \cite{jordan1955,brans1961}, further generalized to the $f(\phi,R)$ and $f(\phi,\mathcal{G})$ gravity \cite{capozziello2011}. For the matter-geometry couplings, the $f(R,T)$ and $f(R,T,R_{\mu\nu}T^{\mu\nu})$ models ($T$ denotes trace of the energy-momentum tensor $T_{\mu\nu}$) with further extensions are good examples studied widely in the literature \cite{haghani2013,odintsov2013,harko2011,houndjo2012,harko2014b,Sharif2013,Kaczmarek2020}.
On the other hand, the dark matter is even more mysterious since very few information is known except the fact that it modifies galactic rotation, it is not baryonic, has no electric and color charges and do not interact with the light \cite{oikonomou2007,shafi2015}. Apart from the particle candidates for the dark matter, such as supersymmetry neutralino, an interesting candidate has been provided by Chamseddine and Mukhanov. In their so-called mimetic modification of the GR, the leading idea was to isolate gravity's conformal degree of freedom by parametrizing physical metric ($g_{\mu\nu}$) in terms of the scalar field ($\phi$) and the auxiliary metric ($\hat{g}_{\mu\nu}$) \cite{Chamseddine2013}. As a consequence, the conformal degree of freedom turns out to be dynamical, even in the absence of the standard matter fluids, and may mimic behavior of the cold dark matter. It is worth to remark here, that the mimetic theory has been recently extensively studied in various contexts \cite{dutta2018,gorji2020,golovnev2014,maleb2015,ganz2019,Bezerra2019,chamseddine2019,Mansoori2021,Cardenas2021}. Moreover, the mimetic approach has been combined with other modifications of gravity, such as $f(R)$, $f(\mathcal{G})$ or $f(R,T)$ theories with emphasis on the cosmological applications \cite{odintsov2015a,oikonomou2015,zhong2018,leon2015,haghani2018,gorji2018,baffou2017,Bhattacharjee2020}. The unimodular extensions of the mimetic approach were also introduced and studied in recent articles \cite{Odintsov2016,nojiri2016}. For introduction and review of the mimetic theory see \cite{sebastiani2017}.
Besides the riddle of the dark Universe, the modified gravity is also important in terms of the early stages of the Universe. In this context, the theory of inflation is by now most possible scenario that describes early Universe history \cite{Linde2007,lyth1999}. This framework, developed in the 80's, provides explanation of important issues in the Big Bang cosmology such as the horizon, flatness and monopole problems \cite{guth1981}. According to the inflation paradigms, early stages of the Universe exhibited exponentially accelerated expansion, which after deceleration led to the standard cosmological eras \cite{hobson2006}. However, no direct proof that this early epoch occurred is given up to date. One should note that inflation is well described by the {\it slow-roll} models which usually contain single scalar field (inflaton) that drives inflation. Moreover, the form of the scalar field potential can be related to the spectral index as well as to the scalar and tensor perturbations as generated during the inflationary epoch. Thus, scalar potential may be consistent with observations. Recent observational data coming from Planck and BICEP2/Keck array heavily constrained scalar-to-tensor ratio and spectral index of primordial curvature perturbations corresponding to the inflation \cite{ade2016}. One should note, that the bounce paradigm may also be viable characteristic of the early Universe and an interesting alternative \cite{bars2014,brandenberger1993,novello2008,cai2014}. From the modified gravity perspective, inflation can be successfully described and the unification of early-time and late-time acceleration eras is also possible \cite{nojiri2011}. In the literature, inflation in the modified gravity is still heavily studied and debated topic. In particular, for the $f(R)$ gravities a successful model of the Starobinsky's inflation exists, which is consistent with the local and large scale constraints of gravity \cite{starobinsky1980}. Alternatively, other modifications of the GR were studied in the context of inflation as well as the unification of accelerated phases of the Universe \cite{nojrii2017,cognola2008}. Note that inflation for the mimetic gravity with its extensions, was also recently investigated \cite{zhong2018,nojiri2014,odintsov2018,lambiase2015,Mansoori2021}.
In the context of the above, herein we combine mimetic idea with the general higher-curvature $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity. We intent to extend previously considered mimetic framework with more general $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity, with respect to the cosmological reconstruction of various models. After brief discussion of the field equations, by using the reconstruction method, we obtain the power law evolution model, the symmetric bounce as well as the transition between radiation and matter dominated phases. Then, in reference the current experimental data, we discuss inflationary model of the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity. By using the reconstruction methods, we obtain Lagrange multiplier and potential responsible for inflationary evolution described by the scale factor $H(N)=\alpha e^{\beta N}+\gamma $. Finally by employing the perfect fluid approach, we show that it is possible to achieve inflationary evolution in the discussed theory, in agreement with the current data.
\section{Theoretical Model}
The main idea behind the mimetic gravity is that the $g_{\mu\nu}$ metric (the fundamental variable in gravity) may be expressed by the new degrees of freedom \cite{lambiase2015}. By doing so, such degrees of freedom can admit wider class of solutions than the standard GR. In particular, the conformal degree of freedom can be isolated by expressing the physical metric in terms of the scalar field $\phi$ and the auxiliary metric $\hat{g}_{\mu\nu}$ as follows:
\begin{align}
g_{\mu\nu}=-\hat{g}^{\alpha\beta}\partial_{\alpha}\phi\partial_{\beta}\phi \hat{g}_{\mu\nu}.
\label{eq:refname1}
\end{align}
Note that the above parametrization posses Weyl symmetry under the conformal transformation, {\it i.e.} $\hat{g}_{\alpha\beta}=e^{\omega(x)}g_{\alpha\beta}$. It is also worth to remark, that instead of using the physical metric $g_{\mu\nu}$ in variation of the gravitational action, one can use variation with respect to the metric $\hat{g}_{\mu\nu}$ and the scalar field $\phi$. As a result, from Eq. (\ref{eq:refname1}) the following constraint can be obtained:
\begin{align}
g^{\mu\nu}(\phi,\hat{g}_{\mu\nu})\partial_{\mu}\phi\partial_{\nu}\phi=-1.
\label{eq:refname2}
\end{align}
In order to impose given constraint at the action level we use the Lagrange multiplier formalism \cite{capoziello2010,makarenko2016}. In the mimetic regime we need to introduce the Lagrange multiplier ($\lambda$) that corresponds to the mimetic constraint \cite{lambiase2015}. Hence, for the mimetic $f(R,R^{\mu\nu}R_{ \mu\nu})$ gravity we have:
\begin{align}
S=\int d^4x\Big[\sqrt{-g}\big( f(R,R_{\mu\nu}R^{\mu\nu})-V(\phi)+\lambda(g^{\mu\nu}\partial_\mu\phi\partial_\nu\phi+1)\big) \Big],
\label{eq:refname3}
\end{align}
where
$f(R,R^{\mu\nu}R_{ \mu\nu})$ is the analytical function of the Ricci scalar $R$ and contraction of the Ricci tensors (the Ricci squared term) $R^{\mu\nu}R_{ \mu\nu}$. Moreover, $V(\phi)$ denotes scalar (mimetic) potential. We note that the equations obtained from the variation with respect to the physical metric $g_{\mu\nu}$ with imposed mimetic constraint are fully equivalent to the equations that one can derive by using action written in terms of the auxiliary metric $\hat{g}_{\mu\nu}$. We emphasize that standard matter fields are not part of our considerations since one of the core characteristics of the mimetic theory is to \textit{mimic} matter content \cite{lambiase2015}. In what follows, the variation of the action (\ref{eq:refname3}) with respect to the components of the metric tensor $g^{\mu\nu}$, gives the following field equation:
\begin{align}\nonumber
f_R R_{\mu\nu}&-\nabla_{\mu}\nabla_{\nu}f_R+g_{\mu\nu}\Box f_R+\Box( R_{\mu\nu}f_Y)+g_{\mu\nu}\nabla^{\alpha}\nabla^{\beta}( R_{\alpha\beta}f_Y)-2\nabla_{\alpha}\nabla_{\beta}(R^{\alpha}_{(\mu}\delta^{\beta}_{\nu)}f_Y)+2R^{\alpha}_{\mu}R_{\alpha\nu}f_Y\\
&-\frac{1}{2}g_{\mu\nu}(f-V(\phi)+\lambda(g^{\alpha\beta}\partial_\alpha\phi\partial_\beta\phi+1))+\lambda \partial_\mu\phi\partial_\nu\phi=0,
\label{eq:refname4}
\end{align}
where $Y=R_{\mu\nu}R^{\mu\nu}$, $f_Y=\frac{\partial f(R,Y)}{\partial R}$ and $f_Y=\frac{\partial f(R,Y)}{\partial Y}$, whereas the indices in brackets denote symmetrization ($A_{(\mu\nu)}=\frac{1}{2}(A_{\mu\nu}+A_{\nu\mu})$). Next, by varying the action with respect to the scalar field $\phi$ the following scalar equation can be obtained:
\begin{align}
2\nabla^{\mu}(\lambda \partial_{\mu}\phi)+V'(\phi)=0,
\label{eq:refname5}
\end{align}
where prime stands for the differentiation with respect to the scalar field ($V'(\phi)=\frac{dV(\phi)}{d\phi}$). It is important to remark, that the variation of the action with respect to $\lambda$ returns mimetic constraint.
Further in this work we assume that the background geometry is described by the FLRW metric, with the following line element:
\begin{align}
ds^2=-dt^2+a^2(t)dx_idx^i,
\label{eq:refname6}
\end{align}
where $i=1,2,3$ and $a(t)$ is the scale factor. For simplicity, we also assume that the scalar field depends only on time, {\it i.e.} $\phi=\phi(t)$. By using the FLRW metric (\ref{eq:refname6}), the corresponding $(t,t)$ component of the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity equations (\ref{eq:refname4}) is:
\begin{align}\nonumber
-3(\Dot{H}+H^2)f_R&+3 H \partial_t f_R+\frac{1}{2}(f+\lambda(\dot{\phi}^2+1)- V(\phi))+6(2 H \ddot{H}-2\dot{H}^2+3H^2\dot{H}-3H^4)f_Y\\ &+6(2 H \dot{H}+3H^3)\partial_t f_Y=0,
\label{eq:refname7}
\end{align}
while the $(i,j)$ component reads:
\begin{align} \nonumber
&(\dot{H}+3H^2)f_R-2 H \partial_t f_R-\partial_{tt}f_R-\frac{1}{2}(f-\lambda(\dot{\phi}^2-1)-V(\phi))-2(2\dddot{H}+12 H \ddot{H}-9H^4+9H^2\dot{H}+6\dot{H}^2)f_Y
\\
&-2(18 H\dot{H}+6H^3+4\ddot{H})\partial_t f_Y-2(2\dot{H}+3H^2)\partial_{tt}f_Y=0.
\label{eq:refname8}
\end{align}
From Eq. (\ref{eq:refname5}) the corresponding scalar equation is given as:
\begin{align}
6H\lambda \dot{\phi}+2(\dot{\lambda}\dot{\phi}+\lambda\Ddot{\phi})-V'(\phi)=0.
\label{eq:refname9}
\end{align}
Note, that {\it dot} denotes differentiation with respect to the cosmic time $t$ and $H=\dot{a}/a$ is the Hubble rate. For brevity, we note that the Ricci scalar and the Ricci tensor squared in the FLRW spacetime are:
\begin{align}
R=6(\dot{H}+2H^2),\;\;\;\;
Y=R_{\mu\nu} R^{\mu\nu}=12\big(\dot{H}^2+3H^2(\dot{H}+H^2)\big).
\label{eq:refname10}
\end{align}
Finally, the mimetic constraint for the metric given by Eq. (\ref{eq:refname6}) is:
\begin{align}
\dot{\phi}^2=1,
\label{eq:refname11}
\end{align}
and yields:
\begin{align}
\phi=t+C,
\label{eq:refname12}
\end{align}
with $C$ being integration constant. Keeping in mind that the scalar field appears only with derivatives, shift symmetry allows for setting constant $C=0$ without losing generality. By using $\phi=t$ the field and scalar equations (\ref{eq:refname7}-\ref{eq:refname9}) reduce to:
\begin{align}\nonumber
-3(\Dot{H}+H^2)f_R&+3 H \partial_t f_R+\frac{1}{2}(f- V(\phi))+6(2 H \ddot{H}-2\dot{H}^2+3H^2\dot{H}-3H^4)f_Y\\ &+6(2 H \ddot{H}+3H^3)\partial_t f_Y+\lambda=0,
\label{eq:refname13}
\end{align}
\begin{align} \nonumber
&(\dot{H}+3H^2)f_R-2 H \partial_t f_R-\partial_{tt}f_R-\frac{1}{2}(f-V(\phi))-2(2\dddot{H}+12 H \ddot{H}-9H^4+9H^2\dot{H}+6\dot{H}^2)f_Y
\\
&-2(18H \dot{H}+6H^3+4\ddot{H})\partial_t f_Y-2(2\dot{H}+3H^2)\partial_{tt}f_Y=0,
\label{eq:refname14}
\end{align}
and
\begin{align}
6H\lambda +2\dot{\lambda}-V'(\phi)=0.
\label{eq:refname15}
\end{align}
Hence, Eq. (\ref{eq:refname14}) can be used to determine potential $V(\phi)$ that generates evolution of the Hubble parameter for the specific $f(R,Y)$ gravity:
\begin{align}\nonumber
V(\phi=t)=&-2(\dot{H}+3H^2)f_R+4 H \partial_t f_R+2\partial_{tt}f_R+f+4\big[(2\dddot{H}+12 H \ddot{H}-9H^4+9H^2\dot{H}+8\dot{H}^2)f_Y
\\
&+(18H \dot{H}+6H^3+4\ddot{H})\partial_t f_Y+(2\dot{H}+3H^2)\partial_{tt}f_Y\big].
\label{eq:refname16}
\end{align}
By combining Eqs. (\ref{eq:refname13}) and (\ref{eq:refname14}), the analytic form of the Lagrange multiplier $\lambda(t)$ is provided:
\begin{align}
\lambda(t)=2\dot{H}f_R-H\partial_t f_R+ \partial_{tt}f_R +4(3 H \ddot{H}+\dddot{H}+\dot{H}^2)f_Y- 2(3H^3 -12 H \dot{H}-4 \ddot{H})\partial_t f_Y+2(2\dot{H}+3H^2)\partial_{tt}f_{Y}.
\label{eq:refname17}
\end{align}
This equation allows to obtain the desired evolution in various ways. Given the Hubble parameter $H(t)$ for the specified $f(R,Y)$ function, one can find corresponding multiplier $\lambda$ and mimetic potential $V(\phi)$. Also, given the form of the mimetic potential and the Lagrange multiplier $\lambda $, one can solve above equations to find proper $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity model that realizes cosmological scenario of interest.
\section{Reconstruction of the FLRW cosmologies}
In the following section, we use reconstruction methods in order to obtain various cosmological scenarios in the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity. The procedure of cosmological reconstruction is widely used in the context of modified theories of gravity and helps with the extraction of physical implications in the theory of interest \cite{nojiri2014,nojiri2006,zubair2016,bamba2012,Hossienkhani2014,rudra2016,Carloni2012,Kaczmarek2020}. In general, obtaining solutions for the GR's extensions may be troublesome, since even small modifications can drastically increase difficulty of equations. The reconstruction scheme used in the extensions of the GR relies on the fact that arbitrary function (or potential) is used in the definition of the modified gravity. Once the Hubble factor $H$ is specified, the field equations are solved to obtain the model of interest. This technique is a good way to survey modified theories of gravity, since satisfying well established cosmological models in gravity's extensions is desired. Moreover, this technique yields difference between the GR and its modifications, namely specific class (or classes) of the extended gravity may describe any given spacetime \cite{nojiri2011}. In this section, once the function $f(R,Y)$ is given, we obtain the Lagrange multiplier $\lambda$ and potential $V$ satisfying specified evolution. In general, by using this technique one can achieve model describing any given Hubble rate such as inflation or transitions between phases for the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ approach.
In the present study, we assume specific form of the $f(R,R_{\mu\nu}R^{\mu\nu})$ function, where Ricci squared term enters with arbitrary power $p$, {\it i. e.}:
\begin{align}
f(R,Y)=R + dY^p.
\label{eq:refname18}
\end{align}
Thus, the rearranged equations for this choice of the $f(R,Y)$ gravity are:
\begin{align}\nonumber
&V(t)=-2(\dot{H}+3H^2)f_R+4 H \partial_t f_R+2\partial_{tt}f_R+f+4dp\big [(2\dddot{H}+12 H \ddot{H}-9H^4+9H^2\dot{H}+8\dot{H}^2)Y^{p-1}
\\
&+(18H \dot{H}+6H^3+4\ddot{H})(p-1) Y^{p-2}\dot{ Y} +(2\dot{H}+3H^2)((p-1)(p-2) Y^{p-3} \dot{Y}^2+(p-1) Y^{p-2} \ddot{Y})\big],
\label{eq:refname19}
\end{align}
and
\begin{align}\nonumber
&\lambda(t)=2\dot{H}f_R-H\partial_t f_R+ \partial_{tt}f_R +4dp(3 H \ddot{H}+\dddot{H}+\dot{H}^2)Y^{p-1}- 2dp(3H^3 -12 H \dot{H}-4 \ddot{H})(p-1)Y
^{p-2}\dot{Y}\\
&+2dp(2\dot{H}+3H^2)\big((p-1)(p-2) Y^{p-3} \dot{Y}^2+(p-1) Y^{p-2} \ddot{Y}\big).
\label{eq:refname20}
\end{align}
As an example, let us consider the power-law scale factor and the corresponding Hubble parameter:
\begin{align}
a(t)=a_0 t^{n},\;\;\; H(t)=\frac{n}{t}.
\label{eq:refname21}
\end{align}
The solutions of this type are very useful in cosmology, especially in describing various phases in the history of the Universe \cite{hobson2006}. We note that for $0<n<1$, the decelerated universe occurs where for dust dominated Universe $n=\frac{2}{3}$ and for the radiation era $n=\frac{1}{2}$. The corresponding Ricci scalar and Ricci squared are:
\begin{align}
R=-6\frac{n}{t^2}(1-2n),\;\;\;\;Y=12n^2\frac{(3n^2-3n+1)}{t^4}.
\label{eq:refname22}
\end{align}
By using this expressions in the field equations, one can reconstruct the Lagrange multiplier and the mimetic potential for the $f(R,R_{\mu\nu}R^{\mu\nu}$ gravity. The general expressions obtained for the Lagrange multiplier and potential are given in the Appendix. For the dust $n=2/3$ and the inverse $p=-1$ power of $Y$, one gets:
\begin{align}
\lambda(t)=d\frac{45}{8}t^4-\frac{4}{3t^2}
\label{eq:refname23}
\end{align}
and
\begin{align}
V(t)=d\frac{135}{16}t^4-\frac{8}{3t^2}.
\label{eq:refname24}
\end{align}
It is worth to mention, that the Ricci squared contribution leads to the higher-order terms in the expressions for $V$ and $\lambda$. In the absence of this term ($d=0$), the potential takes form $\frac{8}{3}t^{-2}$ that is consistent with the result obtained in the standard mimetic gravity, {\it i.e.} $V(t)=C_0t^{-2}$ for constant $C_0=8/3$ \cite{lambiase2015}.
As an another example we consider the exponential symmetric bounce cosmology, associated with the scale factor and the Hubble rate:
\begin{align}
a(t)=e^{\alpha t^2},\;\;\;\; H(t)=2 \alpha t.
\label{eq:refname25}
\end{align}
One of the main characteristics of the bounce cosmologies is the absence of the initial singularity. These theories constitute also interesting alternatives to the well-established inflationary paradigm \cite{Brandenberger2017,brandenberger1993}. For these scenarios, the Universe contraction decreases the effective radius of the Universe to the minimal size. Then, accelerated expansion occurs. In this case the Ricci scalar and Ricci tensor squared are:
\begin{align}
R=-6 (2 \alpha + 8 t^2 \alpha^2),\;\;\;\; Y=48 \alpha^2 (1 + 6 t^2 \alpha + 12 t^4 \alpha^2).
\label{eq:refname26}
\end{align}
The obtained Lagrange multiplier and mimetic potential take form:
\begin{align}
\lambda(t)=4\alpha -\frac{d (3456 \alpha ^5
t^{10}+11376 \alpha ^4
t^8+8208 \alpha ^3 t^6+1788
\alpha ^2 t^4+48 \alpha
t^2-11)}{144 \alpha
^2 (12 \alpha ^2 t^4+6
\alpha t^2+1)^4},
\label{eq:refname27}
\end{align}
\begin{align}
V(t)= -6 (2 \alpha + 8 t^2 \alpha^2) -
2 (2 \alpha + 12 t^2 \alpha^2)+ \frac{d (1152 \alpha ^6
t^{12}+2976 \alpha ^5
t^{10}+816 \alpha ^4
t^8-472 \alpha ^3 t^6-124
\alpha ^2 t^4+14 \alpha
t^2+3)}{16 \alpha ^2
(12 \alpha ^2 t^4+6
\alpha t^2+1)^4},
\label{eq:refname28}
\end{align}
when again we use $p=-1$ and the general form of the potential and the Lagrange multiplier is listed in the Appendix.
In the following part, we provide reconstruction of the model that describes transition between matter-dominated and accelerated phases of the Universe's history. In this scenario, the Hubble rate is given by \cite{houndjo2012}:
\begin{align}
H=g_0+\frac{g_1}{t}.
\label{eq:refname29}
\end{align}
It is important to remark, that for small $t$, $H\approx \frac{g_1}{t}$ and the Universe is filled with the perfect fluid with the EoS (equation of state) parameter $w=-1+2/3g_1$. On the other hand, for the large $t$, the Hubble rate approaches $H\rightarrow g_0$ and the Universe looks like the de-Sitter one \cite{nojiri2011,houndjo2012}. We note that in this scenario, there is no real matter and the contribution of the mimetic $f(R,Y)$ gravity plays the role of the standard matter content.
In the small $t$ limit, the Lagrange multiplier and the mimetic potential are equal to:
\begin{align}
\lambda(t)\approxeq\frac{12 d g_1{}^2 t^6-43 d
g_1 t^6+30 d t^6-648
g_1{}^8+1296 g_1{}^7-1080
g_1{}^6+432 g_1{}^5-72
g_1{}^4}{36 g_1{}^3 (3
g_1{}^2-3 g_1+1){}^2
t^2},
\label{eq:refname30}
\end{align}
\begin{align}
V(t)\approxeq \frac{6 d g_1{}^3 t^6-19 d
g_1{}^2 t^6-15 d g_1 t^6+20
d t^6+648 g_1{}^9-1728
g_1{}^8+1944 g_1{}^7-1152
g_1{}^6+360 g_1{}^5-48
g_1{}^4}{12 g_1{}^3 (3
g_1{}^2-3 g_1+1){}^2
t^2},
\label{eq:refname31}
\end{align}
while in the large $t$ limit one gets:
\begin{align}
\lambda(t)\approxeq0,\;\;\; V(t)\approxeq6g_0^2+\frac{d}{16g_0^4}
\label{eq:refname32}
\end{align}
where again we have chosen $p=-1$. Note also that the full form of the obtained potential and the Lagrange multiplier can be found in the Appendix. Thus, in principle, the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity is able to unify phases dominated by matter with the transition to the late time accelerated evolution even in the absence of real matter for the appropriate mimetic potential and Lagrange multiplier.
To this end, we reconstruct the particular class of $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity that satisfy mimetic field equations, once particular mimetic potential and the Lagrange multiplier are assumed. This is inversion of the previously used reconstruction. In this approach, as an example, we consider the de-Sitter space as specified by the constant Hubble parameter:
\begin{align}
H=H_0.
\label{eq:refname33}
\end{align}
We remark that this spacetime usually serves as a description of the accelerated expansion phase and the early Universe inflation \cite{hobson2006}. By assuming the constant potential $V$ and the multiplier $\lambda$ as:
\begin{align}
V(t)=\alpha H_0,\;\;\;\; \lambda(t)=2 \beta H_0,
\label{eq:refname34}
\end{align}
we obtain the following partial differential equation (PDE):
\begin{align}
3H_0^2 f_R+\frac{1}{2}f+6H_0^2+ \frac{(\beta-\alpha)}{2}H_0-18H_0^2f_Y=0.
\label{eq:refname35}
\end{align}
This PDE has the following solution:
\begin{align}
f(R,Y)=12H_0^2+(\alpha-\beta)H_0+e^{-\frac{R}{6H_0^2}} F(6H_0^2R+Y),
\label{eq:refname36}
\end{align}
where $F$ is an arbitrary function.
In conclusion, the reconstruction technique can work in both ways: one can assume specific functional form of the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ and cosmological evolution to obtain the mimetic potential and the Lagrange multiplier and \emph{vice versa}. In general, by using this procedure one can obtain any cosmological scenario for the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity.
\section{Inflation}
Inflation is believed to be one of the fundamental building blocks of the modern cosmology that may solve some of its problems \cite{guth1981,starobinsky1980}. This is to say, the main goal of the inflationary theory is to explain primordial fluctuations {\it i.e.} the density variations occurring in very early stages of the cosmic evolution. Since inflation is the most plausible scenario for the early Universe, the modified theories of gravity should be able to properly describe this phase of the cosmic history. Moreover, the deviations from GR can be interpreted as a quantum-induced corrections or motivated by the ultra-violet (UV) behavior of the quantum gravity, playing an important role in the inflationary phase of the early Universe \cite{bamba2015}. Thus, viable description of this stage in any modified gravity is desired. Therefore, in the following section we analyse inflation in the context of the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity to explore ability to obtain model comparable with the recent BICEP2/Keck observations \cite{ade2016,bamba2015}. Note, that the cosmological reconstruction can also be successfully applied to the discussed scenario. In this section we obtain the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity that describes inflationary model which is consistent with the recent observational data. Again, the functional form of Eq. (\ref{eq:refname18}) is assumed.
While discussing the inflationary cosmology it is worth to use the number of e-foldings $N$ (intervals for which the scale factor grows by the $N$ factors of $e$) instead of the cosmic time $t$. The relation between the scale factor and the e-folding number $N$ is following, $e^{N}=\frac{a}{a_0}$, where $a_0$ is the initial value of the scale factor in the initial time instance. For brevity, we list transformation rules for the time derivatives with respect to the e-foldings number:
\begin{align}\nonumber
&\frac{d}{dt}=H(N) \frac{d}{dN},\;\;\;\; \frac{d^2}{dt^2}=H^2(N)\frac{d^2}{dN^2}+H(N)H'(N)\frac{d}{dN},\\
&\frac{d^3}{dt^3}=3H^2(N)H'(N)\frac{d^2}{dN^2}+H^2(N)H''(N)\frac{d}{dN}+H(N)H'^2(N)\frac{d}{dN}+H^3(N)\frac{d^3}{dN^3}.
\label{eq:refname37}
\end{align}
We note that prime symbols ($'$) correspond to the derivatives $\frac{dH(N)}{dN}$.
In order to study the {\it slow-roll} indices we use perfect fluid approach developed in \cite{bamba2014}. In this approach extra terms in the gravitational action (\ref{eq:refname1}) can be regarded as the perfect fluid. This formalism allows to obtain the spectral indices independently from the model.
According to this approach, the {\it slow-roll} parameters are given as a functions of the Hubble rate $H$:
\begin{align}\nonumber
\epsilon&=\frac{-4H(N)}{H'(N)}\Big(\frac{H'^2(N)+6H'(N)H(N)+H''(N)H(N)}{H'(N)H(N)+3H^2(N)}\Big)^2,\\
\eta&=-\frac{\Big(9\frac{H'(N)}{H(N)}+3\frac{H''(N)}{H(N)}+\frac{1}{2}\Big(\frac{H'(N)}{H(N)}\Big)^2-\frac{1}{2}\Big(\frac{H''(N)}{H(N)}\Big)^2+\frac{H''(N)}{H'(N)}+\frac{H'''(N)}{H'(N)}\Big)}{2\big(3+\frac{H'(N)}{H(N)}\big)}.
\label{eq:refname38}
\end{align}
By using the above, the spectral index of the curvature perturbations $n_s$ and the scalar-to-tensor ratio $r$ in terms of the {\it slow-roll} parameters can be provided:
\begin{align}
n_s=-6\epsilon+2\eta+1,\;\;\;\;r=16\epsilon. \label{eq:refname39}
\end{align}
We remark that the spectral index describes variation of the density fluctuations with respect to the scale. On the other hand, the scalar-tensor ratio relates spectras of the scalar and tensor perturbations. By using Eqs. (\ref{eq:refname16}), (\ref{eq:refname17}) and (\ref{eq:refname37}), the field equations in terms of the e-foldings number $N$ are:
\begin{align}\nonumber
&\lambda(N)=H(N)^2 f_R''(N)-H(N)^3
f_R'(N) H'(N)+H(N)
f_R'(N) H'(N)+2
f_R(N) H(N) H'(N)+4
H(N)^3 f_Y''
H'(N)\\ \nonumber
&+12 H(N)^2
f_Y'
H'(N)^2+H(N)^3
f_Y' \Big(8
H''(N)+30 H'(N)\Big)-6
H(N)^4
\Big(f_Y'(N)-f_Y''(N)\Big)+4
f_Y(N) H(N) H'(N)^3\\
&+4
f_Y H(N)^3
\Big(H^{(3)}(N)+3
H''(N)\Big)+16
f_Y H(N)^2 H'(N)
\Big(H''(N)+H'(N)\Big),
\label{eq:refname40}
\end{align}
\begin{align}\nonumber
&V(N)=f+2 H(N)^2 f_R''(N)+2
H(N) f_R'(N) H'(N)+4
H(N)^2 f_R'(N)-2
f_R(N) H(N)
\Big(H(N)
(H''(N)+3)+H'(N)
^2\Big)\\ \nonumber
&+8 H(N)^3
f_Y''(N) H'(N)+12
H(N)^4 f_Y''(N)+16
H(N)^3 f_Y'(N)
H''(N)+84 H(N)^3
f_Y'(N) H'(N)+24
H(N)^2 f_Y'(N) \\ \nonumber &\times
H'(N)^2+24 H(N)^4
f_Y'(N)+4
f_Y(N) H(N) \Big(2
H'(N)^3+2 H(N)^2
(H^{(3)}(N)+6
H''(N)\Big)+4 H(N) H'(N)
\Big(2 H''(N)\\&+3
H'(N)\Big) -9
H(N)^3),
\label{eq:refname41}
\end{align}
where we have used relations from Eqs. (\ref{eq:refname37}). Thus, one can obtain the Lagrange multiplier and potential corresponding to the scale factor $H(N)$, effectively reconstructing any given inflationary model. We note, that inverse reconstruction method is also possible once $\lambda(N)$ and $V(N)$ are specified.
As an example of the inflationary reconstruction, we consider a inflationary model described by the Hubble factor \cite{zhong2018}:
\begin{align}
H(N)=\alpha e^{\beta N}+\gamma.
\label{eq:refname42}
\end{align}
The corresponding mimetic potential and the Lagrange multiplier are given by:
\begin{align}\nonumber
\lambda(N)&=\frac{1}{ {36 (\gamma +\alpha
e^{\beta
N})^3}}\Bigg[\alpha \beta e^{\beta
N} (72
(\gamma +\alpha
e^{\beta
N})^4-\frac{d}{{(3 \gamma
^2+\alpha ^2 (\beta
^2+3 \beta +3) e^{2
\beta n}+3 \alpha (\beta
+2) \gamma e^{\beta
n})^4}}\\ \nonumber
& \times\Big[-18 (\beta
^2+3 \beta -6) \gamma
^6+\alpha ^6 (\beta^2
+3 \beta +3)^2
(30 \beta ^2+43 \beta
+12) e^{6 \beta
n}+\alpha ^5 (30 \beta
^6+394 \beta ^5+1806 \beta
^4+4095 \beta ^3\beta\\ \nonumber
& +4941 \beta
^2+2961 +648)
\gamma e^{5 \beta n}+3
\alpha ^4 (3 \beta
^6+69 \beta ^5+494 \beta
^4+1575 \beta ^3+2502 \beta
^2+1920 \beta +540)
\gamma ^2 e^{4 \beta n}+3
\alpha ^3\\ \nonumber&\times (12 \beta
^5+142 \beta ^4+725 \beta
^3+1698 \beta ^2+1830 \beta
+720) \gamma ^3 e^{3
\beta n}+3 \alpha ^2
(10 \beta ^4+90 \beta
^3+432 \beta ^2+825 \beta
+540) \gamma ^4 e^{2
\beta n}\\ &-9 \alpha (2
\beta ^3+3 \beta ^2-37
\beta -72) \gamma ^5
e^{\beta
n})\Big]\Bigg],
\label{eq:refname43}
\end{align}
\begin{align}\nonumber
V(N)&=-\frac{1}{{36
(\gamma +\alpha
e^{\beta
N})^3}}\Bigg[-\frac{3 d (\gamma
+\alpha e^{\beta
N})}{3 \gamma
^2+\alpha ^2 (\beta
^2+3 \beta +3) e^{2
\beta N}+3 \alpha
(\beta +2) \gamma e^{\beta
N}} -216
\alpha \beta e^{\beta
N} (\gamma
+\alpha e^{\beta
N})^4\\ \nonumber&-\frac{1}{{(3
\gamma ^2+\alpha ^2
(\beta ^2+3 \beta
+3) e^{2 \beta
N}+3 \alpha (\beta
+2) \gamma e^{\beta
N})^3}}\Big[4 \alpha
\beta d e^{\beta
N} (3 \gamma
^2+\alpha ^2 (4 \beta
^2+9 \beta +3) e^{2
\beta N}\\ \nonumber &+\alpha
(2 \beta ^2+9 \beta
+6) \gamma e^{\beta
N}) (3
(\beta +4) \gamma ^2+4
\alpha ^2 (\beta ^2+3
\beta +3) e^{2 \beta
N}+\alpha (2
\beta ^2+15 \beta
+24) \gamma e^{\beta
N}) \Big] \\ \nonumber &+\frac{d
(-9 \gamma ^3+3 \alpha
^3 (4 \beta ^3+8 \beta
^2-3) e^{3 \beta
N}+3 \alpha ^2
(4 \beta ^3+12 \beta
^2-9) \gamma e^{2
\beta N}+\alpha
(2 \beta ^3+12 \beta
^2-27) \gamma ^2
e^{\beta
N})}{(3
\gamma ^2+\alpha ^2
(\beta ^2+3 \beta
+3) e^{2 \beta
N}+3 \alpha (\beta
+2) \gamma e^{\beta
N})^2}\\ \nonumber &+\frac{1}{(3
\gamma ^2+\alpha ^2
(\beta ^2+3 \beta
+3) e^{2 \beta
N}+3 \alpha (\beta
+2) \gamma e^{\beta
N})^4}\big(2
\alpha \beta ^2 d e^{\beta
N} (3 \gamma
+\alpha (2 \beta +3)
e^{\beta N})
(-9 (\beta +4) \gamma
^5\\\nonumber &+28 \alpha ^5 (\beta
^2+3 \beta +3)^2 e^{5
\beta N}+4 \alpha
^4 (7 \beta ^4+72
\beta ^3+255 \beta ^2+396
\beta +243) \gamma
e^{4 \beta
N}\\ \nonumber&+\alpha ^3 (8
\beta ^4+141 \beta ^3+786
\beta ^2+1719 \beta
+1368) \gamma ^2 e^{3
\beta N}+3 \alpha
^2 (7 \beta ^3+64
\beta ^2+231 \beta
+264) \gamma ^3 e^{2
\beta N}\\ &+3 \alpha
(2 \beta ^2+15 \beta
+36) \gamma ^4
e^{\beta
N})\big)+72
\alpha ^2 \beta ^2 e^{2
\beta N}
(\gamma +\alpha
e^{\beta
N})^4+72
\alpha \beta ^2 e^{\beta
N} (\gamma
+\alpha e^{\beta
N})^5\Bigg] .
\label{eq:refname44}
\end{align}
Further, the {\it slow-roll} parameters from the perfect fluid approach are:
\begin{align}
\epsilon=-\frac{\alpha \beta e^{\beta n} ((\beta +6) \gamma +2 \alpha
(\beta +3) e^{\beta n})^2}{4 (\gamma +\alpha e^{\beta
n}) (3 \gamma +\alpha (\beta +3) e^{\beta n})^2}
\label{eq:refname45}
\end{align}
and
\begin{align}
\eta=-\frac{\beta ((\beta +6) \gamma ^2+8 \alpha ^2 (\beta +3) e^{2
\beta n}+2 \alpha (4 \beta +15) \gamma e^{\beta n})}{4
(\gamma +\alpha e^{\beta n}) (3 \gamma +\alpha
(\beta +3) e^{\beta n})}.
\label{eq:refname46}
\end{align}
Finally the corresponding spectral index and the scalar-tensor ratios are listed below:
\begin{align}\nonumber
n_s&=\frac{1}{2
(\gamma +\alpha e^{\beta n}) (3 \gamma +\alpha
(\beta +3) e^{\beta n})^2} \Big[-3 (\beta ^2+6 \beta -6) \gamma ^3+2 \alpha ^3 (\beta
+3)^2 (2 \beta +1) e^{3 \beta n}\\
&+2 \alpha ^2 (2 \beta ^3+16
\beta ^2+39 \beta +27) \gamma e^{2 \beta n}+\alpha (2
\beta ^3+3 \beta ^2+12 \beta +54) \gamma ^2 e^{\beta n} \Big],
\label{eq:refname47}
\end{align}
\begin{align}
r=-\frac{4 \alpha \beta e^{\beta n} ((\beta +6) \gamma +2 \alpha
(\beta +3) e^{\beta n})^2}{(\gamma +\alpha e^{\beta
n}) (3 \gamma +\alpha (\beta +3) e^{\beta n})^2}.
\label{eq:refname48}
\end{align}
The inflationary parameters depend on the e-foldings number $N$ and the three free parameters $\alpha,\beta, \gamma$, which are associated with the Hubble rate (\ref{eq:refname42}). We note that this set can be further simplified by introducing $\Upsilon=\alpha/\gamma$. In Fig. (\ref{fig:1}) we present the behavior of the observational indices as a function of the $\beta$ parameter. For exemplary purposes, we have chosen three values of ratio $\Upsilon$. Gray regions indicate parameter range that is compatible with the BICEP2/Keck array data \cite{ade2016}. Based on the obtained results we observe that the spectral index $n_s$ diminishes with increase of the $\beta$, while scalar to tensor ratio $r$ exhibits increasing behavior. This inflationary model was extensively studied by Zhong and collaborators \cite{zhong2018} in mimetic $f(\mathcal{G})$. We refer reader to their work for the detailed survey of the ranges of parameters that are compatible with the recent observational data provided by BICEP2/Keck array \cite{ade2016}.
The main goal of this part of our study was to reconstruct proper inflationary model in the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ theory. For more detailed discussion of other inflationary scenarios in the mimetic gravity, we refer reader to \cite{zhong2018}. We conclude that extension of the GR presented here is comparable with other mimetic gravities and may consitiute interesting alternative \cite{zhong2018,Odintsov2016}.
\begin{figure}[!htb]
\begin{minipage}{0.5\textwidth}
\centering
\includegraphics[width=1.
\linewidth]{fig/3.pdf}
\end{minipage}\hfill
\begin{minipage}{0.5\textwidth}
\centering
\includegraphics[width=1.\linewidth]{fig/4.pdf}
\end{minipage}
\caption{Behavior of the observational indices as a functions of parameter $\beta$ for specific ratios $\Upsilon=\alpha/\gamma$ ($N=60$). Gray regions indicate compatibility of the $n_s$ and $r$ with recent Planck and BICEP2/Keck array data. }
\label{fig:1}
\end{figure}
\section{Summary}
In this work, we presented mimetic extension of the $f(R,R_{\mu\nu}R^{\mu\nu})$ gravity using the Lagrange multiplier formalism with the mimetic scalar potential added to the Lagrangian, where mimetic field isolates conformal degree of freedom. Introduced extension was implemented by using the well-known reconstruction technique in order to obtain models of interest. Assuming functional form as $f(R,Y)=R+dY^p$, we obtained the Lagrange multiplier $\lambda$ and the mimetic potential $U(\phi)$ that satisfy the power law evolution of the Universe. Another reconstructed model described symmetric bounce cosmology. Moreover, we have reconstructed model that satisfy transition between matter dominated and accelerated phase in the history of the Universe, showing that this extension of the mimetic gravity is capable to unify various stages of cosmic history. Additionally, using the inverse reconstruction method and choosing appropriate forms of $\lambda$ and $V$ we have obtained $f(R,Y)$ model describing the de-Sitter model with constant Hubble factor $H_0$. Our work is closed with discussion of the inflationary cosmology in a given regime. By using the Hubble parameter $H(N)=\alpha e^{\beta N}+\gamma$ we get Lagrange multiplier and the mimetic potential which successfully describes inflationary model. The reconstructed inflationary model is phenomenologically viable for the wide range of the parameters when confronted with BICEP2/Keck data \cite{ade2016,zhong2018}.
In conclusion, the mimetic extension of the $f(R,T,R_{\mu\nu}T^{\mu\nu})$ can describe any given early or late-time cosmological model in particularly clear way. In general, the procedure presented here can be extended to other higher-curvature modifications of the general relativity or theories that include coupling with the trace of the energy-momentum tensor such as the $f(R,T,R_{\mu\nu}T^{\mu\nu})$ or $f(\mathcal{G},T)$ approaches \cite{sharif2016,zubair2016}. We also remark that the matter fields were not used in our considerations, since main feature of the mimetic gravity is to \textit{mimic} cosmological behavior driven by the matter fields \cite{lambiase2015,nojiri2016}. In the approach presented here, the mimetic condition is supported by the higher curvature terms of the $f(R,R_{\mu\nu} R^{\mu\nu})$ theory, such that the geometry of the considered theory incarnates matter. Moreover, the mimetic condition can play an important role in further considerations of the various higher order extensions of the general relativity and can be extended by inclusion of the unimodular concept in the considerations \cite{nojiri2016}. We also wish to point out that by choosing an appropriate functions and parameters one can obtain analytical results in the mimetic $f(R,R_{\mu\nu}R^{\mu\nu})$ approach, which generally is hard task in the context of the modified theories of gravity \cite{nojrii2017}. Since there are many other proposals of the modified gravities that satisfy realistic models with sufficient degree of accuracy, the local constraints should be employed, such as the Post-Newtonian or Solar-System tests \cite{Vagnozzi2017}. Future studies should be devoted to this topic, by addressing not only mimetic extension presented here, but also other proposals such as mimetic versions of the $f(R,T)$, $f(R)$ or $f(\mathcal{G})$ gravity.
|
\section{Conclusions}
We present a realistic analysis of $\ensuremath{{t\bar{t}}}{}$ final states using SM and BSM samples and non-$\ensuremath{{t\bar{t}}}{}$ backgrounds corresponding to the integrated luminosity of $1\,\mathrm{ab}^{-1}$, focusing on variables with potential to distinguish the signal of a resonant $\ensuremath{{t\bar{t}}}$ production or $\ensuremath{{t\bar{t}}}$-associated DM pair production at the future LHC energy of 14~TeV.
The models with \ensuremath{{t\bar{t}}}{} production mediated by a vector ($Z'$) or scalar ($y_0$) predict the expected peak in the \ensuremath{{t\bar{t}}}{} invariant mass in all studied 2B0S, 1B1S and 0B2S topologies. The model with \ensuremath{{t\bar{t}}}{}-associated production of a pair of invisible pair of DM particles provide a broad excess in the \ensuremath{{t\bar{t}}}{} mass in all considered topologies. All models predict a broad excess in top quark transverse momentum but also in other spectra. The addition of the semiboosted topologies into the selection can add about 10--15\% events at the $Z'$ mass of 1--1.5~TeV.
The 2D variable $\ensuremath{{H_\mathrm{T}}}+\ensuremath{E^\mathrm{miss}_\mathrm{T}}$ versus the invariant mass of the 4-vector sum of large-$R$ jets $m^\mathrm{vis}_{\sum J}$ performs very well in terms of the signal significance.
Other powerful variables are the $\ensuremath{{t\bar{t}}}$ rest frame angular variables like $\chi^{t\bar{t}}$, $y_{\mathrm{boost}}^{\ttbar{}}$; or $\ensuremath{\delta^{t\bar{t}}}$
and relative dimensionless $\ensuremath{{p^{t}_\mathrm{T}}} / \ensuremath{m^{t\bar{t}}}$, $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}} / \ensuremath{m^{t\bar{t}}}$ and $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}} / \sqrt{\ptto{1}\ptto{2}}$ which also exhibit a large signal significance potential while they also offer lower sensitivity to experimental uncertainties in the jet energy scale determination, being possibly more precise in real measurements. Especially simple yet powerful seems the signal separation for the $\ensuremath{{t\bar{t}}}$-associated production of a pair of invisible DM particles in terms of the 2D distribution of $\ensuremath{E^\mathrm{miss}_\mathrm{T}}$ vs. $\ensuremath{{H_\mathrm{T}}} + \ensuremath{E^\mathrm{miss}_\mathrm{T}}$.
The proposed new 1D or 2D variables may prove useful for signal bump hunting in LHC searches or serve as input for a~further usage in more involved classification approaches. While there do exist modern methods of identifying signal-enriched regions based on multivariate (MV) or machine learning (ML) techniques, the presented 2D BH algorithm may be regarded as a motivational intermediate stage for such more general classification schemes.
In conclusions, we identify statistically and systematics-robust variables with potential to enhance signal significance in sub-regions of these variables. The 2D BH studies lay between standard 1D and multivariate MV/ML techniques.
The 2D variables offer a clear physics insight into regions which can be avoided to enhance signal significance. Also, such variables can be used as input to more powerful MV/ML classifiers.
\section{Introduction}
Search for the $\ensuremath{{t\bar{t}}}$ resonances have been performed by both the ATLAS (\cite{Aaboud:2019roo,Aaboud:2018mjh}) and CMS (\cite{Sirunyan:2018ryr}, also a search for a $tW$ resonance~\cite{Sirunyan:2021fkj}) experiments in both $\ell+$jets and all-hadronic channels, setting limits for masses of beyond-the-standard model (BSM) $Z'$ particles of various scenarios. The focus of this paper is not a study of non-excluded models, but rather the usage of generic scalar and vector resonance models to study possible improvements in reconstructing the $\ensuremath{{t\bar{t}}}$ kinematics in various topologies of the final state, namely regarding the degree of boost of the jets, and identify variables sensitive to the presence of a signal for the selected BSM models.
As the Large Hadron Collider (LHC) at CERN has provided a large dataset and current upgrade works aim increasing the center-of-mass energy as well as the luminosity, boosted topologies in hadronic final states are becoming heavily used tools in both measurements and searches at the LHC. They provide a handle on processes at large momentum transfer, with a chance of observing possible BSM signals.
One of the aims of this study is to employ the usage of events where a boosted $W$-like hadronic jets are reconstructed at medium traverse momenta, in order to study the enhancement of events between the traditional boosted topologies where the boosted jet mass and structure is consistent with a top-jet.
Shapes of several kinematic variables are studied at the particle and detector levels for the presence of a BSM signal in case of a scalar or vector particle decaying into a pair of top quarks, and for the case of an associated production of top quark pairs with a pair of invisible dark matter (DM) particles.
The signal significance is evaluated
for various processes of the \ensuremath{{t\bar{t}}}{} pair production at the central mass system energy of the $pp$ system of 14~TeV, using the \textsc{MadGraph5}\xspace{} event generator and the \textsc{Delphes}\xspace{} parameterized detector simulation.
The presence of a signal peak or other modification of spectra is studied in boosted as well as semiboosted topologies, where, in the latter, one or two boosted jets consistent with the hadronic $W$ candidate are identified, and the top quark candidates are formed by adding a four-momentum of the $b$-tagged jet.
The purpose of this paper is not to optimize details of the $W$ boson tagging algorithms but to explore these usually overlooked topologies in \ensuremath{{t\bar{t}}}{} final states.
While searches for a concrete weak signal usually employ multivariate (MV)
techniques to identify the phase space where the signal can be seen, searches for new physics often look into a single distribution like the invariant mass or missing transverse energy and do not select complicated phase space regions as details of new physics models can vary and shapes of distributions depend on particular model details like the width, mass and spin of a new resonance.
Cross-sections of SM processes are often fitted using simultaneous likelihood fits over many binned variables with detector and modelling systematic uncertainties being allowed to vary and be fitted (profiling). In contrast, in searches for new physics, limits are often set based on a 1D variable using an excess of events over a simulated, (semi) data-driven or fitted background, often validated in control regions, less often based on a MV discriminant due to model dependencies.
We propose an intermediate approach, selecting pairs of variables with a known small correlation in order to quantify a 2D excess of events over the expected background. We suggest dimensionfull as well as dimensionless variables and their pairs as 2D spectra for search for model BSM signals.
\section{Objects reconstruction}
\label{sec:Reconstruction}
Using the \textsc{Delphes}\xspace{} (version {\tt 3.4.1}) detector simulation \cite{deFavereau:2013fsa} with a modified ATLAS card, jets with two distance parameters 0.4 and 1.0 were reconstructed using the anti-$k_t$ algorithm to form corresponding small-$R$ and large-$R$ jets, using the FastJet algorithm~\cite{Cacciari:2011ma} at both particle and detector levels.
The \textsc{Delphes}\xspace{} card was also modified to remove the DM particles from both detector jet clustering as well as from the particle jets.
Although no additional $pp$ interactions (pile-up) were added to the simulated hard-scatter processes, the trimming jet algorithm \cite{Krohn:2009th} as part of the \textsc{Delphes}\xspace{} package was used to obtain jets at both the particle and detector levels with removed soft components, using the parameter of $R_\mathrm{trim} = 0.2$ and modified \ensuremath{{p_\mathrm{T}}}{} fraction parameter $f^{\ensuremath{{p_\mathrm{T}}}}_\mathrm{trim} = 0.03$ (originally 0.05). The trimming algorithm was chosen over the standard non-groomed jets, soft-dropped~\cite{Larkoski:2014wba} and pruned jets~\cite{Ellis:2009me}, with parameters varied, in terms of the narrowness of the mass peaks.
The jet masses shapes of the other algorithms differed in the expected peak positions at $W$ and $t$ masses, although they performed similarly and could be also tuned for the purpose of this analysis. In detail, in the default settings the $W$ and top mass peaks were shifted to higher values by $10$--$20\,$GeV, with a worse shape and shift for the $W$ case which is important for the semiboosted regimes. The new settings help establish both the peaks shape as well as position at the expected values. While the new $\ensuremath{{t\bar{t}}}{}$ invariant mass distribution for the $Z'$ sample with $m_{Z'} = 1\,$TeV shifts down by about $50\,$GeV, the relative peak resolution remains the same.
This procedure leads to a more pronounced reconstructed $W$ as well as top-quark mass peaks in terms of corresponding candidate large-$R$ jet masses, and also to better correlation between the detector and particle levels for derived quantities like the transverse momentum of the top quarks or the \ensuremath{{t\bar{t}}}{} invariant mass.
Custom jet energy scale (JES) correction was applied to large-$R$ jets at the detector level, resulting in an up-scaling correction of 5--15\% depending on $\eta$ and \ensuremath{{p_\mathrm{T}}}{}. Additional custom JES was applied also to the default \textsc{Delphes}\xspace{} JES for small-$R$ jets, resulting in additional correction of 5\% in central rapidities and 4\% at low \ensuremath{{p_\mathrm{T}}}{}. JES closure tests in terms of the ratio of the particle and angularly matched detector-level jet result in agreement within~2\%.
The jet subjettiness variables \cite{Thaler:2010tr} $\tau_{k}$, $k=1,2,3$, and their ratios $\tau_{ij} \equiv \tau_{i}/\tau_{j}$ were employed as provided by the \textsc{Delphes}\xspace{} jet reconstruction properties together with a simple selection on the large-$R$ jet mass to identify jets coming from the hadronic decays of the $W$ boson or a top quark. The large-$R$ jets were tagged as
\begin{itemize}
\item $W$-jets if $ 0.10 < \tau_{21} < 0.60 \, \land \, 0.50 < \tau_{32} < 0.85 \, \land \, m_J \in [70, 110] \,\mathrm{GeV}$;
\item top-jets if $ 0.30 < \tau_{21} < 0.70 \, \land \, 0.30 < \tau_{32} < 0.80 \, \land \, m_J \in [140, 215] \,\mathrm{GeV}$.
\end{itemize}
Illustrated in~Figure~\ref{fig_taus_vs_m} for the $Z'$ sample with $m_{Z'} = 1\,$TeV there are the correlations between the substructure variables and the large-$R$ jet mass, demonstrating the small correlation between $\tau_{21}$ and $\tau_{32}$ but also the worse jet mass resolution at the detector as compared to the particle level.
The above definitions result in the $W$-tagging efficiency of about $0.6$ in the jet \ensuremath{{p_\mathrm{T}}}{} range of $[150,350]\,\mathrm{GeV}$ with mistag rate if about $0.4$; and to the
top-tagging efficiency of about $0.6$ in the jet \ensuremath{{p_\mathrm{T}}}{} range of $[150,350]\,\mathrm{GeV}$ with mistag rate below about $0.2$; with numbers quoted for the \ensuremath{{t\bar{t}}}{} sample, while for the $Z'$ samples the top mistag rate is higher for $\ensuremath{{p_\mathrm{T}}} < 600\,$GeV than in the \ensuremath{{t\bar{t}}}{} samples and falls of again towards the high $\ensuremath{{p_\mathrm{T}}}$.
The $b$-jets at the detector level were used as provided by the tagging efficiency in the default \textsc{Delphes}\xspace{} ATLAS card, while for studies at the particle level, custom tagging based on the presence of a $B$~hadron angularly close to the particle jet was implemented, using $B$~hadrons with $\ensuremath{{p_\mathrm{T}}}{} > 5\,$GeV. Particle-level jets were tagged as $b$-jets if such a $B$~hadron was within $\Delta R = \sqrt{(\Delta\eta)^2 + (\Delta\phi)^2} < 0.4$. Further, in order to account for the practically important case of falsely $b$-tagged jets, a custom random mistagging with a rate of 1\% was added for originally non-$b$-tagged jets.
\begin{figure}[!h]
\begin{center}
\begin{tabular}{ccc}
\includegraphics[width=0.32\textwidth]{figures/migra/migra_zpOnly_NoCuts_ParticleLJetTau21VsMass_NoCuts_ParticleLJetTau21VsMass.pdf} &
\includegraphics[width=0.32\textwidth]{figures/migra/migra_zpOnly_NoCuts_ParticleLJetTau32VsMass_NoCuts_ParticleLJetTau32VsMass.pdf} &
\includegraphics[width=0.32\textwidth]{figures/migra/migra_zpOnly_NoCuts_ParticleLJetTau21VsTau32_NoCuts_ParticleLJetTau21VsTau32.pdf} \\
\includegraphics[width=0.32\textwidth]{figures/migra/migra_zpOnly_NoCuts_DetectorLJetTau21VsMass_NoCuts_DetectorLJetTau21VsMass.pdf} &
\includegraphics[width=0.32\textwidth]{figures/migra/migra_zpOnly_NoCuts_DetectorLJetTau32VsMass_NoCuts_DetectorLJetTau32VsMass.pdf} &
\includegraphics[width=0.32\textwidth]{figures/migra/migra_zpOnly_NoCuts_DetectorLJetTau21VsTau32_NoCuts_DetectorLJetTau21VsTau32.pdf} \\
\end{tabular}
\caption{The $\tau_{21}$ (left) and $\tau_{32}$ (middle) jet structure variables plotted versus the large-$R$ jet mass and the $\tau_{21}$ vs. $\tau_{32}$ for the $Z'$ sample with $m_{Z'} = 1\,$TeV at the particle (top) and detector (bottom) levels. The cross markers are the profile histogram while the solid line stands for the (scaled) projection of the histogram to the $x$ axis. Also indicated is the standard correlation factor~$\rho$.}
\label{fig_taus_vs_m}
\end{center}
\end{figure}
\subsection{Replicas}
The technique of bootstrapping~\cite{Bohm:389738} is employed in order to construct statistically varied pseudo-independent versions (replicas) of 1D as well as 2D histograms, by repeatedly filling 100 replicas by events weighted by random weights drawn from the Poisson distribution of $\mu=1$, \emph{i.e.} as $w \sim \mathrm{Poisson}(1)$.
\noindent This ensures that every event is used in each replica once on average, and enables one to construct statistically correlated replicas across arbitrary histograms (projections) of any dimension over the events. The main aim is to evaluate signal excess over the replicas, obtaining a more robust estimator of the signal significance for each 1D or 2D spectrum in each topology and model.
\section{Results}
\subsection{Pseudo-data study}
For the purpose of creating a realistic mixed sample of a selected SM backgrounds (associated production of a $W$ bosons and $b$-jets), the $\ensuremath{{t\bar{t}}}{}$ sample, as well as a new physics signal, detector-level distributions were created by mixing the samples based on their cross section $\sigma$ using weights $w = \sigma \, L/ N_\mathrm{gen}$ (with $L$ being the luminosity and $N_\mathrm{gen}$ the number of generated events with non-zero matching weights), but also scaling the signal in order to make it more prominent to study its shape and significance.
For the purpose of creating stacked plots consisting of physics and $\ensuremath{{t\bar{t}}}$ background and a selected signal model, the scalar $y_0$ model with the same mass was scaled by a factor of 12 and additional factors of $2^{1/2}$, $2^{3/2}$ and $2^{4.5}$ in the 2B0S, 1B1S and 0B2S topologies, respectively.
Similarly, the $Z'$ model with $m_{Z'} = 1000\,$GeV was scaled by a factor of 10k, with additional $2^{1/2}$, $2^1$ and $2^4$ factors in the same topologies.
The larger $y_0$ production cross section is partially due to the higher gluon-gluon parton luminosity in the production mechanism.
Finally, the sample of the DM pair creation with associated $\ensuremath{{t\bar{t}}}{}$ production was scaled by a factor of 1k and additional factors of $2$ in the 0B2S topology.
The purpose of the additional scaling is to reach a similar level of signal significance across the topologies.
Examples of stacked samples forming predictions to the pseudo-data from the same but statistically independent samples showing a peaked $\ensuremath{m^{t\bar{t}}}$ and the $\chi^{t\bar{t}}$ variable distributions of signal for the $y_0$ model (Figure~\ref{fig_stack_Detector_model_y0_compact_RS}), a broader enhancement in terms of the top quark transverse momentum \ensuremath{{p^{t}_\mathrm{T}}}{} and $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}} / \ensuremath{m^{t\bar{t}}}$ for the $Z'$ model (Figure~\ref{fig_stack_Detector_model_zp_compact_RS}) are shown, together with the large signal in the tail of the $\ensuremath{E^\mathrm{miss}_\mathrm{T}}$ distribution for the DM pair production (Figure~\ref{fig_stack_Detector_models_Met_compact_RS}).
\input{tex/stack_y0_Detector_RS_sel_compact_manual}
\input{tex/stack_zp_Detector_RS_sel_compact_manual}
\input{tex/stack_Met_Detector_RS_sel_compact_manual}
\subsection{Shape comparison}
We present a selection of variables shape comparisons for the 1B1S topology in this section.
The most useful variable for searching for a resonant $\ensuremath{{t\bar{t}}}{}$ production is naturally the invariant mass of the $\ensuremath{{t\bar{t}}}{}$ system, $m^{\ensuremath{{t\bar{t}}}}$. The shape comparison show a clear peak for the vector ($Z'$) as well as the scalar ($y_0$) models, see Figure~\ref{fig_shapes_mtt_1B1S}.
But shapes of other variables are also modified for such a signal, \emph{e.g.} the top quark transverse momentum in~Figure~\ref{fig_shapes_pTt_1B1S}, exhibiting a broader peak rather than just a slope change.
In contrast, the DM model leading to the associated production of a $\ensuremath{{t\bar{t}}}{}$ pair and a pair of DM particles ($\chi_D \bar{\chi}_D$) leads not only to an enhancement of the missing transverse energy where such a signal is often looked for, but also to a more prominent tail in the out-of-plane momentum distribution ($\ensuremath{{p_\mathrm{out}}}$), see~Figure~\ref{fig_shapes_Pout_denser_1B1S}.
The spin of the resonance decaying to a pair of top quarks can be determined from the $\cos\theta^*$ distribution, as provided in~Figure~\ref{fig_shapes_costhetastar_1B1S}, as it leads to a slightly flatter spectrum for the scalar $y_0$ case but a more peaked shape for the case of a $Z'$ production.
Similar observations were found also in the 2B0S and 0B2S event topologies.
\begin{figure}[p]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{figures/shapes/shapes_1B1S_ParticleDiTopMass_denser_1B1S_DetectorDiTopMass_denser_logy_norm.pdf}\\
\end{tabular}
\caption{Shape comparison for the invariant mass of the $\ensuremath{{t\bar{t}}}$ pair in the boosted-semiboosted topology. Left: particle level, right: Delphes ATLAS detector level.}
\label{fig_shapes_mtt_1B1S}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{figures/shapes/shapes_1B1S_ParticleDiTopPout_denser_1B1S_DetectorDiTopPout_denser_logy_norm.pdf}\\
\end{tabular}
\caption{Shape comparison for the out-of-plane momentum $p_\mathrm{out}$ in the boosted-semiboosted topology. Left: particle level, right: Delphes ATLAS detector level.}
\label{fig_shapes_Pout_denser_1B1S}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{figures/shapes/shapes_1B1S_ParticleDiTopCosThetaStar_1B1S_DetectorDiTopCosThetaStar_liny_norm.pdf}\\
\end{tabular}
\caption{Shape comparison for the $\cos\theta^*$ in the boosted-semiboosted topology. Left: particle level, right: Delphes ATLAS detector level.}
\label{fig_shapes_costhetastar_1B1S}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{figures/shapes/shapes_1B1S_ParticleTopPt_denser_1B1S_DetectorTopPt_denser_logy_norm.pdf}\\
\end{tabular}
\caption{Shape comparison for the $\ensuremath{{p_\mathrm{T}}}$ of hadronic top quark candidate in the boosted-semiboosted topology. Left: particle level, right: Delphes ATLAS detector level.}
\label{fig_shapes_pTt_1B1S}
\end{center}
\end{figure}
\section{Samples}
Using the \textsc{MadGraph5}\xspace{} version {\tt 2.6.4} simulation toolkit~\cite{Alwall:2014hca}, proton-proton collision events at $\sqrt{s} = 14$ TeV were generated for the SM process $pp \rightarrow \ensuremath{{t\bar{t}}}{}$ and for the resonant $s$-channel \ensuremath{{t\bar{t}}}{} producing via an additional narrow-width (sub-GeV) vector boson $Z'$ as $pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}{}$ (using the model \cite{FeynModelZprime,Christensen:2008py,Wells:2008xg}). These samples were generated in the all-hadronic $\ensuremath{{t\bar{t}}}$ decay channel and at next-to-leading order (NLO) in QCD in production, using the MLM matching~\cite{Hoche:2006ph}, \emph{i.e.} with additional processes with extra light-flavoured jets produced in the matrix element, matched and resolved for the phase-space overlap of jets generated by the parton shower using \textsc{MadGraph5}\xspace{} defaults settings. The parton shower and hadronization were simulated using \textsc{Pythia}\xspace{}8~\cite{Sjostrand:2014zea}. Masses of the hypothetical $Z'$ particle, serving effectively as a source of semi-boosted and boosted top quarks, were selected as \GeV{750, 800, 900, 1000, 1250 and 1500}. In addition, also a model with a scalar particle decaying to a pair of top quarks $y_0 \rightarrow \ensuremath{{t\bar{t}}}{}$ was adopted~\cite{Christensen:2008py} at the leading-order (LO) in the $\ensuremath{{t\bar{t}}}$ production with the gluon-gluon fusion loop (more details in~\cite{Mattelaer:2015haa,Backovic:2015soa,Neubert:2015fka,Das:2016pbk,Kraml:2017atm,Albert:2017onk,Arina:2017sng,Afik:2018rxl}), with inclusive $\ensuremath{{t\bar{t}}}{}$ decays, selecting the all-hadronic channel later in the analysis. As the last BSM process considered, the possibility of generating a pair of invisible dark matter (DM) fermions $\chi_D$ in association with a $\ensuremath{{t\bar{t}}}{}$ pair using the scalar $y_0$ particle as a~mediator was also studied. For the scalar model, the probed masses were set to $m_{y_0} = 1000\,$GeV with its natural width chosen as $\Gamma_{y_0} =\,10$, $100$ and $300\,$GeV, and for the associated production of the DM particles masses $m_{y_0} = 1000\,$GeV, and $m_{\chi_D} = 10$, $100$ and $300\,$GeV, all with $\Gamma_{y_0} = 10\,$GeV were used. The mass of the top quark was set to \GeV{173} (\textsc{MadGraph5}\xspace{} default). A selection of representative \textsc{MadGraph5}\xspace{} processes are depicted as Feynman diagrams in~Figure~\ref{fig_feynman} while the cross-sections and numbers of generated events for the considered processes are listed in~Table~\ref{tab:xsects}. As typical background processes, the production of hadronically decaying $W^\pm$ bosons in association with a $b\bar{b}$ pair, accompanied by up to two additional light-flavoured jets produced at the matrix element level, was also simulated using the MLM matching with the \textsc{Pythia}\xspace{}8 parton shower using \textsc{MadGraph5}\xspace{} default settings and referred to as $Wbbjj$, and also the production of two $W$ bosons decaying hadronically, accompanied by a $b\bar{b}$ pair at LO, referred to as $WWbb$.
\begin{table}[!]
{\footnotesize
\begin{tabular}{lllrr}
\hline\hline
Sample & Order & Parameters, comment & Events & Cross-section \\ \hline
$pp \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_t = 173\,$GeV, matched & 5,283,343 & 309 pb \\
$60\,\mathrm{GeV} < p_\mathrm{T}^{j1,j2}$ & & control sample & & \\ \hline
$pp \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_t = 173\,$GeV, matched & 1,841,859 & 7.0 pb \\
$200\,\mathrm{GeV} < p_\mathrm{T}^{j1,j2}$ & & & & \\ \hline
$pp \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_t = 173\,$GeV, matched & 1,896,629 & 271 pb \\
$60\,\mathrm{GeV} < p_\mathrm{T}^{j1,j2} < 200\,\mathrm{GeV}$ & & & & \\ \hline
$pp \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_t = 173\,$GeV, matched & 1,544,855 & 30 pb \\
$200\,\mathrm{GeV} < p_\mathrm{T}^{j1}$ & & & & \\
$60\,\mathrm{GeV} < p_\mathrm{T}^{j2} < 200\,\mathrm{GeV}$ & & & & \\ \hline
$pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_{Z'} = 750\,$GeV, matched & 293,402 & 0.403 fb \\
$pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_{Z'} = 800\,$GeV, matched & 289,892 & 0.350 fb \\
$pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_{Z'} = 900\,$GeV, matched & 284,500 & 0.267 fb \\
$pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_{Z'} = 1000\,$GeV, matched & 279,442 & 0.206 fb \\
$pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_{Z'} = 1250\,$GeV, matched & 215,290 & 0.113 fb \\
$pp \rightarrow Z' \rightarrow \ensuremath{{t\bar{t}}}(j)\rightarrow$ hadrons & NLO & $m_{Z'} = 1500\,$GeV, matched & 261,620 & 0.0660 fb\\ \hline
$pp \rightarrow y_0 \rightarrow \ensuremath{{t\bar{t}}}$ incl. & loop SM & $m_{y_0} = 1000\,$GeV, $\Gamma_{y_0} = 10\,$GeV & 1,000,000 & 305 fb \\
$pp \rightarrow y_0 \rightarrow \ensuremath{{t\bar{t}}}$ incl. & loop SM & $m_{y_0} = 1000\,$GeV, $\Gamma_{y_0} = 100\,$GeV & 500,000 & 31.1 fb \\
$pp \rightarrow y_0 \rightarrow \ensuremath{{t\bar{t}}}$ incl. & loop SM & $m_{y_0} = 1000\,$GeV, $\Gamma_{y_0} = 300\,$GeV & 500,000 & 10.0 fb \\ \hline
$pp \rightarrow y_0 \,\ensuremath{{t\bar{t}}} \rightarrow \chi_D \bar{\chi}_D \ensuremath{{t\bar{t}}}\rightarrow$ hadr. & LO & $m_{y_0} = 1\,$TeV, $\Gamma_{y_0} = 10\,$GeV & 500,000 & 2.44 fb \\
& & $m_{\chi_D} = 10\,$GeV & & \\
$pp \rightarrow y_0 \,\ensuremath{{t\bar{t}}} \rightarrow \chi_D \bar{\chi}_D \ensuremath{{t\bar{t}}}\rightarrow$ hadr. & LO & $m_{y_0} = 1\,$TeV, $\Gamma_{y_0} = 10\,$GeV & 1,000,000& 2.29 fb \\
& & $m_{\chi_D} = 100\,$GeV & & \\
$pp \rightarrow y_0 \,\ensuremath{{t\bar{t}}} \rightarrow \chi_D \bar{\chi}_D \ensuremath{{t\bar{t}}}\rightarrow$ hadr. & LO & $m_{y_0} = 1\,$TeV, $\Gamma_{y_0} = 10\,$GeV & 500,000 & 1.24 fb \\
& & $m_{\chi_D} = 300\,$GeV & & \\ \hline
$Wbbjj \rightarrow $ hadr., $p_\mathrm{T}^{j1,j2} > 60\,$GeV & LO & matched 0--2 additional jets & 617,952 & 144 pb \\
$WWbb \rightarrow $ hadr., $p_\mathrm{T}^{j1,j2} > 60\,$GeV & LO & & 2,000,000 & 126 pb \\
\hline\hline
\end{tabular}
}
\caption{Cross sections and numbers of events for the samples used, generated with \textsc{MadGraph5}\xspace{}+\textsc{Pythia}\xspace{}8, with cuts on the transverse momentum of the leading and sub-leading jets indicated as $j1$ and $j2$, respectively.}
\label{tab:xsects}
\end{table}
The multijet background samples are not simulated as in practice they are determined by data-driven techniques in control regions in real data based on vetoing top-tagged jets. In a similar way, this background could be determined for the semi-boosted topologies detailed in Section~\ref{sec:Selection} by considering also the $W$-tag veto regions.
We do not develop novel data-driven techniques for the multijet background estimation but rather study new physics signals appearance and its shape properties above flat enough SM backgrounds from $\ensuremath{{t\bar{t}}}$ or hadronic $W$ processes.
\begin{figure}[!h]
\begin{center}
\begin{tabular}{ccc}
\includegraphics[width=0.99\textwidth]{figures/feynman/tt_SM_cropped.pdf} \\
\includegraphics[width=0.99\textwidth]{figures/feynman/tt_BSM_cropped.pdf} \\
\end{tabular}
\caption{Dominant leading-order Feynman diagrams for SM (top) and selected BSM (bottom) processes of the $\ensuremath{{t\bar{t}}}{}$ pair production. Created using the {\sl FeynMf} package \cite{Ohl:1995kr}.}
\label{fig_feynman}
\end{center}
\end{figure}
\section{Event selection}
\label{sec:Selection}
Two classes of boosted large-$R$-jets are probed for appearance in events,
either top-jets, \emph{i.e.} jets with a mass and substructure consistent with
the hypothesis that the top quark decay products were reconstructed
in such a large-$R$ jet, or $W$-jets, \emph{i.e.} those consistent with the hadronic decay of a $W$ boson. We denote the top-like large-$R$ jets as
boosted (B), and the $W$-like jets as semiboosted (S). Based on their
number in the event, we define in total three topologies of the
all-hadronic $\ensuremath{{t\bar{t}}}{}$ final state denoted as the boosted-boosted
(2B0S), boosted-semiboosted (1B1S) or semiboosted-semiboosted (0B2S),
depicted in~Figure~\ref{fig_cartoons} or as \textsc{Delphes}\xspace{} event display
examples in~Figure~\ref{fig_evt_displ_zp1000_3}.
We do not pursue the resolved topology with dedicated techniques needed to study the combinatorics of $\geq 6$ small-$R$ jets.
The fraction of the considered topologies as a function of the generated mass of the $Z'$
vector boson are shown in~Figure~\ref{fig_topos_graph_migra}, together
with the migration rate between the topologies at the particle
and detector levels, in order to check the stability of the topologies
concept at both levels. The migration is observed to be at the level of $10$\%
in terms of the fraction of events migrating to different
topologies. One can see that while the fractions of semiboosted
topologies (1B1S and 0B2S) decrease with the mass of the resonance, they
still constitute non-negligible fractions of 10--20\% at medium masses
of $Z'$, and about 30\% for the SM $\ensuremath{{t\bar{t}}}{}$ sample.
\noindent The same event selection applied at the particle and detector levels can be summarised as follows:
\begin{itemize}
\item require at least two large-$R$ jets with $\ensuremath{{p_\mathrm{T}}} > 80\,$GeV and $|\eta| < 2.0$;
\item preselect small-$R$ jets with $\ensuremath{{p_\mathrm{T}}} > 25\,$GeV and $|\eta| < 2.5$;
\item require at least two small-$R$ jets tagged as $b$-jets;
\item reject events with an isolated high-$\ensuremath{{p_\mathrm{T}}}$ lepton above $25\,$GeV, in order to remove events from the $pp \rightarrow y_0 \rightarrow \ensuremath{{t\bar{t}}}$ for which, due to the triangle loop in the corresponding diagrams, the \ensuremath{{t\bar{t}}}{} decay mode could not be specified at the generator level.
\item Finally, attempt, in this order, the boosted-boosted (2B0S), boosted-semiboosted (1B1S) or semiboosted-semiboosted (0B2S) topologies reconstruction of the $\ensuremath{{t\bar{t}}}{}$ final state by requiring the corresponding numbers of top-tagged and $W$-tagged jets.
\end{itemize}
\noindent An example of the resulting detector-level yields for the $Z'$ model in the 1B1S topology is listed in~Table~\ref{tab:ex:yields_zp}.
\begin{table}[!h]
\input{yields/yields_zp_1000GeV_1B1S}
\caption{Unweighted and cross-section weighted event yields for the $\ensuremath{{t\bar{t}}}$ and non-$\ensuremath{{t\bar{t}}}$ background samples and the vector $Z'$ model with $m_{Z'} = 1000\,$GeV.}
\label{tab:ex:yields_zp}
\end{table}
\subsection{Signal significance}
The 1D significance of a signal over the total $\ensuremath{{t\bar{t}}}$ and non-$\ensuremath{{t\bar{t}}}$ background is evaluated at the detector level for scalar and vector resonant production of the \ensuremath{{t\bar{t}}}{} pair as well as for the model of paired DM and $\ensuremath{{t\bar{t}}}{}$ production as $N_\mathrm{sig} / \sqrt{\mathrm{Var}_\mathrm{\,bg} + \mathrm{Var}_\mathrm{\,sig}}$.
In order to quantify the signal excess and and find the most sensitive bin area , the 1D \textsc{BumpHunter}\xspace{} algorithm~\cite{Choudalakis:2011qn} is employed, which finds the 1D window of a predefined size in which the probability of the data compatibility with the background-only hypothesis ($p$-value, $p_\mathrm{val}$) is the smallest, constructing a tests statistics $t \equiv - \log p_\mathrm{val}^\mathrm{min}$ which grows with the discrepancy. In detail, the $p$-value for the given area can be computed using the normalized incomplete gamma function as $\Gamma(d,b)$ for the case when the number of data events $d$ is larger than the expected number of background events $b$, or as $1 - \Gamma(d+1,b)$ otherwise, with
\begin{eqnarray}
\Gamma(x,y) \equiv \frac{1}{\Gamma(x)} \int\limits_{0}^{y} \zeta^{x-1} \mathrm{e}^{-\zeta} \,\mathrm{d}\zeta\,,\quad \Gamma(x) \equiv \int\limits_{0}^{\infty} \zeta^{x-1} \mathrm{e}^{-\zeta} \,\mathrm{d}\zeta \,.
\end{eqnarray}
The function appears as a useful replacement for the sum of Poisson weight factors determining the probability of observing $d$ events or more as the sum of Poisson terms for a distribution with mean $b$ as \footnote{There seems to be a typesetting mistake in Eq.~18 in~\cite{Choudalakis:2011qn} which should read $\sum\limits_{n=d}^{\infty} \frac{b^n}{n!}\mathrm{e}^{-b}$.}
\begin{eqnarray}
\sum\limits_{k=d}^{\infty} \frac{b^k}{k!}\mathrm{e}^{-b} = \Gamma(d,b) \,.
\end{eqnarray}
In our adaptation of the 1D \textsc{BumpHunter}\xspace{} algorithm, the following constrains on the 1D area of interest are imposed, with the algorithm details outlined as follows:
\begin{itemize}
\item The initial 1D window width is 2 bins, starting with the window placed at the left of the spectrum.
\item Compute the \textsc{BumpHunter}\xspace{} test statistics~$t$.
\item Add the left- or right-neighbouring bin to the interval based on larger resulting~$t$.
\item Continue as long as the number of bins $N$ of the interval under test in 1D fulfills $ N \leq \frac23 \, n^\mathrm{bins}$ where $ n^\mathrm{bins}$ is the numbers of non-empty bins and as long as $t$ grows.
\item Start again with shifted search window by one bin from left to right.
\item Try all such resulting 1D intervals, return the one with the best BH score.
\end{itemize}
We propose a natural 2D extension of the \textsc{BumpHunter}\xspace{} algorithm following these considerations:
\begin{itemize}
\item The initial 2D window width is $2\times 2$ bins, placing the window at the bottom left of the spectrum.
\item Compute the \textsc{BumpHunter}\xspace{} test statistics~$t$.
\item Add a directly neighbouring bin to the current 2D area (excluding diagonal neighbours) which leads to the largest positive change in~$t$.
\item Continue as long as $t$ grows and while the number of bins $N$ of the 2D area under test its widths in $x$ and $y$ ($w_x$ and $w_y$) fulfill $N < n^\mathrm{bins} \, \land \, w_x \leq \frac23 \, n^\mathrm{bins}_x \, \land \, w_y \leq \frac23 \, n^\mathrm{bins}_y$, where $n^\mathrm{bins}$ is the numbers of non-empty bins and $ n^\mathrm{bins}_{x,y}$ are the number of non-empty bins of the 1D histogram projections along the $x,y$ axes.
\item Start again shifting the search window by one bin gradually in $x$ and $y$ directions from left to right.
\item Try all such 2D areas, return the one with the best BH score.
\end{itemize}
For presentational purposes we further define the BH score as a logarithm of its test statistics, noting that a small shift in the score means a large shift in the actual \textsc{BumpHunter}\xspace{} test statistics.
We present the 1D and 2D significances of signal over the expected background in corresponding histograms.
The resulting histograms of the 1D and 2D \textsc{BumpHunter}\xspace{} (BH) scores $\log t$ as well as the 2D histograms themselves for selected variables are shown in~Figures~\ref{fig_bestVars_y0}--\ref{fig_bestVars_xdxd}, also with indicated 1D and 2D best BH areas.
These are obtained by repeating the \textsc{BumpHunter}\xspace{} procedure over the 100 statistically pseudo-independent replicas of the 1D and 2D spectra.
The best discriminating variables in terms of the \textsc{BumpHunter}\xspace{} score for the three benchmark models are shown in~Tables~\ref{tab:bestBHscores_y0}--\ref{tab:bestBHscores_xdxd}. It is observed that a single 1D variable never scores better than 2D variables.
Histograms of the score over all the variables are shown as 1D plots in~Figures~\ref{fig_bestVars_y0_mean}--\ref{fig_bestVars_xdxd_mean}. In the 2D plots therein, the correlation coefficient for 253 studied 2D spectra built from 23 kinematic variables plotted as function of the best BH score, one can observe denser population of points within approximately $|\bar{\rho}| \lesssim 0.2$, but with a large spread. The $\bar{\rho}$ is the correlation coefficient averaged over 100 pseudo-experiments, computed using the $\ensuremath{{t\bar{t}}}$ control sample. One can conclude the small correlation is not the only property of two variables to possibly provide a~stronger 2D signal significance than the individual 1D variables. Best variables exhibit a mixture of behaviour of those with small as well as large correlations, as well as negative correlations, sometimes negligible for SM $\ensuremath{{t\bar{t}}}$ but largely negative for signal, but in other cases also smaller positive correlation for signal.
Interestingly, for the $y_0$ and $Z'$ models the $\ensuremath{m^{t\bar{t}}}$ variable is not the best standalone variable but is often preferred when paired with another variable, \emph{e.g.} with $\ensuremath{{H_\mathrm{T}}}$, $\ensuremath{E^\mathrm{miss}_\mathrm{T}}$, but also with jets sphericity or relative variables like $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}} / \ensuremath{m^{t\bar{t}}}$. Interestingly, the top quark $\ensuremath{{p_\mathrm{T}}}$ does not seem to help finding signal much, which may be regarded as a good news to worries of tuning new physics into physics models using this variable, while the transverse momentum of the $\ensuremath{{t\bar{t}}}$ system seems to be very sensitive (and harder to model). Other powerful variables turn out to be the sum of the large-$R$ jet masses and the invariant mass of the 4-vector sum of large-$R$ jets $m^\mathrm{vis}_{\sum J}$, which has correlation of about $0.60$ to $\ensuremath{m^{t\bar{t}}}$. Also the top quark $\ensuremath{{p_\mathrm{T}}}$ ratio variable $R^{t1,t2}$ vs. $\ensuremath{m^{t\bar{t}}}$ appears in the tables, as well as sphericity and aplanarity variables.
For the DM signal case, the best variable turns out to be $\ensuremath{E^\mathrm{miss}_\mathrm{T}}$ in combination, as 2D variable, with $\ensuremath{{H_\mathrm{T}}} + \ensuremath{E^\mathrm{miss}_\mathrm{T}}$. Although both variables are highly correlated ($\sim 1$ for the $\ensuremath{{t\bar{t}}}$ and $\sim 0.9$ for the signal samples), the 2D histogram actually helps to separate the signal by moving it off the diagonal, leading to a large significance and 2D BH score. The signal scale factor for this model was chosen so that the 2D \textsc{BumpHunter}\xspace{} significance for this variable is reasonable, which effectively reduced significances of all other variables to much smaller values.
It was checked that besides the missing transverse energy, the best performing variables for the DM signal are similar to those for the $Z'$ model, \emph{e.g.} $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}} / \ensuremath{m^{t\bar{t}}}$, $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}} / \sqrt{p_\mathrm{T}^{t1} \, p_\mathrm{T}^{t2}}$, $H_\mathrm{T}^j$, but also the $\ensuremath{{p_\mathrm{out}}}$ variable, which has a large correlation to $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}}$ ($\rho=0.86$) but did not score prominently for the $y_0$ or $Z'$ models. One can again think of this variable as a useful tuning rather than a search tool, or a proxy to $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}}$.
\input{BestVars}
\input{tables/bestBHscores}
\begin{figure}[p]
\begin{center}
\includegraphics[width=0.99\textwidth]{figures/bhscores/meanBHscore_y0_1000GeV.pdf} \\
\caption{Mean BH scores over the variables for the scalar $y_0$ model with $m_{y_0} = 1000\,$GeV (top left), the vertical dashed lines indicate more precisely the BH score for the best variable. In the 2D plots with the BH score on the $x$-axis and the correlation coefficient between two variables on the $y$-axis, the circle markers show the correlation vs. the 2D BH results while the crosses represent the 1D BH scores plotted at $\bar{\rho} = 1$. Green: the 2B0S, blue: 1B1S and pink: the 0B2S topologies. The insets shows the correlations distribution and the dashed horizontal guiding lines are drawn along $ \pm \sigma_{\bar{\rho}}$, the correlations standard deviation over the studied 2D variables.}
\label{fig_bestVars_y0_mean}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=0.99\textwidth]{figures/bhscores/meanBHscore_zp_1000GeV.pdf} \\
\caption{Mean BH scores over the variables for the vector $Z'$ model with $m_{Z'} = 1000\,$GeV (top left), the vertical dashed lines indicate more precisely the BH score for the best variable. In the 2D plots with the BH score on the $x$-axis and the correlation coefficient between two variables on the $y$-axis, the circle markers show the correlation vs. the 2D BH results while the crosses represent the 1D BH scores plotted at $\bar{\rho} = 1$. Green: the 2B0S, blue: 1B1S and pink: the 0B2S topologies. The insets shows the correlations distribution and the dashed horizontal guiding lines are drawn along $ \pm \sigma_{\bar{\rho}}$, the correlations standard deviation over the studied 2D variables.}
\label{fig_bestVars_zp_mean}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=0.99\textwidth]{figures/bhscores/meanBHscore_xdxdtt_100GeV.pdf}
\caption{Mean BH scores over the variables for the associated production of $\ensuremath{{t\bar{t}}}{}$ with a pair of DM particles with $m_{\chi_D} = 100\,$GeV (top left), the vertical dashed lines indicate more precisely the BH score for the best variable. In the 2D plots with the BH score on the $x$-axis and the correlation coefficient between two variables on the $y$-axis, the circle markers show the correlation vs. the 2D BH results while the crosses represent the 1D BH scores plotted at $\bar{\rho} = 1$. Green: the 2B0S, blue: 1B1S and pink: the 0B2S topologies. The insets shows the correlations distribution and the dashed horizontal guiding lines are drawn along $ \pm \sigma_{\bar{\rho}}$, the correlations standard deviation over the studied 2D variables.}
\label{fig_bestVars_xdxd_mean}
\end{center}
\end{figure}
\section{Variables}
\subsection{Kinematic variables}
Typical variables studied in the $\ensuremath{{t\bar{t}}}{}$ system for SM precision differential cross-section measurements are the
transverse momenta of the two top quarks (both used in this study, with event weights of $0.5$), rapidity of the top quarks ($y^{t}$) and of the \ensuremath{{t\bar{t}}}{} system ($y^{\ensuremath{{t\bar{t}}}{}}$), and the mass ($m^{\ensuremath{{t\bar{t}}}{}}$) and transverse momentum ($\ensuremath{{p^{\ttbar{}}_\mathrm{T}}}$) of the \ensuremath{{t\bar{t}}}{} system; their 2D and 3D extensions and variables related to additional jet activity in the event. Further variables are composed from the two top quarks momenta, like the $\cos\theta^*$ (angle between a~top quark and the $z$ axis in a~frame where the $\ensuremath{{t\bar{t}}}{}$ system has zero momentum along the $z$ axis), the laboratory opening angle between the two top quarks ($\delta_{\ensuremath{{t\bar{t}}}{}}$) and the $\Delta\phi$ between the two top quarks in the transverse plane ($\ensuremath{\Delta\phi^{t\bar{t}}}$).
Other studied observables are the out-of-plane momentum $\ensuremath{{p_\mathrm{out}}}$ based on the direction of one of the top quarks defining a~plane together with the beam ($z$ axis) direction $\hat{z}$, to which the momentum of the other top quark is projected; and the $y_{\mathrm{boost}}^{\ttbar{}}$ and $\chi^{t\bar{t}}$ variables, defined as
\begin{eqnarray}
\ensuremath{{p_\mathrm{out}}} &\equiv& \vec{p}^{\,t, 2} \cdot \frac{\vec{p}^{\,t1} \times \hat{z}}{|\vec{p}^{\,t,1}\times \hat{z}|} \,, \quad \mathrm{and\,\, 1 \leftrightarrow 2} \\
y_{\mathrm{boost}}^{\ttbar{}} &\equiv& \frac12 \left| y^{t2} + y^{t1} \right| \\
\chi^{t\bar{t}} &\equiv& \exp \left| y^{t2} - y^{t1} \right| \,.
\end{eqnarray}
These are sensitive to final state radiation, the boost of the $\ensuremath{{t\bar{t}}}{}$ system and thus also to parton distribution functions; and to new physics via their sensitivity to the production angle in the central-mass-system. We also study the ratio of the two top quark transverse momenta, $R^{t1,t2}{}$, the ratio of the leading to sub-leading top quark \ensuremath{{p_\mathrm{T}}}{}. One can also use event shape variables (originally studied in QCD jet physics) based on the normalized momentum tensor defined using the three-momenta of objects $\mathcal{O}$ in the event (usually jets) as
\begin{eqnarray}
\mathcal{M}_{ij} \equiv \frac{\sum\limits_\mathcal{O} p^\mathcal{O}_i p^\mathcal{O}_j}{\sum\limits_\mathcal{O} \left(\vec{p}^\mathcal{O}\right)^2}
\end{eqnarray}
where $i,j$ run over the three space indices. The matrix has eigenvalues $\lambda_1 \leq \lambda_2 \leq \lambda_3$
used to define aplanarity $\mathcal{A} \equiv \frac32 \lambda_1$ and sphericity $\mathcal{S} \equiv \frac32 (\lambda_1 + \lambda_2)$.
Corresponding aplanarity and sphericity for the small-$R$ jets were studied, labelled as $\mathcal{A}^j$ and~$\mathcal{S}^j$ in plots.
Other global variables are also studied, namely
\begin{itemize}
\item the sum of \ensuremath{{p_\mathrm{T}}}{} of small-$R$ jets $H_\mathrm{T}^{j}$;
\item missing transverse energy $E_\mathrm{T}^\mathrm{miss}$; and the sum of $H_\mathrm{T}^{j} + E_\mathrm{T}^\mathrm{miss}$;
\item the sum of masses of large-$R$ jets $\sum\limits_J m_J$, labelled also as Sum $m_J$ in plots;
\item and the invariant masses of the 4-vector sum of large-$R$ jets 4-momenta $m^\mathrm{vis}_{\sum J}$, labelled also as $m^\mathrm{vis}_{\mathrm{Sum}J}$ in plots;
\end{itemize}
motivated by the fact that invisible BSM particles carrying significant transverse momentum can exhibit tails in $\ensuremath{E^\mathrm{miss}_\mathrm{T}}$, $\ensuremath{{p_\mathrm{out}}}$, or contribute to higher jet activity in the event due to larger energy scale of events with production of high-mass particles. We keep the global event shape variables like aplanarity, sphericity; and $H_\mathrm{T}$ only for small-$R$ jets while the jet-mass related quantities for larger-$R$ jets.
\noindent In addition, dimensionless variables with potential of possessing smaller experimental uncertainties in absolute jet energy scale calibration were studied by constructing ratios relative to $\ensuremath{m^{t\bar{t}}}$ or to the geometric mean of the two top quark transverse momenta $\sqrt{p_\mathrm{T}^{t1} \, p_\mathrm{T}^{t2}}$.
\subsection{Correlations between variables}
Besides the obvious anti-correlations of variables to the relative quantities w.r.t. to such variables, and the negative correlation of $\ensuremath{\Delta\phi^{t\bar{t}}}$ to $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}}$ and $\ensuremath{{p_\mathrm{out}}}$ by construction,
one can make several non-trivial observations. Out of all the variables especially $\ensuremath{{p_\mathrm{out}}}$ (or an absolute value of it) is a very useful one as it has only a small correlation to $\ensuremath{m^{t\bar{t}}}$ (we observe 0.07 for the $\ensuremath{{t\bar{t}}}$ control sample, see also Figure~\ref{fig_corr_vars}) while it exhibits a~large shape change especially in tails for models with additional DM particles produced, see~Figure~\ref{fig_shapes_Pout_denser_1B1S}. It has also been used for MC generators tuning as it is sensitive to the initial and final state radiation~\cite{ATL-PHYS-PUB-2016-020} while caution should be taken in order not to tune to possible new physics effects. We keep variables even with large absolute correlations for detailed study of their performance in terms of a signal significance.
Both $\ensuremath{{p^{t}_\mathrm{T}}}$ and $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}}$ have large correlation to variables quantifying the jet $\ensuremath{{p_\mathrm{T}}}$ activity in the event like $\ensuremath{{H^{j}_\mathrm{T}}}$ and the total large-$R$ jets visible mass $m^\mathrm{vis}_{\sum J}$.
There are also large negative correlations of sphericity and $y_{\mathrm{boost}}^{\ttbar{}}$ to $\ensuremath{m^{t\bar{t}}} / \sqrt{p_\mathrm{T}^{t1} \, p_\mathrm{T}^{t2}}$, of $R^{t1,t2}$ to $\ensuremath{{p^{\ttbar{}}_\mathrm{T}}}$ and $|\ensuremath{{p_\mathrm{out}}}|$; and of $|\cos\theta^*|$ and $\chi^{t\bar{t}}$ to $\ensuremath{{p^{t}_\mathrm{T}}} / \ensuremath{m^{t\bar{t}}}$.
\begin{figure}[p]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.75\textwidth]{figures/corrs/corrs_Detector_1B1S.pdf} \\
\includegraphics[width=1.0\textwidth]{figures/corrs/corrsh_Detector_1B1S.pdf} \\
\end{tabular}
\caption{Example of correlations between kinematic variables at the detector level in the boosted-semiboosted (1B1S) topology evaluated on the $\ensuremath{{t\bar{t}}}{}$ control sample.}
\label{fig_corr_vars}
\end{center}
\end{figure}
|
\section{Introduction}
Each set operator determines the partition of sets to equivalence classes
with equal value of the operator. Let us have some set operator $\alpha $.
Following \cite{Greedoids} we call two sets $X,Y$ \textit{cospanning} if
\alpha (X)=\alpha (Y)$. Thus each set operator generates the cospanning
equivalence relation on sets. Our goal is to investigate cospanning
relations on violator spaces. These spaces were introduced in order to
develop a combinatorial framework encompassing linear programming and other
geometric optimization problems \cite{VS}. Violator spaces are defined by
violator operators, which generalize closure operators \cite{KempnerLevit}.
We also pay special attention to violator spaces with unique bases. In
Section 2, we introduce co-violator spaces based on contracting operators
known also as choice functions. In Section 3, we characterize the cospanning
relation with regards to violator spaces and describe the equivalence
classes of the relation for violator and co-violator spaces. Cospanning
characterizations allow us to obtain some new properties of violator
operators, co-violator operators and their interconnections. In particular,
we show that uniquely generated violator spaces enjoy so-called Krein-Milman
properties.
\subsection{ Violator spaces}
Violator spaces are arisen as a generalization of Linear Programming
problems. LP-type problems have been introduced and analyzed by Matou\v{s
ek, Sharir and Welzl \cite{MSW, SW} as a combinatorial framework that
encompasses linear programming and other geometric optimization problems.
Further, Matou\v{s}ek et al. \cite{VS} define a simpler framework: violator
spaces, which constitute a proper generalization of LP-type problems.
Originally, violator spaces were defined for a set of constraints $E$, where
each subset of constraints $G\subseteq E$ was associated with $\nu (G)$ -
the set of all constraints violating $G$.
The classic example of an LP-type problem is the problem of computing the
smallest enclosing ball of a finite set of points in
\mathbb{R}
^{d}$. Here $E$ is a set of points in
\mathbb{R}
^{d}$, and the violated constraints of some subset of the points $G$ are
exactly the points lying outside the smallest enclosing ball of $G$.
\begin{definition}
\cite{VS} A \textit{violator space} is a pair $(E,\nu)$, where $E$ is a
finite set and $\nu$ is a mapping $2^{E}\rightarrow2^{E}$ such that for all
subsets $X,Y\subseteq E$ the following properties are satisfied:
\textbf{V11}: $X \cap\nu(X) = \emptyset$ (consistency),
\textbf{V22}: $(X\subseteq Y$ and $Y \cap\nu(X) = \emptyset) \Rightarrow
\nu(X)=\nu(Y)$ (locality).
\end{definition}
Let $(E,\nu)$ be a violator space. Define $\varphi(X)=E-\nu(X)$. In what
follows, if $(E,\nu)$ is a violator space and $\varphi(X)=E-\nu(X)$, then
(E, \varphi)$ will be called a violator space as well.
\begin{definition}
(\cite{KempnerLevit}) A \textit{violator space} is a pair $(E,\varphi)$,
where $E$ is a finite set and $\varphi$ is an operator $2^{E
\rightarrow2^{E} $ such that for all subsets $X,Y\subseteq E$ the following
properties are satisfied:
\textbf{V1}: $X\subseteq\varphi(X)$ (extensivity),
\textbf{V2}: $(X\subseteq Y\subseteq\varphi(X))\Rightarrow\varphi
(X)=\varphi(Y)$ (self-convexity).
\end{definition}
Each violator operator $\varphi$ is idempotent. Indeed, extensivity implies
X\subseteq\varphi(X)\subseteq\varphi(X)$. Then, by self-convexity, we
conclude with $\varphi(\varphi(X))=\varphi(X)$.
\begin{lemma}
(\cite{KempnerLevit}) \label{un} Let $(E,\varphi)$ be a violator space. Then
\begin{equation}
\varphi(X)=\varphi(Y)\Rightarrow\varphi(X\cup Y)=\varphi(X)=\varphi(Y)
\label{Union}
\end{equation}
and
\begin{equation}
(X \subseteq Y \subseteq Z) \wedge (\varphi(X)=\varphi(Z)) \Rightarrow
\varphi(X)=\varphi(Y)=\varphi(Z) \label{Convexity}
\end{equation}
for every $X,Y,Z\subseteq E$.
\end{lemma}
Since the second property deals with all sets lying between two given sets,
following \cite{Monjardet} we call the property \textit{convexity}.
\subsection{Uniquely generated violator spaces}
Let $(E,\alpha)$ be an arbitrary space with the operator
\alpha:2^{E}\rightarrow2^{E}$. $B\subseteq E$ is a \textit{generator} of
X\subseteq E$ if $\alpha(B)=\alpha(X)$. For $X\subseteq E$, a \textit{basis}
(minimal generator) of $X$ is a inclusion-minimal set $B\subseteq E$ (not
necessarily included in $X$) with $\alpha(B)=\alpha(X)$. A space $(E,\alpha)$
is \textit{uniquely generated} if every set $X\subseteq E$ has a unique
basis.
\begin{proposition}
\cite{KempnerLevit} \label{UQP} A violator space $(E,\varphi)$ is uniquely
generated if and only if for every $X,Y\subseteq E$
\begin{equation}
\varphi(X)=\varphi(Y)\Rightarrow\varphi(X\cap Y)=\varphi(X)=\varphi(Y)
\label{UQ}
\end{equation}
\end{proposition}
We can rewrite the property (\ref{UQ}) as follows: for every set $X\subseteq
E$ of a uniquely generated violator space $(E,\varphi)$, the basis $B$ of $X$
is the intersection of all generators of $X$:
\begin{equation}
B=\bigcap\{Y\subseteq E:\varphi(Y)=\varphi(X)\}. \label{UQI}
\end{equation}
One of the known examples of a not uniquely generated violator space is the
violator space associated with the smallest enclosing ball problem. A basis
of a set of points is a minimal subset with the same enclosing ball. In
particular, all points of the basis are located on the ball's boundary. For
\mathbb{R}
^{2}$ the set $X$ of the four corners of a square has two bases: the two
pairs of diagonally opposite points. Moreover, one of these pairs is a basis
of the second pair. Thus the equality (\ref{UQI}) does not hold.
For each arbitrary space $(E,\alpha)$ with the operator $\alpha:2^{E
\rightarrow2^{E}$, an element $x$ of a subset $X\subseteq E$ is \textit{an
extreme point} of $X$ if $x\notin\alpha(X-x)$. The set of extreme points of
X$ is denoted by $ex(X)$.
\begin{proposition}
\cite{KempnerLevit} \label{exp} Let $(E,\varphi)$ be a violator space. Then
\begin{equation*}
ex(X)=\bigcap\{B\subseteq X:\varphi(B)=\varphi(X)\}.
\end{equation*}
\end{proposition}
\begin{proposition}
\cite{KempnerLevit} \label{exp-vs} Let $(E,\varphi)$ be a violator space.
Then
\begin{equation*}
ex(\varphi(X))\subseteq ex(X).
\end{equation*}
\end{proposition}
\begin{theorem}
\cite{KempnerLevit} \label{th2} Let $(E,\varphi)$ be a violator space. Then
(E,\varphi)$ is uniquely generated if and only if for every set $X\subseteq
E $, $\varphi (X)=\varphi(ex(X))$.
\end{theorem}
\begin{corollary}
\cite{KempnerLevit} \label{cor_ex} Let $(E,\varphi)$ be a uniquely generated
violator space. Then for every $X\subseteq E$ the set $ex(X)$ is the unique
basis of $X$.
\end{corollary}
\section{Co-violator spaces}
Theorem \ref{th2} and Proposition \ref{exp-vs} show that there is some
duality between extensive ($X \subseteq \varphi(X)$) and contracting ($ex(X)
\subseteq X$) operators. To study this connection we introduce a new type of
spaces.
\begin{definition}
A \textit{co-violator space} is a pair $(E,c)$, where $E$ is a finite set
and $c$ is an operator $2^{E}\rightarrow2^{E}$ such that for all subsets
X,Y\subseteq E$ the following properties are satisfied:
\textbf{CV1}: $c(X)\subseteq X$,
\textbf{CV2}: $(c(X)\subseteq Y\subseteq X)\Rightarrow c(X)=c(Y)$.
\end{definition}
Operators satisfying the property \textbf{CV1} are called contracting
operators.
In social sciences, contracting operators are called choice functions,
usually adding a requirement that $c(X)\neq \emptyset $ for every $X\neq
\emptyset $. The property \textbf{CV2} is called the \textit{outcast propert
} or the \textit{Aizerman property} \cite{Monjardet}.
The properties of co-violator spaces correspond to the corresponding
("mirrored") properties of violator spaces. For instance, every co-violator
operator $c$ is idempotent. Indeed, since $c$ is contracting $c(X)\subseteq
c(X)\subseteq X$. Then, \textbf{CV2} implies $c(c(X))=c(X)$.
Lemma \ref{un} is converted to the following.
\begin{lemma}
\label{co-un} Let $(E,c)$ be a co-violator space. Then
\begin{equation}
c(X)=c(Y)\Rightarrow c(X\cap Y)=c(X)=c(Y) \label{Intersection}
\end{equation}
and
\begin{equation}
(X \subseteq Y \subseteq Z) \wedge (c(X)=c(Z)) \Rightarrow c(X)=c(Y)=c(Z)
\label{Co-Convexity}
\end{equation}
for every $X,Y,Z\subseteq E$.
\end{lemma}
\begin{proof}
Prove (\ref{Intersection}). Let $c(X)=c(Y)$. \textbf{CV1} implies that
c(X)\subseteq X$ and $c(Y)=c(X) \subseteq Y$. Then $c(X) \subseteq X\cap Y
\subseteq X$, that gives (by \textbf{CV2}) $c(X\cap Y)=c(X)$.
To prove (\ref{Co-Convexity}) let $(X\subseteq Y\subseteq
Z)\wedge(c(X)=c(Z)) $. \textbf{CV1} yields $c(Z)=c(X)\subseteq X\subseteq Y
. Then outcast property allows us to get $c(Z)\subseteq Y\subseteq
Z\Rightarrow c(Y)=c(X)=c(Z)$.
\end{proof}
It is easy to see that all the properties of violator spaces hold in their
dual interpretation for co-violator spaces. Since a co-violator operator is
a choice function with outcast properties, the connection between these two
types of spaces may result in better understanding of two theories and in
new findings in each of them.
Connections between contracting and extensive operators were studied in many
works, while most of them were dedicated to connections between choice
functions and closure operators \cite{Ando, Danilov, Monjardet}. Naturally,
extreme point operators were considered as choice functions. But, as we will
see in Proposition \ref{outcast_u}, the extreme point operator of a violator
space satisfies the outcast property, and so it forms a co-violator space,
if and only if the violator space is uniquely generated. We also consider
choice functions investigated in \cite{Libkin}. The \textit{interior
operator }(well-known in topology) is dual to a closure operator. Given an
extensive operator $\varphi :2^{E}\rightarrow 2^{E}$, one can get a
contracting operator $c$: $c(X)=E-\varphi (E-X)$ or $\overline{c(X)}=\varphi
(\overline{X})$.
\begin{proposition}
\label{co-operator} $(E,\varphi)$ is a violator space if and only if $(E,c)$
is a co-violator space, where $c(X)=\overline{\varphi(\overline{X})}$.
\end{proposition}
\begin{proof}
It is easy to see that $\varphi $ is an extensive operator if and only if $c$
is a contracting operator. To prove that $c$ satisfies the outcast property
if and only if $\varphi $ is self-convex one has just to pay attention that:
$c(X) \subseteq Y \subseteq X \Leftrightarrow \overline{X} \subseteq
\overline{Y} \subseteq \overline{c(X)} \Leftrightarrow \overline{X}
\subseteq \overline{Y} \subseteq \varphi(\overline{X}) \Rightarrow \varphi
\overline{X})=\varphi(\overline{Y}) \Leftrightarrow c(X)=c(Y)$. The opposite
direction is proved completely analogously.
\end{proof}
\section{Cospanning relations of violator and co-violator spaces}
Let $E=\left\{ {x_{1},x_{2},...,x_{d}}\right\} $. The graph $H(E)$ is
defined as follows. The vertices are the finite subsets of $E$, two vertices
$A$ and $B$ are adjacent if and only if they differ in exactly one element.
Actually, $H(E)$ is \textit{the hypercube} on $E$ of dimension $d$, since
the hypercube is known to be equivalently considered as the graph on the
Boolean space $\{0,1\}^{d}$ in which two vertices form an edge if and only
if they differ in exactly one position.
Let $(E,\varphi)$ be a violator space. The two sets $X$ and $Y$ are \textit
equivalent} (or \textit{cospanning}) if $\varphi(X)=\varphi(Y)$. In what
follows, $\mathcal{P}$ denotes a partition of $H(E)$ ( or $2^{E}$) into
equivalence classes with regard to this relation, and $[A]_{\varphi}:=\{
\subseteq E:\varphi(X)=\varphi(A)\}$.
\begin{remark}
Note, that the cospanning relation associated with a violator operator
\varphi$ coincides with the cospanning relation associated with an original
violator mapping $\nu$.
\end{remark}
The following theorem characterizes cospanning relations in violator spaces.
\begin{theorem}
\label{T_rel} Let $E$ be a finite set and $R \subseteq 2^{E} \times 2^{E}$
be an equivalence relation on $2^{E}$. Then $R$ is the cospanning relation
of a violator space if and only if the following properties hold for every
X,Y,Z \subseteq E$:
\textbf{R1}: if $(X,Y) \in R$, then $(X,X \cup Y) \in R$
\textbf{R2}: if $X \subseteq Y \subseteq Z$ and $(X,Z) \in R$, then $(X,Y)
\in R$.
\end{theorem}
\begin{proof}
Necessity follows immediately from Lemma \ref{un}.
Let us define an operator $\varphi $ and prove that it satisfies extensivity
and self-convexity. Since $R$ is an equivalence relation, it defines a
partition of $2^{E}$. Then, for each $X\subseteq E$ there is only one class
containing $X$. Thus for every set $X$, we define $\varphi (X)$ as a maximal
element in the class $[X]_{R}$. Notice, that the property \textbf{R1}
implies that each equivalence class has a unique maximal element, so the
partition is well-defined. Hence, we obtain that $X\subseteq \varphi (X)$
and $\varphi (\varphi (X))=\varphi (X)$. Then the self-convexity follows
immediately from \textbf{R2}. It is easy to see that the cospanning relation
w.r.t. $\varphi $ coincides with $R$.
\end{proof}
In conclusion, each equivalence class of the cospanning relation of a
violator space is closed under union (\textbf{R1}) and convex (\textbf{R2}).
The following theorem characterizes equivalence classes of co-violator
spaces.
\begin{theorem}
\label{cv_rel} Let $E$ be a finite set and $R \subseteq 2^{E} \times 2^{E}$
be an equivalence relation on $2^{E}$. Then $R$ is the cospanning relation
of a co-violator space if and only if the following properties hold for
every $X,Y,Z \subseteq E$:
\textbf{R3}: if $(X,Y) \in R$, then $(X,X \cap Y) \in R$
\textbf{R2}: if $X \subseteq Y \subseteq Z$ and $(X,Z) \in R$, then $(X,Y)
\in R$.
\end{theorem}
\begin{proof}
Necessity follows immediately from Lemma \ref{co-un}. By analogy with the
proof of Theorem \ref{T_rel} we define $c(X)$ to be a minimal element in the
class $[X]_{R}$. Since each class is closed under intersection (\textbf{R3
), the partition is well-defined. It is easy to see that operator $c$ is
contracting, satisfies the outcast property, and its cospanning relation
coincides with $R$.
\end{proof}
Consider now both a violator operator $\varphi $ and a co-violator operator
c(X)=\overline{\varphi (\overline{X})}$.
\begin{proposition}
\label{co-co} There is a one-to-one correspondence between an equivalence
class $[X]_{\varphi}$ of $X$ of the cospanning relation associated with a
violator operator $\varphi$ and an equivalence class $[\overline{X}]_{c}$
w.r.t. a co-violator operator $c$, i.e., $A \in [X]_{\varphi}$ if and only
if $\overline{A} \in [\overline{X}]_{c}$.
\end{proposition}
\begin{proof}
Indeed, $A \in [X]_{\varphi} \Leftrightarrow \varphi(X)=\varphi(A)
\Leftrightarrow \overline{c(\overline{X})} = \overline{c(\overline{A})}
\Leftrightarrow c(\overline{X})=c(\overline{A}) \Leftrightarrow \overline{A}
\in [\overline{X}]_{c}$.
\end{proof}
A uniquely generated violator space defines a cospanning relation with
additional property \textbf{R3} (see Proposition \ref{UQP}).
All in all, every uniquely generated violator space is a co-violator space
as well. Each equivalence class of the cospanning relation of a uniquely
generated violator space has an unique minimal element and an unique maximal
element. More precisely, for the sets $A\subseteq B\subseteq E$, let us
define the \textit{interval} $[A,B]$ as $\{C\subseteq E:A\subseteq
C\subseteq B\}$. Then each equivalence class of an uniquely generated
violator space is an interval. We call a partition of $H(E)$ into disjoint
intervals a \textit{hypercube partition}. The following Theorem follows
immediately from Theorem \ref{T_rel} and Proposition \ref{UQP}.
\begin{theorem}
(\cite{Clarkson}) (i) If $(E,\varphi)$ is a uniquely generated violator
space, then $\mathcal{P}$ is a hypercube partition of $H(E)$.
(ii) Every hypercube partition is the partition $\mathcal{P}$ of $H(E)$ into
equivalence classes of a uniquely generated violator space.
\end{theorem}
More specifically \cite{KempnerLevit}, $[A]_{\varphi}=[ex(A),\varphi(A)]$
for every set $A \subseteq E$.
Let us consider now a uniquely generated violator space $(E,\varphi )$ and
the operator $ex$. Since each equivalence class $[A]_{\varphi }$ w.r.t.
operator $\varphi $ is an interval $[ex(A),\varphi (A)]$, we can see that
for each $X\in \lbrack ex(A),\varphi (A)]$ not only $\varphi (X)=\varphi (A)
, but $ex(X)=ex(A)$ as well. Since $\mathcal{P}$ is a hypercube partition of
$H(E)$ we conclude with $[X]_{\varphi }=[X]_{ex}$. Thus the cospanning
partition (quotient set) associated with an operator $\varphi $ coincides
with the cospanning partition associated with a contracting operator $ex$.
Since $ex(X)$ is a minimal element of $[X]$ we immediately obtain the
following
\begin{proposition}
If $(E,\varphi)$ is a uniquely generated violator space, then operator $ex$
satisfies the following properties:
\textbf{X1}: $ex(ex(X))=ex(X)$
\textbf{X2}: $ex(X)=ex(Y)\Rightarrow ex(X\cup Y)=ex(X)=ex(Y)$
\textbf{X3}:$(X \subseteq Y \subseteq Z) \wedge (ex(X)=ex(Z)) \Rightarrow
ex(X)=ex(Y)=ex(Z)$
\textbf{X4}: $ex(X)=ex(Y)\Rightarrow ex(X\cap Y)=ex(X)=ex(Y)$
\end{proposition}
If $(E,\varphi )$ is not a uniquely generated violator space, then the
operator $ex$ may or may not satisfy the properties \textbf{X1}-\textbf{X4}.
Consider the two following examples.
\begin{example}
Let $E=\{1,2,3\}$. Define $\varphi(X)=X$ for each $X\subseteq E$ except
\varphi(\{2\})=\varphi(\{3\})=\{2,3\}$ and $\varphi(\{1,2\})=\varphi(\{1,
\})=\{1,2,3\}$. It is easy to check that $(E,\varphi)$ is a violator space
and the operator $ex$ satisfies \textbf{X1},\textbf{X2}, and \textbf{X4},
but while $ex(\{1\})=ex(\{1,2,3\})=\{1\}$, $ex(\{1\}) \neq ex(\{1,2\})$,
i.e., the operator $ex$ is not convex.
\end{example}
\begin{example}
Let $E=\{1,2,3,4,5,6\}$. Define $\varphi (X)=X$ for each $X\subseteq E$
except $\varphi (\{1\})=\{1,2\}$, $\varphi (\{1,2,3\})=\varphi
(\{1,2,4\})=\{1,2,3,4\}$ and $\varphi (\{1,2,5\})=\varphi
(\{1,2,6\})=\{1,2,5,6\}$. It is easy to check that $(E,\varphi )$ is a
violator space. In addition, $ex(\{1,2,3,4\})=ex(\{1,2,5,6\})=\{1,2\}$,
while $ex(\{1,2\})=\{1\}$. Hence, $ex$ is not idempotent (\textbf{X1}) and
does not satisfy \textbf{X4}. Since $ex(\{1,2,3,4,5,6\})=\{1,2,3,4,5,6\}$
the operator $ex$ does not satisfy \textbf{X2} as well, but, compared to the
previous example, $ex$ is convex.
\end{example}
\begin{proposition}
\label{outcast_u} Let $(E,\varphi)$ be a violator space. The following
assertions are equivalent:
(i) $(E,\varphi)$ is uniquely generated
(ii) \textbf{X5}: $(ex(X) \subseteq Y \subseteq X) \Rightarrow ex(X)=ex(Y)$
(the outcast property)
(iii) \textbf{X6}: $\varphi(ex(X))=\varphi(X)$
(iv) \textbf{X7}: $ex(\varphi(X))=ex(X)$
\end{proposition}
\begin{proof}
If $(E,\varphi)$ is a uniquely generated violator space, then operator $ex$
satisfies \textbf{X5},\textbf{X6} and \textbf{X7}, since
[X]_{\varphi}=[X]_{ex}=[ex(X),\varphi(X)]$.
Before we continue with the proof, it is important to mention that from the
definition of the operator $ex$ it follows that $ex(B)=B$ for each basis $B$.
Further we prove that if a violator space $(E,\varphi )$ satisfies the
property \textbf{X5}, then it is uniquely generated. Suppose that there is a
set $X\subseteq E$ with two bases $B_{1}$ and $B_{2}$. Then $\varphi
(X)=\varphi (B_{1})=\varphi (B_{2})=\varphi (B_{1}\cup B_{2})$. Thus
Proposition \ref{exp} implies $ex(B_{1}\cup B_{2})\subseteq B_{1}\cap B_{2}
. Then we have $ex(B_{1}\cup B_{2})\subseteq B_{1}\subseteq B_{1}\cup B_{2}$
and $ex(B_{1}\cup B_{2})\subseteq B_{2}\subseteq B_{1}\cup B_{2}$, but
ex(B_{1})=B_{1}\neq ex(B_{2})=B_{2}$. In other words, we see that $ex$ does
not satisfy the outcast property.
$(iii)\Rightarrow (i)$ follows from Theorem \ref{th2}.
Now, it is only left to prove that if a violator space $(E,\varphi )$
satisfies the property \textbf{X7}, then it is uniquely generated. Suppose
there is a set $X\subseteq E$ with two bases $B_{1}\neq B_{2}$. Then
\varphi (X)=\varphi (B_{1})=\varphi (B_{2})$, and so $ex(\varphi
(B_{1}))=ex(\varphi (B_{2}))$. Since $ex(B_{1})=B_{1}\neq ex(B_{2})=B_{2}$,
we conclude that the property \textbf{X7} does not hold.
\end{proof}
It is worth reminding that \textbf{X6} and \textbf{X7} are called
Krein-Milman properties. In other words, every uniquely generated violator
space is a Krein-Milman space \cite{KM}.
\section{Conclusion}
Many combinatorial structures are described using operators defined on their
ground sets. For instance, closure spaces are defined by closure operators,
and violator spaces are described by violator operators. In this paper, we
introduced co-violator spaces based on contracting operators known also as
choice functions. Cospanning characterizations of violator spaces allowed us
to obtain some new properties of violator operators, co-violator operators
and their interconnections. In further research, our intent is to extend
this "cospanning" approach to a wider spectrum of combinatorial structures
closure spaces, convex geometries, antimatroids, etc.
|
\section{Introduction}
The spectrum of Vega (= $\alpha$~Lyr = HR~7001 = HD~172167 = HIP~91262; spectral type
A0~V) shows a sharp-line nature indicating a small projected rotational velocity
($v_{\rm e}\sin i \sim 20$~km~s$^{-1}$; where $v_{\rm e}$ is the equatorial rotation
velocity and $i$ is the angle of rotational axis relative to the line of sight),
which is rather unusual among A-type main-sequence stars (many of them showing
$v_{\rm e}\sin i$ typically around $\sim$~100--300~km~s$^{-1}$).
It is nowadays known, however, that this star is actually a rapid rotator with large
$v_{\rm e}$ like other A stars and the apparent smallness of $v_{\rm e}\sin i$ is
simply ascribed to low $i$ (i.e., this star happens to be seen nearly pole-on).
Its intrinsic rotational velocity can be observationally determined by detecting
the gravity darkening effect, because it becomes more exaggerated as $v_{\rm e}$ increases.
The mainstream approach used for this purpose is to analyse the shape of
spectral lines, because lines of a specific group (e.g., weak Fe~{\sc i} lines) show
a characteristic feature (i.e., flat-bottomed profile), which is caused by the lowered
temperature near to the gravity-darkened limb (see, e.g., Fig.~1 in Takeda, Kawanomoto \& Ohishi 2008a).
Alternatively, in order to establish $v_{\rm e}$, the extent of gravity darkening
can be estimated from the brightness distribution on the stellar disk by direct
high-resolution interferometric observations.
\setcounter{table}{0}
\begin{table*}
\begin{minipage}{180mm}
\small
\caption{Previous determinations of Vega's rotation and related parameters.}
\begin{center}
\begin{tabular}{cccccccl}\hline
\hline
Authors & $v_{\rm e}\sin i$ & $v_{\rm e}$ & $i$ & $R_{\rm p}$ & $R_{\rm e}$ & $P$ & Remark \\
& (km~s$^{-1}$) & (km~s$^{-1}$) & (deg) & (${\rm R}_{\odot}$) & (${\rm R}_{\odot}$) & (d) & \\
\hline
Gulliver et al. (1994) & 21.8 & 245 & 5.1 & $\cdots$ & $\cdots$ & $\cdots$ & Line profile \\
Hill et al. (2004) & 21.9 & 160 & 7.9 & $\cdots$ & $\cdots$ & $\cdots$ & Line profile \\
Aufdenberg et al. (2006) & 21.9 & 270 & 4.7 & 2.26 & 2.78 & $\cdots$ & Interferometry \\
Peterson et al. (2006) & 21.5 & 274 & 4.5 & 2.31 & 2.87 & $\cdots$ & Interferometry \\
Takeda et al. (2008b) & $^{*}$22 & 175 & 7.2 & 2.52 & 2.76 & $\cdots$ & Line profile \\
Yoon et al. (2010) & 20.5 & 236 & 5.0 & 2.36 & 2.82 & $\cdots$ & Line profile \\
Hill et al. (2010) & 20.8 & 211 & 5.7 & 2.40 & 2.75 & $\cdots$ & Line profile \\
Monnier et al. (2012) & 21.3 & 197 & 6.2 & 2.42 & 2.73 & $\cdots$ & Interferometry (their Model~3) \\
\hline
Petit et al. (2010) & $\cdots$ & $^{\dagger}$184 & $\cdots$ & $\cdots$ & $\cdots$ & 0.732 & Magnetic modulation \\
Alina et al. (2012) & $\cdots$ & $^{\dagger}$198 & $\cdots$ & $\cdots$ & $\cdots$ & 0.678 & Magnetic modulation \\
Butkovskaya (2014) & $\cdots$ & $^{\dagger}$216 & $\cdots$ & $\cdots$ & $\cdots$ & 0.623 & Magnetic modulation \\
B\"{o}hm et al. (2015) & $\cdots$ & $^{\dagger}$198 & $\cdots$ & $\cdots$ & $\cdots$ & 0.678 & Magnetic modulation \\
\hline
\end{tabular}
\end{center}
In columns 2--7 are given the values of projected rotational velocity, equatorial rotational velocity,
inclination angle of rotational axis, polar radius, equatorial radius, and rotation period, respectively.\\
$^{*}$Assumed value.\\
$^{\dagger}$Derived from $P$ by assuming $R_{\rm e} = 2.8 {\rm R}_{\odot}$
\end{minipage}
\end{table*}
Beginning from 1990s and especially in the period around 2010, quite a few
determinations of Vega's $v_{\rm e}$ based on these two methods have been
tried by various investigators as summarised in Table~1.
However, the resulting literature values of $v_{\rm e}$ considerably differ from
each other as seen from this table. Although the large discrepancy amounting
to $\ga 100$~km~s$^{-1}$ (from $\sim 160$ to $\sim 270$~km~s$^{-1}$) seen
in early 2000s has been mitigated up to the present, they are still diversified
between $\sim 170$~km~s$^{-1}$ and $\sim 230$~km~s$^{-1}$ (which are the published
results since 2008).
Meanwhile, the discovery of magnetic field in Vega by spectropolarimetry
(Ligni\`{e}res et al. 2009) provided a new means to measure $v_{\rm e}$,
because such a Zeeman signature would show cyclic variation due to rotation.
That is, the rotational period ($P$) may be directly evaluated by applying a period
analysis to time-series data of spectropolarimetric observations, from which $v_{\rm e}$
is derived as $v_{\rm e} = 2\pi R_{\rm e}/P$ by using an appropriately assigned
$R_{\rm e}$ (equatorial radius).
Following this policy, Vega's rotation period was determined within several years
after 2010, as shown in Table~1. Although this method is expected to establish $P$
precisely, these published data are not necessarily in good agreement
but somewhat discrepant by $\sim\pm 10$\% (i.e., $\sim \pm 20$~km~s$^{-1}$ around
$v_{\rm e} \sim 200$~km~s$^{-1}$). Therefore, even such an independent technique
(which is essentially different from the other in the sense that any modelling of
gravity-darkened star is not required) has not yet significantly improved
the situation regarding the ambiguity in $v_{\rm e}$.
Accordingly, it is desirable to redetermine $v_{\rm e}$ of Vega with higher reliability
than before, in order to clarify which of the recent results (between ``low-scale'' value
of $\sim$~170--180~km~s$^{-1}$ and ``high-scale'' value of $\sim$~220--230~km~s$^{-1}$)
is more justifiable.
Here, it may be worthwhile to mention the weakpoint of line profile analysis,
which was once employed by the author's group (Takeda, Kawanomoto \& Ohishi 2008b;
hereinafter referred to as Paper~I) to evaluate Vega's $v_{\rm e}$.
According to our experience, to derive $v_{\rm e}$ by searching for the best fit (minimising
$\chi^{2}$) between the observed and modelled line profiles for a selected line feature
(e.g., well-behaved weak Fe~{\sc i} line showing a flat-bottomed profile) is not so hard.
However, there is no way to estimate how much uncertainty is involved in such a
specific solution. Actually, since $\chi^{2}$ residual is a rather broad function of $v_{\rm e}$
and quite vulnerable to a slight imperfection (e.g., improper placement of continuum level,
existence of weak line blending, irregular noise in observed data, etc.), because
extremely subtle difference of profile shape is concerned (typically on the order
of $\sim 10^{-3}$ in unit of the continuum; cf. Figs. 4 and 5 in Paper~I),
an erroneous $v_{\rm e}$ solution is easily brought about (or even no solution
is found). Therefore, it was decided in Paper~I to analyse the profiles of
a large number of lines (87 lines of neutral species and 109 lines of once-ionised
species) with a hope of hitting as many correct solutions as possible.
Nevertheless, from a critical point of view, the result obtained
in Paper~I was not very satisfactory for the following reasons: (i) The final
solution ($v_{\rm e} = 175$~km~s$^{-1}$) was simply selected from 9 models (where
$v_{\rm e}$ was varied from 100 to 300~km~s$^{-1}$ with an increment of 25~km~s$^{-1}$)
as the one corresponding to the highest frequency of $\chi^{2}$ minimum for the case
of neutral lines; so an ambiguity of $\sim$~20~km~s$^{-1}$
due to the coarseness of model grid is inevitable from the start. (ii) While lines of
neutral species yielded a Gaussian-like frequency histogram centred around
175~km~s$^{-}$ (cf. Fig.~6a in Paper~I), those of ionised species (many of them
have ``non-flat-bottom'' profiles) show a near-flat distribution (cf. Fig.~6b in
Paper~I); this means that the latter set of ionised lines were almost useless
because they made no contribution to the determination of $v_{\rm e}$.
Consequently, the conventional line-profile matching in the wavelength domain applied
in Paper~I was not necessarily suitable for such a very delicate problem. In order to make
a further step towards improving the precision, a more efficient approach has to be invoked,
in which many lines of different properties can be effectively combined to increase
the reliability of $v_{\rm e}$ solution while providing a reasonable procedure for
error estimation.
Recently, in an attempt to estimate the intrinsic rotational velocity of Sirius~A,
Takeda (2020; hereinafter referred to as Paper~II) made use of the first zero frequency
($q_{1}$) in the Fourier transform of the line profile. It then revealed that
this quantity can be used for measuring the gravity darkening effect because it
sensitively responds to a slight variation of the line profile; actually, $q_{1}$ was
found to be vary almost monotonically with $v_{\rm e}$ (inducing a gravity darkening).
While how $q_{1}$ reflects a change of $v_{\rm e}$ naturally differs from line to line
depending on its property, it was found to be the sensitivity of line strength
($W$) to temperature ($T$), which is represented by the parameter $K (\equiv \log W/\log T)$,
that essentially controls the $v_{\rm e}$-dependence of $q_{1}$.
Therefore, since information of $v_{\rm e}$ may be extracted from the comparison of the
observed $q_{1}^{\rm obs}$ with a corresponding set of $q_{1}^{\rm cal}(K,v_{\rm e}$
calculated for this line on the models of different $v_{\rm e}$, the best solution
of $v_{\rm e}$ (along with its probable error) can be established by combining
many lines of different $K$.
This technique turned out successful, and in Paper~II was concluded that Sirius~A
is an intrinsically slow rotator ($16 \le v_{\rm e} \la$~30--40~km~s$^{-1}$).
Motivated by this achievement, the author decided to apply this method to analysing
the spectral line profiles of Vega, in order to revisit the task of determining its
$v_{\rm e}$ as done in Paper~I, hoping that a result of higher accuracy would be
obtained, so that the diversified literature values may be verified.
The purpose of this article is to report the outcome of this reinvestigation.
\section{Observational data}
\subsection{Selection of lines and their profiles}
Regarding the basic observational material of Vega, the high-dispersion spectra of
high signal-to-noise ratio (S/N~$\sim 2000$) and high spectral resolving
power ($R \sim 100000$) were used as in Paper~I, which were obtained at Okayama
Astrophysical Observatory by using the HIDES spectrograph attached to the 188~cm
reflector and published by Takeda, Kawanomoto \& Ohishi (2007).
The selection of lines to be used for the analysis was done by following almost the
same procedure as adopted in Paper~II (cf. Sect.~2.2 therein), where it was decided
to employ only lines of neutral and ionised Fe in order to maintain consistency with
Paper~II. As a result, a total of 90 lines (49 Fe~{\sc i} and 41 Fe~{\sc ii} lines)
were eventually sorted out,\footnote{Since the selection criterion adopted in this
study differs from that of Paper~I, the resulting line set is somewhat different.
More precisely, out of 60/52 Fe~{\sc i}/Fe~{\sc ii} lines analysed in Paper~I,
17/16 were discarded, while 6/5 were newly included.} which are listed in Table~2.
The observed profiles of these lines are displayed in Fig.~1, and their original
data are available in ``obsprofs.dat'' of the supplementary material.
The equivalent widths ($W^{\rm obs}$) of these 90 lines were measured by
the Gaussian fitting method, which are in the range of 1~m\AA~$\la W^{\rm obs} \la 40$~m\AA.
As the ``standard'' plane-parallel model atmosphere for Vega, Kurucz's (1993) ATLAS9
model with $T_{\rm eff}$ = 9630~K, $\log g = 3.94$, $v_{\rm t} = 2$~km~s$^{-1}$
(microturbulence), and [X/H]~=~$-0.5$ (metallicity) was adopted in this study as in Paper~I,
which well reproduces the spectral energy distribution. By using this model
along with the atomic data taken from Kurucz \& Bell's (1995) compilation, the abundance
($A^{\rm std}$; called as ``standard abundance'') was derived from $W^{\rm obs}$
for each line.
In the same manner as in Paper~II (cf. Sect.~4.1 therein), the $T$-sensitivity
parameter $K (\equiv {\rm d}\log W/{\rm d}\log T)$ was then evaluated as
\begin{equation}
K \equiv \frac{(W^{+100} - W^{-100})/W^{\rm obs}}{(+100-(-100))/9630},
\end{equation}
where $W^{+100}$ and $W^{-100}$ are the equivalent widths computed from $A^{\rm std}$
by two model atmospheres with only $T_{\rm eff}$ being perturbed by $+100$~K
($T_{\rm eff} = 9730$~K) and $-100$~K ($T_{\rm eff} = 9530$~K), respectively
(while other parameters are kept the same as the standard values).
The ranges of the resulting $K$ values are (roughly) $-20 \la K \la -10$ and
$-5 \la K \la +5$ for Fe~{\sc i} and Fe~{\sc ii} lines, respectively.
\subsection{Zero frequencies of Fourier transforms}
Then, the Fourier transform $d(\sigma)$ of the line depth profile
$D_{\lambda} (\equiv 1 - F_{\lambda}/F_{\rm cont})$ was calculated for each line
as done in Paper~II (cf. Sect.~2.3 therein), and
the 1st and 2nd zero frequencies ($\sigma_{1}$ and $\sigma_{2}$; in unit of wavelength)
were measured from the cuspy features of $|d(\sigma)|$, which were further converted
to wavelength-independent quantities ($q_{1}$ and $q_{2}$; in unit of velocity$^{-1}$)
for convenience by the relation $q \equiv \sigma /(c \lambda)$ ($c$: velocity of light).
The resulting $q_{1}$ and $q_{2}$ are plotted against the line parameters in Fig.~2,
from which the following arguments can be made.
\begin{itemize}
\item
These zero frequencies show an appreciable line-dependent scatter; especially,
those of a fraction of Fe~{\sc ii} lines are remarkably higher in comparison
with the theoretical values expected from the classical rotational broadening
function (cf. Fig.~2a)
\item
This implies that the conventional Fourier analysis of spectral line profiles,
which assumes that the observed profile is expressed by a convolution of the
rotational broadening function and thus the zero frequency of the rotational
broadening function (dependent upon $v_{\rm e}\sin i$) should be simply inherited
in the observed transform equally for any line, is no more feasible for precise
$v_{\rm e} \sin i$ determination in this case.
\item
The cause for this scatter in $q_{1}$ as well as $q_{2}$ is that they tend to
systematically increase with $K$ as shown in Fig.~2b. This is because the
line profile characteristics is determined by this $T$-sensitivity parameter.
That is, a line of small/negative $K$ (e.g., weak Fe~{\sc i} line of low excitation)
shows a boxy {\bf U}-shape, while that of large/positive $K$ (e.g.,
weak Fe~{\sc ii} line of high excitation) has a sharp {\bf V}-shape.
Such a difference in the line profile (even if very subtle) is reflected by
the position of zero frequency, which is actually verified by theoretical
calculations based on the gravity-darkened rotating star model (cf. Sect.~3.3).
\item
These $q_{1}$ and $q{2}$ also show some systematic trends with respect to
$\chi_{\rm low}$ (Fig.~2c) and $W^{\rm obs}$ (Fig.~2d); but they can be
reasonably explained by the dependence of $K$ upon $\chi_{\rm low}$
and $W^{\rm obs}$, as discussed in Appendix~A2 of Paper~II.
Accordingly, it is the difference in $K$ that causes the line-by-line
different characteristics in the profile (and the zero positions).
\end{itemize}
The atomic line data and the values of $W^{\rm obs}$, $K$, $q_{1}$, and $q_{2}$ for
90 lines are presented in Table~2. Besides, more complete data (including $A^{\rm std}$
and the main lobe height as well as the 1st sidelobe height) are summarised in
``obsparms.dat'' of the supplementary material.
\setcounter{figure}{0}
\begin{figure*}
\begin{minipage}{130mm}
\includegraphics[width=13.0cm]{fig1.eps}
\caption{
Observed spectra of finally selected 49 Fe~{\sc i} lines (1st through 3rd panels)
and 41 Fe~{\sc ii} lines (4th through 6th panels), which are arranged in the
increasing order of wavelength within each group of species as in Table~2.
The actual spectral data (normalised flux plotted against the wavelength
displacement relative to the line centre) are shown by lines, while the selected wavelength
portions [$\lambda_{1}$, $\lambda_{2}$] used for calculating the Fourier transforms
are depicted by symbols. Each spectrum (its continuum level is indicated by the horizontal
dashed line) is shifted by 0.02 (2\% of the continuum level)
relative to the adjacent one.
}
\label{fig1}
\end{minipage}
\end{figure*}
\setcounter{figure}{1}
\begin{figure*}
\begin{minipage}{120mm}
\includegraphics[width=12.0cm]{fig2.eps}
\caption{
The 1st-zero frequencies ($q_{1}$: circles) and 2nd-zero frequencies
($q_{2}$: triangles) of Fourier transforms, which were calculated
from the observed profiles of 90 Fe lines, are plotted against
(a) $\lambda$ (wavelength), (b) $K$ ($T$-sensitivity parameter), (c) $\chi_{\rm low}$
(lower excitation potential), and $W^{\rm obs}$ (observed equivalent width).
The filled and open symbols correspond to Fe~{\sc i} and Fe~{\sc ii} lines, respectively.
In panel~(a), the classical $q_{1}(\lambda)$ and $q_{2}(\lambda)$ values
derived from the conventional rotational broadening function corresponding to
$v_{\rm e}\sin i$ = 20, 21, 22, 23, and 24~km~s$^{-1}$ (which were derived
from Eqs.~3, 4, and 6 in Paper~II) are depicted by solid lines.
}
\label{fig2}
\end{minipage}
\end{figure*}
\setcounter{table}{1}
\begin{table}
\scriptsize
\caption{Atomic data and observed quantities of adopted spectral lines.}
\begin{center}
\begin{tabular}{crcrccc}\hline
\hline
$\lambda$ & $\chi_{\rm low}$ & $\log gf$ & $W^{\rm obs}$ & $K$ & $q_{1}^{\rm obs}$ & $q_{2}^{\rm obs}$ \\
(\AA) & (eV) & (dex) & (m\AA) & & (km$^{-1}$s) & (km$^{-1}$s) \\
\hline
\multicolumn{7}{c}{(49 Fe~{\sc i} lines)}\\
3922.911 & 0.052 & $-$1.651 & 25.3 & $-$14.89 & 0.02465 & 0.05085 \\
4014.534 & 3.573 & $-$0.200 & 7.6 & $-$14.48 & 0.02425 & 0.05050 \\
4021.867 & 2.759 & $-$0.660 & 9.4 & $-$15.21 & 0.02276 & 0.05120 \\
4067.978 & 3.211 & $-$0.430 & 7.8 & $-$14.63 & 0.02393 & 0.04988 \\
4175.636 & 2.845 & $-$0.670 & 6.4 & $-$14.82 & 0.02540 & 0.05027 \\
4176.566 & 3.368 & $-$0.620 & 5.2 & $-$15.59 & 0.02517 & 0.04823 \\
4187.038 & 2.449 & $-$0.548 & 15.9 & $-$14.36 & 0.02398 & 0.04980 \\
4195.329 & 3.332 & $-$0.412 & 8.3 & $-$14.90 & 0.02538 & 0.04882 \\
4196.208 & 3.396 & $-$0.740 & 4.2 & $-$14.73 & 0.02371 & 0.04530 \\
4199.095 & 3.047 & +0.250 & 26.5 & $-$12.26 & 0.02503 & 0.05200 \\
4202.028 & 1.485 & $-$0.708 & 33.6 & $-$12.38 & 0.02477 & 0.05236 \\
4210.343 & 2.482 & $-$0.870 & 7.8 & $-$15.85 & 0.02426 & 0.05217 \\
4219.360 & 3.573 & +0.120 & 15.4 & $-$13.31 & 0.02362 & 0.04897 \\
4235.936 & 2.425 & $-$0.341 & 21.6 & $-$13.69 & 0.02568 & 0.05155 \\
4238.021 & 3.417 & $-$1.286 & 2.7 & $-$15.76 & 0.02373 & 0.04730 \\
4238.809 & 3.396 & $-$0.280 & 9.8 & $-$14.59 & 0.02490 & 0.05084 \\
4299.234 & 2.425 & $-$0.430 & 21.1 & $-$13.56 & 0.02584 & 0.05010 \\
4447.718 & 2.223 & $-$1.342 & 5.9 & $-$16.05 & 0.02342 & 0.04301 \\
4466.551 & 2.832 & $-$0.590 & 12.4 & $-$14.96 & 0.02520 & 0.05088 \\
4484.219 & 3.602 & $-$0.720 & 3.4 & $-$15.35 & 0.02078 & 0.04992 \\
4494.563 & 2.198 & $-$1.136 & 7.6 & $-$16.26 & 0.02423 & 0.04895 \\
4528.613 & 2.176 & $-$0.822 & 16.3 & $-$14.93 & 0.02450 & 0.05115 \\
4547.846 & 3.546 & $-$0.780 & 2.0 & $-$16.44 & 0.02646 & 0.05141 \\
4602.940 & 1.485 & $-$1.950 & 2.9 & $-$17.95 & 0.02107 & 0.04621 \\
4611.284 & 3.654 & $-$0.670 & 2.8 & $-$15.21 & 0.02335 & 0.04747 \\
4707.272 & 3.241 & $-$1.080 & 2.7 & $-$15.76 & 0.02289 & 0.05004 \\
4903.308 & 2.882 & $-$1.080 & 3.9 & $-$15.85 & 0.02487 & 0.05074 \\
4918.993 & 2.865 & $-$0.370 & 14.0 & $-$14.24 & 0.02469 & 0.05088 \\
4920.502 & 2.832 & +0.060 & 29.3 & $-$12.08 & 0.02636 & 0.05265 \\
4966.087 & 3.332 & $-$0.890 & 3.7 & $-$16.69 & 0.02211 & 0.05162 \\
5049.819 & 2.279 & $-$1.420 & 4.7 & $-$17.23 & 0.02071 & 0.04464 \\
5133.681 & 4.178 & +0.140 & 10.0 & $-$13.76 & 0.02473 & 0.05017 \\
5162.292 & 4.178 & +0.020 & 8.7 & $-$14.23 & 0.02422 & 0.04991 \\
5281.790 & 3.038 & $-$1.020 & 4.0 & $-$15.46 & 0.02387 & 0.04728 \\
5341.023 & 1.608 & $-$2.060 & 3.5 & $-$17.63 & 0.02246 & 0.06592 \\
5353.373 & 4.103 & $-$0.840 & 1.5 & $-$15.53 & 0.02327 & 0.04369 \\
5364.858 & 4.446 & +0.220 & 7.5 & $-$13.94 & 0.02370 & 0.05068 \\
5367.479 & 4.415 & +0.350 & 9.3 & $-$13.76 & 0.02557 & 0.05168 \\
5371.489 & 0.958 & $-$1.645 & 12.3 & $-$16.95 & 0.02419 & 0.04971 \\
5389.479 & 4.415 & $-$0.410 & 1.9 & $-$15.21 & 0.02572 & 0.04028 \\
5569.618 & 3.417 & $-$0.540 & 4.5 & $-$15.87 & 0.02460 & 0.05192 \\
5572.841 & 3.396 & $-$0.310 & 7.6 & $-$15.01 & 0.02508 & 0.05084 \\
5576.090 & 3.430 & $-$1.000 & 2.2 & $-$14.98 & 0.02342 & 0.04979 \\
5615.644 & 3.332 & $-$0.140 & 14.3 & $-$14.33 & 0.02563 & 0.05133 \\
5633.975 & 4.991 & $-$0.270 & 1.6 & $-$14.59 & 0.02282 & 0.04763 \\
6136.615 & 2.453 & $-$1.400 & 3.8 & $-$17.28 & 0.02402 & 0.04708 \\
6137.694 & 2.588 & $-$1.403 & 2.8 & $-$18.58 & 0.02271 & 0.04905 \\
6191.558 & 2.433 & $-$1.600 & 2.8 & $-$18.58 & 0.02341 & 0.04736 \\
6230.726 & 2.559 & $-$1.281 & 4.0 & $-$16.44 & 0.02328 & 0.05124 \\
\hline
\end{tabular}
\end{center}
\end{table}
\setcounter{table}{1}
\begin{table}
\scriptsize
\caption{(Continued.)}
\begin{center}
\begin{tabular}{crcrccc}\hline
\hline
$\lambda$ & $\chi_{\rm low}$ & $\log gf$ & $W^{\rm obs}$ & $K$ & $q_{1}^{\rm obs}$ & $q_{2}^{\rm obs}$ \\
(\AA) & (eV) & (dex) & (m\AA) & & (km$^{-1}$s) & (km$^{-1}$s) \\
\hline
\multicolumn{7}{c}{(41 Fe~{\sc ii} lines)}\\
4273.326 & 2.704 & $-$3.258 & 17.2 & $-$5.30 & 0.02662 & 0.05143 \\
4278.159 & 2.692 & $-$3.816 & 7.2 & $-$5.98 & 0.02611 & 0.05284 \\
4296.572 & 2.704 & $-$3.010 & 32.6 & $-$4.28 & 0.02764 & 0.05285 \\
4413.601 & 2.676 & $-$3.870 & 4.2 & $-$5.66 & 0.02780 & 0.04958 \\
4451.551 & 6.138 & $-$1.844 & 8.1 & $-$1.19 & 0.03020 & 0.05623 \\
4472.929 & 2.844 & $-$3.430 & 13.2 & $-$5.45 & 0.02412 & 0.05491 \\
4489.183 & 2.828 & $-$2.970 & 31.7 & $-$4.54 & 0.02730 & 0.05278 \\
4491.405 & 2.855 & $-$2.700 & 38.5 & $-$4.12 & 0.02813 & 0.05322 \\
4541.524 & 2.855 & $-$3.050 & 27.7 & $-$4.68 & 0.02737 & 0.05199 \\
4576.340 & 2.844 & $-$3.040 & 27.6 & $-$4.70 & 0.02744 & 0.05185 \\
4582.835 & 2.844 & $-$3.100 & 19.4 & $-$5.20 & 0.02645 & 0.05223 \\
4620.521 & 2.828 & $-$3.280 & 15.8 & $-$5.16 & 0.02702 & 0.05118 \\
4635.316 & 5.956 & $-$1.650 & 15.6 & $-$1.54 & 0.03032 & 0.05428 \\
4666.758 & 2.828 & $-$3.330 & 16.4 & $-$5.56 & 0.02623 & 0.05034 \\
4713.193 & 2.778 & $-$4.932 & 5.9 & $-$6.53 & 0.03818 & 0.06809 \\
4731.453 & 2.891 & $-$3.360 & 20.9 & $-$5.29 & 0.02692 & 0.05184 \\
4913.292 &10.288 & +0.012 & 2.0 & +4.82 & 0.03830 & 0.08123 \\
4948.793 &10.347 & $-$0.008 & 1.9 & +5.07 & 0.03419 & 0.06236 \\
4951.584 &10.307 & +0.175 & 3.3 & +2.92 & 0.03701 & 0.05541 \\
4977.035 &10.360 & +0.041 & 2.1 & +4.59 & 0.03409 & 0.05122 \\
4993.358 & 2.807 & $-$3.650 & 8.4 & $-$6.27 & 0.02610 & 0.05199 \\
5004.195 &10.272 & +0.497 & 6.6 & +2.92 & 0.03372 & 0.05589 \\
5089.214 &10.329 & $-$0.035 & 2.5 & +3.85 & 0.02708 & 0.04168 \\
5106.109 &10.329 & $-$0.276 & 1.1 & 0.00 & 0.04118 & 0.08262 \\
5127.866 & 5.570 & $-$2.535 & 3.7 & $-$2.60 & 0.02657 & 0.07004 \\
5132.669 & 2.807 & $-$4.180 & 3.1 & $-$6.21 & 0.02566 & 0.05284 \\
5149.465 &10.447 & +0.396 & 5.3 & +3.63 & 0.02978 & 0.04636 \\
5203.638 &10.391 & $-$0.046 & 1.9 & +5.07 & 0.02968 & 0.05292 \\
5219.926 &10.522 & $-$0.366 & 1.2 & +3.85 & 0.02856 & 0.04502 \\
5272.397 & 5.956 & $-$2.030 & 7.0 & $-$2.05 & 0.02980 & 0.04995 \\
5291.666 &10.480 & +0.575 & 5.1 & +2.80 & 0.03584 & 0.07705 \\
5387.063 &10.521 & +0.518 & 4.4 & +3.25 & 0.03395 & 0.05187 \\
5529.932 & 6.729 & $-$1.875 & 3.5 & $-$1.36 & 0.02854 & 0.05276 \\
5567.842 & 6.730 & $-$1.887 & 2.1 & 0.00 & 0.03195 & 0.05079 \\
5645.392 &10.561 & +0.085 & 1.9 & +5.07 & 0.03955 & 0.05893 \\
5835.492 & 5.911 & $-$2.372 & 1.4 & 0.00 & 0.02289 & 0.04158 \\
6084.111 & 3.199 & $-$3.808 & 4.0 & $-$7.22 & 0.02622 & 0.05076 \\
6147.741 & 3.889 & $-$2.721 & 14.4 & $-$5.00 & 0.02691 & 0.05213 \\
6149.258 & 3.889 & $-$2.724 & 14.0 & $-$5.14 & 0.02678 & 0.05177 \\
6175.146 & 6.222 & $-$1.983 & 4.0 & $-$2.41 & 0.02596 & 0.04678 \\
6248.898 & 5.511 & $-$2.696 & 3.2 & $-$3.01 & 0.02396 & 0.04336 \\
\hline
\end{tabular}
\end{center}
In columns 1--7 are given the line wavelength, lower excitation potential,
logarithm of oscillator strength times lower level's statistical weight,
observed equivalent width, $T$-sensitivity parameter, observed 1st zero.
frequency, and observed 2nd zero frequency, respectively. The atomic data
are taken from the compilation of Kurucz \& Bell (1995).
\end{table}
\section{Modelling of line profiles}
\subsection{Adopted model parameters}
Regarding the simulation of theoretical line profiles of a gravity-darkened rotating
star, this study follows the same assumptions and procedures (including the adopted set
of parameters for Vega) as described in Paper~I, where the stellar mass ($M$), rotational
velocity at the equator ($v_{\rm e}$), inclination angle of rotation axis ($i$),
polar radius ($R_{\rm p}$), and polar effective temperature ($T_{\rm eff,p}$)
are the fundamental parameters to be specified.
The mass was fixed at $M = 2.3$~M$_{\odot}$, Ten $v_{\rm e}$ values were chosen as
22, 100, 125, 150, $\cdots$ 275, and 300~km~s$^{-1}$, (numbered as models
0, 1, 2, 3, $\cdots$, 8, and 9), and the corresponding $i$ values were derived from the
assumption of $v_{\rm e}\sin i = 22$~km~s$^{-1}$ (which is a reasonable value for Vega).
Based on the requirement of spectral energy distribution, $R_{\rm p}$ and $T_{\rm eff,p}$
can be expressed as 2nd-order polynomials in terms of $v_{\rm e}$ (cf. Eqs. 1 and 2 in Paper~I).
The model parameters for each of the 10 models are summarised in Table~3,
which is the same as Table~1 in Paper~I.
Note that model~0 ($v_{\rm e} = 22$~km~s$^{-1}$ and $i = 90^{\circ}$) is a special model
different from others, in the sense that it is a spherically symmetric rigid model
where the gravity effect (darkening and distortion) is intentionally suppressed.
This model~0 is almost equivalent to the ``standard model'' mentioned in Sect.~2.1.
\setcounter{table}{2}
\begin{table*}
\begin{minipage}{180mm}
\small
\caption{Parameters of adopted models for rotating Vega.}
\begin{center}
\begin{tabular}{cccccccccl}\hline
Model & $v_{\rm e}$ & $i$ & $R_{\rm p}$ & $R_{\rm e}$ & $T_{\rm eff,p}$ & $T_{\rm eff,e}$ & $\log g_{\rm p}$ & $\log g_{\rm e}$ & Remark \\
number & (km~s$^{-1}$) & (deg) & (${\rm R}_{\odot}$) & (${\rm R}_{\odot}$) & (K) & (K) & (cm~s$^{-2}$) & (cm~s$^{-2}$) & \\
\hline
0 & 22 & 90.0 & 2.700 & 2.700 & 9630 & 9630 & 3.937 & 3.937 & Gravity effect suppressed.\\
1 & 100 & 12.7 & 2.640 & 2.722 & 9698 & 9399 & 3.956 & 3.956 & \\
2 & 125 & 10.1 & 2.600 & 2.726 & 9750 & 9281 & 3.969 & 3.884 & \\
3 & 150 & 8.4 & 2.560 & 2.740 & 9806 & 9126 & 3.983 & 3.858 & \\
4 & 175 & 7.2 & 2.520 & 2.763 & 9867 & 8931 & 3.997 & 3.823 & Nominated model in Paper~I. \\
5 & 200 & 6.3 & 2.470 & 2.784 & 9932 & 8695 & 4.014 & 3.783 & Best model concluded in this study. \\
6 & 225 & 5.6 & 2.410 & 2.799 & 10000 & 8416 & 4.035 & 3.736 & \\
7 & 250 & 5.0 & 2.360 & 2.837 & 10074 & 8072 & 4.054 & 3.669 & \\
8 & 275 & 4.6 & 2.300 & 2.869 & 10151 & 7787 & 4.076 & 3.587 & \\
9 & 300 & 4.2 & 2.240 & 2.908 & 10233 & 7546 & 4.099 & 3.477 & \\
\hline
\end{tabular}
\end{center}
Given are the model number, equatorial rotation velocity, inclination angle, radius,
effective temperature, and logarithmic surface gravity at the pole as well as the equator.
These models are the same as adopted in Paper~I (cf. Table~1 therein).
Note that $v_{\rm e}\sin i$ is assumed to be 22~km~s$^{-1}$ in all these models.
\end{minipage}
\end{table*}
\subsection{Simulation of line profiles}
The emergent line flux profile was simulated with the program CALSPEC (cf. Sect.~4.1
in Paper~I) by integrating the intensity profile at each point on the visible disk,
which was generated by using the local model atmosphere corresponding to
$T_{\rm eff}(\Theta)$, $g(\Theta)$, $v_{\rm t}$~=~2~km~s$^{-1}$,
and [X/H] = $-0.5$ (where $\Theta$ is the co-latitude).
Here, a point to notice is how to assign the elemental abundance ($A$).
If $A^{\rm std}$ (standard abundance derived from the classical plane-parallel model)
is simply used, the equivalent width of the calculated line profile ($W^{\rm cal}$)
turns out generally stronger than $W^{\rm obs}$ because of the gravity darkening
effect,\footnote{Although $A^{\rm std}$ was simply used in Paper~II for all models
irrespective of $v_{\rm e}$,it did not cause any serious problem because
$v_{\rm e}$-dependent gravity darkening effect was not so large as to cause
a significant $W^{\rm cal}$ vs. $W^{\rm obs}$ discrepancy in the $v_{\rm e}$
range ($\le 150$~km~s$^{-1}$) inspected therein.}
and this discrepancy progressively increases towards higher $v_{\rm e}$
(as can be recognised in Figs. 4 and 5 in Paper~I).
In Paper~I, this problem was circumvented by renormalising the calculated profile
(cf. Eq. 7 therein) so as to force $W^{\rm cal} = W^{\rm obs}$, although its validity
was not necessarily clear.
Fortunately, this equality does not have to be strictly realised in the present case
of Fourier analysis, because it is the ``characteristics'' of the line shape
that is essential. Accordingly, the following procedure was adopted in this study.
\begin{itemize}
\item
First, the provisional equivalent width $W_{*}^{j}$ ($j = 0, 1, 2, \cdots, 9$)
was calculated with CALSPEC for each model by using $A^{\rm std}$.
\item
Then, the corresponding abundance $A_{*}^{j}$ was derived from $W_{*}^{j}$
with the help of Kurucz's (1993) WIDTH9 program by using the standard plane
parallel model (cf. Sect. 2.1).
\item
The abundance difference defined as $\Delta A^{j} \equiv A^{\rm std} - A_{*}^{j}$
(which is mostly negative) is used as abundance correction to be applied to $A^{\rm std}$.
That is the abundance actually adopted in CALSPEC for calculating the profile
corresponding to model~$j$ is $A^{\rm std} + \Delta A^{j}$.
\end{itemize}
It should be remarked that this procedure is based on two assumptions that
(i) the classical curve of growth ($\log W$ vs. $A$ relation) for the plane-parallel model
is applicable even for the gravity-darkened case, and (ii) the absolute change of
$\log W$ in response to perturbation by $\pm \Delta A$ around $A^{\rm std}$
in this curve of growth is almost the same (i.e., locally linear).
Despite these rough approximations, the discrepancy between $W^{\rm cal}$ and
$W^{\rm obs}$ seen for the case of simply using $A^{\rm std}$ is considerably
reduced by application of this correction ($\Delta A$), as shown in Fig.~3.
\setcounter{figure}{2}
\begin{figure}
\includegraphics[width=8.0cm]{fig3.eps}
\caption{
Graphical illustration describing how the abundance correction applied to the
standard abundance (cf. Sect.~3.2) improves the discrepancy between the observed
($W^{\rm obs}$) and calculated ($W^{\rm cal}$) equivalent widths for each line,
where $W^{\rm cal}$ (upper panels) and $\log (W^{\rm cal}/W^{\rm obs})$ (lower panels)
are plotted against $W^{\rm obs}$. The left-hand panels (a, c) correspond to
the case of using the standard (uncorrected) abundances, while the right-hand panels
(b, d) to the case of using the corrected abundances. The results for models 0, 1, 2, 3,
4, 5, 6, 7, 8, and 9 are denoted by filled circles, filled triangles, filled inverse triangles,
filled squares, open circles, open triangles, open inverse triangles, open squares,
St. Andrew's crosses ($\times$), and Greek crosses (+), respectively.
(The symbols for model~0, models~1--3, models~4--6, and models~7--9 are coloured in
black, blue, green, and red, respectively.)
}
\label{fig3}
\end{figure}
\subsection{Fourier transform and the trend of first zero}
By using such corrected abundances, the theoretical line profiles were simulated
for each of the 10 models and their Fourier transform were computed, from which
$q_{1}^{j}$ and $q_{2}^{j}$ ($j = 0, 1, 2, \cdots, 9$) were measured.
These $q_{1}^{j}$ and $q_{2}^{j}$ values along with the adopted abundance corrections
($\Delta A^{j}$) for all 90 lines are given in ``calparms.dat'' of the supplementary material.
As demonstrative examples, the simulated profiles of Fe~{\sc i} 5133.681 ($K = -13.76$)
and Fe~{\sc ii} 4951.584 ($K = +2.92$) lines and their Fourier transform amplitudes,
which were calculated for models 0, 1, 3, 5, 7, and 9, are illustrated in Fig.~4, where
the observed data are also overplotted for comparison.
It can be seen from Fig.~4 that the behaviours of zero frequency for these two
lines of different $K$ are just the opposite in the sense that $q_{1}$ of
Fe~{\sc i}~5133.681/Fe~{\sc ii}~4951.584 moves towards lower/higher direction
as the gravity-darkening effect is enhanced with an increase in $v_{\rm e}$.
From now on, our discussion focuses only on the first zero frequency ($q_{1}$),
which is less affected by measurement errors or noises in comparison to $q_{2}$.
In order to elucidate the trend of $q_{1}$ as a function of $K$ and $v_{\rm e}$,
the $q_{1}$ values are plotted against $K$ in Fig.~5a--5f (each corresponding to
models 0, 1, 3, 5, 7, and 9, respectively).
Besides, the linear regression lines (determined from the $q_{1}$ vs. $K$ plots
for Fe~{\sc i} and Fe~{\sc ii} lines separately) are also shown in each panel,
and these regression lines for all models are depicted together in Fig.~5g.
An inspection of Fig.~5 reveals the following characteristics.
\begin{itemize}
\item
$q_{1}$ generally increases with an increase in $K$, which was already mentioned
in Sect.~2.2 in reference to Fig.~2b. The $q_{1}$ values for Fe~{\sc i} lines are
generally smaller than those of Fe~{\sc ii} lines because of the difference in $K$.
\item
The slope of the $q_{1}$ vs. $K$ plots is a systematic function of $v_{e}$;
i.e., it becomes progressively steeper with an increase in $v_{\rm e}$ (Fig.~5g).
This is a useful property for estimating $v_{\rm e}$ from the observed
$q_{1}$--$K$ relation.
\item
The sensitivity of $q_{1}$ to a change in $v_{\rm e}$ also depends upon $K$
(cf. Fig.~5h). While $\langle {\rm d}q_{1}/{\rm d}v_{\rm e}\rangle \la 0$ holds
for most lines of $K \la 0$ (all Fe~{\sc i} lines and many Fe~{\sc ii} lines),
a group of high-excitation Fe~{\sc ii} lines ($\chi_{\rm low} \sim 10$~eV;
such as Fe~{\sc ii}~4951.584 in Fig.~4) with positive $K$ indicate
$\langle {\rm d}q_{1}/{\rm d}v_{\rm e}\rangle > 0$.
\end{itemize}
\setcounter{figure}{3}
\begin{figure*}
\begin{minipage}{120mm}
\includegraphics[width=12.0cm]{fig4.eps}
\caption{
Theoretical line profiles (top panels) and their Fourier transform amplitudes
(middle/bottom panels for wide/zoomed view) of Fe~{\sc i} 5133.681 (left-hand side) and
Fe~{\sc ii} 4951.584 (right-hand side) calculated for models 0 (black dashed line),
1 (black solid line), 3 (purple solid line), 5 (blue solid line), 7 (green solid line),
and 9 (red solid line), while the observed data are also overplotted by light-green symbols.
In the middle/bottom panels, the positions of $q_{1}^{\rm classical}$ and $q_{2}^{\rm classical}$
corresponding to the classical rotational broadening function (cf. Eqs.~3, 4, and 6 in Paper~II)
are indicated by vertical dotted lines for comparison.
}
\label{fig4}
\end{minipage}
\end{figure*}
\setcounter{figure}{4}
\begin{figure*}
\begin{minipage}{140mm}
\begin{center}
\includegraphics[width=14.0cm]{fig5.eps}
\caption{
Panels (a)--(f) show the simulated relationships between $q_{1}$ (1st-zero frequency)
and $K$ ($T$-sensitivity parameter) for the 49 Fe~{\sc i} (red filled symbols)
and 41 Fe~{\sc ii} lines (blue open symbols) calculated for models 0, 1, 3, 5, 7,
and 9, respectively. The size and shape of the symbols denote the difference
in line strengths; circles $\cdots$ $W^{\rm obs} <$~5~m\AA,
triangles $\cdots$ 5~m\AA~$\le W^{\rm obs} <$~10~m\AA,
squares $\cdots$ 10~m\AA~$\le W^{\rm obs} <$~20~m\AA, and
diamonds $\cdots$ 20~m\AA~$\le W^{\rm obs} <$~40~m\AA.
The linear regression lines derived from these $q_{1}$ vs. $K$ plots
(separately for Fe~{\sc i} and Fe~{\sc ii}) are also overplotted by solid lines in each
of the panels (a)--(f), and those for all 10 models are put together in panel (g).
The three horizontal dotted lines represent the classical $q_{1}$ values (corresponding
to $v_{\rm e}\sin i$ = 22~km~s$^{-1}$) for the limb-darkening coefficient
($\epsilon$) of 0.3, 0.4, and 0.5 (cf. Eq.~3 in Paper~II).
In panel (h) are plotted the mean gradients $\langle {\rm d}q_{1}/{\rm d}v_{\rm e}\rangle$
(in unit of km$^{-2}$s$^{2}$; averaged over $v_{\rm e}$ between 100 and 300~km~s$^{-1}$)
against $K$, which were computed from the coefficients of quadrature fit
($q_{1} = A + B v_{\rm e} + C v_{\rm e}^{2}$) as $B+2C \times (100+300)/2$ (i.e.,
${\rm d}q_{1}/{\rm d}v_{\rm e}$ at the mid-$v_{\rm e}$).
}
\label{fig5}
\end{center}
\end{minipage}
\end{figure*}
\section{Result and discussion}
\subsection{Rotational velocity of Vega}
Now that the observational data of zero frequencies ($q_{1}^{\rm obs}$) as well as
the corresponding theoretically calculated values ($q_{1}^{{\rm cal},j}$ for
$j = 0, 1, \cdots, 9$) to be compared are all set for 90 lines, we can address
the main task of investigating Vega's rotational velocity, while following
the same procedure as adopted in Paper~II (cf. Sect.~4.3 therein).
The observed $q_{1}^{\rm obs}$ values are plotted against $K$ in Fig.~6.
As seen from this figure, these $q_{1}^{\rm obs}$ data show an increasing tendency
with $K$ and those for Fe~{\sc i} and Fe~{\sc ii} lines are distributed in separate
two groups, which is quite similar to the theoretical predictions mentioned in Sect.~3.3
(cf. Figs.~5a--5f). Therefore, there is a good hope of successfully establishing
$v_{\rm e}$ by comparing $q_{1}^{\rm obs}$ and $q_{1}^{{\rm cal}}$ for many lines altogether.
\setcounter{figure}{5}
\begin{figure}
\begin{minipage}{80mm}
\begin{center}
\includegraphics[width=8.0cm]{fig6.eps}
\caption{
Observed $q_{1}$ values of Fe~{\sc i} and Fe~{\sc ii} lines plotted against $K$,
where the meanings of the symbols are the same as in Fig.~5.
The averaged trends (gradients) of theoretical $q_{1}$ vs. $K$ relations calculated
for models 1, 3, 5, 7, and 9 (determined by linear-regression analysis; cf. Fig.~5)
are also depicted by solid lines, which were multiplied by a scaling factor of
$22/x^{*}$ in order to adjust the difference between the actual
$v_{\rm e}\sin i (\equiv x)$ and the assumed value (22~km~s$^{-1}$) in the profile
calculation (see Table~4 for the $v_{\rm e}$-dependent values of $x^{*}$).
}
\label{fig6}
\end{center}
\end{minipage}
\end{figure}
Since the actual value of $v_{\rm e}\sin i$ (hereinafter denoted as $x$ for simplicity)
is likely to be slightly different from 22~km~s$^{-1}$ assumed for calculating the
modelled profiles, $q_{1}^{{\rm th},j}$ should be multiplied by a scaling factor ($22/x$)
to allow for this possible difference.
The standard deviation defined as
\begin{equation}
\sigma(x^{i},v_{\rm e}^{j}) \equiv
\sqrt{
\frac{ \sum_{n=1}^{N} [q_{1}^{\rm obs}(n) - q_{1}^{{\rm cal},j}(n)(22/x^{i})]^{2}}{N}.
}
\end{equation}
was computed for each combination of ($x^{i}$, $v_{\rm e}^{j}$),
where $x^{i} = 15.0 + 0.2i$ ($i$ = 0, 1, $\cdots$, 75) and $v_{\rm e}^{j} = 100 + 25(j-1)$
($j$ = 1, 2, $\cdots$, 9). Here, $n$ is the index of each line and $N$ is the total
number of the lines used. As in Paper~II, Fe~{\sc i} lines ($N = 49$) and Fe~{\sc ii}
lines ($N = 41$) are treated separately.
The best ($x$, $v_{\rm e}$) solution may be found by searching for the location of
$\sigma$ minimum.
The behaviours of the resulting $\sigma$ (3D surface and contour plots) are displayed
in Fig.~7 (left and right panels are for Fe~{\sc i} and Fe~{\sc ii}, respectively).
The trace line connecting ($x^{*}$, $v_{\rm e}$) is also overplotted by the dashed
line, where $x^{*}$ corresponds to the minimum of $\sigma$ trough for each $v_{\rm e}$
(in Table~4 are given the actual data of $x^{*}$ and the corresponding $\sigma^{*}$).
Besides, the run of $\sigma$ with $v_{\rm e}$ across the tracing is depicted
in Fig.~8a, and the tracings for both species are drawn together in Fig.~8b.
\setcounter{figure}{6}
\begin{figure*}
\begin{minipage}{120mm}
\begin{center}
\includegraphics[width=12.0cm]{fig7.eps}
\caption{
Graphical display of the behaviour of $\sigma$, which is the standard deviation
between the simulated $q_{1}^{\rm cal}(x,v_{\rm e})$ (where $x\equiv v_{\rm e}\sin i$) and
the observed $q_{1}^{\rm obs}$ for each of the Fe lines, where the results for
Fe~{\sc i} and Fe~{\sc ii} lines are separately displayed in the left and
right, respectively. Each set consists of the 3D representation of the $\sigma(x,v_{\rm e})$
surface (upper panel) and the contours of $\sigma$ on the $x$--$v_{\rm e}$ plane (lower panel).
The trace of trough bottom (connection of $x^{*}$ values at the minimum $\sigma$ for each given
$v_{\rm e}$; cf. Table~4) is indicated by the dashed line in the contour plot.
}
\label{fig7}
\end{center}
\end{minipage}
\end{figure*}
\setcounter{figure}{7}
\begin{figure}
\begin{minipage}{80mm}
\begin{center}
\includegraphics[width=8.0cm]{fig8.eps}
\caption{
(a) $\sigma$ vs. $v_{\rm e}$ relation along the trough bottom
for Fe~{\sc i} (filled symbols) and Fe~{\sc ii} (open symbols); in each case,
the position of minimum $\sigma$ (evaluated by interpolation) is indicated by an arrow.
Note that $\sigma_{\rm II}$ ($\sigma$ for Fe~{\sc ii}) is reduced by a factor
of 1/3 in this figure.
(b) The traces of trough bottoms for Fe~{\sc i} and Fe~{\sc ii} (dashed lines in
the contour panels of Fig.~7) are plotted together in the $v_{\rm e}$ vs. $x$ plane
by solid lines (the intersection is shown by the double circle),
while the dashed lines indicate the error bars involved in $x$
($\pm 0.14$~km~s$^{-1}$ for Fe~{\sc i} and $\pm 0.36$~km~s$^{-1}$ for Fe~{\sc ii}).
The minimum positions of $\sigma_{\rm I}$ and $\sigma_{\rm II}$ are also indicated
by Greek cross (+) and St. Andrew's cross ($\times$), respectively.
}
\label{fig8}
\end{center}
\end{minipage}
\end{figure}
An inspection of Fig.~8 yielded satisfactory results, because three kinds
of ($x$, $v_{\rm e}$) solutions turned out consistent with each other: (191, 21.8)
from the minimum of $\sigma_{\rm I}$ (Fig.~8a), (194, 21.5) from the minimum of
$\sigma_{\rm II}$ (Fig.~8a), and (201, 21.5) from the intersection of two trace lines (Fig.~8b).
The uncertainties involved in $x^{*}$ were estimated as $\sim 0.14$~km~s$^{-1}$ (Fe~{\sc i})
and $\sim 0.36$~km~s$^{-1}$ (Fe~{\sc ii}),\footnote{
This estimation is based on the relation $\delta x/x \sim \delta q_{1}/q_{1}$,
where $\delta q_{1} \sim \sigma /\sqrt{N}$. See Sect.~4.3 in Paper~II for more details.}
which are indicated by dashed lines in Fig.~8b. From this figure, errors in $v_{\rm e}$
and $x$ were roughly evaluated (from the size of the parallelogram area embraced by
4 dashed lines around the intersection) as $\sim \pm 15$~km~s$^{-1}$
and $\sim \pm 0.3$~km~s$^{-1}$, respectively.
Consequently, by averaging these three solutions, Vega's equatorial and projected
rotational velocities were concluded as $v_{\rm e} = 195 (\pm 15)$~km~s$^{-1}$ and
$v_{\rm e}\sin i = 21.6 (\pm 0.3)$~km~s$^{-1}$, which further result in
$i = 6.4^{\circ} (\pm 0.5^{\circ})$.
Among the 10 models adopted in this study (cf. Table~3), model~5 ($v_{\rm e} = 200$~km~s$^{-1}$)
is the most preferable model; this can be actually confirmed in Fig.~6, where the linear-regression
lines defined in Fig.~5b--5f are overplotted (after the $v_{\rm e}$-dependent difference
between $x^{*}$ and 22 has been corrected).
\setcounter{table}{3}
\begin{table}
\small
\caption{Behaviours of $\sigma$ trough for Fe~{\sc i} and Fe~{\sc ii} lines.}
\begin{center}
\begin{tabular}{c@{ }c@{ }cccc}\hline
Model & $v_{\rm e}$ & $x^{*}_{\rm I}$ & $x^{*}_{\rm II}$ & $\sigma^{*}_{\rm I}$ & $\sigma^{*}_{\rm II}$ \\
number & (km~s$^{-1}$) & (km~s$^{-1}$) & (km~s$^{-1}$) & (km$^{-1}$s) & (km$^{-1}$s) \\
\hline
1 & 100 & 23.9601 & 21.4741 & 0.0011587 & 0.0034953 \\
2 & 125 & 23.4263 & 21.4789 & 0.0011351 & 0.0034061 \\
3 & 150 & 22.8475 & 21.5036 & 0.0011152 & 0.0033180 \\
4 & 175 & 22.2098 & 21.5175 & 0.0010961 & 0.0032541 \\
5 & 200 & 21.5720 & 21.5339 & 0.0010948 & 0.0032388 \\
6 & 225 & 21.0187 & 21.6385 & 0.0010991 & 0.0032970 \\
7 & 250 & 20.4830 & 21.7954 & 0.0011235 & 0.0034459 \\
8 & 275 & 20.0921 & 22.0538 & 0.0011457 & 0.0036829 \\
9 & 300 & 19.8300 & 22.3655 & 0.0011607 & 0.0039852 \\
\hline
\end{tabular}
\end{center}
These data show the characteristics of the trough in the $\sigma(x,v_{\rm e})$ surface
($x \equiv v_{\rm e}\sin i$) defined by Eq.~2 for each group of Fe~{\sc i} and Fe~{\sc ii} lines.
$x^{*}$ is the $x$ value at the minimum $\sigma(x,v_{\rm e})$ for each given $v_{\rm e}$,
and $\sigma^{*}$ is the corresponding $\sigma(x^{*},v_{\rm e})$. The trace of $x^{*}$ as a
function of $v_{\rm e}$ is shown by the dashed line in the contour plot of Fig.~7.
\end{table}
\subsection{Comparison with previous results}
As mentioned in Sect.~1, although the considerably large differences of Vega's $v_{\rm e}$
amounting to $\ga 100$~km~s$^{-1}$ seen in the literature of early time were reduced
in the more recent results (most of them were published within several years around 2010),
they are still diversified ranging from $\sim 170$ to $\sim 230$~km~s$^{-1}$.
Interestingly, the $v_{\rm e}$ value ($\sim 200$~km~s$^{-1}$) derived in this study
is almost in-between this dispersion.
It may be worth briefly reviewing these literature $v_{\rm e}$ values (published
since Paper~I; cf. Table~1) in comparison with the consequence of this investigation.
\begin{itemize}
\item
Line profile method:\\
Paper~I's result (175~km~s$^{-1}$) based on the conventional profile fitting has been
revised upward by $\sim +20$~km~s$^{-1}$ in this reinvestigation by applying the Fourier
transform method to line profiles.
While Yoon et al.'s (2010) 236~km~s$^{-1}$ is somewhat too large, Hill, Gulliver \&
Adelman's (2010) 211~km~s$^{-1}$ is in tolerable agreement as compared with the preset result.
\item
Interferometry method:\\
Monnier et al.'s (2012) conclusion of $v_{\rm e} = 197$~km~$^{-1}$ (derived from
$v_{\rm e}\sin i$ and $i$ given in their Table~2 as Model~3) based on optical
interferometry is in good agreement with this study. Actually,
Fig.~2 of Monnier et al. (2012) shows that their Model~3 matches well with
model~5 ($v_{\rm e} = 200$~km~s$^{-1}$) of Paper~I.
\item
Magnetic modulation method:\\
Vega's rotational period ($P$) was directly determined by way of detecting the magnetic
modulation based on time-sequence data of spectropolarimetric observations:
0.732~d (Petit et al. 2010), 0.678~d (Alina et al. 2012), 0.623~d (Butkovskaya 2014),
and 0.678~d (B\"{o}hm et al. 2015).
Among these four, it is the $P$ value of 0.678~d derived by both Alina et al. and
B\"{o}hm et al. that is most consistent with the $v_{\rm e}$ result (195~km~s$^{-1}$)
of this investigation, which corresponds to $P = 2\pi R_{\rm e}/v_{\rm e} = 0.685$~d
(where $R_{\rm e} = 2.784$~R$_{\odot}$
for model~5 is adopted).
\end{itemize}
\subsection{Advantage of Fourier analysis}
Finally, some comments may be in order regarding the superiority of exploiting
the zero frequency ($q_{1}$) measured from the Fourier transform of line profiles
in comparison with the ordinary profile fitting approach in the wavelength domain.
The distinct merit of using $q_{1}$ is that it can discern very subtle differences
in the profile shape. Fig.~4 provides a good demonstrative example.
While the profile of Fe~{\sc i} 5133.681 undergoes a comparatively easy-to-detect change
with an increase in $v_{\rm e}$ (Fig.~4a), that of Fe~{\sc ii} 4951.584 is apparently
inert (Fig.~4b), which means that getting information on $v_{\rm e}$
from the profile of the latter is more difficult (this is the reason why Fe~{\sc ii}
lines could not be used for determining $v_{\rm e}$ in Paper~I). However, the situation
is different in the Fourier space, where the shift of $q_{1}$ (reflecting the change of
line profile) is sufficiently detectable with almost the same order of magnitude for
both cases (cf. Figs.~4e and 4f). Accordingly, Fe~{\sc i} as well as Fe~{\sc ii} lines
are equally usable for $v_{\rm e}$ determination if $q_{1}$ is invoked, as done
in this study.
Besides, $q_{1}$ is precisely measurable and easy to handle as a single parameter,
which is a definite advantage from a practical point of view. Actually, $q_{1}$ data
of many lines can be so combined as to improve the precision of $v_{\rm e}$ (while
statistically estimating its error) as done in this paper. Such a treatment would be
difficult in the conventional approach of fitting the observed and theoretical profiles.
\subsection{Line profile classification using $q_{1}$ and $K$}
Another distinct merit of $q_{1}$ is that it provides us with a prospect
for quantitative classification of spectral line shapes founded on a
physically clear basis. Since the discovery around $\sim$~1990 that
a number of spectral lines in Vega (e.g., weak lines of neutral species)
show unusual profiles of square form, there has been a tendency to pay
attention to this specific line group (e.g., compilation of flat-bottomed
lines in Vega by Monier et al. 2017). However, the actual situation of
Vega's spectral lines in general is not so simple as to be dichotomised
into two categories of normal and peculiar profiles; as a matter of fact,
the individual profiles of most lines should more or less have anomalies
of different degree. Unfortunately, detection of such details has been
hardly possible so far, because the judgement of profile peculiarity was
done by simple eye-inspection due to the lack of effective scheme for
describing/measuring the delicate characteristics of line profiles.
The first zero frequency ($q_{1}$) is just what is needed in this context,
which is not only sensitive to a slight difference of line shape but
also easily measurable in the Fourier space. Moreover, thanks to its close
relationship with $K$, the behaviour of $q_{1}$ (representing the line shape
characteristics) can be reasonably explained in terms of the underlying
physical mechanism. We now have a unified understanding as to why different
spectral lines exhibit diversified profiles in Vega, as summarised below.
\begin{itemize}
\item
It is the parameter $K$ (temperature sensitivity) that essentially determines
the observed line shape. The contribution to the important shoulder part of
the profile away from the line centre ($|\Delta \lambda| \la \lambda v_{\rm e}\sin i / c$)
is mainly made by the light coming from near to the
gravity-darkened limb of lowered $T$. Accordingly, lines of $K < 0$, $K \sim 0$,
and $K > 0$ show boxy ({\bf U}-shaped), normally round (like classical rotational
broadening), and rather peaked ({\bf V}-shaped) profiles, each of which result in
appreciably different $q_{1}$ values. For example, in Fig.~4, these three groups
correspond to those of lower $q_{1}$ ($\sim 0.025$~km$^{-1}$s), medium
$q_{1}$ ($\sim 0.03$~km$^{-1}$s), and higher $q_{1}$ ($\sim 0.035$~km$^{-1}$s),
respectively.
\item
The peculiarity degree of the line shape (i.e., departure from the classical
rotationally-broadened profile) is described by $K$, because
$(q_{1} - q_{1}^{\rm classical}) \propto K$
($q_{1}^{\rm classical} \simeq 0.03$~km$^{-1}$s) and the gradient ($>0$)
of this relation progressively increases with $v_{\rm e}$,
as manifested in Fig.~5. As such, the profile of any line in Vega can be
reasonably predicted if $K$ and $v_{\rm e}$ are specified.
\item
As explained in Appendix~A of Paper~II, the value of $K$ for each spectral line
depends upon $\chi_{\rm low}$ (lower excitation potential) and $W$ (equivalent
width). It is important to note that the line strength affects $K$ in the sense
that $|K|$ tends to decrease with an increase in $W$ (i.e., as the line becomes
more saturated), which means that chemical abundances are implicitly involved.
In the present case of A-type stars, $K$ values for Fe~{\sc i} lines are determined
mainly by $W$ while those for Fe~{\sc ii} lines are primarily by $\chi_{\rm low}$
(cf. Fig.~A1 in Paper~II), which are also indicated from Fig.~2c and Fig.~2d.
\item
These behaviours of $K$ in terms of the line parameters reasonably explain why
different spectral lines of Vega reveal various characteristic shapes. For example:
(1) Flat-bottom profiles (manifestation of $K<0$) are seen in Fe~{\sc i} lines
but not in Fe~{\sc ii} lines, because of the distinct difference in $K$ between these
two line groups; i.e., $-20 \la$~$K$(Fe~{\sc i})~$\la -10$ and $-5 \la K$(Fe~{\sc ii})~$\la +5$.
(2) The reason why typical flat-bottomed shape is observed mainly in weak
Fe~{\sc i} lines (e.g., 4707.272, 4903.308 with $W$ of several m\AA) but not
clearly in moderate-strength Fe~{\sc i} lines (e.g., 4202.028, 4920.502 with $W$
of a few tens m\AA) is that the (negative) $K$ values of the former group is
generally lower than those of the latter owing to the dependence upon $W$.
(3) Regarding Fe~{\sc ii} lines, some lines have clearly peaked {\bf V}-shape
(e,g., Fe~{\sc ii}~5004.195 with $\chi_{\rm low}$ = 10.272~eV and $K = +2.92$)
while others exhibit rather rounded profile (e.g., Fe~{\sc ii}~4993.358 with
$\chi_{\rm low}$ = 2.807~eV and $K = -6.27$), which is naturally attributed to the
apparent distinction of $K$ (the sign is inversed) due to the large difference in
$\chi_{\rm low}$.
\end{itemize}
\section{Summary and conclusion}
It is known that the sharp-line star Vega ($v_{\rm e}\sin i \sim 20$~km~s$^{-1}$)
is actually a rapid rotator seen nearly pole-on with low $i$ $(< 10^{\circ})$.
However, its intrinsic rotational velocity is still in dispute,
for which rather diversified values have been published.
In the previous studies (including Paper~I by the author's group), analysis of
spectral line profiles has been often invoked for this purpose, which contain information
on $v_{\rm e}$ via the gravity-darkening effect, However, it is not necessarily
easy to reliably determine $v_{\rm e}$ by direct comparison of observed and
theoretically simulated line profiles. Besides, this approach is not methodologically
effective because it lacks the scope for combining many lines in establishing the solution.
Recently, the author applied in Paper~II the Fourier analysis to the profiles of
many Fe~{\sc i} and Fe~{\sc ii} lines of Sirius~A and estimated its $v_{\rm e}$ by
making use of the first zero ($q_{1}$) of the Fourier transform, which turned out
successful. Therefore, the same approach was decided to adopt in this study to
revisit the task of establishing $v_{\rm e}$ of Vega.
As to the observational data, the same high-dispersion spectra of Vega as adopted
in Paper~I were used. From the Fourier transforms computed from the profiles of
selected 49 Fe~{\sc i} and 41 Fe~{\sc ii} lines, the corresponding zero frequencies
were measured for the analysis. The $K$ values ($T$-sensitivity parameter) of
these Fe lines are in the range of $-20 \la K \la -10$ (Fe~{\sc i} lines) and
$-5 \la K \la +5$ (Fe~{\sc ii} lines).
Regarding the gravity-darkened models of rotating Vega, the model grid (comprising
10 models) arranged in Paper~I was adopted, which cover the $v_{\rm e}$
range of 100--300~km~s$^{-1}$ while assuming $v_{\rm e}\sin i = 22$~km~s$^{-1}$ as fixed.
The theoretical profiles of 90 lines were simulated for each model, from which
Fourier zero frequencies were further evaluated.
An inspection of these $q_{1}^{\rm cal}$ values for the simulated profiles revealed
an increasing tendency with $K$ and the slope of this trend becomes steeper towards
larger $v_{\rm e}$, which suggests that $v_{\rm e}$ is determinable by comparing
$q_{1}^{\rm cal}(K,v_{\rm e})$ with observed $q_{1}^{\rm obs}$ for many lines of different $K$.
It turned out that $v_{\rm e}$ and $v_{\rm e}\sin i$ could be separately established
by the requirement that the standard deviation of the residual between $q_{1}^{\rm cal}$
and $q_{1}^{\rm obs}$ be minimised (while taking into account the difference between
the actual $v_{\rm e}\sin i$ and 22~km~s$^{-1}$ assumed in the model profiles),
and independent analysis applied to two sets of Fe~{\sc i} and Fe~{\sc ii}
lines yielded solutions consistent with each other.
The final parameters of Vega's rotation were concluded to be
$v_{\rm e}\sin i = 21.6 (\pm 0.3)$~km~s$^{-1}$,
$v_{\rm e} = 195 (\pm 15)$~km~s$^{-1}$, and $i = 6.4 (\pm 0.5)^{\circ}$.
\section*{Acknowledgements}
This research has made use of the SIMBAD database, operated by CDS,
Strasbourg, France.
\section*{Data availability}
The data underlying this article are available in the
supplementary materials.
\section*{Supporting information}
Additional Supporting Information may be found in the supplementary materials.
\begin{itemize}
\item
{\bf ReadMe.txt}
\item
{\bf obsparms.dat}
\item
{\bf calparms.dat}
\item
{\bf obsprofs.dat}
\end{itemize}
Please note: Oxford University Press is not responsible for the
content or functionality of any supporting materials supplied by
the authors. Any queries (other than missing material) should be
directed to the corresponding author for the article.
|
\section{Introduction}
\label{sec:intro}
Causal inference draws conclusion about cause and effect relationships through empirical observations. The most widely used statistical framework for causal inference is the \emph{counterfactual} framework. The counterfactual paradigm, also called the \emph{potential outcome} paradigm, was first developed by \cite{Neyman:1923} to study randomized experiments. A generalization allowing the study of causal links with observational data was subsequently carried out by \cite{rubin1974estimating}. Although the process of causal inference is usually complex, it is of extreme importance. In the field of health science, causal inference techniques make it possible to assess the effect of a potential intervention on the health of individuals, in particular in the case of randomized clinical trials. In the field of marketing, such models deal with customers behavioral change caused by a specific treatment, such as a marketing intervention, a courtesy call, targeted advertisement. The counterfactual paradigm assumes that for each individual, there are two potential outcomes, or counterfactuals: i) the potential outcome corresponding to the exposure (treatment), and ii) the potential outcome corresponding to the absence of exposure (control). However, one cannot simultaneously observe treatment and control for a single individual \citep{holland1986statistics}.
In causal inference, the most common population-level estimand is the \emph{average treatment effect}. In the absence of \emph{confounders} (i.e., a variable that influences both the treatment and response variables), this is simply the difference of the two averages between the treatment and control groups. Another estimand include the \emph{individual treatment effect,} also called \emph{conditional average treatment effect} in the context of causal inference, or \emph{uplift} in the context of marketing research. There are different methods to estimate individual treatment effects. The process of estimating these effects has different synonyms like heterogeneous effect modeling, individual treatment effect modeling, or uplift modeling.
Uplift modeling is an important research area in marketing field \citep{radcliffe1999differential, hansotia2002direct, lo2002true, radcliffe2007using}.
This modeling framework has also been proposed to allow for prediction of an individual patient’s response to a medical treatment \citep{jaskowski2012uplift,lamont2018identification}.
Typically, uplift models are developed for randomized experiments, with both the treatment and outcome as binary random variables, where prediction power is the most important issue.
Most common research in the uplift field is based on classification and regression trees (CART) \citep{breiman1984classification}. Unlike other modeling techniques, fitting a decision tree allows each iteration to uniquely partition the sample. This means that each segmentation can be immediately checked against the impact of the treatment. Since the goal of uplift modeling is to find a partition into subgroups of the population, it seems natural to use a decision tree as the method of choice. Then, the idea is to predict the individual uplift by the uplift observed in a terminal node, or by the average when several trees are used, e.g., random forests \citep{Breiman-RandomForest-2001}. However, this is not enough. Indeed, with CART classic division criteria, the search for predictors and their split points is optimized according to the response variable and not the uplift. As practice often shows, response probability and uplift have quite different factors in terms of predictors \citep{radcliffe2011real}. In extreme cases, all uplifts in the leaves might be the same (but not the positive response probabilities). Therefore, various modifications of the split criteria have been proposed in the literature \citep{radcliffe1999differential, hansotia2002direct, rzepakowski2010decision}.
In the case of parametric modeling approaches, the simplest model that can be used to estimate the uplift is logistic regression, because the response variable is binary \citep{lo2002true}. Thus, the underlying optimization problem becomes the maximization of the Binomial likelihood. In this case, the approach does not provide a direct uplift search, but rather the probabilities of positive responses for treated and non-treated are modeled separately. Then, their difference is used as an estimate of the uplift.
However, the solution is not optimized for searching for heterogeneous groups depending on the uplift. Hence, maximizing the likelihood is not necessarily the right way to estimate the uplift. Therefore, changes are required in the optimization problem in order to appropriately estimate the uplift.
In this work, we introduce a new uplift loss function defined by leveraging a connection with the Bayesian interpretation of the relative risk, another treatment effect measure specific to the binary case. Defining an appropriate loss function for uplift also allows to use simple models or any off the shelf method for the related optimization problem, including complex models such as neural networks. When prediction becomes more important than estimation, neural networks become more attractive than classical statistical models. There are several reasons why neural networks are suitable tools for uplift: i) they are flexible models and easy to train with current GPU hardware; ii) with a single hidden layer modeling covariates interactions is straightforward \citep{tsang2018neural}; iii) they are guaranteed to approximate a large class of functions \citep{cybenko1989approximation, hornik1991approximation, pinkus1999approximation}; iv) neural networks perform very well on predictive tasks which is the main objective of uplift; v) a simple network architecture ensures model interpretability for further studies.
Our methodology is developed in the context of a specific neural network architecture in which the proposed loss function can be easily optimized. Our solution uses a representation that resembles the twin networks \citep{bromley1994signature} known in the context of deep learning. This representation helps the fitting process to a great extent. We show our model generalizes Lo's uplift logistic interaction model to a $1$-hidden layer neural network model. In the logistic regression context, one can use \emph{lasso} \citep{Tibshirani_lasso_1996} to produce a sparse model. As more hidden nodes are added to the neural network, lasso becomes choosing the right number of parameters to include in the network (i.e., pruning). However, the fitting mechanism of statistically-driven methods such as the lasso needs to be adapted for stochastic gradient descent to provide a neural network pruning technique.
We propose to use a proximal version of gradient descent and to introduce a scaling factor for structured pruning, which is common in neural networks \citep{ramakrishnan2020differentiable}. This allows to select automatically the number of neurons for each hidden layer. Compared to other existing methods, we thus offer a unified principled approach to obtain a sparse solution that provides well-optimized uplift estimates.
The contributions of this paper are the following: i) defining a new loss function derived from an intuitive interpretation of treatment effects estimation; ii) generalizing the uplift logistic interaction model to a $1$-hidden layer ReLU neural-network; iii) introducing a new twin neural architecture to predict conditional average treatment effects; iv) guiding model selection (or architecture search) with sparse group lasso regularization; v) establishing empirically the validity of the estimation procedure on both synthetic and real-world datasets.
\section{Related work}
\label{sec:notation}
Uplift is a particular case of conditional treatment effect modeling which falls within the potential outcomes framework, also known as the Neyman-Rubin causal model \citep{rubin1974estimating,rosenbaum1983central,holland1986statistics}.
Let $T$ be the binary treatment indicator, and $\boldsymbol X = (X_1,\ldots,X_p)$ be the $p$-dimensional predictors vector. The binary variable $T$ indicates if a unit is exposed to treatment ($T=1$) or control ($T=0$). Let $Y(0)$ and $Y(1)$ be the binary potential outcomes under control and treatment respectively. Assume a distribution $(Y(0),Y(1),\boldsymbol X,T) \sim \mathcal{P}$ from which $n$ iid samples are given as the training observations $\{(y_i, \textbf{x}_i, t_i)\}_{i=1}^n$, where $\mathbf x_i = (x_{i1}, \ldots, x_{ip})$ are realisations of the predictors and $t_i$ the realisation of the treatment for observation $i$. Although each observation $i$ is associated with two potential outcomes, only one of them can be realized as the observed outcome $y_i$. By Assumption~\ref{ass:assumption3}, under the counterfactual consistency, each observation is missing only one potential outcome: the one that corresponds to the absent treatment either $t=0$ or $t=1$.\\
\begin{assumption}
(Consistency) Observed outcome $Y$ is represented using the potential outcomes and treatment assignment indicator as follows:
$$Y=TY(1) + (1-T)Y(0).$$
\label{ass:assumption3}
\end{assumption}
In general, we will assume the following representation of $\mathcal{P}$:
\begin{align*}
&\boldsymbol X \sim \Lambda\\
&T \sim \mathrm{Bernoulli}(e(\mathbf x))\\
&Y(t) \sim \mathrm{Bernoulli}(m_{1t}(\mathbf x))
\end{align*}
where $\Lambda$ is the marginal distribution of $\boldsymbol X$ and $e(\cdot)$ is the propensity score (see Definition~\ref{def:propensity_score}). The probabilities of positive responses for the potential outcomes under control and treatment are given by the functions $m_{1t}(\cdot):\rm I\!R^{p}$ $\rightarrow$ $(0,1)$ for $t=0$ and $t=1$ respectively.
\begin{definition}
(Propensity score) For any $\boldsymbol X=\mathbf x$, the \textit{propensity score} is defined as:
\begin{align}
e(\mathbf x)= \mathrm{Pr}(T_i=1 \mid \boldsymbol X_i = \mathbf x).
\label{def:propensity_score}
\end{align}
\end{definition}
Given the notation above, the conditional average treatment effect (CATE) is defined as follows:
\begin{align}
\mathrm{CATE}(\mathbf x) = \mathbb{E}[Y_i(1) - Y_i(0)| \boldsymbol X_i = \mathbf x]
\label{def:cate}
\end{align}
In order for the CATE to be identifiable, we must make some additional assumptions, standard in the world of causal inference. The propensity score parameter is widely used to estimate treatment effects from observational data \citep{rosenbaum1983central}. Assumption \ref{ass:assumption2} states that each individual has non-zero probabilities of being exposed and being unexposed. This is necessary to make the mean quantities meaningful.
\begin{assumption}
(Overlap) For any $\boldsymbol X=\mathbf x$, the true propensity score is strictly between $0$ and $1$, i.e., for $\epsilon > 0$,
$$\epsilon < e(\mathbf x) < 1-\epsilon.$$
\label{ass:assumption2}
\end{assumption}
For the rest of the paper, we will consider the case of randomized experiments, with $e(\mathbf x)=1/2$, which is common in the uplift literature and is the case of our data. When the data comes from observational studies, combined with the previous assumptions, Assumption \ref{ass:assumption1} allows the identification of the CATE.
\begin{assumption}
(Unconfoundedness) Potential outcomes $Y(0),Y(1)$ are independent ($\independent$) of the treatment assignment indicator $T$ conditioned on all pre-treatment characteristics $\boldsymbol X$, i.e.,
$$Y(0), Y(1) \independent T | \boldsymbol X. $$
\label{ass:assumption1}
\end{assumption}
For randomized experiments, the random variable $ T $ is independent of any pre-treatment characteristics, that is, $Y(0), Y(1), \boldsymbol X \independent T$, which is a stronger assumption than the unconfoundedness assumption.
In the literature, the terms ITE, CATE and uplift often refer to the same quantity, namely the CATE. Indeed, the uplift is a special case of the CATE when the dependent variable $ Y $ is binary $ 0-1 $. Thus, the uplift is defined as the conditional average treatment effect in different sub-populations according to the possible values of the covariates, namely:
\begin{equation}
u(\mathbf x) = \mathrm{Pr}(Y_i=1 \mid \boldsymbol X_i=\mathbf x, T_i=1) - \mathrm{Pr}(Y_i=1 \mid \boldsymbol X_i=\mathbf x, T_i=0).
\label{def:uplift}
\end{equation}
To simplify the notation, we prefer to denote by $m_{yt}(\mathbf x)$ the corresponding conditional probability $\mathrm{Pr}(Y_i=y \mid \boldsymbol X_i=\mathbf x, T_i=t)$. Therefore, the uplift is the difference between the two conditional means $m_{11}(\mathbf x)$ and $m_{10}(\mathbf x)$. The intuitive approach to model uplift is to build two independent models \citep{hansotia2002direct}. This consists of fitting two separate conditional probability models: one model for the treated individuals, and another separate model for the untreated individuals. Then, uplift is the difference between these two conditional probability models. These models are called T-learners (T for ``two models") in the literature \citep{kunzel2019metalearners}. The asset of T-learners is their simplicity, but they do not perform well in practice, because each model focuses on predicting only one class, so the information about the other treatment is never provided to the learning algorithm \citep{radcliffe2011real}. In addition, differences between the covariates distributions in the two treatment groups can lead to bias in treatment effect estimation. There has been efforts in correcting such drawbacks through a combined classification model known as S-learner, for ``single-model" \citep{kunzel2019metalearners}. The idea behind the S-learner is to use the treatment variable as a feature and to add explicit interaction terms between each covariate and the treatment indicator to fit a model, e.g., a logistic regression \citep{lo2002true}. The parameters of the interaction terms measure the additional effect of each covariate due to treatment.
Another related method is known as the X-learner \citep{kunzel2019metalearners}. The X-learner estimates the uplift in three stages. First, $m_{11}(\mathbf x)$ and $m_{10}(\mathbf x)$ are modeled separately as in the case of T-learners. Then, the fitted values $\hat m_{11}(\mathbf x)$ and $\hat m_{10}(\mathbf x)$ are used to impute the ``missing" potential outcomes for each observation, and to create new imputed response variables, $D(1)$ and $D(0)$. These imputed variables are used to fit new models that capture the uplift directly, $\hat u^{(1)}(\mathbf x)$ and $\hat u^{(0)}(\mathbf x)$. The final prediction is given by a weighted average using the propensity score, $e(\mathbf x)\{\hat u^{(1)}(\mathbf x) - \hat u^{(0)}(\mathbf x)\} + \hat u^{(0)}(\mathbf x)$. More recently, \cite{NieWager2020Quasi} introduced the R-learner. The method uses Robinson's transformation \citep{robinson1988root} and assumes ``oracle" estimation of the propensity score $e(\mathbf x)$ and the marginal effect function $m(\mathbf x) = \mathbb{E}[Y|\boldsymbol X=\mathbf x]$ in order to reduce the problem of modeling the uplift to a residual-on-residual ordinary least squares regression. Interestingly, even though the X-learner and R-learner frameworks are valid for continuous and binary $ Y $, both methods aim to find an uplift estimator that minimizes the mean squared error (MSE). The connection with the maximum likelihood framework does not hold in the binary case. Also, in practice, the R-learner is fitted in two stages using cross-fitting to emulate the oracle, thus requiring more observations than S-learners.
The methodology introduced in \cite{belba2019qbased} attacks parameters estimation and addresses the loss-metric mismatch in uplift regression. It can be seen as an S-learner which estimates the model's parameters in two stages. The method first applies the pathwise coordinate descent algorithm \citep{friedman2007pathwise} to compute a sequence of critical regularization values and corresponding (sparse) model parameters. Then, it uses Latin hypercube sampling \citep{mckay2000comparison} to explore the parameters space in order to find the optimal model.
Several proposed non-parametric methods take advantage of grouped observations in order to model the uplift directly. Some $k$-nearest neighbours \citep{cover1967nearest} based methods are adopted for uplift estimation \citep{crump2008nonparametric, Alemi.etal-PersonalizedMedicine-2009, su2012facilitating}. The main idea is to estimate the uplift for an observation based on its neighbourhood containing at least one treated and one control observations. However, these methods quickly become computationally expensive for large datasets.
State-of-the-art proposed methods view random forests as an adaptive neighborhood metric, and estimate the treatment effect at the leaf node \citep{su2009subgroup,chipman2010bart,wager2018estimation}. Therefore, most active research in uplift modeling is in the direction of classification and regression trees \citep{breiman1984classification} where the majority are modified random forests \citep{Breiman-RandomForest-2001}.
In \cite{radcliffe1999differential, radcliffe2011real,hansotia2002direct, rzepakowski2010decision}, modified split criteria that suited the uplift purpose were studied. The criteria used for choosing each split during the growth of the uplift trees is based on maximization of the difference in uplifts between the two child nodes. Within each leaf, uplift is estimated with the difference between the two conditional means. A good estimate of each mean may lead to a poor estimate of the difference \citep{radcliffe2011real}. However, the existing tree-based uplift optimization problems do not take this common misconception into account. Instead, the focus is on maximizing the heterogeneity in treatment effects. Without careful regularization (e.g., honest estimation \citep{athey2019generalized}), splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. In addition, successive splits tend to group together similar extreme values, introducing more variance in the prediction of uplift \citep{zhao2017practically}. Alternatively some models use the transformed outcome \citep{athey2015machine}, an unbiased estimator of the uplift. However, this estimate suffers from higher variance than the difference in conditional means estimator \citep{powers2018some}. In addition, for both estimators, if random noise is larger than the treatment effect, the model will more likely predict random noise instead of uplift. As a result, based on several experiments on real data, and although the literature suggests that tree-based methods are state-of-the-art for uplift \citep{soltys2015ensemble}, the published models overfit the training data and predicting uplift still lacks satisfactory solutions.
\section{An uplift loss function}
We formally define the uplift loss function that will be used to fit our models. Our goal is to regularize the conditional means in order to get a better prediction of the quantity of interest, the uplift. Inspired by the work of \cite{athey2019generalized,kunzel2019metalearners,belba2019qbased,NieWager2020Quasi} which adapt the optimization problem to the uplift context, we propose to define a composite loss function, which can be separated into two pieces: $$\ell(\cdot) = \ell_1(\cdot) + \ell_2(\cdot)$$ and to optimize both simultaneously. Since we generalize the uplift logistic regression, we model the probability of positive response $m_{1t}(\mathbf x)$. Naturally, the first term can be defined as the negative log-likelihood or the binary cross entropy (BCE) loss, with $\mathbf y$ as the response, and $m_{1t}(\mathbf x)$ as the prediction, that is, $$ \ell_1(\mathbf y, \mathbf t \mid \mathbf x) = - \frac{1}{n} \sum_{i=1}^n \Big(y_i \log\{ m_{ 1 t_i}(\mathbf x_i)\} + (1 - y_i)\log\{1 - m_{1 t_i }(\mathbf x_i)\} \Big).$$
We define the second term based on a Bayesian interpretation of another measure of treatment effect, the relative risk. First, let us define the relative risk (or risk ratio) as a function of the conditional means.\\
\begin{definition}
(Relative risk) For any $\boldsymbol X = \mathbf x$ and $m_{10}(\mathbf x) > 0$, the relative risk is defined as follows:
\begin{align*}
\mathrm{RR}(\mathbf x) = \frac{\mathrm{Pr}(Y=1 \mid \boldsymbol X = \mathbf x, T=1)}{\mathrm{Pr}(Y=1 \mid \boldsymbol X = \mathbf x, T=0)} = \frac{m_{11}(\mathbf x)}{m_{10}(\mathbf x)},
\end{align*}
\end{definition}
Relative risk is commonly used to present the results of randomized controlled trials. In the medical context, the uplift is known as the absolute risk (or risk difference). In practice, presentation of both absolute and relative measures is recommended \citep{moher2012consort}. If the relative risk is presented without the absolute measure, in cases where the base rate of the outcome $m_{10}(\mathbf x)$ is low, large or small values of relative risk may not translate to significant effects, and the importance of the effects to the public health can be overestimated. Equivalently, in cases where the base rate of the outcome $m_{10}(\mathbf x)$ is high, values of the relative risk close to 1 may still result in a significant effect, and their effects can be underestimated. Interestingly, the relative risk can be reformulated as:
\begin{align*}
\mathrm{RR}(\mathbf x) = \frac{\mathrm{Pr}(T=1 \mid Y=1, \boldsymbol X = \mathbf x)}{\mathrm{Pr}(T=0 \mid Y=1, \boldsymbol X = \mathbf x)} \biggl(\frac{1-e(\mathbf x)}{e(\mathbf x)}\biggr),
\end{align*}
where the propensity score $e(\mathbf x)$ is given in Definition~\ref{def:propensity_score}. For randomized experiments, the propensity score ratio $\{1-e(\mathbf x)\}/e(\mathbf x)$ is a constant and, written in that form, the relative risk can be interpreted in Bayesian terms as the normalized posterior propensity score ratio (i.e., after observing the outcome). In the particular case where $e(\mathbf x)=1/2$, it is easy to show that $\mathrm{Pr}(T=1 \mid Y=1, \boldsymbol X = \mathbf x) = \mathrm{RR}(\mathbf x)/\{1+\mathrm{RR}(\mathbf x)\}$. Moreover, we have the following equalities:
\begin{align}
\mathrm{Pr}(T=1 \mid Y=1, \boldsymbol X = \mathbf x) &= \frac{m_{11}(\mathbf x)}{m_{11}(\mathbf x)+m_{10}(\mathbf x)}, \\
\mathrm{Pr}(T=1 \mid Y=0, \boldsymbol X = \mathbf x) &= \frac{m_{01}(\mathbf x)}{m_{01}(\mathbf x) + m_{00}(\mathbf x)}.
\end{align}
These two equalities give a lot of information. Without abuse of language, we call them \textit{posterior propensity scores} and we denote by $p_{yt} (\mathbf x)$ the corresponding conditional probability $\mathrm{Pr}(T=t \mid Y=y, \boldsymbol X = \mathbf x)$. The posterior propensity scores are functions of the conditional means.
The quantity $p_{11}(\mathbf x)$ can be seen as the proportion of treated observations among those that had positive outcomes and $p_{01}(\mathbf x)$ can be seen as the proportion of treated observations among those that had negative outcomes. %
We define the second term of our uplift loss as the BCE loss, but this time, using the observed treatment indicator $\mathbf t$ as the ``response" variable, and $p_{y1}(\mathbf x)$ as the ``prediction". Formally, it is given by: $$ \ell_2(\mathbf t, \mathbf y \mid \mathbf x) = -\frac{1}{n} \sum_{i=1}^n \Big(t_i \log\{p_{y_i 1}(\mathbf x_i)\} + (1 - t_i)\log\{1 - p_{y_i 1}(\mathbf x_i)\} \Big).$$
Taken alone, the second loss models the posterior propensity scores as a function of the conditional means (for positive and negative outcomes). Intuitively, if a treatment has a significant positive (resp. negative) effect on a sub-sample of observations, then within the sample of observations that had a positive (resp. negative) response, we expect a higher proportion of treated. Formally, we define the complete uplift loss function in Definition~\ref{def:augmented_loss}.\\
\begin{definition}\label{def:augmented_loss}
Let $m_{yt} \stackrel{\text{def}}{=} m_{yt}(\mathbf x) = \mathrm{Pr}( Y=y |\boldsymbol X =\mathbf x, T=t)$, and $p_{yt} \stackrel{\text{def}}{=} p_{yt} (\mathbf x) = m_{yt} / ( m_{y1} + m_{y0}) $. We define the uplift loss function as follows:
\begin{equation}
\ell(\mathbf y , \mathbf t \mid \mathbf x) = - \frac{1}{n} \sum_{i=1}^n \{y_i \log m_{ 1 t_i} + (1 - y_i)\log m_{0t_i } + t_i \log p_{y_i 1} + (1 - t_i)\log p_{y_i 0} \} .
\label{eq:augmented_loss_function}
\end{equation}
\end{definition}
Although we considered the case where $e(\mathbf x)=1/2$, the development holds for any constant $e(\mathbf x)$. An under-sampling or an over-sampling procedure allows to recover the $e(\mathbf x) = 1/2$ if the constant is below or above $1/2$ respectively. Interestingly, it is possible to find a connection between the uplift loss function \eqref{eq:augmented_loss_function} and the likelihood of the data. Indeed, the described relation between the relative risk and the conditional probabilities $m_{11}(\mathbf x)$ and $m_{10}(\mathbf x)$, as well as the Bayesian interpretation of the posterior propensity scores $p_{yt}(\mathbf x)$ suggest modeling the joint distribution of $Y$ and $T$. Formally, the connection can be shown through the following development.
\begin{align*}
\mathrm{Pr}(Y=y, T=t \mid \mathbf X=\mathbf x) &= \mathrm{Pr}(T=t \mid Y=y, \mathbf X=\mathbf x) \mathrm{Pr}(Y=y\mid \mathbf X=\mathbf x) \\
& = p_{yt}(\mathbf x) \{ m_{y1}(\mathbf x) e(\mathbf x) + m_{y0}(\mathbf x) [1-e(\mathbf x)]\} \\
& = p_{yt}(\mathbf x) \{ m_{y1}(\mathbf x) + m_{y0}(\mathbf x)\} / 2,
\end{align*}
because $e(\mathbf x)=1/2$. Therefore, the likelihood for $n$ observations is proportional to
\begin{align*}
\prod_{i=1}^n p_{y_i 1}^{t_i} p_{y_i 0}^{(1-t_i)} \{m_{11} + m_{10}\}^{y_i} \{m_{01} + m_{00}\}^{(1-y_i)},
\end{align*}
and the log-likelihood is proportional to
\begin{equation}
\sum_{i=1}^n \{y_i \log(m_{11}+m_{10}) + (1 - y_i)\log(m_{01} + m_{00}) + t_i \log p_{y_i 1} + (1 - t_i)\log p_{y_i 0} \}.
\label{eq:log-likelihood_loss_function}
\end{equation}
Notice that the functions \eqref{eq:augmented_loss_function} and \eqref{eq:log-likelihood_loss_function} differ only in that \eqref{eq:augmented_loss_function} uses $m_{1t}$ while \eqref{eq:log-likelihood_loss_function} specifically uses $m_{y1}$ and $m_{y0}$, the conditional means under treatment and control. Traditionally, $m_{1t}$ is more common since in practice, each observation can only be treated or not treated. However, we compared the results by fitting uplift models using both functions. The results being very similar, in the rest of the paper, we only present results for models fitted with the augmented loss function \eqref{eq:augmented_loss_function}. We keep the in-depth analysis of the log-likelihood \eqref{eq:log-likelihood_loss_function} for future work.
The loss function \eqref{eq:augmented_loss_function} can also be interpreted term by term. The first term is simply the binary cross entropy loss w.r.t the conditional means. The second term can be seen as a regularization term on the conditional means. In the second term, the conditional means are represented through the posterior propensity scores. By minimizing the augmented loss, the first term focuses on estimating the conditional means separately while the second term tries to correct for the posterior propensity scores. Since both terms are minimized simultaneously, this can also be seen as a special case of multi-task learning. As we will show later, this new parameter estimation method greatly improves the predictive performance of the underlying uplift models.
\section{A twin neural model for uplift}
Let's start with the uplift interaction model \citep{lo2002true} as a simple preliminary model. This model is based on logistic regression. It is common to add explicit interaction terms between each explanatory variable and the treatment indicator. The parameters of the interaction terms measure the additional effect of each covariate due to treatment. These interactions are important for estimating individual effects since this is what makes it possible to create heterogeneity in the treatment effects.
Logistic regression may be visualised in a model diagram, with a single node to represent the link function, and multiple nodes to represent inputs or outputs. This sort of visualization is very common in neural networks community (see Figure~\ref{fig:int_as_nn}, left panel).
The uplift interaction model can be represented by a fully-connected neural network with no hidden layer, an intercept, $2p+1$ input neurons (covariates, treatment variable and interaction terms) and $1$ output neuron with sigmoid activation function, where $\sigma(z)=1/(1+e^{-z})$, for $z \in \rm I\!R$ (see Figure~\ref{fig:int_as_nn} left panel). Let $\mathbf x = (x_1,...,x_p) \in \rm I\!R^{p}$ be the covariates vector and $t \in \{0,1\}$, a binary variable. Let us further define $\theta_{j}$, for $j=1,\ldots,2p+1$, the coefficient or weight that connects the $j$th input neuron to the output and let $\theta_o \in \mathbb{R}$ be the intercept.
The uplift interaction model can be written as
\begin{equation}
\mu_{1t}(\mathbf x, \boldsymbol{\theta}) = \sigma \Big(\theta_o + \sum_{j=1}^{p} \theta_j x_j + \sum_{j=p+1}^{2p} \theta_j t x_{j-p} + \theta_{2p+1} t \Big), ~~ t \in \{0,1\},
\label{eq:neural_network_1}
\end{equation}
where $\sigma(\cdot)$ represents the sigmoid function and $\boldsymbol{\theta}$ denotes the vector of model parameters. The predicted uplift associated with the covariates vector $\mathbf{x}_{n+1}$ of a future individual is
$$ \hat{u}(\mathbf{x}_{n+1}) = \mu_{11}(\mathbf x_{n+1}, \boldsymbol{\hat{\theta}}) - \mu_{10}(\mathbf x_{n+1}, \boldsymbol{\hat{\theta}}),$$
where $\boldsymbol{\hat{\theta}}$ may be estimated by minimizing a loss function such as the one defined in Equation \eqref{eq:augmented_loss_function}.
\begin{figure}[!ht]
\centering
\input{figures/tikz_int_1hrelu}
\caption{Graphical representation of the uplift interaction model (left panel). Neural network representation of the uplift model (right panel). Note that $h_k = \mathrm{ReLU}(\mathbf x,t)$ for $k=1,\ldots,m$. For $m=2p+1$, Theorem~\ref{theorem} shows the equivalence between the interaction model \eqref{eq:neural_network_1} and a particular case of the $1$-hidden layer neural network \eqref{eq:neural_network_2} with $\mathrm{ReLU}$ activation. In both diagrams, the gray node represents the sigmoid activation function.}
\label{fig:int_as_nn}
\end{figure}
More generally, let $\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta})$ for $t \in \{0,1\}$ be a neural network. We denote by $\mathrm{NN}_{11}(\mathbf x, \boldsymbol{\theta})$ and $\mathrm{NN}_{10}(\mathbf x, \boldsymbol{\theta})$ the conditional mean model for treated and control observations respectively. In what follows, our goal is to generalize the interaction model \eqref{eq:neural_network_1} by a more flexible neural network. We focus on a fully-connected network with an input of size $p+1$ (covariates and treatment variable), and one hidden layer of size $m>1$ with ReLU activation, where $\mathrm{ReLU}(z) = \max\{0,z\}$, for $z \in \rm I\!R$. We assume that the intercept (also called bias term) is inherent in the neural model. The hidden layer is then connected to a single output neuron with a sigmoid activation (see Figure~\ref{fig:int_as_nn} right panel). The output of the neural network $\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta})$ can be written as
\begin{equation}
\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta}) = \sigma \biggl\{ \theta_o^{(2)} + \sum_{k=1}^{m} \theta_k^{(2)} \mathrm{ReLU}\Big(\theta_{o,k}^{(1)} + \sum_{j=1}^{p} \theta_{j,k}^{(1)} x_j + \theta_{p+1,k}^{(1)} t \Big) \biggr\}, ~~ t \in \{0,1\},
\label{eq:neural_network_2}
\end{equation}
where $\theta_{j,k}^{(1)}$ represent the coefficient or weight that connects the $j$th covariate or input neuron to the $k$th hidden neuron and $\theta_k^{(2)}$ represents the coefficient that connects the $k$th hidden neuron to the output. We denote the bias terms for the hidden layer and the output layer by $\theta_{o,k}^{(1)}$, $k=1,\ldots,m$ and $\theta_o^{(2)}$ respectively. Here, $\boldsymbol{\theta}$ contains all of the neural network's coefficients (or parameters). The predicted uplift associated with the covariates vector $\mathbf{x}_{n+1}$ of a future individual is
$$ \hat{u}(\mathbf{x}_{n+1}) = \mathrm{NN}_{11}(\mathbf x_{n+1}, \boldsymbol{\hat{\theta}}) - \mathrm{NN}_{10}(\mathbf x_{n+1}, \boldsymbol{\hat{\theta}}),$$
where $\boldsymbol{\hat{\theta}}$ may be estimated by minimizing a loss function such as the one defined in Equation \eqref{eq:augmented_loss_function}. In the following Theorem, we show that for a judicious choice of the neural network's coefficients matrix, the two models are equivalent.\\
\begin{theorem} \label{theorem}
Let $\mu_{1t}(\mathbf x)$ and $\mathrm{NN}_{1t}(\mathbf x)$ be two uplift models defined as in \eqref{eq:neural_network_1} and \eqref{eq:neural_network_2} respectively. Let $c \in \rm I\!R^+$ be a positive and finite constant and $m=2p+1$. For all $\theta_j \in \rm I\!R$, $j=1,\ldots,2p+1$ and $\theta_o \in \rm I\!R$, there exists a matrix of coefficients $\Big(\theta_{j,k}^{(1)}\Big) \in \rm I\!R^{(p+1) \times (2p+1)}$, such as
\[ \Big(\theta_{j,k}^{(1)}\Big) =
\left(
\begin{array}{cccc|cccc|c}
1 & 0 & \cdots & 0 & 1 & 0 & \cdots & 0 & 0 \\
0 & 1 & \cdots & 0 & 0 & 1 & \cdots & 0 & 0 \\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & \cdots & 1 & 0 & 0 & \cdots & 1 & 0 \\
\hline
0 & 0 & \cdots & 0 & c & c & \cdots & c & 1 \\
\end{array}
\right),
\]
an intercepts vector $\Big(\theta_{o,k}^{(1)}\Big) \in \rm I\!R^{2p+1}$, a vector of coefficients $\Big(\theta_{k}^{(2)}\Big) \in \rm I\!R^{2p+1}$ and an intercept scalar $\theta_o^{(2)} \in \rm I\!R$ such that for all $\mathbf x \in [0,c]^p$ and $t \in \{0,1\}$
\begin{equation}
\mu_{1t}(\mathbf x) = \mathrm{NN}_{1t}(\mathbf x).
\label{eq:theorem_equality}
\end{equation}
\end{theorem}
\begin{proof}
See Appendix \ref{sec:appendix_thoerem_nn_mu}.
\end{proof}
This theorem is interesting since the neural network model is much more flexible and can be seen as a generalization of the interaction model. As we will see in the experimental Section, this flexibility allows a better fit resulting in a higher performance from a prediction point of view.
\paragraph{A twin representation of uplift models.}
For most existing parametric uplift methods, the uplift prediction is computed in several steps: i) the uplift model is fitted; ii) the conditional probabilities are predicted by fixing the treatment variable $T$ to $1$ or $0$; iii) the difference is taken to compute the uplift; iv) the uplift is visualized. The fitted model plays a major role in implementing each of these steps. This can be problematic when the fitted model overfit the data at hand. Therefore, most multi-steps methods require careful regularization. To simplify this task, we propose to combine the whole process into a single step through a twin model. The twin interaction model diagram is visualized in Figure~\ref{fig:siamese_network}. This representation resembles the twin networks \citep{bromley1994signature} in the context of deep learning. Such networks were first introduced in signature verification as an image matching problem, where the task is to compare two hand-written signatures and infer the identity of the writer.
\begin{figure}[!ht]
\centering
\input{figures/tikzsmite}
\caption{Diagram of the twin logistic interaction model. The original interaction model is separated into two sub-components with the same parameters. For the left sub-component, the treatment variable fixed to $1$ and the interaction terms to $\mathbf x$. The treatment variable and the interactions terms are fixed to $0$ for the right sub-component. The sub-components model the conditional means for treated ($\mu_{11}(\mathbf x)$) and for control ($\mu_{10}(\mathbf x)$). The difference gives direct prediction of $u(\mathbf x)$. At the same time, the predicted conditional mean $\mu_{1t}(\mathbf x)$ is based on the actual received treatment for each individual $t \in \{0,1\}$, i.e., $\mu_{1t}(\mathbf x) = t \mu_{11}(\mathbf x) + (1-t) \mu_{10}(\mathbf x)$.}
\label{fig:siamese_network}
\end{figure}
A twin neural network consists of two models that use the same parameters (or weights) while fitted in parallel on two different input vectors to compute comparable outputs. In our case, the input vectors are almost identical, only the treatment variable is changed. The parameters between the twin networks are shared. Weight sharing guarantees that two individuals with similar characteristics are mapped similarly by their respective networks because each network computes the same function. Such networks are mostly known for their application in face recognition \citep{chopra2005learning,taigman2014deepface,parkhi2015deep,schroff2015facenet}, areal-to-ground image matching \citep{lin2015learning}
and large scale video classification \citep{karpathy2014large
, among others.
The twin network representation of the uplift interaction model is easily generalized to neural networks, as show in Figure~\ref{fig:nite}. In our case, instead of comparing two distinct images, we duplicate each observation and fix the treatment variable to $1$ and $0$, while keeping in memory the true value $t$ of the treatment variable. This makes it possible to fit the twins in parallel and to use the true value $t$ to compute $\mathrm{NN}_{1t}(\mathbf x)$. The availability of $\mathrm{NN}_{11}(\mathbf x)$ and $\mathrm{NN}_{10}(\mathbf x)$ allows to compute the uplift loss function \eqref{eq:augmented_loss_function} and to predict the uplift $u(\mathbf x)$ in a single step, which simplifies the fitting process to a great extent.
\input{figures/nnsmite}
\section{Parameter estimation}
Loss functions are generally convex with respect to $\mathrm{NN}(\cdot)$, but not with respect to the model's parameters $\boldsymbol\theta$. Parameter estimation is difficult, and with neural networks, in the best cases we are looking for a good local minimum. Most of the methods are based on some form of gradient descent. Gradient-based optimization is one of the pillars of machine learning. Given the loss function $\ell: \rm I\!R \rightarrow \rm I\!R$, classical gradient descent has the goal of finding (local) minima $\boldsymbol{\hat{\theta}} = \argmin\limits_{\boldsymbol{\theta}} \ell(\boldsymbol{\theta})$ via updates of the form $ \Delta (\boldsymbol{\theta}) = -\eta \nabla \ell$, where $\eta$ is a positive step size (or the learning rate).
The prototypical stochastic optimization method is the stochastic gradient
method \citep{robbins1951stochastic}, which, in the context of minimizing $\ell(\boldsymbol{\theta})$ with a given starting point $\boldsymbol{\theta}^{(0)}$ is
\begin{equation}
\boldsymbol{\theta}^{(q+1)} \leftarrow \boldsymbol{\theta}^{(q)} - \eta \nabla \ell_{i_q} \left(\boldsymbol{\theta}^{(q)}\right),
\end{equation}
where $\ell_{i_q}(\cdot)$ is the loss function evaluated using observation $i_q\in \{1,...,n\}$ which is chosen at random. Here we use the term stochastic in the sense that for each parameters update, only a random sample of the data is used. This method is to be distinguished from a gradient descent method using a stochastic learning rate $\eta_q$.
There is no particular reason to employ information from only one observation per iteration. Instead, one can employ a {\it mini-batch} approach in which a small subset of observations $\mathcal{S}_q \subseteq \{1, ..., n\}$, is chosen randomly in each iteration, leading to
\begin{equation}
\boldsymbol{\theta}^{(q+1)} \leftarrow \boldsymbol{\theta}^{(q)} - \frac{\eta}{|\mathcal{S}_q|} \sum_{i \in \mathcal{S}_q} \nabla \ell_{i} (\boldsymbol{\theta}^{(q)}),
\label{eq:mini-batch-sgd}
\end{equation}
where $\ell_{i}(\cdot)$ is the loss function evaluated using observation $i$ and $|\mathcal{S}_q|$ is the batch size. Such a mini-batch stochastic gradient method has been widely used in practice. There are some fundamental practical and theoretical reasons why stochastic methods have inherent advantages for large-scale machine learning \citep{bottou2018optimization}. Note that optimization algorithms that use only the gradient are called first-order optimization algorithms. Second-order optimization algorithms such as Newton's method use the Hessian matrix. Like all neural network training, we estimate the neural network's parameters using first-order gradient methods.
Neural networks are typically over-parametrized and are prone to overfitting. Hence, regularization is required. Typically the regularization term is a mixture of $L_2$ (Ridge) and $L_1$ (lasso) norms, each with its own regularization constant. Similar to classical regression applications, the intercepts are not penalized. The regularization constants are typically small and serve several roles. The $L_2$ reduces collinearity, and $L_1$ ignores irrelevant parameters, and both are a remedy to the overfitting, especially in over-parametrized models such as deep neural networks \citep{efron2016computer}.
\subsection{Unstructured sparsity}
The parameter estimation process optimizes the uplift loss, so in the interaction model
\begin{equation}
\boldsymbol{\hat{\theta}} = \argmin_{\boldsymbol{\theta}} \ell(\boldsymbol{\theta}) +\lambda \mathcal R(\boldsymbol{\theta}),
\label{eq:optim_uplift_interact}
\end{equation}
where $\ell(\boldsymbol{\theta})$ is the loss, $\lambda \in \rm I\!R^+$ is the regularization constant and $\mathcal R(\boldsymbol{\theta})$ is often convex and probably non-differentiable such as $\norm{\boldsymbol{\theta}}_1$.
The stochastic gradient descent (SGD) for the lasso needs to be modified to make sure the solution remains sparse, so SGD needs to be adapted using the projected gradient update (see \cite{mosci2010solving} for more details). Proximal gradient methods are a generalized form of projection used to solve non-differentiable convex optimization problems. For the reader wishing to have more knowledge on proximal operators, we suggest \cite[Chapter 6]{beck2017first}.
Let's focus only on a univariate $\theta_j$. We adjust SGD by splitting $\theta_j$ into two positive components, $u_j=\theta_j^+$ and $v_j=\theta_j^-$. Here, $\theta_j^+$ and $\theta_j^-$ are the positive and negative parts of $\theta_j \in \rm I\!R$ so $\theta_j= u_j-v_j$ in which $u_j,v_j\geq 0$ and of course $|\theta_j| = u_j+v_j$. The optimization problem \eqref{eq:optim_uplift_interact} can be reformulated as
$$\argmin_{\mathbf u,\mathbf v} \ell(\mathbf u-\mathbf v ) +\lambda \mathbf 1^\top(\mathbf u+\mathbf v).$$
Let's focus only on a single update of the parameter $\theta_j=u_j-v_j$ and a single batch size.
The optimization routine ensures $u_j\geq 0, v_j\geq 0$, by passing the optimizing parameters $\mathbf u, \mathbf v$ through the proximal projection $\mathrm{ReLU}(\cdot)$ to ensure non-negativity.
Our proximal gradient descent has only two basic steps which are iterated until convergence. For given $u_j^{(0)}, v_j^{(0)}$ the modified SGD is
\begin{enumerate}
\item \emph{gradient step}: define intermediate points $\Tilde{u}_j^{(q)}, \Tilde{v}_j^{(q)}$ by taking a gradient step such as
\begin{align}
\Tilde{u}_j^{(q)} &= u_j^{(q)} - \eta \{ \lambda + \nabla \ell_{i_q}(\theta_j^{(q)})\}\nonumber\nonumber\\
\Tilde{v}_j^{(q)} &= v_j^{(q)} - \eta \{ \lambda - \nabla \ell_{i_q}(\theta_j^{(q)})\}
\label{eq:lasso_prox_update1}
\end{align}
\item \emph{projection step}: evaluate the proximal operator at the intermediate points $\Tilde{u}_j^{(q)}, \Tilde{v}_j^{(q)}$ such as
\begin{align}
u_j^{(q+1)} &\leftarrow \mathrm{ReLU}(\Tilde{u}_j^{(q)})\nonumber\\
v_j^{(q+1)} &\leftarrow \mathrm{ReLU}(\Tilde{v}_j^{(q)})\nonumber\\
\theta_j^{(q+1)} &\leftarrow u_j^{(q+1)} - v_j^{(q+1)}
\label{eq:lasso_prox_update2}
\end{align}
\end{enumerate}
Exact zero weight updates appear when $u_j<0$ and $v_j<0$ to enable unstructured sparsity.
\subsection{Structured sparsity}
Group lasso is a generalization of the lasso method, when features are grouped into disjoint blocks with a total of $G<p$ groups \citep{yuan2006model}. The formulation of group lasso allows us to define what may constitute a suitable group. For pruning the entire hidden nodes, it is enough to define the block to be the weights that define each node. Take the $1$-hidden layer uplift neural network $\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta})$ defined in \eqref{eq:neural_network_2}. The neural model can be rewritten as
\begin{equation}
\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta}) = \sigma \biggl\{ \theta_o^{(2)} + \sum_{k=1}^{m} \theta_k^{(2)} \mathrm{ReLU}\Big(s_k \{\theta_{o,k}^{(1)} + \sum_{j=1}^{p} \theta_{j,k}^{(1)} x_j + \theta_{p+1,k}^{(1)} t \} \Big) \biggr\},
\label{eq:neural_network_2_scaling_factor}
\end{equation}
where here, we introduced a scaling factor $s_k \in \rm I\!R$ for $k=1,\ldots,m$ (i.e., one scaling factor per hidden node). Introduction of a scaling factor for structured pruning is common in neural networks, see for instance \cite{ramakrishnan2020differentiable}.
Let $\mathbf s = (s_1, \ldots, s_m)$ be the $m$-dimensional scaling factors vector. We propose the following optimization problem to enforce structured sparsity in the node level
$$\argmin_{\mathbf u, \mathbf v, \mathbf a, \mathbf b} \ell(\mathbf u, \mathbf v, \mathbf a, \mathbf b) + \lambda_1 \mathbf 1^\top({\mathbf a + \mathbf b}) + \lambda_2 \mathcal R(\mathbf u-\mathbf v),$$
where $\lambda_1, \lambda_2 \in \rm I\!R^+$ are the regularization constants and $
\lambda_1 $ controls the amount of structured sparsity. This formulation allows to use a similar lasso proximal SGD development using the introduced scaling factor $s_k$. So we define $s_k= a_k-b_k$, for $k=1,\ldots,m$ in which $a_k,b_k\geq 0$ so that $|s_k| = a_k+b_k$ and $\lambda_1 \norm{\mathbf s}_1 = \lambda_1 \mathbf 1^\top(\mathbf a+\mathbf b)$. This results in the following modified SGD updates for
given $a_k^{(0)}, b_k^{(0)}$ in addition to the lasso proximal updates \eqref{eq:lasso_prox_update1} and \eqref{eq:lasso_prox_update2}
\begin{enumerate}
\item \emph{gradient step}: define intermediate points $\Tilde{a}_k^{(q)}, \Tilde{b}_k^{(q)}$ by taking the gradient step
\begin{align*}
\Tilde{a}_k^{(q)} &= a_k^{(q)} - \eta \{ \lambda_1 + \nabla \ell_{i_q}(s_k^{(q)})\}\\
\Tilde{b}_k^{(q)} &= b_k^{(q)} - \eta \{ \lambda_1 - \nabla \ell_{i_q}(s_k^{(q)})\}
\end{align*}
\item \emph{projection step}: evaluate the proximal operator at the intermediate points $\Tilde{a}_k^{(q)}, \Tilde{b}_k^{(q)}$ such as
\begin{align*}
a_k^{(q+1)} &\leftarrow \mathrm{ReLU}(\Tilde{b}_k^{(q)})\\
b_k^{(q+1)} &\leftarrow \mathrm{ReLU}(\Tilde{a}_k^{(q)})\\
s_k^{(q+1)} &\leftarrow a_k^{(q+1)} - b_k^{(q+1)}
\end{align*}
\end{enumerate}
Pruning the $k$th node happens when $a_k<0$ and $b_k<0$ which enables structured sparsity. This structured sparsity yields an automatic selection of the number of hidden nodes $m$.
\section{Model Evaluation}\label{sec:qini}
In the context of model selection, in practice, given $L$ models and/or hyperparameter settings, we build $L$ estimators $\{ \hat{u}_1, \hat{u}_2, \ldots, \hat{u}_L \}$. We intend to maximize expected prediction performance, using some goodness-of-fit statistic. There are several ways to evaluate prediction performance. However, data-splitting to \emph{training data} and \emph{validation data} is the most widely used in practice \citep{arlot2010survey}. Before fitting the models, the observations are randomly split into training samples $\mathcal T$and validation samples $\mathcal V$. All models are fit on $\mathcal{T}$, but evaluated on $\mathcal{V}$.
Classic evaluation approaches are ineffective for treatment effect estimation, because both treatment and control are not observed in any observation so the true treatment effect is never observed.
{\it Qini coefficient}, which is based on the {\it Qini curve} \citep{radcliffe2007using}, is commonly used in the uplift literature as an alternative to the goodness-of-fit statistic (see Definition~\ref{eq:q:hat}). The Qini curve separates observations into heterogeneous segments in terms of reactions to the treatment and identify sub-groups with most varying predicted uplifts (see Definition~\ref{def:qini_curve}).\\
\begin{definition}
Given a model, let $\hat{u}_{(1)} \geq \hat{u}_{(2)} \geq ... \geq \hat{u}_{(|\mathcal{V}|)}$ be the sorted predicted uplifts on the validation set $\mathcal{V}$. Let $\phi \in [0,1]$ be a given proportion. Define $ N_{\phi} = \{i: \hat{u}_{i} \geq \hat{u}_{(\ceil{\phi |\mathcal{V}|})} \} \subset \lbrace 1, \ldots, |\mathcal{V}| \rbrace$ as the subset of observations with the $\phi |\mathcal{V}| \times 100 \%$ highest predicted uplifts.
The \textit{Qini curve} is the function $g$ of the fraction of population treated $\phi$, where
\begin{equation}
g(\phi) = \biggl(\sum\limits_{i \in N_{\phi}} y_i t_i - \sum\limits_{i \in N_{\phi}} y_i (1-t_i) \biggl\{ \sum\limits_{i \in N_{\phi}} t_i / \sum\limits_{i \in N_{\phi}} (1-t_i) \biggr\} \biggr) / \sum_{i=1}^{|\mathcal{V}|} t_i.
\end{equation}
\label{def:qini_curve}
\end{definition}
In practice, the domain of $\phi \in [0,1]$ is partitioned into $J$ bins, or $J+1$ grid points $0=\phi_1 < \phi_2 < ... < \phi_{J+1} = 1$. The Qini coefficient is an approximation of the area under the Qini curve.\\
\begin{definition}
The {\it Qini coefficient} is given by:
\begin{equation}
\hat{q} = \dfrac{1}{2} \sum_{k=1}^J (\phi_{k+1}-\phi_k)\{Q(\phi_{k+1}) + Q(\phi_{k})\} \times 100\%.
\label{eq:q:hat}
\end{equation}
where $Q(\phi) = g(\phi) - \phi~g(1)$ and $g(1)$ is the average treatment effect in the validation set.
\end{definition}
Unlike the area under the ROC curve, $\hat{q}$ may take negative values. A negative $\hat q$ means the uplift is worse than random targeting. A good uplift model groups the individuals in decreasing uplift bins. This can be measured by the similarity between the theoretical uplift percentiles of predictions compared with empirical percentiles observed in the data. Maximizing the \textit{adjusted Qini coefficient}, given in Definition~\ref{def:qadj}, maximizes the Qini coefficient and simultaneously promotes grouping the individuals in decreasing uplift bins, which in turn result in concave Qini curves \citep{belba2019qbased}.\\
\begin{definition}\label{def:qadj}
Let $B_k$ denote the $k$th bin $(\phi_{k}, \phi_{k+1}] \subseteq (0,1]$, $k=1,\ldots, J$. The \textit{adjusted Qini coefficient} is defined as:
\begin{equation}
\hat{q}_{\mathrm{adj}} = \rho~\mathrm{max}\{ 0, \hat{q}\},
\label{eq:qadj}
\end{equation}
where $\rho$ is the {\it Kendall's uplift rank correlation}:
\begin{equation}
\hat{\rho} = \frac{2}{K(K-1)} \sum_{i<j} \mathrm{sign}(\bar{\hat{u}}_i - \bar{\hat{u}}_j)~\mathrm{sign}(\bar{u}_i - \bar{u}_j),
\label{eq:corr_coeff}
\end{equation}
where $\bar{\hat{u}}_k$ is the average predicted uplift in bin $B_k$, $k \in {1,...,J}$, and $\bar{u}_k$ is the observed uplift in the same bin.
\end{definition}
For the remainder of the paper, we use $\hat{q}_{\mathrm{adj}}$ as the models comparison measure.
\section{Experiments}
We demonstrate the utility of our proposed methods in a simulation study. Each simulation is defined by a data-generating process with a known effect function. Each run of each simulation generates a dataset, and split into training, validation, and test subsets. We use the training data to estimate $L$ different uplift functions $\{\hat{u}_l\}_{l=1}^L$ using $L$ different methods. The models are fine-tuned using the training and validation observations and results are presented for the test set.
\subsection{Data generating process}
We generate synthetic data similar to \cite{powers2018some}. For each experiment, we generate $n$ observations and $p$ covariates. We draw odd-numbered covariates independently from a standard Gaussian distribution. Then, we draw even-numbered covariates independently from a Bernoulli distribution with probability $1/2$. Across all experiments, we define the mean effect function $\mu(\cdot)$ and the treatment effect function $\tau(\cdot)$ for a given noise level $\sigma^2$. Given the elements above, our data generation model is, for $i=1,...,n$,
\begin{align*}
&Y_i^* \mid \mathbf x_i, t_i \sim \mathcal{N}(\mu(\mathbf x_i) + t_i\tau(\mathbf x_i), \sigma^2),\\
&Y_i = \mathbbm{1}(Y_i^* > 0 \mid \mathbf x_i, t_i)
\end{align*}
where $t_i$ is the realisation of the random variable $T_i \sim \mathrm{Bernoulli}(1/2)$ and $Y_i$ is the binary outcome random variable. Following \cite{powers2018some}, within each set of simulations, we make different choices of mean effect function and treatment effect function. In this section, we describe the results associated with the most complex scenario since the conclusions are similar from one scenario to another. For the reader wishing to analyze the results for other scenarios, we present them in Appendix~\ref{sec:appendix_experiments}. We fix $n=20000$, $p=100$ and $\sigma=4$ and define $\mu(\mathbf x)$ and $\tau(\mathbf x)$ as follows
\begin{align*}
\mu(\mathbf x) &= 4 \mathbbm{1}(x_1 > 1)\mathbbm{1}(x_3 > 0) + 4\mathbbm{1}(x_5 > 1)\mathbbm{1}(x_7 > 0) + 2x_8x_9,\\
\tau(\mathbf x) &= \frac{1}{\sqrt{2}} (f_4(\mathbf x) + f_5(\mathbf x)),\\
\end{align*}
where
\begin{align*}
f_4(\mathbf x) &= x_2 x_4 x_6 + 2 x_2 x_4 (1-x_6) +3x_2(1-x_4)x_6 + 4x_2(1-x_4)(1-x_6) + 5(1-x_2)x_4x_6 \\
&+ 6(1-x_2)x_4(1-x_6) + 7 (1-x_2)(1-x_4)x_6 + 8(1-x_2)(1-x_4)(1-x_6),\\
f_5(\mathbf x) &= x_1 + x_3 + x_5 + x_7 + x_8 + x_9 - 2.\\
\end{align*}
Next, we repeat each experiment $20$ times and each run generates a dataset, which is divided into training ($40\%$), validation ($30\%$), and test samples ($30\%$). The models are fitted using the training observations and results are presented for the test set.
\subsection{Regularization}
In Table~\ref{tab:l1_l2_gl_}, we compare the performance of the twin model $\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta})$ (defined in \eqref{eq:neural_network_2_scaling_factor}) using different regularization functions, with and without the scaling factors for pruning (i.e., structured sparsity). Next, we will refer to this model as {\it Twin$_{\mathrm{NN}}$}. We fix the initial number of hidden neurons (or nodes) to $m=512 $. When structured sparsity regularization is used, we denote by $\hat m$ the number of remaining hidden neurons. We run the experiments $20$ times. For fitting the models in each experiment, we vary the hyper-parameters from the following grid: \textit{learning rate} $\eta \in \{ 0.005, 0.01, 0.05, 0.1, 0.2, 0.3 \}$, \textit{structured sparsity constant} $\lambda_1 \in \{0, 0.0001, 0.0005, 0.001, 0.005, 0.01 \}$, \textit{regularization constant} $\lambda_2 \in \{0, 0.0001, 0.0005, 0.001, 0.005, 0.01 \}$ and cross-validate on the $\hat{q}_{\mathrm{adj}}$ for each combination.
\begin{table}[!ht]
\centering
\begin{tabular}{llcc
\hline
Structured Sparsity & $\mathcal R(\cdot)$ & $\hat{m}/m$ & $\hat{q}_{\mathrm{adj}}$ \\%& $\hat{q}_{\mathrm{adj}}(\hat{m})$ [$L_2$] & $\hat{q}_{\mathrm{adj}}(\hat{m})$ [$L_1$] \\
\hline
No & $L_2$ & $512/512$ & $3.15$ \\%& - & -\\
No &$L_1$ & $512/512$ & $3.35$ \\%& - & -\\
Yes & 0 & $218/512$ & $3.09$ \\%& 2.80 & 3.16\\
Yes &$L_2$ & $417/512$ & $3.32$ \\%& 3.01 & 3.34\\
Yes &$L_1$ & $340/512$ & $\bf 3.58$ \\%& 3.01 & 3.31\\
\hline
\end{tabular}
\vspace{2mm}
\caption{Average adjusted Qini (20 runs) for the twin model \eqref{eq:neural_network_2_scaling_factor} with structured pruning of nodes and different regularization functions $\mathcal R(\cdot)$. The $L_1$ regularization of weights provides the highest performance. Note that the maximum standard-error is $0.1$; we do not report them to simplify the Table.
}
\label{tab:l1_l2_gl_}
\end{table}
With the $ L_1 $ regularization, the test-set adjusted Qini coefficient is higher than when the $ L_2 $ regularization is used. When structured sparsity is used alone, the model removes too many neurons (around half) but $\hat{q}_{\mathrm{adj}}$ decreases to $3.09$. On the other hand, when structured sparsity is used in addition to the $ L_2 $ regularization, we see a clear improvement over $L_2$ or structured sparsity alone, and in this case, the number of pruned neurons is smaller. However, it does not outperform the $L_1$ regularization used alone. Finally, the best combination is the $ L_1 $ regularization paired with the structured sparsity achieving an average adjusted Qini coefficient of $3.58$.
Now, when structured sparsity is used to prune the hidden nodes, it is possible to refit a new model with $\hat{m}$ hidden nodes. In our experiments, this did not necessarily have a positive impact on prediction performance. In the case where only structured sparsity is used, $\hat{m} = 218$. For a model with $ 218 $ hidden neurons fitted with the $ L_2 $ regularization, $\hat{q}_{\mathrm{adj}}$ drops to $ 2.80 $. With $ L_1 $, it increases slightly to reach an average of $ 3.16 $. The same results are observed when we first use $ L_2 $ and structured sparsity and then refit models with $417$ hidden neurons. Finally, in the case where we fit the models directly with $ L_1 $ and structured sparsity, the prediction performance is at its maximum. Indeed, when we refit the models with $ \hat{m} = 340 $ hidden nodes, whether with $ L_2 $ or $ L_1 $, the adjusted Qini decreases from $3.58$ to $ 3.01 $ and $ 3.31 $ respectively. Other scenarios not presented here yielded similar conclusions. Therefore, it seems better to fit the models in one step, using the $ L_1 $ regularization and structured sparsity. This make it possible to prune a few nodes from the hidden layer, and to get sparse estimation of the remaining parameters. This seems to improve the prediction performance of the underlying model. For the rest of the paper, we fit our models with $ L_1 $ and structured sparsity.
\subsection{Comparison with benchmark models}
In this section, we compare our models to different benchmarks. First, let us focus on the simple twin model $\mu_{1t}(\mathbf x, \boldsymbol{\theta})$, defined in \eqref{eq:neural_network_1}. The goal is to compare three optimization procedures. The first method optimizes the penalized Binomial likelihood, as in the case of a {\it lasso} logistic regression. This is the baseline model, and we will refer to it as \textit{Logistic} as a reference to Lo's interaction model \citep{lo2002true}. For fair comparison, we use the twin neural architecture. The second method is a two-stage method which
is based on a derivative-free optimization of the $\hat{q}_{\mathrm{adj}}$ and imposes sparsity \citep{belba2019qbased}.
We denote this model by \textit{Qini-based} and our proposed method by {\it Twin$_{\mu}$}. We also compare to lasso versions of the R-learner \citep{NieWager2020Quasi} and X-learner \citep{kunzel2019metalearners}. Table~\ref{tab:param_estimation_results_} shows the averaged results on the test dataset. This experience shows the superiority of using our uplift loss function. Results for other scenarios are presented in Appendix~\ref{sec:appendix_experiments}.\\
\begin{table}[!ht]
\centering
\begin{tabular}{ccccc}
\hline
{\it Logistic} & {\it Qini-based} & {\it Twin$_{\mu}$} & {\it R-Learner }(\texttt{lasso}) & {\it X-Learner }(\texttt{lasso}) \\
\hline
$2.59$ & $2.94$ & $\bf{3.12}$ & $2.83$ & $2.91$\\
\hline
\end{tabular}
\vspace{2mm}
\caption{``Simple" models prediction performance comparison in terms of $\hat{q}_{\mathrm{adj}}$. Note that the maximum standard-error is $0.05$; we do not report them to simplify the Table.}
\label{tab:param_estimation_results_}
\end{table}
Next we compare our twin model $\mathrm{NN}_{1t}(\mathbf x, \boldsymbol{\theta})$, defined in \eqref{eq:neural_network_2_scaling_factor}, with commonly used benchmark models. A method that has proven to be very effective in estimating treatment effects is one based on generalized random forests \citep{athey2019generalized}. This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the uplift. This has the effect of reducing overfitting. Another candidate that we consider in our experiments is also based on random forests and designed for uplift \citep{guelman2012random}. For this method, we consider two different split criteria, one based on the Kullback-Leibler (KL) divergence and one based on the Euclidean distance (ED). We also compare our method to XGboost \citep{chen2015xgboost} versions of the R- and X-learners. In Table~\ref{tab:beyond_linearity_results_}, we compare these benchmark models to different versions of our {\it Twin$_{\mathrm{NN}}$} with different values of $m$ (hidden neurons), with and without structured sparsity. Appendix~\ref{sec:appendix_experiments} provides details about model fine-tuning and other scenarios results. \\
\begin{table}[!ht]
\centering
\begin{tabular}{lccc}
\hline
Method & Structured Sparsity & Hidden-Layer Size ($\hat m/m$) & $\hat{q}_{\mathrm{adj}}$ \\
\hline
{\it Twin$_{\mathrm{NN}}$} & No & $64/64$ & $3.10$\\
{\it Twin$_{\mathrm{NN}}$} & No & $128/128$ & $3.16$\\
{\it Twin$_{\mathrm{NN}}$} & No & $201/201$ & $3.14$\\
{\it Twin$_{\mathrm{NN}}$} & No & $256/256$ & $3.26$\\
{\it Twin$_{\mathrm{NN}}$} & No & $512/512$ & $3.35$\\
{\it Twin$_{\mathrm{NN}}$} & Yes & $340/512$ & $\bf 3.58$\\
\hline
\multicolumn{3}{l}{{\it Causal Forest}}& $2.79$ \\
\multicolumn{3}{l}{{\it Causal Forest (Honest)}}& $\bf 3.07$\\
\multicolumn{3}{l}{{\it Uplift Random Forest (KL)}}&$2.19$\\
\multicolumn{3}{l}{{\it Uplift Random Forest (ED)}}& $2.33$ \\
\multicolumn{3}{l}{{\it R-Learner }(\texttt{XGboost})}& $2.12$\\
\multicolumn{3}{l}{{\it X-Learner }(\texttt{XGboost})}& $2.37$ \\
\hline
\end{tabular}
\vspace{2mm}
\caption{Models prediction performance comparison in terms of $\hat{q}_{\mathrm{adj}}$. Note that the maximum standard-error is $0.1$; we do not report them to simplify the Table.}
\label{tab:beyond_linearity_results_}
\end{table}
Neural networks have two main hyper-parameters that control the architecture or topology of the network: the number of hidden layers and the number of nodes in each hidden layer. The number of nodes can be seen as a hyper-parameter. If we don't use our pruning technique for structured sparsity, we can search for the right number of nodes over a grid. For comparison purposes, we varied the number of nodes using either $m=64, 128, 256$ or $m=512$ nodes. As we can see in Table~\ref{tab:beyond_linearity_results_}, the best model is reached for $m = 512$ (with $\hat{q}_{\mathrm{adj}}=3.35$). Choosing the number of hidden neurons by cross-validation is very common in practice. This is part of architecture search. However, choosing among a few values does not necessarily mean that all the selected model's neurons are useful. As discussed earlier, the use of the scaling factor penalization on the weight matrix allows to automatically fine-tune the number of hidden neurons. This has the effect of increasing the performance of the underlying models from a predictive point of view, as shown in Table~\ref{tab:l1_l2_gl_}. Thus, we suggest to start with a fairly large number of hidden neurons (e.g., $m \geq 2p +1$), and to let the optimization determine the number of active neurons needed for a specific dataset.
Based on these experiments, it appears that the twin neural model fitted with $L_1$ regularization and structured sparsity outperforms all other methods significantly in terms of $\hat{q}_{\mathrm{adj}}$ (e.g., based on Wilcoxon signed-rank test \citep{wilcoxon1992individual}). We also observe that the random forest with honesty criteria performs best in comparison to the other uplift benchmark models.
Finally, the number of hidden layers can also be seen as a hyper-parameter for the neural network architecture. Note that in our experiments, increasing the number of hidden layers did not improve the prediction performance significantly. For instance, we fixed the initial number of hidden neurons to $p+1$ and $p$ for a $2$-hidden layers twin model and observed an average $\hat{q}_{\mathrm{adj}}$ of $3.39$. Results for other scenarios are given in Appendix~\ref{sec:appendix_experiments}.
\section{Application}
We have the rare opportunity to have access to real data from a large scale randomized experiment. Indeed, in the uplift domain, it is not easy to find public benchmarks, apart from the recent CRITEO-UPLIFT1 dataset \citep{Diemert2018}.
The CRITEO-UPLIFT1 dataset is constructed by collecting data from a particular randomized trial procedure where a random part of the population is prevented from being targeted by advertising. The dataset consists of $\approx 14M$ rows, each one representing a user with $p=12$ covariates, a treatment indicator and a binary response variables (visits). Positive responses mean the user visited the advertiser website during the test period (2 weeks). The proportion of treated individuals is $85\%$. We use an undersampling method in order to bring the treatment/control ratio back to 1. Thus, we have about $4M$ observations. We also create two other balanced datasets by randomly sampling $400K$ and $1M$ observations.
We fit the twin-network {\it Twin$_{\mathrm{NN}}$} and optimize the regularized loss function with lasso and structured sparsity. For comparison purposes, we also consider the honest causal forest, which gave the best benchmark in our simulation study. We split the data into training ($40\%$), validation ($30\%$), and test samples ($30\%$). We use the training and the validation sets to fit and fine-tune the models. Then, we select the best model based on the validation set. Using the best model, we score the samples from the test set and compute the adjusted Qini coefficient $\hat{q}_{\mathrm{adj}}$. We repeat the experiment $20$ times and report averaged results as well as their standard-errors in Table~\ref{tab:final_models_results2}.\\
\begin{table}[!ht]
\centering
\begin{tabular}{llc|cc|c}
\hline
Dataset & Sample Size & $p$ &{\it Twin$_{\mathrm{NN}}$} & $\hat{m}/m$ & {\it Causal Forest (Honest)} \\
\hline
Insurance data & 50K & $40$ & $0.19~(0.003)$ & $212/512$ & $0.06~(0.007)$ \\
CRITEO-UPLIFT1 & 100K& $12$ & $0.19~(0.008)$ & $187/512$ & $0.14~(0.011)$ \\
CRITEO-UPLIFT1 & 1M & $12$ & $0.24~(0.007)$ & $127/512$ & $0.18~(0.019)$ \\
CRITEO-UPLIFT1 & 4M & $12$ & $0.28~(0.006)$ & $63/512$ & $0.23~(0.015)$\\
\hline
\end{tabular}
\vspace{2mm}
\caption{Application on real-data. Prediction performance in terms of $\hat{q}_{\mathrm{adj}}$ averaged over $20$ realizations (standard-errors are given in parenthesis).}
\label{tab:final_models_results2}
\end{table}
The same type of analysis is repeated on a data set that was made available to us by a car insurance company with $50K$ customers and $p=40$ covariates. This company was interested in designing strategies to maximize its conversion rate. An experimental acquisition campaign was implemented for $6$ months, for which half of the potential clients were randomly allocated into a treatment group and the other half into a control group. Potential clients under the treatment group were contacted. The goal of the analysis is to propose a predictive model that maximizes the return on investment of future initiatives (i.e., a maximum $\hat{q}_{\mathrm{adj}}$ on a test set). The observed difference in sales rates between the treated and the control groups shows a slightly positive impact of the marketing initiative, that is, $0.55\%$. Results are reported in Table~\ref{tab:final_models_results2}.
For the insurer's data, we also considered a model without hidden layers (i.e., {\it Twin$_{\mu}$}). The objective being model interpretability. In this case, the model is not as efficient as the neural network one (see Table~\ref{tab:final_models_results}). However, it performs better than the causal forest. Therefore, if the practitioner needs an interpretable model, he can use the simple twin model and still get satisfactory results. For comparison purposes, we also fit the {\it Logistic} model. In this case, the model's parameters are estimated by maximizing the penalized Binomial log-likelihood.
\begin{table}[!ht]
\centering
\begin{tabular}{lcccc}
\hline
Method & $\hat{m}/m$ &Training & Validation & Test \\
\hline
{\it Twin$_{\mu}$} & - & $0.112~(0.005)$ & $0.098~(0.004)$ & $0.085~(0.005)$\\
{\it Twin$_{\mathrm{NN}}$} & $212/512$ & $0.223~(0.006)$ & $0.202~(0.003)$ & $0.187~(0.003)$\\
\hline
{\it Logistic} & - & $0.197~(0.029)$ & $0.062~(0.019)$ & $0.021~(0.007)$ \\
{\it Causal Forest (Honest)} & - & $0.573~(0.028)$ & $0.157~(0.022)$ & $0.059~(0.007)$ \\
\hline
\end{tabular}
\vspace{2mm}
\caption{Adjusted Qini coefficients on the insurance data (standard errors are given in parenthesis). The results are averaged over $20$ runs.}
\label{tab:final_models_results}
\end{table}
For these types of marketing initiatives, it makes sense that the overall impact is small. A potential reason is that customers are already interested in the product since it is mandatory to get car insurance. The effect of the call is slightly positive on average, that is to say that during the call, the advisor allows the customer to understand all the details of the coverage and thus the customer is most likely reassured. Therefore, it is rather unlikely that the call would have a negative effect. We believe that for this data set, the situation is similar to the first scenario from the simulations presented in Appendix~\ref{sec:appendix_experiments}. This may explain why it is more difficult to prevent overfitting with the causal forest model.
\section{Conclusion}
We present a meaningful and intuitive twin neural networks architecture for the problem of uplift modeling. We proposed to estimate the model's parameters by optimizing a new loss function. This loss function is built by leveraging a connection with the Bayesian interpretation of the relative risk. The twin neural network performs well on predictive tasks and overfitting is minimized by using a proper regularization term. We applied our method to synthetic and real-world data and we compared it with the state-of-the-art methods for uplift. We modify the learning algorithm to allow for structured sparse solutions, which significantly helped training uplift models. Our results show that the twin models significantly outperform the common approaches to uplift such as random forests in all scenarios.
Our methodological development has been driven by real data from a large scale random experiment where the treatment is the marketing initiative. The methodology is also readily applicable to other types of datasets, such as randomized clinical trial data, for predicting variability in medical treatment response. In addition to the quality of prediction obtained in optimal settings, we observed that even the one-layer version of the proposed architecture performs well compared to random forest models. In this case, the model can be interpreted in the same way as a logistic regression.
This interpretation is of great importance in practice, since the deployment of an interpretable model minimizes risks of applying an inadequate model and facilitates convincing business owners. This is also true in the medical world, where understanding the variability in response to treatment predictions is of great importance for advancing precision medicine, but not at the cost of model explanation.
\section{Old Experiments Section}
\subsection{Fine-tuning}
Deep Learning has enabled remarkable progress over the last years on a variety of tasks, such as image recognition, speech recognition, and machine translation. One crucial aspect for this progress are novel neural architectures. Currently employed architectures have mostly been developed manually by human experts, which is a time-consuming and error-prone process. Because of this, there is growing interest in automated neural architecture search (NAS) methods. NAS methods can be categorized according to three dimensions \citep{elsken2018neural}: search space, search strategy, and performance estimation strategy.
The search space defines which architectures can be represented
in principle. Incorporating prior knowledge about typical properties of architectures well-suited for a task can reduce the size of the search space and simplify the search. SMITE architecture is a Sieamese multi-layer perceptron. Therefore, the search space for the architecture is first defined by the number of hidden layers and the number of hidden units per layer. However, it should not be limited to these two parameters. The choice of loss functions, regularization and activations in deep networks has also a significant effect on the training dynamics and task performance. These functions are therefore part of the architecture and should be included in the search space.
Neural network activations determine the output of a node and are a crucial component of deep learning. When the activation function is non-linear, then a two-layer neural network can be proven to be a universal function approximator. The identity activation function does not satisfy this property. When multiple layers use the identity activation function, the entire network is equivalent to a single-layer model. The Parameteric Rectified Linear Unit (PReLU) \citep{he2015delving} has been proposed as an extension of Rectified Linear Unit (ReLU) \citep{nair2010rectified} in order to improve image classification accuracy. The PReLU function is defined as
\[
f_a (x) =
\begin{cases}
x, ~\mathrm{if}~ x \geq 0\\
ax, ~\mathrm{otherwise}
\end{cases}
\]
where $a \in [0;1]$ is learned in the training via backpropagation. When $a=0$, PReLU simplifies to ReLU and if $a$ is set to a fixed value, PReLU simplifies to Leaky ReLU (LReLU) \citep{maas2013rectifier}.
The search strategy details how to explore the search space
(which is often exponentially large or even unbounded). It encompasses the classical exploration-exploitation trade-off since, on the one hand, it is desirable to find well-performing architectures quickly, while on the other hand, premature convergence to a region of suboptimal architectures should be avoided. In \cite{yang2019evaluation}, based on a benchmark of 8 NAS methods on 5 datasets, they compare methods with different search spaces, using a method’s relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, they find that many NAS techniques struggle to significantly beat the average architecture baseline. Furthermore, it is desirable to have a search strategy because unfortunately, in many deep learning tasks, training is computationally expensive. However, for our application, training SMITE on the real data takes 10 minutes for 100 epochs (on a Dual Intel Xeon 256 Gb RAM and 1 Titan V12GB). Therefore, we believe that grid search is suitable for SMITE architecture search. Table~\ref{tab:NAS} summarizes the hyper-parameters responsible for defining the architecture of SMITE. Hyper-parameter 3 must be defined for each layer. Although hyper-parameters 5-6 do not modify the architecture, they are important in terms of training. We prefer to enumerate them in Table~\ref{tab:NAS} in order to include them in the architecture search. Hyper-parameter 7 is important for reproducible results. Our experimental results are presented in Table~\ref{tab:experiments_low_rank_full_rank} on the real-world marketing campaign dataset (see Section~\ref{sec:application} for the description).
\begin{table}[htb]
\caption{Hyper-parameters that define the architecture of SMITE.}
\centering
\begin{tabular}{|c||l|l|l|}
\hline
\# & Hyper-parameter & Type & Scope\\
\hline
1 & \# of hidden layers & Integer & $\{0,1\}$\\
2 & \# of hidden units & Integer & $\{8, 16, 32, 64, \ldots, 2048\}$\\
3 & Activation & Nominal & $\{$ReLU, LReLU$(0.05)\}$\\
4 & Objective function & Nominal & $\{$TO, IE$\}$ \\
5 & Batch size $m$ & Integer & $\{128,256\}$ \\
6 & Learning rate $\xi$ & Continuous & $\{0.01,0.02,\ldots,0.10\}$ \\
7 & Seed & Integer & $\{1,2,3,\ldots,10\}$ \\
\hline
\end{tabular}
\label{tab:NAS}
\end{table}
The objective of NAS is typically to find architectures that achieve high predictive performance on unseen data. Performance estimation refers to the process of estimating this performance: the simplest option is to perform a standard training and validation of the architecture on data. Here, we use the reference metric for uplift, that is, the adjusted Qini coefficient \eqref{eq:qadj}.
\begin{table}[H]
\centering
\caption{Fine-tuning SMITE with $p=40$ using the insurance policy data. Reported $\hat{q}_{\mathrm{adj}}$ are averaged over $30$ runs on validation sets, with standard errors in parenthesis. Here, the objective function used is IE.}
\begin{tabular}{|c||l|l|l|l|}
\hline
$m$ & No SVD & $r=\min\{m,p\}$ & $r=\ceil{0.5\min\{m,p\}}$ & $r=\ceil{0.1\min\{m,p\}}$\\
\hline
8 & & & & \\
16 & & & & \\
32 & & & & \\
64 & & & & \\
128 & & & & \\
256 & & & & \\
512 & & & & \\
1024 & & & & \\
2048 & & & & \\
\hline
\end{tabular}
\label{tab:experiments_low_rank_full_rank}
\end{table}
\subsection{Data generation}
We evaluate the performance of SMITE using synthetic data generated using two types of mechanisms.
\paragraph{Bootstrap Data.} The underlying treatment effects are inaccessible in real-world data. To create realistic scenarios, we based one of our simulation study data generation on insurance policy data. Moreover, since we compare SMITE to random forests, we want to evaluate its performance when the data generation process comes from a tree-based model. We follow the procedure in the simulation studies of \cite{belba2019qbased}. These simulations allow us to evaluate if SMITE can recover the right representation for data generated from a tree-based model. This also makes comparisons to random forest methods fair.
Several tree-based methods have been suggested in the uplift literature. Here, we use the uplift random forest as the data generation process, which is readily available in the {\bf R} library {\it uplift} \citep{guelman2014R_uplift}. We fit the random-forest on a random sub-sample $\mathcal{D}$ of the real data. We use $100$ trees and the Kullback-Leibler split criterion. Each tree has maximum depth of $5$. Based on the fitted model, we can extract the probabilities $\mu_0(\mathbf x)$ and $\mu_1(\mathbf x)$ for any given value $\mathbf{x}$. We use these probabilities to draw synthetic outcomes for treated and control observations. We start by creating a bootstrap sample $\mathcal{S}$ of size $n_\mathcal{S}$ from $\mathcal{D}$. For each observation $\mathbf{x}_i \in \mathcal{S}$, we generate a random vector $\tilde{y}_i =(\tilde{y}_{i0}, \tilde{y}_{i1})$, where $\tilde{y}_{i0}$ is the binary outcome of a Bernoulli trial with success probability $\mu_0(\mathbf{x}_i)$, and $\tilde{y}_{i1}$ is the binary outcome of a Bernoulli trial with success probability $\mu_1(\mathbf{x}_i)$, $i=1,\ldots, n_\mathcal{S}$. The augmented synthetic dataset $\{ \mathbf x_i , t_i, \tilde{y}_i \}_{i=1}^{n_{\mathcal{S}}}$ is the data of interest in the simulation.
We repeat this process several times. Each bootstrapped data size is $10,000$ observations with $40$ covariates.
\subsection{Experimental results}
For both simulated datasets, we do not consider any out-of-sample subset. In Monte Carlo settings, we are able to generate as much data as we want. Therefore, we first fine-tuned the methods on a single synthetic dataset of same size as the one we used in the repeated simulations presented. We generate $n=10,000$ observations with $p=40$ covariates for the bootstrapped data. For the fully synthetic datasets, we report results from experiments with $n=10,000$ observations and $p=100$ covariates. We split into training and validation and compare the results averaged over $30$ runs for the models that performed best on the validation set.
Results are described in Table \ref{tab:boot_perf} for the bootstrap datasets and Table \ref{tab:synthe_perf} for the synthetic datasets. For the fully synthetic data, we see that Causal Forest (H) and SMITE (TO) had similar performance on the validation set while SMITE (IE) significantly outperformed all the other methods. For the bootstrap data for which the outcomes were generated based on random forest model, it is interesting to see that SMITE (TO) and Causal Forest (H) give similar results. Moreover, once again, the SMITE (IE) model outperforms all the other models.
\input{tables/table_bootstrap.tex}
\input{tables/table_synthetic.tex}
\section{Data generation}
Synthetic data generation depends on the following four elements.
\begin{enumerate}
\item The number $n$ of observations in the sample, and the number $p$ of predictors observed for each observation.
\item The distribution of the predictors. We draw predictors from a multivariate Gaussian distribution.
\item The mean effect function $\mu(\mathbf X)$ and the treatment effect function $\tau(\mathbf X)$. We vary the size of main effects relative to treatment effects.
\item The noise level $\sigma^2$. This corresponds to the conditional variance of the outcome given $\mathbf X$ and $T$.
\end{enumerate}
Given the elements above, our data generation model is, for $i=1,...,n$ :
\begin{align*}
&\mathbf X_i \sim \mathcal{N}_p (0, \Sigma) \\
&T_i \sim \mathrm{Bernoulli}(\theta)\\
&Y_i^* \mid \mathbf X_i, T_i \sim \mathcal{N}(\mu(\mathbf X_i) + T_i\tau(\mathbf X_i), \sigma^2).
\end{align*}
Define $Y_i$ to be the binary outcome as
\begin{align*}
Y_i \mid \mathbf X_i, T_i = \mathbbm{1}(Y_i^* > 0 \mid \mathbf X_i, T_i).
\end{align*}
Hence, $Y_i$ follows a Bernoulli distribution with parameter $\mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i)$ and it is easy to recover the ``true'' treatment effect $u_i^*$ as follows
\begin{align*}
u_i^* &= \mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i = 1) - \mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i = 0)\\
u_i^* &= \Phi \Bigl( \frac{\mu(\mathbf X_i) + \tau(\mathbf X_i)}{\sigma} \Bigr) - \Phi \Big( \frac{\mu(\mathbf X_i)}{\sigma} \Big)
\end{align*}
where $\Phi(.)$ is the standard Gaussian cumulative distribution function. Note that $\mu(.)$ is a nuisance function and $\tau(.)$ is the function that controls the treatment effect. In fact, the sign of $\tau(.)$ and the sign of the treatment effect for a given observation coincide.
We set $\theta$ to $0.5$, which is the case in our real data. For observational data settings, one can define $\theta$ as a function of a subset of the covariates, e.g. $\theta= 1 / \{1 + e^{-(X_1 + X_2)} \}$.
The variance-covariance matrix elements $\Sigma_{ij}$ are defined as follows:
\[\Sigma_{ij} = \begin{cases}
\rho & i\neq j \\
1 & i=j
\end{cases}.
\]
Finally, the mean and treatment effects functions are defined as follows:
$$\mu(\mathbf X) = \sum_{j=1}^J \beta_j X_j + \sum_{j=1}^{p-1} \omega_j X_j X_{j+1},$$
$$\tau(\mathbf X) = \sum_{j=1}^K \gamma_j X_j$$
where $J\leq p$ and $K \leq p$. We set $\beta_j = \beta (-1)^{j}$, $\omega_j = \omega (-1)^j$ and we draw $\gamma_j$ independently from $\mathcal{N}(0,\gamma)$. The constants $\beta$, $\omega$ and $\gamma$ are hyper-parameters.
\bibliographystyle{plainnat}
\section{Introduction}
\label{sec:intro}
Mainly known in the marketing research literature, uplift models \citep{radcliffe1999differential, hansotia2002direct, lo2002true, radcliffe2007using} provide a solution to the problem of isolating the marketing effect. Uplift refers to modeling customers behavioral change that results directly from a specific treatment, such as a marketing intervention (e.g. a courtesy call). Typically, uplift models are developed for randomized controlled trials settings where both the treatment and outcome are binary random variables. Then, the customers can be segmented along two dimensions in function of the potential outcomes and the treatment, giving rise to the following groups \citep{kane2014mining} : a \textit{persuadable} client is an individual who purchases the product only if he receives a call; a \textit{sure thing} client is an individual who purchases the product whether he receives the call or not; a \textit{lost cause} client is an individual who will not purchase the product regardless of whether he receives the call or not; a \textit{do-no-disturb} client is an individual who will not purchase the product only if he receives a call. In general, from a business point of view, interesting customers are the \textit{persuadable} and the \textit{do-not-disturb} clients. The persuadable clients provide incremental sales whereas the do-not-disturb individuals should be avoided because the marketing campaign would have a negative effect on them.
Uplift is a particular case of causal inference and follows the potential outcomes framework \citep{rubin1974estimating, holland1986statistics}. Formally, let $Y$ be the 0-1 binary response variable, $T$ the 0-1 treatment indicator variable and $X_1,\ldots,X_p$ the explanatory variables (predictors). Suppose that $n$ independent units are observed $\{(y_i, \textbf{x}_i, t_i)\}_{i=1}^n$, where $\mathbf x_i = (x_{i1}, \ldots, x_{ip})$ are realisations of the predictors random variables. For $i=1,\ldots,n$, an uplift model estimates
\begin{equation*}
u (\textbf{x}_i)= \mu_1(\mathbf x_i) - \mu_0(\mathbf x_i),
\end{equation*}
where the notation $\mu_{t_i}(\mathbf x_i) = \mathrm{Pr}(Y_i = 1 \mid \mathbf x_i ,T_i=t_i)$ stands for the corresponding conditional probability. In different fields this quantity is alternatively called the individual treatment effect (ITE), conditional average treatment effect (CATE), or uplift. There currently exists a number of methods to estimate individual treatment effects from randomized data. The process of estimating these effects is alternatively referred to as heterogeneous effect modeling, individual treatment effect modeling, or uplift modeling. These approaches can also be used for observational data if certain assumptions are met or if combined with propensity score \citep{rosenbaum1983central} or matching techniques \citep{rubin2006matched}.
\textcolor{blue}{Although by definition, uplift and the conditional average treatment effect are equivalent, they differ from an application point of view, notably in the evaluation of the underlying models.
There are several differences between uplift modeling, and ITE modeling. The first is minimal because in the uplift context, the $Y$ variable is binary while in the ITE context, the $Y$ variable is continuous. This has a direct impact on the quality measurements of the models used. Indeed, the second difference is that in the context of the ITE, the use of the MSE makes sense. On the contrary, it is not used in the uplift context.}
\textcolor{red}{Most existing uplift models are adaptations of classification and regression trees \citep{breiman1984classification} where the majority are modified random forests \citep{Breiman-RandomForest-2001, radcliffe2011real, jaskowski2012uplift, guelman2012random}. Within each leaf, uplift is estimated with the difference between the two conditional means. The belief is that since the uplift is defined as a difference, if we can get good estimates of each mean, we have a good estimate of the difference. However, the existing tree-based uplift optimization problems do not take into account the estimation of each mean. Instead, the focus is on maximizing the heterogeneity in treatment effects. Without careful regularization (e.g. honest estimation \citep{athey2019generalized}), splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. In addition, successive splits tend to group together similar extreme values, introducing more variance in the prediction of uplift \citep{zhao2017practically}. Alternatively some models use the transformed outcome \citep{athey2015machine}, an unbiased estimator of the uplift. However, this estimate suffers from higher variance than the conditional mean estimator \citep{powers2018some}. In addition, for both estimators, if random noise is larger than the treatment effect, the model will more likely predict random noise instead of uplift. This is particularly problematic for uplift trees, as the split criteria is function of the two treatment effects. As a result, based on several experiments on real data, and although the literature suggests that tree-based methods are state-of-the-art for uplift \citep{soltys2015ensemble}, the published models over-fit the training data and predicting uplift still lacks satisfactory solutions.}
\textcolor{blue}{We define a unified framework with the goal of improving uplift models predictions on unseen data. Taking advantage of the flexibility and prediction power of neural networks, we design an optimization problem for which regularization of the weights takes into account both conditional means estimation and adequacy of uplift prediction. To the best of our knowledge, our method proposes a new way to tackle overfitting for uplift models. We present two regularization criteria: one using the transformed outcome, that has been shown to be a consistent estimator of the uplift with respect to the mean squared error \citep{gutierrez2017causal}, and a novel regularization criteria that comes from the Bayes conditional probability theorem.
Empirically, we show that our proposed optimization has the impact to reduce the variance of the predicted uplifts. Furthermore, the flexibility of our proposed method opens a door to connections with the estimation of uplift in observational data settings. Our proposed method is validated on both synthetic and a realworld data-set made available to us by a Canadian insurance company. The experimental results demonstrate its advantages over state-of-the-art tree-based uplift modeling.}
\section{Uplift modeling}
\label{sec:literature}
\subsection{Prerequisites}
Following the potential outcomes framework \citep{holland1986statistics}, throughout the paper, we make the following standard assumptions in causal inference. The binary variable $T$ indicates if a unit is exposed to treatment ($T=1$) or control ($T=0$). Let $Y_0$ and $Y_1$ be the potential outcomes under control and treatment respectively. Although each customer $i$ is associated with two potential outcomes, only one of them can be realized as the observed outcome $y_i$.
{\it Assumption 1.} (Unconfoundedness) Potential outcomes $(Y_0,Y_1)$ are independent ($\independent$) of the treatment assignment indicator $T$ conditioned on all pre-treatment characteristics $(X_1, \ldots, X_p)$, i.e.,
$$\big( Y_0, Y_1 \big) \independent T | X_1,\ldots,X_p. $$
This allows the unbiased estimation of the uplift. Moreover, let us define the \textit{propensity score} $e(\mathbf x)= \mathrm{Pr}(T=1 \mid \mathbf x)$. This parameter is widely used to estimate treatment effects from observational data \citep{rubin1974estimating,rosenbaum1983central,athey2015machine}.
{\it Assumption 2.} (Overlap) For any $\mathbf x$, the true propensity score is strictly between $0$ and $1$, i.e.,
$$0 < e(\mathbf x) < 1.$$
In the context of complete randomization, $e(\mathbf x)$ is a constant.
{\it Assumption 3.} (Consistency) Observed outcome $Y$ is represented using the potential outcomes and treatment assignment indicator as follows,
$$Y=TY_1 + (1-T)Y_0.$$
An uplift model estimates the conditional average treatment effect in various sub-populations as a function of the possible values of the covariates $\mathbf x$,
\begin{align}
u(\mathbf x) &= \mathbb{E} ( Y_1 \mid \mathbf x) - \mathbb{E} ( Y_0 \mid \mathbf x) \nonumber\\
&= \mathbb{E} ( Y \mid \mathbf x, T=1) - \mathbb{E} ( Y \mid \mathbf x, T=0) \nonumber \\
&= \mathrm{Pr} ( Y = 1 \mid \mathbf x, T=1) - \mathrm{Pr} ( Y = 1 \mid \mathbf x, T=0) \nonumber \\
&=\mu_1(\mathbf x) - \mu_0(\mathbf x).
\label{eq:uplift}
\end{align}
Uplift is defined as the difference between the two conditional means $\mu_1(\mathbf x)$ and $\mu_0(\mathbf x)$. Therefore, the intuitive approach to model uplift is to build two classification models \citep{hansotia2002direct,snowden2011implementation,austin2012using}. This consists of fitting two separated conditional probability models: one for the treated individuals, and another for the untreated individuals. The uplift is estimated as the difference between these two conditional probability models. These models have been recently referred to as T-learners (T for ``two models") in the literature \citep{kunzel2019metalearners}. The asset of T-learners is their simplicity. However, they do not perform well in practice \citep{radcliffe2011real}. Both models focus on predicting only a one-class probability and for each model, information about the other treatment is never provided to the learning algorithm. In addition, differences between the covariates distributions in the two treatment groups can lead to bias in treatment effect estimation. Using a unique classification model corrects this drawback (S-learners, for single-model \citep{kunzel2019metalearners}). For instance, a traditional uplift estimation method uses the treatment assignment as a feature, adding explicit interaction terms between each covariate and the treatment indicator, and then trains a classification model to estimate the uplift \citep{lo2002true}. The parameters of the interaction terms measure the additional effect of each covariate because of the treatment. T-learners and S-learners together have been referred to as simulated twins, g-computation, counterfactual regression, or conditional mean regression \citep{schuler2018comparison}.
\subsection{Uplift and machine learning}
Several machine learning methods try to model the uplift directly. Some $k$-nearest neighbours \citep{cover1967nearest} based methods are also adopted for uplift estimation \citep{crump2008nonparametric, Alemi.etal-PersonalizedMedicine-2009, su2012facilitating}. The main idea is to estimate the uplift for an observation based on its neighbourhood containing at least one treated and one control observations. However, these methods quickly become computationally expensive for large datasets. Furthermore, adaptive tree-based methods such as random forests outperform the nearest neighbours approach, especially in the presence of irrelevant covariates \citep{wager2018estimation}.
State-of-the-art proposed methods view the forests as an adaptive neighborhood metric, and estimate the treatment effect at the leaf node \citep{su2009subgroup,chipman2010bart}. Therefore, most active research in uplift modeling is in the direction of classification and regression trees \citep{breiman1984classification} where the majority are modified random forests \citep{Breiman-RandomForest-2001}.
In \cite{radcliffe1999differential, radcliffe2011real,hansotia2002direct, hansotia2002incremental,jaskowski2012uplift}, modified split criteria that suited the uplift purpose were studied. The criteria used for choosing each split during the growth of the uplift trees is based on maximization of the difference in uplifts between the two child nodes. However, in practice, these approaches are prone to overfitting. The splitting criteria introduced are based solely on maximizing differences in uplifts. Therefore, splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. Within each leaf, uplift is estimated with the difference between the two conditional means. The belief is that since the uplift is defined as a difference, if we can get good estimates of each mean, we have a good estimate of the difference. However, the existing tree-based uplift optimization problems do not take into account the estimation of each mean. Instead, the focus is on maximizing the heterogeneity in treatment effects. Without careful regularization (e.g. honest estimation \citep{athey2019generalized}), splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. In addition, successive splits tend to group together similar extreme values, introducing more variance in the prediction of uplift \citep{zhao2017practically}. Alternatively some models use the transformed outcome \citep{athey2015machine}, an unbiased estimator of the uplift. However, this estimate suffers from higher variance than the conditional mean estimator \citep{powers2018some}. In addition, for both estimators, if random noise is larger than the treatment effect, the model will more likely predict random noise instead of uplift. This is particularly problematic for uplift trees, as the split criteria is function of the two treatment effects. As a result, based on several experiments on real data, and although the literature suggests that tree-based methods are state-of-the-art for uplift \citep{soltys2015ensemble}, the published models over-fit the training data and predicting uplift still lacks satisfactory solutions.
\textcolor{blue}{Recently, representation learning approaches have been proposed for estimating individual treatment effect based on observational data. An approach in the literature emphasizes learning a covariate representation that has a balanced distribution across treatment and outcome \citep{johansson2016learning}. Some similar methods try to minimize the distribution difference between treated and control groups in the embedding space \citep{shalit2017estimating, yao2018representation}. Another approach is based on propensity score matching \citep{rosenbaum1983central} with neural networks \citep{alaa2017deep, schwab2018perfect} and \citep{shi2019adapting} use regularized neural networks to estimate the average treatment effect.}
\textcolor{blue}{Dans cet article, nous allons nous concentrer sur les S-learners. Dans la litterature, il y a aussi ce qu'on appelle les R-learners our X-learners, specifiquement developes pour les donnees observationnelles. Nous ne les considérons pas ici car ce type de modèles nécessite d'abord l'ajustement d'un premier modèle qui ne prend pas en compte la variable traitement. Par contre, notre méthodologie est facilement adaptable afin de pouvoir construire un R-learner, simplement en modifiant la fonction objective.}
\subsection{The interaction model}
In order to build our final model, we start with the uplift interaction model \citep{lo2002true}. This model is based on logistic regression. Logistic regression is a well-known parametric model for binary response variables. Given a $p$-dimensional predictor vector $\textbf{x}_i$, $i \in \{1,\ldots,n\},$ logistic intercept $\theta_o \in \rm I\!R$, and logistic regression coefficients $\boldsymbol \beta \in \rm I\!R^p$,
the model is
\begin{equation*}
\mathrm{Pr}(Y_i=1 \mid \textbf{x}_i, \theta_o, \boldsymbol \beta) = \bigl(1 + \mathrm{exp}\{- (\theta_o + \textbf{x}_i^{\top} \boldsymbol \beta) \} \bigr)^{-1}.
\end{equation*}
Throughout the paper, the superscript $^\top$ stands for the transpose of a column vector or matrix.
In the uplift context, one need to add explicit interaction terms between each explanatory variable and the treatment indicator.
Let $\gamma$ denote the treatment effect, $\boldsymbol{\beta}$, the vector of main effects, $\boldsymbol{\delta}$, the vector of interactions effects, and $\theta_o$, the intercept.
The model is
\begin{equation}
\mu_{t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta}) = \mathrm{Pr} (Y_i = 1 \mid \mathbf{x}_i, t_i, \theta_o, \boldsymbol{\theta}) =
\bigl( 1+ \mathrm{exp}\{-(\theta_o + \gamma t_i + \mathbf{x}_i^\top [\boldsymbol{\beta} + t_i \boldsymbol{\delta}] ) \} \bigr)^{-1},
\label{eq:pitheta}
\end{equation}
where $\boldsymbol{\theta} = (\boldsymbol{\beta}, \gamma, \boldsymbol{\delta})$, denotes all model parameters except for the intercept $\theta_o$. The likelihood function associated with the uplift model is
\begin{equation}
\mathcal{L}(\theta_o, \boldsymbol{\theta}) = \prod_{i=1}^{n} \mu_{t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta})^{y_i} \{1-\mu_{t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta})\}^{(1-y_i)},
\label{eq:uplift_likelihood}
\end{equation}
where $\{y_i : i=1,\ldots,n\}$ are the observed response variables.
The maximum likelihood estimates of $(\theta_o, \boldsymbol{\theta})$ will be denoted by $(\hat{\theta}_o, \hat{\boldsymbol{\theta}}),$
with $\hat{\boldsymbol{\theta}} = (\boldsymbol{\hat{\beta}}, \hat{\gamma}, \boldsymbol{\hat{\delta}})$.
The predicted uplift associated with the covariates vector $\mathbf{x}_{n+1}$ of a future individual is estimated by
$$ \hat{u}(\mathbf{x}_{n+1}) = \bigl(1+\mathrm{exp}\{-(\hat{\theta}_o + \hat{\gamma} + \mathbf{x}_{n+1}^\top[ \boldsymbol{\hat{\beta}} + \boldsymbol{\hat{\delta}}] )\}\bigr)^{-1} - \bigl(1+ \mathrm{exp}\{-(\hat{\theta}_o + \mathbf{x}_{n+1}^\top \boldsymbol{\hat{\beta}})\}\bigr)^{-1}.$$
\textcolor{blue}{Logistic regression may be visualised as a graph, with one extra node to represent the transfer function. A logistic regression element may also be described using a linear regression element and a transfer node, by recognising that the first two stages of a logistic regression element form a linear regression element.}
\textcolor{blue}{Evidemment, ce n'est pas le seul modele possible. Qini-based uplift regression \citep{belba2019qbased}.}
\section{Siamese uplift model}
The logistic interaction model \eqref{eq:pitheta} is a single model, i.e. we need to estimate the parameters of a single model. However, in order to predict the uplift for an individual, we must first predict $\mu_1$ by fixing $T=1$ and then predict $\mu_0$ by fixing $T=0$ in order to be able to take the difference of the two conditional means. This has an impact on the model's parameters learning. For instance, the parameters estimation for the logistic regression is based on an iterative algorithm known under the name of iteratively reweighted least squares or IRLS \citep{holland1977robust,green1984iteratively,wolke1988iteratively}. We would like to be able to adapt this type of algorithms in order to orient the objective function towards the training of the uplift in a direct way at each iteration of the algorithm. This implies that we need a model to predict both the conditional means and their difference (the uplift) at each iteration. To achieve this, we propose a Siamese model.
Siamese networks were first introduced in signature verification as an image matching problem \citep{bromley1994signature}, where the task is to compare two hand-written signatures and infer identity. A Siamese neural network consists of twin networks which accept distinct inputs but are joined by an energy function at the top. This function computes some metric between the highest level feature representation on each side. The parameters between the twin networks are tied. Weight tying guarantees that two extremely similar images could not possibly be mapped by their respective networks to very different locations in feature space because each network computes the same function.
\begin{figure}[H]
\centering
\input{figures/tikzsmite}
\caption{
Siamese representation of the uplift interaction model.}
\label{fig:siamese_network}
\end{figure}
Siamese networks are mostly known for their application to face verification and recognition \citep{chopra2005learning,taigman2014deepface,parkhi2015deep,schroff2015facenet}, areal-to-ground image matching \citep{lin2015learning}, stereo matching \citep{zbontar2015computing}], patch matching \citep{han2015matchnet, simo2015discriminative,zagoruyko2015learning}, optical flow \citep{dosovitskiy2015flownet}, large scale video classification \citep{karpathy2014large} and one-shot character recognition \citep{koch2015siamese}. Despite their generality and usefulness in various applications, to the best of our knowledge, Siamese models have never been used for uplift modeling.
Here, we propose a Siamese model called SMITE \footnote{Siamese Model for Individual Treatment Effect} which works in tandem on two almost identical augmented input vectors, and outputs the prediction of the conditional mean $\hat{\mu}_t(\mathbf x) = t \hat{\mu}_1(\mathbf x) + (1-t) \hat{\mu}_0(\mathbf x)$ as well as the one of the uplift $\hat{u}(\mathbf x)$. Using a Sieamese model simplifies the fitting process. On the same forward pass, the model outputs the conditional means as well as the uplift. The idea is that the two input vectors are fed to two identical model sub-components (i.e., with shared weights $\boldsymbol{\theta}$). The inputs contain the covariates vector $\mathbf x$ and, for the left sub-component, the treatment variable fixed to $1$. The treatment variable is fixed to $0$ for the right sub-component as shown in Figure \ref{fig:siamese_network}. The sub-components output the predicted conditional means for treated ($\hat{\mu}_1$) and for control ($\hat{\mu}_0$). The difference gives direct prediction for $\hat{u}(\mathbf x)$. At the same time, the predicted conditional mean $\hat{\mu}_t$ is based on the actual treatment $t \in \{0,1\}$, i.e., $\hat{\mu}_t = t\hat{\mu}_1 + (1-t)\hat{\mu}_0$.
The goal of this paper is to study how we can use the Siamese representation
for flexible uplift estimation that builds on modern machine learning approaches
such as deep learning. Our main result is that we can use this representation
to construct objective functions that capture heterogeneous treatment effects, and that we can accurately estimate the uplift by decomposing these objective functions into a combination of a loss function and a regularization function. Our goal is to estimate the model's parameters by minimizing an objective function of the following form,
\begin{align}
{\boldsymbol{\hat\theta}}_{\alpha} = \argmin_{\boldsymbol{\theta}} (1 - \alpha) \mathcal{L}\Big(\mu(\boldsymbol{\theta})\Big) + \alpha \mathcal{J}\Big(u(\boldsymbol{\theta})\Big),
\label{eq:SMITE_general_objective}
\end{align}
where $\alpha \in [0,1)$ can be seen as a hyper-parameter which controls the trade-off between the two losses in the optimization problem. The term $\mathcal{J}\Big(u(\boldsymbol{\theta})\Big)$ is interpreted as a regularizer on the complexity of the $u(\boldsymbol{\theta})$ function. Equation (\ref{eq:SMITE_general_objective}) encompasses what motivates the SMITE framework to be able to penalize the conditional means regression weights in order to guide the uplift predictions. This is equivalent to jointly optimize both loss functions for a fixed value of $\alpha$. In our experiments, the {\it trade-off} constant $\alpha$ is fine-tuned using cross-validation (see Figure \ref{fig:alphaie}).
\subsection{The transformed outcome}
The transformed outcome variable $Z$, which was proposed by \citep{athey2015machine} is defined as
\begin{equation}
Z = \frac{T Y}{e(\mathbf x)} - \frac{(1 - T) Y}{1 - e(\mathbf x)}.
\label{eq:transformed_outcome}
\end{equation}
This transformation has the key property that, under the complete randomization or unconfoundness assumption \citep{rosenbaum1983central}, its expectation conditional on $(\mathbf X=\mathbf x)$ is equal to $u(\mathbf x)$. Therefore, the transformed outcome is an unbiased estimate of the uplift. With this transformed outcome, one can simply use any supervised learning algorithm to estimate a regression function for the expectation of $Z$ given $\mathbf X$. In our case, with $e(\mathbf x) = 1/2$, $Z$ simplifies to
\[ Z = \begin{cases}
2, & \mathrm{if}~T=1, Y=1 \\
-2, & \mathrm{if}~T=0, Y=1 \\
0 & \mathrm{otherwise.}
\end{cases}
\]
In the uplift context, one can have an interesting interpretation of $Z=2$ and $Z=-2$. Take the first case. This happens when a client is called ($T = 1$) and then buys the insurance policy ($Y = 1$). Thus, for its counterfactual (i.e., $T = 0$), there are two possibilities. Either the client would have bought the policy ($Y = 1$) and in this case would belong to the \textit{sure things} group, or else the client would not have bought the policy ($Y = 0$) in which case would in fact be a \textit{persuadable}. Thus, in the worst case scenario, individuals with $Z = 2$ are \textit{sure things}. The same logic applies for $Z = -2$ where a client could either be a \textit{sure thing} or a \textit{do-not-disturb}. For \textit{lost cause} clients, by definition $Z = 0$. This interpretation does not cover all scenarios for $Z=0$ but gives insights about ordering customers with $Y=1$.
In the same spirit, we introduce $W$, another transformed outcome variable such as
\begin{equation}
W = \frac{(1-T)(1-Y)}{1-e(\mathbf x)} - \frac{T (1-Y)}{e(\mathbf x)}.
\label{eq:transformed_outcome_w}
\end{equation}
In our case, with $e(\mathbf x) = 1/2$, $W$ simplifies to
\[ W = \begin{cases}
2, & \mathrm{if}~T=0, Y=0 \\
-2, & \mathrm{if}~T=1, Y=0 \\
0 & \mathrm{otherwise.}
\end{cases}
\]
Here again, one can have an interesting interpretation of $W=2$ and $W=-2$. Take the first case. This happens when a client is not called ($T = 0$) and then do not buy the insurance policy ($Y = 0$). Thus, for its counterfactual (i.e., $T = 1$), there are two possibilities. Either the client would have bought the policy ($Y = 1$) and in this case would be a \textit{persuadable} client, or else the client would not have bought the policy ($Y = 0$) in which case would in fact be a \textit{lost cause}. Thus, in the worst case scenario, individuals with $W = 2$ are \textit{lost causes}. The same logic applies for $W = -2$ where a client could either be a \textit{lost cause} or a \textit{do-not-disturb}. Here, for \textit{sure things} clients, by definition $W = 0$. Once again, this interpretation does not cover all scenarios for $W=0$ but gives insights about ordering customers with $Y=0$.
We propose to combine $Z$ and $W$ in order to define a new random variable, which can be used as an unbiased estimator of the uplift, but for which the interpretation is straight forward. We define the random variable $R$ such as when $e(\mathbf x) = 1/2$,
\begin{equation}
R = \frac{1}{2} (Z + W) = \begin{cases}
1, & \mathrm{if} ~ T = Y \\
-1, & \mathrm{if} ~ T \neq Y
\end{cases}.
\label{eq:transformed_outcome_sign}
\end{equation}
Here, with regards to the 4 types of clients, we have that for $R=1$, the client is either a \textit{lost cause}, a \textit{sure thing} or a \textit{persuadable} and for $R=-1$, the client is either a \textit{do-not-disturb}, a \textit{lost cause} or a \textit{sure thing}. Therefore, the sign of $R$ and the sign of $u(\mathbf x)$ coincide.
\subsubsection{Regularized conditional means regression}
\label{sec:smite_to}
In deep learning regularization is sometimes hidden in heuristic methods during training. For instance, adding noise to the input data yields to generalization error improvement \citep{bishop1995training,rifai2011adding}. Data augmentation, and early stopping are some other heuristic regularizations widely applied in practice. A more theoretically sound regularization method is dropout \citep{srivastava2014dropout}, a widely-used method for addressing the problem of over-fitting. The idea is to drop units randomly from the neural network during training. It is known that dropout improves the test accuracy compared to conventional regularizers such as $L_1$ \citep{Tibshirani_lasso_1996} and $L_2$ \citep{srivastava2014dropout}. \citep{wager2013dropout} proved that dropout is equivalent to an $L_2$-type regularizer applied after scaling the inputs. These common regularizations encourage the weights to be estimated near zero. In our case, we want regularization on the predicted uplift (by the difference of the conditional means) and not directly applied to the weights. Therefore, we need to define a new regularization function.
Suppose $n$ independent customer observations $\{y_i, \textbf{x}_i, t_i, r_i\}_{i=1}^n$ are at our disposal. Let us denote the model's parameters by $\boldsymbol{\theta}$. Regularization can be seen as a constrained optimization problem, then using the transformed outcome $R$, SMITE solves
\begin{align}
{\boldsymbol{\hat\theta}} & = \argmin_{\boldsymbol{\theta}} ~ \mathcal{L}\Big({\mu_t}(\boldsymbol{\theta}), \mathbf y\Big) \nonumber \\
& \mathrm{subject~to,} \nonumber \\
& \mathcal{J}\Big({\mu_0}(\boldsymbol{\theta}),{\mu_1}(\boldsymbol{\theta}), \mathrm{ReLU}\Big) < \epsilon,
\label{eq:objective}
\end{align}
where $\epsilon > 0$. The quantities $\mu_{0}(\cdot)$ and $\mu_{1}(\cdot)$ are functions of the inputs $\mathbf x$ and could be denoted $\mu_{0}(\mathbf x)$ and $\mu_{1}(\mathbf x)$. To simplify the notation, we prefer to omit this specification throughout the rest of the paper. The rationale behind this constrained formulation is that we want a model capable of direct prediction of the uplift $u(\mathbf x)$, i.e., $\mu_1(\boldsymbol{\hat\theta}) - \mu_0(\boldsymbol{\hat\theta})$, and coherent estimation of $\mu_t(\mathbf x)$, i.e., $\mu_t(\boldsymbol{\hat\theta})$.
Formally, we define $\mathcal{L}(\boldsymbol{\theta}, \mathbf y)$ with respect to the model's parameters $\boldsymbol{\theta}$ as
\begin{equation*}
\mathcal{L}(\boldsymbol{\theta}, \mathbf y) = - \sum_{i=1}^{n} \Big( y_i\mathrm{log} \{\mu_{t_i}(\boldsymbol{\theta})\} + (1-y_i) \mathrm{log} \{1-\mu_{t_i}(\boldsymbol{\theta})\} \Big)
\end{equation*}
where $\mu_{t_i}(\boldsymbol{\theta}) = t_i \mu_{1i}(\boldsymbol{\theta}) + (1-t_i) \mu_{0i}(\boldsymbol{\theta})$ is the model's output of observation $i$ with treatment $t_i$, with $\mu_{1i}(\boldsymbol{\theta})$ and $\mu_{0i}(\boldsymbol{\theta})$ representing the outputs for observation $i$ from the left and right sub-components of the network (see Figure~\ref{fig:siamese_network}).
The function $\mathcal{J}(.)$ in \eqref{eq:objective} is what we refer to as the \textit{direct} uplift regularization function. It is defined as
\begin{equation}
\mathcal{J}(\boldsymbol{\theta}, \mathrm{ReLU}) = \frac{1}{n} \sum_{i=1}^{n} \mathrm{exp}\Big(- r_i \{{\mu_{1i}}(\boldsymbol{\theta }) - {\mu_{0i}}(\boldsymbol{\theta})\} \Big).
\label{eq:direct_uplift_loss}
\end{equation}
It is helpful to rewrite the full constrained problem (\ref{eq:objective}) in his Lagrangian form
\begin{align*}
{\boldsymbol{\hat\theta}}_{\lambda} = \argmin_{\boldsymbol{\theta}} \mathcal{L}(\boldsymbol{\theta}, \mathbf y) + \lambda \mathcal{J}(\boldsymbol{\theta}, \mathrm{ReLU})
\end{align*}
where $\lambda > 0$ is the regularization constant. However, in the SMITE framework, we use the transformation $\alpha = \lambda / (1+\lambda)$. This transformation allows to interpret $\alpha \in [0,1)$ as a hyper-parameter which controls the trade-off between the two losses in the optimization problem. Hence,
\begin{align}
{\boldsymbol{\hat\theta}}_{\alpha} = \argmin_{\boldsymbol{\theta}} (1 - \alpha) \mathcal{L}(\boldsymbol{\theta}, \mathbf y) + \alpha \mathcal{J}(\boldsymbol{\theta}, \mathrm{ReLU}).
\label{eq:SMITE_objective}
\end{align}
\subsubsection{The optimization algorithm}
\label{sec:smite_to_sgd}
Gradient-based optimization is one of the pillars of machine learning. Given an objective function $f: \rm I\!R^n \rightarrow \rm I\!R$, classical gradient descent has the goal of finding
(local) minima $\hat{\boldsymbol{\theta}} = \argmin_{\boldsymbol{\theta}} f(\boldsymbol{\theta})$ via updates of the form $ \Delta (\boldsymbol{\theta}) = -\xi \nabla f$, where $\xi$
is a positive step size (or the learning rate). Two main classes of optimization methods can be applied to solve such problems : (i) {\it stochastic} and (ii) {\it batch}.
The prototypical stochastic optimization method is the stochastic gradient
method \citep{robbins1951stochastic}, which, in the context of minimizing $f(\boldsymbol{\theta})$ and with $\boldsymbol{\theta}^{(0)} \in \rm I\!R^p$ given, is defined by
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \xi \nabla f_{i_k} (\boldsymbol{\theta}^{(k)}).
\end{equation}
Here, for all $k \in \mathbb{N}$, the index $i_k$ is chosen randomly from $\{1,...,n\}$. There is no particular reason to employ information from only one sample point per iteration. Instead, one can employ a {\it mini-batch} approach in which a small subset of samples $\mathcal{S}_k \subseteq \{1, ..., n\}$, is chosen randomly in each iteration, leading to
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \frac{\xi}{|\mathcal{S}_k|} \sum_{i \in \mathcal{S}_k} \nabla f_{i} (\boldsymbol{\theta}^{(k)}).
\label{eq:mini-batch-sgd}
\end{equation}
Here, $|\mathcal{S}_k|$ is the batch size. Such a mini-batch stochastic gradient method has been widely used in practice. On the other hand, the {\it batch} approach, also referred to as the gradient, batch gradient, or full gradient method, uses all $n$ observations in each iteration. There are some fundamental practical and theoretical reasons why stochastic methods have inherent advantages for large-scale machine learning \citep{bottou2018optimization}. In practice, we train the SMITE model by solving (\ref{eq:SMITE_objective}) using the mini-batch stochastic gradient descent (SGD) updates (\ref{eq:mini-batch-sgd}).
Optimization algorithms that use only the gradient are called first-order optimization algorithms. Second-order optimization algorithms such as Newton's method use the Hessian matrix. In this work, as in many deep learning applications where the functions used are complicated, we focus on first-order methods.
\subsection{Exploring other regularization functions}
\label{sec:smite_ie}
The flexibility of SMITE allows us to offer alternatives to the regularization (and loss) functions defined in the general framework (see Figure~\ref{fig:nite}).
SMITE allows to have for each forward pass the estimation of both conditional means. By taking the difference, we get an estimate of the uplift. Now, we want to find a way to regularize this estimate using information from randomization, i.e., the treatment variable $T$. Indeed, suppose a treatment has a significant positive (resp. negative) effect on a sub-sample of customers. This would imply that within the sample of customers that had a positive (resp. negative) response, we expect a higher proportion of treated customers.
For example, suppose the covariate $X$ is a binary variable which represents the gender of each customer and assume the distributions in Table~\ref{tab:example_ie}. Based on the $200$ customers, the average treatment effect is $0$. However, the treatment effect has positive impact on men. Indeed, the uplift for men is $0.1$. Therefore, if we only look at men with $Y=1$, we should observe that more than half ($e(\mathbf x) = 1/2$) were treated. In this example, the proportion is $2/3$.
\begin{table}[H]
\centering
\caption{Example}
\begin{tabular}{|l|cc|cc|cc|}
\hline
& \multicolumn{2}{c|}{Total} & \multicolumn{2}{c|}{$Y=1$} & \multicolumn{2}{c|}{$Y=0$} \\
\hline
& M & F & M & F & M & F \\
\hline
Treatment ($T=1$) & 50 & 50 & 10 & 0 & 40 & 50 \\
Control ($T=0$) & 50 & 50 & 5 & 5 & 45 & 45 \\
\hline
\end{tabular}
\label{tab:example_ie}
\end{table}
More concretely, suppose the data $(y_i, t_i)$, for $i=1,\dots,n$ represents i.i.d. samples from the joint distribution of a pair of binary random variables $(Y,T)$, for which conditional on $\mathbf X=\mathbf x$, $\mathrm{Pr}(T = 1 \mid \mathbf x) = 1 - \mathrm{Pr}(T = 0 \mid \mathbf x) = e(\mathbf x)$. Then, by Bayes theorem,
\begin{align*}
\mathrm{Pr}(T=1 \mid Y=1, \mathbf x) = \frac{ \mathrm{Pr}(Y=1 \mid T=1, \mathbf x) e(\mathbf x)}{\mathrm{Pr}(Y=1 \mid T=0, \mathbf x) \{1- e(\mathbf x)\} + \mathrm{Pr}(Y=1 \mid T=1, \mathbf x)e(\mathbf x)} .
\end{align*}
Inspired by the previous development, we denote the proportion of treated customers among those that had positive outcomes by $p(t=1\mid y=1,\mathbf x)$. When the propensity score is equal to the constant $e(\mathbf x) = 1/2$, this proportion can be written as a function of the conditional means,
\begin{align}
p(t=1\mid y=1,\mathbf x) = \frac{\mu_1(\mathbf x)}{\mu_0(\mathbf x) + \mu_1(\mathbf x)}.
\label{eq:p1}
\end{align}
Although Equation \eqref{eq:p1} is simplified based on the assumption assumption that $e(\mathbf x) = 1/2$, the development holds for any constant $e(\mathbf x)$. Moreover, one can use an under or over-sampling procedure to recover the $e(\mathbf x) = 1/2$ case. Similarly, the proportion of treated customers among those that had negative outcomes is
\begin{align}
p(t=1\mid y=0,\mathbf x) = \frac{(1-\mu_1(\mathbf x))}{(1-\mu_0(\mathbf x)) +(1- \mu_1(\mathbf x))}.
\label{eq:p0}
\end{align}
From these definitions, we propose an alternative to the direct uplift regularization. The \textit{indirect} uplift regularization has the following form
\begin{equation*}
\mathcal{I}(\boldsymbol{\theta},
\mathbf t) = - \sum_{i=1}^{n} \Big( t_i\mathrm{log} \{p_{y_i}(\boldsymbol{\theta})\} + (1-t_i) \mathrm{log} \{1-p_{y_i}(\boldsymbol{\theta})\} \Big),
\end{equation*}
where, $p_{y_i}(\boldsymbol{\theta}) = y_i p_{1i}(\boldsymbol{\theta}) + (1-y_i) p_{0i}(\boldsymbol{\theta})$ is the model's output of observation $i$ with response $y_i$, with $p_{1i}(\boldsymbol{\theta})$ and $p_{0i}(\boldsymbol{\theta})$ representing the estimates for observation $i$ of \eqref{eq:p1} and \eqref{eq:p0}, such as
\begin{align*}
p_{1i}(\boldsymbol{\theta}) &= \frac{\mu_{1i}(\boldsymbol{\theta})}{\mu_{0i}(\boldsymbol{\theta}) +\mu_{1i}(\boldsymbol{\theta})}, \\
p_{0i}(\boldsymbol{\theta}) &= \frac{(1-\mu_{1i}(\boldsymbol{\theta}))}{(1-\mu_{0i}(\boldsymbol{\theta})) +(1- \mu_{1i}(\boldsymbol{\theta}))}.
\end{align*}
The indirect uplift regularization function is simply the binary cross-entropy between a function of the predicted conditional means $p_y(\boldsymbol{\theta})$ and the observed treatment variable $t$.
In what follows, we will refer to the SMITE using $\mathcal{J}$ as SMITE Transformed Outcome (TO) and the SMITE using $\mathcal{I}$ as SMITE Indirect Estimation (IE). In this case, the optimization problem becomes
\begin{align*}
{\boldsymbol{\hat\theta}}_{\alpha} = \argmin_{\boldsymbol{\theta}} (1 - \alpha) \mathcal{L}(\boldsymbol{\theta}, \mathbf y) + \alpha \mathcal{I}(\boldsymbol{\theta}, \mathbf t),
\end{align*}
for $\alpha \in [0,1)$.
\subsection{SMITE likelihood}
Consider the random variables $Z_1 = YT - Y(1-T)$ and $Z_0 = (1-Y)T - (1-Y)(1-T)$.
We have
\[ Z_1 = \left\{ \begin{array}{cc} 1, & Y=1, T=1, \\ -1, & Y=1, T=0\\ 0 & \sim \end{array} \right.
\]
and
\[ Z_0 = \left\{ \begin{array}{cc} 1, & Y=0, T=1, \\ -1, & Y=0, T=0\\ 0 & \sim \end{array} \right.
\]
Now consider the random variable
\[ Z = \left\{ \begin{array}{cc} Z_1, & Y=1 \\ Z_0, & Y=0 \end{array} \right.
\]
Note that $Z= 2T -1 $.
We consider the case $P(T=1)=P(T=0) = 0.5$.
Let $m_{yt} = P( Y=y | T=t)$, and $p_{yt} = m_{yt} / ( m_{yt} + m_{y(1-t)})$.
We have
\begin{align*}
P( Z = 1 | Y=1 ) & = P( Z =1 , Y=1 ) / P(Y = 1) \\
& = \tfrac{\biggl\{P( Z=1, Y=1, T=1 ) + P(Z=1, Y=1, T=0)\biggr\}}{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_1=1 | Y=1, T=1)P(Y=1 | T=1) P(T=1) + P(Z_1=1 | Y=1, T=0)P(Y=1 | T=0) P(T=0) \biggr\}}
{\biggl\{ P(Y=1|T=1)P(T=1) + P(Y=1 | T=0) P(T=0) \biggr\} } \\
& = m_{11} / ( m_{11} + m_{10}) = p_{11}.
\end{align*}
Similarly,
\begin{align*}
P( Z = -1 | Y=1 ) & = P( Z =-1 , Y=1 ) / P(Y = 1) \\
& = \tfrac{\biggl\{P( Z=-1, Y=1, T=1 ) + P(Z=-1, Y=1, T=0)\biggr\}}{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_1=-1 | Y=1, T=1)P(Y=1 | T=1) P(T=1) + P(Z_1=-1 | Y=1, T=0)P(Y=1 | T=0) P(T=0) \biggr\}}
{\biggl\{ P(Y=1|T=1)P(T=1) + P(Y=1 | T=0) P(T=0) \biggr\}} \\
& = m_{10} / ( m_{11} + m_{10}) = 1 - p_{11}.
\end{align*}
Proceeding in the same way, we have
\begin{align*}
P( Z = 1 | Y=0 ) & = P( Z =1 , Y=0 ) / P(Y = 0) \\
& = \tfrac{\biggl\{P( Z=1, Y=0, T=1 ) + P(Z=1, Y=0, T=0)\biggr\} }{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_0=1 | Y=0, T=1)P(Y=0 | T=1) P(T=1) + P(Z_0=1 | Y=0, T=0)P(Y=0 | T=0) P(T=0) \biggr\}}
{\biggl\{ P(Y=0|T=1)P(T=1) + P(Y=0 | T=0) P(T=0) \biggr\}} \\
& = m_{01} / ( m_{01} + m_{00}) = p_{01},
\end{align*}
and
\begin{align*}
P( Z = -1 | Y=0 ) & = P( Z =-1 , Y=0 ) / P(Y = 0) \\
& = \tfrac{\biggl\{P( Z=-1, Y=0, T=1 ) + P(Z=-1, Y=0, T=0)\biggr\}}{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_0=-1 | Y=0, T=1)P(Y=0 | T=1) P(T=1) + P(Z_0=-1 | Y=0, T=0)P(Y=0 | T=0) P(T=0) \biggr\} }
{\biggl\{ P(Y=0|T=1)P(T=1) + P(Y=0 | T=0) P(T=0) \biggr\} }\\
& = m_{00} / ( m_{01} + m_{00}) = 1 - p_{01}.
\end{align*}
Note that $Z = 1$ iff $T=1$ and $Z=-1$ iff $T=0$. This fact will be useful to write the likelihood of the complete data.
This is given by
\begin{multline*}
\prod_{i=1}^n P( z_i, y_i, t_i) = \prod_{i=1}^n P( z_i | y_i) P( y_i | t_i) P(t_i) \\
= 2^{-n}\prod_{i=1}^n p_{11}^{t_i y_i} (1 - p_{11})^{(1-t_i)y_i} p_{01}^{t_i (1-y_i)} (1 - p_{01})^{(1-t_i)(1-y_i)}
m_{11}^{t_i y_i} (1 - m_{11})^{t_i (1-y_i)} m_{10}^{(1-t_i) y_i} (1 - m_{10})^{(1-t_i) (1-y_i)} \\
= 2^{-n}\prod_{i=1}^n p_{y_i 1}^{t_i } (1 - p_{y_i 1})^{(1-t_i)} m_{1t_i}^{y_i} (1 - m_{1t_i})^{(1-y_i)} \\
= 2^{-n} \exp\{ \sum_{i=1}^n t_i \log p_{y_i 1} + (1 - t_i)\log(1 - p_{y_i 1}) \}
\times \exp\{ \sum_{i=1}^n y_i \log m_{ 1 t_i} + (1 - y_i)\log(1 - m_{1 t_i }) \}.
\end{multline*}
\section{SMITE neural network}
\input{figures/nnsmite.tex}
\subsection{Universal approximation theorems}
A neural network (NN; for a definition, see e.g. \cite{pinkus1999approximation} or \cite{calin2020deep}) transforms an input $x\in D\subset\mathbb{R}^{n}$ into an output $\widetilde{y}=\Phi\left( x\right) \in\mathbb{R}$. As such it is a concrete real function of $n$ real variables, $n$ being a priori an arbitrary positive integer. A natural question that has been asked for a long time is the following: can this NN (with an appropriate architecture) approximate functions of a given class (common example: multivariate continuous functions defined on $D=K$, a compact subset of $\mathbb{R}^{n}$, but more general function spaces, such as $L^{p}$ spaces, are usually the proper framework from an abstract mathematical viewpoint) \citep{mhaskar1996system}? In other words, can a NN generate a topologically dense subset of the function class in question? For example, in the case of a normed vector space, given $\epsilon>0$, is there and, if so, can one actually construct at a reasonable cost a NN such that
\[
\left\Vert f-\widetilde{y}\right\Vert <\epsilon\text{ ?}%
\]
A related question (among many other researchers are addressing \citep{fan2020universal}) is: to what extent does the choice of the norm (topology) affect this existence and/or construction?
The proper mathematical interpretation of NN is based on the universal
approximation theorems (UAT). The UAT are roughly divided into two categories:
\begin{enumerate}
\item arbitrary number of neurons in the case of {\it shallow} neural networks (SNN, arbitrary width). This is the classical case. It goes back to the pioneering work of the 1940-1980's;
\item arbitrary number of layers with {\it deep} neural networks (DNN, arbitrary
depth). In practice, the latter have proven to be more efficient
\citep{mhaskar2019function}. It is always possible to build a SNN to emulate a DNN. However the number of neurons grows exponentially in order to obtain a
comparable performance (in terms of error) \citep{telgarsky2016benefits,eldan2016power}.
\end{enumerate}
In this sense, the UAT are representations of classical existence theorems. They implement the construction (i.e. find the appropriate weights) of the approximation of large classes of functions \citep{paluzo2020two}. However, they do not say how one explicitly (with a formula or an algorithm) finds this approximation (weights).
We give a brief summary of some of the conceptual steps that led to today's mathematical understanding of the UAT. For a thorough and extensive historical
bibliography on NN, \cite{schmidhuber2015deep} compiled 888
references. See \cite{goodfellow2016deep} for an elementary well documented
introduction to the subject.
\begin{itemize}
\item The Kolmogorov--Arnold representation (or superposition) theorem \citep{kolmogorov1957representation} states that any continuous real multivariate function can be represented by a superposition of real one-variable continuous functions $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$:
\[
f(x) =\sum_{i=0}^{2n}\Phi_{i}\left( \sum_{j=1}^{n}\phi
_{i,j}(x_{j}) \right)
\]
This theorem gives a partial answer to Hilbert's 13th problem \citep{akashi2001application,khesin2014arnold}. A construction is proposed by \cite{polar2020urysohn}.
\item Prior to the 1980s \citep{mcculloch1943logical,fukushima1975cognitron}, NN architectures were shallow. An understanding of SNN gives a good appreciation of DNN.
\item The Kolmogorov-Arnold theorem gives an exact representation of the function $f(x)$. This representation proves to be very complex (fractal nature) and computationally very slow, therefore unusable in NN. The accuracy condition had to be relaxed for an approximation condition ($L^{p}$
or uniform norm). The optimal solution is sought by iterative methods (the
gradient descent, introduced by Jacques Hadamard in 1908, being the most
popular). Pioneering modern work is due to \cite{cybenko1989approximation} for
sigmoid activation functions (arbitrary width), \cite{funahashi1989approximate}, \cite{hornik1991approximation} (he showed that UAT was
potentially constructible using a multilayer architecture of the NN rather
than a specific choice of the activation function), \cite{barron1993universal} (bounds on smooth function errors), etc...
\item \cite{leshno1993multilayer} and \cite{pinkus1999approximation} have
shown that the universal approximation property (see \cite{kratsios2019universal} for a
definition) is equivalent to a non polynomial activation function.
\end{itemize}
\begin{theorem}
(\cite{cybenko1989approximation,hornik1991approximation,pinkus1999approximation}). Let $n\in\mathbb{N}$, $K\subset\mathbb{R}^{n}$ compact, $f:K\rightarrow\mathbb{R}$ continuous, $\rho:\mathbb{R}\rightarrow\mathbb{R}$ non polynomial and continuous, and let $\epsilon>0$. Then there exists a
two-layer NN $\Phi$ with activation function $\rho$ such that
\[
\left\Vert \Phi-f\right\Vert _{\infty}<\epsilon.
\]
\end{theorem}
We also note that DNN have demonstrated their capabilities in the approximation of functions \citep{liang2016deep,yarotsky2017error,fan2020universal}. They are found to have good experimental (empirical) performance, but it is largely unknown why. This is a very active area of research \cite{caterini2018deep}.
\begin{theorem}
(\cite{lu2017expressive}) Let $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$ be Lebesgue-integrable and $\epsilon>0$. Then there exists a ReLU
network of width $\leq n+4$ such that
\[
\int_{\mathbb{R}^{n}}\left\vert \Phi\left( x\right) -f\left( x\right) \right\vert dx<\epsilon
\]
\end{theorem}
\begin{definition}
(\cite{kidger2020universal}) Let $\rho:\mathbb{R}\rightarrow\mathbb{R}$ and $n,m,k\in\mathbb{N}$. Then let $NN_{n,m,k}$ represent the class of functions $\mathbb{R}^{n}\rightarrow\mathbb{R}^{m}$ described by feedforward neural networks with $n$ neurons in the input layer, $m$ neurons in the output layer, and an arbitrary number of hidden layers, each with $k$ neurons with activation function $\rho$. Every neuron in the output layer has the identity activation function.
\end{definition}
\begin{theorem}
(\cite{kidger2020universal}) Let $\rho:\mathbb{R}\rightarrow\mathbb{R}$ be any non-affine continuous function which is continuously differentiable at
at least one point, with non-zero derivative at that point. Let $K\subset\mathbb{R}^{n}$ be compact. Then NN$_{n,m,n+m+2}$ is dense in $C\left( K,\mathbb{R}^{m}\right)$ with respect to the uniform norm.
\end{theorem}
\subsection{Low-rank approximation}
Due to its fundamental role in Machine Learning, a huge literature is devoted
to the singular value decomposition (SVD) of a matrix. We refer to some of the
well respected texts: \cite{strang1993introduction}, \cite{trefethen1997numerical}, \cite{watkins2004fundamentals}, \cite{gentle2007matrix}, \cite{elden2007matrix}, \cite{lange2010numerical},
and \cite{yanai2011projection} for
instance. Let $A\in \mathbb{R}^{m\times n}$ be a matrix of rank $r$. Its SVD consists in factoring $A$ as
\[
A=U\Sigma V^\top=\sum_{i=1}^{r}\sigma_{i}\mathbf u_{i}\mathbf v_{i}^\top
\]
where $U\in\mathbb{R}^{m\times m}$ and $V\in\mathbb{R}^{n\times n}$ are orthogonal square matrices, and $\Sigma\in\mathbb{R}^{m\times n}$ (\textquotedblleft diagonal\textquotedblright\ rectangular matrix of the ordered singular values: $\sigma_{1}\geq\ldots\geq\sigma _{r}>\sigma_{r+1}=\ldots=0=\sigma_{\min\{m,n\}}$). This
decomposition, and the related principal component analysis (PCA), has several
uses in Data Analysis. Written as $AV=U\Sigma$, SVD says that there exist orthogonal matrices $U$ and $V$ such that $A$ maps the columns $\mathbf v_{i}$ of $V$ (right singular vectors of $A$, directions or principal axes in the PCA parlance) into the columns $\mathbf u_{i}$ of $U$ (left singular vectors of $A$) such as
\[
A\mathbf v_{i}=\sigma_{i}\mathbf u_{i},
\]
the principal components.
Given a positive integer $1\leq k\leq r$, introduce the truncated matrices
$U_{k}=\left[ \mathbf u_{1},\ldots,\mathbf u_{k}\right] \in \mathbb{R}^{m\times k}$, $V_{k}=\left[ \mathbf v_{1},\ldots,\mathbf v_{k}\right] \in \mathbb{R}^{k\times n}$, and $\Sigma_{k}=\mathrm{diag}( \sigma_{1},\ldots,\sigma_{k}) \in \mathbb{R}^{k\times k}$. Then the matrix of rank $k$
\[
A_{k}=U_{k}\Sigma_{k}V_{k}^\top=\sum_{i=1}^{k}\sigma_{i}\mathbf u_{i}\mathbf v_{i}^\top
\]
is an optimal lower-rank approximation of $A$ in the sense of both the
Frobenius norm and the $L_2$-norm. Indeed, the Eckart--Young--Mirsky theorem (for
a proof, see \cite{markovsky2012low}) states that among all matrices $B\in\mathbb{R}^{m\times n}$ of a given rank $k\leq r$, the one closest to $A$ is
$\hat{A}=A_{k}$, i.e.
\begin{align*}
\inf_{\mathrm{rank}(B)=k}\left\Vert A-B\right\Vert _{F} & =\left\Vert
A-\hat{A}\right\Vert _{F}=\sqrt{\sum_{i=k+1}^{r}\sigma_{i}^{2}},\\
\inf_{\mathrm{rank}(B)=k}\left\Vert A-B\right\Vert _{2} & =\left\Vert
A-\hat{A}\right\Vert _{2}=\sigma_{k+1}.
\end{align*}
Furthermore, the approximating matrix $\hat{A}$ is unique if and only if
$\sigma_{k+1}\neq\sigma_{k}$ \citep{markovsky2012low}. The theorem was proved in 1907 by Schmidt (of the Gram-Schmidt procedure) for the Frobenius norm \citep{leon2013gram}. A constructive solution was given by \cite{eckart1936approximation}. This solution was
shown by \cite{mirsky1960symmetric} to hold for all unitarily invariant norms. A further generalization addressing collinearity issues was proposed by \cite{golub1987generalization}. See \cite{stewart1993early} for historical details.
One interpretation of the decomposition $A_{k}=U_{k}\Sigma_{k}V_{k}^{T}$ is as
a compression/learning linear procedure of the data matrix $A$ (in other
words, a dimension reduction method). Indeed, since the SVD extracts data in
the direction with the highest variance (and going down to those with the
lowest variance), matrix $U_{k}$ is the compression matrix in the first $k$
directions, matrix $\Sigma_{k}$ contains the amount of error (cost function)
we tolerate, and $V_{k}$ is the decompression matrix. If we keep all the non
zero singular values ($k=r$), there is theoretically no loss in the
compression (assuming that the data in $A$ are exact). If we drop the small
singular values (set them to $0$), it is not possible to reconstruct the
initial matrix $A$ exactly, but we obtain a matrix $A_{k}$ that is close. In
other words, the SVD decomposition is a way to control the loss of information
from $A$ under a predetermined level.
The assumption (routinely validated in image processing, for example) in the
above approach is that the components (columns of $U$) of the small singular
values contain little information (mostly \textquotedblleft
noise\textquotedblright) and should be discarded. However, the low-rank matrix approximation is easily seen not to be optimal if used to solve the linear regression problem (hence as a predictive tool) $y=A\beta+\epsilon$, that is to minimize
\begin{equation}
\inf_{\mathrm{rank}(B)=k}\left\Vert \mathbf y-B\boldsymbol \beta\right\Vert .
\label{eq:svd_regression}
\end{equation}
While the SVD picks the parameter $k-$vector $\boldsymbol \beta$ of minimum norm
\citep{elden2007matrix,lange2010numerical},
\begin{equation}
\hat{\boldsymbol \beta}=A_{k}^{\dag} \mathbf y=\sum_{i=1}^{k}\frac{1}{\sigma_{i}} \mathbf u_{i}^\top \mathbf y \mathbf v_{i},
\label{eq:svd_reg_solution}
\end{equation}
where $A_{k}^{\dag}$ is the pseudoinverse of $A_{k}$, the general solution of
\eqref{eq:svd_regression} is unknown \citep{xiang2012optimal}. See also \cite{talwalkar2010matrix} for a discussion in the Machine Learning context.
Some authors \citep{jolliffe1982note,hadi1998some} warned against the misconceived (mainly due to collinearity considerations) use of \eqref{eq:svd_reg_solution} in an ordinary linear regression setting. Dropping some of the small singular values may result in ignoring significant variables, mainly because the approximation does not take into account the response variable $\mathbf y$. A partial solution is provided by the ridge regression \citep{hoerl1970ridge}. However, still from a predictive viewpoint, SVD may prove to be an interesting approach in deep neural networks \citep{bermeitinger2019singular}. Used for initialization in the first layer, SVD gave promising experimental results. Theoretical issues still need to be addressed.
\subsection{Low-rank and SMITE}
\textit{We do not necessarily study low rank approximations of the matrix of weights. For this you will need indeed to get bounds on the distance between the estimated function and the approximation function. In Statistics, usually we just measure the statistical difference between two models (and not the distance between the two models). So instead of thinking of a reduced approximation of the weight matrix, we might just think of a model with a lower-rank matrix of weights. This in general will lead to a non-parametric estimator with lower variance and hopefully, only slightly increased bias. In this setup, methods to measure the statistical difference between the full model and the reduced model should be applied for each dataset. This is in contrast to the general approximation theory, where bounds on the approximation error over all possible datasets need to be found.}
We are interested in determining how complex the network ought to be to theoretically guarantee that the network would approximate an unknown target function $f$ up to a given accuracy $\epsilon > 0$. To measure the accuracy, we need a norm $\norm{\cdot}$ on some normed linear space. We work with $L_2$ norm.
We focus on a feed-forward neural network with one fully-connected hidden layer. Here the input layer (with $n$ neurons) and the hidden layer (with $m$ neurons) are assumed to be fully connected with a weight matrix $\mathbf W \in \rm I\!R^{n \times m}$ with rank at most $p \leq \mathrm{min}(n,m)$. The domain for the input vector $\mathbf x$ is the $n$-dimensional hypercube $I^n = [0, 1]^n$, and the output layer only contains one neuron. The neural network can be expressed as
$$\hat y = \hat f(\mathbf x\mid \mathbf W) = \sum_{j=1}^m \alpha_j h(\mathbf w_j^\top\mathbf x + \beta_j) = \boldsymbol{\alpha} h(\mathbf x \mathbf{W} + \boldsymbol{\beta}).$$
Here $h(\cdot)$ is the activation function, $\mathbf w_j \in \rm I\!R^n$ denotes the $j$-th column of the weight matrix $\mathbf W$, and $\alpha_j, \beta_j \in \rm I\!R$ for $j = 1, \ldots, m$. Suppose $p$ is the rank of the matrix $\mathbf W$. We can denote it $\mathbf W_p$ and define the following trained model
$$\hat y_p = \boldsymbol{\alpha} h(\mathbf x \mathbf{W_p} + \boldsymbol{\beta}).$$
Let $r < p$ be an integer and define
$$\hat y_r = \boldsymbol{\alpha} h(\mathbf x \mathbf{\hat W_r} + \boldsymbol{\beta}),$$
be the prediction from the model where $\mathbf{W_p}$ has been approximated with $\mathbf{\hat W_r}$ of rank $r$, that is, $\mathbf{\hat W_r} = \mathbf{\hat U \hat \Sigma_r \hat V^{\top}}$ where $\mathbf{\hat{U}, \hat{\Sigma_r}, \hat{V}^\top} \in \rm I\!R^{n\times r}, \rm I\!R^{r\times r}, \rm I\!R^{r \times m}$. This approximates the original weights $\mathbf W_p$ into reconstructed approximated weights $\mathbf{\hat{W_r}}$.
The truncation process always keeps the first $r$ rows of $\mathbf U$ and columns of $\mathbf V^\top$ (see Figure~\ref{fig:svd}). The number of non-zero parameters in the truncated SVD is $r( n + m + 1)$. The compression ratio is the ratio between parameters after and before the decomposition which is calculated as $r( n + m + 1)/( nm )$. In addition, the weight matrix does not need to be reconstructed during inference where the input $\mathbf x$ can be directly multiplied with $\mathbf{\hat{U}, \hat{\Sigma_r}, \hat{V}^\top}$ matrices, which takes lesser computation than direct multiplication of $\hat{\mathbf W_r}$ and $\mathbf x$. Therefore, we have the following development
\begin{align*}
\norm{ \hat y_p - \hat y_r }_2 &= \norm{\boldsymbol{\alpha} h(\mathbf x \mathbf{W_p} + \boldsymbol{\beta}) - \boldsymbol{\alpha} h(\mathbf x \mathbf{\hat W_r} + \boldsymbol{\beta})}_2 \\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{h(\mathbf x \mathbf{W_p} + \boldsymbol{\beta}) - h(\mathbf x \mathbf{\hat W_r} + \boldsymbol{\beta})}_2\\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{\mathbf x \mathbf{W_p} - \mathbf x \mathbf{\hat W_r}}_2 ~ \text{because $h(\cdot)$ is $1$-Lipschitz continuous}\\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{\mathbf x}_2 \norm{\mathbf{W_p} - \mathbf{\hat W_r}}_2\\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{\mathbf x}_2 \sigma_{r+1} ~ \text{by Eckart-Young Theorem \citep{eckart1936approximation}}\\
&\leq \Big( \sum_{j=1}^m \alpha_j ^ 2 \Big)^{1/2} \sqrt{n} \sigma_{r+1}
\end{align*}
where $\sigma_{r+1}$ is the $(r+1)$th singular value of $\mathbf{W_p}$.
\section{Evaluating uplift models}
The goal of uplift modeling is to separate all observations into heterogeneous segments that differ as much as possible with regard to the reactions to the treatment and, thus, identify (sub)groups with very different uplifts. Then, the uplifts in the resulting segments serve as a prediction for the future behavior of the observations. However, the \textit{ground truth} is not observed in the case of individual treatment effect estimation. \textcolor{blue}{Estimating the true metric values is not always necessary to conduct valid model selection or hyperparameter tuning. It maybe possible to obtain a better evaluation metric under an objective specific to selection and tuning \citep{saito2019counterfactual}.} Therefore, in order to evaluate uplift models, we refer to known metrics in the uplift literature, the {\it Qini coefficient} and the {\it Kendall's uplift rank correlation}. The computation of both statistics requires the construction of the {\it Qini curve} \citep{radcliffe2007using}. These statistics are commonly used as goodness-of-fit measures for uplift models.
\textcolor{blue}{In previous studies \citep{gutierrez2017causal, schuler2018comparison, alaa2019validating}, the evaluation of an uplift predictor $\hat{u}(\cdot)$ has been formulated as accurately estimating the following ground-truth performance metric $\mathcal{R}(\hat{u}) = \mathbbm{E}[( \hat u - u )^2]$. This metric is also termed as the expected precision in estimation of heterogeneous effect (PEHE) \citep{hill2011bayesian}. This approach is intuitive and ideal. However, realizations of the true uplift are never observable, and thus, accurate performance estimation is difficult. A natural workaround is to replace $\hat{u}(\cdot)$ with an estimate derived from a validation set $\mathcal{V}$ such as $\mathcal{\Tilde{R}}(\hat{u}) = \frac{1}{|\mathcal{V}|}\sum_{i \in \mathcal{V}}( \hat{u}_i - \Tilde{u}_i )^2$, where $\Tilde{u}$ is a {\it plug-in} estimate of $u$ estimated using data in the validation set.}
For a given model, let $\hat{u}_{(1)} \geq \hat{u}_{(2)} \geq ... \geq \hat{u}_{(n)}$ be the sorted predicted uplifts. Let $\phi \in [0,1]$ be a given proportion and let $ N_{\phi} = \{i: \hat{u}_{i} \geq \hat{u}_{\ceil{\phi n}} \} \subset \lbrace 1, \ldots, n \rbrace$ be the subset of observations with the $\phi n \times 100 \%$ highest predicted uplifts $\hat{u}_i$ (here $\ceil{s}$ denotes the smallest integer larger or equal to $s\in \rm I\!R$). The \textit{Qini curve} is defined as a function $f$ of the fraction of population targeted $\phi$, where
\begin{equation*}
f(\phi) = \frac{1}{n_t}\biggl(\sum\limits_{i \in N_{\phi}} y_i t_i - \sum\limits_{i \in N_{\phi}} y_i (1-t_i) \biggl\{ \sum\limits_{i \in N_{\phi}} t_i / \sum\limits_{i \in N_{\phi}} (1-t_i) \biggr\} \biggr),
\end{equation*}
where $n_t = \sum_{i=1}^n t_i$ is the number of treated customers, with $f(0)=0$ and $f(1)$ is the average treatment effect. In other words, $f(\phi)$ represents the incremental number of positive responses for a fraction $\phi$ or targeted customers relative to the total number of targeted customers in the sample. The Qini helps visualize the performance of an uplift model, a bit like the ROC curve of a binary classification model. Following the same logic, a straight line between the points $(0,0)$ and $(1, f(1))$ defines a benchmark to compare the performance of the model to a strategy that would randomly target individuals. This means that if a proportion $\phi$ of the population is targeted, we expect to observe an uplift of $\phi f(1)$.
The {\it Qini coefficient} is defined as
\begin{equation*}
q = \int_0^1 Q(\phi) ~\mathrm{d}\phi = \int_0^1 \{f(\phi) - \phi~f(1)\} ~\mathrm{d}\phi,
\end{equation*}
where $Q(\phi) = f(\phi) - \phi~f(1)$. This area can be numerically approximated using a Riemann method such as the trapezoid rule formula: the domain of $\phi \in [0,1]$ is partitioned into $K$ bins, or $K+1$ grid points $0=\phi_1 < \phi_2 < ... < \phi_{K+1} = 1$, to approximate $q$ by
\begin{equation}
\hat{q} = \dfrac{1}{2} \sum_{k=1}^K (\phi_{k+1}-\phi_k)\{Q(\phi_{k+1}) + Q(\phi_{k})\}.
\label{eq:q:hat}
\end{equation}
It is important to note, unlike the area under the ROC curve, $\hat{q}$ could be negative. This would simply mean that a strategy following the model in question does worse than random targeting.
Measuring the goodness-of-fit of an uplift model can also be achieved by the similarity between the theoretical uplift percentiles (given by the model's predictions) and empirical percentiles (observed in the data) based on the model. This can be approximated by the {\it Kendall's uplift rank correlation} \citep{belba2019qbased} defined as
\begin{equation}
\hat{\rho} = \frac{2}{K(K-1)} \sum_{i<j} \mathrm{sign}(\bar{\hat{u}}_i - \bar{\hat{u}}_j)~\mathrm{sign}(\bar{u}_i - \bar{u}_j),
\label{eq:corr_coeff}
\end{equation}
where $\bar{\hat{u}}_k$ is the average predicted uplift in bin $k$, $k \in {1,...,K}$, and $\bar{u}_k$ is the observed uplift in the same bin.
\section{Experiments}
\label{sec:experiments}
We conduct a simulation study to examine the performance SMITE. We compare its performance with existing uplift methods using three types of datasets: i) a real-world dataset from a marketing campaign (see Section~\ref{sec:application} for the analysis); ii) bootstrapped versions of the real-world dataset with synthetic outcomes generated from tree-based models, and iii) synthetic datasets generated from parametric models. First, we show results for the SMITE architecture search.
\subsection{Fine-tuning SMITE}
Deep Learning has enabled remarkable progress over the last years on a variety of tasks, such as image recognition, speech recognition, and machine translation. One crucial aspect
for this progress are novel neural architectures. Currently employed architectures have mostly been developed manually by human experts, which is a time-consuming and error-prone process. Because of this, there is growing interest in automated neural architecture search (NAS) methods. NAS methods can be categorized according to three dimensions \citep{elsken2018neural}: search space, search strategy, and performance estimation strategy.
\paragraph{Search space.}
The search space defines which architectures can be represented
in principle. Incorporating prior knowledge about typical properties of architectures well-suited for a task can reduce the size of the search space and simplify the search. SMITE architecture is a Sieamese multi-layer perceptron. Therefore, the search space for the architecture is first defined by the number of hidden layers and the number of hidden units per layer. However, it should not be limited to these two parameters. The choice of loss functions, regularization and activations in deep networks has also a significant effect on the training dynamics and task performance. These functions are therefore part of the architecture and should be included in the search space. For example, instead of using (\ref{eq:direct_uplift_loss}), one can think of the Hinge regularization instead.
Neural network activations determine the output of a node and are a crucial component of deep learning. When the activation function is non-linear, then a two-layer neural network can be proven to be a universal function approximator \citep{cybenko1989approximation}. The identity activation function does not satisfy this property. When multiple layers use the identity activation function, the entire network is equivalent to a single-layer model. The Parameteric Rectified Linear Unit (PReLU) \citep{he2015delving} has been proposed as an extension of Rectified Linear Unit (ReLU) \citep{nair2010rectified} in order to improve image classification accuracy. The PReLU function is defined as
\[
f_a (x) =
\begin{cases}
x, ~\mathrm{if}~ x \geq 0\\
ax, ~\mathrm{otherwise}
\end{cases}
\]
where $a \in [0;1]$ is learned in the training via backpropagation. When $a=0$, PReLU simplifies to ReLU and if $a$ is set to a fixed value, PReLU simplifies to Leaky ReLU (LReLU) \citep{maas2013rectifier}.
\paragraph{Search strategy.}
The search strategy details how to explore the search space
(which is often exponentially large or even unbounded). It encompasses the classical exploration-exploitation trade-off since, on the one hand, it is desirable to find
well-performing architectures quickly, while on the other hand, premature convergence to a region of suboptimal architectures should be avoided. In \cite{yang2019evaluation}, based on a benchmark of 8 NAS methods on 5 datasets, they compare methods with different search spaces, using a method’s relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, they find that many NAS techniques struggle to significantly beat the average architecture baseline. Furthermore, it is desirable to have a search strategy because unfortunately, in many deep learning tasks, training is computationally expensive. However, for our application, training SMITE on the real data takes 10 minutes for 100 epochs (on a Dual Intel Xeon 256 Gb RAM and 1 Titan V12GB). Therefore, we believe that grid search is suitable for SMITE architecture search. Table~\ref{tab:NAS} summarizes the hyper-parameters responsible for defining the architecture of SMITE. Hyper-parameter 3 must be defined for each layer. Although hyper-parameters 5-7 do not modify the architecture, they are important in terms of training. We prefer to enumerate them in this Table in order to include them in the architecture search. Hyper-parameter 8 is important for reproducible results. Our experimental results are presented in Section~\ref{sec:experiments}.
\begin{table}[htb]
\caption{Hyper-parameters that define the architecture of SMITE.}
\centering
\begin{tabular}{|c||l|l|l|}
\hline
\# & Hyper-parameter & Type & Scope\\
\hline
1 & \# of hidden layers & Integer & $\{0,1\}$\\
2 & \# of hidden units & Integer & $\{32,64,128,256,512,1024\}$\\
3 & Activation & Nominal & $\{$ReLU, LReLU$(0.05)\}$\\
4 & Regularization function & Nominal & $\{$TO, IE$\}$ \\
5 & Batch size $m$ & Integer & $\{128,256\}$ \\
6 & Learning rate $\xi$ & Continuous & $\{0.01,0.02,0.03,0.04,0.05\}$ \\
7 & Trade-off constant $\alpha$ & Continuous & $\{0,0.1,0.2,...,1\}$ \\
8 & Seed & Integer & $\{1,2,3,...,10\}$ \\
\hline
\end{tabular}
\label{tab:NAS}
\end{table}
\paragraph{Performance estimation strategy.}
The objective of NAS is typically to find architectures that achieve high predictive performance on unseen data. Performance
estimation refers to the process of estimating this performance: the simplest option is to perform a standard training and validation of the architecture on data.
\textcolor{blue}{HERE WE NEED TO SHOW RESULTS FROM THE ARCHITECTURE SEARCH AND SHOW THE FINAL ARCHITECTURE THAT WE WILL USE IN THE EXPERIMENTS.}
\textcolor{red}{The backbone of the SMITE architecture is a $6$-layers feed forward MLP with the following input neurons sizes: $\{200,200,300,100,50,10\}$. The first two layers are linearly connected while the rest are connected through a leaky ReLu activation function. Both the depth and the width of this network were explored using a random grid search on data generated with the framework described in Section~\ref{sec:NAS}. Two hyper-parameters are fine-tuned for each set of experiments: the learning rate $\xi$ and the trade-off constant $\alpha$. To find an adequate candidate for $\alpha$, we follow the following protocol: the dataset $D$ is split into two subsets ($D_T$,$D_H$) containing respectively $70\%$ and $30\%$ of $D$. $D_H$ remains unseen during any of our fine-tuning or model selection process. We then repeat $k$-times a random separation of $D_T$ into two datasets denoted ($D^{i}_{\mathrm{train}}$,$D^{i}_{\mathrm{valid}}$) with sizes $60\%$ and $40\%$ of $D_T$. At a fixed learning rate of $\xi=0.03$, we fit the network on $D^{i}_{\mathrm{train}}$ for different values of $\alpha$, from $0$ to $1$ with $0.1$ increments. We measure the average Qini coefficient $\hat q$ obtained on $\{D^{i}_{valid}\}_{i \in [1,k]}$ and we estimated a $95\%$ confidence interval (CI) for $\hat q$. The selected $\alpha$ corresponds to the highest average as long as the $95\%$ CI lower bound remains higher than $0$. The resultant of this protocol is illustrated in Figure \ref{fig:alphaie}. Once $\alpha$ is fixed, we follow then the same process for the learning rate $\xi$.}
\input{figures/alpha.tex}
\subsection{Data generation}
We evaluate the performance of SMITE using synthetic data generated using two types of mechanisms.
\paragraph{Bootstrap Data.} The underlying treatment effects are inaccessible in real-world data. To create realistic scenarios, we based one of our simulation study data generation on insurance policy data. Moreover, since we compare SMITE to random forests, we want to evaluate its performance when the data generation process comes from a tree-based model. We follow the procedure in the simulation studies of \cite{belba2019qbased}. These simulations allow us to evaluate if SMITE can recover the right representation for data generated from a tree-based model. This also makes comparisons to random forest methods fair.
Several tree-based methods have been suggested in the uplift literature. Here, we use the uplift random forest \citep{guelman2012random} as the data generation process, which is readily available in the {\bf R} library {\bf uplift}. We fit the random-forest on a random sub-sample $\mathcal{D}$ of the real data. We use $100$ trees and the Kullback-Leibler split criterion. Each tree has maximum depth of $5$. Based on the fitted model, we can extract the probabilities $\mu_0(\mathbf x)$ and $\mu_1(\mathbf x)$ for any given value $\mathbf{x}$. We use these probabilities to draw synthetic outcomes for treated and control observations. We start by creating a bootstrap sample $\mathcal{S}$ of size $n_\mathcal{S}$ from $\mathcal{D}$. For each observation $\mathbf{x}_i \in \mathcal{S}$, we generate a random vector $\tilde{y}_i =(\tilde{y}_{i0}, \tilde{y}_{i1})$, where $\tilde{y}_{i0}$ is the binary outcome of a Bernoulli trial with success probability $\mu_0(\mathbf{x}_i)$, and $\tilde{y}_{i1}$ is the binary outcome of a Bernoulli trial with success probability $\mu_1(\mathbf{x}_i)$, $i=1,\ldots, n_\mathcal{S}$. The augmented synthetic dataset $\{ \mathbf x_i , t_i, \tilde{y}_i \}_{i=1}^{n_{\mathcal{S}}}$ is the data of interest in the simulation.
We repeat this process several times. Each bootstrapped data size is $10,000$ observations with $40$ covariates.
\paragraph{Synthetic Data.} We evaluate the performance of SMITE models using synthetic data generated in a similar way to the simulation studies of \cite{tian2014simple}. For the synthetic data generation, there are four elements:
\begin{enumerate}
\item The number $n$ of observations in the sample, and the number $p$ of predictors observed for each observation.
\item The distribution of the predictors. We draw predictors from a multivariate Gaussian distribution.
\item The mean effect function $\mu(\mathbf X)$ and the treatment effect function $\tau(\mathbf X)$. We vary the size of main effects relative to treatment effects.
\item The noise level $\sigma^2$. This corresponds to the conditional variance of the outcome given $\mathbf X$ and $T$.
\end{enumerate}
Given the elements above, our data generation model is, for $i=1,...,n$ :
\begin{align*}
&\mathbf X_i \sim \mathcal{N}_p (0, \Sigma) \\
&T_i \sim \mathrm{Bernoulli}(\theta)\\
&Y_i^* \mid \mathbf X_i, T_i \sim \mathcal{N}(\mu(\mathbf X_i) + T_i\tau(\mathbf X_i), \sigma^2).
\end{align*}
Define $Y_i$ to be the binary outcome as
\begin{align*}
Y_i \mid \mathbf X_i, T_i = \mathbbm{1}(Y_i^* > 0 \mid \mathbf X_i, T_i).
\end{align*}
Hence, $Y_i$ follows a Bernoulli distribution with parameter $\mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i)$ and it is easy to recover the ``true'' uplift $u_i^*$ as follows
\begin{align*}
u_i^* &= \mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i = 1) - \mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i = 0)\\
u_i^* &= \Phi \Bigl( \frac{\mu(\mathbf X_i) + \tau(\mathbf X_i)}{\sigma} \Bigr) - \Phi \Big( \frac{\mu(\mathbf X_i)}{\sigma} \Big)
\end{align*}
where $\Phi(.)$ is the standard Gaussian cumulative distribution function. Note that $\mu(.)$ is a nuisance function and $\tau(.)$ is the function that controls the treatment effect. In fact, the sign of $\tau(.)$ and the sign of the uplift for a given observation coincide.
We set $\theta$ to $0.5$, which is the case in our real data. The variance-covariance matrix elements $\Sigma_{ij}$ are defined as follows:
\[\Sigma_{ij} = \begin{cases}
\rho & i\neq j \\
1 & i=j
\end{cases}.
\]
Finally, the mean and treatment effects functions are defined as follows:
$$\mu(\mathbf X) = \sum_{j=1}^J \beta_j X_j + \sum_{j=1}^{p-1} \omega_j X_j X_{j+1},$$
$$\tau(\mathbf X) = \sum_{j=1}^K \gamma_j X_j$$
where $J\leq p$ and $K \leq p$. We set $\beta_j = \beta (-1)^{j}$, $\omega_j = \omega (-1)^j$ and we draw $\gamma_j$ independently from $\mathcal{N}(0,\gamma)$. The constants $\beta$, $\omega$ and $\gamma$ are hyper-parameters.
In this case, we generate datasets of sizes $n=20,000$ observations with $100$ covariates.
\subsection{Benchmarks and implementation}
Ensemble methods are amongst the most powerful for the uplift case \citep{soltys2015ensemble}. A method that has proven to be very effective in estimating ITE is one based on generalized random forests \citep{athey2019generalized}. This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the treatment effects. This has the effect of reducing overfitting. Another candidate that we consider in our experiments is also based on random forests that were designed for uplift \citep{guelman2012random, jaskowski2012uplift}. For this method, we consider two different split criteria, one based on Kullback-Leibler divergence and one based on Euclidean distance.
For random forests (RF), we use the implementation from the \textbf{R} libraries \textbf{grf} and \textbf{uplift} which were developed by the authors of the methods respectively. Implementations of these causal forests have several useful parameters that make it easy to fine tune them. In particular, \textbf{grf} provides cross-validation procedures for hyper-parameters tuning which we use to get the best possible candidates to compare our results to. We also use the \textbf{R} library \textbf{tools4uplift} for computing the model's goodness-of-fit measures. SMITE is implemented in \textbf{Python}, so we reimplemented the Qini (\ref{eq:q:hat}) and Kendall (\ref{eq:corr_coeff}) coefficients functions. In total, we consider $4$ different random forests to compare to SMITE: causal forest, honest causal forest (H), RF with Kullback-Leibler (KL) and RF with Euclidean distance (ED) split criteria.
\subsection{Experimental results}
For both simulated datasets, we do not consider any out-of-sample subset. In Monte Carlo settings, we are able to generate as much data as we want. Therefore, we first fine-tuned the methods on a single synthetic dataset of same size as the one we used in the repeated simulations presented. We generate $n=10,000$ observations with $p=40$ covariates for the bootstrapped data. For the fully synthetic datasets, we report results from experiments with $n=10,000$ observations and $p=100$ covariates. We split into training and validation and compare the results averaged over $30$ runs for the models that performed best on the validation set. Results are described in Table \ref{tab:boot_perf} for the bootstrap datasets and Table \ref{tab:synthe_perf} for the synthetic datasets. For the fully synthetic data, we see that Causal Forest (H) and SMITE (TO) had similar performance on the validation set while SMITE (IE) significantly outperformed all the other methods. For the bootstrap data for which the outcomes were generated based on random forest model, it is interesting to see that SMITE (TO) and Causal Forest (H) give similar results. Moreover, once again, the SMITE (IE) model outperforms all the other models.
\input{tables/table_bootstrap.tex}
\textcolor{blue}{HERE WE NEED TO SHOW ESTIMATED AND PREDICTED UPLIFTS DISTRIBUTIONS PER METHOD. I THINK THIS WILL SHOW THAT RF METHODS COVER A RANGE THAT IS A LOT LARGER THAN SMITE. WE CAN ALSO COMPUTE THE EMPIRICAL VARIANCES AND COMPUTE A CONFIDENCE INTERVAL FOR THE VARIANCE OF THE PREDICTED UPLIFTS TO SEE IF THERE ARE SIGNIFICANTLY LARGER THANT SMITE'S VARIANCE.}
\input{tables/table_synthetic.tex}
\section{Application}
\label{sec:application}
A data set with $n=50,000$ customers and $p=40$ covariates from an insurer is at our disposal to evaluate the performance of our methods. This company was interested in designing strategies to maximize its conversion rate. An experimental acquisition campaign was implemented for 6 months, for which half of the potential clients were randomly allocated into a treatment group and the other half into a control group. Potential clients under the treatment group were contacted using a different (and more costly) strategy. The goal of the analysis was to be able to select who among the clients portfolio should be contacted in that way going forward in order to maximize the benefits. The observed difference in sales rates between the treated group and the control shows some evidence of a positive impact of the treatment ($\mathrm{ATE}=0.55\%$).
\textcolor{blue}{DESCRIBE THE DATA AND GIVE MORE DETAILS ON THE APPLICATION. ADD FIGURES FOR BEST FINAL MODELS (Qini Curves).}
For all models we follow the same learning process and performance metrics computations. For the real-world dataset, a sample of $30\%$ is dedicated to out-of-sample performance measures. We do not score these customers until all models are properly fine-tuned. The remaining $70\%$ is used for fine-tuning. We repeat the experiment $30$ times. We randomly split the subset into $60\%$ for training and $40\%$ for validation. We select the model which performs best on the validation set. Using that model, we score the out-of-sample customers and compute the performance metrics. Results are presented in Table \ref{tab:realdata_perf}.
\input{tables/table_real_data.tex}
We observe that the random forests tend to overfit. Indeed, based on the out-of-sample results from the hold out set, the models seem to be close to random targeting. This might be due to the small (in magnitude) overall impact of the initiative. On the other hand, we observe a good adequacy of SMITE for both TO and IE. Overall, the results of SMITE (IE) appear to outperform all other methods. It is also interesting to note that the distribution of the ITE obtained by the model SMITE (IE) are in line with what was logically expected (see Figure \ref{fig:pred_uplift_evol}: giving what appears to be a better service only slightly increases the propensity of buying.
\input{figures/uplift_evolutions.tex}
\section{Conclusion}
\label{sec:conclusion}
In this paper we present a meaningful and intuitive neural networks architecture for the problem of uplift modeling.
We apply our methods to both real-world and synthetic data and we compare with what is, to the best of our knowledge, state-of-the-art methods for uplift. Our results show that the SMITE models significantly outperform random forests designed for ITE and uplift in all of our scenarios. For future work, we want to study the theoretical properties of the SMITE optimization problem as well as including more causal inference related problems such as multiple treatments, time series, and the use of observational data.
\bibliographystyle{plainnat}
\section{Introduction}
\label{sec:intro}
\textit{Uplift} is a particular case of causal inference and follows the potential outcomes framework \citep{rubin1974estimating, holland1986statistics}. Uplift models are developed for randomized controlled trials (RCTs) settings where both the treatment and outcome are binary random variables. Uplift refers to modeling customers behavioral change that results directly from a specific treatment, such as a marketing intervention (e.g. a courtesy call). The customers can be segmented along two dimensions in function of the potential outcomes and the treatment, giving rise to the following groups \citep{kane2014mining} : a \textit{persuadable} client is an individual who purchases the product only if he receives a call; a \textit{sure thing} client is an individual who purchases the product whether he receives the call or not; a \textit{lost cause} client is an individual who will not purchase the product regardless of whether he receives the call or not; a \textit{do-no-disturb} client is an individual who will not purchase the product only if he receives a call. In general, from a business point of view, interesting customers are the \textit{persuadable} and the \textit{do-not-disturb} clients. The persuadable clients provide incremental sales whereas the do-not-disturb individuals should be avoided because the marketing campaign would have a negative effect on them. Mainly known in the marketing research literature, uplift models \citep{radcliffe1999differential, hansotia2002direct, lo2002true, radcliffe2007using} provide a solution to the problem of isolating the marketing effect. Instead of modeling the positive response probability, uplift attempts to model the difference between conditional probabilities in the treatment and control groups.
Most existing uplift models are adaptations of classification and regression trees \citep{breiman1984classification} where the majority are modified random forests \citep{Breiman-RandomForest-2001, radcliffe2011real, jaskowski2012uplift, guelman2012random}. Within each leaf, uplift is estimated with the difference between the two conditional means. The belief is that since the uplift is defined as a difference, if we can get good estimates of each mean, we have a good estimate of the difference. However, the existing tree-based uplift optimization problems do not take into account the estimation of each mean. Instead, the focus is on maximizing the heterogeneity in treatment effects. Without careful regularization (e.g. honest estimation \citep{athey2019generalized}), splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. In addition, successive splits tend to group together similar extreme values, introducing more variance in the prediction of uplift \citep{zhao2017practically}. Alternatively some models use the transformed outcome \citep{athey2015machine}, an unbiased estimator of the uplift. However, this estimate suffers from higher variance than the conditional mean estimator \citep{powers2018some}. In addition, for both estimators, if random noise is larger than the treatment effect, the model will more likely predict random noise instead of uplift. This is particularly problematic for uplift trees, as the split criteria is function of the two treatment effects. As a result, based on several experiments on real data, and although the literature suggests that tree-based methods are state-of-the-art for uplift \citep{soltys2015ensemble}, the published models over-fit the training data and predicting uplift still lacks satisfactory solutions.
We define a unified framework with the goal of improving uplift models predictions on unseen data. Taking advantage of the flexibility and prediction power of neural networks, we design an optimization problem for which regularization of the weights takes into account both conditional means estimation and adequacy of uplift prediction. To the best of our knowledge, our method proposes a new way to tackle overfitting for uplift models. We present two regularization criteria: one using the transformed outcome, that has been shown to be a consistent estimator of the uplift with respect to the mean squared error \citep{gutierrez2017causal}, and a novel regularization criteria that comes from the Bayes conditional probability theorem.
Empirically, we show that our proposed optimization has the impact to reduce the variance of the predicted uplifts. Furthermore, the flexibility of our proposed method opens a door to connections with the estimation of uplift in observational data settings. Our proposed method is validated on both synthetic and a realworld data-set made available to us by a Canadian insurance company. The experimental results demonstrate its advantages over state-of-the-art tree-based uplift modeling.
\section{Uplift Modeling}
\label{sec:literature}
To formalize the problem, let $Y$ $\in$ $\lbrace0,1\rbrace$ be a binary response variable, $\mathbf X=(X_{1},\ldots,X_{p})$ a vector of covariates, and $T\in\lbrace0,1\rbrace$ the treatment indicator variable. The binary variable $T$ indicates if a customer is exposed to treatment ($T=1$) or control ($T=0$). The observed outcome $Y$ can be described as
$$Y=TY_1 + (1-T)Y_0,$$
where $Y_0$ and $Y_1$ are the potential outcomes under control and treatment respectively. Suppose $n$ independent customer observations $\{y_i, \textbf{x}_i, t_i\}_{i=1}^n$ are at our disposal. Although each customer $i$ is associated with two potential outcomes, only one of them can be realized as the observed outcome $y_i$. For RCTs settings, $T$ is independent ($\independent$) of all pre-treatment characteristics, i.e.,
\begin{equation}
\big( Y_0, Y_1, \mathbf X \big) \independent T.
\label{eq:randomization}
\end{equation}
This allows the unbiased estimation of the average treatment effect. Moreover, the \textit{propensity score} $e(\mathbf x)= \mathrm{Pr}(T=1 \mid \mathbf X=\mathbf x)$ is a constant. In our data, $e(\mathbf x) = 1/2$. An uplift model estimates the conditional average treatment effect in various sub-populations as a function of the possible values of the covariates $(\mathbf X=\mathbf x)$,
\begin{align}
u(\mathbf x) &= \mathbb{E} ( Y_1 \mid \mathbf X=\mathbf x) - \mathbb{E} ( Y_0 \mid \mathbf X=\mathbf x) \nonumber\\
&= \mathbb{E} ( Y \mid T=1, \mathbf X=\mathbf x) - \mathbb{E} ( Y \mid T=0, \mathbf X=\mathbf x) \nonumber \\
&= \mathrm{Pr} ( Y = 1 \mid T=1, \mathbf X=\mathbf x) - \mathrm{Pr} ( Y = 1 \mid T=0, \mathbf X=\mathbf x) \nonumber \\
&=m_1(\mathbf x) - m_0(\mathbf x).
\label{eq:uplift}
\end{align}
Uplift is defined as the difference between the two conditional means $m_1(\mathbf x)$ and $m_0(\mathbf x)$. Therefore, the intuitive approach to model uplift is to build two classification models \citep{hansotia2002direct}. This consists of fitting two separated conditional probability models: one for the treated individuals, and another for the untreated individuals. The uplift is estimated as the difference between these two conditional probability models. The asset of this technique is its simplicity. However, it does not perform well in practice \citep{radcliffe2011real}. Both models focus on predicting only a one-class probability and for each model, information about the other treatment is never provided to the learning algorithm. In addition, differences between the covariates distributions in the two treatment groups can lead to bias in treatment effect estimation. Using a unique classification model corrects this drawback. Some traditional uplift estimation methods use the treatment assignment as a feature, adding explicit interaction terms between each covariate and the treatment indicator, and then train classification models to estimate the uplift \citep{lo2002true}. The parameters of the interaction terms measure the additional effect of each covariate because of the treatment.
The transformed outcome variable $Z$, which was proposed by \citep{athey2015machine} is defined as
\begin{equation}
Z = \frac{T Y}{e(\mathbf x)} - \frac{(1 - T) Y}{1 - e(\mathbf x)}.
\label{eq:transformed_outcome}
\end{equation}
This transformation has the key property that, under the complete randomization or unconfoundness assumption \citep{rosenbaum1983central}, its expectation conditional on $(\mathbf X=\mathbf x)$ is equal to $u(\mathbf x)$. Therefore, the transformed outcome is an unbiased estimate of the uplift. With this transformed outcome, one can simply use any supervised learning algorithm to estimate a regression function for the expectation of $Z$ given $\mathbf X$. In our case, with $e(\mathbf x) = 1/2$, $Z$ simplifies to
\[ Z = \begin{cases}
2, & \mathrm{if}~T=1, Y=1 \\
-2, & \mathrm{if}~T=0, Y=1 \\
0 & \mathrm{otherwise.}
\end{cases}
\]
In the uplift context, one can have an interesting interpretation of $Z=2$ and $Z=-2$. Take the first case. This happens when a client is called ($T = 1$) and then buys the insurance policy ($Y = 1$). Thus, for its counterfactual (i.e., $T = 0$), there are two possibilities. Either the client would have bought the policy ($Y = 1$) and in this case would belong to the \textit{sure things} group, or else the client would not have bought the policy ($Y = 0$) in which case would in fact be a \textit{persuadable}. Thus, in the worst case scenario, individuals with $Z = 2$ are \textit{sure things}. The same logic applies for $Z = -2$ where a client could either be a \textit{sure thing} or a \textit{do-not-disturb}. For \textit{lost cause} clients, by definition $Z = 0$. This interpretation does not cover all scenarios for $Z=0$ but gives insights about ordering customers with $Y=1$.
Several machine learning methods try to model the uplift directly. Some $k$-nearest neighbours \citep{cover1967nearest} based methods are also adopted for uplift estimation \citep{crump2008nonparametric, Alemi.etal-PersonalizedMedicine-2009, su2012facilitating}. The main idea is to estimate the uplift for an observation based on its neighbourhood containing at least one treated and one control observations. However, these methods quickly become computationally expensive for large datasets. Furthermore, adaptive tree-based methods such as random forests outperform the nearest neighbours approach, especially in the presence of irrelevant covariates \citep{wager2018estimation}. Therefore, most active research in uplift modeling is in the direction of classification and regression trees \citep{breiman1984classification}. State-of-the-art proposed methods view the forests as an adaptive neighborhood metric, and estimate the treatment effect at the leaf node \citep{su2009subgroup, chipman2010bart, guelman2012random}. In \cite{radcliffe1999differential, radcliffe2011real} or \cite{ hansotia2002direct, hansotia2002incremental}, modified split criteria that suited the uplift purpose were studied. The criteria used for choosing each split during the growth of the uplift trees is based on maximization of the difference in uplifts between the two child nodes. However, in practice, these approaches are prone to overfitting. The splitting criteria introduced are based solely on maximizing differences in uplifts. Therefore, splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. In addition, successive splits tend to group together similar extreme values, introducing more variance in the prediction of uplift \citep{zhao2017practically}.
Recently, representation learning approaches have been proposed for estimating individual treatment effect based on observational data. An approach in the literature emphasizes learning a covariate representation that has a balanced distribution across treatment and outcome \citep{johansson2016learning}. Some similar methods try to minimize the distribution difference between treated and control groups in the embedding space \citep{shalit2017estimating, yao2018representation}. Another approach is based on propensity score matching \citep{rosenbaum1983central} with neural networks \citep{alaa2017deep, schwab2018perfect} and \citep{shi2019adapting} use regularized neural networks to estimate the average treatment effect.
\section{Methodology}
\label{sec:methodology}
Classification models with binary outcomes are relatively easy to develop\footnote{Depending on the complexity of the data, any supervised learning algorithm used properly tends to give satisfactory results.}. However, uplift models are not classification models. The goal of uplift modeling is to separate all observations into heterogeneous groups that differ as much as possible with regard to the reactions to the treatment and, thus, identify subgroups with different uplifts. Then, the uplifts in the resulting groups serve as a prediction for the future behavior of the observations. Therefore, a model that learns only the binary classification task fails to properly capture the uplift because it is not designed for this task. But, since the uplift is defined as a difference between two conditional probabilities, it is important to put emphasis on the estimation of these probabilities. On the other hand, methods that aim to directly predict the uplift (such as tree-based methods which maximize heterogeneity) fail to generalize because they focus only on heterogeneity and do not take into account the estimation of the conditional probabilities.
In recent years, deep neural networks have become very popular thanks to their ability to model very complex structures. They have achieved remarkable results, for example, in the fields of computer vision or natural language processing. Taking advantage of the flexibility and predictive power of neural networks, we propose a \textit{Siamese} \citep{bromley1994signature} architecture called SMITE \footnote{Siamese Model for Individual Treatment Effect} which works in tandem on two almost identical augmented input vectors $(\mathbf x,\mathbf{1})$ and $(\mathbf x,\mathbf{0})$, and outputs the prediction of the conditional mean $m_t(\mathbf x) = tm_1(\mathbf x) + (1-t)m_0(\mathbf x)$ as well as the one of the uplift $u(\mathbf x)$. Using a Sieamese architecture simplifies the fitting process. On the same forward pass, the model outputs the conditional means as well as the uplift. The idea is that the two input vectors are fed to two identical neural network sub-components (i.e., with shared weights $\boldsymbol{\theta}$). The inputs contain the covariates vector $\mathbf x$ and, for the top network sub-component, the treatment variable fixed to $1$. The treatment variable is fixed to $0$ for the bottom network sub-component as shown in Figure \ref{fig:nite}.
\input{figures/framework.tex}
The sub-components output the predicted conditional means for treated ($\mu_1$) and for control ($\mu_0$). The difference gives direct prediction for $u(\mathbf x)$. At the same time, the predicted conditional mean $\mu_t$ is based on the actual treatment $t \in \{0,1\}$, i.e., $\mu_t = t\mu_1 + (1-t)\mu_0$.
The conditional mean $\mu_t$ is compared to the observed outcome $y$ based on binary cross-entropy loss $\mathcal{L}$ or BCE. Then, we propose two regularization criteria. The first one compares the uplift prediction to the observed transformed outcome $z$ based on the mean squared error loss $\mathcal{J}$. The second one uses the observed treatment variable $t$ and another BCE loss $\mathcal{I}$ (see Section~\ref{sec:smite_ie}). We give the details in the following sections. Essentially, back-propagating both losses gradients through the network consists in fitting a {\it regularized} conditional means regression.
\subsection{Regularized conditional means regression}
\label{sec:smite_to}
In deep learning regularization is sometimes hidden in heuristic methods during training. For instance, adding noise to the input data yields to generalization error improvement \citep{bishop1995training,rifai2011adding}. Data augmentation, and early stopping are some other heuristic regularizations widely applied in practice. A more theoretically sound regularization method is dropout \citep{srivastava2014dropout}, a widely-used method for addressing the problem of over-fitting. The idea is to drop units randomly from the neural network during training. It is known that dropout improves the test accuracy compared to conventional regularizers such as $L_1$ \citep{Tibshirani_lasso_1996} and $L_2$ \citep{srivastava2014dropout}. \citep{wager2013dropout} proved that dropout is equivalent to an $L_2$-type regularizer applied after scaling the inputs. These common regularizations encourage the weights to be estimated near zero. In our case, we want regularization on the predicted uplift (by the difference of the conditional means) and not directly applied to the weights. Therefore, we need to define a new regularization function.
Suppose $n$ independent customer observations $\{y_i, \textbf{x}_i, t_i, z_i\}_{i=1}^n$ are at our disposal. Let us denote the model's parameters by $\boldsymbol{\theta}$. Regularization can be seen as a constrained optimization problem, then using the transformed outcome $Z$, SMITE solves
\begin{align}
{\boldsymbol{\hat\theta}} & = \argmin_{\boldsymbol{\theta}} ~ \mathcal{L}\Big({\mu_t}(\boldsymbol{\theta}), \mathbf y\Big) \nonumber \\
& \mathrm{subject~to,} \nonumber \\
& \mathcal{J}\Big({\mu_0}(\boldsymbol{\theta}),{\mu_1}(\boldsymbol{\theta}), \mathbf z\Big) < \epsilon,
\label{eq:objective}
\end{align}
where $\epsilon > 0$. The quantities $\mu_{0}(\cdot)$ and $\mu_{1}(\cdot)$ are functions of the inputs $\mathbf x$ and could be denoted $\mu_{0}(\mathbf x)$ and $\mu_{1}(\mathbf x)$. To simplify the notation, we prefer to omit this specification throughout the rest of the paper. The rationale behind this constrained formulation is that we want a model capable of direct prediction of the uplift $u(\mathbf x)$, i.e., $\mu_1(\boldsymbol{\hat\theta}) - \mu_0(\boldsymbol{\hat\theta})$, and coherent estimation of $m_t(\mathbf x)$, i.e., $\mu_t(\boldsymbol{\hat\theta})$.
Formally, we define $\mathcal{L}(\boldsymbol{\theta}, \mathbf y)$ with respect to the network's parameters $\boldsymbol{\theta}$ as
\begin{equation*}
\mathcal{L}(\boldsymbol{\theta}, \mathbf y) = - \sum_{i=1}^{n} \Big( y_i\mathrm{log} \{\mu_{t_i}(\boldsymbol{\theta})\} + (1-y_i) \mathrm{log} \{1-\mu_{t_i}(\boldsymbol{\theta})\} \Big)
\end{equation*}
where $\mu_{t_i}(\boldsymbol{\theta}) = t_i \mu_{1i}(\boldsymbol{\theta}) + (1-t_i) \mu_{0i}(\boldsymbol{\theta})$ is the model's output of observation $i$ with treatment $t_i$, with $\mu_{1i}(\boldsymbol{\theta})$ and $\mu_{0i}(\boldsymbol{\theta})$ representing the outputs for observation $i$ from the top and bottom sub-components of the network (see Figure~\ref{fig:nite}).
The function $\mathcal{J}(.)$ in \eqref{eq:objective} is what we refer to as the \textit{direct} uplift regularization function. It is defined as a distance between the predicted uplift and the transformed outcome, such as
\begin{equation}
\mathcal{J}(\boldsymbol{\theta}, \mathbf z) = \frac{1}{n} \sum_{i=1}^{n} \Big( z_i - \{{\mu_{1i}}(\boldsymbol{\theta }) - {\mu_{0i}}(\boldsymbol{\theta})\} \Big)^2.
\label{eq:direct_uplift_loss}
\end{equation}
It is helpful to rewrite the full constrained problem (\ref{eq:objective}) in his Lagrangian form
\begin{align*}
{\boldsymbol{\hat\theta}}_{\lambda} = \argmin_{\boldsymbol{\theta}} \mathcal{L}(\boldsymbol{\theta}, \mathbf y) + \lambda \mathcal{J}(\boldsymbol{\theta}, \mathbf z)
\end{align*}
where $\lambda > 0$ is the regularization constant. However, in the SMITE framework, we use the transformation $\alpha = \lambda / (1+\lambda)$. This transformation allows to interpret $\alpha \in [0,1)$ as a hyper-parameter which controls the trade-off between the two losses in the optimization problem. Hence,
\begin{align}
{\boldsymbol{\hat\theta}}_{\alpha} = \argmin_{\boldsymbol{\theta}} (1 - \alpha) \mathcal{L}(\boldsymbol{\theta}, \mathbf y) + \alpha \mathcal{J}(\boldsymbol{\theta}, \mathbf z).
\label{eq:SMITE_objective}
\end{align}
Equation (\ref{eq:SMITE_objective}) encompasses what motivates the SMITE framework to be able to penalize the conditional means regression weights in order to guide the uplift predictions. This is equivalent to jointly optimize both loss functions for a fixed value of $\alpha$. In our experiments, the {\it trade-off} constant $\alpha$ is fine-tuned using cross-validation (see Figure \ref{fig:alphaie}).
\subsection{The optimization algorithm}
\label{sec:smite_to_sgd}
Gradient-based optimization is one of the pillars of machine learning. Given an objective function $f: \rm I\!R^n \rightarrow \rm I\!R$, classical gradient descent has the goal of finding
(local) minima $\hat{\boldsymbol{\theta}} = \argmin_{\boldsymbol{\theta}} f(\boldsymbol{\theta})$ via updates of the form $ \Delta (\boldsymbol{\theta}) = -\xi \nabla f$, where $\xi$
is a positive step size (or the learning rate). Two main classes of optimization methods can be applied to solve such problems : (i) {\it stochastic} and (ii) {\it batch}.
The prototypical stochastic optimization method is the stochastic gradient
method \citep{robbins1951stochastic}, which, in the context of minimizing $f(\boldsymbol{\theta})$ and with $\boldsymbol{\theta}^{(0)} \in \rm I\!R^p$ given, is defined by
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \xi \nabla f_{i_k} (\boldsymbol{\theta}^{(k)}).
\end{equation}
Here, for all $k \in \mathbb{N}$, the index $i_k$ is chosen randomly from $\{1,...,n\}$. There is no particular reason to employ information from only one sample point per iteration. Instead, one can employ a {\it mini-batch} approach in which a small subset of samples $\mathcal{S}_k \subseteq \{1, ..., n\}$, is chosen randomly in each iteration, leading to
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \frac{\xi}{|\mathcal{S}_k|} \sum_{i \in \mathcal{S}_k} \nabla f_{i} (\boldsymbol{\theta}^{(k)}).
\label{eq:mini-batch-sgd}
\end{equation}
Here, $|\mathcal{S}_k|$ is the batch size which we will later denote by $m$. Such a mini-batch stochastic gradient method has been widely used in practice. On the other hand, the {\it batch} approach, also referred to as the gradient, batch gradient, or full gradient method, uses all $n$ observations in each iteration. There are some fundamental practical and theoretical reasons why stochastic methods have inherent advantages for large-scale machine learning \citep{bottou2018optimization}. In practice, we train the SMITE model by solving (\ref{eq:SMITE_objective}) using the mini-batch stochastic gradient descent (SGD) updates (\ref{eq:mini-batch-sgd}).
Linear regression can be regarded as the simplest neural network, with no hidden layer, a linear activation function and the MSE loss function. For logistic regression, the sigmoid activation replaces the linear one and the BCE loss is used instead of the MSE. In order to get insight about the SMITE framework, it is important to study SGD updates in linear/logistic regression contexts. To do so, let us define a simple SMITE framework. For $\alpha \in [0,1)$, let
\begin{equation}
f_{\alpha}(\boldsymbol{\theta}) = \alpha \mathcal{J}(\boldsymbol{\theta}, \mathbf z)+ (1 - \alpha) \mathcal{L}(\boldsymbol{\theta}, \mathbf y)
\label{eq:obj_simple}
\end{equation}
be the objective function of the SMITE model with no hidden layer. Without loss of generality, suppose that we only have one predictor at our disposal and let
\[
\mathbf{y} =
\begin{bmatrix}
y_1 \\ y_2 \\ \vdots \\ y_m
\end{bmatrix},~
\mathbf{z} =
\begin{bmatrix}
z_1 \\ z_2 \\ \vdots \\ z_m
\end{bmatrix},~
\mathbf X =
\begin{bmatrix}
1 & x_1 & t_1 \\ 1 & x_2 & t_2 \\ \vdots & \vdots & \vdots \\ 1 & x_m & t_m
\end{bmatrix},~
\boldsymbol{\theta} =
\begin{bmatrix}
\theta_0 \\ \theta_1 \\ \gamma
\end{bmatrix},
\]
where $m$ is the batch size, $\mathbf{y}$ is the binary outcome, $\mathbf{z}$ is the transformed outcome and $\mathbf X$ is the design matrix. The parameters of $\boldsymbol{\theta}$ represent the intercept ($\theta_0$), the predictor's coefficient ($\theta_1$) and the treatment coefficient ($\gamma$) which we represent with a different symbol because it plays a different role than $\theta_1$. Indeed, if $\gamma = 0$, the uplift is $0$ for all observations.
Define the sigmoid function as $\sigma (a) = 1 / (1 + e^{-a})$ for $a \in \rm I\!R$. For $i=1,...,m$, let
\[
\begin{cases}
s_i = \sigma( \theta_0 + \theta_1 x_i + \gamma t_i), \\
u_i = s_{1i} - s_{0i},
\end{cases}\]
where
\[
\begin{cases}
s_{1i} = \sigma( \theta_0 + \theta_1 x_i + \gamma), ~ \mathrm{and} \\
s_{0i} = \sigma( \theta_0 + \theta_1 x_i).
\end{cases}
\]
The SMITE framework models simultaneously $\mathbbm{E}[Y_i \mid \mathbf x_i]$ with $s_i$ and $\mathbbm{E}[Z_i \mid \mathbf x_i]$ with $u_i$. The objective function (\ref{eq:obj_simple}) can be written as follows
\begin{equation}
f_{\alpha}(\boldsymbol{\theta}) = \frac{\alpha }{m} \sum_{i=1}^m (z_i - u_i)^2
- \frac{(1 - \alpha)}{m} \sum_{i=1}^m \Big( y_i (\theta_0 + \theta_1 x_i + \gamma t_i) - \mathrm{log}\{1 + e^{\theta_0 + \theta_1 x_i + \gamma t_i} \} \Big).
\label{eq:obj_simple_details}
\end{equation}
It can be shown that each SGD iteration leads to the following parameters update
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \xi_1 \mathbf{V}^{(k)^{\top}}(\mathbf{u}^{(k)} - \mathbf{z}) - \xi_2 \mathbf X^{\top}(\mathbf{s}^{(k)} - \mathbf{y}),
\label{eq:sgd_update}
\end{equation}
where $\xi_1 = 2\alpha\xi/m^2$ and $\xi_2 = (1-\alpha)\xi/m^2$, with
\[
\mathbf{u} =
\begin{bmatrix}
u_1 \\ u_2 \\ \vdots \\ u_m
\end{bmatrix},~
\mathbf{s} =
\begin{bmatrix}
s_1 \\ s_2 \\ \vdots \\ s_m
\end{bmatrix},~
\mathbf{V} = \mathrm{diag}(\mathbf{v}_1) \mathbf{P} - \mathrm{diag}(\mathbf{v}_0) \mathbf{N}
\]
where
\[
\mathbf{v}_1 =
\begin{bmatrix}
v_{11} \\ v_{12} \\ \vdots \\ v_{1m}
\end{bmatrix},~
\mathbf{v}_0 =
\begin{bmatrix}
v_{01} \\ v_{02} \\ \vdots \\ v_{0m}
\end{bmatrix},~
\mathbf{P} =
\begin{bmatrix}
1 & x_1 & 1 \\ 1 & x_2 & 1 \\ \vdots & \vdots & \vdots \\ 1 & x_m & 1
\end{bmatrix},~
\mathbf{N} =
\begin{bmatrix}
1 & x_1 & 0 \\ 1 & x_2 & 0 \\ \vdots & \vdots & \vdots \\ 1 & x_m & 0
\end{bmatrix},
\]
with
\[
\begin{cases}
v_{1i} = s_{1i} ( 1 - s_{1i}), \\
v_{0i} = s_{0i} ( 1 - s_{0i}).
\end{cases}
\]
The quantities $v_{1i}$ and $v_{0i}$ can be interpreted as the variance of $Y_i$ for treated and control groups respectively. Therefore, when looking at the gradient that comes from the MSE loss and the elements of the matrix $\mathbf{V}$, we can see that for $\theta_0$ and $\theta_1$, the updates are weighted with the difference in variances between the treated and control groups $(v_{1i} - v_{0i})$, whereas for the treatment coefficient $\gamma$, the updates are weighted with $v_{1i}$ only. On the other hand, the gradient that comes from the BCE loss does not depend on the uplift and behaves exactly like in the classic logistic regression problem. Thus, by looking at Equation (\ref{eq:sgd_update}), we can interpret the part of the gradients linked to the MSE as a regularization to the part linked to the BCE loss.
Optimization algorithms that use only the gradient are called first-order optimization algorithms. Second-order optimization algorithms such as Newton's method use the Hessian matrix. In this work, as in many deep learning applications where the functions used are complicated, we focus on first-order methods.
\subsection{An alternative to the $Z$ transform}
\label{sec:smite_ie}
The flexibility of SMITE allows us to offer alternatives to the regularization (and loss) functions defined in the general framework (see Figure~\ref{fig:nite}).
SMITE allows to have for each forward pass the estimation of both conditional means. By taking the difference, we get an estimate of the uplift. Now, we want to find a way to regularize this estimate using information from randomization, i.e., the treatment variable $T$. Indeed, suppose a treatment has a significant positive (resp. negative) effect on a sub-sample of customers. This would imply that within the sample of customers that had a positive (resp. negative) response, we expect a higher proportion of treated customers.
For example, suppose the covariate $X$ is a binary variable which represents the gender of each customer and assume the distributions in Table~\ref{tab:example_ie}. Based on the $200$ customers, the average treatment effect is $0$. However, the treatment effect has positive impact on men. Indeed, the uplift for men is $0.1$. Therefore, if we only look at men with $Y=1$, we should observe that more than half ($e(\mathbf x) = 1/2$) were treated. In this example, the proportion is $2/3$.
\begin{table}[H]
\centering
\caption{Example}
\begin{tabular}{|l|cc|cc|cc|}
\hline
& \multicolumn{2}{c|}{Total} & \multicolumn{2}{c|}{$Y=1$} & \multicolumn{2}{c|}{$Y=0$} \\
\hline
& M & F & M & F & M & F \\
\hline
Treatment ($T=1$) & 50 & 50 & 10 & 0 & 40 & 50 \\
Control ($T=0$) & 50 & 50 & 5 & 5 & 45 & 45 \\
\hline
\end{tabular}
\label{tab:example_ie}
\end{table}
More concretely, suppose the data $(y_i, t_i)$, for $i=1,\dots,n$ represents i.i.d. samples from the joint distribution of a pair of binary random variables $(Y,T)$, for which conditional on $\mathbf X=\mathbf x$, $\mathrm{Pr}(T = 1 \mid \mathbf x) = 1 - \mathrm{Pr}(T = 0 \mid \mathbf x) = e(\mathbf x)$. Then, by Bayes theorem,
\begin{align*}
\mathrm{Pr}(T=1 \mid Y=1, \mathbf x) = \frac{ \mathrm{Pr}(Y=1 \mid T=1, \mathbf x) e(\mathbf x)}{\mathrm{Pr}(Y=1 \mid T=0, \mathbf x) \{1- e(\mathbf x)\} + \mathrm{Pr}(Y=1 \mid T=1, \mathbf x)e(\mathbf x)} .
\end{align*}
Inspired by the previous development, we denote the proportion of treated customers among those that had positive outcomes by $p(t=1\mid y=1,\mathbf x)$. When the propensity score is equal to the constant $e(\mathbf x) = 1/2$, this proportion can be written as a function of the conditional means,
\begin{align}
p(t=1\mid y=1,\mathbf x) = \frac{m_1(\mathbf x)}{m_0(\mathbf x) + m_1(\mathbf x)}.
\label{eq:p1}
\end{align}
Although Equation \eqref{eq:p1} is simplified based on the assumption assumption that $e(\mathbf x) = 1/2$, the development holds for any constant $e(\mathbf x)$. Moreover, one can use an under or over-sampling procedure to recover the $e(\mathbf x) = 1/2$ case. Similarly, the proportion of treated customers among those that had negative outcomes is
\begin{align}
p(t=1\mid y=0,\mathbf x) = \frac{(1-m_1(\mathbf x))}{(1-m_0(\mathbf x)) +(1- m_1(\mathbf x))}.
\label{eq:p0}
\end{align}
From these definitions, we propose an alternative to the direct uplift regularization. The \textit{indirect} uplift regularization has the following form
\begin{equation*}
\mathcal{I}(\boldsymbol{\theta},
\mathbf t) = - \sum_{i=1}^{n} \Big( t_i\mathrm{log} \{p_{y_i}(\boldsymbol{\theta})\} + (1-t_i) \mathrm{log} \{1-p_{y_i}(\boldsymbol{\theta})\} \Big),
\end{equation*}
where, $p_{y_i}(\boldsymbol{\theta}) = y_i p_{1i}(\boldsymbol{\theta}) + (1-y_i) p_{0i}(\boldsymbol{\theta})$ is the model's output of observation $i$ with response $y_i$, with $p_{1i}(\boldsymbol{\theta})$ and $p_{0i}(\boldsymbol{\theta})$ representing the estimates for observation $i$ of \eqref{eq:p1} and \eqref{eq:p0}, such as
\begin{align*}
p_{1i}(\boldsymbol{\theta}) &= \frac{\mu_{1i}(\boldsymbol{\theta})}{\mu_{0i}(\boldsymbol{\theta}) +\mu_{1i}(\boldsymbol{\theta})}, \\
p_{0i}(\boldsymbol{\theta}) &= \frac{(1-\mu_{1i}(\boldsymbol{\theta}))}{(1-\mu_{0i}(\boldsymbol{\theta})) +(1- \mu_{1i}(\boldsymbol{\theta}))}.
\end{align*}
The indirect uplift regularization function is simply the binary cross-entropy between a function of the predicted conditional means $p_y(\boldsymbol{\theta})$ and the observed treatment variable $t$.
The associated learning curves are smoother than the ones of the direct regularization using $Z$ (see Figure \ref{fig:instabilty}), which improves the training of the proposed uplift models. In what follows, we will refer to the SMITE using $\mathcal{J}$ as SMITE Transformed Outcome (TO) and the SMITE using $\mathcal{I}$ as SMITE Indirect Estimation (IE). In this case, the optimization problem becomes
\begin{align*}
{\boldsymbol{\hat\theta}}_{\alpha} = \argmin_{\boldsymbol{\theta}} (1 - \alpha) \mathcal{L}(\boldsymbol{\theta}, \mathbf y) + \alpha \mathcal{I}(\boldsymbol{\theta}, \mathbf t),
\end{align*}
for $\alpha \in [0,1)$.
\input{figures/losses.tex}
\section{SMITE Architecture}
\label{sec:NAS}
Deep Learning has enabled remarkable progress over the last years on a variety of tasks, such as image recognition, speech recognition, and machine translation. One crucial aspect
for this progress are novel neural architectures. Currently employed architectures have mostly been developed manually by human experts, which is a time-consuming and error-prone process. Because of this, there is growing interest in automated neural architecture search (NAS) methods. NAS methods can be categorized according to three dimensions \citep{elsken2018neural}: search space, search strategy, and performance estimation strategy.
\paragraph{Search space.}
The search space defines which architectures can be represented
in principle. Incorporating prior knowledge about typical properties of architectures well-suited for a task can reduce the size of the search space and simplify the search. SMITE architecture is a Sieamese multi-layer perceptron. Therefore, the search space for the architecture is first defined by the number of hidden layers and the number of hidden units per layer. However, it should not be limited to these two parameters. The choice of loss functions, regularization and activations in deep networks has also a significant effect on the training dynamics and task performance. These functions are therefore part of the architecture and should be included in the search space. For example, instead of using the MSE (\ref{eq:direct_uplift_loss}), one can think of the $L_1$ regularization instead. Other functions have been proposed in the deep learning literature and can be included in the search space as well, such as Foothill function \citep{belbahri2019foothill} that has interesting properties and is flexible enough to approximate the $L_1$ loss and the MSE.
Neural network activations determine the output of a node and are a crucial component of deep learning. When the activation function is non-linear, then a two-layer neural network can be proven to be a universal function approximator \citep{cybenko1989approximation}. The identity activation function does not satisfy this property. When multiple layers use the identity activation function, the entire network is equivalent to a single-layer model. The Parameteric Rectified Linear Unit (PReLU) \citep{he2015delving} has been proposed as an extension of Rectified Linear Unit (ReLU) \citep{nair2010rectified} in order to improve image classification accuracy. The PReLU function is defined as
\[
f_a (x) =
\begin{cases}
x, ~\mathrm{if}~ x \geq 0\\
ax, ~\mathrm{otherwise}
\end{cases}
\]
where $a \in [0;1]$ is learned in the training via backpropagation. When $a=0$, PReLU simplifies to ReLU and if $a$ is set to a fixed value, PReLU simplifies to Leaky ReLU (LReLU) \citep{maas2013rectifier}.
\paragraph{Search strategy.}
The search strategy details how to explore the search space
(which is often exponentially large or even unbounded). It encompasses the classical exploration-exploitation trade-off since, on the one hand, it is desirable to find
well-performing architectures quickly, while on the other hand, premature convergence to a region of suboptimal architectures should be avoided. In \cite{yang2019evaluation}, based on a benchmark of 8 NAS methods on 5 datasets, they compare methods with different search spaces, using a method’s relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, they find that many NAS techniques struggle to significantly beat the average architecture baseline. Furthermore, it is desirable to have a search strategy because unfortunately, in many deep learning tasks, training is computationally expensive. However, for our application, training SMITE on the real data takes 10 minutes for 100 epochs (on a Dual Intel Xeon 256 Gb RAM and 1 Titan V12GB). Therefore, we believe that grid search is suitable for SMITE architecture search. Table~\ref{tab:NAS} summarizes the hyper-parameters responsible for defining the architecture of SMITE. Hyper-parameter 3 must be defined for each layer. Although hyper-parameters 6-8 do not modify the architecture, they are important in terms of training. We prefer to enumerate them in this Table in order to include them in the architecture search. Hyper-parameter 9 is important for reproducible results. Our experimental results are presented in Section~\ref{sec:experiments}.
\begin{table}[htb]
\caption{Hyper-parameters that define the architecture of SMITE.}
\centering
\begin{tabular}{|c||l|l|l|}
\hline
\# & Hyper-parameter & Type & Scope\\
\hline
1 & \# of hidden layers & Integer & $\{0,1,2,...,10\}$\\
2 & \# of hidden units per layer & Integer & $\{32,64,128,256,512,1024\}$\\
3 & Activation & Nominal/Continuous & $\{$ReLU, LReLU$(a)\}$\\
4 & Regularization function & Nominal & $\{$SMITE (TO), SMITE (IE)$\}$ \\
5 & Regularization function for SMITE (TO) & Nominal/Integer & $\{L_1$,$L_2$,Foothill$\}$\\
6 & Batch size $m$ & Integer & $\{128,256\}$ \\
7 & Learning rate $\xi$ & Continuous & $\{0.01,0.02,0.03,0.04,0.05\}$ \\
8 & Trade-off constant $\alpha$ & Continuous & $\{0,0.1,0.2,...,1\}$ \\
9 & Seed & Integer & $\{1,2,3,...,10\}$ \\
\hline
\end{tabular}
\label{tab:NAS}
\end{table}
\paragraph{Performance estimation strategy.}
The objective of NAS is typically to find architectures that achieve high predictive performance on unseen data. Performance
estimation refers to the process of estimating this performance: the simplest option is to perform a standard training and validation of the architecture on data.
The goal of uplift modeling is to separate all observations into heterogeneous segments that differ as much as possible with regard to the reactions to the treatment and, thus, identify (sub)groups with very different uplifts. Then, the uplifts in the resulting segments serve as a prediction for the future behavior of the
observations. However, the \textit{ground truth} is not observed in the case of individual treatment effect estimation. Therefore, in order to evaluate uplift models, we refer to known metrics in the uplift literature, the {\it Qini coefficient} and the {\it Kendall's uplift rank correlation}. The computation of both statistics requires the construction of the {\it Qini curve} \citep{radcliffe2007using}. These statistics are commonly used as goodness-of-fit measures for uplift models.
For a given model, let $\hat{u}_{(1)} \geq \hat{u}_{(2)} \geq ... \geq \hat{u}_{(n)}$ be the sorted predicted uplifts. Let $\phi \in [0,1]$ be a given proportion and let $ N_{\phi} = \{i: \hat{u}_{i} \geq \hat{u}_{\ceil{\phi n}} \} \subset \lbrace 1, \ldots, n \rbrace$ be the subset of observations with the $\phi n \times 100 \%$ highest predicted uplifts $\hat{u}_i$ (here $\ceil{s}$ denotes the smallest integer larger or equal to $s\in \rm I\!R$). The \textit{Qini curve} is defined as a function $f$ of the fraction of population targeted $\phi$, where
\begin{equation*}
f(\phi) = \frac{1}{n_t}\biggl(\sum\limits_{i \in N_{\phi}} y_i t_i - \sum\limits_{i \in N_{\phi}} y_i (1-t_i) \biggl\{ \sum\limits_{i \in N_{\phi}} t_i / \sum\limits_{i \in N_{\phi}} (1-t_i) \biggr\} \biggr),
\end{equation*}
where $n_t = \sum_{i=1}^n t_i$ is the number of treated customers, with $f(0)=0$ and $f(1)$ is the average treatment effect. In other words, $f(\phi)$ represents the incremental number of positive responses for a fraction $\phi$ or targeted customers relative to the total number of targeted customers in the sample. The Qini helps visualize the performance of an uplift model, a bit like the ROC curve of a binary classification model. Following the same logic, a straight line between the points $(0,0)$ and $(1, f(1))$ defines a benchmark to compare the performance of the model to a strategy that would randomly target individuals. This means that if a proportion $\phi$ of the population is targeted, we expect to observe an uplift of $\phi f(1)$.
The {\it Qini coefficient} is defined as
\begin{equation*}
q = \int_0^1 Q(\phi) ~\mathrm{d}\phi = \int_0^1 \{f(\phi) - \phi~f(1)\} ~\mathrm{d}\phi,
\end{equation*}
where $Q(\phi) = f(\phi) - \phi~f(1)$. This area can be numerically approximated using a Riemann method such as the trapezoid rule formula: the domain of $\phi \in [0,1]$ is partitioned into $K$ bins, or $K+1$ grid points $0=\phi_1 < \phi_2 < ... < \phi_{K+1} = 1$, to approximate $q$ by
\begin{equation}
\hat{q} = \dfrac{1}{2} \sum_{k=1}^K (\phi_{k+1}-\phi_k)\{Q(\phi_{k+1}) + Q(\phi_{k})\}.
\label{eq:q:hat}
\end{equation}
It is important to note, unlike the area under the ROC curve, $\hat{q}$ could be negative. This would simply mean that a strategy following the model in question does worse than random targeting.
Measuring the goodness-of-fit of an uplift model can also be achieved by the similarity between the theoretical uplift percentiles (given by the model's predictions) and empirical percentiles (observed in the data) based on the model. This can be approximated by the {\it Kendall's uplift rank correlation} \citep{belba2019qbased} defined as
\begin{equation}
\hat{\rho} = \frac{2}{K(K-1)} \sum_{i<j} \mathrm{sign}(\bar{\hat{u}}_i - \bar{\hat{u}}_j)~\mathrm{sign}(\bar{u}_i - \bar{u}_j),
\label{eq:corr_coeff}
\end{equation}
where $\bar{\hat{u}}_k$ is the average predicted uplift in bin $k$, $k \in {1,...,K}$, and $\bar{u}_k$ is the observed uplift in the same bin.
\section{Experiments}
\label{sec:experiments}
We conduct a simulation study to examine the performance SMITE. We compare its performance with existing uplift methods using three types of datasets: i) a real-world dataset from a marketing campaign (see Section~\ref{sec:application} for the analysis); ii) bootstrapped versions of the real-world dataset with synthetic outcomes generated from tree-based models, and iii) synthetic datasets generated from parametric models. First, we show results for the SMITE architecture search.
\subsection{Fine-tuning SMITE}
\textcolor{blue}{HERE WE NEED TO SHOW RESULTS FROM THE ARCHITECTURE SEARCH AND SHOW THE FINAL ARCHITECTURE THAT WE WILL USE IN THE EXPERIMENTS.}
\textcolor{red}{The backbone of the SMITE architecture is a $6$-layers feed forward MLP with the following input neurons sizes: $\{200,200,300,100,50,10\}$. The first two layers are linearly connected while the rest are connected through a leaky ReLu activation function. Both the depth and the width of this network were explored using a random grid search on data generated with the framework described in Section~\ref{sec:NAS}. Two hyper-parameters are fine-tuned for each set of experiments: the learning rate $\xi$ and the trade-off constant $\alpha$. To find an adequate candidate for $\alpha$, we follow the following protocol: the dataset $D$ is split into two subsets ($D_T$,$D_H$) containing respectively $70\%$ and $30\%$ of $D$. $D_H$ remains unseen during any of our fine-tuning or model selection process. We then repeat $k$-times a random separation of $D_T$ into two datasets denoted ($D^{i}_{\mathrm{train}}$,$D^{i}_{\mathrm{valid}}$) with sizes $60\%$ and $40\%$ of $D_T$. At a fixed learning rate of $\xi=0.03$, we fit the network on $D^{i}_{\mathrm{train}}$ for different values of $\alpha$, from $0$ to $1$ with $0.1$ increments. We measure the average Qini coefficient $\hat q$ obtained on $\{D^{i}_{valid}\}_{i \in [1,k]}$ and we estimated a $95\%$ confidence interval (CI) for $\hat q$. The selected $\alpha$ corresponds to the highest average as long as the $95\%$ CI lower bound remains higher than $0$. The resultant of this protocol is illustrated in Figure \ref{fig:alphaie}. Once $\alpha$ is fixed, we follow then the same process for the learning rate $\xi$.}
\input{figures/alpha.tex}
\subsection{Data generation}
We evaluate the performance of SMITE using synthetic data generated using two types of mechanisms.
\paragraph{Bootstrap Data.} The underlying treatment effects are inaccessible in real-world data. To create realistic scenarios, we based one of our simulation study data generation on insurance policy data. Moreover, since we compare SMITE to random forests, we want to evaluate its performance when the data generation process comes from a tree-based model. We follow the procedure in the simulation studies of \cite{belba2019qbased}. These simulations allow us to evaluate if SMITE can recover the right representation for data generated from a tree-based model. This also makes comparisons to random forest methods fair.
Several tree-based methods have been suggested in the uplift literature. Here, we use the uplift random forest \citep{guelman2012random} as the data generation process, which is readily available in the {\bf R} library {\bf uplift}. We fit the random-forest on a random sub-sample $\mathcal{D}$ of the real data. We use $100$ trees and the Kullback-Leibler split criterion. Each tree has maximum depth of $5$. Based on the fitted model, we can extract the probabilities $\mu_0(\mathbf x)$ and $\mu_1(\mathbf x)$ for any given value $\mathbf{x}$. We use these probabilities to draw synthetic outcomes for treated and control observations. We start by creating a bootstrap sample $\mathcal{S}$ of size $n_\mathcal{S}$ from $\mathcal{D}$. For each observation $\mathbf{x}_i \in \mathcal{S}$, we generate a random vector $\tilde{y}_i =(\tilde{y}_{i0}, \tilde{y}_{i1})$, where $\tilde{y}_{i0}$ is the binary outcome of a Bernoulli trial with success probability $\mu_0(\mathbf{x}_i)$, and $\tilde{y}_{i1}$ is the binary outcome of a Bernoulli trial with success probability $\mu_1(\mathbf{x}_i)$, $i=1,\ldots, n_\mathcal{S}$. The augmented synthetic dataset $\{ \mathbf x_i , t_i, \tilde{y}_i \}_{i=1}^{n_{\mathcal{S}}}$ is the data of interest in the simulation.
We repeat this process several times. Each bootstrapped data size is $10,000$ observations with $40$ covariates.
\paragraph{Synthetic Data.} We evaluate the performance of SMITE models using synthetic data generated in a similar way to the simulation studies of \cite{tian2014simple}. For the synthetic data generation, there are four elements:
\begin{enumerate}
\item The number $n$ of observations in the sample, and the number $p$ of predictors observed for each observation.
\item The distribution of the predictors. We draw predictors from a multivariate Gaussian distribution.
\item The mean effect function $\mu(\mathbf X)$ and the treatment effect function $\tau(\mathbf X)$. We vary the size of main effects relative to treatment effects.
\item The noise level $\sigma^2$. This corresponds to the conditional variance of the outcome given $\mathbf X$ and $T$.
\end{enumerate}
Given the elements above, our data generation model is, for $i=1,...,n$ :
\begin{align*}
&\mathbf X_i \sim \mathcal{N}_p (0, \Sigma) \\
&T_i \sim \mathrm{Bernoulli}(\theta)\\
&Y_i^* \mid \mathbf X_i, T_i \sim \mathcal{N}(\mu(\mathbf X_i) + T_i\tau(\mathbf X_i), \sigma^2).
\end{align*}
Define $Y_i$ to be the binary outcome as
\begin{align*}
Y_i \mid \mathbf X_i, T_i = \mathbbm{1}(Y_i^* > 0 \mid \mathbf X_i, T_i).
\end{align*}
Hence, $Y_i$ follows a Bernoulli distribution with parameter $\mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i)$ and it is easy to recover the ``true'' uplift $u_i^*$ as follows
\begin{align*}
u_i^* &= \mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i = 1) - \mathrm{Pr}(Y_i^* > 0 \mid \mathbf X_i, T_i = 0)\\
u_i^* &= \Phi \Bigl( \frac{\mu(\mathbf X_i) + \tau(\mathbf X_i)}{\sigma} \Bigr) - \Phi \Big( \frac{\mu(\mathbf X_i)}{\sigma} \Big)
\end{align*}
where $\Phi(.)$ is the standard Gaussian cumulative distribution function. Note that $\mu(.)$ is a nuisance function and $\tau(.)$ is the function that controls the treatment effect. In fact, the sign of $\tau(.)$ and the sign of the uplift for a given observation coincide.
We set $\theta$ to $0.5$, which is the case in our real data. The variance-covariance matrix elements $\Sigma_{ij}$ are defined as follows:
\[\Sigma_{ij} = \begin{cases}
\rho & i\neq j \\
1 & i=j
\end{cases}.
\]
Finally, the mean and treatment effects functions are defined as follows:
$$\mu(\mathbf X) = \sum_{j=1}^J \beta_j X_j + \sum_{j=1}^{p-1} \omega_j X_j X_{j+1},$$
$$\tau(\mathbf X) = \sum_{j=1}^K \gamma_j X_j$$
where $J\leq p$ and $K \leq p$. We set $\beta_j = \beta (-1)^{j}$, $\omega_j = \omega (-1)^j$ and we draw $\gamma_j$ independently from $\mathcal{N}(0,\gamma)$. The constants $\beta$, $\omega$ and $\gamma$ are hyper-parameters.
In this case, we generate datasets of sizes $n=20,000$ observations with $100$ covariates.
\subsection{Benchmarks and implementation}
Ensemble methods are amongst the most powerful for the uplift case \citep{soltys2015ensemble}. A method that has proven to be very effective in estimating ITE is one based on generalized random forests \citep{athey2019generalized}. This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the treatment effects. This has the effect of reducing overfitting. Another candidate that we consider in our experiments is also based on random forests that were designed for uplift \citep{guelman2012random, jaskowski2012uplift}. For this method, we consider two different split criteria, one based on Kullback-Leibler divergence and one based on Euclidean distance.
For random forests (RF), we use the implementation from the \textbf{R} libraries \textbf{grf} and \textbf{uplift} which were developed by the authors of the methods respectively. Implementations of these causal forests have several useful parameters that make it easy to fine tune them. In particular, \textbf{grf} provides cross-validation procedures for hyper-parameters tuning which we use to get the best possible candidates to compare our results to. We also use the \textbf{R} library \textbf{tools4uplift} for computing the model's goodness-of-fit measures. SMITE is implemented in \textbf{Python}, so we reimplemented the Qini (\ref{eq:q:hat}) and Kendall (\ref{eq:corr_coeff}) coefficients functions. In total, we consider $4$ different random forests to compare to SMITE: causal forest, honest causal forest (H), RF with Kullback-Leibler (KL) and RF with Euclidean distance (ED) split criteria.
\subsection{Experimental results}
For both simulated datasets, we do not consider any out-of-sample subset. In Monte Carlo settings, we are able to generate as much data as we want. Therefore, we first fine-tuned the methods on a single synthetic dataset of same size as the one we used in the repeated simulations presented. We generate $n=10,000$ observations with $p=40$ covariates for the bootstrapped data. For the fully synthetic datasets, we report results from experiments with $n=10,000$ observations and $p=100$ covariates. We split into training and validation and compare the results averaged over $30$ runs for the models that performed best on the validation set. Results are described in Table \ref{tab:boot_perf} for the bootstrap datasets and Table \ref{tab:synthe_perf} for the synthetic datasets. For the fully synthetic data, we see that Causal Forest (H) and SMITE (TO) had similar performance on the validation set while SMITE (IE) significantly outperformed all the other methods. For the bootstrap data for which the outcomes were generated based on random forest model, it is interesting to see that SMITE (TO) and Causal Forest (H) give similar results. Moreover, once again, the SMITE (IE) model outperforms all the other models.
\input{tables/table_bootstrap.tex}
\textcolor{blue}{HERE WE NEED TO SHOW ESTIMATED AND PREDICTED UPLIFTS DISTRIBUTIONS PER METHOD. I THINK THIS WILL SHOW THAT RF METHODS COVER A RANGE THAT IS A LOT LARGER THAN SMITE. WE CAN ALSO COMPUTE THE EMPIRICAL VARIANCES AND COMPUTE A CONFIDENCE INTERVAL FOR THE VARIANCE OF THE PREDICTED UPLIFTS TO SEE IF THERE ARE SIGNIFICANTLY LARGER THANT SMITE'S VARIANCE.}
\input{tables/table_synthetic.tex}
\section{Application}
\label{sec:application}
A data set with $n=50,000$ customers and $p=40$ covariates from an insurer is at our disposal to evaluate the performance of our methods. This company was interested in designing strategies to maximize its conversion rate. An experimental acquisition campaign was implemented for 6 months, for which half of the potential clients were randomly allocated into a treatment group and the other half into a control group. Potential clients under the treatment group were contacted using a different (and more costly) strategy. The goal of the analysis was to be able to select who among the clients portfolio should be contacted in that way going forward in order to maximize the benefits. The observed difference in sales rates between the treated group and the control shows some evidence of a positive impact of the treatment ($\mathrm{ATE}=0.55\%$).
\textcolor{blue}{DESCRIBE THE DATA AND GIVE MORE DETAILS ON THE APPLICATION. ADD FIGURES FOR BEST FINAL MODELS (Qini Curves).}
For all models we follow the same learning process and performance metrics computations. For the real-world dataset, a sample of $30\%$ is dedicated to out-of-sample performance measures. We do not score these customers until all models are properly fine-tuned. The remaining $70\%$ is used for fine-tuning. We repeat the experiment $30$ times. We randomly split the subset into $60\%$ for training and $40\%$ for validation. We select the model which performs best on the validation set. Using that model, we score the out-of-sample customers and compute the performance metrics. Results are presented in Table \ref{tab:realdata_perf}.
\input{tables/table_real_data.tex}
We observe that the random forests tend to overfit. Indeed, based on the out-of-sample results from the hold out set, the models seem to be close to random targeting. This might be due to the small (in magnitude) overall impact of the initiative. On the other hand, we observe a good adequacy of SMITE for both TO and IE. Overall, the results of SMITE (IE) appear to outperform all other methods. It is also interesting to note that the distribution of the ITE obtained by the model SMITE (IE) are in line with what was logically expected (see Figure \ref{fig:pred_uplift_evol}: giving what appears to be a better service only slightly increases the propensity of buying.
\input{figures/uplift_evolutions.tex}
\section{Conclusion}
\label{sec:conclusion}
In this paper we present a meaningful and intuitive neural networks architecture for the problem of uplift modeling.
We apply our methods to both real-world and synthetic data and we compare with what is, to the best of our knowledge, state-of-the-art methods for uplift. Our results show that the SMITE models significantly outperform random forests designed for ITE and uplift in all of our scenarios. For future work, we want to study the theoretical properties of the SMITE optimization problem as well as including more causal inference related problems such as multiple treatments, time series, and the use of observational data.
\bibliographystyle{plainnat}
\section{Experiments}
\label{sec:appendix_experiments}
In this section, we present in more detail the experimental protocol as well as the results associated with the different scenarios.
Each simulation is defined by a data-generating process with a known effect function. Each run of each simulation generates a dataset, and split into training, validation, and test subsets. We use the training data to estimate $L$ different uplift functions $\{\hat{u}_l\}_{l=1}^L$ using $L$ different methods. For each of those we calculate the validation metric $\hat{q}_{\mathrm{adj}}$ using the data in validation set. The models selected by optimizing each validation-set metric are then used to predict the uplift on the test set.
\paragraph{Data generating process}
We generate synthetic data similar to \cite{powers2018some}. For each experiment, we generate $n$ observations and $p$ covariates. We draw odd-numbered covariates independently from a standard Gaussian distribution. Then, we draw even-numbered covariates independently from a Bernoulli distribution with probability $1/2$. Across all experiments, we define the mean effect function $\mu(\cdot)$ and the treatment effect function $\tau(\cdot)$ for a given noise level $\sigma^2$. Given the elements above, our data generation model is, for $i=1,...,n$,
\begin{align*}
&Y_i^* \mid \mathbf x_i, t_i \sim \mathcal{N}(\mu(\mathbf x_i) + t_i\tau(\mathbf x_i), \sigma^2),\\
&Y_i = \mathbbm{1}(Y_i^* > 0 \mid \mathbf x_i, t_i)
\end{align*}
where $t_i$ is the realisation of the random variable $T_i \sim \mathrm{Bernoulli}(1/2)$ and $Y_i$ is the binary outcome random variable. Therefore, $Y_i$ follows a Bernoulli distribution with parameter $\mathrm{Pr}(Y_i^* > 0 \mid \mathbf x_i, t_i)$ and it is easy to recover the ``true'' uplift $u_i^*$, that is,
\begin{align}
u_i^* &= \mathrm{Pr}(Y_i^* > 0 \mid \mathbf x_i, T_i = 1) - \mathrm{Pr}(Y_i^* > 0 \mid \mathbf x_i, T_i = 0) \nonumber \\
u_i^* &= \Phi \Bigl( \frac{\mu(\mathbf x_i) + \tau(\mathbf x_i)}{\sigma} \Bigr) - \Phi \Big( \frac{\mu(\mathbf x_i)}{\sigma} \Big) \label{eq:sim_true_uplift}
\end{align}
for $i=1,\ldots,n$, where $\Phi(.)$ is the standard Gaussian cumulative distribution function.
Following \cite{powers2018some}, within each set of simulations, we make different choices of mean effect function and treatment effect function, each represents a wide variety of functional forms, univariate and multivariate, additive and interactive, linear and piecewise constant. The functions are
\begin{align*}
f_1(\mathbf x) &= 0 \\
f_2(\mathbf x) &= 5 \mathbbm{1}(x_1 > 1) - 5\\
f_3(\mathbf x) &= 2 x_1 - 4 \\
f_4(\mathbf x) &= x_2 x_4 x_6 + 2 x_2 x_4 (1-x_6) +3x_2(1-x_4)x_6 + 4x_2(1-x_4)(1-x_6) + 5(1-x_2)x_4x_6 \\
&+ 6(1-x_2)x_4(1-x_6) + 7 (1-x_2)(1-x_4)x_6 + 8(1-x_2)(1-x_4)(1-x_6)\\
f_5(\mathbf x) &= x_1 + x_3 + x_5 + x_7 + x_8 + x_9 - 2\\
f_6(\mathbf x) &= 4 \mathbbm{1}(x_1 > 1)\mathbbm{1}(x_3 > 0) + 4\mathbbm{1}(x_5 > 1)\mathbbm{1}(x_7 > 0) + 2x_8x_9\\
f_7(\mathbf x) &= \frac{1}{2} (x_1^2 + x_2 + x_3^2 + x_4 + x_5^2 + x_6 + x_7^2 + x_8 + x_9^2 - 11)\\
f_8(\mathbf x) &= \frac{1}{\sqrt{2}} (f_4(\mathbf x) + f_5(\mathbf x)).
\end{align*}
Table~\ref{tab:scenarios} gives the mean and treatment effect functions for the different randomized simulations, in terms of the eight functions above. In Figure~\ref{fig:qini_curves_scene}, we show the Qini curves associated with the ``true" uplift $u^*$.
\begin{table}[!ht]
\centering
\caption{Specifications for the 5 simulation scenarios. The rows of the table correspond, respectively, to the sample size, dimensionality, mean effect function, treatment effect function and noise level.}
\label{tab:scenarios}
\begin{tabular}{l|ccccc}
Scenarios & 1 & 2 & 3 & 4 & 5\\
Parameters & & & & & \\
\hline
$n$ & 10000 & 20000 & 20000 & 20000 & 20000 \\
$p$ & 200 & 100 & 100 & 100 & 100 \\
$\mu(\mathbf x)$ & $f_7(\mathbf x)$ & $f_3(\mathbf x)$ & $f_1(\mathbf x)$ & $f_2(\mathbf x)$ & $f_6(\mathbf x)$\\
$\tau(\mathbf x)$ & $f_4(\mathbf x)$ & $f_5(\mathbf x)$ & $f_6(\mathbf x)$ &$ f_7(\mathbf x)$ & $f_8(\mathbf x)$\\
$\sigma$ & $1/2$ & $1$ & $1$ & $2$ & $4$\\
\hline
\end{tabular}
\end{table}
The defined scenarios are interesting because they represent different situations that we can face with real data. Indeed, scenario 1, although it seems simpler than the others (with a small noise level), is in fact a case that happens often, especially for marketing campaigns. In this scenario, by construction, there are no do-not-disturb clients. This results in an increasing monotonic Qini curve. In scenario 2, we introduce a group of do-not-disturb clients. In addition, we increase the noise level as well as the number of observations, but we reduce the number of variables. The average treatment effect is negative but a small group of persuadables is generated. In scenario 3, the main effect function is zero. Thus, the uplift is entirely guided by the treatment effect function. In scenario 4, we introduce a quadratic treatment effect function and scenario 5 is the most complex one.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.49\textwidth]{figures/simulations/Qini_Curve_Scene_1.pdf}
\includegraphics[width=0.49\textwidth]{figures/simulations/Qini_Curve_Scene_2.pdf}
\includegraphics[width=0.49\textwidth]{figures/simulations/Qini_Curve_Scene_3.pdf}
\includegraphics[width=0.49\textwidth]{figures/simulations/Qini_Curve_Scene_4.pdf}
\caption{Qini curves based on the ``true" uplift with respect to scenarios 1-4 of Table~\ref{tab:scenarios}.}
\label{fig:qini_curves_scene}
\end{figure}
\paragraph{Benchmarks and implementation}
Each run of each simulation generates a dataset, which we split into training ($40\%$), validation ($30\%$), and test samples ($30\%$). We use the training and the validation sets to fit and fine-tune the models. Note that fine-tuning is specific to the type of the model, e.g. for a neural network, we fine-tune the learning rate and the regularization constant but for random forests, we fine-tune the number of trees as well as the depth of each tree. Then, we score the test samples with the fine-tuned models and compute performance metrics. We repeat each experiment $20$ times.
For each simulated dataset, we implement the following models:
\begin{enumerate}
\item[(a)] a multivariate logistic regression, i.e. a twin model with no hidden layer that optimizes the Binomial likelihood. This is the baseline model, and we will refer to it as \textit{Logistic}. For fair comparison, we use the twin neural architecture and $L_1$ regularization.
\item[(b)] our twin models, i.e. models (with and without hidden layers) that optimize the augmented loss function \eqref{eq:augmented_loss_function} to estimate the regression parameters. We denote these models by \textit{Twin$_{\mu}$} and \textit{Twin$_{\mathrm{NN}}$}.
\item[(c)] a Qini-based uplift regression model that uses several LHS structures to search for the optimal parameters (see \citep{belba2019qbased} for more details). We denote this model by \textit{Qini-based}.
\item[(d)] two types of random forests designed for causal inference (see \citep{athey2019generalized} for more details). This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the uplift. We denote these models by \textit{Causal Forest} and \textit{Causal Forest (Honest)} without and with honest estimation respectively.
\item[(e)] two types of random forests designed for uplift (see \citep{rzepakowski2010decision,guelman2012random} for more details). This method uses different split criteria. We denote these models by \textit{Uplift Random Forest (KL)} and \textit{Uplift Random Forest (ED)} for Kullback-Leibler and Euclidean split criterion respectively.
\end{enumerate}
\paragraph{Results} All results are reported in Table~\ref{tab:all_scenarios}.
\begin{table}[!ht]
\centering
\begin{tabular}{l|c|ccccc}
\hline
Method & Hidden-Layer Size & 1 & 2 & 3 & 4 & 5 \\
\hline
{\it Logistic} && $0.75$ & $1.80$ & $3.50$ & $0.98$ & $2.59$\\
{\it Qini-based} && $1.02$ & $\bf2.68$ & $4.36$ & $1.09$ & $2.94$\\
{\it Twin$_{\mu}$} && $\bf1.24$ & $2.32$ & $\bf4.41$ & $\bf1.20$ & $\bf3.12$\\
\hline
{\it Twin$_{\mathrm{NN}}$}
& $64$ & $1.54$ & $2.61$ & $\bf 4.91$ & $1.22$ & $3.10$ \\
& $128$ & $1.56$ & $2.54$ & $4.80$ & $1.28$ & $3.16$\\
& $201$ & $\bf 1.63$ & $\bf 2.67$ & $4.86$ & $1.30$ & $3.14$\\
& $256$ & $1.57$ & $2.65$ & $4.78$ & $\bf 1.32$ & $3.26$ \\
& $512$ & $1.60$ & $2.49$ & $4.64$ & $1.18$ & $\bf 3.35$\\
\hline
& $340/512$ & $1.67$ & $2.63$ & $5.03$ & $1.35$ & $3.58$\\
\hline
{\it Causal Forest} & &$0.75$ & $2.22$ & $4.42$ & $0.94$ & $2.79$ \\
{\it Causal Forest (Honest)} && $0.75$ & $2.51$ & $4.53$ & $1.14$ & $3.07$\\
{\it Uplift Random Forest (KL)} && $0.74$ & $2.52$ & $4.02$ & $1.01$& $2.19$ \\
{\it Uplift Random Forest (ED)} && $0.68$ & $2.42$ & $4.18$ & $0.99$ & $2.33$ \\
\hline
\end{tabular}
\vspace{2mm}
\caption{Summary: models comparison in terms of $\hat{q}_{\mathrm{adj}}$.}
\label{tab:all_scenarios}
\end{table}
\paragraph{The optimization criterion}
In this part of the analysis, we focus on the simple model (i.e. no hidden layer {\color{magenta} [do you mean the logistic regression with interaction terms, better to refer with eqref to the right model to be precise]}), $\mu_{1t}(\mathbf x, \boldsymbol{\theta})$, defined in \eqref{eq:neural_network_1}. The goal is to compare three different ways to estimate the parameters. The {\it Logistic} method optimizes the penalized Binomial likelihood, as in the case of a {\it lasso} logistic regression. The {\it Qini-based} is a two-stage method which also makes use of the $L_1$ regularization. The methodology is based on $\hat{q}_{\mathrm{adj}}$ and imposing sparsity using the lasso. The $\hat{q}_{\mathrm{adj}}$ is a difficult statistic to compute, maximizing it directly is not an easy task. Therefore, the {\it Qini-based} first applies the pathwise coordinate descent algorithm \citep{friedman2007pathwise} to compute a sequence of critical regularization values and corresponding (sparse) model parameters. Then, it uses a derivative-free optimization algorithm (Latin Hypercube Sampling or LHS) to explore the parameters space in order to find the optimal model. The {\it Augmented loss} method optimizes our augmented loss function defined in \eqref{eq:augmented_loss_function} with $L_1$ regularization. Table~\ref{tab:param_estimation_results} shows the averaged results on the test dataset, for different scenarios.
\begin{table}[!ht]
\centering
\caption{Comparison of parameter estimation for the $\mu_{1t}(\mathbf x, \boldsymbol{\theta})$ model in terms of $\hat{q}_{\mathrm{adj}}$.}
\label{tab:param_estimation_results}
\begin{tabular}{l|ccccc}
Scenarios & 1 & 2 & 3 & 4 & 5\\
Method & & & & & \\
\hline
{\it Logistic + $L_1$} & $0.75$ & $1.80$ & $3.50$ & $0.98$ & $2.59$\\
{\it Qini-based} & $1.02$ & $\bf2.68$ & $4.36$ & $1.09$ & $2.94$\\
{\it Augmented loss + $L_1$} & $\bf1.24$ & $2.32$ & $\bf4.41$ & $\bf1.20$ & $\bf3.12$\\
\hline
\end{tabular}
\end{table}
If we pick “winners” in each of the simulation scenarios based on which
method has the highest $\hat{q}_{\mathrm{adj}}$, {\it AL} would win all scenarios except for the second one. In scenario 2, the {\it Qini-based} model has the highest $\hat{q}_{\mathrm{adj}}$. In general both methods outperform the {\it Logistic} model.
\paragraph{Beyond linearity}
Ensemble methods are amongst the most powerful for the uplift case \citep{soltys2015ensemble}. A method that has proven to be very effective in estimating individual treatment effects is one based on generalized random forests \citep{athey2019generalized}. This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the uplift. This has the effect of reducing overfitting. Another candidate that we consider in our experiments is also based on random forests and designed for uplift \citep{rzepakowski2010decision,guelman2012random}. For this method, we consider two different split criteria, one based on the Kullback-Leibler divergence and one based on the Euclidean distance. In Table~\ref{tab:beyond_linearity_results}, we compare these methods to our neural-network version defined in \eqref{eq:neural_network_2}, that is, $\mathrm{NN}_{1t}(\mathbf x, \mathbf{W})$ with $m=2p+1$ hidden neurons. Here, we only consider $L_1$ regularization on the model's parameters.
\begin{table}[!ht]
\centering
\caption{{\color{magenta} [bad table caption, why is it called AL? what is CF? URF? remember a table caption must be self contained.] Rethink about abbreviations, do not abbreviate if it is not necessary, otherwise it explodes your caption.}Comparison of non-linear models in terms of $\hat{q}_{\mathrm{adj}}$. The {\it AL} method optimizes the $\mathrm{NN}_{1t}(\mathbf x, \mathbf{W})$ model with $m=2p+1$ hidden neurons and $L_1$ regularization.}
\label{tab:beyond_linearity_results}
\begin{tabular}{l|ccccc}
Scenarios & 1 & 2 & 3 & 4 & 5\\
Method & & & & &\\
\hline
{\it AL$_{(p+1)\times m}$} & $\bf 1.63$ & $\bf 2.67$ & $\bf 4.86$ & $\bf 1.30$ & $\bf 3.14$\\
{\it CF} & $0.75$ & $2.22$ & $4.42$ & $0.94$ & $2.79$ \\
{\it CF$_{h}$} & $0.75$ & $2.51$ & $4.53$ & $1.14$ & $3.07$\\
{\it URF$_{\mathrm{KL}}$} & $0.74$ & $2.52$ & $4.02$ & $1.01$& $2.19$ \\
{\it URF$_{\mathrm{ED}}$} & $0.68$ & $2.42$ & $4.18$ & $0.99$ & $2.33$ \\
\hline
\end{tabular}
\end{table}
We observe that the random forests tend to overfit for several scenarios, with the exception of the honesty criteria which seems to mitigate this problem. On the other hand, we observe a good adequacy of the twin model in all scenarios. Overall, the results of {\it AL} appear to outperform all other methods significantly (based on Wilcoxon signed-rank test \citep{wilcoxon1992individual}).
\paragraph{Parameter size}
{\color{magenta} [this section is messy. simplify your paragraphs. I got lost.]}
Neural networks have two main hyper-parameters that control the architecture or topology of the network: the number of hidden layers and the number of nodes in each hidden layer. In the previous simulations, we fixed the number of hidden layers to $1$ and the number of nodes in the hidden layer to $2p+1$ as per Theorem~\ref{theorem}. However, the number of nodes can be seen as a hyper-parameter as well.
Here, we vary the number of nodes to assess the impact on predictive performance, using either $m=64, 128, 256$ or $m=512$ nodes. In Table~\ref{tab:hidden_neurons_results}, we report the averaged results. As we can see, there is not one optimal value of $m$ among the ones we compare that gives the best model for all scenarios. Depending on the scenario, we observe clear differences in terms of performance.
For scenario 1, in which $p = 200$, we see that the best model is reached for $m = 512$; for scenario 2 and scenario 3, the model performs best with $m = 64$ and $m = 256$ respectively, and finally with $m = 256$ and $m = 512$ for scenarios 4 and 5. Therefore, in practice, for a specific dataset, it is important to consider the number of neurons in the hidden layer as a hyper-parameter that can be fine-tuned by cross-validation.
\begin{table}[!ht]
\centering
\caption{Fine-tuning the number of nodes in terms of $\hat{q}_{\mathrm{adj}}$. The {\it AL} method optimizes the $\mathrm{NN}_{1t}(\mathbf x, \mathbf{W})$ model with $m$ hidden neurons and $L_1$ regularization.}
\label{tab:hidden_neurons_results}
\begin{tabular}{l|l|ccccc}
Scenarios &$m$ & 1 & 2 & 3 & 4 & 5\\
Method &&&&&&\\
\hline
{\it AL$_{(p+1)\times m}$}
& $64$ & $1.54$ & $2.61$ & $\bf 4.91$ & $1.22$ & $3.10$ \\
& $128$ & $1.56$ & $2.54$ & $4.80$ & $1.28$ & $3.16$\\
& $256$ & $1.57$ & $\bf 2.65$ & $4.78$ & $\bf 1.32$ & $3.26$ \\
& $512$ & $\bf 1.60$ & $2.49$ & $4.64$ & $1.18$ & $\bf 3.35$\\
\hline
\end{tabular}
\end{table}
Choosing the number of hidden neurons by cross-validation is very common in practice. This is part of architecture search. However, choosing among a few values does not necessarily mean that all the selected model's neurons are useful. As discussed in Section~\ref{sec:choice_reg_function}, the use of the group lasso penalization on the weight matrix allows to automatically fine-tune the number of hidden neurons (i.e. pruning). This has the effect of increasing the performance of the underlying models from a predictive point of view, as shown in Table~\ref{tab:hidden_neurons_pruning}. Thus, we suggest to start with a fairly large number of hidden neurons (e.g. $m \geq 2p +1$), and to let the optimization determine the number of active neurons needed for a specific dataset.
\begin{table}[!ht]
\centering
\caption{Performance when automatically pruning the number of nodes. The {\it AL} method optimizes the $\mathrm{NN}_{1t}(\mathbf x, \mathbf{W})$ model with $m=512$ hidden neurons and $L_1$ + \emph{group lasso} regularization functions.}
\label{tab:hidden_neurons_pruning}
\begin{tabular}{l|l|ccccc}
Scenarios & & 1 & 2 & 3 & 4 & 5\\
\hline
{\it AL$_{(p+1)\times 512}$}
& $\hat{m}$ & $422$ & $97$ & $48$ & $178$ & $340$ \\
& $\hat{q}_{\mathrm{adj}}$ & $1.67$ & $2.63$ & $5.03$ & $1.35$ & $3.58$\\
\hline
\end{tabular}
\end{table}
\paragraph{Summary} The number of hidden layers can also be seen as a hyper-parameter for the neural network architecture. In our experiments, increasing the number of hidden layers did not improve the performance significantly. When we fix the initial number of hidden neurons to $p+1$ and $p$ for the $2$ hidden layer, in scenarios 1 and 5, the average $\hat{q}_{\mathrm{adj}}$ are $1.59$ and $3.39$ respectively. We report the averaged results as well as a summary of all previous experiments in Table~\ref{tab:hidden_layers_results}. For random forests uplift models, we report the best score between the four used methods. For scenarios 1, 3, 4 and 5, the best results are obtained with the $1$-hidden layer twin model. For scenario 2, the {\it Qini-based} uplift regression achieves the highest $\hat{q}_{\mathrm{adj}}$, tying with the $1$-hidden layer twin model (i.e. $2.68$ vs. $2.67$).
{\color{magenta}[This is not summary. The conclusion is what? Scnarios change the performance? number of hidden units are not important? lasso and group lasso finds the moinimal $m$? ]}
\begin{table}[H]
\centering
\caption{Summary: performance of fitted models in terms of $\hat{q}_{\mathrm{adj}}$.}
\label{tab:hidden_layers_results}
\begin{tabular}{l|l|ccccc}
Scenarios &$h$ & 1 & 2 & 3 & 4 & 5\\
Method &&&&&&\\
\hline
{\it AL + $L_1$} & $0$ & $1.24$ & $2.32$ & $4.41$ & $1.20$ & $3.12$\\
{\it Logistic + $L_1$} & - & $0.75$ & $1.80$ & $3.50$ & $0.98$ & $2.59$\\
{\it Qini-based} & - & $1.02$ & $\bf 2.68$ & $4.36$ & $1.09$ & $2.94$\\
\hline
{\it AL + $L_1$} & $1$ & $1.63$ & $2.67$ & $4.91$ & $1.32$ & $3.35$\\
{\it AL + $L_1$ + \emph{group lasso}} & $1$ & $\bf 1.67$ & $2.63$ & $\bf5.03$ & $\bf1.35$ & $\bf3.58$\\
Best {\it RF} & - & $0.75$ & $2.52$ & $4.53$ & $ 1.14$ & $3.07$\\
\hline
\end{tabular}
\end{table}
\subsection{Alternative losses and activations}
The choice of loss functions and activations in deep networks has a significant effect on the training dynamics and task performance.
\subsubsection{Definitions}
Foothill function \citep{belbahri2019foothill} is defined as
\begin{align}
p_{\alpha,\beta}(x) = \alpha x~ \mathrm{tanh}\left( \frac{\beta x}{2} \right).
\label{eq:reswish}
\end{align}
where $\mathrm{tanh}(.)$ is the hyperbolic tangent function, $\alpha > 0$ is a shape parameter and $\beta > 0$ is a scale parameter.
The function is symmetric about $0$ (see Figure \ref{fig:foothill}, left panel). The first and the second derivatives (see Figure \ref{fig:foothill}, right panel) of foothill are
\begin{align*}
\frac{dp_{\alpha,\beta}(x)}{dx} = \alpha ~ \mathrm{tanh}\left( \frac{\beta x}{2} \right) + \frac{1}{2} \alpha \beta x ~ \mathrm{sech}^2\left( \frac{\beta x}{2} \right), \\
\frac{d^2p_{\alpha,\beta}(x)}{d^2x} = \frac{1}{2} \alpha \beta~ \mathrm{sech}^2\left( \frac{\beta x}{2} \right) \left\{ 2 - \beta x ~ \mathrm{tanh}\left( \frac{\beta x}{2} \right) \right\},
\end{align*}
where $\mathrm{sech}(.)$ is the hyperbolic secant function.
\begin{figure}[htb]
\centering
\begin{tikzpicture}[scale=0.5]
\begin{axis}[
axis x line=middle, axis y line=middle, samples = 200,
ylabel={$p_{\alpha,\beta}(x)$},ylabel style={font=\Large, yshift=3pt},
ymin=-0.1, ymax=2.5, ytick={0,1,2},
xmin=-3, xmax=3, xtick={-1, 0, 1}
]
\addplot[black, domain=-2.4:2.4, dashed]{1*2*(x-0)*((1+pow(e,-1*(x-0)))^(-1) - 0.5)};
\addplot[black, domain=-2:2, smooth]{1*2*(x-0)*((1+pow(e,-50*(x-0)))^(-1) - 0.5)};
\end{axis}
\end{tikzpicture}
\hspace{0.5cm
\begin{tikzpicture}[scale=0.5]
\begin{axis}[
axis x line=middle, axis y line=middle, samples = 100, ylabel={$\frac{d^{(l)}p_{1,1}(x)}{d^{(l)}x}$},ylabel style={font=\Large, yshift=3pt}, ymin=-2, ymax=2, ytick={-1,0,1}, xtick={-2.4, -1, 0, 1, 2.4},
xmin=-4, xmax=4
]
\addplot[black, domain=-4:4, dashed]{2*(1+pow(e,-1*x))^(-1)*(1+1*x*pow(e,-1*x)*(1+pow(e,-1*x))^(-1)))-1};
\addplot[black, domain=-4:4, smooth]{2 * (-1^2 * x * pow(e, -1*x) * (1+pow(e,-1*x))^(-2) + 2 * 1^2 * x * pow(e, -2*1*x) * (1+pow(e,-1*x))^(-3) + 2 * 1 * pow(e, -1*x) * (1+pow(e,-1*x))^(-2)) };
\end{axis}
\end{tikzpicture}
\caption{Left panel: Regularization function \eqref{eq:reswish} for $\alpha=1$, $\beta=1$ (solid line) and $\alpha=1$, $\beta = 50$ (dashed line). Right panel: The first (dashed line) and the second (solid line) derivatives of the regularization function \eqref{eq:reswish} for $\alpha=1$ and $\beta=1$.
}
\label{fig:foothill}
\end{figure}
\subsubsection{Properties}
The regularization function (\ref{eq:reswish}) has several interesting properties. It is infinitely differentiable and symmetric about the origin, $$ p_{\alpha,\beta}(x) = p_{\alpha,\beta}(-x).$$
Also, it is flexible enough to approximate the lasso \cite{Tibshirani_lasso_1996} and Ridge penalties \cite{hoerl1970ridge} for particular values of $\alpha$ and $\beta$. The following properties suggest that this function could be considered as a quasiconvex alternative to the elastic net penalty \cite{zou2005regularization}.
\begin{property} For $\alpha=1$ and $\beta \rightarrow \infty$, the foothill penalty (\ref{eq:reswish}) converges to the lasso penalty.
\end{property}
\emph{Proof.} For $x>0$, it is easy to see that
\begin{flalign*}
\lim_{\beta \rightarrow +\infty}& \mathrm{tanh}\left( \frac{\beta x}{2} \right) = 1,\\
\lim_{\beta \rightarrow +\infty}& p_{\alpha, \beta}(x) = x.
\end{flalign*}
Equivalently, for negative $x$, as $p_{\alpha, \beta}(x)$ is symmetric about the origin, then $\lim_{\beta \rightarrow +\infty} p_{\alpha, \beta}(x) = -x$, which is equivalent to $p_{\alpha, \beta}(x) \rightarrow |x|$ when $\beta \rightarrow +\infty$.
\begin{property} For $\alpha>0$, $\beta>0$, and $\beta = 2 / \alpha$ the foothill penalty (\ref{eq:reswish}) approximates the Ridge penalty in a given interval $[-c ; c]$.
\end{property}
\emph{Proof.} Let us study this property formally. Take the Taylor expansion of (\ref{eq:reswish}),
\begin{align}
p_{\alpha, \beta}(x) \approx \frac{\alpha\beta}{2} x^2 - \frac{\alpha\beta^3}{24} x^4 + \frac{\alpha\beta^5}{240} x^6 + \mathrm{O}(x^8).
\label{eq:taylor}
\end{align}
And, for a given $c>0$,
\begin{align}
\int_{0}^{c} \left( \frac{\alpha\beta}{2} x^2 - p_{\alpha, \beta}(x) \right)^2 dx \approx \frac{\alpha^2\beta^6}{5184} c^9 + \mathrm{O}(c^{11}).
\label{eq:approxridge}
\end{align}
The integral in (\ref{eq:approxridge}) diverges if $c$ tends to infinity, but for a finite positive number $c$, one can numerically estimate the minimal distance between the $L_2$ norm and (\ref{eq:reswish}) with a tiny approximation error. This can be achieved by taking $\beta = 2 / \alpha$ and (\ref{eq:approxridge}) becomes
\begin{align}
\int_{0}^{c} \left( x^2 - p_{\alpha, \beta}(x) \right)^2 dx \approx \frac{1}{81\alpha^4} c^9 + \mathrm{O}(c^{11}) = \varepsilon_c .
\label{eq:approxridge2}
\end{align}
For large values of $\alpha$, the error $\varepsilon_c$ is negligible, see for example Figure \ref{fig:approxridge} where the regularization function (\ref{eq:reswish}) approximates the Ridge penalty almost perfectly within $[-5 ; 5]$. Furthermore, for fixed parameters, note that
\begin{align*}
\lim_{x \rightarrow +\infty} p_{\alpha, \beta}(x) - \alpha x = 0, ~~\mathrm{and}~~
\lim_{x \rightarrow -\infty} p_{\alpha, \beta}(x) + \alpha x = 0.
\end{align*}
Hence it is also interesting to note that (\ref{eq:reswish}) behaves like a polynomial function for small values of $x$, and like a linear function for large values. Therefore, using it as a loss function (instead of a regularization), (\ref{eq:reswish}) behaves like the Huber loss used in practice for robust estimation \cite{huber1964robust}. Figure \ref{fig:approxridge} shows that (\ref{eq:reswish}) is bounded between the Huber loss and the squared error loss.
\begin{figure}[htb]
\centering
\includegraphics[scale=0.4]{figures/fig/huber_like_foothill.pdf}
\caption{Plots of the Ridge, the foothill penalty with $\alpha = 16$ and $\beta=0.125$ and twice the Huber loss. The solid blue line represents the foothill penalty, the dashed line represents the Ridge one (upper bound) and Huber (lower bound) is represented in dotted line.}
\label{fig:approxridge}
\end{figure}
\begin{property} Saddle points of $p_{\alpha, \beta}(x)$ are $x_0 \approx \pm 2.3994/\beta$ and $p_{\alpha, \beta}(x_0) = \frac{2 \alpha }{\beta}$.
\end{property}
\emph{Proof.} Indeed, the second order derivative vanishes at
\begin{align*}
2 - \beta x ~ \mathrm{tanh}\left( \frac{\beta x}{2} \right) = 0,\\
\end{align*}
which is solved by an iterative method for $\beta x \approx \pm 2.3994$. This implies
\begin{align*}
\beta x_0 ~ \mathrm{tanh}\left( \frac{\beta x_0}{2} \right) = 2,
\end{align*}
or equivalently
\begin{align*}
p_{\alpha, \beta}(x_0) = \frac{2 \alpha }{\beta}.
\end{align*}
\begin{property} The function $p_{\alpha, \beta}(x)$ is quasiconvex.
\end{property}
\emph{Proof.} It is straight forward to show that $p_{\alpha, \beta}(x)$ is decreasing from $-\infty$ to $0$ and increasing from $0$ to $+\infty$ and any monotonic function is quasiconvex (see Figure \ref{fig:foothill}, right panel).
\vspace{0.5cm}
Table \ref{tab:relation} suggests that foothill has the flexibility to be used for feature selection regularizer such as the lasso or used only to shrink the estimator in order to prevent over-fitting like the Ridge. Finally, it also can be used as a loss function for robust regression as an alternative to the Huber loss.
\begin{table}[htb]
\caption{Relationship to other functions} \label{tab:relation}
\begin{center}
\begin{tabular}{l@{\hspace{0.1in}}l@{\hspace{0.1in}}l@{\hspace{0.1in}}l}
& Shape $\alpha$ & Scale $\beta$ & Function\\
\hline
Lasso & $1$ & $+\infty$ & $p_{\alpha, \beta}(x)=|x|$\\
Ridge & $+\infty$ & $2 / \alpha$ & $p_{\alpha, \beta}(x)=x^2$\\
Huber & $< +\infty$ & $2 / \alpha$ & $p_{\alpha,\beta}(x) = \alpha x~ \mathrm{tanh}\left( \frac{x}{\alpha} \right)$\\
Foothill & $1$ & $2$ & $p_{\alpha, \beta}(x)=x \mathrm{tanh}(x)$\\
\end{tabular}
\end{center}
\end{table}
\section*{Siamese Model for Individual Treatment Effect}
{\centering
Mouloud Belbahri \& Alejandro Murua \\
Department of Mathematics and Statistics,
University of Montreal
}
\vspace{30pt}
Consider the random variables $Z_1 = YT - Y(1-T)$ and $Z_0 = (1-Y)T - (1-Y)(1-T)$.
We have
\[ Z_1 = \left\{ \begin{array}{cc} 1, & Y=1, T=1, \\ -1, & Y=1, T=0\\ 0 & \sim \end{array} \right.
\]
and
\[ Z_0 = \left\{ \begin{array}{cc} 1, & Y=0, T=1, \\ -1, & Y=0, T=0\\ 0 & \sim \end{array} \right.
\]
Now consider the random variable
\[ Z = \left\{ \begin{array}{cc} Z_1, & Y=1 \\ Z_0, & Y=0 \end{array} \right.
\]
Note that $Z= 2T -1 $.
We consider the case $P(T=1)=P(T=0) = 0.5$.
Let $m_{yt} = P( Y=y | T=t)$, and $p_{yt} = m_{yt} / ( m_{yt} + m_{y(1-t)})$.
We have
\begin{align*}
P( Z = 1 | Y=1 ) & = P( Z =1 , Y=1 ) / P(Y = 1) \\
& = \tfrac{\biggl\{P( Z=1, Y=1, T=1 ) + P(Z=1, Y=1, T=0)\biggr\}}{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_1=1 | Y=1, T=1)P(Y=1 | T=1) P(T=1) + P(Z_1=1 | Y=1, T=0)P(Y=1 | T=0) P(T=0) \biggr\}}
{\biggl\{ P(Y=1|T=1)P(T=1) + P(Y=1 | T=0) P(T=0) \biggr\} } \\
& = m_{11} / ( m_{11} + m_{10}) = p_{11}.
\end{align*}
Similarly,
\begin{align*}
P( Z = -1 | Y=1 ) & = P( Z =-1 , Y=1 ) / P(Y = 1) \\
& = \tfrac{\biggl\{P( Z=-1, Y=1, T=1 ) + P(Z=-1, Y=1, T=0)\biggr\}}{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_1=-1 | Y=1, T=1)P(Y=1 | T=1) P(T=1) + P(Z_1=-1 | Y=1, T=0)P(Y=1 | T=0) P(T=0) \biggr\}}
{\biggl\{ P(Y=1|T=1)P(T=1) + P(Y=1 | T=0) P(T=0) \biggr\}} \\
& = m_{10} / ( m_{11} + m_{10}) = 1 - p_{11}.
\end{align*}
Proceeding in the same way, we have
\begin{align*}
P( Z = 1 | Y=0 ) & = P( Z =1 , Y=0 ) / P(Y = 0) \\
& = \tfrac{\biggl\{P( Z=1, Y=0, T=1 ) + P(Z=1, Y=0, T=0)\biggr\} }{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_0=1 | Y=0, T=1)P(Y=0 | T=1) P(T=1) + P(Z_0=1 | Y=0, T=0)P(Y=0 | T=0) P(T=0) \biggr\}}
{\biggl\{ P(Y=0|T=1)P(T=1) + P(Y=0 | T=0) P(T=0) \biggr\}} \\
& = m_{01} / ( m_{01} + m_{00}) = p_{01},
\end{align*}
and
\begin{align*}
P( Z = -1 | Y=0 ) & = P( Z =-1 , Y=0 ) / P(Y = 0) \\
& = \tfrac{\biggl\{P( Z=-1, Y=0, T=1 ) + P(Z=-1, Y=0, T=0)\biggr\}}{ P(Y=1)} \\
& = \tfrac{\biggl\{P(Z_0=-1 | Y=0, T=1)P(Y=0 | T=1) P(T=1) + P(Z_0=-1 | Y=0, T=0)P(Y=0 | T=0) P(T=0) \biggr\} }
{\biggl\{ P(Y=0|T=1)P(T=1) + P(Y=0 | T=0) P(T=0) \biggr\} }\\
& = m_{00} / ( m_{01} + m_{00}) = 1 - p_{01}.
\end{align*}
Notons que $Z = 1$ iff $T=1$ and $Z=-1$ iff $T=0$. This fact will be useful to write the likelihood of the complete data.
This is given by
\begin{multline*}
\prod_{i=1}^n P( z_i, y_i, t_i) = \prod_{i=1}^n P( z_i | y_i) P( y_i | t_i) P(t_i) \\
= 2^{-n}\prod_{i=1}^n p_{11}^{t_i y_i} (1 - p_{11})^{(1-t_i)y_i} p_{01}^{t_i (1-y_i)} (1 - p_{01})^{(1-t_i)(1-y_i)}
m_{11}^{t_i y_i} (1 - m_{11})^{t_i (1-y_i)} m_{10}^{(1-t_i) y_i} (1 - m_{10})^{(1-t_i) (1-y_i)} \\
= 2^{-n}\prod_{i=1}^n p_{y_i 1}^{t_i } (1 - p_{y_i 1})^{(1-t_i)} m_{1t_i}^{y_i} (1 - m_{1t_i})^{(1-y_i)} \\
= 2^{-n} \exp\{ \sum_{i=1}^n t_i \log p_{y_i 1} + (1 - t_i)\log(1 - p_{y_i 1}) \}
\times \exp\{ \sum_{i=1}^n y_i \log m_{ 1 t_i 1} + (1 - y_i)\log(1 - m_{1 t_i }) \}.
\end{multline*}
\end{document}
\section{Universal approximation theorems}
A neural network (NN; for a definition, see e.g. \cite{pinkus1999approximation} or \cite{calin2020deep}) transforms an input $x\in D\subset\mathbb{R}^{n}$ into an output $\widetilde{y}=\Phi\left( x\right) \in\mathbb{R}$. As such it is a concrete real function of $n$ real variables, $n$ being a priori an arbitrary positive integer. A natural question that has been asked for a long time is the following: can this NN (with an appropriate architecture) approximate functions of a given class (common example: multivariate continuous functions defined on $D=K$, a compact subset of $\mathbb{R}^{n}$, but more general function spaces, such as $L^{p}$ spaces, are usually the proper framework from an abstract mathematical viewpoint) \citep{mhaskar1996system}? In other words, can a NN generate a topologically dense subset of the function class in question? For example, in the case of a normed vector space, given $\epsilon>0$, is there and, if so, can one actually construct at a reasonable cost a NN such that
\[
\left\Vert f-\widetilde{y}\right\Vert <\epsilon\text{ ?}%
\]
A related question (among many other researchers are addressing \citep{fan2020universal}) is: to what extent does the choice of the norm (topology) affect this existence and/or construction?
The proper mathematical interpretation of NN is based on the universal
approximation theorems (UAT). The UAT are roughly divided into two categories:
\begin{enumerate}
\item arbitrary number of neurons in the case of {\it shallow} neural networks (SNN, arbitrary width). This is the classical case. It goes back to the pioneering work of the 1940-1980's;
\item arbitrary number of layers with {\it deep} neural networks (DNN, arbitrary
depth). In practice, the latter have proven to be more efficient
\citep{mhaskar2019function}. It is always possible to build a SNN to emulate a DNN. However the number of neurons grows exponentially in order to obtain a
comparable performance (in terms of error) \citep{telgarsky2016benefits,eldan2016power}.
\end{enumerate}
In this sense, the UAT are representations of classical existence theorems. They implement the construction (i.e. find the appropriate weights) of the approximation of large classes of functions \citep{paluzo2020two}. However, they do not say how one explicitly (with a formula or an algorithm) finds this approximation (weights).
We give a brief summary of some of the conceptual steps that led to today's mathematical understanding of the UAT.
\begin{itemize}
\item The Kolmogorov--Arnold representation (or superposition) theorem \citep{kolmogorov1957representation} states that any continuous real multivariate function can be represented by a superposition of real one-variable continuous functions $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$:
\[
f(x) =\sum_{i=0}^{2n}\Phi_{i}\left( \sum_{j=1}^{n}\phi
_{i,j}(x_{j}) \right)
\]
This theorem gives a partial answer to Hilbert's 13th problem \citep{akashi2001application,khesin2014arnold}. A construction is proposed by \cite{polar2020urysohn}.
\item Prior to the 1980s \citep{mcculloch1943logical,fukushima1975cognitron}, NN architectures were shallow. An understanding of SNN gives a good appreciation of DNN.
\item The Kolmogorov-Arnold theorem gives an exact representation of the function $f(x)$. This representation proves to be very complex (fractal nature) and computationally very slow, therefore unusable in NN. The accuracy condition had to be relaxed for an approximation condition ($L^{p}$
or uniform norm). The optimal solution is sought by iterative methods (the
gradient descent, introduced by Jacques Hadamard in 1908, being the most
popular). Pioneering modern work is due to \cite{cybenko1989approximation} for
sigmoid activation functions (arbitrary width), \cite{funahashi1989approximate}, \cite{hornik1991approximation} (he showed that UAT was
potentially constructible using a multilayer architecture of the NN rather
than a specific choice of the activation function), \cite{barron1993universal} (bounds on smooth function errors), etc...
\item \cite{leshno1993multilayer} and \cite{pinkus1999approximation} have
shown that the universal approximation property (see \cite{kratsios2019universal} for a
definition) is equivalent to a non polynomial activation function.
\end{itemize}
\begin{theorem}
(\cite{cybenko1989approximation,hornik1991approximation,pinkus1999approximation}). Let $n\in\mathbb{N}$, $K\subset\mathbb{R}^{n}$ compact, $f:K\rightarrow\mathbb{R}$ continuous, $\rho:\mathbb{R}\rightarrow\mathbb{R}$ non polynomial and continuous, and let $\epsilon>0$. Then there exists a
two-layer NN $\Phi$ with activation function $\rho$ such that
\[
\left\Vert \Phi-f\right\Vert _{\infty}<\epsilon.
\]
\end{theorem}
We also note that DNN have demonstrated their capabilities in the approximation of functions \citep{liang2016deep,yarotsky2017error,fan2020universal}. They are found to have good experimental (empirical) performance, but it is largely unknown why. This is a very active area of research \cite{caterini2018deep}.
\begin{theorem}
(\cite{lu2017expressive}) Let $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$ be Lebesgue-integrable and $\epsilon>0$. Then there exists a ReLU
network of width $\leq n+4$ such that
\[
\int_{\mathbb{R}^{n}}\left\vert \Phi\left( x\right) -f\left( x\right) \right\vert dx<\epsilon
\]
\end{theorem}
\begin{definition}
(\cite{kidger2020universal}) Let $\rho:\mathbb{R}\rightarrow\mathbb{R}$ and $n,m,k\in\mathbb{N}$. Then let $NN_{n,m,k}$ represent the class of functions $\mathbb{R}^{n}\rightarrow\mathbb{R}^{m}$ described by feedforward neural networks with $n$ neurons in the input layer, $m$ neurons in the output layer, and an arbitrary number of hidden layers, each with $k$ neurons with activation function $\rho$. Every neuron in the output layer has the identity activation function.
\end{definition}
\begin{theorem}
(\cite{kidger2020universal}) Let $\rho:\mathbb{R}\rightarrow\mathbb{R}$ be any non-affine continuous function which is continuously differentiable at
at least one point, with non-zero derivative at that point. Let $K\subset\mathbb{R}^{n}$ be compact. Then NN$_{n,m,n+m+2}$ is dense in $C\left( K,\mathbb{R}^{m}\right)$ with respect to the uniform norm.
\end{theorem}
\section{Low-rank approximation}
Due to its fundamental role in Machine Learning, a huge literature is devoted
to the singular value decomposition (SVD) of a matrix. We refer to some of the
well respected texts: \cite{strang1993introduction}, \cite{trefethen1997numerical}, \cite{watkins2004fundamentals}, \cite{gentle2007matrix}, \cite{elden2007matrix}, \cite{lange2010numerical},
and \cite{yanai2011projection} for
instance. Let $A\in \mathbb{R}^{m\times n}$ be a matrix of rank $r$. Its SVD consists in factoring $A$ as
\[
A=U\Sigma V^\top=\sum_{i=1}^{r}\sigma_{i}\mathbf u_{i}\mathbf v_{i}^\top
\]
where $U\in\mathbb{R}^{m\times m}$ and $V\in\mathbb{R}^{n\times n}$ are orthogonal square matrices, and $\Sigma\in\mathbb{R}^{m\times n}$ (\textquotedblleft diagonal\textquotedblright\ rectangular matrix of the ordered singular values: $\sigma_{1}\geq\ldots\geq\sigma _{r}>\sigma_{r+1}=\ldots=0=\sigma_{\min\{m,n\}}$). This
decomposition, and the related principal component analysis (PCA), has several
uses in Data Analysis. Written as $AV=U\Sigma$, SVD says that there exist orthogonal matrices $U$ and $V$ such that $A$ maps the columns $\mathbf v_{i}$ of $V$ (right singular vectors of $A$, directions or principal axes in the PCA parlance) into the columns $\mathbf u_{i}$ of $U$ (left singular vectors of $A$) such as
\[
A\mathbf v_{i}=\sigma_{i}\mathbf u_{i},
\]
the principal components.
Given a positive integer $1\leq k\leq r$, introduce the truncated matrices
$U_{k}=\left[ \mathbf u_{1},\ldots,\mathbf u_{k}\right] \in \mathbb{R}^{m\times k}$, $V_{k}=\left[ \mathbf v_{1},\ldots,\mathbf v_{k}\right] \in \mathbb{R}^{k\times n}$, and $\Sigma_{k}=\mathrm{diag}( \sigma_{1},\ldots,\sigma_{k}) \in \mathbb{R}^{k\times k}$. Then the matrix of rank $k$
\[
A_{k}=U_{k}\Sigma_{k}V_{k}^\top=\sum_{i=1}^{k}\sigma_{i}\mathbf u_{i}\mathbf v_{i}^\top
\]
is an optimal lower-rank approximation of $A$ in the sense of both the
Frobenius norm and the $L_2$-norm. Indeed, the Eckart--Young--Mirsky theorem (for
a proof, see \cite{markovsky2012low}) states that among all matrices $B\in\mathbb{R}^{m\times n}$ of a given rank $k\leq r$, the one closest to $A$ is
$\hat{A}=A_{k}$, i.e.
\begin{align*}
\inf_{\mathrm{rank}(B)=k}\left\Vert A-B\right\Vert _{F} & =\left\Vert
A-\hat{A}\right\Vert _{F}=\sqrt{\sum_{i=k+1}^{r}\sigma_{i}^{2}},\\
\inf_{\mathrm{rank}(B)=k}\left\Vert A-B\right\Vert _{2} & =\left\Vert
A-\hat{A}\right\Vert _{2}=\sigma_{k+1}.
\end{align*}
Furthermore, the approximating matrix $\hat{A}$ is unique if and only if
$\sigma_{k+1}\neq\sigma_{k}$ \citep{markovsky2012low}. The theorem was proved in 1907 by Schmidt (of the Gram-Schmidt procedure) for the Frobenius norm \citep{leon2013gram}. A constructive solution was given by \cite{eckart1936approximation}. This solution was
shown by \cite{mirsky1960symmetric} to hold for all unitarily invariant norms. A further generalization addressing collinearity issues was proposed by \cite{golub1987generalization}. See \cite{stewart1993early} for historical details.
One interpretation of the decomposition $A_{k}=U_{k}\Sigma_{k}V_{k}^{T}$ is as
a compression/learning linear procedure of the data matrix $A$ (in other
words, a dimension reduction method). Indeed, since the SVD extracts data in
the direction with the highest variance (and going down to those with the
lowest variance), matrix $U_{k}$ is the compression matrix in the first $k$
directions, matrix $\Sigma_{k}$ contains the amount of error (cost function)
we tolerate, and $V_{k}$ is the decompression matrix. If we keep all the non
zero singular values ($k=r$), there is theoretically no loss in the
compression (assuming that the data in $A$ are exact). If we drop the small
singular values (set them to $0$), it is not possible to reconstruct the
initial matrix $A$ exactly, but we obtain a matrix $A_{k}$ that is close. In
other words, the SVD decomposition is a way to control the loss of information
from $A$ under a predetermined level.
The assumption (routinely validated in image processing, for example) in the
above approach is that the components (columns of $U$) of the small singular
values contain little information (mostly \textquotedblleft
noise\textquotedblright) and should be discarded. However, the low-rank matrix approximation is easily seen not to be optimal if used to solve the linear regression problem (hence as a predictive tool) $y=A\beta+\epsilon$, that is to minimize
\begin{equation}
\inf_{\mathrm{rank}(B)=k}\left\Vert \mathbf y-B\boldsymbol \beta\right\Vert .
\label{eq:svd_regression}
\end{equation}
While the SVD picks the parameter $k-$vector $\boldsymbol \beta$ of minimum norm
\citep{elden2007matrix,lange2010numerical},
\begin{equation}
\hat{\boldsymbol \beta}=A_{k}^{\dag} \mathbf y=\sum_{i=1}^{k}\frac{1}{\sigma_{i}} \mathbf u_{i}^\top \mathbf y \mathbf v_{i},
\label{eq:svd_reg_solution}
\end{equation}
where $A_{k}^{\dag}$ is the pseudoinverse of $A_{k}$, the general solution of
\eqref{eq:svd_regression} is unknown \citep{xiang2012optimal}. See also \cite{talwalkar2010matrix} for a discussion in the Machine Learning context.
Some authors \citep{jolliffe1982note,hadi1998some} warned against the misconceived (mainly due to collinearity considerations) use of \eqref{eq:svd_reg_solution} in an ordinary linear regression setting. Dropping some of the small singular values may result in ignoring significant variables, mainly because the approximation does not take into account the response variable $\mathbf y$. A partial solution is provided by the ridge regression \citep{hoerl1970ridge}. However, still from a predictive viewpoint, SVD may prove to be an interesting approach in deep neural networks \citep{bermeitinger2019singular}. Used for initialization in the first layer, SVD gave promising experimental results. Theoretical issues still need to be addressed.
\newpage
\subsection{Low-rank and SMITE}
\textit{Comment from Alejandro Murua: We do not necessarily study low rank approximations of the matrix of weights. For this you will need indeed to get bounds on the distance between the estimated function and the approximation function. In Statistics, usually we just measure the statistical difference between two models (and not the distance between the two models). So instead of thinking of a reduced approximation of the weight matrix, we might just think of a model with a lower-rank matrix of weights. This in general will lead to a non-parametric estimator with lower variance and hopefully, only slightly increased bias. In this setup, methods to measure the statistical difference between the full model and the reduced model should be applied for each dataset. This is in contrast to the general approximation theory, where bounds on the approximation error over all possible datasets need to be found.}
We are interested in determining how complex the network ought to be to theoretically guarantee that the network would approximate an unknown target function $f$ up to a given accuracy $\epsilon > 0$. To measure the accuracy, we need a norm $\norm{\cdot}$ on some normed linear space. We work with $L_2$ norm.
We focus on a feed-forward neural network with one fully-connected hidden layer. Here the input layer (with $n$ neurons) and the hidden layer (with $m$ neurons) are assumed to be fully connected with a weight matrix $\mathbf W \in \rm I\!R^{n \times m}$ with rank at most $p \leq \mathrm{min}(n,m)$. The domain for the input vector $\mathbf x$ is the $n$-dimensional hypercube $I^n = [0, 1]^n$, and the output layer only contains one neuron. The neural network can be expressed as
$$\hat y = \hat f(\mathbf x\mid \mathbf W) = \sum_{j=1}^m \alpha_j h(\mathbf w_j^\top\mathbf x + \beta_j) = \boldsymbol{\alpha} h(\mathbf x \mathbf{W} + \boldsymbol{\beta}).$$
Here $h(\cdot)$ is the activation function, $\mathbf w_j \in \rm I\!R^n$ denotes the $j$-th column of the weight matrix $\mathbf W$, and $\alpha_j, \beta_j \in \rm I\!R$ for $j = 1, \ldots, m$. Suppose $p$ is the rank of the matrix $\mathbf W$. We can denote it $\mathbf W_p$ and define the following trained model
$$\hat y_p = \boldsymbol{\alpha} h(\mathbf x \mathbf{W_p} + \boldsymbol{\beta}).$$
Let $r < p$ be an integer and define
$$\hat y_r = \boldsymbol{\alpha} h(\mathbf x \mathbf{\hat W_r} + \boldsymbol{\beta}),$$
be the prediction from the model where $\mathbf{W_p}$ has been approximated with $\mathbf{\hat W_r}$ of rank $r$, that is, $\mathbf{\hat W_r} = \mathbf{\hat U \hat \Sigma_r \hat V^{\top}}$ where $\mathbf{\hat{U}, \hat{\Sigma_r}, \hat{V}^\top} \in \rm I\!R^{n\times r}, \rm I\!R^{r\times r}, \rm I\!R^{r \times m}$. This approximates the original weights $\mathbf W_p$ into reconstructed approximated weights $\mathbf{\hat{W_r}}$.
The truncation process always keeps the first $r$ rows of $\mathbf U$ and columns of $\mathbf V^\top$. The number of non-zero parameters in the truncated SVD is $r( n + m + 1)$. The compression ratio is the ratio between parameters after and before the decomposition which is calculated as $r( n + m + 1)/( nm )$. In addition, the weight matrix does not need to be reconstructed during inference where the input $\mathbf x$ can be directly multiplied with $\mathbf{\hat{U}, \hat{\Sigma_r}, \hat{V}^\top}$ matrices, which takes lesser computation than direct multiplication of $\hat{\mathbf W_r}$ and $\mathbf x$. Therefore, we have the following development
\begin{align*}
\norm{ \hat y_p - \hat y_r }_2 &= \norm{\boldsymbol{\alpha} h(\mathbf x \mathbf{W_p} + \boldsymbol{\beta}) - \boldsymbol{\alpha} h(\mathbf x \mathbf{\hat W_r} + \boldsymbol{\beta})}_2 \\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{h(\mathbf x \mathbf{W_p} + \boldsymbol{\beta}) - h(\mathbf x \mathbf{\hat W_r} + \boldsymbol{\beta})}_2\\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{\mathbf x \mathbf{W_p} - \mathbf x \mathbf{\hat W_r}}_2 ~ \text{because $h(\cdot)$ is $1$-Lipschitz continuous}\\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{\mathbf x}_2 \norm{\mathbf{W_p} - \mathbf{\hat W_r}}_2\\
&\leq \norm{\boldsymbol{\alpha}}_2 \norm{\mathbf x}_2 \sigma_{r+1} ~ \text{by Eckart-Young Theorem \citep{eckart1936approximation}}\\
&\leq \Big( \sum_{j=1}^m \alpha_j ^ 2 \Big)^{1/2} \sqrt{n} \sigma_{r+1}
\end{align*}
where $\sigma_{r+1}$ is the $(r+1)$th singular value of $\mathbf{W_p}$.
\subsection*{Random Forests and IE loss}
\begin{table}[H]
\centering
\begin{tabular}{lrrrrrrrrrr}
\toprule
max\_depth & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\
n\_estimators & & & & & & & & & & \\
\midrule
10 & 0.0000 & 0.0000 & 0.4627 & 0.0000 & 0.0000 & 0.0847 & 0.6873 & 0.1620 & 0.6227 & 0.8690 \\
60 & 0.0000 & 0.6076 & 0.2521 & 0.3716 & 0.7808 & 0.6221 & 0.9801 & 0.4200 & \textcolor{blue}{0.6480} & 0.6447 \\
110 & 0.2358 & 0.2152 & 0.4802 & 0.2579 & 0.6272 & 0.5698 & 0.7611 & 0.8179 & 0.9595 & 0.6621 \\
160 & 0.7447 & 0.1843 & 0.2604 & 0.1395 & 0.6537 & 0.7644 & 0.7623 & 0.8682 & 0.6104 & 0.5373 \\
210 & 0.6333 & 0.1896 & 0.3153 & 0.3784 & 0.5498 & 0.9748 & 0.5699 & 0.8745 & \bf 1.1953 & 0.6199 \\
\bottomrule
\end{tabular}
\caption{$\hat{q}_{\mathrm{adj}}$}
\label{tab:my_label}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{lrrrrrrrrrr}
\toprule
max\_depth & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\
n\_estimators & & & & & & & & & & \\
\midrule
10 & 0.91698 & 0.89777 & 0.87211 & 0.88674 & 0.87344 & 0.86945 & 0.86670 & 0.86730 & 0.86579 & 0.86359 \\
60 & 0.91405 & 0.89896 & 0.88380 & 0.88023 & 0.87313 & 0.86804 & 0.86168 & 0.86205 & \bf 0.85797 & 0.85802 \\
110 & 0.91569 & 0.89789 & 0.88678 & 0.88219 & 0.87444 & 0.87028 & 0.86275 & 0.86258 & 0.86004 & 0.86072 \\
160 & 0.91891 & 0.89967 & 0.88847 & 0.88106 & 0.87363 & 0.87015 & 0.86320 & 0.86221 & 0.86109 & 0.86088 \\
210 & 0.91967 & 0.89970 & 0.88769 & 0.87966 & 0.87184 & 0.86772 & 0.86179 & 0.86061 & \textcolor{blue}{0.85979} & 0.85980 \\
\bottomrule
\end{tabular}
\caption{IE Loss}
\label{tab:my_label}
\end{table}
\section{Writing}
The top priority lies in writing the text. The methodology has already been developed and literature review is complete. Therefore, we just need to follow the detailed plan to tell the story of SMITE.
\subsection{Introduction/Conclusion/Abstract}
In the introduction section, we need to answer the following questions :
\begin{enumerate}
\item What is the problem ? \textcolor{red}{DONE}
\item Why is it important ? \textcolor{blue}{Almost done.}
\item How others propose to solve the problem ? \textcolor{blue}{Almost done.}
\item What are the shortages of existing methods ? \textcolor{blue}{Almost done.}
\item What do we propose to solve these shortages ? \textcolor{blue}{Almost done.}
\end{enumerate}
The introduction, the conclusion and the abstract must be written last. We must clearly cite our contributions in the introduction and the benefits of using SMITE.
\subsection{Uplift modeling}
Here we define the uplift framework more clearly and discuss existing methods. This section is already written and is composed of the following three subsections:
\begin{enumerate}
\item Prerequisites \textcolor{red}{DONE}
\item Uplift and machine learning \textcolor{red}{DONE}
\item The interaction model \textcolor{blue}{Almost done. We need to cite and describe the Qini-based paper here.}
\end{enumerate}
\subsection{Siamese uplift model}
In this section, we introduce our methodology in the context of the interaction model that we want to represent as a Siamese model (see Figure~\ref{fig:siamese_network}). We start by giving some references on Siamese networks. \textcolor{red}{DONE}
\begin{figure}[H]
\centering
\input{figures/tikzsmite}
\caption{Siamese representation of the uplift interaction model.}
\label{fig:siamese_network}
\end{figure}
The model can be written as follows:
\begin{align*}
&\mu_{t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta}) = \mathrm{Pr} (Y_i = 1 \mid \mathbf{x}_i, t_i, \theta_o, \boldsymbol{\theta}) =
\bigl( 1+ \mathrm{exp}\{-(\theta_o + \gamma t_i + \mathbf{x}_i^\top [\boldsymbol{\beta} + t_i \boldsymbol{\delta}] ) \} \bigr)^{-1},\\
&u(\mathbf x_i) = \mu_1(\mathbf x_i) - \mu_0(\mathbf x_i)
\end{align*}
where $\boldsymbol{\theta} = (\boldsymbol{\beta}, \gamma, \boldsymbol{\delta})$, denotes all model parameters except for the intercept $\theta_o$.
Some parts of the following subsections need work:
\begin{enumerate}
\item The transformed outcome \textcolor{blue}{Almost done. Introduced the different transformed outcomes but need to elaborate a bit more on $Z$ and why MSE is not useful for binary $Y$.}
\begin{enumerate}
\item[a.1)] Regularized conditional means regression \textcolor{blue}{Almost done. We need to justify the exponential loss more clearly for SMITE TO.}
\item[a.2)] The optimization algorithm \textcolor{red}{DONE. Elementary explanation of SGD and related references.}
\end{enumerate}
\item Exploring other regularization functions \textcolor{blue}{This part needs work. We need Alejandro's input on the justification of SMITE IE and we can also discuss other types of losses.} \textcolor{red}{Check Alejandro's development of the IE likelihood.}
\end{enumerate}
\subsection{SMITE neural network}
\input{figures/nnsmite.tex}
In this section, we generalize to the neural networks version of SMITE. \textcolor{blue}{We will need Alejandro's input on what is next. Based on discussions with Vahid, this section will be arranged as follows:}
\begin{enumerate}
\item Universal approximation theorems \textcolor{red}{DONE}
\item Low-rank approximation \textcolor{red}{DONE}
\item Low-rank SMITE \textcolor{blue}{TO DO. This part can be related to model selection ! Needs more discussions.}
\end{enumerate}
The first two subsections are brief literature reviews on the subject. The last one needs a clear connection with SMITE. Here are Alejandro's comments on the subject:
\textit{We do not necessarily study low rank approximations of the matrix of weights. For this you will need indeed to get bounds on the distance between the estimated function and the approximation function. In Statistics, usually we just measure the statistical difference between two models (and not the distance between the two models). So instead of thinking of a reduced approximation of the weight matrix, we might just think of a model with a lower-rank matrix of weights. This in general will lead to a non-parametric estimator with lower variance and hopefully, only slightly increased bias. In this setup, methods to measure the statistical difference between the full model and the reduced model should be applied for each dataset. This is in contrast to the general approximation theory, where bounds on the approximation error over all possible datasets need to be found.}
\subsection{Evaluating uplift models}
This section introduces the Qini measures. I believe we should also define the ITE related metrics such as $\mu$-risk, $\tau$-risk and $v$ that stands for {\it value} which is somehow related to the gain chart sometimes used in marketing. \textcolor{blue}{TO DO}
\textcolor{red}{On Kendall and Qini coefficients (and associated curves), we need to discuss and develop some geometrical intuition behind these statistics. We can present these with a different way of thinking.}
\subsection{Experiments}
We already have results but I think we need to discuss which experiments we want to run in order to be able to improve this part of the work. \textcolor{blue}{TO DO}
\section{Coding/Experiments}
All codes are ready and clean. We only need to run the necessary experiments and list the Figures we want to show in the paper. \textcolor{blue}{TO DO}
\section{Data}
The problem with public datasets used in ITE literature is that they are observational. We need to decide on which datasets we work. I think that we could also use the dataset used in the Qini-based paper to be able to compare the results without re-running all the experiments. \textcolor{blue}{TO DO}
\end{document}
\section{Introduction}
\label{sec:intro}
Uplift modeling is an important research area in marketing field \citep{radcliffe1999differential, hansotia2002direct, lo2002true, radcliffe2007using} and refers to modeling customers behavioral change that results directly from a specific treatment, such as a marketing intervention, a courtesy call, targeted advertisement. Typically, uplift models are developed for randomized controlled trials, where both the treatment and outcome are binary random variables.
Customers are segmented into different groups depending on their potential outcome \citep{kane2014mining}: \textit{persuadable}, \textit{sure thing}, \textit{lost cause}, \textit{do-no-disturb} client. A persuadable is an individual who purchases the product only if he receives a call; a sure thing client is an individual who purchases the product whether he receives the call or not; a lost cause client is an individual who will not purchase the product regardless of whether he receives the call or not; a do-no-disturb client is an individual who will not purchase the product if he receives a call. In general, from marketing standpoint interesting customers are the \textit{persuadable} and the \textit{do-not-disturb} clients. The persuadable clients provide incremental sales whereas the do-not-disturb individuals respond negatively to the marketing campaign.
Uplift fits within the causal modeling framework \citep{rubin1974estimating, holland1986statistics}. Let $Y$ be the binary response variable, $T$ be the binary treatment indicator, and $X_1,\ldots,X_p$ be the predictors. Suppose that $n$ independent iid samples are observed $\{(y_i, \textbf{x}_i, t_i)\}_{i=1}^n$, where $\mathbf x_i = (x_{i1}, \ldots, x_{ip})$ are realisations of the predictors. The uplift model estimates
\begin{equation*}
u (\textbf{x}_i)= m_{11}(\mathbf x_i) - m_{10}(\mathbf x_i),
\end{equation*}
where the notation $m_{yt}(\mathbf x) = \mathrm{Pr}(Y = y \mid \mathbf x ,T=t)$ stands for the corresponding conditional probability. This difference in conditional probabilities is named \emph{individual treatment effect,} \emph{conditional average treatment effect} in the context of causal inference, or \emph{uplift} in the context of marketing research. There are different methods to estimate individual treatment effects from randomized data. The process of estimating these effects is alternatively referred to as heterogeneous effect modeling, individual treatment effect modeling, or uplift modeling.
Since the response variable is binary, the simplest parametric model that can be used is a logistic regression. Thus, the underlying optimization problem becomes the maximization of the Binomial likelihood. On the other hand, maximizing the likelihood is not necessarily appropriate to estimate the uplift \citep{belba2019qbased}. So, very quickly, research in the uplift field turned to non-parametric models. Indeed, most of the existing methods are based on classification and regression trees \citep{breiman1984classification}. The idea is to predict the individual uplift by the uplift observed in a terminal node, or by the average when several trees are used, e.g. random forests \citep{Breiman-RandomForest-2001}. Various modifications of the split criteria have been proposed in the literature \citep{radcliffe1999differential, hansotia2002incremental, hansotia2002direct, jaskowski2012uplift, radcliffe2011real}. Thus, there is not a lot of research on defining new loss functions.
However, there are several advantages to defining an appropriate loss function. First, this would allow to re-explore the use of simple parametric models for which parameter estimation becomes appropriate for uplift. This gives a considerable advantage over non-parametric methods, since interpretation becomes simpler. This is not negligible in practice since the deployment of a model that can be interpreted minimizes the risks that a business can take. Then, when prediction becomes more important, it suffices to define a more complex model. Having a well-defined loss function allows any off the shelf method to be used for the related optimization problem. One way to have a more complex model based on a method known for its prediction power is to turn to the use of a neural network. Indeed, when working within the neural networks framework, it is possible to define a simple model (i.e. no hidden layer), but it is also possible to model statistical interactions between predictors easily \citep{tsang2018neural}, or to model polynomial functions \citep{andoni2014learning} while maintaining a very competitive prediction power.
For these reasons, in this work we focus on defining new loss functions for uplift. We develop our methodology for neural networks because of prediction power being the most important issue for uplift in practice and that it generalizes the logistic regression for which interpretation of the odds ratios is well-known. However, using our losses, our procedure can be easily generalized to other models. Empirically, we show that our proposed optimization has the impact to improve prediction performance of uplift models. Furthermore, the flexibility of our proposed method opens a door to connections with the estimation of uplift in observational data settings. We evaluate the performance of the proposed loss functions and models on both synthetic datasets and a real-world dataset made available to us by a Canadian insurance company. The experimental results demonstrate significant improvement over state-of-the-art tree-based uplift modeling.
\section{Uplift modeling}
\label{sec:literature}
Uplift is the difference between the two conditional means $m_{11}(\mathbf x)$ and $m_{10}(\mathbf x)$. The intuitive approach to model uplift is to build two classification models \citep{hansotia2002direct,snowden2011implementation,austin2012using}. This consists of fitting two separate conditional probability models: one model for the treated individuals, and another separate model for the untreated individuals. Then, uplift is the difference between these two conditional probability models. These models are called T-learners (T for ``two models") in the literature \citep{kunzel2019metalearners}. The asset of T-learners is their simplicity, but they do not perform well in practice, because each model focuses on predicting only a one-class probability and for each model information about the other treatment is never provided to the learning algorithm \citep{radcliffe2011real}. In addition, differences between the covariates distributions in the two treatment groups can lead to bias in treatment effect estimation. There has been efforts in correcting such drawbacks through a combined classification model known as S-learners, for ``single-model" \citep{kunzel2019metalearners}). The idea behind the S-learner is to use the treatment variable as a feature and to add explicit interaction terms between each covariate and the treatment indicator to fit a model, e.g. a logistic regression \citep{lo2002true}. The parameters of the interaction terms measure the additional effect of each covariate due to treatment.
\subsection{General assumptions}
Following the potential outcomes framework \citep{holland1986statistics} we make the following standard assumptions. The binary variable $T$ indicates if a unit is exposed to treatment ($T=1$) or control ($T=0$). Let $Y_0$ and $Y_1$ be the potential outcomes under control and treatment respectively. Although each customer $i$ is associated with two potential outcomes, only one of them can be realized as the observed outcome $y_i$.
{\it Assumption 1.} (Unconfoundedness) Potential outcomes $(Y_0,Y_1)$ are independent ($\independent$) of the treatment assignment indicator $T$ conditioned on all pre-treatment characteristics $(X_1, \ldots, X_p)$, i.e.,
$$\big( Y_0, Y_1 \big) \independent T | X_1,\ldots,X_p. $$
This allows the unbiased estimation of the uplift. Define the \textit{propensity score} $e(\mathbf x)= \mathrm{Pr}(T=1 \mid \mathbf x)$. This parameter is widely used to estimate treatment effects from observational data \citep{rubin1974estimating,rosenbaum1983central,athey2015machine}.
{\it Assumption 2.} (Overlap) For any $\mathbf x$, the true propensity score is strictly between $0$ and $1$, i.e.,
$$0 < e(\mathbf x) < 1.$$
In the context of complete randomization, $e(\mathbf x)$ is a constant.
{\it Assumption 3.} (Consistency) Observed outcome $Y$ is represented using the potential outcomes and treatment assignment indicator as follows,
$$Y=TY_1 + (1-T)Y_0.$$
An uplift model estimates the conditional average treatment effect in various sub-populations as a function of the possible values of the covariates $\mathbf x$,
\begin{align}
u(\mathbf x) &= \mathbb{E} ( Y_1 \mid \mathbf x) - \mathbb{E} ( Y_0 \mid \mathbf x)
=m_{11}(\mathbf x) - m_{10}(\mathbf x).
\label{eq:uplift}
\end{align}
\subsection{Uplift and machine learning}
Several machine learning methods try to model the uplift directly. Some $k$-nearest neighbours \citep{cover1967nearest} based methods are adopted for uplift estimation \citep{crump2008nonparametric, Alemi.etal-PersonalizedMedicine-2009, su2012facilitating}. The main idea is to estimate the uplift for an observation based on its neighbourhood containing at least one treated and one control observations. However, these methods quickly become computationally expensive for large datasets. On the contrary, adaptive tree-based methods such as random forests outperform the nearest neighbours approach, especially in the presence of irrelevant covariates \citep{wager2018estimation}.
State-of-the-art proposed methods view the forests as an adaptive neighborhood metric, and estimate the treatment effect at the leaf node \citep{su2009subgroup,chipman2010bart}. Therefore, most active research in uplift modeling is in the direction of classification and regression trees \citep{breiman1984classification} where the majority are modified random forests \citep{Breiman-RandomForest-2001}.
In \cite{radcliffe1999differential, radcliffe2011real,hansotia2002direct, hansotia2002incremental,jaskowski2012uplift}, modified split criteria that suited the uplift purpose were studied. The criteria used for choosing each split during the growth of the uplift trees is based on maximization of the difference in uplifts between the two child nodes. However, in practice, these approaches are prone to overfitting. The splitting criteria introduced are based solely on maximizing differences in uplifts. Therefore, splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. Within each leaf, uplift is estimated with the difference between the two conditional means. The belief is that since the uplift is defined as a difference, if we can get good estimates of each mean, we have a good estimate of the difference. However, the existing tree-based uplift optimization problems do not take into account the estimation of each mean. Instead, the focus is on maximizing the heterogeneity in treatment effects. Without careful regularization (e.g. honest estimation \citep{athey2019generalized}), splits are likely to be placed next to extreme values because outliers of any treatment can influence the choice of a split point. In addition, successive splits tend to group together similar extreme values, introducing more variance in the prediction of uplift \citep{zhao2017practically}. Alternatively some models use the transformed outcome \citep{athey2015machine}, an unbiased estimator of the uplift. However, this estimate suffers from higher variance than the conditional mean estimator \citep{powers2018some}. In addition, for both estimators, if random noise is larger than the treatment effect, the model will more likely predict random noise instead of uplift. This is particularly problematic for uplift trees, as the split criteria is function of the two treatment effects. As a result, based on several experiments on real data, and although the literature suggests that tree-based methods are state-of-the-art for uplift \citep{soltys2015ensemble}, the published models over-fit the training data and predicting uplift still lacks satisfactory solutions.
\section{Methodology}
Let's start with the uplift interaction model \citep{lo2002true} as a simple preliminary model. This model is based on logistic regression. Given a $p$-dimensional predictor vector $\textbf{x}_i$, $i \in \{1,\ldots,n\},$ {\color{magenta} [Make sure $\theta_o$ is a good notation. I prefer $\theta_0$.]} logistic intercept $\theta_o \in \rm I\!R$, and logistic regression coefficients $\boldsymbol \beta \in \rm I\!R^p$,
the model is
\begin{equation*}
\mathrm{Pr}(Y_i=1 \mid \textbf{x}_i, \theta_o, \boldsymbol \beta) = \bigl(1 + \mathrm{exp}\{- (\theta_o + \textbf{x}_i^{\top} \boldsymbol \beta) \} \bigr)^{-1}.
\end{equation*}
Throughout the paper, the superscript $^\top$ stands for the transpose of a column vector or matrix. In the uplift context, one needs to add explicit interaction terms between each explanatory variable and the treatment indicator {\color{magenta} [Why? explain more and maybe spend some sentences to explain and elaborate the reason here why interaction model is better than without interaction other than flexibility]}. Let $\gamma$ denote the treatment effect, $\boldsymbol{\beta}$, the vector of main effects, $\boldsymbol{\delta}$, the vector of interactions effects, and $\theta_o$, the intercept.
The model is
\begin{equation}
\mu_{1t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta}) = \mathrm{Pr} (Y_i = 1 \mid \mathbf{x}_i, t_i, \theta_o, \boldsymbol{\theta}) =
\bigl( 1+ \mathrm{exp}\{-(\theta_o + \gamma t_i + \mathbf{x}_i^\top [\boldsymbol{\beta} + t_i \boldsymbol{\delta}] ) \} \bigr)^{-1},
\label{eq:pitheta}
\end{equation}
where $\boldsymbol{\theta} = (\boldsymbol{\beta}, \gamma, \boldsymbol{\delta})$, denotes all model parameters except for the intercept $\theta_o$. The likelihood function associated with the interaction uplift model is
\begin{equation}
\mathcal{L}(\theta_o, \boldsymbol{\theta}) = \prod_{i=1}^{n} \mu_{1t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta})^{y_i} \{1-\mu_{1t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta})\}^{(1-y_i)},
\label{eq:uplift_likelihood}
\end{equation}
where $\{y_i : i=1,\ldots,n\}$ are the observed response variables.
The maximum likelihood estimates of $(\theta_o, \boldsymbol{\theta})$ will be denoted by $(\hat{\theta}_o, \hat{\boldsymbol{\theta}}),$
with $\hat{\boldsymbol{\theta}} = (\boldsymbol{\hat{\beta}}, \hat{\gamma}, \boldsymbol{\hat{\delta}})$.
The predicted uplift associated with the covariates vector $\mathbf{x}_{n+1}$ of a future individual is estimated by
$$ \hat{u}(\mathbf{x}_{n+1}) = \mu_{11}(\mathbf x_{n+1}, \hat{\theta}_o, \hat{\boldsymbol{\theta}}) - \mu_{10}(\mathbf x_{n+1}, \hat{\theta}_o, \hat{\boldsymbol{\theta}}).$$
The interaction model is easy to fit using standard statistical packages. In addition, the variance in comparison with T-learners {\color{magenta} [I recommend to replace T-learner with ``twin learner"]} is reduced, and the homogeneity of observations with similar properties is kept irrespective of whether they have been treated or not {\color{magenta} [It is not clear what you wanted to say here]}. However, there is still some disadvantage: the solution is not optimized with respect to finding maximally different groups. Instead, the parameters are estimated with respect to the likelihood. The methodology introduced in \cite{belba2019qbased} attacks parameter estimation framework in uplift regression. This methodology is based on uplift goodness-of-fit measures and imposing sparsity using the lasso of \cite{Tibshirani_lasso_1996}. {\color{magenta} [checked, writing is acceptable.]}
The uplift goodness-of-fit measure is a difficult statistic to compute (see Section~\ref{sec:qini} for the definition), maximizing it directly is not an easy task. Therefore, \cite{belba2019qbased} first apply the pathwise coordinate descent algorithm \citep{friedman2007pathwise} to the uplift model to compute a sequence of critical regularization values $\lambda_1 < \cdots < \lambda_{\min\{n,p'\}}$ and corresponding (sparse) model parameters $\{ (\hat \theta_o(\lambda_j), \boldsymbol{\hat\beta}(\lambda_j), \hat\gamma(\lambda_j), \boldsymbol{\hat\delta}(\lambda_j)) \}_{j=1}^{\min\{n,p'\}}$ of dimension $p'$, with $p'=2p+1$. Then, using a derivative-free optimization algorithm to explore the parameters space in order to find the maximum uplift. Empirical results show that estimating the regression parameters by maximizing the uplift goodness-of-fit improves the uplift performance.
We generalize this idea further. However, instead of maximizing the goodness-of-fit measure, we would like our method to be completely driven by gradient of the objective function {\color{magenta} [ Why?]}. To do so, we propose two alternatives to the likelihood function \eqref{eq:uplift_likelihood}. Our method is easy to implement and both objective functions have clear interpretations. {\color{magenta} [This is a very important paragraph. You must sell your method here and attract the attention of the reader to your method. Sell your method as a businessman here. I think you argument remain weak and you need to make it clear why somebody must prefer your method.]}
\subsection{The transformed outcome}
\label{sec:transformed_outcome}
The transformed outcome variable $Z_1$, which was proposed by \citep{athey2015machine} is defined as
\begin{equation}
Z_1 = \frac{T Y}{e(\mathbf x)} - \frac{(1 - T) Y}{1 - e(\mathbf x)}.
\label{eq:transformed_outcome}
\end{equation}
This transformation has the key property that, under the complete randomization or unconfoundedness assumption, its expectation conditional on $(\mathbf X=\mathbf x)$ is equal to $u(\mathbf x)$. Therefore, the transformed outcome is an unbiased estimate of the uplift. In our case, with $e(\mathbf x) = 1/2$, $Z_1$ simplifies to
\[ Z_1 = \left\{ \begin{array}{cc}
2 & \mathrm{if}~T=1, Y=1, \\
-2 & \mathrm{if}~T=0, Y=1, \\
0 & \mathrm{otherwise.}
\end{array} \right.
\]
Observing $Z_1=2$ and $Z_1=-2$ involve an interesting interpretation in the uplift context. For instance $Z=2$ for happens when a client is called ($T = 1$) and then buys the insurance policy ($Y = 1$). Thus, for its counterfactual (i.e., $T = 0$), there are two possibilities i) either the client bought the policy ($Y = 1$) and in this case it belongs to the \textit{sure things} group, or ii) the client did not buy the policy ($Y = 0$) so belongs to \textit{persuadable} group {\color{magenta} [re-read my sentence revision, make sure it is correct]}. Thus, in the worst case scenario, individuals with $Z_1 = 2$ are \textit{sure things}. The same logic applies for $Z_1 = -2$ where a client could either be a \textit{sure thing} or a \textit{do-not-disturb}. For \textit{lost cause} clients, by definition $Z_1 = 0$. This interpretation does not cover all scenarios for $Z_1=0$ but gives insights about ordering customers with $Y=1$. {\color{magenta} [ this paragraph is confusing. clean it up. you try to interpret and make sure the reader understand the meaning, but you confuse the reader further. Divide it perhaps into two paragraphs?]}
In the same spirit, we introduce $Z_0$, another transformed outcome variable which gives insights about ordering customers with $Y=0$, such as
\begin{equation}
Z_0 = \frac{T (1-Y)}{e(\mathbf x)} - \frac{(1-T)(1-Y)}{1-e(\mathbf x)}.
\label{eq:transformed_outcome_w}
\end{equation}
In our case, with $e(\mathbf x) = 1/2$, $Z_0$ simplifies to
\[ Z_0 =
\left\{ \begin{array}{cc}
2 & \mathrm{if}~T=1, Y=0, \\
-2 & \mathrm{if}~T=0, Y=0, \\
0 & \mathrm{otherwise.}
\end{array} \right.
\]
Here again, an interesting interpretation of $Z_0=2$ and $Z_0=-2$ raises. If $Z_0=2$happens only when a client is called ($T = 1$) but did not buy the insurance ($Y = 0$). Thus, for its counterfactual (i.e., $T = 0$), there are two possibilities i) either the client bought the policy ($Y = 1$) so belongs to the \textit{do-not-disturb} group, or ii) the client did not buy the policy ($Y = 0$) annd belong to the \textit{lost cause}. In the best case scenario, individuals with $Z_0 = 2$ is a \textit{lost cause}. The same logic applies for $Z_0 = -2$ where a client could either be a \textit{lost cause} or a \textit{persuadable}. Here, for \textit{sure things} clients, by definition $Z_0 = 0$. {\color{magenta} [follow the same paragraph structure as in $Z_1$. Your last sentence is confusing: why do you talk about ``sure thing" here? isn't it obvious?]}
\subsection{Augmented data conditional mean regression}
\label{sec:smite_ie}
{\color{magenta} [before going to the development you need to motivate the reader, why do you want this section, why data augmentation in this form is meaningful, etc. The development below are simple algebraic manipulation, you may declare a corollary and put the details in Appendix.]}
The fundamental problem of causal inference can be thought of as a missing data problem, or more specifically, a missing potential outcomes problem. Indeed, under the counterfactual consistency assumption, each participant is missing one potential outcome: that corresponding to the treatment not received.
If the potential outcomes were fully observed, then causal inference could be made in a trivial way. While often not explicitly discussed in the uplift literature, the links between causal inference and missing data can provide additional insight.
To make causal inference using a counterfactual framework, we must now find a way to impute the missing potential outcomes either implicitly or explicitly.
Amongst all the missing data imputation techniques, data augmentation is not so popular in the literature and very few articles mentioned the use of the technique to account for missing data problems.
The term {\it data augmentation} refers to methods for constructing iterative optimization or sampling algorithms via the introduction of unobserved data or latent variables. For deterministic algorithms, the method was popularized in the general statistical community by the seminal article by \cite{dempster1977maximum} on the EM algorithm for maximizing a likelihood function or, more generally, a posterior density.
Instead of using the transformed outcome directly as a response variable, we can think of it as an auxiliary or latent variable and augment the data.
Recall the transformed outcomes $Z_1$ \eqref{eq:transformed_outcome} and $Z_0$ \eqref{eq:transformed_outcome_w} defined in Section~\ref{sec:transformed_outcome}. Now consider the random variable
\[ Z = \left\{ \begin{array}{cc}
Z_1/2 & \mathrm{if}~ Y=1, \\
Z_0/2 & \mathrm{if}~ Y=0.
\end{array} \right.
\]
Now, suppose $n$ independent customer observations $\{y_i, \textbf{x}_i, t_i, z_i\}_{i=1}^n$ are at our disposal, where $\{z_i\}_{i=1}^n$ are realisations of the transformed outcome $Z$.
Note that $Z= 2T -1 $.
Again, we consider the case where $e(\mathbf x)=1/2$, hence, $\mathrm{Pr}(T=1)=\mathrm{Pr}(T=0)=1/2$. Although the following corollary is simplified based on the assumption assumption that $e(\mathbf x) = 1/2$, the development holds for any constant $e(\mathbf x)$. Moreover, one can use an under or over-sampling procedure to recover the $e(\mathbf x) = 1/2$ case.
\begin{corollary}
Let $m_{yt} = \mathrm{Pr}( Y=y | T=t)$, and $p_{yt} = m_{yt} / ( m_{yt} + m_{y(1-t)})$ and suppose that the propensity score $e(\mathbf x)=1/2$. Then, the augmented data $\{y_i, t_i, z_i\}_{i=1}^n$ likelihood can be written as
\begin{align}
\prod_{i=1}^n \mathrm{Pr}( z_i, y_i, t_i) = 2^{-n}\prod_{i=1}^n p_{y_i 1}^{t_i } (1 - p_{y_i 1})^{(1-t_i)} m_{1t_i}^{y_i} (1 - m_{1t_i})^{(1-y_i)}.
\label{eq:adata_lik}
\end{align}
\end{corollary}
\begin{corollary}
Let $\gamma$ denote the treatment effect, $\boldsymbol{\beta}$, the vector of main effects, $\boldsymbol{\delta}$, the vector of interactions effects, and $\theta_o$, the intercept.
Define the uplift model as
\begin{equation}
\mu_{1t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta}) = \mathrm{Pr} (Y_i = 1 \mid \mathbf{x}_i, t_i, \theta_o, \boldsymbol{\theta}) =
\bigl( 1+ \mathrm{exp}\{-(\theta_o + \gamma t_i + \mathbf{x}_i^\top [\boldsymbol{\beta} + t_i \boldsymbol{\delta}] ) \} \bigr)^{-1},
\label{eq:pitheta}
\end{equation}
where $\boldsymbol{\theta} = (\boldsymbol{\beta}, \gamma, \boldsymbol{\delta})$, denotes all model parameters except for the intercept $\theta_o$. The parameters are estimated by maximizing the likelihood \eqref{eq:adata_lik}. This is equivalent to minimizing the following objective function
\begin{equation}
- \sum_{i=1}^n \Big(y_i \log m_{ 1 t_i} + (1 - y_i)\log(1 - m_{1 t_i }) \Big) - \sum_{i=1}^n \Big(t_i \log p_{y_i 1} + (1 - t_i)\log(1 - p_{y_i 1}) \Big).
\label{eq:IE_neg_likelihood}
\end{equation}
\end{corollary}
We have
\begin{align*}
\mathrm{Pr}( Z = 1 | Y=1 ) & = \mathrm{Pr}( Z =1 , Y=1 ) / \mathrm{Pr}(Y = 1)\\
& = \{\mathrm{Pr}( Z=1, Y=1, T=1 ) + \mathrm{Pr}(Z=1, Y=1, T=0)\} / \mathrm{Pr}(Y=1) \\
& = \tfrac{\mathrm{Pr}(Z_1=2 | Y=1, T=1)\mathrm{Pr}(Y=1 | T=1) \mathrm{Pr}(T=1) + \mathrm{Pr}(Z_1=2 | Y=1, T=0)\mathrm{Pr}(Y=1 | T=0) \mathrm{Pr}(T=0)}
{ \mathrm{Pr}(Y=1|T=1)\mathrm{Pr}(T=1) + \mathrm{Pr}(Y=1 | T=0) \mathrm{Pr}(T=0) } \\
& = m_{11} / ( m_{11} + m_{10}) = p_{11}.
\end{align*}
Similarly,
\begin{align*}
\mathrm{Pr}( Z = -1 | Y=1 ) & = \mathrm{Pr}( Z =-1 , Y=1 ) / \mathrm{Pr}(Y = 1) \\
& = \{\mathrm{Pr}( Z=-1, Y=1, T=1 ) + \mathrm{Pr}(Z=-1, Y=1, T=0)\} / \mathrm{Pr}(Y=1) \\
& = \tfrac{\mathrm{Pr}(Z_1=-2 | Y=1, T=1)\mathrm{Pr}(Y=1 | T=1) \mathrm{Pr}(T=1) + \mathrm{Pr}(Z_1=-2 | Y=1, T=0)\mathrm{Pr}(Y=1 | T=0) \mathrm{Pr}(T=0) }
{ \mathrm{Pr}(Y=1|T=1)P(T=1) + \mathrm{Pr}(Y=1 | T=0) \mathrm{Pr}(T=0) } \\
& = m_{10} / ( m_{11} + m_{10}) = 1 - p_{11}.
\end{align*}
Proceeding in the same way, we have
\begin{align*}
\mathrm{Pr}( Z = 1 | Y=0 ) & = \mathrm{Pr}( Z =1 , Y=0 ) / \mathrm{Pr}(Y = 0) \\
& = \{ \mathrm{Pr}( Z=1, Y=0, T=1 ) + \mathrm{Pr}(Z=1, Y=0, T=0)\}/ \mathrm{Pr}(Y=0) \\
& = \tfrac{ \mathrm{Pr}(Z_0=2 | Y=0, T=1) \mathrm{Pr}(Y=0 | T=1) \mathrm{Pr}(T=1) + \mathrm{Pr}(Z_0=2 | Y=0, T=0) \mathrm{Pr}(Y=0 | T=0) \mathrm{Pr}(T=0) }
{ \mathrm{Pr}(Y=0|T=1) \mathrm{Pr}(T=1) + \mathrm{Pr}(Y=0 | T=0) \mathrm{Pr}(T=0)} \\
& = m_{01} / ( m_{01} + m_{00}) = p_{01},
\end{align*}
and
\begin{align*}
\mathrm{Pr}( Z = -1 | Y=0 ) & = \mathrm{Pr}( Z =-1 , Y=0 ) / \mathrm{Pr}(Y = 0) \\
& = \{ \mathrm{Pr}( Z=-1, Y=0, T=1 ) + \mathrm{Pr}(Z=-1, Y=0, T=0)\}/ \mathrm{Pr}(Y=0) \\
& = \tfrac{ \mathrm{Pr}(Z_0=-2 | Y=0, T=1) \mathrm{Pr}(Y=0 | T=1) \mathrm{Pr}(T=1) + \mathrm{Pr}(Z_0=-2 | Y=0, T=0) \mathrm{Pr}(Y=0 | T=0) \mathrm{Pr}(T=0) }
{ \mathrm{Pr}(Y=0|T=1) \mathrm{Pr}(T=1) + \mathrm{Pr}(Y=0 | T=0) \mathrm{Pr}(T=0) }\\
& = m_{00} / ( m_{01} + m_{00}) = 1 - p_{01}.
\end{align*}
Note that $Z = 1$ iff $T=1$ and $Z=-1$ iff $T=0$. This fact will be useful to write the likelihood of the complete data.
This is given by
\begin{align}
&\prod_{i=1}^n \mathrm{Pr}( z_i, y_i, t_i) = \prod_{i=1}^n \mathrm{Pr}( z_i | y_i) \mathrm{Pr}( y_i | t_i) \mathrm{Pr}(t_i) \label{eq:IE_likelihood}\\
&= 2^{-n}\prod_{i=1}^n p_{11}^{t_i y_i} (1 - p_{11})^{(1-t_i)y_i} p_{01}^{t_i (1-y_i)} (1 - p_{01})^{(1-t_i)(1-y_i)}
m_{11}^{t_i y_i} (1 - m_{11})^{t_i (1-y_i)} m_{10}^{(1-t_i) y_i} (1 - m_{10})^{(1-t_i) (1-y_i)} \nonumber \\
&= 2^{-n}\prod_{i=1}^n p_{y_i 1}^{t_i } (1 - p_{y_i 1})^{(1-t_i)} m_{1t_i}^{y_i} (1 - m_{1t_i})^{(1-y_i)} \nonumber \\
&= 2^{-n} \exp\biggl\{ \sum_{i=1}^n \Big( t_i \log p_{y_i 1} + (1 - t_i)\log(1 - p_{y_i 1}) \Big) \biggr\}
\exp\biggl\{ \sum_{i=1}^n \Big( y_i \log m_{ 1 t_i} + (1 - y_i)\log(1 - m_{1 t_i }) \Big) \biggr\}. \nonumber
\end{align}
Now, by rearranging the terms and taking the negative logarithm of the likelihood, maximizing \eqref{eq:IE_likelihood} is equivalent to minimizing the following function, with respect to the model's parameters, that is,
\begin{equation}
- \sum_{i=1}^n \Big(y_i \log m_{ 1 t_i} + (1 - y_i)\log(1 - m_{1 t_i }) \Big) - \sum_{i=1}^n \Big(t_i \log p_{y_i 1} + (1 - t_i)\log(1 - p_{y_i 1}) \Big).
\label{eq:IE_neg_likelihood}
\end{equation}
If we model $m_{1t_i}$ by $\mu_{1t_i}(\theta_o,\boldsymbol{\theta})$ as in Section \ref{sec:smite_to}, the first term of \eqref{eq:IE_neg_likelihood} is equivalent to the one of \eqref{eq:SMITE_objective}, that is, $\ell(\theta_o,\boldsymbol{\theta})$. The second term of \eqref{eq:IE_neg_likelihood}, which is also a function of $\mu_{1t_i}(\theta_o,\boldsymbol{\theta})$ can be interpreted as the regularization term. We can denote this function as $\mathcal{I}(\theta_o,\boldsymbol{\theta})$. In this case, our \textit{indirect} regularized conditional means regression solves
\begin{align}
(\hat{\theta}_o,{\boldsymbol{\hat\theta}}) = \argmin_{\theta_o,\boldsymbol{\theta}} \ell(\theta_o,\boldsymbol{\theta}) + \mathcal{I}(\theta_o,\boldsymbol{\theta}).
\label{eq:SMITE_objective_IE}
\end{align}
{\color{magenta} where is $\lambda$? is it equal one? then it is a special case, mention it!}
{\color{magenta} [this section is badly written. it is not clear what is the purpose of the section until the end of the section. It appears to me that you develop a model that becomes a special case of the Lagragian form. In this case, why you do not start with this section and put the Largragian after the model to claim generalization? Think again about this section. You need to bring out the message of the section earlier.]}
\subsection{Regularized conditional means regression}
\label{sec:smite_to}
We propose to combine $Z_1$ and $Z_0$ in order to define a new random variable, which can be used as an unbiased estimator of the uplift, but still carry a meaningful interpretation. We define the random variable $S$ such as when $e(\mathbf x) = 1/2$,
\begin{equation}
S = \frac{1}{2} (Z_1 - Z_0) =
\left\{ \begin{array}{cc}
1 & \mathrm{if} ~ T = Y, \\
-1 & \mathrm{if} ~ T \neq Y.
\end{array} \right.
\label{eq:transformed_outcome_sign}
\end{equation}
Here, with regards to the 4 types of clients, {\color{magenta} [which 4 types of clients?]} we have that for $S=1$, the client is either a \textit{lost cause}, a \textit{sure thing} or a \textit{persuadable} (i.e. $u(\mathbf x) \geq 0$) and for $S=-1$, the client is either a \textit{do-not-disturb}, a \textit{lost cause} or a \textit{sure thing} (i.e. $u(\mathbf x) \leq 0$). Therefore, the random variable $S$ can be interpreted as the sign of $u(\mathbf x)$. Next, we will use this transformed outcome in order to penalize the logistic regression for uplift.
Suppose $n$ independent customer observations $\{y_i, \textbf{x}_i, t_i, s_i\}_{i=1}^n$ are at our disposal, where $\{s_i\}_{i=1}^n$ are realisations of the transformed outcome $S$. Let us denote the model parameters by $(\theta_o, \boldsymbol{\theta})$. Regularization can be seen as a constrained optimization problem; using the transformed outcome $S$ the optimization problem reduces to
\begin{align}
{(\hat{\theta}_o,\boldsymbol{\hat\theta})} & = \argmin_{\boldsymbol{\theta}} ~ \ell\Big({\mu_{1t}}(\theta_o,\boldsymbol{\theta}) \mid \mathbf y, \mathbf x, \mathbf t \Big) \nonumber \\
& \mathrm{s.t.} \nonumber \\
& \sign \Big(u_i(\theta_o,\boldsymbol{\theta})\Big) = s_i \mathrm{~for~} i=1,\ldots,n.
\label{eq:objective}
\end{align}
The quantities $\mu_{1t}(\cdot)$ and $u_i(\cdot)$ are functions of the covariates $\mathbf x$ and could be denoted $\mu_{1t}(\mathbf x)$ and $u(\mathbf x_i)$. {\color{magenta} [you need to talk about $\ell$ here.]} To simplify the notation, we prefer to omit this specification throughout the rest of the paper. This optimization predicts the uplift $u(\mathbf x)$ and the sign of $u(\mathbf x)$ is determined by transformed outcome $S$ {\color{magenta} [what is the point of this sentence?]}. Formally, we define the negative log-likelihood $\ell(\theta_o,\boldsymbol{\theta})$ with respect to the model parameters as
\begin{equation*}
\ell(\theta_o,\boldsymbol{\theta}) = - \sum_{i=1}^{n} \Big( y_i\mathrm{log} \{\mu_{1t_i}(\theta_o,\boldsymbol{\theta})\} + (1-y_i) \mathrm{log} \{1-\mu_{1t_i}(\theta_o,\boldsymbol{\theta})\} \Big),
\end{equation*}
where $\mu_{1t_i}(\theta_o,\boldsymbol{\theta})$ is defined in \eqref{eq:pitheta}. The constraint in \eqref{eq:objective} is what we will next refer to as the \textit{direct} uplift regularization. We propose to use the exponential loss such as
\begin{equation}
\mathcal{J}(\theta_o,\boldsymbol{\theta}) = \frac{1}{n} \sum_{i=1}^{n} \mathrm{exp}\Big(- s_i u_i(\theta_o,\boldsymbol{\theta}) \Big).
\label{eq:direct_uplift_loss}
\end{equation}
where $u_i(\theta_o,\boldsymbol{\theta}) = {\mu_{11_i}}(\theta_o,\boldsymbol{\theta }) - {\mu_{10_i}}(\theta_o,\boldsymbol{\theta})$. It is helpful to rewrite the Lagrangian dual of (\ref{eq:objective}) in his as
\begin{align*}
(\hat{\theta}_o,{\boldsymbol{\hat\theta}})_{\lambda} = \argmin_{\theta_o,\boldsymbol{\theta}} \ell(\theta_o,\boldsymbol{\theta}) + \lambda \mathcal{J}(\theta_o,\boldsymbol{\theta})
\end{align*}
where $\lambda > 0$ is the regularization constant. However, in our framework, we use the transformation $\alpha = \lambda / (1+\lambda)$. This transformation allows to interpret $\alpha \in [0,1)$ as a hyper-parameter which controls the trade-off between the two losses in the optimization problem. Hence, our direct regularized conditional means regression solves
\begin{align}
(\hat{\theta}_o,{\boldsymbol{\hat\theta}})_{\alpha} = \argmin_{\theta_o,\boldsymbol{\theta}} (1 - \alpha) \ell(\theta_o,\boldsymbol{\theta}) + \alpha \mathcal{J}(\theta_o,\boldsymbol{\theta}).
\label{eq:SMITE_objective}
\end{align}
\section{Graphical representation of the uplift model}
Logistic regression may be visualised in a model diagram, with a single node to represent the link function, and multiple nodes to represent input or outputs. This sort of vosialuzation is very common in neural networks community. In Figure~\ref{fig:int_as_nn}, we show a the diagram of the uplift interaction model \eqref{eq:pitheta}.
\begin{figure}[!ht]
\centering
\input{figures/tikz_int_1hrelu}
\caption{Graphical representation of the uplift interaction model (left panel). Neural network representation of the uplift model (right panel). Note that $h_j = \mathrm{ReLU}(\mathbf x,t)$ for $j=1,\ldots,m$ and for $m=2p+1$, Theorem~\ref{theorem} shows the equivalence between the interaction model \eqref{eq:neural_network_1} and a particular case of the neural network with $1$ hidden layer and $\mathrm{ReLU}$ activation \eqref{eq:neural_network_2}.
}
\label{fig:int_as_nn}
\end{figure}
\subsection{Uplift neural model}
The uplift interaction model \eqref{eq:pitheta}
can be represented by a fully-connected neural network with no hidden layer, a bias, $2p+1$ input neurons and $1$ output neuron with sigmoid activation function (see Figure~\ref{fig:int_as_nn} left panel). Without loss of generality, we define $\mathbf x = (x_1,...,x_p) \in [0,c]^{p}$ for $c>0$ and $t \in \{0,1\}$, a binary variable. Let us further define $\theta_{j}$, for $j=1,\ldots,2p+1$, the weight that connects the $j$th neuron to the output and let $\theta_o \in \mathbb{R}$ be the bias.
The output of the uplift interaction model neural network $\mathrm{NN}_1$ can be written as
\begin{equation}
\mathrm{NN}_1(\mathbf x,t) = \sigma \Big(\theta_o + \sum_{j=1}^{p} \theta_j x_j + \sum_{j=p+1}^{2p} \theta_j t x_{j-p} + \theta_{2p+1} t \Big),
\label{eq:neural_network_1}
\end{equation}
where $\sigma(\cdot)$ represents the sigmoid function. Now, define $\mathrm{NN}_2$, a fully connected network, with an input of size $p+1$, and one hidden layer of size $2p+1$ with a bias attached to each hidden node as well as a ReLU activation function $\mathrm{ReLU}(\cdot)$, where for $z \in \rm I\!R$, we have $\mathrm{ReLU}(z) = \max\{0,z\}$. The hidden layer is then connected to a single output neuron with a sigmoid activation function (see Figure~\ref{fig:int_as_nn} right panel). The output of the neural network $\mathrm{NN}_2$ can be written as
\begin{equation}
\mathrm{NN}_2(\mathbf x,t) = \sigma \biggl\{ w_o^{(2)} + \sum_{k=1}^{2p+1} w_k^{(2)} \mathrm{ReLU}\Big(w_{o,k}^{(1)} + \sum_{j=1}^{p} w_{j,k}^{(1)} x_j + w_{p+1,k}^{(1)} t \Big) \biggr\},
\label{eq:neural_network_2}
\end{equation}
where $w_{j,k}^{(1)}$ represent the weight that connects the $j$th input neuron to the $k$th hidden neuron and $w_k^{(2)}$ represents the weight that connects the $k$th hidden neuron to the output. We denote the biases for the hiddden layer and the output layer by $w_{o,k}^{(1)}$, $k=1,\ldots,2p+1$ and $w_o^{(2)}$ respectively.
\begin{theorem} \label{theorem}
Let $\mathrm{NN}_1(\mathbf x,t)$ and $\mathrm{NN}_2(\mathbf x,t)$ be two neural networks defined as in \eqref{eq:neural_network_1} and \eqref{eq:neural_network_2} respectively. Let $c \in \rm I\!R^+$ be a positive constant. For all $\theta_j \in \rm I\!R$, $j=1,\ldots,2p+1$ and $\theta_o \in \rm I\!R$, there exists a matrix of weights $\Big(w_{j,k}^{(1)}\Big) \in \rm I\!R^{(p+1) \times (2p+1)}$, a bias vector $\Big(w_{o,k}^{(1)}\Big) \in \rm I\!R^{2p+1}$, a vector of weights $\Big(w_{k}^{(2)}\Big) \in \rm I\!R^{2p+1}$ and a bias scalar $w_o^{(2)} \in \rm I\!R$ such that for all $\mathbf x \in [0,c]^p$ and $t \in \{0,1\}$
\begin{equation}
\mathrm{NN}_1(\mathbf x,t) = \mathrm{NN}_2(\mathbf x,t).
\label{eq:theorem_equality}
\end{equation}
\end{theorem}
{\color{magenta} put the proof in the Appendix.}
\begin{proof}
We prove the theorem by finding explicitly the weights of $\mathrm{NN}_2(\mathbf x,t)$ that verify the equality \eqref{eq:theorem_equality}.
First, we fix the vector of weights that connects the hidden layer to the output such as
$$w_{k}^{(2)} = \theta_j,$$
for $k=j$, for $k,j=1,\ldots,2p+1$. Moreover, we fix the bias scalar $w_o^{(2)}$ to be equal to the bias scalar from the first network, that is, $w_o^{(2)} = \theta_o$. Therefore, $\mathrm{NN}_2(\mathbf x,t)$ can be written as
\begin{equation}
\mathrm{NN}_2(\mathbf x,t) = \sigma \biggl\{ \theta_o + \sum_{k=1}^{2p+1} \theta_k r\Big(w_{o,k}^{(1)} + \sum_{j=1}^{p} w_{j,k}^{(1)} x_j + w_{p+1,k}^{(1)} t \Big) \biggr\}.
\label{eq:neural_network2_proof}
\end{equation}
Next, we need to define the right structure for the matrix of weights $\Big(w_{j,k}^{(1)}\Big) \in \rm I\!R^{(p+1) \times (2p+1)}$ to be able to recover the
$\mathrm{NN}_1(\mathbf x,t)$ model. First, let us fix the bias vector $\Big(w_{o,k}^{(1)}\Big) \in \rm I\!R^{2p+1}$ such as
\[ w_{o,k}^{(1)} =
\left\{ \begin{array}{cl}
0 & \mathrm{if}~k \in \{1,\ldots,p\}\\
-c & \mathrm{if}~k \in \{p+1, \ldots, 2p\}\\
0 & \mathrm{if}~k=2p+1
\end{array} \right.
\]
Finally, we give the following representation to the matrix of weights $\Big(w_{j,k}^{(1)}\Big) \in \rm I\!R^{(p+1) \times (2p+1)}$, that is,
\[ \Big(w_{j,k}^{(1)}\Big) =
\left(
\begin{array}{cccc|cccc|c}
1 & 0 & \cdots & 0 & 1 & 0 & \cdots & 0 & 0 \\
0 & 1 & \cdots & 0 & 0 & 1 & \cdots & 0 & 0 \\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & \cdots & 1 & 0 & 0 & \cdots & 1 & 0 \\
\hline
0 & 0 & \cdots & 0 & c & c & \cdots & c & 1 \\
\end{array}
\right)
\]
Therefore, based on these attributions, we have the following equalities. For $k \in \{1, \ldots, p\}$,
\begin{equation}
r\Big(w_{o,k}^{(1)} + \sum_{j=1}^{p} w_{j,k}^{(1)} x_j + w_{p+1,k}^{(1)} t \Big) = r(x_k) = x_k.
\label{eq:1st_equality}
\end{equation}
Then, for $k \in \{p+1, \ldots, 2p\}$,
\begin{equation}
r\Big(w_{o,k}^{(1)} + \sum_{j=1}^{p} w_{j,k}^{(1)} x_j + w_{p+1,k}^{(1)} t \Big) = r(-c + x_{k-p} + ct) = t x_{k-p}.
\label{eq:2nd_equality}
\end{equation}
In \eqref{eq:2nd_equality}, the outcome depends on the value of $t$ and the second equality comes from the fact that we have
\[ r(-c + x_{k-p} + ct) =
\left\{ \begin{array}{ll}
r(x_{k-p} - c) = 0 & \mathrm{if}~t = 0,\\
r(x_{k-p}) = x_{k-p} & \mathrm{if}~t=1,
\end{array} \right.
\]
therefore, $r(-c + x_{k-p} + ct) = t x_{k-p}$. Finally, for $k=2p+1$, we have
\begin{equation}
r\Big(w_{o,k}^{(1)} + \sum_{j=1}^{p} w_{j,k}^{(1)} x_j + w_{p+1,k}^{(1)} t \Big) = r(t) = t.
\label{eq:3rd_equality}
\end{equation}
Hence, the desired result is obtained by simply replacing \eqref{eq:1st_equality}, \eqref{eq:2nd_equality} and \eqref{eq:3rd_equality} in \eqref{eq:neural_network2_proof}.
\end{proof}
\subsection{Twin neural model for uplift}
After the uplift model parameters are estimated, the uplift prediction is computed in three steps. i) step the conditional probabilities are estimated by fixing the treatment variable $T$ to $1$ or $0$ ii) the difference is taken to compute the uplift iii) the uplift is visualized. The model plays a major role in implementing these three steps. The twin neural model diagram is shown in Figure~\ref{fig:siamese_network}. This representation reminds us resembles the Siamese model {\color{magenta} [cite the siamese model]} \cite{} in the context of deep learning.
\begin{figure}[!ht]
\centering
\input{figures/tikzsmite}
\caption{Diagram of the twin logistic interaction model {\color{magenta} [add the ReLU hidde layers to the diagram]}.}
\label{fig:siamese_network}
\end{figure}
Twin networks were first introduced in signature verification as an image matching problem \citep{bromley1994signature}, where the task is to compare two hand-written signatures and infer identity. A twin neural network consists of two networks which are fed with distinct inputs, but their output are matched by an energy function. This function computes some metric between the highest level feature representation on each side. The parameters between the twin networks are tied. Weight tying guarantees that two extremely similar images could not possibly be mapped by their respective networks to different locations in feature space because each network computes the same function. Such networks are mostly known for their application to face verification and recognition \citep{chopra2005learning,taigman2014deepface,parkhi2015deep,schroff2015facenet}, areal-to-ground image matching \citep{lin2015learning}, stereo matching \citep{zbontar2015computing}], patch matching \citep{han2015matchnet, simo2015discriminative,zagoruyko2015learning}, optical flow \citep{dosovitskiy2015flownet}, large scale video classification \citep{karpathy2014large} and one-shot character recognition \citep{koch2015siamese}. To the best of our knowledge we are the first to propose a twin model in the uplift estimation. Our experiments show that this approach bridge the gap between performance and interpretability {\color{magenta} [Shall we move this sentence upper to introduction or methodology section?]}.
The twin network representation of the uplift interaction model is easily generalized to neural networks, as show in Figure~\ref{fig:nite}. There are several reasons why neural networks are suitable tool for uplift i) they are flexible models and easy to train with current GPU hardware. ii) with a single hidden layer modeling input interactions is straightforward \citep{tsang2018neural} iii) they are guaranteed to approximate a large class of functions \citep{cybenko1989approximation, hornik1991approximation, pinkus1999approximation} iv) neural networks perform very well on predictive tasks which is the main objective of uplift v) a simple network architecture ensures model interpretability for further studies.
\input{figures/nnsmite.tex}
{\color{magenta} [Revise caption to Figure 3. Add ReLU layer to Figure 3.]}
Here, we propose a twin neural network for uplift which is composed of two collaborative models two almost identical input vectors. This twin model outputs the estimate of the conditional mean $\mu_{1t}(\mathbf x)$ and at the same time predicts uplift $u(\mathbf x)$. This model simplifies the fitting process to a great extent. After feeding the network with input values, the model computes the conditional means, as well as the uplift. This is why two input vectors are fed to almost identical sub-models with shared parameters.
The inputs contain the covariates vector $\mathbf x$ and, for the left sub-component, the treatment variable fixed to $1$. The treatment variable is fixed to $0$ for the right sub-component as shown in Figures~ \ref{fig:siamese_network} and \ref{fig:nite}. The sub-components output the predicted conditional means for treated ($\mu_{11}(\mathbf x)$) and for control ($\mu_{10}(\mathbf x)$). The difference gives direct prediction of $u(\mathbf x)$. At the same time, the predicted conditional mean $\mu_{1t}(\mathbf x)$ is based on the actual received treatment for each individual $t \in \{0,1\}$, i.e., $\mu_{1t}(\mathbf x) = t \mu_{11}(\mathbf x) + (1-t) \mu_{10}(\mathbf x)$.
{\color{magenta} This paragraph is confusing. I did not get what you wanted to say. Re-write.}
\subsection{Parameter estimation}
\label{sec:smite_to_sgd}
Gradient-based optimization is one of the pillars of machine learning. Given an objective function $f: \rm I\!R^n \rightarrow \rm I\!R$, classical gradient descent has the goal of finding
(local) minima $\hat{\boldsymbol{\theta}} = \argmin_{\boldsymbol{\theta}} f(\boldsymbol{\theta})$ via updates of the form $ \Delta (\boldsymbol{\theta}) = -\xi \nabla f$, where $\xi$
is a positive step size (or the learning rate). Two main classes of optimization methods can be applied to solve such problems : (i) {\it stochastic} and (ii) {\it batch}.
The prototypical stochastic optimization method is the stochastic gradient
method \citep{robbins1951stochastic}, which, in the context of minimizing $f(\boldsymbol{\theta})$ and with $\boldsymbol{\theta}^{(0)} \in \rm I\!R^p$ given, is defined by
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \xi \nabla f_{i_k} (\boldsymbol{\theta}^{(k)}).
\end{equation}
Here, for all $k \in \mathbb{N}$, the index $i_k$ is chosen randomly from $\{1,...,n\}$. There is no particular reason to employ information from only one sample point per iteration. Instead, one can employ a {\it mini-batch} approach in which a small subset of samples $\mathcal{S}_k \subseteq \{1, ..., n\}$, is chosen randomly in each iteration, leading to
\begin{equation}
\boldsymbol{\theta}^{(k+1)} \leftarrow \boldsymbol{\theta}^{(k)} - \frac{\xi}{|\mathcal{S}_k|} \sum_{i \in \mathcal{S}_k} \nabla f_{i} (\boldsymbol{\theta}^{(k)}).
\label{eq:mini-batch-sgd}
\end{equation}
Here, $|\mathcal{S}_k|$ is the batch size. Such a mini-batch stochastic gradient method has been widely used in practice. There are some fundamental practical and theoretical reasons why stochastic methods have inherent advantages for large-scale machine learning \citep{bottou2018optimization}. In our experiments, we train our models by solving (\ref{eq:SMITE_objective}) and/or (\ref{eq:SMITE_objective_IE}) using the mini-batch stochastic gradient descent (SGD) updates as in (\ref{eq:mini-batch-sgd}). Note that optimization algorithms that use only the gradient are called first-order optimization algorithms. Second-order optimization algorithms such as Newton's method use the Hessian matrix. Like all neural network training, we train the network using first-order gradient methods.
\section{Evaluating uplift models}\label{sec:qini}
In the context of model selection, in practice, given $M$ models and/or hyperparameter settings, we build $M$ estimators $\{ \hat{u}_1, \hat{u}_2, \ldots, \hat{u}_M \}$. The quantity of interest in this case is the expected prediction risk of the model when it is applied to new data, according to some goodness-of-fit statistic. The idea of model selection is to estimate this expected risk for each of our models and find the model that optimizes it. There are several ways to estimate this risk, including information criteria, but data-splitting is the easiest and most widely applicable \citep{arlot2010survey}. Before fitting the models, the observations are randomly split into training and validation samples. The models are fit on the training sample $\mathcal{T}$ and evaluated on the validation sample $\mathcal{V}$. In cross-validation, this process is repeated round-robin across different random splits of the data and the estimated risks are averaged per model.
Classic evaluation approaches break down for treatment effect estimation because the true treatment effect is never observed in any sample. Despite this, we can attempt to use prediction error to select among treatment effect models.
Assuming the treatment effect model is built by regressing the outcomes onto the covariates and treatment to obtain $\hat\mu_0$ and $\hat\mu_1$, straightforward goodness-of-fit statistic often referred to as the $\mu$-risk can be compared for different models. This statistic is defined for continuous outcome $Y$ and can be estimated by $\frac{1}{|\mathcal{V}|}\sum_{i \in \mathcal{V}}( \hat{\mu}_{t_i}(\mathbf x_i) - y_i )^2$ with observations from the validation set $\mathcal{V}$. In the context of a binary response variables, one can use the Deviance instead.
An alternative is to maximize the {\it value} of the treatment policy $d(\mathbf x) = \mathbbm{1}(u(\mathbf x) > 0)$ which indicates which individuals we expect to benefit from the treatment \citep{zhao2017uplift}. The value of a decision policy $\hat d$ is estimated by $\hat v = \frac{1}{|\mathcal{V}|}\sum_{i \in \mathcal{V}} \frac{\mathbbm{1}(\hat d(\mathbf x_i) = t_i) y_i}{e(\mathbf x_i)}$. In other words, the value is the expected outcome of an individual when all individuals are treated according to the policy $\hat d$. If larger values of the outcome are more desirable (e.g. lifespan, click-through rate, approval ratings), then the policy that maximizes the value is optimal. If all we are interested is the treatment decision (and not the treatment effect itself), then we are already justified in maximizing value. However, a drawback of this metric is that it only utilizes a portion of the data. Indeed, it throws away individuals whose treatments do not match $\hat d(\mathbf x_i)$.
The goal of uplift modeling is to separate all observations into heterogeneous segments that differ as much as possible with regard to the reactions to the treatment and, thus, identify (sub)groups with very different uplifts. In the uplift literature, the {\it Qini coefficient} is commonly used as a goodness-of-fit measure for uplift models. The computation of this statistic requires the construction of the {\it Qini curve} \citep{radcliffe2007using}.
For a given model, let $\hat{u}_{(1)} \geq \hat{u}_{(2)} \geq ... \geq \hat{u}_{(|\mathcal{V}|)}$ be the sorted predicted uplifts on the validation set $\mathcal{V}$. Let $\phi \in [0,1]$ be a given proportion and let $ N_{\phi} = \{i: \hat{u}_{i} \geq \hat{u}_{(\ceil{\phi |\mathcal{V}|})} \} \subset \lbrace 1, \ldots, |\mathcal{V}| \rbrace$ be the subset of observations with the $\phi |\mathcal{V}| \times 100 \%$ highest predicted uplifts $\hat{u}_i$ (here $\ceil{s}$ denotes the smallest integer larger or equal to $s\in \rm I\!R$).
The \textit{Qini curve} is defined as a function $f$ of the fraction of population targeted $\phi$, where
\begin{equation*}
f(\phi) = \frac{1}{n_t}\biggl(\sum\limits_{i \in N_{\phi}} y_i t_i - \sum\limits_{i \in N_{\phi}} y_i (1-t_i) \biggl\{ \sum\limits_{i \in N_{\phi}} t_i / \sum\limits_{i \in N_{\phi}} (1-t_i) \biggr\} \biggr),
\end{equation*}
where $n_t = \sum_{i=1}^{|\mathcal{V}|} t_i$ is the number of treated customers, with $f(0)=0$ and $f(1)$ is the average treatment effect. In other words, $f(\phi)$ represents the incremental number of positive responses for a fraction $\phi$ or targeted customers relative to the total number of targeted customers in the sample. The Qini helps visualize the performance of an uplift model, a bit like the ROC curve of a binary classification model. Following the same logic, a straight line between the points $(0,0)$ and $(1, f(1))$ defines a benchmark to compare the performance of the model to a strategy that would randomly target individuals. This means that if a proportion $\phi$ of the population is targeted, we expect to observe an uplift of $\phi f(1)$.
The {\it Qini coefficient} \citep{radcliffe2007using} is defined as
\begin{equation*}
q = \int_0^1 Q(\phi) ~\mathrm{d}\phi = \int_0^1 \{f(\phi) - \phi~f(1)\} ~\mathrm{d}\phi,
\end{equation*}
where $Q(\phi) = f(\phi) - \phi~f(1)$. This area can be numerically approximated using a Riemann method such as the trapezoid rule formula: the domain of $\phi \in [0,1]$ is partitioned into $K$ bins, or $K+1$ grid points $0=\phi_1 < \phi_2 < ... < \phi_{K+1} = 1$, to approximate $q$ by its empirical estimation
\begin{equation}
\hat{q} = \dfrac{1}{2} \sum_{k=1}^K (\phi_{k+1}-\phi_k)\{Q(\phi_{k+1}) + Q(\phi_{k})\}.
\label{eq:q:hat}
\end{equation}
It is important to note, unlike the area under the ROC curve, $\hat{q}$ could be negative. This would simply mean that a strategy following the model in question does worse than random targeting.
Measuring the goodness-of-fit of an uplift model can also be achieved by the similarity between the theoretical uplift percentiles (given by the model's predictions) and empirical percentiles (observed in the data) based on the model. This can be approximated by the {\it Kendall's uplift rank correlation} \citep{belba2019qbased} defined as
\begin{equation}
\hat{\rho} = \frac{2}{K(K-1)} \sum_{i<j} \mathrm{sign}(\bar{\hat{u}}_i - \bar{\hat{u}}_j)~\mathrm{sign}(\bar{u}_i - \bar{u}_j),
\label{eq:corr_coeff}
\end{equation}
where $\bar{\hat{u}}_k$ is the average predicted uplift in bin $k$, $k \in {1,...,K}$, and $\bar{u}_k$ is the observed uplift in the same bin.
The \textit{adjusted Qini coefficient} represents a trade-off between maximizing the area under the Qini curve and grouping the individuals in decreasing uplift bins. It is given by
\begin{equation}
\hat{q}_{\mathrm{adj}} = \rho~\mathrm{max}\{ 0, \hat{q}\}.
\label{eq:qadj}
\end{equation}
\section{Experiments}
\label{sec:experiments}
We demonstrate the utility of our proposed methods using simulations. Each simulation is defined by a data-generating process with a known effect function. Each run of each simulation generates
a dataset, which we split into training, validation, and test samples. We use
the training data to estimate $M$ different uplift functions $\{\hat{u}_m\}_{m=1}^M$ using
$M$ different algorithms (e.g. Qini-based uplift regression, SMITE$_{\mathrm{TO}}$ or causal forest). For each of those we calculate the validation metric $\hat{q}_{\mathrm{adj}}$ using the data in validation set. The models selected by optimizing each validation-set metric are then used to estimate the uplift on the test set. Each simulation is repeated multiple times.
\subsection{Data generating process}
We generate synthetic data in a similar way to the simulation studies of \cite{tian2014simple}, inspired by scenarios defined in \cite{powers2018some}. For each scenario, we generate $n$ observations and $p$ covariates. We draw odd-numbered covariates independently from a standard Gaussian distribution. We draw even-numbered covariates independently from a
Bernoulli distribution with probability $1/2$. Accross all scenarios, we define the mean effect function $\mu(\cdot)$ and the treatment effect function $\tau(\cdot)$. We also control the noise level $\sigma^2$.
Given the elements above, our data generation model is, for $i=1,...,n$ :
\begin{align*}
Y_i^* \mid \mathbf x_i, t_i \sim \mathcal{N}(\mu(\mathbf x_i) + t_i\tau(\mathbf x_i), \sigma^2).
\end{align*}
where $t_i$ is the realisation of the random variable $T_i \sim \mathrm{Bernoulli}(1/2)$. Then, we define $Y_i$ to be the binary outcome as $
Y_i = \mathbbm{1}(Y_i^* > 0 \mid \mathbf x_i, t_i)$.
Hence, $Y_i$ follows a Bernoulli distribution with parameter $\mathrm{Pr}(Y_i^* > 0 \mid \mathbf x_i, t_i)$ and it is easy to recover the ``true'' uplift $u_i^*$ as follows
\begin{align}
u_i^* &= \mathrm{Pr}(Y_i^* > 0 \mid \mathbf x_i, T_i = 1) - \mathrm{Pr}(Y_i^* > 0 \mid \mathbf x_i, T_i = 0) \nonumber \\
u_i^* &= \Phi \Bigl( \frac{\mu(\mathbf x_i) + \tau(\mathbf x_i)}{\sigma} \Bigr) - \Phi \Big( \frac{\mu(\mathbf x_i)}{\sigma} \Big) \label{eq:sim_true_uplift}
\end{align}
for $i=1,\ldots,n$, where $\Phi(.)$ is the standard Gaussian cumulative distribution function. Following \cite{powers2018some}, within each set of simulations, we make different choices of mean effect function and treatment effect function, meant to represent a wide variety of functional forms: both univariate and multivariate; both additive and interactive; both linear and piecewise constant. The functions are:
\begin{align*}
f_1(\mathbf x) &= 0 \\
f_2(\mathbf x) &= 5 \mathbbm{1}(x_1 > 1) - 5\\
f_3(\mathbf x) &= 2 x_1 - 4 \\
f_4(\mathbf x) &= x_2 x_4 x_6 + 2 x_2 x_4 (1-x_6) +3x_2(1-x_4)x_6 + 4x_2(1-x_4)(1-x_6) + 5(1-x_2)x_4x_6 \\
&+ 6(1-x_2)x_4(1-x_6) + 7 (1-x_2)(1-x_4)x_6 + 8(1-x_2)(1-x_4)(1-x_6)\\
f_5(\mathbf x) &= x_1 + x_3 + x_5 + x_7 + x_8 + x_9 - 2\\
f_6(\mathbf x) &= 4 \mathbbm{1}(x_1 > 1)\mathbbm{1}(x_3 > 0) + 4\mathbbm{1}(x_5 > 1)\mathbbm{1}(x_7 > 0) + 2x_8x_9\\
f_7(\mathbf x) &= \frac{1}{2} (x_1^2 + x_2 + x_3^2 + x_4 + x_5^2 + x_6 + x_7^2 + x_8 + x_9^2 - 11)\\
f_8(\mathbf x) &= \frac{1}{\sqrt{2}} (f_4(\mathbf x) + f_5(\mathbf x)).
\end{align*}
Table~\ref{tab:scenarios} gives the mean and treatment effect functions for the different randomized simulations, in terms of the eight functions above. In Figure~\ref{fig:qini_curves_scene}, we show the Qini curves associated with the ``true" uplift $u^*$ \eqref{eq:sim_true_uplift}.
\begin{table}[!ht]
\centering
\caption{Specifications for the 5 simulation scenarios. The rows of the table correspond, respectively, to the sample size, dimensionality, mean effect function, treatment effect function, noise level and the adjusted Qini coefficient based off the ``true" uplift $u^*$ \eqref{eq:sim_true_uplift}.}
\label{tab:scenarios}
\begin{tabular}{l|ccccc}
Scenarios & 1 & 2 & 3 & 4 & 5\\
\hline
$n$ & 10000 & 20000 & 20000 & 20000 & 20000 \\
$p$ & 200 & 100 & 100 & 100 & 100 \\
$\mu(\mathbf x)$ & $f_7(\mathbf x)$ & $f_3(\mathbf x)$ & $f_1(\mathbf x)$ & $f_2(\mathbf x)$ & $f_6(\mathbf x)$\\
$\tau(\mathbf x)$ & $f_4(\mathbf x)$ & $f_5(\mathbf x)$ & $f_6(\mathbf x)$ &$ f_7(\mathbf x)$ & $f_8(\mathbf x)$\\
$\sigma$ & $1/2$ & $1$ & $1$ & $2$ & $4$\\
$\hat{q}_{\mathrm{adj}} (u^*)$ & $1.80$ & $3.00$ & $10.00$ & $ 2.16$ & $3.68$\\
\hline
\end{tabular}
\end{table}
The defined scenarios are interesting because they represent different situations that we can face with real data. Indeed, scenario 1, although it seems simpler than the others (with a small noise level), is in fact a case that happens often, especially for marketing campaigns. In this scenario, by construction, the uplift is positive or zero. Thus, there are no do-not-disturb clients. This results in an increasing monotonic Qini curve. In scenario 2, we introduce a group of do-not-disturb clients. In addition, we increase the noise level as well as the number of observations, but we reduce the number of variables. Overall, the treatment effect is negative but a small group of persuadables is generated. In scenario 3, it is interesting to note that the main effect function is zero. Thus, the uplift is entirely guided by the treatment effect function. In scenario 4, we introduce a quadratic treatment effect function and scenario 5 is the most complex one. Using these scenarios, we believe it make it possible to compare the different methods rigorously.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.3\textwidth]{figures/simulations/Qini_Curve_Scene_1.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Qini_Curve_Scene_2.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Qini_Curve_Scene_3.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Qini_Curve_Scene_4.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Qini_Curve_Scene_5.pdf}
\caption{Qini curve based on the ``true" uplift with respect to the five scenarios of Table~\ref{tab:scenarios}.}
\label{fig:qini_curves_scene}
\end{figure}
\subsection{Benchmarks and implementation}
Each run of each simulation generates a dataset, which we split into training ($40\%$), validation ($30\%$), and test samples ($30\%$). We use the training and the validation sets to fit and fine-tune the models by maximizing the adjusted Qini coefficient. Note that fine-tuning is specific to the type of the model, e.g. for our transformed outcome model, we fine-tune the learning rate and the trade-off constant but for random forests, we fine-tune the number of trees as well as the depth of each tree. Then, we score the test samples with the fine-tuned models and compute the related adjusted Qini coefficients. We repeat each experiment $20$ times.
For each simulated dataset, we implement the following models:
\begin{enumerate}
\item[(a)] a multivariate logistic regression without penalization as in (\ref{eq:uplift_likelihood}). This is the baseline model, and we will refer to it as \textit{Logistic}.
\item[(b)] a Qini-based uplift regression model that uses several LHS structures to search for the optimal parameters (see \citep{belba2019qbased} for more details). We denote this model by \textit{Qini-based}.
\item[(c)] two types of random forests designed for causal inference (see \citep{athey2019generalized} for more details). This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the uplift. We denote these models by \textit{CF} and \textit{CF$_h$} without and with honest estimation respectively.
\item[(d)] two types of random forests designed for uplift (see \citep{jaskowski2012uplift} for more details). This method uses different split criteria. We denote these models by \textit{ URF$_{\mathrm{KL}}$} and \textit{ URF$_{\mathrm{ED}}$} for Kullback-Leibler divergence and Euclidean distance split criterion respectively.
\item[(e)] our SMITE regression model that uses the TO objective function to estimate the regression parameters as explained in Section~\ref{sec:smite_to}. We denote this model by \textit{TO$_{\hat{\alpha}}$}.
\item[(f)] our SMITE regression model that uses the IE objective function to estimate the regression parameters as explained in Section~\ref{sec:smite_ie}. We denote this model by \textit{IE}.
\end{enumerate}
In our experiments, we make use of available implementations from the authors of the benchmarks. We use the \textbf{R} libraries \textit{grf} \citep{tibshirani2019R_grf} and \textit{uplift} \citep{guelman2014R_uplift} for the random forests implementations. We also use the \textbf{R} library \textit{tools4uplift} \citep{belba2020R_tools4uplift} to fit the Qini-based uplift regression \citep{belba2019qbased} and for computing the models goodness-of-fit measure. Note that the {\it Logistic}, \textit{TO$_{\hat{\alpha}}$} and \textit{IE} methods were implemented with \textbf{Pytorch} in \textbf{Python}.
\subsection{Fine-tuning the trade-off constant $\alpha$}
We start our simulation study by analyzing the impact of the trade-off constant $\alpha$ on SMITE$_{\mathrm{TO}}$ performance. Recall the objective function related to the SMITE$_{\mathrm{TO}}$ model
$$ (1 - \alpha) \ell(\theta_o,\boldsymbol{\theta}) + \alpha \mathcal{J}(\theta_o,\boldsymbol{\theta}), $$
with $\alpha \in [0;1)$. In this experiment, we want to evaluate the impact of $\alpha$ on the predictive performance of the model.
To do this, for each run, we fix $\alpha$ and fine-tune the learning rate. We choose the model which gives the best $\hat{q}_{\mathrm{adj}}$ on the validation set. Next, we score the test samples and report the $\hat{q}_{\mathrm{adj}}$ for the test set. We repeat this for each $\alpha$ value between $0$ and $1$, with increments of $0.05$. The reported results are averaged over $20$ runs.
Looking at the curves in Figure~\ref{fig:alpha_scene}, we see that the trade-off constant gives flexibility to the model to adapt to the underlying data generating process. We advise to consider $\alpha$ as a hyper-parameter and to select the optimal $\hat\alpha$ by cross-validation.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.3\textwidth]{figures/simulations/Alpha_Scene1.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Alpha_Scene2.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Alpha_Scene3.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Alpha_Scene4.pdf}
\includegraphics[width=0.3\textwidth]{figures/simulations/Alpha_Scene5.pdf}
\caption{Average adjusted Qini coefficient as a function of the trade-off constant $\alpha$ for SMITE$_{\mathrm{TO}}$.}
\label{fig:alpha_scene}
\end{figure}
\subsection{Parameters estimation}
In this part of the analysis, we focus on the simple model
\begin{equation*}
\mu_{1t_i}(\mathbf x_i, \theta_o, \boldsymbol{\theta}) =
\bigl( 1+ \mathrm{exp}\{-(\theta_o + \gamma t_i + \mathbf{x}_i^\top [\boldsymbol{\beta} + t_i \boldsymbol{\delta}] ) \} \bigr)^{-1},
\end{equation*}
for $i=1,\ldots n$, where $\boldsymbol{\theta} = (\boldsymbol{\beta}, \gamma, \boldsymbol{\delta})$, denotes all model parameters except for the intercept $\theta_o$. We compare $4$ different methods to estimate the parameters, that is, with \textit{Logistic, Qini-based, TO$_{\hat{\alpha}}$}, and {\it IE}. Table~\ref{tab:param_estimation_results} shows the averaged results on the out-of-sample dataset.
\begin{table}[!ht]
\centering
\caption{Comparison of parameters estimation. The metric used is the adjusted Qini coefficient.}
\label{tab:param_estimation_results}
\begin{tabular}{l|ccccc}
Scenarios & 1 & 2 & 3 & 4 & 5\\
\hline
{\it Logistic} & $0.75$ & $1.80$ & $7.50$ & $0.98$ & $2.59$\\
{\it Qini-based} & $0.82$ & $\bf2.68$ & $8.36$ & $1.09$ & $2.94$\\
{\it TO$_{\hat{\alpha},(2p+1)}$} & $0.96$ & $2.06$ & $\bf8.84$ & $\bf1.22$ & $3.05$\\
{\it IE$_{(2p+1)}$} & $\bf1.24$ & $2.32$ & $8.41$ & $1.20$ & $\bf3.12$\\
\hline
\end{tabular}
\end{table}
If we pick “winners” in each of the simulation scenarios based on which
method has the highest $\hat{q}_{\mathrm{adj}}$, {\it IE} would win Scenarios
1 and 5, tying with the regularized transformed outcome model {\it TO$_{\hat{\alpha}}$} in Scenario 4. The {\it TO$_{\hat{\alpha}}$} model would win Scenario 3, for which the generated data was purely guided by the treatment effect function and the optimal trade-off constant $\alpha$ is close to $1$ as seen in Figure~\ref{fig:alpha_scene}. In Scenario 2, the {\it Qini-based} model has the highest $\hat{q}_{\mathrm{adj}}$ while the {\it TO$_{\hat{\alpha}}$} model performance is lower than the one of {\it IE}. In general all of the methods outperform the {\it Logistic} model.
\subsection{Beyond linearity}
Ensemble methods are amongst the most powerful for the uplift case \citep{soltys2015ensemble}. A method that has proven to be very effective in estimating individual treatment effects is one based on generalized random forests \citep{athey2019generalized}. This method uses honest estimation, i.e., it does not use the same information for the partition of the covariates space and for the estimation of the uplift. This has the effect of reducing overfitting. Another candidate that we consider in our experiments is also based on random forests and designed for uplift \citep{jaskowski2012uplift}. For this method, we consider two different split criteria, one based on Kullback-Leibler divergence and one based on Euclidean distance. In Table~\ref{tab:beyond_linearity_results}, we compare these methods to our neural-network version of TO$_{\hat{\alpha}}$ and {\it IE} models.
\begin{table}[!ht]
\centering
\caption{Comparison of non-linear models. The metric used is the adjusted Qini coefficient.}
\label{tab:beyond_linearity_results}
\begin{tabular}{l|ccccc}
Scenarios & 1 & 2 & 3 & 4 & 5\\
\hline
{\it TO$_{\hat{\alpha},(p+1)\times(2p+1)}$} & $1.59$ & $2.01$ & $9.09$ & $1.23$ & $3.01$ \\
{\it IE$_{(p+1)\times(2p+1)}$} & $\bf 1.63$ & $\bf 2.67$ & $\bf 9.86$ & $\bf 1.30$ & $\bf 3.14$\\
{\it CF} & $0.75$ & $2.22$ & $9.42$ & $0.94$ & $2.79$ \\
{\it CF$_{h}$} & $0.75$ & $2.51$ & $9.53$ & $1.14$ & $3.07$\\
{\it URF$_{\mathrm{KL}}$} & $0.74$ & $2.52$ & $8.92$ & $1.01$& $1.99$ \\
{\it URF$_{\mathrm{ED}}$} & $0.68$ & $2.42$ & $9.18$ & $0.99$ & $2.33$ \\
\hline
\end{tabular}
\end{table}
We observe that the random forests tend to overfit for several scenarios, with the exception of the honesty criteria which seems to mitigate this problem. On the other hand, we observe a good adequacy of SMITE models in all scenarios. Overall, the results of {\it IE} appear to outperform all other methods significantly (based on Wilcoxon signed-rank test \citep{wilcoxon1992individual}).
\subsection{Impact of the number of parameters on uplift prediction}
Neural networks have two main hyper-parameters that control the architecture or topology of the network: the number of hidden layers and the number of nodes in each hidden layer. In the previous simulations, we fixed the number of hidden layers to $1$ and the number of nodes in the hidden layer to $2p+1$ as per Theorem~\ref{theorem}. However, the number of nodes can be seen as a hyper-parameter as well. Here, we vary the number of nodes to assess the impact on predictive performance, using either $m=64, 128, 256$ or $m=512$ nodes. In Table~\ref{tab:hidden_neurons_results}, we report the averaged results. As we can see, there is not one optimal value of $m$ among the ones we compare that gives the best model for all scenarios. Depending on the scenario, we observe clear differences in terms of performance. However, between {\it TO} and {\it IE}, there are some similarities. For scenario 1, in which $p = 200$, we see that the best model is reached for $m = 512$; for scenario 3, the models perform best with $m = 64$, and finally with $m = 256$ and $m = 128$ for scenarios 4 and 5. Only scenario 2 gives two different values of $m$ for {\it TO} and {\it IE}, that is, $m = 128$ and $m = 256$ respectively. Therefore, in practice, for a specific dataset, it is important to consider the number of neurons in the hidden layer as a hyper-parameter and take the time to choose the optimal value by cross-validation.
\begin{table}[!ht]
\centering
\caption{Fine-tuning the number of nodes. The metric used is the adjusted Qini coefficient.}
\label{tab:hidden_neurons_results}
\begin{tabular}{l|l|ccccc}
Scenarios &$m$ & 1 & 2 & 3 & 4 & 5\\
\hline
{\it TO$_{\hat\alpha,(p+1)\times m}$}
& $64$ & $1.49$ & $2.01$ & $\bf 9.23$ & $1.16$ & $2.90$\\
& $128$ & $1.55$ & $\bf 2.08$ & $9.17$ & $1.20$ & $\bf 3.11$\\
& $256$ & $1.54$ & $2.03$ & $9.19$ & $\bf 1.23$ & $3.02$ \\
& $512$ & $\bf 1.61$ & $2.01$ & $9.05$ & $1.10$ & $2.98$\\
\hline
{\it IE$_{(p+1)\times m}$}
& $64$ & $1.54$ & $2.61$ & $\bf 9.91$ & $1.22$ & $3.10$ \\
& $128$ & $1.56$ & $2.54$ & $9.80$ & $1.28$ & $\bf 3.16$\\
& $256$ & $1.57$ & $\bf 2.65$ & $9.78$ & $\bf 1.32$ & $3.13$ \\
& $512$ & $\bf 1.60$ & $2.49$ & $9.64$ & $1.18$ & $3.15$\\
\hline
\end{tabular}
\end{table}
As stated before, the number of hidden layers can also be seen as a hyper-parameter for the neural network architecture. In our case, we want to evaluate the performance of our models when using a $2$-hidden layers neural network. We fix the number of hidden neurons to $p+1$ and $p$ for the $2$ hidden layers respectively. We report the averaged results as well as a summary of all previous experiments in Table~\ref{tab:hidden_layers_results}. For random forests uplift models, we report the best score between the four used methods. We see that for the most complex scenario, that is, scenario 5, adding a second hidden layer improves the prediction performance of both models {\it TO$_{\hat\alpha}$} and {\it IE} achieving $\hat{q}_{\mathrm{adj}}$ of $3.12$ and $3.29$ respectively, both outperforming the {\it Best RF} result, that is a $\hat{q}_{\mathrm{adj}}$ of $3.07$. For scenarios 1, 3 and 4, the best results are attained with the $1$-hidden layer {\it IE} model. Finaly, for scenario $2$, the {\it Qini-based} uplift regression achieves the highest $\hat{q}_{\mathrm{adj}}$, tying with the {\it IE} model (i.e. $2.68$ vs. $2.67$).
\begin{table}[!ht]
\centering
\caption{Summary. The metric used is the adjusted Qini coefficient.}
\label{tab:hidden_layers_results}
\begin{tabular}{l|l|ccccc}
Scenarios &$h$ & 1 & 2 & 3 & 4 & 5\\
\hline
{\it TO$_{\hat\alpha}$}& $0$ & $0.96$ & $2.06$ & $8.84$ & $1.22$ & $3.05$\\
& $1$ & $1.61$ & $2.08$ & $9.23$ & $1.23$ & $3.11$ \\
& $2$ & $1.44$ & $2.03$ & $8.99$ & $1.19$ & $3.12$ \\
\hline
{\it IE} & $0$ & $1.24$ & $2.32$ & $8.41$ & $1.20$ & $3.12$\\
& $1$ & $\bf 1.63$ & $2.67$ & $\bf 9.91$ & $\bf 1.32$ & $3.16$\\
& $2$ & $1.49$ & $2.36$ & $9.12$ & $1.28$ & $\bf 3.29$ \\
\hline
{\it Logistic} & - & $0.75$ & $1.80$ & $7.50$ & $0.98$ & $2.59$\\
{\it Qini-based} & - & $0.82$ & $\bf 2.68$ & $8.36$ & $1.09$ & $2.94$\\
Best {\it RF} & - & $0.75$ & $2.52$ & $9.53$ & $ 1.14$ & $3.07$\\
\hline
\end{tabular}
\end{table}
\newpage
\section{Application}
\label{sec:application}
A data set with $n=50,000$ customers and $p=40$ covariates from an insurer is at our disposal to evaluate the performance of our methods. This company was interested in designing strategies to maximize its conversion rate. An experimental acquisition campaign was implemented for 6 months, for which half of the potential clients were randomly allocated into a treatment group and the other half into a control group. Potential clients under the treatment group were contacted using a different (and more costly) strategy. The goal of the analysis was to be able to select who among the clients portfolio should be contacted in that way going forward in order to maximize the benefits. The observed difference in sales rates between the treated group and the control shows some evidence of a positive impact of the marketing initiative ($\mathrm{ATE}=0.55\%$).
We fit the SMITE models using {\it TO$_{\alpha}$} and {\it IE} losses to the data using the methodology described in the previous sections. We compare models with $0,1$ and $2$ hidden layers. For comparison purposes, we also considered the honest causal forest.
We split the data into training ($40\%$), validation ($30\%$), and test samples ($30\%$). We use the training and the validation sets to fit and fine-tune the models. Then, we select the best model based on the validation set. Using the best model, we score the samples from the test set and compute the adjusted Qini coefficient $\hat{q}_{\mathrm{adj}}$. We repeat the experiment $30$ times and report averaged results as well as their standard-errors in Table~\ref{tab:final_models_results}.
\begin{table}[!ht]
\centering
\caption{Final models performance: averaged $\hat{q}_{\mathrm{adj}}$ (S.E) over $30$ runs.}
\label{tab:final_models_results}
\begin{tabular}{l|cccc}
Method & $\hat{\alpha}$ & Training & Validation & Test \\
\hline
{\it TO$_{\hat{\alpha},(81)}$} & $0.5$ & $0.251~(0.028)$ & $0.112~(0.026)$ & $0.037~(0.011)$\\
{\it IE$_{(81)}$} & - & $0.331~(0.029)$ & $0.115~(0.022)$ & $0.040~(0.013)$\\
\hline
{\it TO$_{\hat{\alpha},(41)\times(81)}$} & $0.3$ & $0.539~(0.042)$ & $0.181~(0.028)$ & $\bf 0.104~(0.024) $ \\
{\it IE$_{(41)\times(81)}$} & - & $0.705~(0.043)$ & $0.210~(0.029)$ & $0.095~(0.018)$\\
\hline
{\it TO$_{\hat{\alpha},(41)\times(41) + (41)\times(40)}$} & $0.3$ & $0.928~(0.066)$ & $0.221~(0.032)$ & $0.095~(0.026)$ \\
{\it IE$_{(41)\times(41) + (41)\times(40)}$} & - & $\bf 0.944~(0.065)$ & $0.254~(0.034)$ & $0.091~(0.021)$ \\
\hline
{\it CF$_{h}$} & - & $0.873~(0.028)$ & $\bf 0.257~(0.022)$ & $0.019~(0.007)$ \\
\hline
\end{tabular}
\end{table}
In these kinds of marketing initiatives, it makes sense that the overall impact is close to zero. Indeed, customers are already interested in the product because it is mandatory to have insurance for their car. Also, the effect of the call is positive on average, that is to say that during the call, the advisor allows the customer to understand all the details of his coverage and thus reassure him about the product he buys. Thus, it is rather unlikely that the call could have a negative effect. We therefore think that for this data set, we have a situation similar to Scenario 1 of the simulations. This may explain why the causal forest does not produce good predictions here. On the other hand, both {\it IE} and {\it TO$_{\hat\alpha}$} work well on this data set. The $1$-hidden layer neural network version of the models improves significantly the fit for both loss functions. Adding a second layer does not seem to increase significantly the performance of the models.
\section{Conclusion}
\label{sec:conclusion}
\bibliographystyle{plainnat}
|
\section*{Nomenclature}
\addcontentsline{toc}{section}{Nomenclature}
\begin{IEEEdescription}[\IEEEusemathlabelsep\IEEEsetlabelwidth{$V_1,V_2,V_3$}]
\item[BCSE] Branch current state estimation
\item[DSSE] Distribution system state estimation
\item[MPE] Mean percentage error
\item[PCP] Principle component pursuit
\item[PCA] Principle component analysis
\item[SM] Smart meter
\item[WLS] Weighted least squares
\item[$G$] Gain matrix
\item[$H$] Jacobian matrix
\item[$h_i$] Measurement function that maps state values to the measurement variable $i$
\item[$\pmb{I_{re}},\pmb{I_{im}}$] Current real and imaginary values for all the branches
\item[$J$] Sum of squared residuals
\item[$L$] Weight parameter for penalizing deviations from SM measurements
\item[$M_U$] Voltage observation matrix
\item[$M_U^*$] Refined post-mitigation voltage matrix
\item[$M_P$] Nodal active power injection matrix
\item[$M_P^*$] Refined post-mitigation active power matrix
\item[$M_Q$] Nodal reactive power injection matrix
\item[$M_Q^*$] Refined post-mitigation reactive power matrix
\item[$M_{MV}$] Synchronized sensor measurements
\item[$M_{z}$] Measurement vector
\item[$M_{PS}$] Pseudo measurements
\item[$P_i(t_j)$] Measured active power at node $i$ at time $t_j$
\item[$Q_i(t_j)$] Measured reactive power at node $i$ at time $t_j$
\item[$R$] Branch resistance matrix of the system
\item[$U_0$] Squared voltage magnitude of substation
\item[$U_i(t_j)$] Measured voltage magnitude squared at node $i$ at time $t_j$
\item[$W$] Weight matrix
\item[$X$] Branch reactance matrix of the system
\item[$\pmb{x_s}$] System state vector
\item[$Y_M, Y_S$] Interim matrices using the latest solution updates
\item[$Z_M, Z_S$] Interim matrices using the full history of the solution trajectory
\item[$\alpha,\beta$] Auxiliary matrices
\item[$\Delta S_U$] Asynchrony voltage error matrix
\item[$\Delta E_U$] Voltage measurement error matrix
\item[$\Delta S_P$] Asynchrony active power error matrix
\item[$\Delta E_P$] Active power measurement error matrix
\item[$\Delta S_Q$] Asynchrony reactive power error matrix
\item[$\Delta E_Q$] Reactive power measurement error matrix
\item[$\delta_U, \delta_P, \delta_Q$] Standard deviations of voltage, active power, and reactive power measurement errors
\item[$\Gamma(\cdot,\cdot)$] Differentiable function for low rank matrices
\item[$\Gamma_T$] Total approximate sparsity norm for all the SM datasets
\item[$||\cdot||_{*}$] Nuclear norm operation
\item[$||\cdot||_{1}$] 1-norm operation
\item[$||\cdot||_{F}$] Frobenius norm operation
\item[$<\cdot,\cdot>$] Frobenius inner product
\item[$\lambda_U, \lambda_P, \lambda_Q$] Balanced parameters for voltage, active power, and reactive power measurements
\item[$\mu_U,\nu_U$] Smoothness parameter
\item[$\omega_1, \omega_2, \omega_3$] Non-negative weights
\item[$\Psi(\cdot,\cdot)$] Differentiable function for sparse error matrices
\item[$\sigma_i^2$] Error variance of sensor $i$
\item[$\tau(\cdot,\cdot)$] Aggregate gradient factor
\item[$\zeta_j(A)$] $j$'th singular value of an arbitrary matrix $A$
\end{IEEEdescription}
\IEEEpeerreviewmaketitle
\section{Introduction}
The wide-scale deployment of smart meters (SMs) provides a unique opportunity for utilities to enhance their situational awareness capabilities in distribution grids. By 2018, more than 150 million customers across the U.S. were equipped with SMs \cite{EIA}. On the other hand, SMs are commonly counted among low-quality sensors. Specifically, SMs are asynchronous due to mismatching in sampling time among sensors in the grid, which can limit their applicability in real-time system monitoring \cite{Lin2019}.
Most previous works on distribution grid state estimation have assumed that SMs are perfectly time-aligned \cite{Primadianto2017,Deh2018}. Only few works have studied the impact of time misalignment and asynchrony of various sensors on grid monitoring and situational awareness: In \cite{Antonios2019,Schwefel2018}, the statistical characteristics of time misalignment in distribution grid sensors have been estimated using Markov-modulated models. In \cite{Lin2019}, exponential load variation trends are exploited for developing confidence intervals for SM data samples in distribution system state estimation (DSSE) to compensate for time delays and asynchrony. In \cite{Carvaro2019}, a dynamic DSSE formulation is proposed for multitude of asynchronous sensors, which has proven bounded estimation errors. In \cite{Bolognani2015,Alimardani2015}, meter clock synchronization errors are captured through Gaussian probability distributions and represented in DSSE. This idea was also applied in \cite{Ni2016} to model measurement errors in grid monitoring. Most solutions proposed for mitigating SM data quality issues rely on \textit{a priori} knowledge of error distribution structure and parameters, which can be difficult to acquire due to information scarcity.
In this paper, we propose a SM data recovery technique that is capable of mitigating the impact of asynchrony error in grid monitoring. Our method has three novel features: (1) We have noted that a rise in SM asynchrony results in a loss of mutual temporal correlation in their time-series data streams. This loss of temporal correlation can be translated into an increase in the rank of \textit{observation matrices}, which store the measurement data from multiple SMs. Thus, we propose to cast the asynchrony error mitigation problem as a low rank matrix recovery process. For this purpose, we have leveraged principle component pursuit (PCP) techniques \cite{Bouwmans2014,Rodriguez2013}. PCP employs data-centric optimization for decomposing SM datasets to identify and separate asynchrony error term from raw data. The main idea is that by manipulating the SM data and reducing the rank of the observation matrices, we will enhance the temporal correlation among the SMs which rolls back the adverse impact of asynchrony. (2) In addition to asynchronous errors, SM data has measurement errors that result from the imprecision (i.e., noise) of the measuring devices. Typically, SMs have a relative measurement error of about 1$\%$. Further, unlike image datasets, synchronous SM measurements and asynchronous errors cannot be exactly low rank and exactly sparse. These data properties hinder the applications of state-of-the-art low rank data recovery methods to deal with SM asynchrony errors, such as robust principal component analysis (PCA) \cite{Zhou2010}. To deal with these problems, we utilize a relaxation to PCP that introduces an entry-wise noise term to represent SM measurement errors in the objective function and eliminate rank-1 constraints. (3) SMs are multi-modal, meaning that they can measure several different variables, including nodal voltage magnitude and nodal average active power (plus nodal reactive power, in some cases.) To mitigate the impact of sensor asynchrony, data recovery needs to be conducted over all measurement datasets simultaneously. However, since these multi-modal datasets are inherently interdependent due to the grid physics, a coordination scheme is required to revise all the datasets while capturing their dependencies. To achieve this, we propose a new multi-objective data recovery formulation that refines voltage magnitude, active/reactive power measurements (and pseudo-measurements), concurrently. The dependencies among these datasets are captured via approximate DistFlow-based constraints \cite{Gilbert1998,Baran1989}. We have developed a Nesterov-based technique to solve the PCP-based multi-objective optimization for recovering multiple SM datasets \cite{Aybat2011}.
The main contributions of this paper are summarized as follows:
\begin{itemize}
\item An important observation from real data is presented: asynchrony results in loss of temporal correlation among neighboring SMs. This observation can be quantified using the rank of the nodal voltage observation matrix.
\item A novel low rank-based data recovery method is developed to fully mitigate asynchronization error in grid monitoring based on our observation.
\item The proposed method considers various specific properties of SM data for enhancing the quality of the recovered data and ensure consistency with grid physics: 1) SMs can measure several different asynchronous variables; 2) SM measurements are statistically interdependent; 3) small entry-wise measurement errors exist within SM measurements.
\item Our method handles SM asynchrony issue without needing high-resolution reference sensors, such as micro-PMUs, which are unavailable in most practical distribution systems.
\item The proposed solution has been tested using real SM data and feeder models to verify its performance.
\end{itemize}
The rest of the paper is constructed as follows: Section \ref{sec:MO} presents the proposed multi-objective data recovery method and our approximate first-order solution; Section \ref{sec:dsse} demonstrates the application of data recovery in grid monitoring; Section \ref{sec:result} analyzes numerical results and verification of the proposed models; finally, Section \ref{sec:con} presents the paper conclusions.
\section{Multi-Objective SM Data Recovery Strategy}\label{sec:MO}
In this section, we lay out our data recovery solution for mitigating the errors caused by the asynchronous nature of SMs in distribution grids. This includes key ideas in developing a multi-objective optimization formulation, along with an approximate first-order algorithm to solve the model.
\subsection{Rationale}
The available data from SMs can be organized into \textit{observation matrices}. These matrices capture the time-series measurements of several sensors within a given time window $[t_1,t_m]$. For example, the voltage observation matrix is as follows:
\begin{equation}
\label{eq:MU}
M_U = \left[
\begin{array}{ccc}
U_1(t_1) & \cdots & U_N(t_1)\\
\vdots & \ddots & \vdots\\
U_1(t_m) & \cdots & U_N(t_m)
\end{array}
\right]
\end{equation}
where, $U_i(t_j)$ is the measured voltage magnitude squared at node $i$ and at time $t_j$. Note that each column of $M_U$ corresponds to an SM. The observation matrices can be constructed at feeder-, lateral-, or service transformer-levels.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{rank_v.png}
\caption{Rank increase in $M_U$ due to SM asynchrony.}
\label{fig:rank}
\end{figure}
Our PCP-based data recovery model is based on a key observation from real data: asynchrony among SMs leads to an increase in the rank of $M_U$. The increase in rank is caused by loss of temporal correlation among SMs, which translates into a decrease in statistical correlations in columns of $M_U$ (i.e., the columns lose linear dependency.) This observation can be backed-up by numerical experiments, as shown in Fig. \ref{fig:rank}. This figure shows the average rank of $M_U$ at various time windows (measured for a grid lateral) as a function of strength of SM asynchrony (measured in terms of variance of time misalignment distribution.) As is observed, the rank of the observation matrix increases as the SM asynchrony intensifies. Note that this observation can be found on the data from SMs with diverse resolutions, including 15 minutes, 30 minutes, and 60 minutes.
To fully capture and mitigate the impact of SM asynchrony, similar observation matrices can be defined for nodal active and reactive power injection measurements, denoted as $M_P$ and $M_Q$, respectively:
\begin{equation}
\label{eq:MP}
M_P = \left[
\begin{array}{ccc}
P_1(t_1) & \cdots & P_N(t_1)\\
\vdots & \ddots & \vdots\\
P_1(t_m) & \cdots & P_N(t_m)
\end{array}
\right]
\end{equation}
\begin{equation}
\label{eq:MQ}
M_Q = \left[
\begin{array}{ccc}
Q_1(t_1) & \cdots & Q_N(t_1)\\
\vdots & \ddots & \vdots\\
Q_1(t_m) & \cdots & Q_N(t_m)
\end{array}
\right]
\end{equation}
where, $P_i(t_j)$ and $Q_i(t_j)$ are active and reactive power measurements at node $i$ and time $t_j$, respectively. Note that in general SMs are capable of measuring both average active and reactive powers. However, in many cases, this function is not activated for residential sensors. Thus, in case the reactive power data is unavailable, \textit{pseudo-measurements} can be applied instead to construct an approximate $M_Q$. Note that our method is robust to gross sparse errors, thus, it can handle the uncertainty of pseudo-measurements and low quality data.
\subsection{Data Recovery Model}
The main component of asynchrony error mitigation is to compensate for the loss of temporal correlation among SMs. Since this loss can be detected via the changes in the ranks of the observation matrices, asynchrony error mitigation can be written as a low rank matrix recovery model. To consider both asynchrony errors and small entry-wise measurement errors in SM data, our data recovery approach models an observation matrix (i.e., asynchrony voltage magnitude matrix) as the summation of three components: a low rank voltage magnitude matrix, an asynchrony error matrix, and a measurement error matrix. The goal is to identify unknown voltage magnitude matrix and asynchrony error matrix within the datasets in the presence of entry-wise noise. The model is shown below:
\begin{equation}
\label{eq:modelU}
M_U = M_U^* + \Delta S_U + \Delta E_U
\end{equation}
where, $M_U^*$ represents the \textit{refined post-mitigation} voltage magnitude matrix which has a low rank, $\Delta S_U$ is the asynchrony error matrix, and $\Delta E_U$ represents entry-wise measurement errors. It should be noted that measurement error is different from asynchrony error, as mentioned in previous work \cite{Alimardani2015}. The same representation applies to both active and reactive measurements and pseudo-measurements, as follows:
\begin{equation}
\label{eq:modelP}
M_P = M_P^* + \Delta S_P + \Delta E_P
\end{equation}
\begin{equation}
\label{eq:modelQ}
M_Q = M_Q^* + \Delta S_Q + \Delta E_Q
\end{equation}
where, the sub-components are defined similar to \eqref{eq:modelU}. The objective of the data recovery process is to revise the SM data in a way that the ranks of observation matrices are minimized (i.e., temporal correlations among SMs are maximized), while the extent of changes made in the original data is kept at a minimum level. This goal can be represented using three objective functions, corresponding to the available datasets, $M_U$, $M_P$, and $M_Q$, as follows:
\begin{equation}
\label{eq:all_obj}
\begin{cases}
f_U = ||M_U^*||_{*} + \lambda_U||\Delta S_U||_1\\
f_P = ||M_P^*||_{*} + \lambda_P||\Delta S_P||_1\\
f_Q = ||M_Q^*||_{*} + \lambda_Q||\Delta S_Q||_1
\end{cases}
\end{equation}
where, $||\cdot||_{*}$ and $||\cdot||_1$ are the nuclear norm and 1-norm (i.e., sparsity norm) operations, respectively. These norms are calculated as follows \cite{Phillips1996}:
\begin{equation}
\label{eq:nuc}
||A||_{*} = \sum_j\zeta_j(A)
\end{equation}
\begin{equation}
\label{eq:1-n}
||A||_1 = \max_j\sum_j|A(i,j)|
\end{equation}
where, $\zeta_j(A)$ denotes the $j$'th singular value of an arbitrary matrix $A$. Further, $\lambda_U$, $\lambda_P$, and $\lambda_Q$ are tunable parameters that are leveraged to balance out the two competing components of the objective functions: minimizing the rank of the recovered data versus the amount of changes made in the data during the recovery process. Mathematically, this means that by minimizing $f_U$, $f_P$, and $f_Q$, the data recovery process effectively minimizes the ranks of $M_U^*$, $M_P^*$, and $M_Q^*$. At the same time, the changes made in the data are kept small by penalizing the sparsity norm of matrices $\Delta S_U$, $\Delta S_P$, and $\Delta S_Q$. The three objectives $f_U$, $f_P$, and $f_Q$ are evaluated over the datasets that are generated by the same system (e.g., same feeder, lateral, or service transformer). However, these three datasets are not independent from each other due to the power flow constraints. Thus, the re-calibration of these three datasets cannot be performed separately using conventional low rank data recovery methods, such as robust PCA and PCP. To address this problem, we propose a multi-objective PCP-based model that can jointly refine three the SM datasets. The objective function minimizes the ranks of recovered data to realize the best achievable SM re-alignment. Moreover, to incorporate the inherent interdependencies of the three objectives, power flow equations are added as the constraints of the model. The proposed multi-objective optimization is as follows:
\begin{equation}
\label{eq:obj}
\min_{M_U^*,M_P^*,M_Q^*}\ \{f_U,f_P,f_Q\}
\end{equation}
\begin{equation}
\label{eq:constU}
s.t.\ \ ||M_U - M_U^* - \Delta S_U ||_F \leq \delta_U
\end{equation}
\begin{equation}
\label{eq:constP}
||M_P - M_P^* - \Delta S_P||_F \leq \delta_P
\end{equation}
\begin{equation}
\label{eq:constQ}
||M_Q - M_Q^* - \Delta S_Q ||_F \leq \delta_Q
\end{equation}
\begin{equation}
\label{eq:DistFlow}
M_U^* = M_P^* \cdot R^T + M_Q^* \cdot X^T + 1_{m \cross N}U_0
\end{equation}
where, $||\cdot||_F$ denotes the Frobenius norm of matrix, defined as follows:
\begin{equation}
\label{eq:frob}
||A||_F = \sqrt{\sum_i\sum_j A(i,j)}
\end{equation}
In addition, parameters $\delta_U$, $\delta_P$, and $\delta_Q$ are the standard deviations of the measurement/pseudo-measurement errors (obtained using knowledge of sensor tolerance or pseudo-measurement confidence intervals), matrices $R$ and $X$ represent the branch resistance and branch reactance of the network, respectively \cite{Qu2020}. $U_0$ is the primary voltage magnitude squared for the transformer to which the SMs are connected. The rationale behind constraints \eqref{eq:constU}, \eqref{eq:constP}, and \eqref{eq:constQ} is that the refined components (i.e., $M_U^*$, $M_P^*$, $M_Q^*$) are not exactly low rank and the asynchrony error components (i.e., $\Delta S_U$, $\Delta S_P$, $\Delta S_Q$) are not exactly sparse. Such soft constraints allow for slight deviations in the recovered data to compensate for SM measurement errors, which are consistent with our knowledge of measurement device confidence levels. Also, these allow utilities to minimize asynchrony error with noisy practical SM data, which particularly pertains to reactive power data that may be unavailable for residential customers. Constraint \eqref{eq:DistFlow} is obtained from the linear DistFlow in matrix form \cite{Baran1989}, which can enforce network physics and capture the inherent dependencies among datasets. The goal of this constraint is to ensure that the recovered SM data is feasible in power engineering context.
Our method follows the line of low rank data recovery techniques that have been commonly used in many areas \cite{PCAreview}. Unlike the black box methods that lack interpretability, the proposed model has a solid mathematical foundation to recover a low rank SM data matrix in the presence of gross asynchrony errors. Also, the dependencies among the datasets are basically encoded into the solution through a set of linear equality constraints. Such power flow models can be applied for arbitrary distribution systems. Note that the model is extendable to unbalanced systems in a straightforward way (i.e., full three-phase DistFlow is leveraged). Further, the proposed data recovery model makes no assumptions on system topology or load distribution, which ensures the performance of this model in other distribution systems.
\subsection{Solution Strategy}
A major challenge in solving the proposed data recovery model is the existence of power flow constraints \eqref{eq:DistFlow} that hinders the application of the existing closed-form dual solvers \cite{Zhou2010}. Another complication is that \eqref{eq:obj} has three non-smooth objective functions, which makes the problem non-differentiable. To efficiently tackle these challenges, we present a first-order Nesterov-like algorithm to solve the proposed multi-objective data recovery framework \cite{Nesterov2005}. The basic idea of our solution is to approximate the non-smooth objectives with differentiable surrogates. By applying this idea, the following surrogate components can be written for $f_U$ \cite{Aybat2011}:
\begin{equation}
\label{eq:MUS}
||M_U^{*}||_{*} \approx \Gamma(M_U^{*},\mu_U) = \max_{||\alpha||_2 \leq 1} <M_U^{*},\alpha> - \frac{\mu_U}{2}||\alpha||^2_{F}
\end{equation}
\begin{equation}
\label{eq:DSS}
||\Delta S_U||_{1} \approx \Psi(\Delta S_U,\nu_U) = \max_{||\beta||_{\infty} \leq 1} <\Delta S_U,\beta> - \frac{\nu_U}{2}||\beta||^2_{F}
\end{equation}
where, $\alpha$ and $\beta$ are auxiliary matrices, $\mu_U$ and $\nu_U$ are smoothness parameters, and $<\cdot,\cdot>$ is the Frobenius inner product \cite{Phillips1996}, calculated as follows:
\begin{equation}
\label{eq:frobinner}
<A,B> = \sum_i\sum_j A(i,j)\cdot B(i,j)
\end{equation}
Note that the non-differentiable norms are replaced with differentiable functions $\Gamma(\cdot,\cdot)$ and $\Psi(\cdot,\cdot)$ in \eqref{eq:MUS} and \eqref{eq:DSS}. The Lipschitz constants for the gradients of $\Gamma(\cdot,\cdot)$ and $\Psi(\cdot,\cdot)$ equal $\frac{1}{\mu_U}$ and $\frac{1}{\nu_U}$, respectively. Similar smooth surrogates are defined and calculated for the objectives $f_P$ and $f_Q$. By adopting this approximate alternative, the objectives in optimization \eqref{eq:obj} can be rewritten as a single-objective weighted averaging process by using a scalarization method \cite{MOO}. Since the relaxed problem is convex, the single-objective formulation is guaranteed to track all the Pareto-optimal solutions, given valid weight assignment to the objectives \cite{Deb2001}. The single-objective formulation can be rearranged as follows:
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{overall.pdf}
\caption{Overall structure of the solution for grid monitoring.}
\label{fig:monitor}
\end{figure*}
\begin{equation}
\label{eq:weightedsum}
\min_{M_U^*,M_P^*,M_Q^*}\ \Gamma_T(M_U^*,M_P^*,M_Q^*) + \Psi_T(\Delta S_U,\Delta S_P,\Delta S_Q)
\end{equation}
\begin{equation}
\label{eq:constraints1}
s.t.\ \ \eqref{eq:constU}-\eqref{eq:DistFlow}
\end{equation}
here, the new objective function consists of two component: (I) $\Gamma_T$ quantifies the total approximate nuclear norm for all the SM datasets:
\begin{equation}
\label{eq:GammaT}
\begin{split}
&\Gamma(M_U^*,M_P^*,M_Q^*) =\\
&\omega_1\Gamma(M_U^{*},\mu_U) + \omega_2\Gamma(M_P^{*},\mu_P) + \omega_3\Gamma(M_Q^{*},\mu_Q)
\end{split}
\end{equation}
where, $\omega_1$, $\omega_2$, and $\omega_3$ are the non-negative user-defined weights assigned to $f_U$, $f_P$, and $f_Q$, respectively. Assigning a larger weight to the objective function indicates that the function has a higher priority compared to a function with a smaller weight. Further, $ \omega_1 + \omega_2 + \omega_3 = 1$ needs to hold to ensure Pareto-optimality. (II) $\Gamma_T$ is the total approximate sparsity norm for all the SM datasets, as follows:
\begin{equation}
\label{eq:PsiT}
\begin{split}
&\Psi(\Delta S_U,\Delta S_P,\Delta S_Q) =\\
&\omega_1\lambda_U\Psi(\Delta S_U,\nu_U) + \omega_2\lambda_P\Psi(\Delta S_P,\nu_P) + \omega_3\lambda_Q\Psi(\Delta S_Q,\nu_Q)
\end{split}
\end{equation}
This new data recovery formulation \eqref{eq:weightedsum} is both convex and differentiable. Given the new model, the Nesterov algorithm entails the following steps to solve SM data recovery problem:
\noindent\textbf{Step I -} \textit{Initialization}: $k\leftarrow 0$ (counter initialization); $M_U^*(0)\leftarrow M_U$, $M_P^*(0)\leftarrow M_P$, $M_Q^*(0) \leftarrow M_Q$, $\Delta S_U \leftarrow 0_{m\times N}$, $\Delta S_P\leftarrow 0_{m\times N}$, and $\Delta S_Q\leftarrow 0_{m\times N}$ (solution initialization).
\noindent\textbf{Step II -} \textit{Component-wise gradient calculation}: Obtain the gradients of components \eqref{eq:MUS} and \eqref{eq:DSS} for all the objective functions in the data recovery problem. As shown in \cite{Nesterov2005}, these gradients can be computed as follows:
\begin{equation}
\label{eq:g_MUS}
\nabla\Gamma(M_U^{*}(k),\mu_U) = \alpha^{*}(\mu_U)
\end{equation}
\begin{equation}
\label{eq:g_DSS}
\nabla\Psi(\Delta S_U(k),\nu_U) = \beta^{*}(\nu_U)
\end{equation}
where, $\alpha^{*}$ and $\beta^{*}$ are the optimal solutions of \eqref{eq:MUS} and \eqref{eq:DSS}, respectively, obtained for the latest values of $M_U^{*}$ and $\Delta S_U$ at iteration $k$. Similar gradient values can be obtained for surrogate components of active/reactive power data.
\noindent\textbf{Step III -} \textit{Aggregate gradient computation}: Insert the obtained gradients in Step II, to form the overall gradient values for the weighted averaging problem \eqref{eq:weightedsum}:
\begin{equation}
\label{eq:g_w1}
\nabla\Gamma_T(M_U^*,M_P^*,M_Q^*) =
\omega_1\alpha^{*}(\mu_U) + \omega_2\alpha^{*}(\mu_P) + \omega_3\alpha^{*}(\mu_Q)
\end{equation}
\begin{equation}
\label{eq:g_w2}
\nabla\Psi_T(\Delta S_U,\Delta S_P,\Delta S_Q) = \omega_1\beta^{*}(\nu_U) + \omega_2\beta^{*}(\nu_P) + \omega_3\beta^{*}(\nu_Q)
\end{equation}
\noindent\textbf{Step IV -} \textit{Interim variable updates}: This step in the algorithm defines and updates several interim variables. These variables will be leveraged in the data refinement step. The idea is to apply gradient descent using the aggregate gradient components, obtained in Step III, while at the same time penalize deviations from the original measurements. Four interim matrices are defined: $Y_M$, $Y_S$, $Z_M$, and $Z_S$. While $Y_M$ and $Y_S$ are computed using the latest solution updates, on the other hand, $Z_M$ and $Z_S$ are obtained using the full history of the solution trajectory. Accordingly, the update process for $[Y_M,Y_S]$ is a convex and tractable optimization process, as follows:
\begin{equation}
\label{eq:update1}
\begin{split}
&[Y_M,Y_S] = \argminA_{M,S} \{<\nabla\Gamma_T(M_U^*(k),M_P^*(k),M_Q^*(k)),M>\\
& + <\nabla\Psi_T(\Delta S_U,\Delta S_P,\Delta S_Q),S> + \frac{L}{2}(||\Delta M ||_F^2 + ||\Delta S||_F^2)\}
\end{split}
\end{equation}
\begin{equation}
\label{eq:constraints2}
s.t.\ \ \eqref{eq:constU}-\eqref{eq:DistFlow}
\end{equation}
where, $L$ is a weight parameter used for penalizing deviations from SM measurements. Here, the deviation from the original data are denoted as $\Delta M$ and $\Delta S$ (e.g., $\Delta M = M - [(M_U^*(0),M_P^*(0),M_Q^*(0)]$). Similarly, a convex optimization process is defined for updating $[Z_M,Z_S]$, considering full solver trajectory:
\begin{equation}
\label{eq:update3}
[Z_M,Z_S] = \argminA_{M,S} \{\tau(M,S) + \frac{L}{2}(||\Delta M ||_F^2 + ||\Delta S||_F^2)\}
\end{equation}
\begin{equation}
\label{eq:constraints3}
s.t.\ \ \eqref{eq:constU}-\eqref{eq:DistFlow}
\end{equation}
where, $\tau(M,S)$ is an average aggregate gradient factor with respect to solver history, defined as follows:
\begin{equation}
\label{eq:update2}
\begin{split}
\tau (M,S) =\sum_{i=0}^k <\nabla\Gamma_T(M_U^*(k),M_P^*(k),M_Q^*(k)),M> +\\ <\nabla\Psi_T(\Delta S_U,\Delta S_P,\Delta S_Q),S>
\end{split}
\end{equation}
\noindent\textbf{Step V -} \textit{Data refinement}: Apply a weighted averaging process using the updated interim variables, from Step IV, to refine the SM data. Based on the suggestion in \cite{Aybat2011}, this weighted update process is written as follows:
\begin{equation}
\label{eq:dataupdate1}
\left[
\begin{array}{c}
M_U^*(k+1)\\
M_P^*(k+1)\\
M_Q^*(k+1)
\end{array}
\right]\leftarrow (\frac{k+1}{k+3}) Y_M + (\frac{2}{k+3})Z_M
\end{equation}
\begin{equation}
\label{eq:dataupdate2}
\left[
\begin{array}{c}
\Delta S_U(k+1)\\
\Delta S_P(k+1)\\
\Delta S_Q(k+1)
\end{array}
\right]\leftarrow (\frac{k+1}{k+3}) Y_S + (\frac{2}{k+3})Z_S
\end{equation}
\noindent\textbf{Step V-}\textit{Iterate and terminate}: $k\leftarrow k + 1$ and go to Step II until the maximum number of iterations is reached. Output the refined SM datasets, $M_U^*$, $M_P^*$, and $M_Q^*$, after algorithm convergence.
\section{Enhancing Grid Monitoring Robustness to SM Asynchrony Error}\label{sec:dsse}
Fig. \ref{fig:monitor} shows how our proposed data recovery technique can be integrated into grid monitoring systems as a pre-processor. The refined data is continuously fed to a branch current state estimation (BCSE) module to monitor the grid states in real-time, including the real and imaginary parts of currents of all branches \cite{Wang2004}. The BCSE method leverages a weighted least squares (WLS)-based solver to minimize the sum of squared residuals ($J$). This problem can be formulated as an optimization task over the distribution network given the recovered data samples $M_U^{*}$, $M_p^{*}$, $M_Q^{*}$ from our multi-objective PCP-based model, as follows:
\begin{equation}
\label{eq:bcse0}
\begin{split}
\min_{\pmb{x_s}} J &= \sum_{i} W_{i,i} (M_z(i) - h_i(\pmb{x_s}))^2\\
s.t. \ \ \ &M_z = \left[
\begin{array}{c}
M_{MV} \\
M_U^{*}(:)\\
M_P^{*}(:)\\
M_Q^{*}(:)\\
M_{PS}
\end{array}
\right]\\
W = &\left[
\begin{array}{ccccc}
W_{MV} & \pmb{0} & \pmb{0}& \pmb{0}& \pmb{0}\\
\pmb{0} & W_U &\pmb{0}& \pmb{0}& \pmb{0}\\
\pmb{0}& \pmb{0} &W_P& \pmb{0}& \pmb{0}\\
\pmb{0} & \pmb{0}& \pmb{0} & W_Q& \pmb{0}\\
\pmb{0} & \pmb{0}& \pmb{0} & \pmb{0}& W_{PS}
\end{array}
\right]\\
&\pmb{x_s} = \left[
\begin{array}{c}
\pmb{I_{re}} \\
\pmb{I_{im}}
\end{array}
\right]
\end{split}
\end{equation}
where, $\pmb{x_s}$ is the grid state vector that contains current real/imaginary values for all the branches of the distribution system ($\pmb{I_{re}}/\pmb{I_{im}}$), and $M_z$ is the measurement vector. The measurement data includes the MV network synchronized sensor measurements ($M_{MV}$), including SCADA and $\mu$PMUs, if available, the refined SM data, $M_U^{*}$, $M_P^{*}$, $M_Q^{*}$, and the pseudo measurements $M_{PS}$ that can generated by our previous work \cite{yuanyx}. $h_i$ is the measurement function that maps state values to the $i$'th measurement variable, which is obtained based on the power flow equation. Furthermore, $W$ is a weight matrix that represents the solver's confidence level in each element of $M_z$. The matrix $W$ includes the measurement confidence weights, consisting of sub-matrces $W_{MV}$, $W_U$, $W_P$, $W_Q$, and $W_{PS}$ corresponding to $M_{MV}$, $M_U^{*}$, $M_P^{*}$, $M_Q^{*}$, and $M_{PS}$, respectively. These weight values are determined by the nominal accuracy levels of the senors as $W_{i,i} = \frac{1}{\sigma_i^2}$, where $\sigma_i^2$ is the $i$'th sensor error variance \cite{Baran1995}. The purpose of the weights is to devalue the importance of unreliable data sources in grid monitoring.
The WLS-based solution employs a gradient-based algorithm to find the optimal solutions for \eqref{eq:bcse0} (i.e., $\nabla_{\pmb{x_s}J = 0}$) \cite{RS2009}. The algorithm involves the following steps to estimate the states of the grid:
\noindent\textbf{Step I} - \textit{Receive input data}: Receive the recovered SM data, $M_U^{*}$, $M_P^{*}$, and $M_Q^{*}$ (see Section \ref{sec:MO}), and the latest measurement data from the primary network, $M_{MV}$. Concatenate the input data to form the measurement vector, $M_z$.
\noindent\textbf{Step II} - \textit{State initialization}: $k\leftarrow 0$; initialize the values of the states through randomization, $\pmb{x_s}(k)$ (to speed up the BCSE solver the values of states can be initialized using the solutions from the last time step.)
\noindent\textbf{Step III} - \textit{Jacobian computation}: Update the \textit{Jacobian matrix}, $H$, using the gradients of the measurement function. The Jacobian captures the sensitivity of the measurements to the state variables:
\begin{equation}
\label{eq:jac}
H_{i,j} = \frac{\partial h_i(\pmb{x_s}(k))}{\partial\pmb{x_s}_j}
\end{equation}
The Jacobian matrix can be conveniently calculated for the BCSE method for feeders with known topology (e.g., see \cite{Wang2004} for details on how Jacobian can be obtained for various types of measurement functions.)
\noindent\textbf{Step IV} - \textit{Gain matrix computation}: Leverage the Jacobian matrix from Step III to obtain the gain matrix, $G$, as follows:
\begin{equation}
\label{eq:bcse3}
G(\pmb{x_s}(k)) = H^\top(\pmb{x_s}(k))WH(\pmb{x_s}(k))
\end{equation}
\noindent\textbf{Step V} - \textit{State update}: Update the values of the states using the gain matrix within the first order Newton-Gauss method, as follows:
\begin{equation}
\label{eq:bcse4}
\pmb{x_s}(k + 1) \leftarrow \pmb{x_s}(k)+G^{-1}H^\top W (M_z - \pmb{h}(\pmb{x_s}(k)))
\end{equation}
\noindent\textbf{Step VI} - \textit{Iterate and terminate}: $k\leftarrow k + 1$; go back to Step III until convergence, i.e., $k \geq k_{max}$, with $ k_{max}$ being a user-defined maximum number of iteration for the BCSE algorithm.
\noindent\textbf{Step VII} - \textit{Roll the time window}: At the new time point, the data recovery is performed using the latest measurement data, according to \ref{sec:MO}. Go back to Step I.
\begin{figure*}[tbp]
\centering
\includegraphics[width=2\columnwidth]{ec2_no_line_kaveh.pdf}
\caption{164-node feeder topology.}
\label{fig:feeder}
\end{figure*}
\section{Numerical Results}\label{sec:result}
The proposed data recovery and grid monitoring framework has been tested and validated using a fully observable feeder model shown in Fig. \ref{fig:feeder}. This feeder represents an unbalanced utility network in U.S. MidWest and consists of 164 nodes, which is publicly available online \cite{Test_system}. The details of the system model include network topology, line parameters, and standard electric components. The system has an average of 30\% solar-power-to-peak-load penetration level. The solar data is adopted from \cite{data}. The nodal time-series load demand is aggregated using a real-world 1-second-resolution household dataset and utilized as the input of the power flow analysis \cite{data}. The computed voltages are treated as the voltage measurements. The resolution of the SM measurements is 15-minute. To simulate realistic asynchronous SM measurements, we randomly sample the 1-second resolution data at 15-min rate at each node to represent SM measurements. Thus, in this work, the SM asynchrony strength of each customer can be anywhere between 0 to 900 s. Fig. \ref{fig:pv} and \ref{fig:load} show the original solar and load time-series data in a day at different nodes of the system. User-defined parameters within the proposed data recovery model, including coefficients of the optimization solver, have been tuned through try-outs over historical/simulation datasets. Basically, the values of these parameters are chosen when the residual of branch current state estimation is minimized. It should be noted that the high computational budget of this strategy does not impact the real-time performance of the proposed method since this parameter calibration is an offline process.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{PV.png}
\caption{PV generation data.}
\label{fig:pv}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{load.png}
\caption{Consumption data.}
\label{fig:load}
\end{figure}
The case study is conducted on a standard PC with an Intel(R) Xeon(R) CPU running at 3.70 GHz and with 32.0 GB of RAM. Based on 500 Monte Carlo simulations, the average computational time is around 23 s, which is feasible in real-time applications. Fig. \ref{fig:V_e}, \ref{fig:P_e}, and \ref{fig:Q_e} show the average error histograms of the proposed data recovery method for voltage, active power, and reactive power, respectively. The error is calculated by comparing the actual values of various variables with the solutions of the recovery model. As can be observed, the recovery error values are maintained within low levels, which confirms the acceptable performance of the data recovery framework. Specifically, the mean average errors are 0.11\%, 2.03\%, and 1.27\% for voltage, active power, and reactive power, respectively. This also demonstrates that the proposed data refinement framework has the best performance over the SM voltage dataset, among the three datasets. This outcome is consistent with the correlation-driven nature of the data recovery model (i.e., nodal voltage measurements are highly correlated, which facilitates better refinement.)
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Verror.png}
\caption{Voltage recovery error.}
\label{fig:V_e}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Perror.png}
\caption{Active power recovery error.}
\label{fig:P_e}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Qerror.png}
\caption{Reactive power recovery error.}
\label{fig:Q_e}
\end{figure}
Fig. \ref{fig:Vt} compares the average value of recovered voltage data from the data refinement framework with the actual nodal voltage average (assuming synchronized sensors) within a sample time-window. As is observed in this figure, the developed algorithm closely follows the underlying signal. Fig. \ref{fig:PQt} shows a similar concept for active and reactive power datasets. As observed in this figure, the data recovery framework is basically an approximate identity mapping between the recovered data and the underlying (ideal) data. This corroborates the satisfactory performance of the model over real data in time domain.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Vt.png}
\caption{Recovered average voltage data versus real (synchronized) time-series data.}
\label{fig:Vt}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{Scat.png}
\caption{Recovered nodal active/reactive power data versus real (synchronized) data.}
\label{fig:PQt}
\end{figure}
Fig. \ref{fig:PF} shows the histogram of power flow error with and without leveraging the DisFlow equations within the proposed data recovery framework. As can be observed, having the DistFlow equations as constraints within the multi-objective data refinement model has resulted in a significant reduction in power flow errors. This demonstrates that the proposed method is able to output data that is consistent with network physics, while capturing the dependencies among all SM datasets.
\begin{figure}
\centering
\subfloat[Without DistFlow\label{fig:woPF}]{
\includegraphics[width=0.5\textwidth]{PFpre.png}
}
\hfill
\subfloat[With DistFlow\label{fig:WPF}]{
\includegraphics[width=0.5\textwidth]{PFpost.png}
}
\caption{Power flow error with and without DistFlow constraints.}
\label{fig:PF}
\end{figure}
Finally, Fig. \ref{fig:bcse_e} depicts the histogram of system monitoring error after applying the data recovery framework. The mean percentage error (MPE) criterion is utilized to evaluate the performance of BCSE with our data recovery method, which is calculated by comparing the real state values ($\pmb{x_s}$), obtained from power flow simulations on the feeder model, with the estimated state values ($\hat{\pmb{x}}_{\pmb{s}}$), coming from the BCSE, as follows:
\begin{equation}
\label{eq:mape}
E = 100 \times \sum_{i}\frac{\pmb{x_s}(i) - \hat{\pmb{x}}_{\pmb{s}}(i)}{\pmb{x_s}(i)}
\end{equation}
As is observed in Fig. \ref{fig:bcse_e}, the DSSE error value is maintained at low levels, which demonstrates the successful integration of the data recovery solution into grid monitoring, which allows us to track the behavior of the feeder accurately. The mean estimation error value is 0.87\%.
To further demonstrate the performance of the proposed SM data recovery method, We have conducted numerical comparisons with two previous methods, including a previous smart meter asynchrony mitigation method \cite{Alimardani2015} and a state-of-the-art low rank data recovery method \cite{Bouwmans2014}. The three methods are simulated with the same real-world datasets to calculate the accuracy of the methods. The comparison result is shown in \ref{fig:bcompare}. As demonstrated in the figure, in terms of voltage, the average recovery errors are 0.11\%, 0.877\%, and 1.34\% for the proposed solution, \cite{Alimardani2015} and \cite{Bouwmans2014}, respectively. In terms of active power, the average recovery errors are 2.03\%, 5.84\%, and 6.48\%, respectively. Hence, based on this dataset, the proposed method can achieve a better performance compared to the previous works.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{HistV.png}
\caption{BCSE error distribution.}
\label{fig:bcse_e}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{nonsy_compare.pdf}
\caption{Comparison results between \cite{Alimardani2015}, \cite{Bouwmans2014}, and the proposed method.}
\label{fig:bcompare}
\end{figure}
\section{Conclusions}\label{sec:con}
In this paper, we have presented a multi-objective data recovery method to mitigate the impacts of SM asynchrony issues in distribution system real-time monitoring. The proposed method is able to refine voltage, active power, and reactive power datasets simultaneously within the same framework via a multi-objective formulation. The inherent dependencies among these measurements are captured by using DistFlow equations. Our solution considers both asynchrony errors and measurement errors, thus making the model more widely applicable to practical distribution systems. A first-order algorithm is presented to solve the proposed multi-objective data recovery model. This algorithm is based on Nesterov method for approximating non-differentiable optimization problems with smooth surrogates. To evaluate the proposed method, a real 164-node utility feeder with real data is utilized. The results show that SM asynchrony error mitigation is possible using the proposed method with good accuracy. In this work, the mean average data recovery error are about 0.11\%, 2.03\%, and 1.27\% for voltage magnitude, active power, and reactive power, respectively. Also, it can be observed that the DistFlow constraints can significantly reduce the inconsistency of recovered data with power flow equations. Based on the proposed data recovery method, the system state estimation error is less than 1$\%$.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\subsection{Microbial association network and its importance}
The gut is one of the most significant habitats of a myriad of microbial communities that play critical roles in their host's health. Well-balanced gut microbial communities provide many health benefits, such as maintaining metabolic homeostasis and high functioning immune system \citep{martinez2016gut,kim2016gut,cani2019microbial}. The imbalance of the gut microbiome (dysbiosis) has been related to a variety of human diseases
\citep{cho2012human,lynch2016human}.
The gut microbial balance is maintained by complex microbial interactions such as metabolites consumption, production, and exchange. Microbiome dysbiosis occurs when these interactions are interrupted by environmental alterations such as diet change, antibiotic consumption, and chemical exposure. These changes may deplete nutrients for beneficial microbes and create favorable surroundings for disease-causing bacteria to flourish. Nevertheless, some microbes help the microbial communities to maintain their stability under the environmental changes by providing energy sources and necessary metabolites \citep{zhang2019facing}. Because of the complexity of the functional roles of microbes, identifying microbial association networks, that is, microbe-microbe interaction networks, is crucial for fundamental understanding of the gut microbiome, a key contributor to the host's health.
\subsection{Motivating application: quantitative microbiome profiling data}
Microbiome data collected from 16S ribosomal RNA (rRNA) sequencing are compositional in that each subject has an arbitrary total microbial count determined by the sequencing instrument \citep{gloor2017microbiome}. Hence, a quantitative comparison of microbial abundances cannot be made across subjects as only the information on relative abundances within a subject are available from such data.
Furthermore, compositional data raise a concern for biased estimates of association since a change in absolute abundance of one microbe affects the relative abundance of all the microbes \citep{vandeputte2017quantitative}. The recently developed quantitative microbiome profiling (QMP) techniques account for these compositional limitations by adjusting microbial counts from 16S rRNA sequencing using cell counts and sequencing depths. In this work, we utilize this recent development by considering the QMP data of \citet{vandeputte2017quantitative} with $n=106$ healthy subjects' gut microbiome. We focus on estimating genus-level association networks with the aim of understanding the overall configurations of healthy gut microbial communities and their interactions.
\subsection{Graphical models and network estimation}
Gaussian graphical model is a popular tool for modeling an association network via an undirected graph, where an edge between the two nodes represents the conditional dependence, and an absence of an edge represents conditional independence. Under the Gaussian assumption, this graph structure is fully encoded in the concentration matrix (the inverse covariance matrix) as a zero off-diagonal entry is equivalent to the conditional independence between the corresponding variables. Thus, multiple methods focus on sparse estimation of concentration matrices. Neighborhood selection \citep{meinshausen2006high} recovers the sparse graph structure by performing $L_{1}$-regularized regression of each node on the rest. \citet{yuan2007model,banerjee2008model,dahl2008covariance,friedman2008sparse} directly estimate the sparse concentration matrix by optimizing the $L_{1}$-penalized log-likelihood function, the so-called graphical lasso. \citet{wang2012bayesian} propose a Bayesian counterpart of the graphical lasso using the Laplace prior on the off-diagonal elements of the concentration matrix. \citet{roverato2002hyper,dobra2011bayesian,lenkoski2011computational} consider a G-Wishart prior for the concentration matrix, of which the posterior inference is computationally more expensive than \cite{wang2012bayesian}. For better scalability, \citet{wang2015scaling} develop a continuous spike-and-slab prior for the off-diagonals of the concentration matrix. Furthermore, the Gaussian graphical models can be extended to non-Gaussian data via latent Gaussian copula models. \citet{liu2012high} consider Gaussian copula model for skewed continuous distributions. \citet{fan2017high} consider extension to mixed binary-continuous variables via latent Gaussian copula. \citet{dobra2011copula} consider a Bayesian latent Gaussian copula for graph estimation with binary and ordinal variables, where they approximate the likelihood function using the extended rank likelihood \citep{hoff2007extending}.
Despite the significant advancements in Gaussian graphical models, they are not appropriate for estimating microbial association networks. Microbiome data obtained from high-throughput sequencing are heavily right skewed and zero-inflated. The zeros, furthermore, are not necessarily absolute, but are often due to the limited sequencing depth. Thus, direct application of Gaussian graphical models to zero-inflated sequencing data leads to inaccurate estimation and inference. To address these challenges, several graphical models for zero-inflated data have been proposed. \citet{osborne2020latent} model microbial counts using Dirichlet-multinomial distribution with a latent Gaussian graphical model. \citet{zhou2020identification} consider a zero-inflated latent Ising model for microbial association network estimation. \citet{mcdavid2019graphical} propose a multivariate hurdle model, which is a mixture of degenerate (at 0) and Gaussian distributions. SPIEC-EASI \citep{kurtz2015sparse} is a two-stage inference procedure specifically designed for compositional microbiome data. \citet{yoon2019spring} propose Semi-Parametric Rank-based approach for INference in Graphical model (SPRING) based on truncated latent Gaussian copula \citep{yoon2020sparse}. \citet{ma2020joint} proposes truncated Gaussian graphical model.
\subsection{The major limitation of existing network estimation models and our proposal}
The aforementioned microbial network estimation models share a common limitation: they do not take advantage of additionally available evolutionary information for reverse-engineering the graph structure.
The information on microbes' genetic similarities is available in a form of a phylogenetic tree, however to our knowledge the phylogenetic tree is not taken into account by existing methods for estimation of microbial networks. Since microbial interactions, positive (e.g., mutualism) or negative (e.g., competition), increase with the increase in microbes' genetic similarity \citep{rohr2014, peralta2016merging}, evolutionary information encoded in a phylogenetic tree has great potential in improving the accuracy of microbial associations network estimation.
In this work, we propose a Bayesian truncated Gaussian copula graphical model for microbial association networks that takes advantage of available evolutionary information. Our major contributions are three-fold. First, we provide a general framework for incorporating evolutionary history into the estimation of microbe-microbe association networks. We model the phylogenetic tree as a Gaussian diffusion process in the latent space,
which allows us to represent the microbes and their ancestors as (correlated) Gaussian vectors. Our framework is not limited to the phylogenetic tree and can accommodate any prior knowledge that is expressed in a tree, e.g., a taxonomic rank tree. We formulate the prior probability model on graph so that the microbes that are closer to each other on the tree have a higher edge inclusion probability. Our simulation study reveals that our approach significantly improves the graph estimation accuracy compared to the methods that do not take advantage of the tree structure (Section~\ref{sec:simulation}). Second, the proposed model effectively handles zero-inflation resulting from limited sequencing depth. We consider the observed zeros as truncated realizations of unobserved random quantities that are below certain thresholds. In particular, we establish a Bayesian formulation of the truncated Gaussian copula model \citep{yoon2020sparse} and develop an efficient Gibbs sampling algorithm. Third, the proposed approach facilitates the statistical inference on the estimated network. For each pair of nodes, an edge connectivity is immediately available from the posterior sample, which provides a convenient way to control the posterior expected FDR \citep{mitra2013bayesian,peterson2015bayesian}. Finally, while our model is designed for quantitative microbiome data, it can also be applied to compositional data using modified central log ratio transformation \citep{yoon2019spring}.
The rest of this paper is organized as follows. In Section \ref{sec:model}, we introduce the proposed graphical model. In Section \ref{sec:inference}, we discuss posterior inference. In Section \ref{sec:simulation}, we evaluate the graph estimation accuracy of the proposed model on simulated datasets. In Section \ref{sec:application}, we analyze quantitative microbiome profiling data of \citet{vandeputte2017quantitative}, and compare our results to SPRING \citep{yoon2019spring} and SPIEC-EASI \citep{kurtz2015sparse}.
\section{Bayesian truncated Gaussian copula graphical model}\label{sec:model}
In Section \ref{ssec:copula}, we discuss the semiparametric modeling of conditional dependencies for zero-inflated data through a truncated Gaussian copula model with a sparse concentration matrix. In Section \ref{ssec:tree}, we introduce a prior model that incorporates the phylogenetic tree to facilitate posterior inference of microbial associations. The complete hierarchical model is summarized in Figure \ref{fig:schema}.
\begin{figure}[t]
\ctikzfig{tikz/BCG_scheme}
\caption{Schematic illustration of the phylogenetically informed Bayesian truncated Gaussian copula graphical model. Hyperparameters that are held constant are given in boxes with dashed-line. The quantities that need posterior inference are illustrated in boxes with solid-line. The ellipse with solid line represents observed data. }
\label{fig:schema}
\end{figure}
\subsection{Truncated Gaussian copula graphical model}\label{ssec:copula}
Let $\boldsymbol{x}=(x_{1},\ldots,x_{p})^{\top}$ denote the zero-inflated abundances of $p$ microbes. This is either directly the counts resulting from quantitative microbiome profiling, e.g. motivating data from \citet{vandeputte2017quantitative}, or transformed
compositional microbiome data via the modified central log-ratio transformation \citep{yoon2019spring}.
We propose to model $\boldsymbol{x}$ such that only the microbial abundances that are larger than certain thresholds can be observed. Specifically, we assume that there exist latent $\boldsymbol{x}^{*}=(x_{1}^{*},\ldots,x_{p}^{*})^{\top}$ representing the true abundances such that
\begin{flalign}\label{eq:truncation}
x_{j} &= 1(x_{j}^{*}>c_{j})x_{j}^{*}, \quad j=1,\ldots,p,
\end{flalign}
where $1(\cdot)$ is the indicator function and $c_{j}$ is the unknown threshold for the $j$th variable.
We call $x_j$ a \textit{truncated} variable if $x_{j}^{*}$ is less than $c_j$, and an \textit{observed} variable, otherwise. Let $F_{j}$ be the marginal cumulative distribution function (cdf) of the $j$th latent variable $x_{j}^{*}$, $\Phi$ be the cdf of standard Gaussian, and $f_{j}=\Phi^{-1}\circ F_{j}$, where we assume that $F_{j}$'s are continuous.
The truncated Gaussian copula model \citep{yoon2020sparse} assumes
\bfl{\label{eq:motr}
& z_{j} = f_{j} (x_{j}^{*}), \quad j=1\ldots,p,\\ \label{eq:copula}
&\boldsymbol{z}=(z_{1},\ldots,z_{p})^{\top} \sim \textup{N}_{p}(\boldsymbol{0},\boldsymbol{\Omega}^{-1}),
}
where $\boldsymbol{\Omega}^{-1} \succ 0$ is the positive definite correlation matrix.
Since $f_{j}$'s are monotone continuous, we can write \eqref{eq:truncation} as $x_{ij}=1\{f_{j}(x_{ij}^{*})>f_{j}(c_{j}) \}x_{ij}^{*}$ $=1(z_j>\delta_j )x_{ij}^{*}$, where $\delta_j=f_j(c_j)$. We denote the truncated and the observed sub-vectors of $\boldsymbol{x}$ by $\boldsymbol{x}_{t}\in \mathbb{R}^{p_{t}}$ and $\boldsymbol{x}_{o} \in \mathbb{R}^{p_{o}}$, respectively, where $p_{t}+p_{o}=p$. Likewise, let $\boldsymbol{z}_{t}$ and $\boldsymbol{z}_{o}$ be the corresponding latent Gaussian vectors, and let $\boldsymbol{\delta}_{t}$ and $\boldsymbol{\delta}_{o}$ be their thresholds.
Given the thresholds, the conditional distribution of $\boldsymbol{z}$ is given by
\begin{flalign}\label{eq:density}
p(\boldsymbol{z}_{o},\boldsymbol{z}_{t}|\boldsymbol{z}_{o}>\boldsymbol{\delta}_{o},\boldsymbol{z}_{t}<\boldsymbol{\delta}_{t},\boldsymbol{\Omega})=
\frac{\textup{N}_{p}(\boldsymbol{z}_{o},\boldsymbol{z}_{t}|\boldsymbol{0},\boldsymbol{\Omega}^{-1}) }{\PP(\boldsymbol{z}_{o}>\boldsymbol{\delta}_{o},\boldsymbol{z}_{t}<\boldsymbol{\delta}_{t}|\boldsymbol{\delta},\boldsymbol{\Omega})} 1(\boldsymbol{z}_{o}>\boldsymbol{\delta}_{o}) 1(\boldsymbol{z}_{t}<\boldsymbol{\delta}_{t}).
\end{flalign}
Unlike the approach in \citet{dobra2011copula} that only uses the relative ranks of the observed data, we condition on the observed value of $\boldsymbol{z}_{o}$, which subsequently allows us to sample truncated variables $\boldsymbol{z}_{t}$ from the posterior distribution as discussed in Section \ref{sec:inference}.
Because of the multivariate normality of $\boldsymbol{z}$, zero entries of $\boldsymbol{\Omega}=[\omega_{jk}]_{1\leq j, k\leq p}$ imply the conditional indepdendence between the corresponding variables. The dependency structure of $\boldsymbol{z}$ can be graphically summarized as an undirected graph $G=(\boldsymbol{z},\boldsymbol{E})$ with an adjacency matrix $\boldsymbol{E}=[e_{jk}]_{1\leq j, k\leq p}$, where nodes $z_j$ and $z_k$ are connected (denoted by $e_{jk}=1$) if $\omega_{jk}\neq0$.
Consequently, learning the graph structure (adjacency matrix) $\boldsymbol{E}$ is equivalent to finding the sparse pattern of $\boldsymbol{\Omega}$. To encourage sparsity, we follow a similar strategy as in \cite{wang2015scaling} by assigning a spike-and-slab prior on the off-diagonal elements of $\boldsymbol{\Omega}$ and an exponential prior on the diagonal elements,
\bfl{
\label{eq:SSSL2}
p(\boldsymbol{\Omega}|\boldsymbol{E},v_{0}^2) = C(\boldsymbol{E},v_{0}^2)^{-1}
1(\boldsymbol{\Omega}\succ 0) & \prod_{j<k} \Big\{ (1-e_{jk})\textup{N}(\omega_{jk}|0,v_{0}^2) + e_{jk}\textup{N}(\omega_{jk}|0,hv_{0}^2) \Big\} \\
\nonumber
\times & \prod_{j=1}^{p} \textup{Exp} (\omega_{jj}|\frac{\lambda}{2}),
}
where $ \textup{Exp} (\cdot|\lambda)$ is the exponential density function with rate parameter $\lambda$, $v_{0}^{2}$ is the spike variance, $h\gg1$ is a large constant such that the slab variance $hv_{0}^{2}\gg v_{0}^{2}$, and $C(\boldsymbol{E},v_{0}^2)$ is the normalizing constant.
We impose an improper uniform prior on the thresholds $\boldsymbol{\delta}=(\delta_{1},\ldots,\delta_{p})^{\top}\sim\pi(\boldsymbol{\delta})\propto 1$, a conjugate inverse-gamma prior on the spike variance $v_{0}^2 \sim \textup{IG}(a_{v_{0}},b_{v_{0}})$, and Bernoulli-like priors on the edge indicators
\bfl{
\label{eq:eprior}
p(\boldsymbol{E})\propto C(\boldsymbol{E},v_{0}^2)\prod_{j<k} \mbox{Ber}(e_{jk}|\pi_{jk}).
}
Including the normalizing constant $C(\boldsymbol{E},v_{0}^2)$ in \eqref{eq:eprior} serves to cancel out that in \eqref{eq:SSSL2}, facilitating the posterior computation in updating $\boldsymbol{E}$ \citep{wang2015scaling}.
\subsection{Incorporating phylogenetic tree}\label{ssec:tree}
Evolution plays an important role in shaping the interaction patterns of microbes \citep{peralta2016merging}. We will exploit the evolution footprints in identifying microbial association networks through a novel phylogenetic tree prior. The proposed prior is a distribution on edge inclusion probabilities $\boldsymbol{\Pi}=[\pi_{jk}]_{1\leq j, k\leq p}$ that encourages the interactions, positive (e.g., mutualism) or negative (e.g., competition), of phylogenetically similar microbes as they tend to be phenotypically/functionally correlated \citep{martiny2015microbiomes,xiao2018predictive,zhou2021bayesian}. The prior is constructed by first embedding the network in $L$-dimensional Euclidean space through the latent position model \citep{hoff2002latent}, and then arranging the latent positions according to the phylogenetic tree.
\textbf{Latent position model.}
We introduce a latent position $\boldsymbol{t}_j=(t_{1j},\dots,t_{Lj})^{\top}\in \mathbb{R}^L$ for each node $z_j$, and link it to the edge inclusion probability $\pi_{jk}$ through a probit link function
\bfln{
\pi_{jk} = \Phi( \boldsymbol{t}_{j}^{\top} \boldsymbol{t}_{k} ), \quad j<k.
}
The inner product $\boldsymbol{t}_{j}^{\top} \boldsymbol{t}_{k}$ measures the similarity between $\boldsymbol{t}_{j}$ and $\boldsymbol{t}_{k}$, with larger inner product leading to higher prior inclusion probability. We assign a prior on $\boldsymbol{t}_j$'s to encourage the interactions between phylogenetically similar microbes.
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{diffusiontree.pdf}\\
\caption{An illustrative example of a phylgenetic tree with $p=4$ microbes (left) and the corresponding diffusion process in $\mathbb{R}^{1}$ with $\sigma^2=3$ (right). The new branch $t_2$ is split from $t_1$ at the first divergence time $s_1=0.2$. Then new branches $t_3$ and $t_4$ are split from $t_1$ and $t_2$ at the divergence time $s_2=0.55$ and $s_3=0.85$, respectively. The green circle and blue triangle are the most common ancestors of the pairs $(t_1,t_3)$ and $(t_2,t_4)$, whose heights $s_2$ and $s_3$ are the correlations of the pairs, respectively. } \label{fig:diffusiontree}
\end{figure}
\textbf{Phylogenetic tree.}
Let $\boldsymbol{T}=[\boldsymbol{t}_{1},\ldots,\boldsymbol{t}_{p}]\in \mathbb{R}^{L\times p}$ and let $\boldsymbol{t}^\ell=(t_{1\ell},\dots,t_{p\ell})$ be the $\ell$th row of $\boldsymbol{T}$. We assume $\boldsymbol{t}^{\ell}\buildrel iid \over\sim \textup{N}_{p}(\boldsymbol{0},\sigma^2\boldsymbol{H})$ for $\ell=1,\dots,L$, where $\boldsymbol{H}$ is a correlation matrix that reflects the phylogenetic similarity. Our specific choice of $\boldsymbol{H}$ is motivated by the following diffusion process. Let $\mathcal{T}$ be a phylogenetic tree with terminal nodes representing the $p$ microbes under investigation and internal nodes representing their common ancestors. Starting from time 0 at the origin (root), the first branch $\boldsymbol{t}_{1}$ follows a Brownian motion with variance $\sigma^2$ until the divergence time $s_1 \in [0,1]$. Then it splits into two branches, $\boldsymbol{t}_{1}$ and $\boldsymbol{t}_{2}$, each following the same Brownian motion independently before they split at times $s_2,s_3\in [0,1]$ resulting in $\boldsymbol{t}_{3}$ and $\boldsymbol{t}_{4}$. This process repeats until the $p$ terminal nodes are reached at time 1. An illustrative example of the diffusion process is provided in Figure \ref{fig:diffusiontree}
This diffusion process defines a centered multivariate Gaussian distribution on the terminal nodes of $\mathcal{T}$ with covariance matrix $\sigma^2 \boldsymbol{H}$. We define the \textit{height} of each node (split for internal nodes) as its distance from the root (time from 0). The correlation of two terminal nodes equals the height of their most recent common ancestor, which is large for a phylogenetically similar microbes. This multivariate Gaussian prior, together with the latent position model, achieves the desired prior distribution of $\pi_{jk}$ that encourages interactions between phylogenetically similar microbes.
Lastly, we assign a conjugate inverse-gamma prior for the variance parameter $\sigma^2\sim \textup{IG}(a_{\sigma^2},b_{\sigma^2})$.
\section{Posterior inference}\label{sec:inference}
The proposed model is parameterized by $\{\boldsymbol{z},\boldsymbol{\delta}, \boldsymbol{\Omega}, \boldsymbol{E},v_{0}^2,\boldsymbol{T},\sigma^2\}$ of which the posterior distribution is not available in closed form.
We use a Markov chain Monte Carlo (MCMC) algorithm to draw posterior samples from the intractable posterior distribution. Section \ref{ssec:samplezt} discusses Gibbs steps for sampling $\boldsymbol{z}$ and $\boldsymbol{\delta}$ from their full conditional distributions. Section \ref{ssec:samplelatent} describes Gibbs steps for $\boldsymbol{T}$ and $\sigma^2$. In Section \ref{supp:spikeandslab} of the Supplemenatry Materials, we provide the Gibbs steps for $\boldsymbol{\Omega}, \boldsymbol{E},v_{0}^2$ by following \cite{wang2015scaling}.
\subsection{Full conditionals of truncated observations and thresholds} \label{ssec:samplezt}
Let $\boldsymbol{x}_{1},\ldots,\boldsymbol{x}_{n} \in \mathbb{R}^{p}$ be a sample from the model \eqref{eq:truncation}--\eqref{eq:copula} and let $\boldsymbol{z}_{1},\ldots,\boldsymbol{z}_{n} \in \mathbb{R}^{p}$ be the corresponding latent variables. As before, we use subscripts $t$ and $o$, respectively, to denote the truncated and observed sub-vectors (e.g., $\boldsymbol{x}_{i,t}$ and $\boldsymbol{x}_{i,o}$ are the truncated and observed sub-vectors of $\boldsymbol{x}_{i}$).
For observed $\boldsymbol{x}_{i,o}$, the corresponding Gaussian variables are defined as $z_{ij,o} =f_j(x_{ij,o})=\Phi^{-1}\circ F_{j}(x_{ij,o})$. A natural estimator for the unknown function $F_{j}$ is the scaled empirical cdf $\widehat{F}_{j}(c)=\{n/(n+1)\}\sum_{i=1}^{n} n^{-1}1(x_{ij}\leq c)$ \citep{klaassen1997efficient}, where the constant term $n/(n+1)$ is needed to make $\Phi^{-1}$ finite. Thus, the estimator of $f_j$ is given by $\Phi^{-1}\circ\widehat F_j$, and we set $\widehat{z}_{ij,o}=\Phi^{-1}\circ \widehat{F}_{j}(x_{ij,o})$. An alternative approach to estimate $f_j$ using B-spline basis functions has been considered in \citet{mulgrave2020bayesian}; however, we use the empirical cdf for computational efficiency.
Given $\boldsymbol{z}_{i,o}=\widehat{\boldsymbol{z}}_{i,o}$, we sample $\boldsymbol{z}_{i,t}$ from a truncated multivariate Gaussian distribution derived from \eqref{eq:density},
\bfln
p(\boldsymbol{z}_{i,t}|\widehat{\boldsymbol{z}}_{i,o},\boldsymbol{z}_{i,t}<\boldsymbol{\delta}_{t},\boldsymbol{\Omega})=
\frac{\textup{N}_{p_{i,t}}(\boldsymbol{z}_{i,t}|\boldsymbol{\mu}_{i},\boldsymbol{\Delta}_{i}) }{\PP(\boldsymbol{z}_{i,t}<\boldsymbol{\delta}_{t}|\widehat{\boldsymbol{z}}_{i,o},\boldsymbol{\Omega},\boldsymbol{\delta})} 1(\boldsymbol{z}_{i,t}<\boldsymbol{\delta}_{i,t}),
}
where $p_{i,t}$ is the number of truncated variables of $\boldsymbol{x}_{i}$, and $\boldsymbol{\mu}_{i}$, $\boldsymbol{\Delta}_{i}$ are the mean and covariance matrix of $\boldsymbol{z}_{i,t}$ given $\boldsymbol{z}_{i,o}=\widehat{\boldsymbol{z}}_{i,o}$ (detailed expressions are provided in Section \ref{supp:samplezt} of the Supplementary Materials).
The conditional pdf of $\boldsymbol{\delta}$ given ${\boldsymbol{z}}_{1},\ldots,\boldsymbol{z}_{n}$ is proportional to the $n$-product of indicator functions of \eqref{eq:density}. That is, we have the independent uniform full conditional distributions of $\delta_{j}$'s as
\bfln{
p(\boldsymbol{\delta}| {\boldsymbol{z}}_{1},\ldots,{\boldsymbol{z}}_{n},\boldsymbol{\Omega}) &\propto
\prod_{i=1}^{n}
\textup{N}_{p}({\boldsymbol{z}}_{i,t},\widehat{\boldsymbol{z}}_{i,o},|\boldsymbol{\Omega})
1({\boldsymbol{z}}_{i,t}<\boldsymbol{\delta}_{i,t}) 1(\widehat{\boldsymbol{z}}_{i,o}>\boldsymbol{\delta}_{i,o}) ,\\
&\propto \prod_{j=1}^{p} 1({z}_{j,t}^{\max}<\delta_{j}< \widehat{z}_{j,o}^{\min} ),
}
where ${z}_{j,t}^{\max}=\max_{i} {z}_{ij,t}$ and $\widehat{z}_{j,o}^{\min}=\min_{i} \widehat{z}_{ij,o}$ are the maximum and minimum of the truncated and observed components of the $j$th Gaussian variable, respectively.
\subsection{Full conditionals of latent positions and the tree scale}\label{ssec:samplelatent}
Let $\boldsymbol{T}_{-j}$ be the submatrix of $\boldsymbol{T}$ without the $j$th column.
The full conditional distribution of $\boldsymbol{t}_{j}$ is given by,
\bfln{
p(\boldsymbol{t}_{j}|\boldsymbol{T}_{-j},\boldsymbol{E},\sigma^2) \propto \left[ \prod_{k \neq j } \{ \Phi( \boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j} )\}^{e_{kj}} \{1- \Phi(\boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j} )\}^{1-e_{kj}} \right]
\textup{N}_{L}(\boldsymbol{t}_{j}|\boldsymbol{\theta}_{j},\boldsymbol{\Psi}_{j}), \quad j=1,\ldots,p,
}
where $\boldsymbol{\theta}_{j}$ and $\boldsymbol{\Psi}_{j}$ are mean and covariance matrix of $\boldsymbol{t}_{j}$ given $\boldsymbol{T}_{-j}$ (detailed expressions are provided in Section \ref{supp:samplelatent} of the Supplementary Materials). We update $\boldsymbol{t}_{j}$ using the data augmentation technique of \citet{albert1993bayesian} by introducing the auxiliary data $\boldsymbol{y}_{j} \in \mathbb{R}^{p-1}$. Let $\textup{TN}(\mu,\sigma^2,e)$ be $\textup{N}(\mu,\sigma^2)$ truncated to be positive if $e=1$ and negative if $e=0$. Conditining on $\boldsymbol{T}$, each component $y_{kj}$ of $\boldsymbol{y}_{j}$ follows $\textup{TN}(\boldsymbol{t}_{k}^{\top}\boldsymbol{t}_{j},1,e_{kj})$, and the resulting augmented pdf of $\boldsymbol{t}_{j}$ and $\boldsymbol{y}_{j}$ is
\bfln{
p(\boldsymbol{t}_{j},\boldsymbol{y}_{j}&|\boldsymbol{T}_{-j},\boldsymbol{E},\sigma^2) \propto \\
& \prod_{k \neq j}
\left\{ 1(y_{kj}>0, e_{kj}=1) + 1(y_{kj}<0,e_{kj}=0) \right\} \textup{N}(y_{kj}|\boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j}, 1) \textup{N}_{L}(\boldsymbol{t}_{j}|\boldsymbol{\theta}_{j},\boldsymbol{\Psi}_{j}).
}
We obtain a posterior sample of $\boldsymbol{T}$ by alternately sampling $\boldsymbol{y}_{j}$ and $\boldsymbol{t}_{j}$ for $j=1,\ldots,p$. Conditional on $\boldsymbol{T}$, we independently sample $y_{kj}$ from $\textup{TN}(\boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j},1,e_{kj})$ for $k\neq j$. Then, conditional on $\boldsymbol{y}_{j}$, we have the Gaussian full conditional of $\boldsymbol{t}_{j}$ as
\begin{flalign}\nonumber
p(\boldsymbol{t}_{j} |\boldsymbol{y}_{j}, \boldsymbol{T}_{-j},\boldsymbol{E},\sigma^2) \propto
\textup{N}_{p-1}(\boldsymbol{y}_{j}|\boldsymbol{T}_{-j}^{\top} \boldsymbol{t}_{j}, \boldsymbol{I}_{p-1}) \textup{N}_{L}(\boldsymbol{t}_{j}|\boldsymbol{\theta}_{j},\boldsymbol{\Psi}_{j}).
\end{flalign}
Accordingly, we draw $\boldsymbol{t}_{j}$ from $\textup{N}_{L}( \boldsymbol{\gamma}_{j}, \boldsymbol{\Delta}_{j} )$, where
\bfln{
\boldsymbol{\Delta}_{j} = \left( \boldsymbol{T}_{-j}\boldsymbol{T}_{-j}^{\top} + \boldsymbol{\Psi}_{j}^{-1} \right)^{-1}, \quad
\boldsymbol{\gamma}_{j} = \boldsymbol{\Delta}_{j}
\left( \boldsymbol{T}_{-j} \boldsymbol{y}_{j} + \boldsymbol{\Psi}_{j}^{-1} \boldsymbol{\theta}_{j} \right).
}
The edge inclusion probabilities are updated as $\pi_{jk}=\Phi(\boldsymbol{t}_{j}^{\top}\boldsymbol{t}_{k})$, $1\leq j<k \leq p$.
Conditional on $\boldsymbol{T}$, we sample the tree scale parameter $\sigma^2$ from
\bfln{
\sigma^2|\boldsymbol{T} \sim \textup{IG}\left(pL/2 + a_{\sigma^2}, \vect(\boldsymbol{T})^{\top}(\boldsymbol{H}\otimes \boldsymbol{I}_{L})^{-1}\vect(\boldsymbol{T})/2 + b_{\sigma^2}\right),
}
where $\vect(\boldsymbol{T})$ is the vector obtained by stacking the columns of $\boldsymbol{T}$ and $\otimes$ is the Kronecker product.
For sampling concentration matrix and graph, we follow the block Gibbs sampler of \citet{wang2015scaling}. For completeness, we provide the block Gibbs sampling algorithm in Section \ref{supp:spikeandslab} of the Supplementary Materials. Upon the completion of the MCMC, we compute the posterior mean of the edge inclusion probabilities for each pair of nodes, $\widehat{\pi}_{jk}=\sum_{s=1}^{S} e_{jk}^{(s)}/S$, where the superscript indexes posterior samples. We obtain the estimated graph by selecting edges for which $\widehat{\pi}_{jk}$ is larger than some cutoff.
We choose the cutoff to control the posterior expected FDR \citep{mitra2013bayesian,peterson2015bayesian} at prespecified level $\alpha$, where the posterior expected FDR is a decreasing function of cutoff $c$ defined as
\bfl{\label{eq:postfdr}
\textup{E}(\text{FDR}_{c}|\text{data})=\frac{\sum_{j<k}(1-\widehat{\pi}_{jk})1(\widehat{\pi}_{jk}> c)}{\sum_{j<k}1(\widehat{\pi}_{jk}> c)}.
}
\section{Simulation}\label{sec:simulation}
We simulate microbiome data following the data generation mechanism proposed in \cite{yoon2019spring}, which allows to obtain synthetic samples that exactly follow the empirical marginal cumulative distributions of measured microbiome count data while respecting user-specified microbial dependencies via $\boldsymbol{\Omega}$.
Specifically, we randomly generate 10 phylogenetic trees $\mathcal{T}_{1},\ldots,\mathcal{T}_{10}$ with $p=50$ terminal nodes by the \textsf{R} package \textsf{ape} \citep{paradis2019ape}, constituting 10 simulation scenarios. For each tree, the latent positions of terminal nodes, $\boldsymbol{t}_{1},\ldots,\boldsymbol{t}_{p}$, are generated from the diffusion process as described in Section \ref{ssec:tree} with $\sigma^2=3$ and $L=2$. The true graph adjacency matrix $\boldsymbol{E}_0$ is obtained by independently generating $e_{jk} \sim \textup{Bernoulli} (\pi_{jk})$ for $1\leq j<k \leq p$ with $\pi_{jk} = \Phi(\boldsymbol{t}_{j}^{\top} \boldsymbol{t}_{k})$. The trees and the true graphs are plotted in Section \ref{supp:phylotrees} of the Supplementary Materials.
Given $\boldsymbol{E}_0$, the concentration matrix $\boldsymbol{\Omega}$ is drawn from G-Wishart$(\boldsymbol{I}_{p},4)$ \citep{roverato2002hyper}.
To obtain empirical cdfs, we use the quantitative microbiome profiling data of \citet{vandeputte2017quantitative} from $n=106$ subjects, more detailed description of the data is provided in Section \ref{sec:application}. We select $p = 50$ genera (variables) of which 6 genera have no observed zero counts, and 44 genera have 20\% to 70\% zero counts across samples.
Given the empirical cdf $\widehat{F}_{j}$ of each selected genera, $j=1,\ldots, p$, and the concentration matrix $\boldsymbol{\Omega}$, we generate $n=106$ independent latent Gaussian vectors $\boldsymbol{z}_{i}\sim\textup{N}_{p}(\boldsymbol{0}, \boldsymbol{\Omega}^{-1})$. The final data $\boldsymbol{x}_1,\dots,\boldsymbol{x}_n$ are obtained as
\bfln{
x_{ij} = \widehat{F}_{j}^{-1}\circ \Phi(z_{ij}), \quad i=1,\ldots,n, \quad j=1,\ldots,p.
}
We consider 50 independent replications of this data generating process for each scenario.
We compare the performance of the the proposed phylogenetically-informed Bayesian Copula Graphical model (PhyloBCG) with SPIEC-EASI \citep{kurtz2015sparse} and SPRING \citep{yoon2019spring}. Additionally, we also consider two special cases of PhyloBCG with the following simplification to the prior model of graph:
\begin{flalign*}
{\textup{Oracle}} :~& \pi_{jk} = \binom{p}{2}^{-1}|\boldsymbol{E}_0|;\\
{\textup{Dist}}:~& \pi_{jk} = \exp(-\gamma d_{jk} ), \quad \gamma \sim \textup{Exp} (1);
\end{flalign*}
where $|\boldsymbol{E}_0|$ is the number of true edges in the underlying graph, and $d_{jk}$ is the tree distance between terminal nodes $j$ and $k$, defined as the sum of the branch lengths to their most recent common ancestor. We refer to the first model as ``${\textup{Oracle}} $'' since it uses the knowledge of true graph sparsity, however it does not use any tree information. We refer to the second model as ``${\textup{Dist}}$'', as it directly incorporates tree distances between the terminal nodes: the edge inclusion probability is higher (smaller) when the corresponding nodes are closer (farther) to each other on the tree. ${\textup{Dist}}$ model thus takes into account the information from the tree, however, the tree information is deterministically incorporated to the model in contrast to the stochastic incorporation of PhyloBCG.
To implement SPIEC-EASI and SPRING, we use the corresponding R packages \citep{kurtz2017spieceasi,yoon2019springpackage} with sparsity parameters tuned over 100 values.
For PhyloBCG, Oracle and Dist, the hyperparameters are fixed as $a_{\sigma^2}=b_{\sigma^2}=a_{v_{0}^{2}}=b_{v_{0}^{2}}=0.001$, $h=2500$, $\lambda=1$, and $L=2$. PhyloBCG is relatively robust to the hyperparameter setting; see the sensitivity analyses in Section \ref{supp:sensitivity} of the Supplementary Materials. We obtain an MCMC posterior sample of size $S=5000$ after 500 burn-in iterations. The graph estimate is obtained by thresholding the mean of posterior inclusion probability $\widehat{\pi}_{jk}$ at $c_{0.05}$, the smallest $c$ that controls the posterior expected FDR \eqref{eq:postfdr} at level 0.05.
We assess the graph recovery performance (accuracy in estimating $\boldsymbol{E}_{0}$) of each method using Matthews correlation coefficient \citep{matthews1975comparison}, true positive rate, and false positive rate, which will be denoted by MCC, TPR, and FPR, respectively. Ranging from -1 to 1, a larger value of MCC represents a better network estimation accuracy, where the two boundary values indicate completely correct (+1) and wrong (-1) edge selection, respectively.
Figure \ref{fig:simsum} summarizes the mean values of these metrics for each of the 10 phylogenetic trees based on 50 replications. The phylogenetic trees are sorted in terms of the global clustering coefficient \citep{wasserman1994social} of the true graph from largest value ($\mathcal{T}_{1}$) to lowest value ($\mathcal{T}_{10}$). A large value of the global clustering coefficient indicates a presence of microbial communities, with dense interactions within the same community and sparse interactions across communities. A small value of the global clustering coefficient indicates a random interaction pattern close to what will be expected with Erdos-Renyi random graph. Thus, we anticipate the phylogenetic tree to be more informative for network estimation when the global clustering coefficient is larger.
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{simsum_cdf_all_ejk_fdr_5.pdf}\\
\caption{Averages of Matthews correlation coefficient (MCC), true positive rates (TPR) and false positive rates (FPR) with 2 standard error bars. The phylogenetic trees are ordered in term of the global clustering coefficients. Averages are taken over 50 replicated data sets.} \label{fig:simsum}
\end{figure}
Figure \ref{fig:simsum} supports that incorporating phylogenetic tree information improves the network estimation accuracy, with Phylo and Dist having higher MCC, higher TPR, and similar FPR values when compared to other methods. As expected, the value of MCC for the tree-based methods decreases as the phylogenetic tree becomes less informative (larger tree index). For PhyloBCG, this trend is driven by the decreasing TPR, whereas, for Dist, it is the increasing FPR. Although PhyloBCG and Dist both use the evolutionary information, PhyloBCG shows significantly better performances than Dist possibly due to the flexibility of the latent space embedding.
Note that the FPR of PhyloBCG has larger variability for trees $\mathcal{T}_{5}$ and $\mathcal{T}_{10}$ than others, with $\mathcal{T}_5$ leading to the largest mean FPR. The reason for increased FPR in these settings is the discrepancy between the phylogenetic tree and the true graph. Recall that the true edge inclusion probabilities $\pi_{jk}$ are obtained from the Gaussian latent positions rather than directly from the tree, thus allowing the true graph to deviate, sometimes significantly, from the phylogenetic tree. To illustrate this phenomenon, Figure \ref{fig:violin_hjk} in the Supplementary Materials shows the upper triangular part of the true tree correlation matrix $\boldsymbol{H}$ defined in Section~\ref{ssec:tree} against the edge indicators for $\mathcal{T}_{1}-\mathcal{T}_{10}$. Both $\mathcal{T}_{5}$ and $\mathcal{T}_{10}$ show a large number of disconnected edges, $e_{jk}=0$, with high tree correlation values which may contribute to the increase in FPR. Nevertheless, for $\mathcal{T}_{5}$, ${\textup{PhyloBCG}}$ still shows favorable performance, having much larger TPRs than SPIEC-EASI and SPRING. Under $\mathcal{T}_{10}$, PhyloBCG shows comparable performance to Oracle. Considering that the prior edge inclusion probability of ${\textup{Oracle}} $ relies on the knowledge of the true number of connected edges, whereas the proposed PhyloBCG model does not and performs at least as good as the ${\textup{Oracle}} $, we conclude that PhyloBCG adapts well to the unknown sparsity of the underlying graph.
\section{Application to quantitative gut microbiome profiling data}\label{sec:application}
\subsection{Data and phylogenetic tree}
We focus on estimating genus-level association network of the QMP data \citep{vandeputte2017quantitative} that consists of $n=106$ healthy subjects' gut microbiome. We use the data as processed in \citet{yoon2019spring}, which can be obtained from the R-package \textsf{SPRING} \citep{yoon2019springpackage}. Among the 91 genera, there are 33 genera with missing names and 4 genera without available phylogenetic information on the National Center for Biotechnology Information (NCBI) database on which we based the phylogenetic tree construction. Consequently, we consider $p=54$ genera and obtain their phylogenetic tree based on the NCBI taxonomy database using the platform PhyloT\footnote{https://phylot.biobyte.de}. As the database does not provide divergence times of branches, we match the branch lengths with the taxonomic ranks as illustrated in Figure \ref{fig:qmptreemembership}. These genera have up to 70\% zeros (14 have more than 50\% zeros).
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth]{figures/membership.tree.4chains.pdf}
\caption{The phylogenetic tree of 54 genera. Different numbers (colors) represent three microbial communities derived from the graph estimated by the proposed PhyloBCG; the asterisk indicates a stand-alone node. }\label{fig:qmptreemembership}
\end{figure}
\subsection{Analysis}
For PhyloBCG, we use the same hyperparameter values as in Section \ref{sec:simulation} and run 4 parallel Markov chains for 100,000 iterations after 25,000 burn-in iterations. We then concatenate the 4 chains and obtain a posterior sample of size 10,000 by retaining every 40th iteration, from which we compute the posterior means $\widehat{\pi}_{jk}$ and $\widehat{\boldsymbol{\Omega}}$.
The microbial association network is estimated by controlling the posterior expected FDR at 0.1 which results in $c_{0.1}=0.719$.
SPIEC-EASI and SPRING are tuned using 100 sparsity parameter values. The default stability threshold \citep{liu2010stability} is changed from 0.1 to 0.2 to avoid overly sparse network estimates. The recovered networks are shown in Figure \ref{fig:qmpnetwork}.
\afterpage{
\begin{figure}[ht!]
\centering
\includegraphics[width=0.91\textwidth]{figures/network.tree.4chains.pdf}
\includegraphics[width=0.90\textwidth]{network.sp_1e-3.pdf}
\includegraphics[width=0.90\textwidth]{RMPnetwork.se_1e-3.pdf}
\caption{ Estimated microbial association networks of PhyloBCG, SPRING, and SPIEC-EASI with 54 genera in QMP data \citep{vandeputte2017quantitative}.
Communities are separately estimated with the graph estimate of each model, and node colors represent the memberships.}\label{fig:qmpnetwork}
\end{figure}
}
\subsubsection{Overall network summary and interpretation}
We first compare estimated networks in terms of their density and community structure. The estimated network from PhyloBCG appears to be denser and have much more definitive communities than those from SPIEC-EASI and SPRING.
While we do not know the true network and community structure of the gut microbiota in the study population, the following reasons support our belief that the additional findings of microbial interactions and communities from PhyloBCG are biologically meaningful.
First, microbes are known to form communities \citep{pflughoeft2012human}. Applying the edge proximity measure of \citet{newman2004finding} to the estimated network by PhyloBCG, we find three evident microbial communities which are marked with different colors in Figures \ref{fig:qmptreemembership} and \ref{fig:qmpnetwork}. Posterior mean latent positions, illustrated in Figure \ref{fig:post_latentpositions}, also form three distinct clusters that consistently match the estimated microbial communities.
Interestingly, the genera within each of these communities tend to share unique characteristics. On the one hand, most of the genera from the top two communities in Figure \ref{fig:qmptreemembership} are obligate anaerobes which only survive in the absence of oxygen.
On the other hand, the community located at the bottom of Figure \ref{fig:qmptreemembership} contains genera with species that need or at least can tolerate oxygen. For example, \textit{Streptococcus} and \textit{Enterococcus} contain facultative anaerobic species that are able to utilize oxygen as a source of energy, but can also generate energy anaerobically in an oxygen-deficient environment \citep{fisher2009ecology,clewell1981streptococcus}. All the members of \textit{Haemophilus} are facultatively anaerobic species or aerobic species, where aerobic species need oxygen to survive \citep{cooke20171611haemophilus}. \textit{Lactobacillus} contains aerotolerant and microaerophilic species \citep{zheng2020lactobacillus}. Aerotolerant species do not need oxygen and use anaerobic fermentation to generate energy, but oxygen is not toxic to them. Microaerophiles need oxygen to survive but require low oxygen concentration to thrive and are damaged by high oxygen level, e.g., atmospheric oxygen level. Furthermore, the bottom community has very few interactions with the two top communities, possibly because of their distinct living environments. By contrast, the estimated microbial interaction networks from SPIEC-EASI and SPRING do not present obvious communities; this lack of community structure is, based on existing literature \citep{rohr2014, peralta2016merging}, unlikely.
Second, while the phylogenetic tree prior helps identify additional interactions and communities, it does not dictate the posterior inference. Some of the genera from the top right community in Figure \ref{fig:qmptreemembership} (e.g., \textit{Succinivibrio} and \textit{Prevotella}) are not phylogenetically similar to each other. Similarly, the other two communities also contains phylogenetically distant genera. This suggests that the phylogenetic tree prior does not override the information of associations that are strongly supported by the data.
Third, from the simulation study, we have seen that PhyloBCG is much more powerful in detecting interactions than SPIEC-EASI and SPRING especially when there is a clear community structure, while also having comparable FPR.
Given that the communities of the estimated network by PhyloBCG seem biologically plausible, the additional interactions found by PhyloBCG are more likely to be true positives than false discoveries, some of which will be explained in detail in the next section.
\begin{figure}[t!]
\centering
\includegraphics[width=1\textwidth]{figures/latent_positions.pdf}
\caption{ Posterior mean latent positions of 54 genera in QMP data \citep{vandeputte2017quantitative}. Community memberships estimated with the graph estimate of PhyloBCG are marked by colored symbols. Black squares represent stand alone genera. }\label{fig:post_latentpositions}
\end{figure}
\subsubsection{Detailed explanation of interactions}
\textbf{All models} identify strong negative partial correlations between \textit{Dialister} and \textit{Phascolarctobacterium}. This finding is in agreement with multiple published results. The original QMP study \citep{vandeputte2017quantitative} finds a strong negative correlation between these two genera. \citet{naderpoor2019faecal} report that \textit{Phascolarctobacterium} (\textit{Dialister}) is positively (negatively) correlated with insulin sensitivity. Consistently, \citet{pedrogo2018gut} indirectly observe a trade-off relationship between the two genera from obese groups.
Besides, strong positive partial correlations between \textit{Oscillospira} and \textit{Ruminococcus}
are found by all methods as well, which is consistent with the finding in \citet{chen2020high}.
\textbf{PhyloBCG and SPRING} detect relatively strong partial correlations for the pairs (\textit{Mitsuokella}, \textit{Prevotella}) and (\textit{Ruminococcus}, \textit{Blautia}). In the phylogenetic tree displayed in Figure \ref{fig:qmptreemembership}, \textit{Ruminococcus} and \textit{Blautia} are relatively close to each other, being the members of the same order, Clostridales. \textit{Mitsuokella} and \textit{Prevotella} are, however, phylogenetically distant from each other, indicating that the QMP data present a strong association between them and that the tree prior of the proposed PhyloBCG does not dominate the inference. These findings agree with the network analysis of \citet{ramayo2016phylogenetic}, where they also detect positive partial correlations in the pairs (\textit{Mitsuokella}, \textit{Prevotella}) and (\textit{Ruminococcus}, \textit{Blautia}).
Additionally, the pairs (\textit{Oscillospira}, \textit{Butyricimonas}) and (\textit{Eubacterium}, \textit{Peptococcus}) also show relatively strong partial correlations under the two models. Both pairs are known to be related to diet and leanness. \citet{oh2020association} discover positive correlations between body weight and each of \textit{Eubacterium} and \textit{Peptococcus}. \citet{garcia2018shifts} find negative correlations between unhealthy diet (high intake of saturated fat and refined carbohydrates) and each of \textit{Oscillospira} and \textit{Butyricimonas}.
Furthermore, positive partial correlations for the pairs (\textit{Bacteroides}, \textit{Bilophila}) and (\textit{Akkermansia}, \textit{Methanobrevibacter}) are observed, where these results match the analyses in \citet{vandeputte2016stool} and \citet{vandeputte2017quantitative}, respectively.
\begin{table}[t]
\centering
\caption{Selected pairs of genera with strong association as identified by PhyloBCG. ``--'' indicates that no significant partial correlation is found by the corresponding.}\label{tab:pairs}
\resizebox{1\textwidth}{!}{
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{ll|ccc|l}
\multicolumn{2}{c|}{\multirow{2}{*}{Pairs of Microbial Genera}}&\multicolumn{3}{c|}{Partial Correlations}& \multirow{2}{*}{Reference} \\\cline{3-5}
&& PhyloBCG & SPRING & SPIEC-EAIS & \\
\hline
Dialister & Phascolarctobacterium & -0.384 & -0.185 & -0.252 & \citet{vandeputte2017quantitative}, \citet{naderpoor2019faecal} \\
Oscillospira & Ruminococcus & 0.129 & 0.320 & 0.335 & \citet{chen2020high} \\
Mitsuokella & Prevotella & 0.155 & 0.152 & -- & \citet{ramayo2016phylogenetic} \\
Ruminococcus & Blautia & 0.115 & 0.048 & -- & \citet{ramayo2016phylogenetic} \\
Oscillospira & Butyricimonas & 0.155 & 0.335 & -- & \citet{garcia2018shifts}, \citet{thomaz2021influence} \\
Eubacterium & Peptococcus & 0.178 & 0.083 & -- & \citet{oh2020association} \\
Bacteroides & Bilophila & 0.137 & 0.221 & -- & \citet{vandeputte2017quantitative} \\
Akkermansia & Methanobrevibacter & 0.202 & 0.354 & 0.054 & \citet{vandeputte2016stool} \\
Blautia & Methanobrevibacter & -0.062 & -- & -0.329 & \citet{garcia2018shifts}, \citet{muller2019distal} \\
Prevotella & Bacteroides & -0.107 & -- & -0.019 & \cite{vandeputte2017quantitative} \\
& & & & & \citet{ley2016prevotella}, \citet{johnson2017microbiome} \\
Veillonella & Streptococcus & 0.149 & -- & -- & \citet{anbalagan2017next}, \citet{chen2020gut}\\
& & & & & \citet{den2013diversity}, \citet{egland2004interspecies} \\
& & & & & \citet{zoetendal2012human}, \citet{van2014immunomodulatory} \\
Bifidobacterium & Holdemania & -0.167 & -- & -- & \citet{liu2017fructooligosaccharide}, \citet{yang2018dietary} \\
& & & & & \citet{wang2020rational} \\
\hline
\end{tabular}
}
\end{table}
\textbf{PhyloBCG and SPIEC-EASI} find negative partial correlations for the pairs (\textit{Blautia}, \textit{Methanobrevibacter}) and (\textit{Prevotella}, \textit{Bacteroides}). \textit{Blautia} and \textit{Methanobrevibacter} are known to be positively and negatively related to dietary fiber intake, respectively \citep{garcia2018shifts}. \citet{muller2019distal} suggest that their inverse relationship is possibly due to substrate competition as both use hydrogen as energy source.
For \textit{Prevotella} and \textit{Bacteroides}, \citet{lozupone2012diversity} find the trade-off between these two genera -- carbohydrates (including simple sugars) focused diet increases \textit{Prevotella} and decreases \textit{Bacteroides} whereas protein and fat focused diet has the opposite effects on them. Their trade-off relationship is also discussed in \citet{ley2016prevotella} and \citet{johnson2017microbiome}. On the contrary, \citet{vandeputte2017quantitative} claim that their negative association is an artifact of using compositional rather than quantitative microbiome data for analyses .
\textbf{PhyloBCG} uniquely discovers positive partial correlations for the pairs (\textit{Veillonella}, \textit{Streptoccocus}) and (\textit{Bifidobacterium}, \textit{Holdemania}). The estimated positive partial correlation between \textit{Veillonella} and \textit{Streptoccocus} is consistent with that of the gut microbial network analysis of \citet{chen2020gut}. \citet{anbalagan2017next} demonstrate their mutualistic relationship: \textit{Streptoccocus} uses glucose as a source of carbon and release lactic acid, whereas \textit{Veillonella} utilizes lactic acid as carbon and energy source for growth. There are also many studies reporting their co-occurrence and mutualism \citep{den2013diversity,egland2004interspecies, zoetendal2012human,van2014immunomodulatory}.
For \textit{Bifidobacterium} and \textit{Holdemania}, \citet{liu2017fructooligosaccharide} report that prebiotic supplement significantly increases relative abundance of beneficial \textit{Bifidobacterium} and decreases \textit{Holdemania}, where \textit{Holdemania} is reported to be associated with unhealthy gut and antibiotic use \citep{yang2018dietary}. \citet{wang2020rational} discuss the underlying mechanism of the trade-off relationship.
In summary, we find these uniquely identified pairs by the proposed PhyloBCG to be well supported by existing literature.
All the genera pairs discussed above are summarized in Table \ref{tab:pairs} with their estimated partial correlations and supporting references.
\section{Discussion}
In this work, we propose a phylogenetically informed Bayesian truncated copula graphical model for estimating microbial association networks with QMP data. The proposed method explicitly accounts for the zero-inflated nature of the QMP data and incorporates the microbial evolutionary information through the diffusion process and latent position model.
Simulation study with various phylogenetic tree structures reveals that the phylogenetic prior significantly improves network estimation accuracy. In particular, the proposed model shows much larger true positive rates while having comparable false positive rates to existing microbial network estimation models. Our sensitivity analysis shows reasonably stable performance under various hyperparameter settings. Also, we find that the proposed model is robust to tree misspecification in that the phylogenetic tree does not override conditional dependence supported by data. In application to the QMP data analysis, the proposed model identifies several unique genus-level conditional dependencies that are missed by existing microbial network estimation methods.
Although the proposed model is developed for microbial association network estimation, the established formulation of the truncated Gaussian copula can be directly applied to other zero-inflated data, such as single-cell RNA sequencing data. Furthermore, the framework for incorporating evolutionary information is not limited to undirected graph estimation and can be extended to the directed graph estimation models.
\section*{Acknowledgements}
The authors thank Grace Yoon for constructive discussions on QMP data analysis.
This work has been partially supported by the Texas A\&M Institute of Data Science (TAMIDS) and the Texas A\&M Strategic Transformative Research Program. Gaynanova's research was partially supported by the National Science Foundation (NSF CAREER DMS-2044823). Ni's research was partially supported by the National Science Foundation (NSF DMS-1918851). The \textsf{R} programs that reproduce the presented results are available at \url{https://github.com/heech31/phyloBCG}.
\section{Introduction}
\subsection{Microbial association network and its importance}
The gut is one of the most significant habitats of a myriad of microbial communities that play critical roles in their host's health. Well-balanced gut microbial communities provide many health benefits, such as maintaining metabolic homeostasis and high functioning immune system \citep{martinez2016gut,kim2016gut,cani2019microbial}. The imbalance of the gut microbiome (dysbiosis) has been related to a variety of human diseases
\citep{cho2012human,lynch2016human}.
The gut microbial balance is maintained by complex microbial interactions such as metabolites consumption, production, and exchange. Microbiome dysbiosis occurs when these interactions are interrupted by environmental alterations such as diet change, antibiotic consumption, and chemical exposure. These changes may deplete nutrients for beneficial microbes and create favorable surroundings for disease-causing bacteria to flourish. Nevertheless, some microbes help the microbial communities to maintain their stability under the environmental changes by providing energy sources and necessary metabolites \citep{zhang2019facing}. Because of the complexity of the functional roles of microbes, identifying microbial association networks, that is, microbe-microbe interaction networks, is crucial for fundamental understanding of the gut microbiome, a key contributor to the host's health.
\subsection{Motivating application: quantitative microbiome profiling data}
Microbiome data collected from 16S ribosomal RNA (rRNA) sequencing are compositional in that each subject has an arbitrary total microbial count determined by the sequencing instrument \citep{gloor2017microbiome}. Hence, a quantitative comparison of microbial abundances cannot be made across subjects as only the information on relative abundances within a subject are available from such data.
Furthermore, compositional data raise a concern for biased estimates of association since a change in absolute abundance of one microbe affects the relative abundance of all the microbes \citep{vandeputte2017quantitative}. The recently developed quantitative microbiome profiling (QMP) techniques account for these compositional limitations by adjusting microbial counts from 16S rRNA sequencing using cell counts and sequencing depths. In this work, we utilize this recent development by considering the QMP data of \citet{vandeputte2017quantitative} with $n=106$ healthy subjects' gut microbiome. We focus on estimating genus-level association networks with the aim of understanding the overall configurations of healthy gut microbial communities and their interactions.
\subsection{Graphical models and network estimation}
Gaussian graphical model is a popular tool for modeling an association network via an undirected graph, where an edge between the two nodes represents the conditional dependence, and an absence of an edge represents conditional independence. Under the Gaussian assumption, this graph structure is fully encoded in the concentration matrix (the inverse covariance matrix) as a zero off-diagonal entry is equivalent to the conditional independence between the corresponding variables. Thus, multiple methods focus on sparse estimation of concentration matrices. Neighborhood selection \citep{meinshausen2006high} recovers the sparse graph structure by performing $L_{1}$-regularized regression of each node on the rest. \citet{yuan2007model,banerjee2008model,dahl2008covariance,friedman2008sparse} directly estimate the sparse concentration matrix by optimizing the $L_{1}$-penalized log-likelihood function, the so-called graphical lasso. \citet{wang2012bayesian} propose a Bayesian counterpart of the graphical lasso using the Laplace prior on the off-diagonal elements of the concentration matrix. \citet{roverato2002hyper,dobra2011bayesian,lenkoski2011computational} consider a G-Wishart prior for the concentration matrix, of which the posterior inference is computationally more expensive than \cite{wang2012bayesian}. For better scalability, \citet{wang2015scaling} develop a continuous spike-and-slab prior for the off-diagonals of the concentration matrix. Furthermore, the Gaussian graphical models can be extended to non-Gaussian data via latent Gaussian copula models. \citet{liu2012high} consider Gaussian copula model for skewed continuous distributions. \citet{fan2017high} consider extension to mixed binary-continuous variables via latent Gaussian copula. \citet{dobra2011copula} consider a Bayesian latent Gaussian copula for graph estimation with binary and ordinal variables, where they approximate the likelihood function using the extended rank likelihood \citep{hoff2007extending}.
Despite the significant advancements in Gaussian graphical models, they are not appropriate for estimating microbial association networks. Microbiome data obtained from high-throughput sequencing are heavily right skewed and zero-inflated. The zeros, furthermore, are not necessarily absolute, but are often due to the limited sequencing depth. Thus, direct application of Gaussian graphical models to zero-inflated sequencing data leads to inaccurate estimation and inference. To address these challenges, several graphical models for zero-inflated data have been proposed. \citet{osborne2020latent} model microbial counts using Dirichlet-multinomial distribution with a latent Gaussian graphical model. \citet{zhou2020identification} consider a zero-inflated latent Ising model for microbial association network estimation. \citet{mcdavid2019graphical} propose a multivariate hurdle model, which is a mixture of degenerate (at 0) and Gaussian distributions. SPIEC-EASI \citep{kurtz2015sparse} is a two-stage inference procedure specifically designed for compositional microbiome data. \citet{yoon2019spring} propose Semi-Parametric Rank-based approach for INference in Graphical model (SPRING) based on truncated latent Gaussian copula \citep{yoon2020sparse}. \citet{ma2020joint} proposes truncated Gaussian graphical model.
\subsection{The major limitation of existing network estimation models and our proposal}
The aforementioned microbial network estimation models share a common limitation: they do not take advantage of additionally available evolutionary information for reverse-engineering the graph structure.
The information on microbes' genetic similarities is available in a form of a phylogenetic tree, however to our knowledge the phylogenetic tree is not taken into account by existing methods for estimation of microbial networks. Since microbial interactions, positive (e.g., mutualism) or negative (e.g., competition), increase with the increase in microbes' genetic similarity \citep{rohr2014, peralta2016merging}, evolutionary information encoded in a phylogenetic tree has great potential in improving the accuracy of microbial associations network estimation.
In this work, we propose a Bayesian truncated Gaussian copula graphical model for microbial association networks that takes advantage of available evolutionary information. Our major contributions are three-fold. First, we provide a general framework for incorporating evolutionary history into the estimation of microbe-microbe association networks. We model the phylogenetic tree as a Gaussian diffusion process in the latent space,
which allows us to represent the microbes and their ancestors as (correlated) Gaussian vectors. Our framework is not limited to the phylogenetic tree and can accommodate any prior knowledge that is expressed in a tree, e.g., a taxonomic rank tree. We formulate the prior probability model on graph so that the microbes that are closer to each other on the tree have a higher edge inclusion probability. Our simulation study reveals that our approach significantly improves the graph estimation accuracy compared to the methods that do not take advantage of the tree structure (Section~\ref{sec:simulation}). Second, the proposed model effectively handles zero-inflation resulting from limited sequencing depth. We consider the observed zeros as truncated realizations of unobserved random quantities that are below certain thresholds. In particular, we establish a Bayesian formulation of the truncated Gaussian copula model \citep{yoon2020sparse} and develop an efficient Gibbs sampling algorithm. Third, the proposed approach facilitates the statistical inference on the estimated network. For each pair of nodes, an edge connectivity is immediately available from the posterior sample, which provides a convenient way to control the posterior expected FDR \citep{mitra2013bayesian,peterson2015bayesian}. Finally, while our model is designed for quantitative microbiome data, it can also be applied to compositional data using modified central log ratio transformation \citep{yoon2019spring}.
The rest of this paper is organized as follows. In Section \ref{sec:model}, we introduce the proposed graphical model. In Section \ref{sec:inference}, we discuss posterior inference. In Section \ref{sec:simulation}, we evaluate the graph estimation accuracy of the proposed model on simulated datasets. In Section \ref{sec:application}, we analyze quantitative microbiome profiling data of \citet{vandeputte2017quantitative}, and compare our results to SPRING \citep{yoon2019spring} and SPIEC-EASI \citep{kurtz2015sparse}.
\section{Bayesian truncated Gaussian copula graphical model}\label{sec:model}
In Section \ref{ssec:copula}, we discuss the semiparametric modeling of conditional dependencies for zero-inflated data through a truncated Gaussian copula model with a sparse concentration matrix. In Section \ref{ssec:tree}, we introduce a prior model that incorporates the phylogenetic tree to facilitate posterior inference of microbial associations. The complete hierarchical model is summarized in Figure \ref{fig:schema}.
\begin{figure}[t]
\ctikzfig{tikz/BCG_scheme}
\caption{Schematic illustration of the phylogenetically informed Bayesian truncated Gaussian copula graphical model. Hyperparameters that are held constant are given in boxes with dashed-line. The quantities that need posterior inference are illustrated in boxes with solid-line. The ellipse with solid line represents observed data. }
\label{fig:schema}
\end{figure}
\subsection{Truncated Gaussian copula graphical model}\label{ssec:copula}
Let $\boldsymbol{x}=(x_{1},\ldots,x_{p})^{\top}$ denote the zero-inflated abundances of $p$ microbes. This is either directly the counts resulting from quantitative microbiome profiling, e.g. motivating data from \citet{vandeputte2017quantitative}, or transformed
compositional microbiome data via the modified central log-ratio transformation \citep{yoon2019spring}.
We propose to model $\boldsymbol{x}$ such that only the microbial abundances that are larger than certain thresholds can be observed. Specifically, we assume that there exist latent $\boldsymbol{x}^{*}=(x_{1}^{*},\ldots,x_{p}^{*})^{\top}$ representing the true abundances such that
\begin{flalign}\label{eq:truncation}
x_{j} &= 1(x_{j}^{*}>c_{j})x_{j}^{*}, \quad j=1,\ldots,p,
\end{flalign}
where $1(\cdot)$ is the indicator function and $c_{j}$ is the unknown threshold for the $j$th variable.
We call $x_j$ a \textit{truncated} variable if $x_{j}^{*}$ is less than $c_j$, and an \textit{observed} variable, otherwise. Let $F_{j}$ be the marginal cumulative distribution function (cdf) of the $j$th latent variable $x_{j}^{*}$, $\Phi$ be the cdf of standard Gaussian, and $f_{j}=\Phi^{-1}\circ F_{j}$, where we assume that $F_{j}$'s are continuous.
The truncated Gaussian copula model \citep{yoon2020sparse} assumes
\bfl{\label{eq:motr}
& z_{j} = f_{j} (x_{j}^{*}), \quad j=1\ldots,p,\\ \label{eq:copula}
&\boldsymbol{z}=(z_{1},\ldots,z_{p})^{\top} \sim \textup{N}_{p}(\boldsymbol{0},\boldsymbol{\Omega}^{-1}),
}
where $\boldsymbol{\Omega}^{-1} \succ 0$ is the positive definite correlation matrix.
Since $f_{j}$'s are monotone continuous, we can write \eqref{eq:truncation} as $x_{ij}=1\{f_{j}(x_{ij}^{*})>f_{j}(c_{j}) \}x_{ij}^{*}$ $=1(z_j>\delta_j )x_{ij}^{*}$, where $\delta_j=f_j(c_j)$. We denote the truncated and the observed sub-vectors of $\boldsymbol{x}$ by $\boldsymbol{x}_{t}\in \mathbb{R}^{p_{t}}$ and $\boldsymbol{x}_{o} \in \mathbb{R}^{p_{o}}$, respectively, where $p_{t}+p_{o}=p$. Likewise, let $\boldsymbol{z}_{t}$ and $\boldsymbol{z}_{o}$ be the corresponding latent Gaussian vectors, and let $\boldsymbol{\delta}_{t}$ and $\boldsymbol{\delta}_{o}$ be their thresholds.
Given the thresholds, the conditional distribution of $\boldsymbol{z}$ is given by
\begin{flalign}\label{eq:density}
p(\boldsymbol{z}_{o},\boldsymbol{z}_{t}|\boldsymbol{z}_{o}>\boldsymbol{\delta}_{o},\boldsymbol{z}_{t}<\boldsymbol{\delta}_{t},\boldsymbol{\Omega})=
\frac{\textup{N}_{p}(\boldsymbol{z}_{o},\boldsymbol{z}_{t}|\boldsymbol{0},\boldsymbol{\Omega}^{-1}) }{\PP(\boldsymbol{z}_{o}>\boldsymbol{\delta}_{o},\boldsymbol{z}_{t}<\boldsymbol{\delta}_{t}|\boldsymbol{\delta},\boldsymbol{\Omega})} 1(\boldsymbol{z}_{o}>\boldsymbol{\delta}_{o}) 1(\boldsymbol{z}_{t}<\boldsymbol{\delta}_{t}).
\end{flalign}
Unlike the approach in \citet{dobra2011copula} that only uses the relative ranks of the observed data, we condition on the observed value of $\boldsymbol{z}_{o}$, which subsequently allows us to sample truncated variables $\boldsymbol{z}_{t}$ from the posterior distribution as discussed in Section \ref{sec:inference}.
Because of the multivariate normality of $\boldsymbol{z}$, zero entries of $\boldsymbol{\Omega}=[\omega_{jk}]_{1\leq j, k\leq p}$ imply the conditional indepdendence between the corresponding variables. The dependency structure of $\boldsymbol{z}$ can be graphically summarized as an undirected graph $G=(\boldsymbol{z},\boldsymbol{E})$ with an adjacency matrix $\boldsymbol{E}=[e_{jk}]_{1\leq j, k\leq p}$, where nodes $z_j$ and $z_k$ are connected (denoted by $e_{jk}=1$) if $\omega_{jk}\neq0$.
Consequently, learning the graph structure (adjacency matrix) $\boldsymbol{E}$ is equivalent to finding the sparse pattern of $\boldsymbol{\Omega}$. To encourage sparsity, we follow a similar strategy as in \cite{wang2015scaling} by assigning a spike-and-slab prior on the off-diagonal elements of $\boldsymbol{\Omega}$ and an exponential prior on the diagonal elements,
\bfl{
\label{eq:SSSL2}
p(\boldsymbol{\Omega}|\boldsymbol{E},v_{0}^2) = C(\boldsymbol{E},v_{0}^2)^{-1}
1(\boldsymbol{\Omega}\succ 0) & \prod_{j<k} \Big\{ (1-e_{jk})\textup{N}(\omega_{jk}|0,v_{0}^2) + e_{jk}\textup{N}(\omega_{jk}|0,hv_{0}^2) \Big\} \\
\nonumber
\times & \prod_{j=1}^{p} \textup{Exp} (\omega_{jj}|\frac{\lambda}{2}),
}
where $ \textup{Exp} (\cdot|\lambda)$ is the exponential density function with rate parameter $\lambda$, $v_{0}^{2}$ is the spike variance, $h\gg1$ is a large constant such that the slab variance $hv_{0}^{2}\gg v_{0}^{2}$, and $C(\boldsymbol{E},v_{0}^2)$ is the normalizing constant.
We impose an improper uniform prior on the thresholds $\boldsymbol{\delta}=(\delta_{1},\ldots,\delta_{p})^{\top}\sim\pi(\boldsymbol{\delta})\propto 1$, a conjugate inverse-gamma prior on the spike variance $v_{0}^2 \sim \textup{IG}(a_{v_{0}},b_{v_{0}})$, and Bernoulli-like priors on the edge indicators
\bfl{
\label{eq:eprior}
p(\boldsymbol{E})\propto C(\boldsymbol{E},v_{0}^2)\prod_{j<k} \mbox{Ber}(e_{jk}|\pi_{jk}).
}
Including the normalizing constant $C(\boldsymbol{E},v_{0}^2)$ in \eqref{eq:eprior} serves to cancel out that in \eqref{eq:SSSL2}, facilitating the posterior computation in updating $\boldsymbol{E}$ \citep{wang2015scaling}.
\subsection{Incorporating phylogenetic tree}\label{ssec:tree}
Evolution plays an important role in shaping the interaction patterns of microbes \citep{peralta2016merging}. We will exploit the evolution footprints in identifying microbial association networks through a novel phylogenetic tree prior. The proposed prior is a distribution on edge inclusion probabilities $\boldsymbol{\Pi}=[\pi_{jk}]_{1\leq j, k\leq p}$ that encourages the interactions, positive (e.g., mutualism) or negative (e.g., competition), of phylogenetically similar microbes as they tend to be phenotypically/functionally correlated \citep{martiny2015microbiomes,xiao2018predictive,zhou2021bayesian}. The prior is constructed by first embedding the network in $L$-dimensional Euclidean space through the latent position model \citep{hoff2002latent}, and then arranging the latent positions according to the phylogenetic tree.
\textbf{Latent position model.}
We introduce a latent position $\boldsymbol{t}_j=(t_{1j},\dots,t_{Lj})^{\top}\in \mathbb{R}^L$ for each node $z_j$, and link it to the edge inclusion probability $\pi_{jk}$ through a probit link function
\bfln{
\pi_{jk} = \Phi( \boldsymbol{t}_{j}^{\top} \boldsymbol{t}_{k} ), \quad j<k.
}
The inner product $\boldsymbol{t}_{j}^{\top} \boldsymbol{t}_{k}$ measures the similarity between $\boldsymbol{t}_{j}$ and $\boldsymbol{t}_{k}$, with larger inner product leading to higher prior inclusion probability. We assign a prior on $\boldsymbol{t}_j$'s to encourage the interactions between phylogenetically similar microbes.
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{diffusiontree.pdf}\\
\caption{An illustrative example of a phylgenetic tree with $p=4$ microbes (left) and the corresponding diffusion process in $\mathbb{R}^{1}$ with $\sigma^2=3$ (right). The new branch $t_2$ is split from $t_1$ at the first divergence time $s_1=0.2$. Then new branches $t_3$ and $t_4$ are split from $t_1$ and $t_2$ at the divergence time $s_2=0.55$ and $s_3=0.85$, respectively. The green circle and blue triangle are the most common ancestors of the pairs $(t_1,t_3)$ and $(t_2,t_4)$, whose heights $s_2$ and $s_3$ are the correlations of the pairs, respectively. } \label{fig:diffusiontree}
\end{figure}
\textbf{Phylogenetic tree.}
Let $\boldsymbol{T}=[\boldsymbol{t}_{1},\ldots,\boldsymbol{t}_{p}]\in \mathbb{R}^{L\times p}$ and let $\boldsymbol{t}^\ell=(t_{1\ell},\dots,t_{p\ell})$ be the $\ell$th row of $\boldsymbol{T}$. We assume $\boldsymbol{t}^{\ell}\buildrel iid \over\sim \textup{N}_{p}(\boldsymbol{0},\sigma^2\boldsymbol{H})$ for $\ell=1,\dots,L$, where $\boldsymbol{H}$ is a correlation matrix that reflects the phylogenetic similarity. Our specific choice of $\boldsymbol{H}$ is motivated by the following diffusion process. Let $\mathcal{T}$ be a phylogenetic tree with terminal nodes representing the $p$ microbes under investigation and internal nodes representing their common ancestors. Starting from time 0 at the origin (root), the first branch $\boldsymbol{t}_{1}$ follows a Brownian motion with variance $\sigma^2$ until the divergence time $s_1 \in [0,1]$. Then it splits into two branches, $\boldsymbol{t}_{1}$ and $\boldsymbol{t}_{2}$, each following the same Brownian motion independently before they split at times $s_2,s_3\in [0,1]$ resulting in $\boldsymbol{t}_{3}$ and $\boldsymbol{t}_{4}$. This process repeats until the $p$ terminal nodes are reached at time 1. An illustrative example of the diffusion process is provided in Figure \ref{fig:diffusiontree}
This diffusion process defines a centered multivariate Gaussian distribution on the terminal nodes of $\mathcal{T}$ with covariance matrix $\sigma^2 \boldsymbol{H}$. We define the \textit{height} of each node (split for internal nodes) as its distance from the root (time from 0). The correlation of two terminal nodes equals the height of their most recent common ancestor, which is large for a phylogenetically similar microbes. This multivariate Gaussian prior, together with the latent position model, achieves the desired prior distribution of $\pi_{jk}$ that encourages interactions between phylogenetically similar microbes.
Lastly, we assign a conjugate inverse-gamma prior for the variance parameter $\sigma^2\sim \textup{IG}(a_{\sigma^2},b_{\sigma^2})$.
\section{Posterior inference}\label{sec:inference}
The proposed model is parameterized by $\{\boldsymbol{z},\boldsymbol{\delta}, \boldsymbol{\Omega}, \boldsymbol{E},v_{0}^2,\boldsymbol{T},\sigma^2\}$ of which the posterior distribution is not available in closed form.
We use a Markov chain Monte Carlo (MCMC) algorithm to draw posterior samples from the intractable posterior distribution. Section \ref{ssec:samplezt} discusses Gibbs steps for sampling $\boldsymbol{z}$ and $\boldsymbol{\delta}$ from their full conditional distributions. Section \ref{ssec:samplelatent} describes Gibbs steps for $\boldsymbol{T}$ and $\sigma^2$. In Section \ref{supp:spikeandslab} of the Supplemenatry Materials, we provide the Gibbs steps for $\boldsymbol{\Omega}, \boldsymbol{E},v_{0}^2$ by following \cite{wang2015scaling}.
\subsection{Full conditionals of truncated observations and thresholds} \label{ssec:samplezt}
Let $\boldsymbol{x}_{1},\ldots,\boldsymbol{x}_{n} \in \mathbb{R}^{p}$ be a sample from the model \eqref{eq:truncation}--\eqref{eq:copula} and let $\boldsymbol{z}_{1},\ldots,\boldsymbol{z}_{n} \in \mathbb{R}^{p}$ be the corresponding latent variables. As before, we use subscripts $t$ and $o$, respectively, to denote the truncated and observed sub-vectors (e.g., $\boldsymbol{x}_{i,t}$ and $\boldsymbol{x}_{i,o}$ are the truncated and observed sub-vectors of $\boldsymbol{x}_{i}$).
For observed $\boldsymbol{x}_{i,o}$, the corresponding Gaussian variables are defined as $z_{ij,o} =f_j(x_{ij,o})=\Phi^{-1}\circ F_{j}(x_{ij,o})$. A natural estimator for the unknown function $F_{j}$ is the scaled empirical cdf $\widehat{F}_{j}(c)=\{n/(n+1)\}\sum_{i=1}^{n} n^{-1}1(x_{ij}\leq c)$ \citep{klaassen1997efficient}, where the constant term $n/(n+1)$ is needed to make $\Phi^{-1}$ finite. Thus, the estimator of $f_j$ is given by $\Phi^{-1}\circ\widehat F_j$, and we set $\widehat{z}_{ij,o}=\Phi^{-1}\circ \widehat{F}_{j}(x_{ij,o})$. An alternative approach to estimate $f_j$ using B-spline basis functions has been considered in \citet{mulgrave2020bayesian}; however, we use the empirical cdf for computational efficiency.
Given $\boldsymbol{z}_{i,o}=\widehat{\boldsymbol{z}}_{i,o}$, we sample $\boldsymbol{z}_{i,t}$ from a truncated multivariate Gaussian distribution derived from \eqref{eq:density},
\bfln
p(\boldsymbol{z}_{i,t}|\widehat{\boldsymbol{z}}_{i,o},\boldsymbol{z}_{i,t}<\boldsymbol{\delta}_{t},\boldsymbol{\Omega})=
\frac{\textup{N}_{p_{i,t}}(\boldsymbol{z}_{i,t}|\boldsymbol{\mu}_{i},\boldsymbol{\Delta}_{i}) }{\PP(\boldsymbol{z}_{i,t}<\boldsymbol{\delta}_{t}|\widehat{\boldsymbol{z}}_{i,o},\boldsymbol{\Omega},\boldsymbol{\delta})} 1(\boldsymbol{z}_{i,t}<\boldsymbol{\delta}_{i,t}),
}
where $p_{i,t}$ is the number of truncated variables of $\boldsymbol{x}_{i}$, and $\boldsymbol{\mu}_{i}$, $\boldsymbol{\Delta}_{i}$ are the mean and covariance matrix of $\boldsymbol{z}_{i,t}$ given $\boldsymbol{z}_{i,o}=\widehat{\boldsymbol{z}}_{i,o}$ (detailed expressions are provided in Section \ref{supp:samplezt} of the Supplementary Materials).
The conditional pdf of $\boldsymbol{\delta}$ given ${\boldsymbol{z}}_{1},\ldots,\boldsymbol{z}_{n}$ is proportional to the $n$-product of indicator functions of \eqref{eq:density}. That is, we have the independent uniform full conditional distributions of $\delta_{j}$'s as
\bfln{
p(\boldsymbol{\delta}| {\boldsymbol{z}}_{1},\ldots,{\boldsymbol{z}}_{n},\boldsymbol{\Omega}) &\propto
\prod_{i=1}^{n}
\textup{N}_{p}({\boldsymbol{z}}_{i,t},\widehat{\boldsymbol{z}}_{i,o},|\boldsymbol{\Omega})
1({\boldsymbol{z}}_{i,t}<\boldsymbol{\delta}_{i,t}) 1(\widehat{\boldsymbol{z}}_{i,o}>\boldsymbol{\delta}_{i,o}) ,\\
&\propto \prod_{j=1}^{p} 1({z}_{j,t}^{\max}<\delta_{j}< \widehat{z}_{j,o}^{\min} ),
}
where ${z}_{j,t}^{\max}=\max_{i} {z}_{ij,t}$ and $\widehat{z}_{j,o}^{\min}=\min_{i} \widehat{z}_{ij,o}$ are the maximum and minimum of the truncated and observed components of the $j$th Gaussian variable, respectively.
\subsection{Full conditionals of latent positions and the tree scale}\label{ssec:samplelatent}
Let $\boldsymbol{T}_{-j}$ be the submatrix of $\boldsymbol{T}$ without the $j$th column.
The full conditional distribution of $\boldsymbol{t}_{j}$ is given by,
\bfln{
p(\boldsymbol{t}_{j}|\boldsymbol{T}_{-j},\boldsymbol{E},\sigma^2) \propto \left[ \prod_{k \neq j } \{ \Phi( \boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j} )\}^{e_{kj}} \{1- \Phi(\boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j} )\}^{1-e_{kj}} \right]
\textup{N}_{L}(\boldsymbol{t}_{j}|\boldsymbol{\theta}_{j},\boldsymbol{\Psi}_{j}), \quad j=1,\ldots,p,
}
where $\boldsymbol{\theta}_{j}$ and $\boldsymbol{\Psi}_{j}$ are mean and covariance matrix of $\boldsymbol{t}_{j}$ given $\boldsymbol{T}_{-j}$ (detailed expressions are provided in Section \ref{supp:samplelatent} of the Supplementary Materials). We update $\boldsymbol{t}_{j}$ using the data augmentation technique of \citet{albert1993bayesian} by introducing the auxiliary data $\boldsymbol{y}_{j} \in \mathbb{R}^{p-1}$. Let $\textup{TN}(\mu,\sigma^2,e)$ be $\textup{N}(\mu,\sigma^2)$ truncated to be positive if $e=1$ and negative if $e=0$. Conditining on $\boldsymbol{T}$, each component $y_{kj}$ of $\boldsymbol{y}_{j}$ follows $\textup{TN}(\boldsymbol{t}_{k}^{\top}\boldsymbol{t}_{j},1,e_{kj})$, and the resulting augmented pdf of $\boldsymbol{t}_{j}$ and $\boldsymbol{y}_{j}$ is
\bfln{
p(\boldsymbol{t}_{j},\boldsymbol{y}_{j}&|\boldsymbol{T}_{-j},\boldsymbol{E},\sigma^2) \propto \\
& \prod_{k \neq j}
\left\{ 1(y_{kj}>0, e_{kj}=1) + 1(y_{kj}<0,e_{kj}=0) \right\} \textup{N}(y_{kj}|\boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j}, 1) \textup{N}_{L}(\boldsymbol{t}_{j}|\boldsymbol{\theta}_{j},\boldsymbol{\Psi}_{j}).
}
We obtain a posterior sample of $\boldsymbol{T}$ by alternately sampling $\boldsymbol{y}_{j}$ and $\boldsymbol{t}_{j}$ for $j=1,\ldots,p$. Conditional on $\boldsymbol{T}$, we independently sample $y_{kj}$ from $\textup{TN}(\boldsymbol{t}_{k}^{\top} \boldsymbol{t}_{j},1,e_{kj})$ for $k\neq j$. Then, conditional on $\boldsymbol{y}_{j}$, we have the Gaussian full conditional of $\boldsymbol{t}_{j}$ as
\begin{flalign}\nonumber
p(\boldsymbol{t}_{j} |\boldsymbol{y}_{j}, \boldsymbol{T}_{-j},\boldsymbol{E},\sigma^2) \propto
\textup{N}_{p-1}(\boldsymbol{y}_{j}|\boldsymbol{T}_{-j}^{\top} \boldsymbol{t}_{j}, \boldsymbol{I}_{p-1}) \textup{N}_{L}(\boldsymbol{t}_{j}|\boldsymbol{\theta}_{j},\boldsymbol{\Psi}_{j}).
\end{flalign}
Accordingly, we draw $\boldsymbol{t}_{j}$ from $\textup{N}_{L}( \boldsymbol{\gamma}_{j}, \boldsymbol{\Delta}_{j} )$, where
\bfln{
\boldsymbol{\Delta}_{j} = \left( \boldsymbol{T}_{-j}\boldsymbol{T}_{-j}^{\top} + \boldsymbol{\Psi}_{j}^{-1} \right)^{-1}, \quad
\boldsymbol{\gamma}_{j} = \boldsymbol{\Delta}_{j}
\left( \boldsymbol{T}_{-j} \boldsymbol{y}_{j} + \boldsymbol{\Psi}_{j}^{-1} \boldsymbol{\theta}_{j} \right).
}
The edge inclusion probabilities are updated as $\pi_{jk}=\Phi(\boldsymbol{t}_{j}^{\top}\boldsymbol{t}_{k})$, $1\leq j<k \leq p$.
Conditional on $\boldsymbol{T}$, we sample the tree scale parameter $\sigma^2$ from
\bfln{
\sigma^2|\boldsymbol{T} \sim \textup{IG}\left(pL/2 + a_{\sigma^2}, \vect(\boldsymbol{T})^{\top}(\boldsymbol{H}\otimes \boldsymbol{I}_{L})^{-1}\vect(\boldsymbol{T})/2 + b_{\sigma^2}\right),
}
where $\vect(\boldsymbol{T})$ is the vector obtained by stacking the columns of $\boldsymbol{T}$ and $\otimes$ is the Kronecker product.
For sampling concentration matrix and graph, we follow the block Gibbs sampler of \citet{wang2015scaling}. For completeness, we provide the block Gibbs sampling algorithm in Section \ref{supp:spikeandslab} of the Supplementary Materials. Upon the completion of the MCMC, we compute the posterior mean of the edge inclusion probabilities for each pair of nodes, $\widehat{\pi}_{jk}=\sum_{s=1}^{S} e_{jk}^{(s)}/S$, where the superscript indexes posterior samples. We obtain the estimated graph by selecting edges for which $\widehat{\pi}_{jk}$ is larger than some cutoff.
We choose the cutoff to control the posterior expected FDR \citep{mitra2013bayesian,peterson2015bayesian} at prespecified level $\alpha$, where the posterior expected FDR is a decreasing function of cutoff $c$ defined as
\bfl{\label{eq:postfdr}
\textup{E}(\text{FDR}_{c}|\text{data})=\frac{\sum_{j<k}(1-\widehat{\pi}_{jk})1(\widehat{\pi}_{jk}> c)}{\sum_{j<k}1(\widehat{\pi}_{jk}> c)}.
}
\section{Simulation}\label{sec:simulation}
We simulate microbiome data following the data generation mechanism proposed in \cite{yoon2019spring}, which allows to obtain synthetic samples that exactly follow the empirical marginal cumulative distributions of measured microbiome count data while respecting user-specified microbial dependencies via $\boldsymbol{\Omega}$.
Specifically, we randomly generate 10 phylogenetic trees $\mathcal{T}_{1},\ldots,\mathcal{T}_{10}$ with $p=50$ terminal nodes by the \textsf{R} package \textsf{ape} \citep{paradis2019ape}, constituting 10 simulation scenarios. For each tree, the latent positions of terminal nodes, $\boldsymbol{t}_{1},\ldots,\boldsymbol{t}_{p}$, are generated from the diffusion process as described in Section \ref{ssec:tree} with $\sigma^2=3$ and $L=2$. The true graph adjacency matrix $\boldsymbol{E}_0$ is obtained by independently generating $e_{jk} \sim \textup{Bernoulli} (\pi_{jk})$ for $1\leq j<k \leq p$ with $\pi_{jk} = \Phi(\boldsymbol{t}_{j}^{\top} \boldsymbol{t}_{k})$. The trees and the true graphs are plotted in Section \ref{supp:phylotrees} of the Supplementary Materials.
Given $\boldsymbol{E}_0$, the concentration matrix $\boldsymbol{\Omega}$ is drawn from G-Wishart$(\boldsymbol{I}_{p},4)$ \citep{roverato2002hyper}.
To obtain empirical cdfs, we use the quantitative microbiome profiling data of \citet{vandeputte2017quantitative} from $n=106$ subjects, more detailed description of the data is provided in Section \ref{sec:application}. We select $p = 50$ genera (variables) of which 6 genera have no observed zero counts, and 44 genera have 20\% to 70\% zero counts across samples.
Given the empirical cdf $\widehat{F}_{j}$ of each selected genera, $j=1,\ldots, p$, and the concentration matrix $\boldsymbol{\Omega}$, we generate $n=106$ independent latent Gaussian vectors $\boldsymbol{z}_{i}\sim\textup{N}_{p}(\boldsymbol{0}, \boldsymbol{\Omega}^{-1})$. The final data $\boldsymbol{x}_1,\dots,\boldsymbol{x}_n$ are obtained as
\bfln{
x_{ij} = \widehat{F}_{j}^{-1}\circ \Phi(z_{ij}), \quad i=1,\ldots,n, \quad j=1,\ldots,p.
}
We consider 50 independent replications of this data generating process for each scenario.
We compare the performance of the the proposed phylogenetically-informed Bayesian Copula Graphical model (PhyloBCG) with SPIEC-EASI \citep{kurtz2015sparse} and SPRING \citep{yoon2019spring}. Additionally, we also consider two special cases of PhyloBCG with the following simplification to the prior model of graph:
\begin{flalign*}
{\textup{Oracle}} :~& \pi_{jk} = \binom{p}{2}^{-1}|\boldsymbol{E}_0|;\\
{\textup{Dist}}:~& \pi_{jk} = \exp(-\gamma d_{jk} ), \quad \gamma \sim \textup{Exp} (1);
\end{flalign*}
where $|\boldsymbol{E}_0|$ is the number of true edges in the underlying graph, and $d_{jk}$ is the tree distance between terminal nodes $j$ and $k$, defined as the sum of the branch lengths to their most recent common ancestor. We refer to the first model as ``${\textup{Oracle}} $'' since it uses the knowledge of true graph sparsity, however it does not use any tree information. We refer to the second model as ``${\textup{Dist}}$'', as it directly incorporates tree distances between the terminal nodes: the edge inclusion probability is higher (smaller) when the corresponding nodes are closer (farther) to each other on the tree. ${\textup{Dist}}$ model thus takes into account the information from the tree, however, the tree information is deterministically incorporated to the model in contrast to the stochastic incorporation of PhyloBCG.
To implement SPIEC-EASI and SPRING, we use the corresponding R packages \citep{kurtz2017spieceasi,yoon2019springpackage} with sparsity parameters tuned over 100 values.
For PhyloBCG, Oracle and Dist, the hyperparameters are fixed as $a_{\sigma^2}=b_{\sigma^2}=a_{v_{0}^{2}}=b_{v_{0}^{2}}=0.001$, $h=2500$, $\lambda=1$, and $L=2$. PhyloBCG is relatively robust to the hyperparameter setting; see the sensitivity analyses in Section \ref{supp:sensitivity} of the Supplementary Materials. We obtain an MCMC posterior sample of size $S=5000$ after 500 burn-in iterations. The graph estimate is obtained by thresholding the mean of posterior inclusion probability $\widehat{\pi}_{jk}$ at $c_{0.05}$, the smallest $c$ that controls the posterior expected FDR \eqref{eq:postfdr} at level 0.05.
We assess the graph recovery performance (accuracy in estimating $\boldsymbol{E}_{0}$) of each method using Matthews correlation coefficient \citep{matthews1975comparison}, true positive rate, and false positive rate, which will be denoted by MCC, TPR, and FPR, respectively. Ranging from -1 to 1, a larger value of MCC represents a better network estimation accuracy, where the two boundary values indicate completely correct (+1) and wrong (-1) edge selection, respectively.
Figure \ref{fig:simsum} summarizes the mean values of these metrics for each of the 10 phylogenetic trees based on 50 replications. The phylogenetic trees are sorted in terms of the global clustering coefficient \citep{wasserman1994social} of the true graph from largest value ($\mathcal{T}_{1}$) to lowest value ($\mathcal{T}_{10}$). A large value of the global clustering coefficient indicates a presence of microbial communities, with dense interactions within the same community and sparse interactions across communities. A small value of the global clustering coefficient indicates a random interaction pattern close to what will be expected with Erdos-Renyi random graph. Thus, we anticipate the phylogenetic tree to be more informative for network estimation when the global clustering coefficient is larger.
\begin{figure}[!t]
\centering
\includegraphics[width=1\textwidth]{simsum_cdf_all_ejk_fdr_5.pdf}\\
\caption{Averages of Matthews correlation coefficient (MCC), true positive rates (TPR) and false positive rates (FPR) with 2 standard error bars. The phylogenetic trees are ordered in term of the global clustering coefficients. Averages are taken over 50 replicated data sets.} \label{fig:simsum}
\end{figure}
Figure \ref{fig:simsum} supports that incorporating phylogenetic tree information improves the network estimation accuracy, with Phylo and Dist having higher MCC, higher TPR, and similar FPR values when compared to other methods. As expected, the value of MCC for the tree-based methods decreases as the phylogenetic tree becomes less informative (larger tree index). For PhyloBCG, this trend is driven by the decreasing TPR, whereas, for Dist, it is the increasing FPR. Although PhyloBCG and Dist both use the evolutionary information, PhyloBCG shows significantly better performances than Dist possibly due to the flexibility of the latent space embedding.
Note that the FPR of PhyloBCG has larger variability for trees $\mathcal{T}_{5}$ and $\mathcal{T}_{10}$ than others, with $\mathcal{T}_5$ leading to the largest mean FPR. The reason for increased FPR in these settings is the discrepancy between the phylogenetic tree and the true graph. Recall that the true edge inclusion probabilities $\pi_{jk}$ are obtained from the Gaussian latent positions rather than directly from the tree, thus allowing the true graph to deviate, sometimes significantly, from the phylogenetic tree. To illustrate this phenomenon, Figure \ref{fig:violin_hjk} in the Supplementary Materials shows the upper triangular part of the true tree correlation matrix $\boldsymbol{H}$ defined in Section~\ref{ssec:tree} against the edge indicators for $\mathcal{T}_{1}-\mathcal{T}_{10}$. Both $\mathcal{T}_{5}$ and $\mathcal{T}_{10}$ show a large number of disconnected edges, $e_{jk}=0$, with high tree correlation values which may contribute to the increase in FPR. Nevertheless, for $\mathcal{T}_{5}$, ${\textup{PhyloBCG}}$ still shows favorable performance, having much larger TPRs than SPIEC-EASI and SPRING. Under $\mathcal{T}_{10}$, PhyloBCG shows comparable performance to Oracle. Considering that the prior edge inclusion probability of ${\textup{Oracle}} $ relies on the knowledge of the true number of connected edges, whereas the proposed PhyloBCG model does not and performs at least as good as the ${\textup{Oracle}} $, we conclude that PhyloBCG adapts well to the unknown sparsity of the underlying graph.
\section{Application to quantitative gut microbiome profiling data}\label{sec:application}
\subsection{Data and phylogenetic tree}
We focus on estimating genus-level association network of the QMP data \citep{vandeputte2017quantitative} that consists of $n=106$ healthy subjects' gut microbiome. We use the data as processed in \citet{yoon2019spring}, which can be obtained from the R-package \textsf{SPRING} \citep{yoon2019springpackage}. Among the 91 genera, there are 33 genera with missing names and 4 genera without available phylogenetic information on the National Center for Biotechnology Information (NCBI) database on which we based the phylogenetic tree construction. Consequently, we consider $p=54$ genera and obtain their phylogenetic tree based on the NCBI taxonomy database using the platform PhyloT\footnote{https://phylot.biobyte.de}. As the database does not provide divergence times of branches, we match the branch lengths with the taxonomic ranks as illustrated in Figure \ref{fig:qmptreemembership}. These genera have up to 70\% zeros (14 have more than 50\% zeros).
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth]{figures/membership.tree.4chains.pdf}
\caption{The phylogenetic tree of 54 genera. Different numbers (colors) represent three microbial communities derived from the graph estimated by the proposed PhyloBCG; the asterisk indicates a stand-alone node. }\label{fig:qmptreemembership}
\end{figure}
\subsection{Analysis}
For PhyloBCG, we use the same hyperparameter values as in Section \ref{sec:simulation} and run 4 parallel Markov chains for 100,000 iterations after 25,000 burn-in iterations. We then concatenate the 4 chains and obtain a posterior sample of size 10,000 by retaining every 40th iteration, from which we compute the posterior means $\widehat{\pi}_{jk}$ and $\widehat{\boldsymbol{\Omega}}$.
The microbial association network is estimated by controlling the posterior expected FDR at 0.1 which results in $c_{0.1}=0.719$.
SPIEC-EASI and SPRING are tuned using 100 sparsity parameter values. The default stability threshold \citep{liu2010stability} is changed from 0.1 to 0.2 to avoid overly sparse network estimates. The recovered networks are shown in Figure \ref{fig:qmpnetwork}.
\afterpage{
\begin{figure}[ht!]
\centering
\includegraphics[width=0.91\textwidth]{figures/network.tree.4chains.pdf}
\includegraphics[width=0.90\textwidth]{network.sp_1e-3.pdf}
\includegraphics[width=0.90\textwidth]{RMPnetwork.se_1e-3.pdf}
\caption{ Estimated microbial association networks of PhyloBCG, SPRING, and SPIEC-EASI with 54 genera in QMP data \citep{vandeputte2017quantitative}.
Communities are separately estimated with the graph estimate of each model, and node colors represent the memberships.}\label{fig:qmpnetwork}
\end{figure}
}
\subsubsection{Overall network summary and interpretation}
We first compare estimated networks in terms of their density and community structure. The estimated network from PhyloBCG appears to be denser and have much more definitive communities than those from SPIEC-EASI and SPRING.
While we do not know the true network and community structure of the gut microbiota in the study population, the following reasons support our belief that the additional findings of microbial interactions and communities from PhyloBCG are biologically meaningful.
First, microbes are known to form communities \citep{pflughoeft2012human}. Applying the edge proximity measure of \citet{newman2004finding} to the estimated network by PhyloBCG, we find three evident microbial communities which are marked with different colors in Figures \ref{fig:qmptreemembership} and \ref{fig:qmpnetwork}. Posterior mean latent positions, illustrated in Figure \ref{fig:post_latentpositions}, also form three distinct clusters that consistently match the estimated microbial communities.
Interestingly, the genera within each of these communities tend to share unique characteristics. On the one hand, most of the genera from the top two communities in Figure \ref{fig:qmptreemembership} are obligate anaerobes which only survive in the absence of oxygen.
On the other hand, the community located at the bottom of Figure \ref{fig:qmptreemembership} contains genera with species that need or at least can tolerate oxygen. For example, \textit{Streptococcus} and \textit{Enterococcus} contain facultative anaerobic species that are able to utilize oxygen as a source of energy, but can also generate energy anaerobically in an oxygen-deficient environment \citep{fisher2009ecology,clewell1981streptococcus}. All the members of \textit{Haemophilus} are facultatively anaerobic species or aerobic species, where aerobic species need oxygen to survive \citep{cooke20171611haemophilus}. \textit{Lactobacillus} contains aerotolerant and microaerophilic species \citep{zheng2020lactobacillus}. Aerotolerant species do not need oxygen and use anaerobic fermentation to generate energy, but oxygen is not toxic to them. Microaerophiles need oxygen to survive but require low oxygen concentration to thrive and are damaged by high oxygen level, e.g., atmospheric oxygen level. Furthermore, the bottom community has very few interactions with the two top communities, possibly because of their distinct living environments. By contrast, the estimated microbial interaction networks from SPIEC-EASI and SPRING do not present obvious communities; this lack of community structure is, based on existing literature \citep{rohr2014, peralta2016merging}, unlikely.
Second, while the phylogenetic tree prior helps identify additional interactions and communities, it does not dictate the posterior inference. Some of the genera from the top right community in Figure \ref{fig:qmptreemembership} (e.g., \textit{Succinivibrio} and \textit{Prevotella}) are not phylogenetically similar to each other. Similarly, the other two communities also contains phylogenetically distant genera. This suggests that the phylogenetic tree prior does not override the information of associations that are strongly supported by the data.
Third, from the simulation study, we have seen that PhyloBCG is much more powerful in detecting interactions than SPIEC-EASI and SPRING especially when there is a clear community structure, while also having comparable FPR.
Given that the communities of the estimated network by PhyloBCG seem biologically plausible, the additional interactions found by PhyloBCG are more likely to be true positives than false discoveries, some of which will be explained in detail in the next section.
\begin{figure}[t!]
\centering
\includegraphics[width=1\textwidth]{figures/latent_positions.pdf}
\caption{ Posterior mean latent positions of 54 genera in QMP data \citep{vandeputte2017quantitative}. Community memberships estimated with the graph estimate of PhyloBCG are marked by colored symbols. Black squares represent stand alone genera. }\label{fig:post_latentpositions}
\end{figure}
\subsubsection{Detailed explanation of interactions}
\textbf{All models} identify strong negative partial correlations between \textit{Dialister} and \textit{Phascolarctobacterium}. This finding is in agreement with multiple published results. The original QMP study \citep{vandeputte2017quantitative} finds a strong negative correlation between these two genera. \citet{naderpoor2019faecal} report that \textit{Phascolarctobacterium} (\textit{Dialister}) is positively (negatively) correlated with insulin sensitivity. Consistently, \citet{pedrogo2018gut} indirectly observe a trade-off relationship between the two genera from obese groups.
Besides, strong positive partial correlations between \textit{Oscillospira} and \textit{Ruminococcus}
are found by all methods as well, which is consistent with the finding in \citet{chen2020high}.
\textbf{PhyloBCG and SPRING} detect relatively strong partial correlations for the pairs (\textit{Mitsuokella}, \textit{Prevotella}) and (\textit{Ruminococcus}, \textit{Blautia}). In the phylogenetic tree displayed in Figure \ref{fig:qmptreemembership}, \textit{Ruminococcus} and \textit{Blautia} are relatively close to each other, being the members of the same order, Clostridales. \textit{Mitsuokella} and \textit{Prevotella} are, however, phylogenetically distant from each other, indicating that the QMP data present a strong association between them and that the tree prior of the proposed PhyloBCG does not dominate the inference. These findings agree with the network analysis of \citet{ramayo2016phylogenetic}, where they also detect positive partial correlations in the pairs (\textit{Mitsuokella}, \textit{Prevotella}) and (\textit{Ruminococcus}, \textit{Blautia}).
Additionally, the pairs (\textit{Oscillospira}, \textit{Butyricimonas}) and (\textit{Eubacterium}, \textit{Peptococcus}) also show relatively strong partial correlations under the two models. Both pairs are known to be related to diet and leanness. \citet{oh2020association} discover positive correlations between body weight and each of \textit{Eubacterium} and \textit{Peptococcus}. \citet{garcia2018shifts} find negative correlations between unhealthy diet (high intake of saturated fat and refined carbohydrates) and each of \textit{Oscillospira} and \textit{Butyricimonas}.
Furthermore, positive partial correlations for the pairs (\textit{Bacteroides}, \textit{Bilophila}) and (\textit{Akkermansia}, \textit{Methanobrevibacter}) are observed, where these results match the analyses in \citet{vandeputte2016stool} and \citet{vandeputte2017quantitative}, respectively.
\begin{table}[t]
\centering
\caption{Selected pairs of genera with strong association as identified by PhyloBCG. ``--'' indicates that no significant partial correlation is found by the corresponding.}\label{tab:pairs}
\resizebox{1\textwidth}{!}{
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{ll|ccc|l}
\multicolumn{2}{c|}{\multirow{2}{*}{Pairs of Microbial Genera}}&\multicolumn{3}{c|}{Partial Correlations}& \multirow{2}{*}{Reference} \\\cline{3-5}
&& PhyloBCG & SPRING & SPIEC-EAIS & \\
\hline
Dialister & Phascolarctobacterium & -0.384 & -0.185 & -0.252 & \citet{vandeputte2017quantitative}, \citet{naderpoor2019faecal} \\
Oscillospira & Ruminococcus & 0.129 & 0.320 & 0.335 & \citet{chen2020high} \\
Mitsuokella & Prevotella & 0.155 & 0.152 & -- & \citet{ramayo2016phylogenetic} \\
Ruminococcus & Blautia & 0.115 & 0.048 & -- & \citet{ramayo2016phylogenetic} \\
Oscillospira & Butyricimonas & 0.155 & 0.335 & -- & \citet{garcia2018shifts}, \citet{thomaz2021influence} \\
Eubacterium & Peptococcus & 0.178 & 0.083 & -- & \citet{oh2020association} \\
Bacteroides & Bilophila & 0.137 & 0.221 & -- & \citet{vandeputte2017quantitative} \\
Akkermansia & Methanobrevibacter & 0.202 & 0.354 & 0.054 & \citet{vandeputte2016stool} \\
Blautia & Methanobrevibacter & -0.062 & -- & -0.329 & \citet{garcia2018shifts}, \citet{muller2019distal} \\
Prevotella & Bacteroides & -0.107 & -- & -0.019 & \cite{vandeputte2017quantitative} \\
& & & & & \citet{ley2016prevotella}, \citet{johnson2017microbiome} \\
Veillonella & Streptococcus & 0.149 & -- & -- & \citet{anbalagan2017next}, \citet{chen2020gut}\\
& & & & & \citet{den2013diversity}, \citet{egland2004interspecies} \\
& & & & & \citet{zoetendal2012human}, \citet{van2014immunomodulatory} \\
Bifidobacterium & Holdemania & -0.167 & -- & -- & \citet{liu2017fructooligosaccharide}, \citet{yang2018dietary} \\
& & & & & \citet{wang2020rational} \\
\hline
\end{tabular}
}
\end{table}
\textbf{PhyloBCG and SPIEC-EASI} find negative partial correlations for the pairs (\textit{Blautia}, \textit{Methanobrevibacter}) and (\textit{Prevotella}, \textit{Bacteroides}). \textit{Blautia} and \textit{Methanobrevibacter} are known to be positively and negatively related to dietary fiber intake, respectively \citep{garcia2018shifts}. \citet{muller2019distal} suggest that their inverse relationship is possibly due to substrate competition as both use hydrogen as energy source.
For \textit{Prevotella} and \textit{Bacteroides}, \citet{lozupone2012diversity} find the trade-off between these two genera -- carbohydrates (including simple sugars) focused diet increases \textit{Prevotella} and decreases \textit{Bacteroides} whereas protein and fat focused diet has the opposite effects on them. Their trade-off relationship is also discussed in \citet{ley2016prevotella} and \citet{johnson2017microbiome}. On the contrary, \citet{vandeputte2017quantitative} claim that their negative association is an artifact of using compositional rather than quantitative microbiome data for analyses .
\textbf{PhyloBCG} uniquely discovers positive partial correlations for the pairs (\textit{Veillonella}, \textit{Streptoccocus}) and (\textit{Bifidobacterium}, \textit{Holdemania}). The estimated positive partial correlation between \textit{Veillonella} and \textit{Streptoccocus} is consistent with that of the gut microbial network analysis of \citet{chen2020gut}. \citet{anbalagan2017next} demonstrate their mutualistic relationship: \textit{Streptoccocus} uses glucose as a source of carbon and release lactic acid, whereas \textit{Veillonella} utilizes lactic acid as carbon and energy source for growth. There are also many studies reporting their co-occurrence and mutualism \citep{den2013diversity,egland2004interspecies, zoetendal2012human,van2014immunomodulatory}.
For \textit{Bifidobacterium} and \textit{Holdemania}, \citet{liu2017fructooligosaccharide} report that prebiotic supplement significantly increases relative abundance of beneficial \textit{Bifidobacterium} and decreases \textit{Holdemania}, where \textit{Holdemania} is reported to be associated with unhealthy gut and antibiotic use \citep{yang2018dietary}. \citet{wang2020rational} discuss the underlying mechanism of the trade-off relationship.
In summary, we find these uniquely identified pairs by the proposed PhyloBCG to be well supported by existing literature.
All the genera pairs discussed above are summarized in Table \ref{tab:pairs} with their estimated partial correlations and supporting references.
\section{Discussion}
In this work, we propose a phylogenetically informed Bayesian truncated copula graphical model for estimating microbial association networks with QMP data. The proposed method explicitly accounts for the zero-inflated nature of the QMP data and incorporates the microbial evolutionary information through the diffusion process and latent position model.
Simulation study with various phylogenetic tree structures reveals that the phylogenetic prior significantly improves network estimation accuracy. In particular, the proposed model shows much larger true positive rates while having comparable false positive rates to existing microbial network estimation models. Our sensitivity analysis shows reasonably stable performance under various hyperparameter settings. Also, we find that the proposed model is robust to tree misspecification in that the phylogenetic tree does not override conditional dependence supported by data. In application to the QMP data analysis, the proposed model identifies several unique genus-level conditional dependencies that are missed by existing microbial network estimation methods.
Although the proposed model is developed for microbial association network estimation, the established formulation of the truncated Gaussian copula can be directly applied to other zero-inflated data, such as single-cell RNA sequencing data. Furthermore, the framework for incorporating evolutionary information is not limited to undirected graph estimation and can be extended to the directed graph estimation models.
\section*{Acknowledgements}
The authors thank Grace Yoon for constructive discussions on QMP data analysis.
This work has been partially supported by the Texas A\&M Institute of Data Science (TAMIDS) and the Texas A\&M Strategic Transformative Research Program. Gaynanova's research was partially supported by the National Science Foundation (NSF CAREER DMS-2044823). Ni's research was partially supported by the National Science Foundation (NSF DMS-1918851). The \textsf{R} programs that reproduce the presented results are available at \url{https://github.com/heech31/phyloBCG}.
|
\section{Introduction}
We study the distributions of the number of the occurrences of the words (the distribution of words for short) of finite alphabet,
which plays an important role in information theory, ergodic theory, computer science, and DNA analysis,
see \cite{{JaquetandSzpankowski},{waterman},{bertheRigo},{lothire},{bookRobin}}.
The data compression scheme LZ are based on the statistics of the occurrences of words \cite{{LZ77},{Ziv78},{shields96}} and
are applied to nonparametric statistics \cite{LV2008}.
R\'{e}gnier and Szpankowski \cite{RegnierSpankowski98} derived generating functions of the distributions of words
in a finite sample under finite Markov processes.
Goulden and Jackson \cite{Goulden83} and Bassino et al.~\cite{Bassinoetal2011} obtained the
generating functions of the distributions of words via cluster method and inclusion-exclusion principle.
In \cite{{nielsen},{li},{feller1}} average return time of words are analyzed.
Classical studies on the generating functions of the distributions of words \cite{{Guibas81},{RegnierSpankowski98},{JaquetandSzpankowski},{Goulden83},{Bassinoetal2011},{feller1}}
involve the variable of sample size and are in the form of rational function \(f/g\) where \(f\) and \(g\) are polynomial.
For example let \(F(z):=\sum_n f(n)z^n\), where \(f(n)\) is the number of binary strings of length \(n\) that does not contain
the word \(01\).
Eq.~(\ref{eqGuibas}) for \(|z|<1\) is given as a special case of the theorem in Guibas and Odlyzko \cite{Guibas81} (pp.61 \cite{Flajolet2009}).
Since
\begin{align}
F(z) &=\frac{1}{1-2z+z^2}\label{eqGuibas}\\
&=\frac{1}{(1-z)^2}=(\sum_n z^n)^2=\sum_n (n+1)z^n,\nonumber
\end{align}
we have \(f(n)=n+1\).
For example, \(f(3)=4\) and the four words, 000, 100, 110, and 111 do not contain the word 01 among the strings of length 3.
In general, the multivariate generating functions of the distributions of words with the variables of the number of words and sample size
are rational functions and the asymptotic formula of the distributions of
words are obtained from the generating functions \cite{{JaquetandSzpankowski},{Flajolet2009}}. However, except for simple cases, the exact value of the coefficients of the expansion of the rational functions are complicated and
we do not have a simple formula of the exact distributions of words from the rational functions \cite{feller1}.
In this paper, for each fixed sample size, we show the finite dimensional generating functions of the probabilities of the number of nonoverlapping words by inclusion-exclusion principle and present the explicit formula of the exact distributions of words
(Theorem~\ref{th-main2}).
The number of the appearance of \(01\) is almost same to the half of the number of runs.
For example the number of runs and the word \(01\) in \(00110011\) are 4 and 2, respectively.
Statistical tests based on the number of the appearance of words are considered to be a generalization of the run tests.
Let \(w_0w_1\cdots\) be a pseudo random numbers where \(w_n=f(w_{n-1})\) for all \(n\) and \(f\) is a pseudo random number generator.
Then \(w_{n-1}\ne w_n\) for all \(n\) and the empirical distribution of the word \(ww\) in the sequence \(w_0w_1\cdots\) will be different from the fair-coin flipping if \(|w|\) is large.
It is unpractical to test the number of the occurrence of the word \(ww\) when \(|w|\) is large since
we need large sample to detect the small difference between the empirical distribution and null hypothesis with Kolmogorov-Smirnov test.
To overcome this difficulty, in Section~\ref{secpartial} we introduce partial word that allows the special letter ? \cite{berstelBoasson}.
Here ? represents arbitrary letter.
Roughly speaking partial words are sparse pattern of letters.
The partial word has two contradicting properties, i.e., both the probability and the length of the word are large.
Our theorems can be applied to nonoverlapping partial words.
We can detect the difference of the distributions of large size sparse pattern of letters (nonoverlapping partial words) with
practical sample size.
In Section~\ref{secPow} we consider statistical tests based on the distributions of words (sliding sampling test) and the distributions of the number of block wise occurrence of words (block wise sampling test).
We show that the power function of the sliding sampling test is much larger than that of the block wise sampling test.
In Section~\ref{sec-pseudo} we apply the distribution of words to statistical test for random numbers.
Finally in Section~\ref{dna}, we report the experiment on the computation of the distribution of words in human DNA size string.
The preliminary versions of the paper have been presented at \cite{{takahashi2018MSJ},{takahashi2018ergod},{takahashi2018ProbSympo},{takahashi2018SITA},{takahashiBernoulliPoster2020},{takahashiPosterBio}}.
For the proof of Theorem~\ref{th-main}, see \cite{takahashi2018ProbSympo}.
\section{Main results}
Let \(X^n:=X_1\ldots X_n\) be random variables that take value in finite alphabet \({\cal A}\) and
\(N(w_1,\ldots,w_l; X^n)\) be the vector of the number of the appearance of the word \(w_1,\ldots,w_l\) in an arbitrary position of \(X^n\).
For example \(N(10,11;1011101)=(2,2)\).
Let \(|x|\) be the length of \(x\).
A word \(x\) is called overlapping if there is a word \(z\) such that \(x\) appears at least 2 times in \(z\) and \(|z|<2|x|\) otherwise
\(x\) is called nonoverlapping.
A pair of words \(x,y\) is called overlapping if there is a word \(z\) such that \(x\) and \(y\) appear in \(z\) and \(|z| < |x|+|y|\)
otherwise the pair is called nonoverlapping.
A word \(x\) is overlapping if and only if \((x,x)\) is overlapping.
A finite set of words \(S\) is called nonoverlapping if every pair \((x,y),\ x,y\in S\) are nonoverlapping.
If a finite set of words \(S\) is nonoverlapping, each element of \(S\) is nonoverlapping since \((x,x), x\in S\) is nonoverlapping.
For example \(11,\ \{10,01\},\ \text{and }\{00,11\}\) are overlapping, and \(10\text{ and } \{00111,00101\}\) are nonoverlapping.
\begin{theorem}\label{th-main2}
Let \(X_1X_2\cdots X_n\) be an i.i.d.~process of fixed sample size \(n\) of finite alphabet.
Let \(w_1,\ldots, w_l\) be the set of nonoverlapping words of finite alphabet.
Let \(m_i:=|w_i|\) and \(P(w_i)\) be the probability of \(w_i\) for \(i=1,\ldots, l\).
Let
\begin{align}
&A(k_1,\ldots,k_l)=\dbinom{n-\sum_i m_i k_i + \sum_i k_i}{k_1,\ldots, k_l}\prod_{i=1}^l P^{k_i}(w_i),\nonumber\\
&B(k_1,\ldots,k_l)=P(\sum_{i=1}^n I_{X_i^{i+m_i-1}=w_j}=k_j,\ j=1,\ldots,l),\label{eq-A}\\
&F_A(z_1,\ldots,z_l)=\sum_{k_1,\ldots,k_l}A(k_1,\ldots,k_l)z^{k_1}\cdots z^{k_l},\text{ and}\nonumber\\
&F_B(z_1,\ldots,z_l)=\sum_{k_1,\ldots,k_l}B(k_1,\ldots,k_l)z^{k_1}\cdots z^{k_l}.\nonumber
\end{align}
Then
\[
F_A(z_1,z_2,\ldots,z_l)=F_B(z_1+1, z_2+1,\ldots,z_l+1),
\]
and
\begin{align}\label{eqmain}
& P(N(w_1,\ldots,w_l;X^n)=(s_1,\ldots,s_l))\nonumber\\
& =\sum_{k_1,\ldots,k_l\ : \ s_1\leq k_1,\ldots,s_l\leq k_l,\ 0\leq n-\sum_i m_i k_i} \
\nonumber\\
& (-1)^{\sum_i k_i-s_i}\dbinom{n-\sum_i m_i k_i + \sum_i k_i}{s_1,\ldots, s_l,k_1-s_1,\ldots k_l-s_l}\prod_{i=1}^l P^{k_i}(w_i) .
\end{align}
\end{theorem}
Theorem~\ref{th-main2} is equivalent to Theorem~\ref{th-main} when \(l=1\).
\begin{theorem}\label{th-main}
Let \(X_1X_2\cdots X_n\) be an i.i.d.~process of fixed sample size \(n\) of finite alphabet.
Let \(w_1\sqsubset w_2\sqsubset\cdots \sqsubset w_l\) be strictly increasing nonoverlapping words of finite alphabet, i.e., \(w_i\) is a prefix of \(w_j\) and \(m_i<m_j\), where \(m_i\) is the length of \(w_i\), for all \(i<j\).
Let \(P(w_i)\) be the probability of \(w_i\) for \(i=1,\ldots, l\).
Let \(A,B,F_A,\) and \(F_B\) be the functions defined with the same manner in (\ref{eq-A}).
Then
\[
F_A(z_1,z_2,\ldots,z_l)=F_B(z_1+1, z_1+z_2+1,\ldots, z_1+\cdots+z_l+1).
\]
\end{theorem}
With slight modification of Theorem~\ref{th-main}, we can compute the number of the occurrence of the overlapping increasing words.
For example, let us consider increasing overlapping words 11, 111, 1111 and the number of their occurrences.
Let 011, 0111, 01111 then these words are increasing nonoverlapping words.
The number of occurrences 11, 111, 1111 in sample of length \(n\) is equivalent to the number of occurrences 011, 0111, 01111 in sample of length \(n+1\) that starts with 0.
In \cite{RegnierSpankowski98}, expectation, variance, and CLTs (central limit theorems) for the occurrences of words are shown. We show the higher moments for nonoverlapping words.
\begin{theorem}\label{th-moments}
Let \(w\) be a nonoverlapping word.
\[\forall t\ E(N^t_w)=\sum_{s=1}^{\min \{T,t\}} A_{t,s}\dbinom{n-s|w|+s}{s}P^s(w).\]
\[A_{t,s}=\sum_r \dbinom{s}{r}r^t(-1)^{s-r},\ T=\max \{t\in{\mathbb{N}}\ \vert\ n-t|w|\geq 0\}.\]
\end{theorem}
In the above theorem, \(A_{t,s}\) is the number of surjective functions from \(\{1,2,\ldots,t\}\to\{1,2,\ldots,s\}\) for \(t,s\in{\mathbb{N}}\), see \cite{Riordan}.
\section{nonoverlapping partial words}\label{secpartial}
We introduce the symbol \(?\) to represent arbitrary symbols.
Let \({\cal A}\) be a finite alphabet.
A word consists of extended alphabet \({\cal A}\cup\{?\}\) is called {\it partial word} \cite{berstelBoasson}.
The word \(w'\) consists of \({\cal A}\) is
called a realization of the partial word \(w\) if \(w'\) consists of \({\cal A}\) and coincides with \(w\) except for the symbol \(?\).
A partial word is called nonoverlapping if the set of the realization is nonoverlapping.
For example \(001? 1\) is a nonoverlapping partial word with alphabet \({\cal A}=\{0,1\}\) and its realizations are 00101 and 00111.
We write \(w_1\sqsubset w_2\) for two partial words if \(w_1\) is a prefix of \(w_2\) with the alphabet \({\cal A}\cup\{?\}\).
For example \(01?1 \sqsubset 01?11\) but \(01?1 \not\sqsubset 01111\).
The probability of partial word \(w\) is defined by
\[P(w):=\sum_{w'\text{: realization of }w}P(w').\]
Let \(\tilde{w}\) be the word obtained by removing the letter ? from the partial word \(w\).
For example \(\tilde{w}=011\) if \(w=01?1\).
\begin{prop}
Let \(P\) be i.i.d. and \(w\) be a partial word. Then
\[P(w)=P(\tilde{w}).\]
In particular if \(P\) is the fair-coin flipping,
\[P(w)=2^{-|\tilde{w}|}.\]
\end{prop}
The theorems above still hold for nonoverlapping partial words.
\begin{cor}
Theorem~\ref{th-main2} and Theorem~\ref{th-moments} holds for nonoverlapping partial words.
Theorem~\ref{th-main} holds for nonoverlapping increasing partial words.
\end{cor}
We can find many nonoverlapping partial words.
For example, \(0^{m+1}(1?^m)^n 1\) are nonoverlapping for all \(n,m\).
Here \(w^m\) is the \(m\) times concatenation of the word \(w\).
For example, \(0^3(1?^2 )^21=0001?? 1?? 1\).
We can construct large size partial words that have large probabilities
\begin{prop}
Let \(P\) be the fair-coin flipping and
\begin{equation}
w(m):= 0^{m}(1?^{m-1})^{m-1} 1
\end{equation}
then \(w(m)\) is nonoverlapping for any \(m\),
\(|w(m)|=m^2+1\), and \(P(w(m))=2^{-2m}\).
\end{prop}
For example let \(m=10\) then we have 101 length partial word and its probability is \(2^{-20}\) for fair-coin flipping, which is significantly large compared to its length.
The author expect that \(w(m)\) achieves almost maximum probability among the nonoverlapping partial words with length \(|w(m)|\).
\begin{cor}
Let \(P\) be the fair-coin flipping.
There is a sequence of nonoverlapping partial words \(w_1, w_2,\ldots\) with increasing length \(|w_n|<|w_{n+1}|\) and \(lim_n |w_n|=\infty\) such that
\[\lim_{n\to \infty} \frac{-\log P(w_n)}{|w_n|}=0.\]
\end{cor}
\section{Power function of nonoverlapping words tests}\label{secPow}
In \cite{{Billingsley},{RegnierSpankowski98}}, CLT for the occurrences of words is shown.
\[ P(\frac{N_w-E(N_w)}{\sqrt{V_w}}<x)\to \frac{1}{\sqrt{2\pi}}\int^x_{-\infty} e^{-\frac{1}{2}x^2} dx,\]
where \(w\) is a nonoverlapping word,
\begin{align}\label{exp-variance}
&E(N_w)=(n-|w|+1)P(w) \text{ and} \nonumber\\
&V(N_w)=E(N_w)+(n-2|w|+2)(n-2|w|+1)P^2(w)\nonumber\\
& \hspace{1.4cm}-E^2(N_w).
\end{align}
Let
\[N'_w:= \sum_{i=1}^{\lfloor n/|w|\rfloor} I_{X_{i*|w|}^{(i+1)*|w|-1}=w}.\]
\(N'_w\) obeys binomial law if the process is i.i.d.
We call \(N_w\) {\it sliding block sampling} and \(N'_w\) {\it block-wise sampling}.
As an application of CLT approximation, we compare power functions of sliding block sampling \(N_w\) and block-wise sampling \(N'_w\).
We consider the following test for sliding block sampling:
Let \(E_\theta=E(N_w)\) and \(V_\theta=V(N_w)\) if \(P(w)=\theta\).
Null hypothesis: \(P(w)=\theta^*\) vs alternative hypothesis \(P(w)<\theta^*\).
Reject null hypothesis if and only if \(N_w<E_{\theta^*} -5\sqrt{V_{\theta^*}}\).
The likelihood of the critical region is called power function, i.e.,
\begin{equation}\label{eqPowfunc}
Pow(\theta):=P_\theta(N_w<E_{\theta^*} -5\sqrt{V_{\theta^*}})\text{ for }\theta\leq \theta^*.
\end{equation}
From (\ref{eqPowfunc}) power function will be large if the variance is small.
We construct a test for block-wise sampling:
Null hypothesis: \(P(w)=\theta^*\) vs alternative hypothesis \(P(w)<\theta^*\).
Reject null hypothesis if and only if \(N'_w<E'_{\theta^*} -5\sqrt{V'_{\theta^*}}\), where \(E'_\theta = \lfloor n/|w|\rfloor \theta\) and \(V'_\theta = \lfloor n/|w|\rfloor \theta (1-\theta)\).
The following table shows powers of sliding block sampling and block wise sampling at \(\theta=0.2, 0.18, 0.16\) under the condition that
alphabet size is 2 (binary data), \(\theta^*=0.25, |w|=2,\text{ and } n=500\).
\begin{center}
\begin{tabular}{@{} cccc @{}}
\hline
\(\theta\) &0.2& 0.18& 0.16 \\
\hline
Power of Sliding block &0.316007 &0.860057& 0.995681 \\
Power of Block wise & 0.000295 & 0.002939 & 0.021481\\
\hline
\end{tabular}
\end{center}
Figure~\ref{fig-A} shows the graph of power functions for sliding block words test and block-wise sampling.
\vspace{1cm}
\begin{figure}[h]
\scalebox{0.5}
{\includegraphics[0in,0in][4in,4in]{PowerPlot.pdf}}
\caption{Comparison of power functions: sliding block test (black line) vs block-wise test (red line).
\(P(w)< 0.25\) vs \(P(w)=0.25\) (null hypothesis). \(|w|=2, n=500\).
}\label{fig-A}
\end{figure}
\section{Experiments on statistical tests for pseudo random numbers}\label{sec-pseudo}
In \cite{NIST2010}, a battery of statistical tests for pseudo random number generators are proposed, and
chi-square test is recommended to test the pseudo random numbers with
\(N_w\) for nonoverlapping word \(w\).
In this section, we apply Kolmogorov Smirnov test to the empirical distribution of pseudo random numbers with
the true distribution of nonoverlapping words (\ref{eqmain}).
Let sample size \(n= 1600\) and \(l=1\) in (\ref{eqmain}) and null hypothesis \(P\) be fair-coin flipping.
For each nonoverlapping words \(w=10\) and \(w=11110\),
we consider the three distributions, 1) true distributions of \(N_w\) (eq.~(\ref{eqmain}) in Theorem~\ref{th-main2})
2) binomial distributions
\( \binom{n}{k}p^k(1-p)^{n-k},\ p=2^{-|w|},\ k=1,\ldots,n\), and
3) empirical distributions of \(\sum_{i=1}^n I_{X_i^{i+|w|-1}=w}\) generated by Monte Carlo method with BSD RNG {\bf random},
200000 times iteration of random sampling.
Figure~\ref{fig-dist} shows the graph of the three distributions for \(w=11110\).
The expectation of binomial distribution is \(pn\), which is almost same to the expectation of \(N_w\).
However the variance of binomial distribution and \(N_w\) are different.
Sliding block sampling \(N_w\) have strong correlations even if the process is i.i.d.
For example, if a nonoverlapping word has occurred at some position, then the same nonoverlapping word do not occur in the next position.
From Figure~\ref{fig-dist}, we see that binomial distributions has large variance compared to the true distributions.
This is because, in the binomial model, the correlations of words are not considered.
We see that the binomial model approximations of the distributions of the words are not appropriate.
Figure~\ref{fig-dist} shows that the empirical distributions (Monte Carlo method) is different from the true distribution.
We see that BSD RNG {\bf random} do not simulate the sliding block sampling correctly.
\vspace*{3cm}
\begin{figure}[ht]
\scalebox{0.44}{
\includegraphics[0in,0in][4in,4in]{Exw5.pdf}}
\caption{Comparison of distributions:
the distributions for \(w=11110\) with sample size \(n= 1600\).
Black, red, and blue lines show true distribution, binomial, and Monte Carlo simulation, respectively.}\label{fig-dist}
\end{figure}
From the experiment for \(w=11110\) we have obtained
\[\sup_{0\leq x<\infty}| F_t(x)-F(x)|=0.302073,\]
where \(F_t(x)\) is the empirical cumulative distribution generated by Monte Carlo method with BSD RNG {\bf random}
with \(t=200000\) times random sampling.
\(F(x)\) is the cumulative distributions of (\ref{eqmain}).
The p-value of the Kolmogorov-Smirnov test is almost zero,
\[P(\sup_{0\leq x<\infty}| F_t(x)-F(x)|\geq 0.302513 )\approx 0,\]
where \(P\) is the null hypothesis.
The sliding block sampling tests reject BSD RNG {\bf random}.
The p-values of the Kolmogorov-Smirnov test for BSD RNG with \(w=10,\ w=11110,\ t=200000\) and \(n=1600\) are summarized in the following table.
\begin{center}
\begin{tabular}{@{} cccc @{}}
\hline
BSD RNG &w=10& w=11110 \\
\hline
\(\sup_{0\leq x<\infty}| F_t(x)-F(x)|\) &0.012216 & 0.302073\\
p-value & 0 & 0\\
\hline
\end{tabular}
\end{center}
The sliding block words tests do not reject MT RNG \cite{MT98} under the same condition above.
The p-values of the Kolmogorov-Smirnov test for MT RNG with \(w=10,\ w=11110,\ t=200000\) and \(n=1600\) are summarized in the following table.
\begin{center}
\begin{tabular}{@{} cccc @{}}
\hline
MT RNG &w=10& w=11110 \\
\hline
\(\sup_{0\leq x<\infty}| F_t(x)-F(x)|\) & 0.001376& 0.001409\\
p-value &0.843306 & 0.822066\\
\hline
\end{tabular}
\end{center}
\section{Computation of the distribution of words in human DNA}\label{dna}
Let \(A=\{a,b,c,d\}, |X|=32\cdot 10^8\), and the probability of each letter is 1/4.
\(X\) is the human DNA size and \(P(w)=4^{-|w|}\).
If \(|w| \geq 14\), our algorithm can compute the exact distribution of \(w\)
within few seconds with desktop computer and we identified that the exact distribution (\ref{eqmain}) is numerically almost same to Poisson distribution,
\(Po(N_w=k)=e^{-\lambda}\frac{\lambda^k}{k!}\) and \(\lambda =E(N_w)\).
In case that \(|w|\) is small, the exact distributions will be well approximated with CLT.
\section{Proofs}
Proof of Theorem~\ref{th-main2})
For simplicity we prove the theorem for \(l=1\).
The proof of the general case is similar.
Let \(m=|w|\).
Since \(w\) is nonoverlapping,
the number of possible allocations such that \(k\) times appearance of \(w\) in the string of length \(n\) is
\[\dbinom{n-mk+k}{k}.\]
This is because if we replace each \(w\) with additional extra symbol \(\alpha\) in the string of length \(n\) then the problem reduces to
choosing \(k\) \(\alpha\)'s among the string of length \(n-mk+k\).
Let
\begin{equation}\label{basicFuncA}
A(k)=\dbinom{n-mk+k}{k}P^{k}(w).
\end{equation}
\(A\) is not the probability of \(k\) \(w\)'s occurrence in the string, since
we allow any letters in the remaining place except for the appearance of \(w\).
For example \(A\) may count the event that
\(w\) appear more than \(k\) times.
Let \(B(t)\) be the probability that nonoverlapping words \(w\) appear \(k\) times.
We have the following identity,
\[
A(k) =\sum_{k\leq t} B(t) \dbinom{t}{k}.
\]
Let \(F_A(z):=\sum_{k}A(k) z^{k}\) and
\(F_B(z):=\sum_{k}B(k) z^{k}\) be generating functions for \(A\) and \(B\) respectively.
Then
\begin{align*}
F_A(z) & =\sum_k z^k \sum_{k\leq t} B(t) \dbinom{t}{k}\\
& = \sum_{t} B(t) \sum_{k\leq t} \dbinom{t}{k} z^k\\
&= \sum_{t} B(t) (z+1)^{t} \\
& = F_B(z+1).
\end{align*}
In the above second equality, we changed the order of the sum.
We have
\begin{align*}
&F_B(z) =F_A(z-1)\\
&=\sum_{k~:~n-mk\geq 0} \dbinom{n-mk+k}{k}(z-1)^kP^k(w)\\
&=\sum_{k,t~:~n-mk\geq 0, t\leq k} \dbinom{n-mk+k}{k}\dbinom{k}{t}z^t(-1)^{k-t}P^k(w)\\
&=\sum_t z^t\sum_{k~:~t\leq k, n-mk\geq 0} (-1)^{k-t} \dbinom{n-mk+k}{t,k-t}P^k(w),
\end{align*}
and (\ref{eqmain}) .
\hfill\hbox{\rule{6pt}{6pt}}
\bigskip
Proof of Theorem~\ref{th-moments})
For simplicity let
\(Y_i=I_{X_i^{i+|w|-1}}=w\).
Since \(w\) is nonoverlapping,
\(Y_i Y_j =Y_i\) if \(i=j\). \(Y_i Y_j =0\) if \(\{i,i+1,\ldots,i+|w|-1\}\cap \{j,j+1,\ldots,j+|w|-1\}\ne\emptyset\).
We say that \(\{i,i+1,\ldots,i+|w|-1\}\) is the coordinate of \(Y_i\).
We say that a subset of \(\{Y_i\}_{1\leq i\leq n-|w|+1}\) is disjoint if their coordinate are disjoint.
Let \(Y_{i,j}=Y_i\) for all \(1\leq j\leq t\).
Then \((\sum_i Y_i)^t=\prod_{j=1}^t\sum_i Y_{i,j}=\sum_i \prod_{j=1}^t Y_{i,j}\).
Note that \(E( \prod_{j=1}^t Y_{i,j})=P^s(w)\) if and only if there is a disjoint set \(Y_{n(j)}, 1\leq j\leq s\) such that \(\prod_{j=1}^t Y_{i,j}=\prod_{j=1}^s Y_{n(j)}\).
Observe that the number of possible combination of disjoint sets of \(Y_{n(j)}, 1\leq j\leq s\) such that \(\prod_{j=1}^t Y_{i,j}=\prod_{j=1}^s Y_{n(j)}\) is
\(A_{t,s}\dbinom{n-s|w|+s}{s}\). Note that there is no disjoint sets of \(Y_{n(j)}, 1\leq j\leq s\) if \(s>\max \{t\in{\mathbb{N}}\ \vert\ n-t|w|\geq 0\}\).
From the linearity of the expectation, we have the theorem.
\hfill\hbox{\rule{6pt}{6pt}}
\section*{Acknowledgment}
The author thanks for a helpful discussion with Prof. S.~Akiyama and Prof.~M.~Hirayama at Tsukuba University. This work was supported by the Research Institute for Mathematical Sciences, an International Joint Usage/Research Center located in Kyoto University, and Ergodic theory and related fields in Osaka University.
{\small
|
\section{The OEIS}\label{SecI}
The {\em On-Line Encyclopedia of Integer Sequences} or {\em OEIS} \cite{OEIS} is a free website (\url{https://oeis.org}) containing information about 350,000 number sequences. You will probably first encounter it when trying to identify a sequence that has come up in your work. If your sequence is recognized, the response will tell you the first 100 or sometimes 10,000 terms, give a definition, properties, formulas, references, links, computer programs, etc., as appropriate. If it is not recognized, the system will invite you to submit it if you think it is of general interest, so that the next person who looks it up will find it - and be grateful to you.
Ron Graham called the OEIS a ``fingerprint file for mathematics''. It has also been described as, ``pound for pound, one of the most useful mathematics sites on the web''. If you have been struggling with a sequence, and the OEIS tells you what it is, you will understand why it is so popular. Of course, if it tells you that your problem was solved forty years ago, you may be unhappy, but it is better to find out now rather than later. The database has been around, in one form or another, for 57 years, so if your sequence is not yet included, there is a moderate chance it is new (of course this is not a proof).
\section{The OEIS as a source of problems}\label{Sec2Prob}
The entries are constantly being updated.
Every day we get a hundred or so submissions of new sequences, and another hundred
comments on existing entries (new formulas, references, additional terms, etc.).
The new sequences are often sent in by non-mathematicians,
and are a great source of problems to work on. You can see the current submissions at \url{https://oeis.org/draft}.
Often enough you will see a sequence that is so interesting you want to drop everything else and work on it.
Well, go ahead! Many research papers have been born in this way. The ``Yellowstone Permutation''
\seqnum{A098550}
is one of my favorite examples: it was quite a challenge to prove that it {\em is}
a permutation of the positive integers, and the entry has a link to a paper \cite{Yellow} that a group of us wrote
analysing it.
Some of the things you might work on after seeing an interesting sequence on the drafts stack are:
\begin{itemize}
\item Is the sequence well-defined?
\item Does it contain infinitely many terms?
\item How fast does it grow? The ``graph'' button in every entry is helpful here.
\item Is there a formula or generating function?
\item Or you may see a conjecture or question that you think you can answer.
\item How would you program it to generate more terms?
\end{itemize}
Studying the ``drafts'' stack is an endless source of fun (and sleepless nights).
\section{The OEIS as a reference work}\label{SecRef}
Even when you know what the sequence is, you may still look it up, to find out what is presently known about it.
We try to make sure the sequences are well-supplied with references, especially any recent articles.
The coverage is broad: besides the obvious fields like combinatorics, graph theory,
group theory, number theory, computer science, recreational mathematics, there are
large numbers of sequences from physics and chemistry.
This makes the OEIS extremely useful as a reference work.
Earlier this year a senior group theorist told us he had been working on a problem for 20 years, but when he looked it up in the OEIS he found a reference that he was not aware of.
Another great resource are the computer programs. An important sequence will have programs to generate it in C, Java, MAGMA, Maple, Mathematica, PARI, Python, Sage, etc. This is very helpful for numerical investigations.
\section{Submitting your sequence or comment}\label{Sec2Sub}
If you have an interesting number sequence that is not in the OEIS, you should definitely submit it.
Having a sequence in the OEIS is something you can be proud of.
You will be joining an enterprise that has been running
for almost 60 years, and to which over 12,000 people have already contributed.
Contributions come from almost every country, and we have been called one of the most successful
international collaborations.
You must register before you can contribute, using your real name,
the name you would use in a scientific publication
(see \href{https://oeis.org/wiki/Special:RequestAccount}{Request Account}).
The OEIS is not a ``social media''. One of
the reasons for the success of the database is that we have high standards, all contributions are
refereed by the editors, and accuracy is of primary importance.
See \href{https://oeis.org/wiki/Overview_of_the_contribution_process}{here} for more about the submission process.
The OEIS Wiki also has a page showing examples of
\href{https://oeis.org/wiki/Examples_of_what_not_to_submit}{What not to submit}!
\section{The OEIS Wiki}\label{SecW}
The OEIS Wiki (\url{https://oeis.org/wiki}) has a great deal of useful information for users, especially in the ``Information'' section. There is a general index to the sequences, a style sheet, a Q\&A page, an FAQ page, pages called
\href{https://oeis.org/wiki/How_to_add_a_comment,_more_terms,_or_a_b-file_(short_version)}{How to add a comment, more terms, or a b-file, ...},
\href{https://oeis.org/wiki/Instructions_For_General_Users}{Instructions for general users},
\href{https://oeis.org/wiki/The_multi-faceted_reach_of_the_OEIS}{The multi-faceted reach of the OEIS}, and so on.
\section{Citations of the OEIS}\label{SecCite}
An especially important part of the OEIS Wiki is the section
(see \url{https://oeis.org/wiki/Works_Citing_OEIS}) that lists citations of the OEIS in
the literature or on the web. There are now about 10,000 citations, which often say things like ``This theorem would not have been discovered
without the help of the OEIS''.
You can help keep these pages up-to-date. If you come across a paper that mentions a number sequence,
check if the sequence is in the OEIS, add it if it isn't, and make sure the entry has a reference to the paper.
If the context seems new, consider adding a comment saying ``Arises in the spectral analysis of cobweb singularities,
see A. Spider et al. (2021)." If you come across an article that references the OEIS, make sure the paper
is listed in the ``Works Citing the OEIS'' pages on the Wiki. And if you happen to see a comment that ``This sequence is not in the OEIS", then add it at once.
When you write a paper yourself, using information from the OEIS, don't forget to mention us in your references list,
typically by saying
The OEIS Foundation Inc., Entry A123456. Published electronically at https://oeis.org, 2021,
\noindent
and also mention it in any relevant OEIS entries.
Many authors ``forget'' to do this, but it will help your career by drawing attention to your paper.
\section{``Link Rot''}\label{SecL}
Don't get me started! The OEIS serves as a guide to the literature on a huge number of subjects,
and we have hundreds of thousands of links. And every day many of these links break.
System administrators feel they are not doing their job if they don't change all their URLs every
couple of years. Or, as happened last year, a major university will decide to delete all the faculty home pages, with no warning. That broke several hundred of our links.
Pages on individual's web sites are the most fragile of all.
If you are a frequent user of the OEIS you will often run into this problem. You can help by locating a replacement
URL if the site has simply moved, or by adding a new link to a copy of the missing page on the wonderful Wayback Machine, run by the Internet Archive \cite{WBM}. A better solution is to ask the author of the page
for permission to put a local copy of the page on the OEIS server. We have a strong reputation, and hope to be around for a long time. Almost everyone we have asked has agreed.
But it is a time-consuming business, and you could help.
\section{Other topics}\label{SecOT}
\vspace*{+.2in}
\noindent{\bf ``Superseeker''.} This is a program that runs on our server, and tries very hard to identify a sequence.
It runs several programs that try to guess a formula or recurrence (including the powerful
Salvy-Zimmermann program {\em gfun} \cite{gfun}). It also transforms
the sequence in a hundred ways and looks up the results in the OEIS, hoping for a match.
To use it, send an email
to \href{mailto:<EMAIL>}{\tt <EMAIL>}.
with a blank subject, containing a single line of the form
lookup 0 1 3 7 11 15 23 35 43 47
\noindent
(with no commas).
Since this uses many resources on our server, please use it sparingly.
One of my current goals is to strengthen Superseeker. If you are interested in helping, please contact me.
\vspace*{+.2in}
\noindent{\bf The Sequence Fans Mailing List.} Any registered user of the OEIS can join, and
messages go out to a few hundred fellow sequence-lovers. This can be even more powerful than Superseeker if you are really desperate to identify a sequence. There is a link on the OEIS Wiki.
\vspace*{+.4in}
\noindent{\bf Triangles and arrays of numbers.}
The OEIS also includes triangles and arrays of numbers. Triangles are read by rows. Pascal's triangle
becomes \seqnum{A007318}: $1, ~1, 1, ~1, 2, 1,$ $1, 3, 3, 1, ~1, 4, 6, 4, 1,$ $1, 5, 10, 10, 5, 1, \ldots$.
Arrays are read by antidiagonals. The table of Nim-sums $m \oplus n$ \cite{BON},
$$
\begin{array}{ccccccc}
0 & 1 & 2 & 3 & 4 & 5 & \ldots \\
1 & 0 & 3 & 2 & 5 & 4 & \ldots \\
2 & 3 & 0 & 1 & 6 & 7 & \ldots \\
3 & 2 & 1 & 0 & 7 & 6 & \ldots \\
4 & 5 & 6 & 7 & 0 & 1 & \ldots \\
. & . & . & . & . & . & \ldots \\
\end{array}
$$
becomes \seqnum{A003987}, $0, ~1, 1,~ 2, 0, 2,$ $ 3, 3, 3, 3, $ $ 4, 2, 0, 2, 4, \ldots$.
\vspace{0.2in}
The OEIS aims for a broad coverage of integer sequences arising in science, especially in mathematics (of course including several versions of the famous subway stops sequence, since they have been published on tests; but not the number of pages in the $n$th issue of these {\em Notices} since 1954, because no one has mentioned it before). Our motto is accuracy and completeness, combined with good judgment.
We have a great many contributors, a great many entries, and we hope you will join us at \url{https:oeis.org}.
|
\section{Introduction} \label{sec:introduction}
\IEEEPARstart{C}{oronary} artery disease accounts for three hundred thousand deaths per year. More than one million individuals suffered from acute coronary events last year in the United States~\cite{virani2020heart}.
Although effective therapies are available and several novel treatments are in development, predicting and preventing acute coronary events still presents a major challenge~\cite{franco2011challenges}.
In addition to physical examination and blood biomarkers, physicians diagnose coronary artery disease in symptomatic patients using various non-invasive imaging tests such as nuclear stress tests, CT coronary angiography, and PET-CT.
Similarly, the guidance of percutaneous coronary interventions relies increasingly on invasive imaging modalities, which may offer a window of opportunity to prospectively identify lesions with a propensity for subsequent acute events.
Since coronary atherosclerosis is the primary underlying cause of acute myocardial infarction, the adoption of sensitive tools for vessel and plaque characterization is essential~\cite{arbab2012acute}.
Intravascular (IV) optical coherence tomography (OCT) acquires high resolution images of the subsurface microstructure of coronary atherosclerotic lesions available in catheterization laboratories~\cite{jang2001visualization,okamura2011first}, enabling the development and application of new therapeutic procedures to improve clinical outcomes~\cite{meneveau2016optical,pazdernik2020donor}. Nonetheless, the interpretation and characterization of OCT images require extensive training and experience due to the image contrast, signal statistics, and speckle characteristics.
These difficulties result in modest intra- and inter-reader agreement and limit adoption of OCT~\cite{manfrini2006sources}.
Furthermore, image processing algorithms used in the clinical setting are currently limited to lumen segmentation without extracting the remaining embedded information in the images~\cite{dallan2021comparison}.
Notably, the automatic detection of anatomical layers within the vessel wall and other features beyond the lumen promises to accelerate OCT utilization in clinical settings and translational research by eliminating the need for extensive training and time-consuming manual segmentation.
In parallel to conventional OCT (Fig.~\ref{fig:annotations}.A), polarization sensitive (PS) OCT performs polarimetry and produces images of tissue birefringence and depolarization (Fig.~\ref{fig:annotations}.B-C).
Nadkarni et al.~\cite{nadkarni2007measurement} demonstrated that fibrillar tissues such as interstitial collagen and arterial smooth muscle cell arrays are birefringent. Villiger et al.~\cite{villiger2018coronary} showed that tissues containing lipid particles, macrophage accumulations, or cholesterol crystals exhibit depolarization.
Intimal thickness is an important early clinical sign of atherosclerosis~\cite{tuzcu2001high} and can be measured by detecting the outer boundary of the intima (i.e., the internal elastic lamina (IEL)) based on the birefringence difference between intima and media layers.
Moreover, Villiger et al.~\cite{villiger2018coronary} characterized human coronary atherosclerosis by utilizing PS-OCT signatures and showed their correlation with histological features. The foregoing studies suggest that PS-OCT may facilitate automated image analyses and objective tissue characterization.
Our present work builds on the success of deep learning methods for segmentation to automatically characterize anatomical layers.
This paper proposes a convolutional neural network and optimizes its performance using a new multi-term loss function to utilize the advantage of PS-OCT image contrast and to analyze coronary artery images automatically.
The multi-term loss function includes two common segmentation loss terms, i.e., weighted cross-entropy loss and generalized multi-class Dice loss.
In addition to the common segmentation loss terms, a boundary loss term focuses on the accuracy of the model only within the pixels close to the boundaries between the anatomical layers.
Similarly, ``Boundary Cardinality loss'' penalizes the model from a topological point of view when the number of anatomical layers is different between the model's prediction and ground-truth by counting the number of boundary pixels along the radial axis.
Additionally, ``Attending Physician loss'' uses an independently-trained critique model, which distinguishes the low- and high-quality labels, as a loss term.
\begin{table}
\centering
\caption{The definition of the six exclusive labels that are based on the manual expert annotations and shown in Fig.~\ref{fig:annotations}.D.}
\input{Table/tab_class_label_def}
\label{tab:class_label_def}
\end{table}
\begin{figure*}
\centerline{\includegraphics[width=\textwidth]{Figure/fig_architecture_small.pdf}}
\caption{The proposed model architecture.
The auxiliary critique model (a.k.a. Attending Physician) concatenates the image ($\boldsymbol{\mathcal{I}}$) and ground-truth label ($\boldsymbol{\mathcal{Y}}$) as its input to predict the quality level of manual labeling.
After training, this model was used as one of the loss terms.
The main model has a multi-channel image as the input and a multi-class probability prediction as the output ($\hat{\boldsymbol{\mathcal{Y}}}$).}
\label{fig:architecture_small}
\end{figure*}
We evaluated our method on a set of 984 images in 57 patients and compared its performance to state-of-the-art algorithms.
Our work improves the boundary detection of coronary lumen and two anatomical layers while also detecting the guidewire and plaque shadow in addition to handling the thickened vessel wall cases excluded by other studies.
\section{Related Work} \label{sec:related_works}
The clear need for automated segmentation and interpretation of intravascular OCT images has motivated the development of methods that tackle the challenges and artifacts presented by typical intravascular OCT data.
OCT axial A-lines are independently acquired along the radial direction of the vessel in cylindrical coordinates as the probe rotates. Simultaneously, the core of the catheter is pulled back through the vessel, resulting in a helical scan pattern.
Contemporary IV-OCT reconstructs individual A-lines from measurements in the frequency domain and visualizes the logarithm of the power of the reconstructed signal on the dB scale.
Prior signal processing techniques aimed to analyze A-lines independently in order to exploit the rich characteristics of the OCT signal~\cite{moraes2013automatic,bologna2019automatic,macedo2015bifurcation,macedo2016robust,akbar2019automated}.
Researchers proposed various classical segmentation methods to detect the vessel lumen in IV-OCT images.
The region-based active contour segmentation methods with level-set energy functions utilize the prior cross-sectional information~\cite{ughi2012automatic, joseph2016automatic, cao2017automatic}.
While methods based on dynamic programming~\cite{zhu2021automatic, cao2017automatic} or small artificial neural network~\cite{nam2016automated} are able to correct the remaining artifacts after the application of the primary methods, level-set methods perform poorly in low signal-to-noise (SNR) regions and result in non-smooth and imprecise lumen boundaries.
On the other hand, graphical models such as graph-cuts are suitable solutions for this problem because the lumen boundary and the interfaces of anatomical layers do not intersect~\cite{wang2014fully,zahnd2017contour,lee2015fast,chen2018quantitative}.
While other graphical models such as Markov random fields~\cite{tsantis2012automatic} tackle the vessel wall segmentation problem from a similar point of view, the counterpart physics-based methods simulate the segmentation as a physical problem~\cite{roy2015lumen,olender2018mechanical,yang2020novel}.
Despite the success of graphical methods, difficulties arise in low SNR regions with speckle, anatomical anomalies, and external objects, which limits the practical applications of these models and causes a cascade of increasing errors that require follow-up manual corrections done by expert annotators in the post-processing stage~\cite{pazdernik2018early}.
In recent years, new machine learning models have been emerging as a solution to many medical image analysis problems.
Yong et al.~\cite{yong2017linear} used a regression deep learning (DL) network to detect the vessel lumen along the radial direction in polar coordinates.
Gharaibeh et al.~\cite{gharaibeh2019coronary} segmented vessel lumen and coronary calcifications in IV-OCT images using a U-Net architecture and post-processed the output with a conditional random field model.
Abdolmanafi et al.~\cite{abdolmanafi2018characterization, abdolmanafi2019fully} used pre-trained convolutional neural networks (CNN) to segment the anatomical layers of the coronary arteries using transfer learning.
Our model builds on these previous approaches and extends the state-of-the-art DL methods to detect anatomical layers and image artifacts in both healthy and thickened vessel walls.
\section{Methods} \label{sec:methods}
As mentioned in the description of the manual annotation, one of the four dataset batches was revised extensively to provide a curated set of high-quality segmentations.
This practice resembles the scenario of an attending physician reviewing and scoring manual annotations done by resident physicians to provide constructive feedback for training.
By analogy, we trained a model to critique the multi-class labels conditioned on their input images by distinguishing between the initial and final revisions of the labels.
The trained model was used as the attending physician loss term ($\mathcal{L}_{AP}$) to critique the quality of predicted labels by the main model.
Therefore, we trained the main model from both types of manual segmentations by introducing the critique model (Fig.~\ref{fig:architecture_small}).
\subsection{Segmentation Loss}
We developed a multi-term multivariate loss function that includes novel loss terms.
The first loss term is the weighted cross-entropy function,
\begin{align}
\mathcal{L}_{WCE} (\mathbf{y}, \hat{\mathbf{y}}) = - \frac{1}{N} \sum_{c=1}^{N_c} \omega_c \sum_{i,j} \left[ \mathbf{y}_{ijc} \log \hat{\mathbf{y}}_{ijc} \right],
\end{align}
that measures the cross-entropy between the target label ($y$) and predicted label probabilities ($\hat{y}$) of all $N$ pixels where $i$ and $j$ are 2D matrix indices and $c$ is the class index .
Each pixel's cross-entropy is then weighted proportionately to the inverse of its class population ($\omega_c=\lVert \mathbf{Y} \rVert_1 / \lVert \mathbf{Y}_c \rVert_1$).
The second loss term is a multi-class version of the generalized Dice loss function~\cite{milletari2016v},
\begin{align}
\mathcal{L}_{Dice} (\mathbf{y}, \hat{\mathbf{y}}) = 1 - \frac{2}{N_c} \sum_{c=1}^{N_c} \frac{\sum_{i,j} \left[ \mathbf{y}_{ijc} \times \hat{\mathbf{y}}_{ijc} \right] + \epsilon}{\sum_{i,j} \left[\mathbf{y}_{ijc}^2 + \hat{\mathbf{y}}_{ijc}^2 \right] + \epsilon},
\label{eq_dice_loss}
\end{align}
which measures each label's segmentation accuracy similar to the Dice coefficient.
Dice loss uses the prediction probability (e.g. softmax of logits) instead of the classification result and ranges from zero to one, with zero corresponding to the most accurate result.
In comparison to the Dice coefficient, the argument of the maxima is omitted in order to make the function differentiable, and we added a small constant ($\epsilon$) for numerical stability.
\subsection{Boundary Loss}
The weighted cross-entropy and Dice loss terms are only marginally affected by the errors on the boundary because the boundary pixels are a small portion of the target objects.
The third loss term is based on the boundary segmentation accuracy and focuses the network’s attention on the close vicinity of label boundaries.
The boundary precision loss term ($\mathcal{L}_{BP}$) utilizes a boundary neighborhood mask,
\begin{align}
\begin{split}
\pmb{\beta} (\mathbf{y}) =&
\left( \bigvee_{c=1}^{N_c} \left[ \left( \mathbf{1}_{b \times 1} \ast \mathbf{y}_k \right) \oplus \left( \mathbf{y}_k \right) \right] \right) \bigvee \\
&\left( \bigvee_{c=1}^{N_c} \left[ \left( \mathbf{1}_{b \times 1} \ast (1 - \mathbf{y}_k) \right) \oplus \left( 1 - \mathbf{y}_k \right) \right] \right),
\end{split}
\end{align}
which masks the cross-entropy loss values of the pixels that are not in close vicinity of the label boundaries along the radial axis (e.g. $> b$, $b=10$ pixels).
The boundary precision loss term,
\begin{align}
\mathcal{L}_{BP} (\mathbf{y}, \hat{\mathbf{y}}, \pmb{\beta}) =&
\frac{-1}
{\sum_{i,j} \pmb{\beta}_{ij}} \sum_{c=1}^{N_c} \sum_{i,j} \pmb{\beta}_{ij} \mathbf{y}_{ijc} \log \hat{\mathbf{y}}_{ijc},
\end{align}
is differentiable with respect to the model parameters as long as $\pmb{\beta}$ is a function of the ground-truth target label.
\subsection{Attending Physician Loss}
The training of a critique model involves a loss function that measures the distance of model parameters from the optimal solution in the parameter space.
Arjovsky et al.~\cite{arjovsky2017wasserstein} proposed \textit{Wasserstein-1} (a.k.a. Earth-Mover) distance,
\begin{align}
W_1(P,P') = \inf_{\gamma \in \Pi(P,P')} \mathbb{E}_{(x,x') \sim \gamma} \lVert x - x' \rVert^1 ,
\label{eq.wasserstein1}
\end{align}
where $\Pi(P,P')$ is the set of all joint distributions $\gamma(x,x')$ that their marginal distributions are equal to $P$ and $P'$.
Wasserstein-1 is the optimal cost of transporting a mass with distribution $P$ to another mass with distribution $P'$ when the transport cost and transport distance are linearly related.
Stable learning with a meaningful learning curve that avoids common problems, including mode collapse, can be obtained when \textit{Wasserstein-1} distance is adopted~\cite{arjovsky2017wasserstein}.
Since the infimum in (\ref{eq.wasserstein1}) is intractable, Kantorovich and Rubinstein~\cite{villani2008optimal} proposed a tractable dual problem,
\begin{align}
W_1(P,P') = \sup_{\lVert f \rVert_L \leq 1} \{
\mathbb{E}_{x \sim P} [f(x)] - \mathbb{E}_{x' \sim P'} [f(x')] \},
\end{align}
where $f$ is a 1-Lipschitz function, mapping the support of $P$ and $P'$ to real numbers.
Similarly, we can select $f$ from a family of parameterized functions ($\{f_w\}_{w \in \mathcal{W}}$) that are at least K-Lipschitz for a constant K and optimize (\ref{eq.wasserstein1}) over the functional parameter space,
\begin{align}
W_1(P,P') &= \max_{w \in \mathcal{W}}
\{\mathbb{E}_{x \sim P} [f_w(x)] - \mathbb{E}_{x' \sim P'} [f_w(x')]\},
\end{align}
\begin{align}
\begin{split}
\nabla_w W_1(P,P') = \max_{w \in \mathcal{W}}
\{ &\mathbb{E}_{x \sim P} [\nabla_w f_w(x)] - \\
&\mathbb{E}_{x' \sim P'} [\nabla_w f_w(x')]\}.
\end{split}
\end{align}
The requirement of $f$ being K-Lipschitz for the function family of deep neural networks can be imposed by clipping the parameter values with an absolute value upper limit \cite{arjovsky2017wasserstein} or enforcing the gradient of parameters to be 1 almost everywhere through a gradient penalty loss term \cite{gulrajani2017improved}.
Gulrajani et al.~\cite{gulrajani2017improved} showed that the gradient constraining method improves the learning process compared to the weight clipping method.
We observed a similar effect while training our Attending Physician model, which is then used as the fourth loss term ($\mathcal{L}_{AP}$).
\subsection{Topological Loss}
The last loss term examines the labels from a topological point of view.
Ideally, the predicted labels along A-lines are composed of three or four connected components without any void, starting with the lumen label in the center and ending with the outside label.
The area between the lumen and outside labels should be occupied by two adjacent solid anatomical layers (i.e., intima and media) or by one of the artifact labels (i.e., guidewire or plaque shadows).
These configurations are distinguishable in terms of the number of label boundaries along the radial direction.
The soft boundary cardinality loss term ($\mathcal{L}_{BC}$) penalizes the discrepancy between the predicted and ground-truth labels based on the number of boundary pixels along the radial axis.
We propose ``soft argmax'',
\begin{align}
\mathbf{S} (\mathbf{y}) = 1 + \tanh{\left( M \left[ \mathbf{y} - \max_{c} \mathbf{y}_{ijc} \right] \right)},
\end{align}
which is a differentiable proxy for the arguments of the maxima and a saturated equivalent of the softmax function, i.e., $e^\mathbf{x}/ \lVert e^\mathbf{x_i} \rVert_1$.
The soft argmax admits the predicted class probabilities at each pixel and maps the probability of the most probable class and other classes to $\sim 1$ and $\sim 0$, respectively.
The level of saturation is controlled by the large number $M$ and the precision of the probability values.
Since the value of the soft argmax for a given class changes between two adjacent pixels at the label boundary, the soft boundary set cardinality along the radial axis,
\begin{align}
\mathbf{BC} (\mathbf{S}) = \frac{1}{2} \sum_i \sum_c \lvert \mathbf{S}(\mathbf{y})_{(i+1)jc} - \mathbf{S}(\mathbf{y})_{ijc} \rvert,
\end{align}
approximates the number of class boundaries in each A-line.
The boundary cardinality loss function ($\mathcal{L}_{BC}$),
\begin{align}
\mathcal{L}_{BC} (\mathbf{y}, \hat{\mathbf{y}}) = \sigma(\mathbf{BC} (\mathbf{S}(\mathbf{y})), \mathbf{BC} (\mathbf{S}(\hat{\mathbf{y}}))),
\end{align}
compares the prediction and ground-truth labels with respect to the number of boundaries, where $\sigma$ measures the difference between two $\mathbf{BC}$ vectors (e.g. norm 1).
We considered 1, 100, and $100/\epsilon$ for $M$, where $\epsilon$ is the small number used for mathematical stability in \ref{eq_dice_loss} and sofmax value clipping.
For $\sigma$, we considered $\lVert . \rVert_1$, $\lVert . \rVert_2$, and $\max(.)$.
Based on the validation dataset and the convexity of $\mathcal{L}_{BC}$, norm-1 ($\lVert . \rVert_1$) and $100/\epsilon$ are the optimal choices for $\sigma$ and $M$, respectively.
The final loss function combines all five loss terms,
\begin{align}
\begin{split}
\mathcal{L} =& \lambda_{WCE} \mathcal{L}_{WCE} + \lambda_{Dice} \mathcal{L}_{Dice} + \\
& \lambda_{BP} \mathcal{L}_{BP} + \lambda_{AP} \mathcal{L}_{AP} + \lambda_{BC} \mathcal{L}_{BC},
\end{split}
\end{align}
in which loss term weights ($\lambda_.$) are selected within the range $[10^{-3}, 10^{3}]$ and optimized over their logarithmically-spaced multidimensional grid using greedy algorithms.
\subsection{CNN Architecture}
The proposed network architecture scheme is based on the U-Net \cite{ronneberger2015u} and deep residual learning \cite{he2016deep} models (Fig.~\ref{fig:architecture_small}).
The auxiliary critique model was trained independently to distinguish low- and high-quality labels.
Subsequently, the main model combines the trained critique model and other loss terms to segment three anatomical layers and two shadows.
The optimized architecture contains multi-scale encoder and decoder sections with skip connections at each scale.
The convolutional complex contains three convolutional layers with a $3 \times 3$ pixel kernel size and a leaky version of the rectified linear unit (L-ReLU) activation function, which has a negative slope coefficient of 0.3.
These three convolutional layers compute the residual values by using an internal skip connection.
The max-pooling layers with a $2 \times 2$ pixel kernel size are applied after convolutional complexes in the encoding section for down-scaling while the counterpart deconvolutional layers are applied for bi-linear up-scaling within the decoding section.
The encoding output and decoding input are connected through two convolutional complexes that operate at the latent representation level.
The layers within each of the three scales and the latent representation layers have 8, 8, 16, and 16 features, respectively.
The critique model architecture accepts the concatenation of image channels and output label channels as the input and applies three convolutional complexes with 32, 64, and 128 features, respectively.
Each complex consists of two convolutional layers with a $3 \times 3$ pixel kernel size and the ReLU activation function followed by a max-pooling with a $2 \times 2$ pixel kernel size.
The last complex's output is flattened and fed into a three-layer dense neural network with 1024, 256, and 128 hidden nodes and ReLU activation function.
The final output has one feature and uses the hyperbolic tangent activation function (Fig.~\ref{fig:architecture}).
\subsection{Training and Implementation}
We randomly divided the final annotated dataset between training, validation, and hold-out testing dataset by selecting 45, 6, and 6 patients (80\%/10\%/10\%), respectively.
Augmentation included random mirroring, rotation, image intensity distribution manipulations, and spatial scaling .
An element from the power set of the image augmentation set was applied to each given PS-OCT cross-section with randomly selected transformation parameters based on independent and identically uniformly distributed augmentation probabilities.
The geometric transformations were defined in the Cartesian coordinate system, but they were implemented and applied in the polar coordinate system.
The data augmentation methods were implemented and executed on GPU to improve the model's run-time.
We implemented our model in Python using Keras\textsuperscript{\texttrademark} and Tensorflow\textsuperscript{\texttrademark}.
We commonly used RMSprop optimizer with $10^{-3}-10^{-4}$ learning rate and mini-batch size of 20 per GPU.
The GPU memory size was the limiting factor in the learning rate and mini-batch size selection.
We used two NVIDIA\textsuperscript{\textregistered} GeForce\textsuperscript{\textregistered} RTX 2080 Ti or four NVIDIA\textsuperscript{\textregistered} Tesla\textsuperscript{\textregistered} V100.
\subsection{Post-processing}
We applied a post-processing procedure to the model output to enforce known topology of the segmentations.
Initially, small objects and holes within each class were removed, and their interfaces were smoothed.
Then, a set of logical operations was applied to impose the topological relationships between the classes in the polar coordinate system.
The proposed set includes the following operations:
\begin{itemize}
\item Lumen is a single connected object without any 2D void. The same rule applies to both guidewire shadow and outside.
\item Guidewire and plaque shadows are confined between the lumen, the outside, and two A-lines.
\item The order of layers from proximal to distal ends is lumen, intima, media, and outside.
\end{itemize}
\subsection{Performance Metrics}
Based on the ground-truth labels, we evaluated the performance of multi-class prediction models using accuracy and Dice coefficient,
\begin{align}
Accuracy(\mathbf{Y}, \Hat{\mathbf{Y}}) &= \frac{1}{\lVert \mathcal{C} \rVert}\sum_{c\in \mathcal{C}} \frac{ \lVert \mathbf{Y} \rVert - \lVert \mathbf{Y}_c \bigtriangleup \Hat{\mathbf{Y}}_c \rVert}{\lVert \mathbf{Y} \rVert}, \\
Dice(Y, \Hat{Y}) &= \frac{2}{\lVert \mathcal{C} \rVert}\sum_{c\in \mathcal{C}} \frac{\lVert \mathbf{Y}_c \cap \Hat{\mathbf{Y}}_c\rVert}{\lVert \mathbf{Y}_c \rVert + \lVert {\mathbf{Y}}_c \rVert},
\end{align}
where $\lVert . \rVert$ is the set cardinality, $\Hat{\mathbf{Y}}_c$ is the set of predicted pixels as class $c$, $\mathbf{Y}_c$ is the set of pixels in ground-truth as class $c$, and $\mathcal{C}$ is the set of classes.
Furthermore, we evaluated the precision of inter-class boundaries using the average distance error (ADE) along the radial direction and modified Hausdorff distance (MHD) \cite{dubuisson1994modified} in 2D within the cross-section,
\begin{align}
d(a, \mathcal{B}) &= \inf_{b \in \mathcal{B}} \{ \lVert a - b \rVert_2 \}, \\
ADE(\Hat{\mathcal{B}}; \mathcal{B}) &= \frac{1}{\lVert \Hat{\mathcal{B}} \rVert} \sum_{a \in \Hat{\mathcal{B}}} d(a, \mathcal{B}), \\
MHD(\Hat{\mathcal{B}}, \mathcal{B}) &= \max \{ ADE(\Hat{\mathcal{B}}; \mathcal{B}) , ADE(\mathcal{B}; \Hat{\mathcal{B}}) \},
\end{align}
where $\mathcal{B}$ and $\Hat{\mathcal{B}}$ are the set of boundary pixels in the ground-truth and prediction, respectively, and $\lVert . \rVert_2$ is the Euclidean norm.
\section{Data}
We demonstrate the method on images from an intravascular polarimetry pilot study, which included two cohorts and enrolled a total of 57 (30+27) patients who underwent percutaneous coronary intervention and PS-OCT imaging at the Erasmus University Medical Center in Rotterdam.
Of the 57 pullbacks, only segments of native vessel wall or containing old stents from previous interventions were included in this study.
The imaging system consists of ``FastView'' intravascular catheters (Terumo Co., Tokyo, Japan) interfaced with our custom-built PS-OCT system, operating at 1300 nm central wavelength similar to commercially available clinical IV-OCT systems.
The wavelength scanning range was 110 nm, achieving a radial resolution below 10 $\mu$m, assuming a tissue refractive index of 1.34. The dimension of the pixels in the reconstructed tomograms in the radial direction were 4.2 $\mu$m and 4.43 $\mu$m, respectively, for the two cohorts.
The catheter's rotation speed was 100 RPS, with 1024 radial scans per rotation, and pullbacks were performed at 10 mm/s or 20 mm/s, at the operator’s discretion.
Non-ionic contrast solution was injected at a rate of 3-4 mL/s during the pullback to displace coronary blood and obtain an unperturbed view of the vessel wall.
Intravascular polarimetry was performed based on our earlier work~\cite{villiger2018coronary,otsuka2019intravascular, otsuka2020intravascular,otsuka2020polarimetric}.
Briefly, an electro-optic polarization modulator was used to alternate the polarization state of the light incident on the tissue between consecutive depth scans and a polarization-diverse receiver enabled determination of the detected light's polarization state and intensity.
Polarimetric analysis employed spectral binning ~\cite{villiger2013spectral} to reconstruct maps of tissue birefringence and depolarization.
Birefringence is the difference in the refractive index experienced by orthogonal polarization states aligned and orthogonal to the tissue optic axis, respectively.
Tissue depolarization measures the randomness of the detected light’s polarization state using the complement to one of the degree of polarization.
Initially, an expert interventional cardiologist excluded partial segments of 3D pullbacks that suffered from artifacts caused by insufficient blood clearing.
The qualified pullback segments added up to 3936 mm of pullbacks at a 100 or 200 $\mu$m pitch.
Subsequently, the expert annotated a total of 984 cross-sections spaced 4 mm apart using our in-house Matlab graphical user interface (Fig.~\ref{fig:annotations}.D).
The manual annotations included the outer boundaries of the lumen, tunica intima (i.e., internal elastic lamina (IEL)), and tunica media (i.e., external elastic lamina (EEL)).
The location of IEL and EEL within the plaque and guidewire shadows were extrapolated based on their visible segments (Fig.~\ref{fig:annotations}.E).
Additionally, A-lines containing plaque, guidewire, stent struts, side branches, or thrombus were identified and used for segmentation or selective analysis without influencing the main label categories.
Consequently, as summarized in Table~\ref{tab:class_label_def}, the manual annotations were converted into six exclusive labels: outside, lumen, visible intima, visible media, plaque shadow, and guidewire shadow (Fig.~\ref{fig:annotations}.F).
To manage the workload, we annotated the total dataset in four separate batches and through three phases: initial annotation, high-precision annotation, and annotation approval.
One of the batches was revised extensively at the pixel-level, requiring four times as long as other batches.
The high-accuracy batch, in combination with its initial annotation, was utilized to train the proposed critique model and its resulting loss term.
\begin{table}
\centering
\caption{Loss terms ablation study.}
\input{Table/tab_loss_term_results.tex}
\label{tab:loss_term_results}
\end{table}
\section{Results} \label{sec:results}
We evaluated the advantage of using PS-OCT to visualize and disseminate anatomical layers compared to the conventional IV-OCT in a controlled experiment in confirmation with the evidence provided by our previous studies~\cite{villiger2018coronary,villiger2018repeatability}.
The optimized architecture with the weighted cross-entropy loss function ($\mathcal{L}_{WCE}$) was trained using IV-OCT or PS-OCT, one at a time.
The Dice coefficient of the Media class for the $\mathcal{L}_{WCE}$ model trained on the conventional IV-OCT and PS-OCT were 62.7\% and 70.7\%, respectively, while the same metric for our final method was 79.5\%.
The performance improvement confirms that PS-OCT provides auxiliary information and facilitates the detection of anatomical layers.
\begin{figure*}
\centerline{\includegraphics[width=.95\textwidth]{Figure/fig_qualitative_cf_V5.pdf}}
\caption{Qualitative assessment of PS-OCT cross-sections.
The annotations of our model and the ground-truth are overlaid on the gray-scale intensity image in blue and red outlines, respectively.
See the text for detailed discussion.
Scale bar: 1 mm.}
\label{fig:qualitative_cf}
\end{figure*}
We compared the model's automated annotation results to the expert's ground-truth annotations in Fig.~\ref{fig:qualitative_cf} to qualitatively characterize our model, illustrate the model's strengths, and identify possible areas of improvement.
Our model's annotations and the ground-truth are overlaid on the gray-scale intensity image in blue and red outlines, respectively.
The most common complication for boundary annotation, particularly for the outer intima and outer media, is the presence of thick plaques or calcium (e.g. Fig.~\ref{fig:qualitative_cf}.A\, green arrow) and thickened vessel walls (e.g. Fig.~\ref{fig:qualitative_cf}.G, green arrow) that cause significant reduction in the detected signal.
The background signal and statistical noise characteristics within the plaque regions impede the model's objective to annotate the anatomical layers and result in higher annotation variability (e.g. Fig.~\ref{fig:qualitative_cf}.F, yellow arrow; Fig.~\ref{fig:qualitative_cf}.G, yellow arrow; Fig.~\ref{fig:qualitative_cf}.K, yellow arrow).
Nonetheless, whenever the image information supports the ground-truth boundaries, the model matches well with the expert annotations even in these challenging cases (e.g. Fig.~\ref{fig:qualitative_cf}.D, both arrows).
Correspondingly, the boundaries detected by the model may conform with the underlying multi-dimensional images more accurately than the ground-truth annotations (e.g. Fig.~\ref{fig:qualitative_cf}.F, green arrow), suggesting inconsistencies in the manual ground-truth segmentation.
The guidewire obstructs the probing light, causing a fuzzy signal at its boundaries, resulting in imprecise automatic and manual boundary detection (e.g. Fig.~\ref{fig:qualitative_cf}.B, green arrow).
Moreover, the physical proximity of the vessel lumen with the guidewire and catheter leads to perturbed pixel-level delineation of the lumen boundary (e.g. Fig.~\ref{fig:qualitative_cf}.B, green arrow; Fig.~\ref{fig:qualitative_cf}.I, green arrow).
Side branches can appear in various locations of the field of view and could be expected to exhibit confusing features, yet our model analyzes these cases in concordance with the ground truth annotation.
Such vessels might appear outside the vessel wall (e.g. Fig.~\ref{fig:qualitative_cf}.B, yellow arrow; Fig.~\ref{fig:qualitative_cf}.E, green arrow), directly adjacent to the vessel wall boundary (e.g. Fig.~\ref{fig:qualitative_cf}.I, yellow arrow), inside the intima (e.g. Fig.~\ref{fig:qualitative_cf}.C, yellow arrow), or in direct communication with the lumen (e.g. Fig.~\ref{fig:qualitative_cf}.L, yellow arrow).
Even though non-ionic contrast solution is injected during catheter pull-back to displace blood, residues of blood may persist in the vessel lumen vicinity (e.g. Fig.~\ref{fig:qualitative_cf}.A, yellow arrow; Fig.~\ref{fig:qualitative_cf}.E, yellow arrow).
Blood clearance can be incomplete, especially at the onset or the end of contrast injection Fig.~\ref{fig:qualitative_cf}.H, both arrows).
Still, in all these cases, our model successfully detects the lumen outer boundaries.
Equivalently, the dark and bright tissue patterns (e.g. Fig.~\ref{fig:qualitative_cf}.C, green arrow; and Fig.~\ref{fig:qualitative_cf}.K, green arrow) are observed beyond the media layer and mimic the multi-layer vessel wall structures but they do not distract the automatic boundary allocations.
Finally, while our study only included intravascular imaging prior to intervention, previously embedded stents are commonly encountered, owing to the high recurrence rate of acute coronary syndrome and myocardial infarction. Depending on the specific stent material and patient history, stents might appear embedded in the vessel wall (e.g. Fig.~\ref{fig:qualitative_cf}.J, green arrow) or protruding into the lumen (e.g. Fig.~\ref{fig:qualitative_cf}.J, yellow arrow; Fig.~\ref{fig:qualitative_cf}.L, green arrow).
Stents generate diverse and strong image artifacts that impede the model's ability to correctly detect the boundaries.
Exact layer segmentation behind stents presents challenges even for expert readers. With the exception of neointimal hyperplasia, previously stented segments are unlikely to reside in the culprit segment. Such segments were included in our data set merely to train the model to ignore the ensuing artifacts. Notably, there exists a distinct class of models designed to detect stent struts and verify correct stent deployment~\cite{tsantis2012automatic,ughi2012automatic,yang2020novel}).
\begin{table}
\centering
\caption{The performance of our multi-label classification model based on different performance metrics.}
\input{Table/tab_class_accuracy}
\label{tab:class_accuracy}
\end{table}
Furthermore, we conducted an ablation study to examine the individual effects of the proposed loss terms, i.e.,
\begin{enumerate}
\item The soft boundary cardinality loss term ($\mathcal{L}_{BC}$).
\item The Attending Physician (a.k.a. Wasserstein critique model) loss term ($\mathcal{L}_{AP}$),
\item The boundary precision loss term ($\mathcal{L}_{BP}$),
\item The generalized soft multi-class dice loss term ($\mathcal{L}_{Dice}$), and
\item The weighted cross-entropy loss term ($\mathcal{L}_{WCE}$),
\end{enumerate}
We measured the accuracy, Dice coefficient, and modified Hausdorff distance (MHD) on the hold-out test dataset.
The implementation and optimization of loss functions were devised and supervised by monitoring the Dice coefficient due to the unbalanced class populations, except for $\mathcal{L}_{BC}$ that we additionally monitored by the boundary error based on MHD.
The results of the ablation study on the loss terms are tabulated in Table~\ref{tab:loss_term_results} and confirm that each loss term contributes to the performance of the method.
While the aggregate performance metrics summarized in Tables~\ref{tab:loss_term_results} and Supplementary Table~\ref{tab:lumen_cf_result} allowed us to perform model comparisons and optimizations, the itemized performance measurements enable a more thorough inspection of the model.
We highlight possible improvement areas based on Table~\ref{tab:class_accuracy} by considering different characteristics of each label, such as the label population and ground-truth noise distribution.
The high-performance segmentation of the lumen and outside classes is a confirmation that we reached the limit of the ability to segment these classes by training models using a single-reader ground-truth.
On the other hand, the media and shadow classes suffer from higher inter- and intra-reader ground-truth variability because they are located in the areas where signal quality is decaying and have lower performance compared to other classes.
Nonetheless, these classes are great targets for future work and may benefit from unsupervised methodologies and ground-truth annotations done by a panel of readers.
Even though PS-OCT has been demonstrated with a custom imaging console, Xiong et al.~\cite{xiong2019constrained} proposed a new method that may be compatible with existing imaging consoles and could accelerate the clinical translation of using anatomical layer segmentation based on tissue polarization properties.
\begin{table}
\centering
\caption{Comparison of our model and other studies that detect the outer boundary of lumen, intima, and media.}
\input{Table/tab_boundary_result_cf}
\label{tab:boundary_result_cf}
\end{table}
For comparison with previous segmentation efforts, we compiled the reported performance of previous studies that developed segmentation methods for the lumen in Supplementary Table \ref{tab:lumen_cf_result} and for those that segmented the two additional anatomical layers in Table \ref{tab:boundary_result_cf}.
There are many methods (\cite{ughi2012automatic,cao2017automatic,tung2011automatical,wang2011automatic,moraes2013automatic,bologna2019automatic,macedo2015bifurcation,tsantis2012automatic,yong2017linear,macedo2016robust,olender2018mechanical,chen2018quantitative,yang2020novel,akbar2019automated,abdolmanafi2019fully,gharaibeh2019coronary,olender2019simultaneous}) that extract lumen with Dice 95-95\% and our method outperform them all at 99\%.
For more details, see Supplementary Table~\ref{tab:lumen_cf_result}, where we compare based on the metrics reported in the papers.
Moreover, Table~\ref{tab:boundary_result_cf} indicates that our model achieves comparable performance in the outer lumen detection task and demonstrates lower absolute distance error (ADE) for both outer intima and media boundaries compared to the two other reports accomplishing this task.
Similar to other methods~\cite{zahnd2017contour,chen2018quantitative}, we excluded thickened vessel walls from evaluation of the outer boundaries in Table~\ref{tab:boundary_result_cf}.
However, thickened vessel walls are very common, especially in the population of patients likely to undergo intravascular imaging.
Importantly, our model is able to segment cross-sections including thickened vessel wall segments with a higher ADE (2.60, 16.9, and 20.85 $\mu m$ for outer the lumen, intima, and media, respectively) but it is still comparable to the previous methods that do not even attempt to handle thickened vessel wall segments.
\section{Discussion} \label{sec:discussion}
PS-OCT complements the IV-OCT backscatter intensity signal by measuring the polarization state of reflected light and reconstructing tissue birefringence and depolarization signals. These polarimetric signals provide a more detailed characterization of the vessel wall and can help to differentiate tissue layers that have comparable scattering properties but distinct polarization features. PS-OCT enriches the visualization of anatomical layers and facilitates the downstream image processing tasks.
We proposed a convolutional neural network model with a new multi-term loss function to segment vessel lumen, intima layer, and media layer, that can leverage the increased contrast available to PS-OCT.
Furthermore, the model works on all plaque types and correctly segments the intima and media boundaries behind thickened vessel walls, as long as the plaque is not opaque.
Conversely, segments of lipid-rich or calcified plaques that impede detection of the outer anatomical layers are identified as plaque shadows. The model also segments guidewire shadows.
Our comprehensive multi-class image segmentation model can support many downstream image analysis tasks.
Automated and objective image segmentation can serve to compute anatomical metrics and enable the formulation of clinical research questions that are currently impractical to address due to the workload of manual segmentation.
Crucially, automated segmentation also enables evaluation of tissue polarization properties in distinct anatomical areas, which previously relied on tedious manual segmentation~\cite{villiger2018coronary,otsuka2019intravascular, otsuka2020intravascular}.
In a clinical setting, automated segmentation could provide a simplified view of the culprit vessel and complement the currently used lumen diameter with information on the location, extent, and polarimetry-informed composition of plaques.
This information could refine and optimize stent sizing.
\section{Conclusion} \label{sec:conclusion}
We proposed a method for the segmentation of IV-PS-OCT images. The method performance compares favorably with state-of-the-art baseline algorithms, which operate on conventional IV-OCT images. The additional polarization contrast available to IV-PS-OCT affords improved segmentation across a wide range of atherosclerotic lesion types and could be used for refined lesion characterization or guiding interventions.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{sec:intro}
\input{sections/intro}
\section{Related work}\label{sec:related}
\input{sections/related}
\section{ANDREAS Architecture}\label{sec:architecture}
\input{sections/architecture}
\section{ANDREAS Optimizer}\label{sec:optimizer}
\input{sections/model}\label{sec:model}
\input{sections/algorithm}
\section{Experimental Analysis}\label{sec:exp}
\input{sections/experimental_analysis}
\section{Conclusions and Future Work}\label{sec:conclusions}
\input{sections/conclusion}
\section*{
Acknowledgements}
\vspace{-0.2cm}
This work has been partially funded by the European Union’s Horizon 2020 research and innovation programme under the TETRAMAX grant agreement no 761349.
\bibliographystyle{splncs04}
\subsection{Assumptions and Problem Description}
This section addresses the capacity allocation problem solved by the Job Optimizer at every rescheduling point along with the main underpinning assumptions.
The viewpoint of a CSP running a data center is considered.
The data center is assumed to be composed by a set of heterogeneous nodes denoted by $\mathcal{N} = \{n_1, n_2, ..., n_N\}$, with $N = |\mathcal{N}|$.
Each node is characterized by a specific type and number of available GPUs. The number of GPUs available on node $n$ is identified by a set $\mathcal{G}_n = \{1, 2, ..., G_n\}$. The node is characterized also by a cost per unit time $c_{ng}$, when $g$ GPUs are in use. GPU homogeneity, however, is not preserved across different nodes.
Jobs are submitted continuously over time by several users, are profiled and executed by ANDREAS on a suitable node.
Each node $n \in \mathcal{N}$ can host multiple virtualized workloads, assigning in a dedicated fashion the available GPUs to jobs; a job, in turn, runs on a single node. In addition, since the number of available resources is limited, jobs that cannot be accommodated in the system are entered into a waiting set until the next rescheduling point. Let $\mathcal{J}$ represent the set of submitted jobs. Each job $j\in \mathcal{J}$ is characterized by a due date $d_j$, a tardiness penalty $\omega_j$, and a penalty to pay for postponing a job $\rho$.
Furthermore, each job is associated with an execution time $t_{jng}$ that depends on the job, the number of GPUs $g$ and their type, which in turn depends on the considered node $n$.
The parameters and variables of the problem that are considered are reported in \autoref{tab:MS_PBdescr}.
\input{tables_and_algorithms/MStable}
For each rescheduling point, the problem to be solved is a capacity allocation problem where for each job $j \in \mathcal{J}$ it must be decided whether to execute it right away or to postpone it. In the former case, the system also has to determine in which node to host the job and the resources (GPUs) to assign to its execution. If jobs are preempted, they will restart from the last snapshot (under the assumption that a snapshot of the DL model weight is taken every few epochs).
The primary objective of ANDREAS is to minimize costs,
including tardiness-caused penalties and energy related execution costs (including, e.g., air conditioning, UPS overhead, etc.). Nonetheless, the Job Optimizer at each rescheduling point considers a problem with a much narrower scope making it challenging to estimate the impact of decisions on the overall cost.
For this reason, the following proxy function, denoted by $f_{\text{OBJ}}$, is considered:
\begin{footnotesize}
\begin{equation}
\min \left[\sum_{j \in \mathcal{J}}\left(\omega_j\tau_j + \rho\omega_j\hat{\tau}_j\right) +
\sum_{j \in \mathcal{J}, n \in \mathcal{N}}\alpha_{jn}\pi_{jn}\right]
\label{eq:MS_ObjFunc}
\end{equation}
\end{footnotesize}
In particular, the proxy is expressed by two terms: the first
one represents the job tardiness and the worst-case tardiness penalty costs (where $\hat{\tau}_j$ is the tardiness a job may incur in if it is postponed to the next period).
The second term, in turn, takes into account operation costs of the fastest jobs on all nodes; $\alpha_{jn}$ is 1 if job $j$ is the first-ending job on node $n$ while $\pi_{jn}$ represents its execution cost.
This function has been designed with the precise aim of penalizing configurations with delayed jobs and, at the same time, incentivizing solutions with short-duration jobs. In this way, ANDREAS can provide a lower average job latency.
In addition, since the Job Optimizer is invoked every time the fastest job finishes its execution, such a proxy function tends to promote higher system responsiveness and better control over the quality of the overall schedule.
For a matter of space, the complete Mixed Integer NonLinear Programming (MINLP) formulation of the problem is reported in Appendix A.
\subsection{ANDREAS Randomized Greedy algorithm}
The randomized greedy heuristic that we have developed for the capacity allocation problem is based on the following criteria and assumptions.
\begin{itemize}
\item Jobs are selected according to their pressure, which is an index that measures how close they are to the due date when executed with the fastest configuration. For each job $j \in \mathcal{J}$, the pressure $\Delta_j$ is:
\begin{footnotesize}
\begin{equation}\label{eq:pressure}
\Delta_j = T_c + \min_{\substack{n \in \mathcal{N}\\ g \in \mathcal{G}_n}}{\left\{t_{jng}\right\}} - d_j,
\end{equation}
\end{footnotesize}
where $T_c$ denotes the current scheduling time.
\item The best configuration $\left(n,g\right) \in \mathcal{N} \times \mathcal{G}_n$ for each selected job is either (i) the cheapest configuration such that the job is executed before its due date, if such a configuration exists, or (ii) the fastest available configuration if, independently from the selected setup, it is not possible to execute the job before its due date. Since in our framework the time unit execution cost of a job on any available configuration is always lower than the penalty incurred if the job due date is violated, it is reasonable to choose the cheapest configuration as long as the due date can be met, while, if it is violated, it is more convenient to complete the job execution as fast as possible to reduce the corresponding penalty.
\item Execution costs increase linearly with the number of GPUs, as demonstrated by GPU-based servers energy profiles \cite{FfDL}, while the speedup of the processing time is sublinear in the number of GPUs, as observed in GPU-based application benchmarks~\cite{closer19}.
\end{itemize}
The randomized greedy algorithm, which is summarized in \autoref{lst:RG}, receives as input the set of jobs $\mathcal{J}$ in the queue, their pressures and a maximum number of iterations, and returns a solution specifying for every job
whether to execute it and the corresponding configuration.
At each iteration, a candidate solution $S$ is constructed via randomized greedy choices and is stored in $S_{best}$ if $f_{\text{OBJ}}(S) < f_{\text{OBJ}}(S_{best})$.
\input{tables_and_algorithms/RG}
At line 6, the set $\mathcal{J}$ of submitted jobs is sorted, producing as output a new set denoted by $\mathcal{J}_s$. Note that job $j$ precedes job $k$ in $\mathcal{J}_s$ if and only if $\Delta_j > \Delta_k$, i.e., job $j$ is more likely to violate its due date. As a first randomization step, some jobs in $\mathcal{J}_s$ can be swapped, with probability inversely proportional to their tardiness weight.
For all jobs $j$ in the sorted list $\mathcal{J}_s$, the set of configurations (i.e., pairs made by a node and a certain number of GPUs) such that the job can be executed within its due date (determined by the condition $t_{jng} + T_c < d_j$) is defined, in line 8, as:
\begin{footnotesize}
\[
D^*_j = \left\{(n,g) \in \mathcal{N} \times \mathcal{G}_n
:~ T_c + t_{jng} < d_j\right\}.
\]
\end{footnotesize}
The set $D^*_j$ is used to select the best configuration for job $j$, according to the following rules. Since one of the main contributions to the total cost of the schedule is the penalty for due date violations, we always try to select configurations that guarantee to complete the jobs without tardiness (or with the least possible delay, if there is no possibility to match the required due date). Therefore, we define the optimal configuration as:
\begin{footnotesize}
\begin{equation*}
\left(n^*, g^*\right) =
\begin{cases}
\argmin_{D^*_j}\left\{t_{jng}c_{ng}\right\} &\text{ if } D^*_j \neq \emptyset \\
\argmin_{n \in \mathcal{N}, g \in \mathcal{G}_n}\left\{t_{jng}\right\} &\text{ otherwise.}
\end{cases}
\end{equation*}
\end{footnotesize}
In the first case, the set $D^*_j$ is not empty, which entails that there exist configurations $(n,g)$ such that the job can be executed before its due date. In this case, $(n^*,g^*)$ is identified by the cheapest configuration, i.e., by the element in $D^*_j$ with minimum cost (identified by the product between the time-unit cost $c_{ng}$ and the expected execution time $t_{jng}$). In the second case, the job due date is violated with any available configuration, therefore we select the one that guarantees the minimum expected execution time, in order to reduce the tardiness as much as possible.
The algorithm explores the space of possible configuration by introducing some randomness in the the configuration selection process, where, instead of choosing the configuration according to the rule defined above, it selects as candidate configuration for job $j$ one of the $(n, g)$ with lower cost, with probability inversely proportional to the cost itself. Having denoted by $(n^*,g^*)$ the selected configuration for job $j$, the assignment proceeds as follows:
\begin{enumerate}
\item The algorithm tries to assign job $j$ to the required node $n^*$, provided that the number of available GPUs in $n^*$ is enough to assign $g^*$ of them to job $j$.
\item If this assignment is not feasible, it tries to assign job $j$ to a suboptimal configuration available in an open node. Iteratively, it loops over the elements in $D^*_j$ until it finds a configuration that fits in any node $n' \in \mathcal{N}_O$.
\end{enumerate}
\subsubsection*{subject to}
\begin{subequations}\label{eq:constraints2}
\scriptsize
\begin{align}
\label{MSha}
&\sum_{g \in \mathcal{G}_n}x_{jng} ~\leq~ z_{jn} &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MShc}
&z_{jn} ~\leq~ w_n &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MShd}
&x_{jng} ~\leq~ z_{jn} &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MShe}
&\sum_{n \in \mathcal{N}}z_{jn} ~\leq~ 1 &\forall &j \in \mathcal{J}\\
\label{MShg}
&\sum_{\substack{n \in \mathcal{N} \\ g \in \mathcal{G}}}x_{jng}
~=~
\sum_{n \in \mathcal{N}}z_{jn} &\forall &j \in \mathcal{J}\\
\label{MShh}
&\sum_{g \in \mathcal{G}_n}y_{gn} ~=~ w_n &\forall & n \in \mathcal{N}\\
\label{MShi}
&\sum_{\substack{j \in \mathcal{J}\\ \bar{g}\in \mathcal{G}_n}}\bar{g}x_{jn\bar{g}}
~\leq~
gy_{gn} + G_n(1-y_{gn}) &\forall & g \in \mathcal{G}_n,
\forall n \in \mathcal{N}\\
\label{MShj}
&gy_{gn}
~\leq~
\sum_{\substack{j \in \mathcal{J}\\ \bar{g} \in \mathcal{G}_n}}\bar{g}x_{jn\bar{g}} &\forall &g \in \mathcal{G}_n,
\forall n \in \mathcal{N}\\
\label{MShk}
&\sum_{\substack{g \in \mathcal{G}_n\\ n \in \mathcal{N}}}t_{jgn}x_{jng}
~\leq~
d_j + \tau_j &\forall &j \in \mathcal{J}\\
\label{MShl}
&\left(H + M_j\right)
(1 - \sum_{n \in \mathcal{N}}z_{jn})
~\leq~
d_j + \hat{\tau}_j &\forall &j \in \mathcal{J}\\
\label{MShm}
&\sum_{g \in \mathcal{G}_n}t_{jgn}c_{ng}x_{jng}
~\leq~
\pi_{jn} &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MShn}
&\sum_{j \in \mathcal{J}}\alpha_{jn} ~=~ w_n &\forall &n \in \mathcal{N}\\
\label{MSho}
&\alpha_{jn} ~\leq~ z_{jn} &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MShp}
&\sum_{n \in \mathcal{N}}w_n = \min\left\{N,J\right\} & &\\
\label{MShq}
&w_n \in \{0,1\} &\forall &n \in \mathcal{N}\\
\label{MShr}
&y_{gn} \in \{0,1\} &\forall &g \in \mathcal{G}_n
\forall n \in \mathcal{N}\\
\label{MShs}
&z_{jn} \in \{0,1\} &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MSht}
&x_{jng} \in \{0,1\} &\forall &j \in \mathcal{J},
\forall g \in \mathcal{G}_n,
\forall n \in \mathcal{N}\\
\label{MShu}
&\tau_j \geq 0 &\forall &j \in \mathcal{J}\\
\label{MShv}
&\hat{\tau}_j \geq 0 &\forall &j \in \mathcal{J}\\
\label{MShw}
&\pi_{jn} \geq 0 &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}\\
\label{MShx}
&\alpha_{jn} \in \{0,1\} &\forall &j \in \mathcal{J},
\forall n \in \mathcal{N}
\end{align}
\end{subequations}
The proposed mathematical programming formulation is non-linear,
due to the last term of the objective function \eqref{eq:OF}, representing the
sum of deployment costs of the first-ending jobs on each node.
Constraints \eqref{MSha} enforce that, for all jobs $j \in \mathcal{J}$
and for all nodes $n \in \mathcal{N}$, at most one configuration (expressed
by a number $g \in \mathcal{G}_n$ of GPUs) is selected and, in particular, that
this configuration can be selected only if job $j$ is actually deployed on node
$n$, i.e., if $z_{jn}$ is equal to 1.
Indeed, if this happens, the sum at left-hand-side, must be less than or equal to one, meaning that at most one value $g$ can be selected as the number of GPUs assigned to job $j$ on
node $n$.
Constraints \eqref{MShc} prescribe that a job $j$ can be deployed on
a node $n$ only if the node is chosen. Constraints
\eqref{MShd} enforce, in turn, that, if a job $j$ is executed on node $n$ (and
therefore $z_{jn}$ is equal to 1), at most one number $g$ of GPUs can be
assigned to that job, while no GPUs can be assigned to $j$ on $n$ if the job is
not deployed on that node.
Constraints \eqref{MShe} state that every job $j \in \mathcal{J}$ can
be deployed at most on a node $n \in \mathcal{N}$.
Since $x_{jng}$ is equal to 1 only if job $j$ is executed on $n$ with
$g$ GPUs and since, by the previous constraints, each job can be deployed at
most on a pair node/number of GPUs, the sum at left-hand-side of Constraints \eqref{MShg} is equal to 1 if job $j$ is executed.
The same holds for the sum at right-hand-side, since $z_{jn}$ is 1 if job $j$ is assigned to node $n$.
The equalities, therefore, enforce that the two variables $x_{jng}$ and $z_{jn}$ give
the same information about the execution of job $j$.
Constraints \eqref{MShh} prescribe that exactly one
configuration, represented by a number $g \in \mathcal{G}_n$ of GPUs, must be
selected on each node $n \in \mathcal{N}$ if the corresponding node is chosen.
The sum at the left-hand-side of Constraints \eqref{MShi} corresponds
to the total number of GPUs assigned to all jobs on a given node. Indeed,
variable $x_{jn\bar{g}}$ is equal to 1, for a fixed job $j$ and node $n$, only
if that job is deployed on that node with exactly $\bar{g}$ GPUs. Since,
for each running job, only one number $\bar{g}$ of GPUs can be selected,
as prescribed by the previous constraints, the sum over $\bar{g} \in
\mathcal{G}_n$ of $\bar{g}x_{jn\bar{g}}$ turns out to be equal to the number of GPUs assigned
to job $j$. The overall sum represents, therefore, the total number of GPUs
assigned to all jobs running on $n$. This number must be less than or equal to
$g$, if $g$ is the number of GPUs selected on node $n$, which happens
when variable $y_{gn}$ is equal to 1. If, in turn, $y_{gn}$ is 0, the inequality prescribes that the total number of GPUs assigned to
jobs on node $n$ is less than or equal to the maximum number
of available GPUs on the node itself.
On the other hand, Constraints \eqref{MShj} enforce the total number
of GPUs assigned to jobs on node $n$ to be greater than or equal to $g$.
Therefore, by combining \eqref{MShi} and \eqref{MShj}, the total number of
GPUs assigned to jobs on node $n$ must be exactly equal to $g$, if $g$ is the
number of GPUs selected on that node.
Constraints \eqref{MShk} and \eqref{MShl} are used to define the
tardiness and the worst-case tardiness of every job $j \in \mathcal{J}$. In
particular, Constraints \eqref{MShk} state that the total execution time of
job $j$ on the selected configuration must be less than or equal to the sum
between its due date and its tardiness, which entails that $\tau_j$ is
greater than or equal to the execution time of job $j$ minus its due date.
At the left-hand-side of Constraints \eqref{MShl}, the sum over all
nodes of $z_{jn}$ is equal to 1 when job $j$ is executed and it is 0 when job $j$ is postponed, since $z_{jn}$ is 0 for all nodes. In the first
case, the sum between the due date and the worst-case tardiness of job $j$ is
simply prescribed to be a non-negative number. In the second case,
the worst-case tardiness is defined as the sum of the scheduling interval
$H$ and the maximum execution time of job $j$, minus its due date.
Constraints \eqref{MShm} prescribe the deployment cost of job $j$ on
node $n$ to be greater than or equal to the energy cost of GPUs used on that
node, multiplied by the amount of time spent by job $j$ on $n$. Since
$x_{jng}$ is equal to 1 if job $j$ is executed on $n$ with $g$ GPUs and since
only one of those variables can be equal to one for each job, the sum
over all $g \in \mathcal{G}_n$ of $t_{jgn}x_{jng}$ is equal to the execution time
of $j$ with the unique number of GPUs it is assigned to. Moreover, the term $t_{jgn}x_{jng}$ is multiplied by $c_{ng}$ to compute the energy cost of the selected configuration.
Constraints \eqref{MShn} and \eqref{MSho} state that there must be,
on each selected node, exactly one $j \in \mathcal{J}$ being the
first-ending job. Constraint \eqref{MShp}, where $J$ and $N$ denote
the cardinalities of $\mathcal{J}$ and $\mathcal{N}$, respectively, is
used to enforce the usage of nodes. Jobs that are postponed are indeed more likely to
violate their due date. In order to avoid the resulting penalty costs, and since
job deployments can be redefined in a following rescheduling thanks to
preemption, it is always better to execute jobs as soon as resources are
available. Finally, Constraints \eqref{MShq} to \eqref{MShx} define the
domain of the variables.
\subsection{Profiling Analysis}\label{sec:exp_profiling}
\input{sections/exp_profiling}
\subsection{Simulations}\label{sec:exp_simulations}
\input{sections/exp_simulations}
\subsection{Validation in the ARMIDA setup}\label{sec:exp_real}
\input{sections/exp_real}
|
\section{Introduction}
\label{introduction}
Multiple-input multiple-output (MIMO) technology is vital for modern wireless communication systems to support explosively growing throughput requirement ~\cite{marzetta2010,larsson2014,bolei2018}. In general, the maximum a \textit{posterior} (MAP) detector delivers the optimal detection performance but has an exponential computational complexity, which is infeasible for large-sized MIMO systems~\cite{1998Multiuser}. Alternatively, suboptimal detection algorithms are implemented to achieve a better tradeoff between accuracy and complexity. The linear detectors, such as matched filter (MF), zero-forcing (ZF), and linear minimum mean-squared error (LMMSE), are with low complexity but exhibit poor performance compared to the MAP detector. On the other hand, iterative detection algorithms, e.g., approximate message passing (AMP)~\cite{donoho2009message,bayati2011}, sphere decoding (SD)~\cite{renqiu2006}, soft interference cancellation (SIC)~\cite{xiaodong1999,choi2000}, can achieve good performance with moderate complexity under some practical scenarios. All these detectors are model-specific and require complete knowledge of channel state information (CSI), which are prone to error propagation and suffer from serious performance deterioration if the system model mismatches the real transmission model or if the imperfect CSI is presented.
Over the last decade, deep learning (DL) has made profound technical revolution to many areas, such as computer vision~\cite{krizhevsky2012} and speech recognition~\cite{hinton2012}. Inspired by these successes, DL has been applied to the design of communication systems recently, including physical layer processing~\cite{qin2019deep} (channel estimation~\cite{yang2019MIMO,ye2020deep} and symbol detection~\cite{o2017introduction}), and resource management~\cite{ye2019deep,liang2019deep}, etc. Among all DL-based applications, MIMO detection is one of the most crucial and fundamental issues. DL-based detectors could learn to map the received signals into the transmitted symbols from training data and achieve better performance than the traditional detection algorithms~\cite{ye2018power}.
Generally, DL-based detectors can be divided into two categories: data-driven DL detectors based on deep neural networks (DNNs) and model-driven DL detectors from unfolding iterative detection algorithms. Data-driven DL detectors use DNN architectures to implement symbol detection~\cite{farsad2018,nir2021}. These DNN embedded architectures are model independent and can recover transmitted symbols in various scenarios with high precision if properly trained. However, such properties come at the price of a large number of trainable parameters and training samples. On the other hand, model-driven DL detectors are designed from the traditional iterative detection algorithms, where each layer of the network represents a single iteration with some trainable variables added~\cite{he2019model,Samuel2019Learning,hedetec2020}. The resulting detectors tend to have better performance and faster convergence compared to original iterative detection algorithms~\cite{hedetec2020}. However, current model-driven DL detectors are established on the premise that channel model is linear and CSI is available, limiting their application in complicated environments.
Despite their great success, data-driven DL detectors are considered as black boxes for signal reception and only experimental evaluation is available to demonstrate their performance. It is desired to understand internal mechanism of DL-based MIMO detection and provide a general design guidance. In fact, there exits a lot of literature on analyzing internal mechanisms of DNNs. The pioneering works in~\cite{cybenko1989approximation,hornik1989multilayer} have proved that any continuous function on a compact set can approximated with any precision by a DNN with sigmoid activation function. Recently, it has been proved in~\cite{montufar2014number,arora2018understanding} that DNNs with rectified linear units (ReLU DNNs) can also approximate to a large family of functions. Furthermore, DL-based channel estimation has been proved to converges to the minimum mean-squared error (MMSE) estimator as the size of training set increases in~\cite{qiang2020channel}. However, MIMO detection is a classification problem and the analysis of DL-based channel estimation in~\cite{qiang2020channel} cannot be directly generalized to DL-based MIMO detection. To the best of the authors' knowledge, there is no analytical interpretation to the advantages and disadvantages of DL-based MIMO detection.
In this paper, we analyze the performance of DL-based MIMO detection including the data-driven and the model-driven DL detectors. Our contributions are listed as follows.
\begin{itemize}
\item We prove that the data-driven DL detector with ReLU DNN can well approximate the MAP detector under sufficiently large training set in MIMO systems. The rate of convergence of the DL detector to the MAP detector scales at least polynomially fast with the size of training samples.
\item We show that the data-driven DL detector requires no CSI to approach the MAP detector for time-invariant channels and is robust to CSI uncertainty. For time-varying channels, the data-driven DL detector requires perfect CSI to converge to the MAP detector and is sensitive to CSI uncertainty.
\item We prove that the model-driven DL detector may asymptotically approach to the optimal one that minimizes the mean-squared error (MSE) or the expectation of Kullback-Leibler (KL) divergence as the size of training set increases if the original iterative detection algorithm is properly designed. In general, the model-driven DL detector requires much less training data but has lower detection accuracy than the data-driven DL detector.
\end{itemize}
The rest of this paper is organized as follows. The system model and the traditional MIMO detection algorithms are introduced in Section~\ref{preliminary}. The performance analysis of the data-driven and the model-driven DL detectors is presented in Sections~\ref{linear-sec} and~\ref{model-driven}, respectively. Simulation results are provided in Section~\ref{simulation} followed by the conclusions in Section~\ref{conclusion-sec}.
\textit{Notations}: We use lowercase letters and capital letters in boldface to denote vectors and matrices, respectively. The positive integer set, natural number set, real number set, and complex number set are denoted by $\mathbb{N}$, $\mathbb{Z}$, $\mathbb{R}$, and $\mathbb{C}$, respectively. The real and imaginary parts of a complex matrix or vector are defined by $\Re(\cdot)$ and $\Im(\cdot)$, respectively. $\mathbf{I}_{M}$ denotes the $M\times M$ identity matrix. Notations $(\cdot)^{T}$ and $(\cdot)^{H}$ represent the transpose and Hermitian of a matrix or a vector, respectively. $\mathbb{E}\{\cdot\}$ denotes the expectation, $\mathrm{tr}\{\cdot\}$ denotes the trace of a matrix, and $\mathrm{vec}(\cdot)$ denotes the vectorization of a matrix. The cardinality of a set is denoted by $|\cdot|$. Notations $\|\cdot\|_{2}$ and $\|\cdot\|_{\infty}$ represent the $2$-norm and supremum-norm of a vector or a matrix, respectively. Notation $\lceil\cdot\rceil$ represents the ceiling of a real number. Notation $\mathbf{1}_{\mathcal{A}}(x)$ is an indicator function of set $\mathcal{A}$, where $\mathbf{1}_{\mathcal{A}}(x)=1$ if $x\in\mathcal{A}$ and $\mathbf{1}_{\mathcal{A}}(x)=0$ if $x\notin\mathcal{A}$.
\section{Traditional MIMO Detection}\label{preliminary}
In this section, we first introduce a MIMO communication system and then present some traditional MIMO detection algorithms.
\subsection{System Model}
Let us consider a standard linear MIMO system with $d_{t}$ transmit and $d_{r}$ receive antennas. The $d_{r}\times 1$ received signal vector at the BS is
\begin{align}\label{signal_model}
\bar{\mathbf{x}}=\bar{\mathbf{H}}\bar{\mathbf{s}}+\bar{\mathbf{n}},
\end{align}
where $\bar{\mathbf{H}}\in\mathbb{C}^{d_{r}\times d_{t}}$ is the channel matrix, $\bar{\mathbf{s}}\in\bar{\mathbb{S}}^{d_{t}}$ is a transmitted symbol vector of mutually independent elements drawn from a discrete constellation $\bar{\mathbb{S}}$, and $\bar{\mathbf{n}}\in\mathbb{C}^{d_{t}}$ is an independent zero-mean Gaussian noise vector with element-wise variance $\sigma_{n}^{2}$.
To avoid handling complex values in MIMO detection, we re-parameterize~\eqref{signal_model} into a real-valued signal model,
\begin{align}\label{signal_model2}
\mathbf{x}=\mathbf{H}\mathbf{s}+\mathbf{n},
\end{align}
where
\begin{align}\nonumber
\mathbf{x}=\begin{bmatrix}\Re(\bar{\mathbf{x}})\\\Im(\bar{\mathbf{x}})\end{bmatrix}, \mathbf{s}=\begin{bmatrix}\Re(\bar{\mathbf{s}})\\\Im(\bar{\mathbf{s}})\end{bmatrix},
\mathbf{n}=\begin{bmatrix}\Re(\bar{\mathbf{n}})\\\Im(\bar{\mathbf{n}})\end{bmatrix},
\end{align}
and
\begin{align}\label{real_part}
\mathbf{H}=\begin{bmatrix}\Re(\bar{\mathbf{H}})&-\Im(\bar{\mathbf{H}})\\\Im(\bar{\mathbf{H}})&\Re(\bar{\mathbf{H}})\end{bmatrix}.
\end{align}
Denote $\mathbb{S}=\Re(\bar{\mathbb{S}})$ as the real part of $\bar{\mathbb{S}}$ and assume that $\Re(\bar{\mathbb{S}})=\Im(\bar{\mathbb{S}})$. Then, we have $\mathbf{s}\in\mathbb{S}^{2d_{t}}$ in~\eqref{signal_model2}.
\subsection{Traditional MIMO Detection}\label{traditional}
The following examples are some traditional MIMO detection algorithms.
\subsubsection{MAP Detector} Let $p_{o}(\mathbf{s}|\mathbf{x})$ be the \textit{posterior} probability of $\mathbf{s}$ given $\mathbf{x}$. The MAP detector is optimal in terms of minimizing the error probability of symbol detection given $\mathbf{x}$~\cite{kay1993fundamentals}, i.e.,
\begin{align}\label{map_detec}
\mathbf{s}_{\mathrm{MAP}}=\arg\max_{\mathbf{s}\in\mathbb{S}^{2d_{t}}}p_{o}(\mathbf{s}|\mathbf{x}).
\end{align}
The MAP detector in~\eqref{map_detec} is equivalent to the maximum likelihood (ML) detector when the transmitted symbols are with equal probability. However, there are two reasons that prevent the MAP detector from practical applications: 1) the optimization in~\eqref{map_detec} requires an exhausted search of $|\mathbb{S}|^{2d_{t}}$ different possible input combinations and is computationally infeasible especially when $d_{t}$ is large; 2) accurate knowledge of $p_{o}(\mathbf{s}|\mathbf{x})$ is required to implement the MAP detector, which is sometimes very hard.
\subsubsection{ZF Detector}
A common strategy for decoding $\mathbf{s}$ with affordable computational complexity is to utilize the ZF detector~\cite{1998Multiuser},
\begin{align}\label{zf_detec}
\mathbf{s}_{\mathrm{ZF}} = (\mathbf{H}^{T}\mathbf{H})^{-1}\mathbf{H}^{T}\mathbf{x}.
\end{align}
The ZF detector involves only simple matrix computations and is easy to implement in practice. However, such simplicity comes at the cost of low accuracy. The performance of the ZF detector degrades significantly when the MIMO system is nonlinear or when only imperfect CSI is available.
\subsubsection{Iterative Detector}
To better balance computational complexity and accuracy, iterative detector is used for MIMO detection. Any iterative detection algorithm, such as AMP~\cite{bayati2011} and SIC~\cite{choi2000} algorithms, can be expressed in a cascaded function as
\begin{align}\label{iterative}
\mathbf{f}_{\mathrm{iter}}(\mathbf{H},\mathbf{x})=\tilde{\mathcal{A}}_{l_{u}}\circ\tilde{\mathcal{A}}_{l_{u}-1}\circ\cdots\circ\tilde{\mathcal{A}}_{1}(\mathbf{x},\mathbf{H}),
\end{align}
where $l_{u}\in\mathbb{N}$ is the iteration number and $\tilde{\mathcal{A}_{i}}$ is the $i$-th iteration for $i\in\{1,\ldots,l_{u}\}$. Generally, each iteration in~\eqref{iterative}
is with low complexity and increasing $l_{u}$ can continuously improve the detection accuracy until the iterative detector converges. For most of the traditional iterative detectors, $\tilde{\mathcal{A}}_{i}$ keeps the same for different $i$'s and their performance is suboptimal compared to the MAP detector, which has much room for improvement.
\section{Data-driven DL Detector}\label{linear-sec}
With powerful learning ability, the data-driven DL detector can establish a stable and precise model to achieve the performance comparable with the MAP detector and has been used for MIMO detection. In this section, the performance of the data-driven DL detector is analyzed from a theoretical perspective via statistical learning theory.
\subsection{Basic Setting of Data-driven DL Detector without CSI}\label{basic}
Let us consider a data-driven DL detector, $\mathcal{D}$, with a fully-connected ReLU DNN, where only the received signal $\mathbf{x}$ is available.\footnote{The fully-connected ReLU DNN is the basis for most of current state-of-the-art DNNs~\cite{szegedy2015going,he2016deep}. In this respect, we choose the fully-connected ReLU DNN as an example to analyze the performance of the data-driven DL detector, which can be easily extended to other more advanced network structures.} To consider MIMO detection for general systems, we extend the linear model in~\eqref{signal_model} to the following statistical models
\begin{align}\label{nonlinear1}
\mathbf{x}=\mathbf{f}_{\mathrm{nlr}}(\mathbf{H}\mathbf{s}+\mathbf{n}),
\end{align}
and
\begin{align}\label{nonlinear2}
\mathbf{x}=\mathbf{H}\mathbf{f}_{\mathrm{nlt}}(\mathbf{s})+\mathbf{n},
\end{align}
where $\mathbf{f}_{\mathrm{nlr}}(\cdot):\mathbb{R}^{2d_{r}}\rightarrow\mathbb{R}^{2d_{r}}$ and $\mathbf{f}_{\mathrm{nlt}}(\cdot):\mathbb{R}^{2d_{t}}\rightarrow\mathbb{R}^{2d_{t}}$ are the unknown distortions imposed on the received and transmitted signals, respectively, e.g., imperfect power amplifier (PA) at the transmitters~\cite{costa1999impact,costa2002m} or quantization error of analog-to-digital converter (ADC) at the receivers~\cite{XULIANG2019}.
A major concern for $\mathcal{D}$ is that the constellations in communication systems are generally not taken as the targets of DNNs. To comply with standard processing in DL methods, we need to re-parameterize $\mathbf{s}$ using one-hot mapping. Let $\mathbf{s}_{i}\in\mathbb{S}^{2}$ be the $2$-dimensional vector of real-valued symbols transmitted at the $i$-th antenna. Stacking all the symbols at transmitted antenna, $\mathbf{s}$ can be expressed as
\begin{align}\label{one-hot}
\mathbf{s}=(s_{j_{1}},s_{j_{2}},\ldots,s_{j_{d_{t}}},s_{j_{d_{t}+1}},\ldots,s_{j_{2d_{t}}}),
\end{align}
where $\{(j_{1},\ldots,j_{2d_{t}})\in\mathbb{N}^{2d_{t}}:0\leq j_{i}, j_{d_{t}+i}\leq (|\mathbb{S}|-1)\ \forall i=1,\ldots,d_{t}\}$ and $\mathbf{s}_{i}=(s_{j_{i}},s_{j_{d_{t}+i}})$ for $s_{j_{i}}\in\mathbb{S}$ and $s_{j_{d_{t}+i}}\in\mathbb{S}$.
For notation convenience, we associate a unit vector $\mathbf{u}\in\mathbb{R}^{|\mathbb{S}|^{2d_{t}}}$ with each $\mathbf{s}\in\mathbb{S}^{2d_{t}}$ and the index of nonzero element of $\mathbf{u}$ can be derived from $\big[\sum_{i=0}^{d_{t}-1}(j_{d_{t}+i}|S|+j_{i})|\mathbb{S}|^{2i}+1\big]$.
In this way, $\mathbf{u}$ is a bijective transformation of $\mathbf{s}$ and we have $p_{o}(\mathbf{s}|\mathbf{x})=p_{o}(\mathbf{u}|\mathbf{x})$. Hence, one can also set $\mathbf{u}$ as the target for MIMO detection. The input-output sample set of $\mathcal{D}$ is then defined by
\begin{align}
\mathcal{Z}=\{(\mathbf{x}_{m}, \mathbf{u}_{m})\,|\,\mathbf{x}\in\mathbb{R}^{2d_{r}},\mathbf{u}\in\mathbb{R}^{|\mathbb{S}|^{2d_{t}}}, m=1,\ldots,|\mathcal{Z}|\}
\end{align}
and samples in $\mathcal{Z}$ are independent and identically distributed (i.i.d.).
The DNN of $\mathcal{D}$ consists of input and output layers, $l\in\mathbb{N}$ hidden layers, and neuron assignment $\mathbf{d}=(d_{0},d_{1},\ldots,d_{l},d_{l+1})\in\mathbb{N}^{l+1}$ with $d_{0}=2d_{r}$ and $d_{l+1}=|\mathbb{S}|^{2d_{t}}$. The depth, width, and size of $\mathcal{D}$ are defined by $l$, $\max\{d_{1},\ldots,d_{l}\}$, and $d_{u}=\sum_{i}^{l}d_{i}$, respectively.
Let
\begin{align}\label{parameter}
\Theta=\{\boldsymbol{\theta}=(\mathrm{vec}(\mathbf{W}_{0}),\mathbf{b}_{0},\ldots,\mathrm{vec}(\mathbf{W}_{l}),\mathbf{b}_{l})\in\mathbb{R}^{d_{s}}\}
\end{align}
be the set of all parameters of $\mathcal{D}$, where $d_{s}=\sum_{i=0}^{l}d_{i+1}\times (d_{i}+1)$, $\mathbf{W}_{i}\in\mathbb{R}^{d_{i+1}\times d_{i}}$ is the weight matrix connecting the $i$-th layer to the $(i+1)$-th layer, and $\mathbf{b}_{i}\in\mathbb{R}^{d_{i+1}}$ is the bias vector of the $(i+1)$-th layer for $i\in\{0,\ldots,l\}$.
For a fixed $\mathbf{d}$,
\begin{align}\label{network_output}
\mathbf{p}_{\boldsymbol{\theta}}(\mathbf{x})=\psi_{d_{l+1}}\circ\mathcal{A}_{l}\circ\varphi_{d_{l}}\circ\mathcal{A}_{l-1}\circ\varphi_{d_{l-1}}\circ\cdots\circ\varphi_{d_{1}}\circ\mathcal{A}_{0}(\mathbf{x})
\end{align}
is the underlying function of $\mathcal{D}$, where $\circ$ denotes the function composition, $\psi_{d_{l+1}}:\mathbb{R}^{d_{l+1}}\rightarrow\mathbb{R}^{d_{l+1}}$ is the entry-wise softmax function, $\mathcal{A}_{i}:\mathbb{R}^{d_{i}}\rightarrow\mathbb{R}^{d_{i+1}}$ is the affine transformation with weight $\mathbf{W}_{i}$ and bias $\mathbf{b}_{i}$, and $\varphi_{d_{i}}:\mathbb{R}^{d_{i}}\rightarrow\mathbb{R}^{d_{i}}$ is the entry-wise ReLU activation function for $i\in\{0,\ldots,l\}$.
Denote $\tilde{\mathbf{x}}_{i}=[x_{i,1},\ldots,x_{i,d_{i}}]^{T}$ as the output of the $i$-th layer with $\tilde{\mathbf{x}}_{0}=\mathbf{x}$ and $\mathcal{A}_{i-1}(\tilde{\mathbf{x}}_{i-1})=(x_{\mathcal{A}_{i-1},1},\ldots,x_{\mathcal{A}_{i-1},d_{i}})\in\mathbb{R}^{d_{i}}$. The ReLU activation function and softmax function can be expressed as
\begin{align}\label{relu}
\varphi_{d_{i}}(\mathcal{A}_{i-1}(\tilde{\mathbf{x}}_{i-1}))=(\max\{0,x_{\mathcal{A}_{i-1},1}\},\ldots,\max\{0,x_{\mathcal{A}_{i-1},d_{i}}\}),
\end{align} and
\begin{align}\label{softmax}
\psi_{d_{l+1}}(\mathcal{A}_{l}(\tilde{\mathbf{x}}_{l}))=\bigg(\frac{\exp(x_{\mathcal{A},1})}{\sum_{j=1}^{d_{l+1}}\exp(x_{\mathcal{A},j})},\ldots,\frac{\exp(x_{\mathcal{A},d_{l+1}})}{\sum_{j=1}^{d_{l+1}}\exp(x_{\mathcal{A},j})}\bigg),
\end{align} respectively.
From~\eqref{relu}, the neurons in $\mathcal{D}$ have only two states: zero output or replicating input. All possible states of neurons in $\mathcal{D}$ can be represented by a set $\mathcal{K}\subseteq\{0,1\}^{d_{u}}$ when $\boldsymbol{\theta}$ is fixed. Each element in $\mathcal{K}$ is a $d_{u}$-dimensional vector with its entries being either $0$ or $1$. Similar to~\cite{montufar2014number}, the input space of $\mathcal{D}$ is partitioned into linear regions according to different states. Denote $\mathcal{X}$ and $\mathcal{X}_{k}$ as the input space and the input region corresponding to the $k$-th state, respectively. It is obvious that
\begin{align}\label{partition}
&\mathcal{X}_{k}\subseteq\mathcal{X},\ k=1,\ldots,K=|\mathcal{K}|,\nonumber\\
&\mathcal{X}=\cup_{k=1}^{K}\mathcal{X}_{k}.
\end{align}
For $\mathbf{x}\in\mathcal{X}_{k}$, we know $\mathcal{A}_{i}(\tilde{\mathbf{x}}_{i})$ in~\eqref{network_output} satisfies
\begin{align}\label{hyperplane1}
\mathcal{A}_{i}(\tilde{\mathbf{x}}_{i})=\left\{\begin{array}{lcl}\mathbf{W}_{0}\mathbf{x}+\mathbf{b}_{0}, & & i=0,\\
\tilde{\mathbf{W}}_{i}\mathcal{A}_{i-1}(\tilde{\mathbf{x}}_{i-1})+\mathbf{b}_{i},& &i\geq 1,\end{array}\right.
\end{align}
where $\tilde{\mathbf{W}}_{i}=\mathbf{W}_{i}\mathbf{\Lambda}_{i}$ and $\mathbf{\Lambda}_{i}$ is an $\mathbb{R}^{d_{i}\times d_{i}}$ diagonal matrix whose diagonal element is either $0$ or $1$. Moreover, $\mathbf{\Lambda}_{0}=\mathbf{I}_{d_{0}}$ and $\tilde{\mathbf{W}}_{0}=\mathbf{W}_{0}\mathbf{\Lambda}_{0}$. The diagonal elements of $\mathbf{\Lambda}_{i}$ correspond to the states of neurons at the $i$-th layer. By expanding $\mathcal{A}_{i}(\tilde{\mathbf{x}}_{i})$ recursively, we further obtain
\begin{align}\label{hyperplane}
\mathcal{A}_{i}(\tilde{\mathbf{x}}_{i})&=\prod_{j=0}^{i}\tilde{\mathbf{W}}_{j}\mathbf{x}+\sum_{j=0}^{i-1}\bigg(\prod_{p=0}^{j}\tilde{\mathbf{W}}_{i-p}\bigg)\mathbf{b}_{i-1-j}+\mathbf{b}_{i}=\hat{\mathbf{W}}_{i}\mathbf{x}+\hat{\mathbf{b}}_{i},
\end{align}
where $\hat{\mathbf{W}}_{i}=\prod_{j=0}^{i}\tilde{\mathbf{W}}_{j}$ and
$\hat{\mathbf{b}}_{i}=\sum_{j=0}^{i-1}\left(\prod_{p=0}^{j}\tilde{\mathbf{W}}_{i-p}\right)\mathbf{b}_{i-1-j}+\mathbf{b}_{i}$. Let $\mathbf{f}_{\boldsymbol{\theta}}\left(\mathbf{x}\right)=\mathcal{A}_{l}(\tilde{\mathbf{x}}_{l})$ be the input at the last layer. Using~\eqref{hyperplane}, we could derive the explicit form of $\mathbf{f}_{\boldsymbol{\theta}}(\mathbf{x})$ as
\begin{align}\label{network} \mathbf{f}_{\boldsymbol{\theta}}(\mathbf{x})=\mathbf{W}_{\mathcal{X}_{k}}\mathbf{x}+\mathbf{b}_{\mathcal{X}_{k}}
\end{align}
for any $\mathbf{x}\in\mathcal{X}_{k}$, where $\mathbf{W}_{\mathcal{X}_{k}}=\hat{\mathbf{W}}_{l}$ and $\mathbf{b}_{\mathcal{X}_{k}}=\hat{\mathbf{b}}_{l}$. From~\eqref{network}, $\mathbf{f}_{\boldsymbol{\theta}}(\mathbf{x})$ turns into an affine function for any $\mathbf{x}\in\mathcal{X}_{k}$ and a piecewise linear function for any $\mathbf{x}\in\mathcal{X}$.
Applying the softmax function to $\mathbf{f}_{\boldsymbol{\theta}}(\mathbf{x})$ yields
\begin{align}\label{network2} \mathbf{p}_{\boldsymbol{\theta}}(\mathbf{x})=\psi_{d_{l+1}}(\mathbf{f}_{\boldsymbol{\theta}}(\mathbf{x}))=\psi_{d_{l+1}}(\mathbf{W}_{\mathcal{X}_{k}}\mathbf{x}+\mathbf{b}_{\mathcal{X}_{k}}).
\end{align}
Each entry of $\mathbf{p}_{\boldsymbol{\theta}}\left(\mathbf{x}\right)$ is restricted in the range $[0,1]$ and the sum of all these entries is equal to $1$, as shown in~\eqref{softmax}. Since $\mathbf{u}$ is the target for detection, $\mathbf{p}_{\boldsymbol{\theta}}\left(\mathbf{x}\right)$ can be regarded as a set of estimated \textit{posterior} probabilities for all possible $\mathbf{u}$ given $\mathbf{x}$. The goal of the data-driven DL detector is to approximate $p_{o}(\mathbf{u}|\mathbf{x})$ by optimizing $\boldsymbol{\theta}$ within some feasible set.
\subsection{Performance Analysis without CSI}\label{wCSI}
Unlike model-based detectors in~\eqref{map_detec} and~\eqref{zf_detec}, it is difficult to derive an explicit analytical form of $\mathbf{p}_{\boldsymbol{\theta}}(\mathbf{x})$ and the performance analysis of the data-driven DL detector is not straightforward.
Let $\tilde{\mathbf{u}}_{i}\in\mathbb{R}^{d_{l+1}}$ denote the unit vector whose $i$-th entry is nonzero. Note that $(\tilde{\mathbf{u}}_{1},\ldots,\tilde{\mathbf{u}}_{d_{l+1}})$ is an i.i.d. multinomial random variable with probability $((p(\tilde{u}_{1}),\ldots,p(\tilde{u}_{d_{l+1}}))$.
Denote
\begin{align}\label{posterior}
p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})=\prod_{i=1}^{d_{l+1}}p_{\boldsymbol{\theta},i}(\mathbf{x})^{\mathbf{1}_{\tilde{\mathbf{u}}_{i}}(\mathbf{u})}
\end{align}
as the estimated \textit{posterior} probability of the data-driven DL detector, where $p_{\boldsymbol{\theta},i}(\mathbf{x})$ is the $i$-th entry of $\mathbf{p}_{\boldsymbol{\theta}}(\mathbf{x})$.
To find $\boldsymbol{\theta}$ that leads to the optimal data-driven DL detector, we need a loss function to measure the distance between $p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})$ and $p_{o}(\mathbf{u}|\mathbf{x})$. Typically, the KL divergence of $p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})$ and $p_{o}(\mathbf{u}|\mathbf{x})$ is adopted, which is defined as
\begin{align}\label{kl-info}
D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})=\mathbb{E}\Big\{\ln\frac{p_{o}(\mathbf{u}|\mathbf{x})}{p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})}\Big\}.
\end{align}
Note that $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})$ is non-negative and is equal to zero if and only if $p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})=p_{o}(\mathbf{u}|\mathbf{x})$~\cite{cover1999elements}. Though the KL information is not a distance function, its convergence often implies the same trend in other metrics.
A convenient distance metric is the Hellinger metric:
\begin{align}
\gamma(p_{o},p_{\boldsymbol{\theta}})=\Big(\frac{1}{2}\mathbb{E}\big\{(p_{o}(\mathbf{u}|\mathbf{x})^{1/2}-p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})^{1/2})^{2}\big\}\Big)^{1/2}
\end{align}
and the following inequality
\begin{align}\label{helling}
\gamma^{2}(p_{o},p_{\boldsymbol{\theta}})\leq \frac{1}{2}D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})
\end{align}
holds between $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})$ and $\gamma(p_{o},p_{\boldsymbol{\theta}})$~\cite[Lemma 1.3]{vandeer2006}. From~\eqref{helling}, decreasing the KL information reduces the distance between $p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})$ and $p_{o}(\mathbf{u}|\mathbf{x})$ in the Hellinger metric. In this respect, the optimal data-driven DL detector derived from minimizing the KL divergence, $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})$, also implies that $p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})$ is in close proximity of $p_{o}(\mathbf{u}|\mathbf{x})$.
Let $\Theta_{R}=\{\boldsymbol{\theta}\,|\,\|\boldsymbol{\theta}\|_{\infty}\leq R, R\geq 1\}$ be the bounded subset of $\Theta$ and the performance of the data-driven DL detector will be evaluated within $\Theta_{R}$. Denote
\begin{align}
J(p_{o})=\mathbb{E}\{\ln p_{o}(\mathbf{u}|\mathbf{x})\}
\end{align}
as the expectation of $\ln p_{o}(\mathbf{u}|\mathbf{x})$. The optimal data-driven DL detector that minimizes $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})$ within $\Theta_{R}$ can be expressed as
\begin{align}\label{mini1}
\boldsymbol{\theta}_{o}&=\arg\min_{\boldsymbol{\theta}\in\Theta_{R}}D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})=\arg\max_{\boldsymbol{\theta}\in\Theta_{R}}J(p_{\boldsymbol{\theta}}),
\end{align}
where $J(p_{\boldsymbol{\theta}})=\mathbb{E}\{\ln p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})\}$. Eq.\,\eqref{mini1} indicates that minimizing $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}})$ is equivalent to maximizing the log-likelihood $J(p_{\boldsymbol{\theta}})$.
However, the optimization over $J(p_{\boldsymbol{\theta}})$ in~\eqref{mini1} is difficult to implement in practice. Generally,
\begin{align}\label{empirical}
J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})=\frac{1}{|\mathcal{Z}|}\sum_{(\mathbf{x}_{m},\mathbf{u}_{m})\in\mathcal{Z}}\ln p_{\boldsymbol{\theta}}(\mathbf{u}_{m}|\mathbf{x}_{m})
\end{align}
is applied to optimize $\boldsymbol{\theta}$ with respect to (w.r.t.) $\mathcal{Z}$, and the corresponding maximum log-likelihood detector is given by
\begin{align}\label{min2}
\boldsymbol{\theta}_{\mathcal{Z}}=\arg\max_{\boldsymbol{\theta}\in\Theta_{R}}J_{\mathcal{Z}}(p_{\boldsymbol{\theta}}).
\end{align}
Obviously,
\begin{align}
\sum_{(\mathbf{x}_{m},\mathbf{u}_{m})\in\mathcal{Z}}\ln \frac{p_{\boldsymbol{\theta}_{\mathcal{Z}}}(\mathbf{u}_{m}|\mathbf{x}_{m})}{p_{\boldsymbol{\theta}}(\mathbf{u}_{m}|\mathbf{x}_{m})}\geq 0.
\end{align}
The detected symbol of the data-driven DL detector can then be expressed as
\begin{align}\label{dl_det}
\mathbf{s}_{\mathrm{DL}}=\arg\max_{\mathbf{s}\in\mathbb{S}^{d_{l+1}}} p_{\boldsymbol{\theta}_{\mathcal{Z}}}(\mathbf{f}_{\mathrm{map}}(\mathbf{s})|\mathbf{x}),
\end{align}
where $\mathbf{f}_{\mathrm{map}}(\mathbf{s})$ is one-to-one mapping from $\mathbf{s}$ to $\mathbf{u}$.
According to~\eqref{kl-info}, $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}_{\mathcal{Z}}})$ can be decomposed into
\begin{align}\label{decom}
D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}_{\mathcal{Z}}})=[J(p_{o})-J(p_{\boldsymbol{\theta}_{o}})]+[J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})].
\end{align}
The first term $[J(p_{o})-J(p_{\boldsymbol{\theta}_{o}})]$ in~\eqref{decom} is non-negative and is independent of $\mathcal{Z}$, referred to as the approximation error. The second term $[J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]$ in~\eqref{decom} is also non-negative and is determined by $\mathcal{Z}$, called as the generalization error.
Denote $f(\mathbf{x})$ as an $\mathbb{R}^{d}\rightarrow\mathbb{R}$ function and $\ell_{2}$ be the finite 2-norm space of $f(\mathbf{x})$ with
\begin{align}
\|f(\mathbf{x})\|_{2}=\big[\mathbb{E}\{f^{2}(\mathbf{x})\}\big]^{1/2}<+\infty.
\end{align}
The following theorem, proved in Appendix~\ref{apex-a}, demonstrates that the approximation error in~\eqref{decom} can be narrowed down with any precision by ReLU DNNs.
\begin{theorem}\label{theorem1}
If $\ln p_{o}(\mathbf{u}|\mathbf{x})\in\ell_{2}$, then there exits an optimized DL estimator built on a ReLU DNN of $\boldsymbol{\theta}\in\Theta_{R}$ with sufficiently large $R$ and at most $
\lceil\log_{2}(d_{0}+1)\rceil
$ hidden layers such that
\begin{align}\label{theo1}
J(p_{o})-J(p_{\boldsymbol{\theta}_{o}})\leq\varepsilon
\end{align}
for any $\varepsilon>0$.
\end{theorem}
\begin{remark}\label{remark3}
From Theorem~\ref{theorem1}, the data-driven DL detector with bigger network size is more powerful at function representation and tends to have lower approximation error and better performance.
\end{remark}
\begin{remark}
Theorem~\ref{theorem1} indicates that the data-driven DL detector is model independent and is capable of approximating any target \textit{posterior} distribution. Therefore, the data-driven DL detector is a preferred choice compared to other model-based MIMO detection algorithms if no specific channel model is known a priori or if complicated nonlinear systems are presented.
\end{remark}
Next, we will discuss the convergence of the generalization error in~\eqref{decom}. The following two auxiliary lemmas, proved in Appendixes~\ref{apex-b} and~\ref{apex-c}, respectively, are presented first.
\begin{lemma}\label{lemma-1}
Let $\alpha=R\|\mathbf{d}\|_{\infty}$, $\beta=\alpha/(\alpha-1)$, and
\begin{align}
\nu=\mathbb{E}\big\{\big[(\ln d_{l+1}+1)(\alpha^{l+1}(\|\mathbf{x}\|_{2}+\beta)-\beta)\big]^{2}\big\}.
\end{align}
Assume that $\mathbb{E}\{\|\mathbf{x}\|_{2}^{2}\}$ is finite and so is $\nu$. For any $\varepsilon>0$ and
$|\mathcal{Z}|\geq 4\nu/\varepsilon^{2}$,
we have
\begin{align}\label{lemma1}
\mathbf{P}\big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J(p_{\boldsymbol{\theta}})|>\varepsilon\big)\leq 4\mathbf{P}\big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(p_{\boldsymbol{\theta}})|>\frac{\varepsilon}{4}\big),
\end{align}
where $\mathbf{P}$ is the distribution of training samples in $\mathcal{Z}$ and
\begin{align}
J_{\mathcal{Z}}^{\circ}(p_{\boldsymbol{\theta}})=\frac{1}{|\mathcal{Z}|}\sum_{m=1}^{|\mathcal{Z}|}\omega_{m}\ln p_{\boldsymbol{\theta}}(\mathbf{u}_{m}|\mathbf{x}_{m})
\end{align}
with $\{\omega_{1},\ldots,\omega_{|\mathcal{Z}|}\}$ a Rademacher sequence.
\end{lemma}
\begin{lemma}\label{lemma2}
Assume that $\frac{1}{|\mathcal{Z}|}\sum_{m=1}^{|\mathcal{Z}|}\|\mathbf{x}_{m}\|_{2}^{2}\leq\delta^{2}$. For any $\boldsymbol{\theta},\boldsymbol{\lambda}\in\Theta_{R}$, we have
\begin{align}
&|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{\boldsymbol{\lambda}})|\leq 3^{l+1}\|\mathbf{d}\|_{\infty}\alpha^{l}(\delta+\beta)\|\boldsymbol{\theta}-\boldsymbol{\lambda}\|_{\infty}.
\end{align}
\end{lemma}
If there exits a collection of functions $\mathbf{p}_{1}(\mathbf{x}),\ldots,\mathbf{p}_{C}(\mathbf{x})\in\mathcal{D}$ with their parameters belonging to $\Theta_{R}$ for $C\in\mathbb{N}$, the functions in this collection satisfy
\begin{align}\label{cover}
|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{j})|\leq\varepsilon,\ \forall j\in\{1,\ldots,C\},
\end{align}
for any $\mathbf{p}_{\boldsymbol{\theta}}(\mathbf{x})\in\mathcal{D}$ with $\boldsymbol{\theta}\in\Theta_{R}$ and $\varepsilon>0$, where
\begin{align}
p_{j}(\mathbf{u}|\mathbf{x})=\prod_{i=1}^{d_{l+1}}p_{j,i}(\mathbf{x})^{\mathbf{1}_{\tilde{\mathbf{u}}_{i}}(\mathbf{u})}
\end{align}
and $p_{j,i}(\mathbf{x})$ is the $i$-th entry of $\mathbf{p}_{j}(\mathbf{x})$.
Define the covering number $C(\varepsilon, \Theta_{R})$ as the smallest value of $C\in\mathbb{N}$ that satisfies~\eqref{cover}. According to Lemma~\ref{lemma2} and~\cite[Lemma~3]{qiang2020channel}, the upper bound on $\ln C(\varepsilon, \Theta_{R})$ is given by
\begin{align}\label{covering}
\ln C(\varepsilon, \Theta_{R})\leq d_{s}\ln\Big[\frac{3^{l+1}4\|\mathbf{d}\|_{\infty}\alpha^{l}(\delta+\beta)}{\varepsilon}\Big]
\end{align}
for any $\varepsilon>0$.
Following~\eqref{covering}, the next theorem, proved in Appendix~\ref{apex-d}, demonstrates the rate of convergence of the generalization error in~\eqref{decom}.
\begin{theorem}\label{theorem2}
Let $\alpha=R\|\mathbf{d}\|_{\infty}$, $\beta=\alpha/(\alpha-1)$, $\mu=\mathbb{E}\{\|\mathbf{x}\|_{2}^{2}\}$, and
\begin{align}
\nu=\mathbb{E}\big\{\big[(\ln d_{l+1}+1)(\alpha^{l+1}(\|\mathbf{x}\|_{2}+\beta)-\beta)\big]^{2}\big\}.
\end{align} Denote $\sigma$ as the variance of $\|\mathbf{x}\|_{2}^{2}$. Let $\delta_{1}=[(\ln d_{l+1}+1)(\alpha^{l+1}\delta+\beta)-\beta)]^{2}$ and $\delta_{2}=3^{l+1}2^{6}\alpha^{l}(\delta+\beta)$ for any $\delta^{2}\geq\mu$. For any $\varepsilon>0$, we have
\begin{align}
\mathbf{P}([J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]>\varepsilon)\leq 8\mathrm{exp}\Big(-\frac{|\mathcal{Z}|\varepsilon^{2}}{1024\delta_{1}}\Big)+\frac{4\sigma^{2}}{|\mathcal{Z}|(\delta^{2}-\mu)^{2}}
\end{align}
if $|\mathcal{Z}|\geq 16\nu/\varepsilon^{2}$ and $|\mathcal{Z}|\geq (1024\delta_{1}d_{s}\ln\frac{\delta_{2}}{\varepsilon})/\varepsilon^{2}$.
\end{theorem}
\begin{remark}
Though the approximation error can be reduced by enlarging network size as indicated in Theorem~\ref{theorem1}, Theorem~\ref{theorem2} demonstrates that the rate of the convergence of the generalization error will decrease. Therefore, a tradeoff between the generalization error and the approximation error should be carefully balanced when implementing the data-driven DL detector.
\end{remark}
\begin{remark}
Theorem~\ref{theorem2} demonstrates that the rate of convergence of $J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})$ to $J(p_{\boldsymbol{\theta}_{o}})$ grows at least polynomially with $|\mathcal{Z}|$ under a fixed network structure.
\end{remark}
The following corollary, proved in Appendix~\ref{apex-e}, presents our main conclusion on the performance of the data-driven DL detector.
\begin{corollary}\label{corollary}
For any $\varepsilon>0$ and sufficiently large $R$, there exits a DL detector built on a ReLU DNN with at most
$\lceil\log_{2}(d_{0}+1)\rceil
$ hidden layers and $\boldsymbol{\theta}\in\Theta_{R}$
such that
\begin{align}\label{coro1-1}
\lim_{|\mathcal{Z}|\rightarrow+\infty}\mathbf{P}(D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}_{\mathcal{Z}}})>\varepsilon)=0.
\end{align}
\end{corollary}
\begin{remark}\label{remark2}
Corollary~\ref{corollary} and~\eqref{helling} show that
\begin{align}\label{con}
p_{\boldsymbol{\theta}_{\mathcal{Z}}}(\mathbf{u}|\mathbf{x})\approx p_{o}(\mathbf{u}|\mathbf{x}),
\end{align}
if $|\mathcal{Z}|$ is sufficiently large and the network structure is suitably configured. Therefore, the data-driven DL detector can learn to fit $p_{o}(\mathbf{u}|\mathbf{x})$ perfectly and is very suitable for communication systems with unknown nonlinear detrimental effects.
\end{remark}
\subsection{Performance Analysis with CSI}\label{incomplete}
\begin{figure*}[t]
\vskip 0.2in
\centering
\subfigure[]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\columnwidth]{simulation_sys1.pdf}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\columnwidth]{simulation_sys2.pdf}
\end{minipage}
}
\caption{Network structures for the data-driven DL detectors with CSI and without CSI.}
\label{figure6-2}
\vskip -0.2in
\end{figure*}
The data-driven DL detector derived in~\eqref{dl_det} is model independent and requires no CSI to learn a detection mapping approaching MAP performance. However, it is only applicable when the channel is static and it does not perform well for varying channels~\cite{Samuel2019Learning}. Next, we will show the necessity of CSI for DL-based MIMO detection to generalize over the whole distribution of $\mathbf{H}$.
From~\eqref{con}, the data-driven detector takes the \textit{posterior} probability, $p_{o}(\mathbf{u}|\mathbf{x})$, as the target and we have
\begin{align}\label{ml1}
p_{o}(\mathbf{u}|\mathbf{x})=p_{o}(\mathbf{u}|\mathbf{f}_{\mathrm{nlr}}(\mathbf{H}\mathbf{s}+\mathbf{n}))
\end{align}
and
\begin{align}
p_{o}(\mathbf{u}|\mathbf{x})=p_{o}(\mathbf{u}|(\mathbf{H}\mathbf{f}_{\mathrm{nlt}}(\mathbf{s})+\mathbf{n}))
\end{align}
according to~\eqref{nonlinear1} and~\eqref{nonlinear2}, respectively. To exploit $\mathbf{s}_{\mathrm{DL}}$ in different scenarios, we discuss two channel states:
\begin{itemize}
\item In time-invariant channels, $\mathbf{H}$ remains deterministic and constant during the training and detecting phases.
\item In time-varying channels, $\mathbf{H}$ is randomly generated from a known continuous distribution and changes in each realization of one training sample.
\end{itemize}
In the time-invariant channels, channel matrices in training data and deployed environment are identical, and $\mathbf{s}$ can be recovered with low detection error using the minimum distance rule. In view of $\mathbf{s}_{\mathrm{DL}}\approx\mathbf{s}_{\mathrm{MAP}}$ from Corollary~\ref{corollary}, the performance of the data-driven DL detector without CSI is not affected by the state of $\mathbf{H}$ in the time-invariant channels.
In time-varying channels, channel matrices in training data and deployed environment are different due to randomness. As a consequence, $p_{\boldsymbol{\theta}_{\mathcal{Z}}}(\mathbf{u}|\mathbf{x})$ learned from training data is inconsistent with the real \textit{posterior} probability in the deployed environment and $\mathbf{s}$ is indistinguishable using the maximum likelihood (ML) rule. The data-driven DL detector is unable to decouple transmitted symbols from time-varying channels owing to a lack of CSI and using a single network without the information of $\mathbf{H}$ cannot generalize the entire distribution of possible channels for MIMO detection.
To alleviate the impact of time-varying channels, let us consider the case that channel $\mathbf{H}$ is already known. The MAP detector in~\eqref{map_detec} is converted into
\begin{align}\label{map1}
\mathbf{s}_{\mathrm{MAP}}=\arg\max_{\mathbf{s}\in\mathbb{S}^{2d_{t}}}p_{o}(\mathbf{s}|\,\mathbf{H},\mathbf{x}),
\end{align}
where $p_{o}(\mathbf{s}|\,\mathbf{H},\mathbf{x})$ is the \textit{posterior} probability of $\mathbf{s}$ given $\mathbf{H}$ and $\mathbf{x}$.
To approximate the MAP detector in~\eqref{map1}, the data-driven DL detector should take both $\mathbf{H}$ and $\mathbf{x}$ as its input and other structures of the data-driven DL detector remain the same. The network structures for the data-driven DL detectors with and without CSI are illustrated in Fig.~\ref{figure6-2}. Similar to Corollary~\ref{corollary}, we can also prove that the data-driven DL detector with CSI can well approximate the MAP detector in~\eqref{map1}. Therefore, the data-driven DL detector manages to generalize over all possible realizations of $\mathbf{H}$ by incorporating CSI into the network design.
\begin{remark}
In time-varying channels, CSI is essential to the data-driven DL detector to detect $\mathbf{s}$ over all possible realizations of the channel. Moreover, the data-driven DL detector with CSI is also model independent and can achieve the MAP comparable performance over various scenarios. In this situation, the data-driven DL detector tends to have a large network structure by taking $\mathbf{H}$ as the input and thus requires enormous train samples to converge, as shown in Theorem~\ref{theorem2}. The data-driven DL detector with CSI yields optimal accuracy but is prohibitive for large-sized MIMO systems.
\end{remark}
\section{Model-driven DL Detector}\label{model-driven}
Simply implementing the DL detector in a data-driven fashion is ineffective in some practical scenarios, especially for large-sized MIMO systems. An alternative way is to integrate the model expert knowledge into the network structure and design model-driven DL detectors. Instead of using a conventional DNN, model-driven DL detectors usually associate iterative detection algorithms with DL, which can mitigate the impact of time-varying channels, accelerate the convergence, and reduce the complexity~\cite{hedetec2020}. Each layer of the model-driven DL detector operates a single iteration in~\eqref{iterative} and adds some trainable variables to enhance the detection performance.
Suppose that the MIMO channel model is linear as shown in~\eqref{signal_model2} and channel $\mathbf{H}$ is known at the receiver. Model-driven DL detector $\mathcal{D}_{u}$ is based on the iteration detector in~\eqref{iterative} and is given by
\begin{align}
\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{H},\mathbf{x})=\tilde{\mathcal{A}}_{\boldsymbol{\vartheta}_{l_{u}}}\circ\tilde{\mathcal{A}}_{\boldsymbol{\vartheta}_{l_{u}-1}}\circ\cdots\circ\tilde{\mathcal{A}}_{\boldsymbol{\vartheta}_{0}},
\end{align}
where $\boldsymbol{\vartheta}_{i}$ is the trainable parameter, $\boldsymbol{\vartheta}=(\boldsymbol{\vartheta}_{0},\ldots,\boldsymbol{\vartheta}_{l_{u}})$ is the parameter of $\mathcal{D}_{u}$, and $\tilde{\mathcal{A}}_{\boldsymbol{\vartheta}_{i}}$ is the computation at the $i$-th iteration that is parameterized by $\boldsymbol{\vartheta}_{i}$ for $i\in\{0,\ldots,l_{u}\}$. In particular, $\tilde{\mathcal{A}}_{\boldsymbol{\vartheta}_{i}}$ depends on underlying iterative detection algorithm, i.e., $\mathbf{f}_{\mathrm{iter}}$ in~\eqref{iterative}. The training sample set of $\mathcal{D}_{u}$ is defined by
\begin{align}
\Omega=\{(\mathbf{x}_{m},\mathbf{H}_{m},\mathbf{s}_{m})|\mathbf{x}\in\mathbb{R}^{2d_{r}},\mathbf{H}\in\mathbb{R}^{2d_{r}\times 2d_{t}},\mathbf{s}\in\mathbb{S}^{2d_{t}},m=1,\ldots,|\Omega|\}.
\end{align}
Let $J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x},\mathbf{H}))$ be the non-negative loss function between $\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x},\mathbf{H})$ and $\mathbf{s}$ and $\Theta_{u}=\{\boldsymbol{\vartheta}\in\mathbb{R}^{\tilde{d}_{u}}\}$ be the parameter set for $\tilde{d}_{u}\in\mathbb{N}$.\footnote{The selection of $J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x},\mathbf{H}))$ is not fixed, which may be MSE or KL divergence.} The goal of $\mathcal{D}_{u}$ is to optimize $\boldsymbol{\vartheta}$ by minimizing $J(\mathbf{f}_{\boldsymbol{\vartheta}})$ within $\Theta_{u}$,
\begin{align}
\boldsymbol{\vartheta}_{o}=\arg\min_{\boldsymbol{\vartheta}\in\Theta_{u}}\mathbb{E}\{J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x},\mathbf{H}))\}=\arg\min_{\boldsymbol{\vartheta}\in\Theta_{u}}J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}}),
\end{align}
where $\boldsymbol{\vartheta}_{o}$ is the optimized parameter and $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_{o}})$ has the lowest mean loss over all $\boldsymbol{\vartheta}\in\Theta_{u}$.
However, it is difficult to obtain the explicit form of $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}})$ and the empirical mean
\begin{align}
J_{\Omega}(\mathbf{f}_{\boldsymbol{\vartheta}})=\frac{1}{|\Omega|}\sum_{(\mathbf{x}_{m},\mathbf{H}_{m},\mathbf{h}_{m},)\in\Omega}J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x}_{m},\mathbf{H}_{m}))
\end{align}
is typically used to optimize $\boldsymbol{\vartheta}$ w.r.t. $\Omega$. Denote
\begin{align}\label{unfolded}
\boldsymbol{\vartheta}_{\Omega} = \arg\min_{\boldsymbol{\vartheta}\in\Theta_{u}}J_{\Omega}(\mathbf{f}_{\boldsymbol{\vartheta}}),
\end{align}
and
\begin{align}\label{unfolded3}
J_{\Omega}(\mathbf{f}_{\boldsymbol{\vartheta}_{\Omega}})=\min_{\boldsymbol{\vartheta}\in\Theta_{u}}J_{\Omega}(\mathbf{f}_{\boldsymbol{\vartheta}_{\Omega}}).
\end{align}
From~\eqref{unfolded} and~\eqref{unfolded3}, $\mathbf{f}_{\boldsymbol{\vartheta}_{\Omega}}(\mathbf{x},\mathbf{H})$ is the optimal model-driven DL detector trained by $\Omega$. Therefore, we can evaluate the performance of the model-driven DL detector by quantifying the distance between $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_{\Omega}})$ and $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_{o}})$.
For an integer $C\in\mathbb{N}$ and a collection of functions $\mathbf{f}_{1}(\mathbf{x},\mathbf{H}),\ldots,\mathbf{f}_{C}(\mathbf{x},\mathbf{H})\in\mathcal{D}_{u}$, let $C_{u}(\epsilon,\Theta_{u})$ be the smallest value of $C$ such that
\begin{align}
\min_{j\in\{1,\ldots,C\}}\big|J_{\Omega}(\mathbf{f}_{\boldsymbol{\vartheta}})-J_{\Omega}(\mathbf{f}_{j})\big|\leq\varepsilon
\end{align}
for $\varepsilon>0$ and any $\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x},\mathbf{H})\in\mathcal{D}_{u}$. The following theorem, proved in Appendix~\ref{apex-f}, shows the relationship between $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_\Omega})$ and $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_o})$.
\begin{theorem}\label{theorem3}
For any $\varepsilon>0$ and $\delta_{u}>0$, we have
\begin{align}\label{unfolded_network}
\mathbf{P}_{u}([J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_\Omega})-J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_o})]>\varepsilon
)\leq8\exp\left(\ln C_{u}-\frac{|\Omega|\varepsilon^{2}}{512\delta_{u}}\right)+\mathbf{P}_{\Omega},
\end{align}
where $\mathbf{P}_{u}$ denotes the distribution of the training samples in $\Omega$ and
\begin{align}
\mathbf{P}_{\Omega}=\mathbf{P}_{u}\bigg(\frac{1}{|\Omega|}\sum_{m=1}^{|\Omega|}[J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x}_{m},\mathbf{H}_{m}))]^{2}\geq\delta_{u}\bigg).
\end{align}
\end{theorem}
In general, $\frac{1}{|\Omega|}\sum_{m=1}^{|\Omega|}[J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x}_{m},\mathbf{H}_{m}))]^{2}$ converges to its mean value as $|\Omega|$ increases, and $|J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_\Omega})-J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_o})|$ also converges to zero if $\ln C_{u}$ in~\eqref{unfolded_network} is finite. Therefore, $\mathbf{f}_{\boldsymbol{\vartheta}_{\Omega}}(\mathbf{x},\mathbf{H})$, asymptotically approaches to the optimal model-driven DL detector, $\mathbf{f}_{\boldsymbol{\vartheta}_{o}}(\mathbf{x},\mathbf{H})$, as the size of training data increases. Furthermore, the model-driven DL detector $\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x},\mathbf{H})$ is equivalent to underlying iterative detection algorithm $\mathbf{f}(\mathbf{x},\mathbf{H})$ if $\boldsymbol{\vartheta}$ is set to be specific value so that $\hat{\mathcal{A}}_{\boldsymbol{\vartheta}_{i}}=\hat{\mathcal{A}}_{i}$ for $i\in\{0,\ldots,l_{u}\}$. Hence, the performance of the model-driven DL detectors is better or at least equal to that of its underlying iterative detection algorithm, as indicated by Theorem~\ref{theorem3}.
\begin{remark}
In general, the dimension of the parameter space of the model-driven detector, $\tilde{d}_{u}$, is much smaller than that of the data-driven DL detector. Theorem~\ref{theorem3} demonstrates that model-driven DL detectors require far fewer train samples to converge, making them more suitable to large-sized MIMO systems. The performance of the model-driven DL detectors is determined by underlying iterative detection algorithms, while most of these algorithms cannot guarantee the convergence to the MAP detector except under some special scenarios. Generally, there exits the performance gap between the model-driven and the data-driven DL detectors.
\end{remark}
\begin{remark}
The model-driven DL detectors are customized for specific MIMO systems and thus do not capture the model independent property of DL. On the other hand, the model-driven DL detectors may become unreliable and divergent where the presumptive system model mismatches the real environment, which severely limits their applications.
\end{remark}
\section{Simulation Results}\label{simulation}
In this section, computer simulation is provided to verify that the data-driven DL detector asymptotically approaches to the MAP detector under linear and nonlinear MIMO systems. Moreover, simulation results shows that CSI is essential for the data-driven DL detector to achieve the MAP comparable performance over the time-varying channels. Moreover, simulation results demonstrate that the underlying iterative detection algorithm is the determinant factor that affects the performance of the model-driven DL detector.
\subsection{Simulation Setting}
The SNR is defined as
\begin{align}
\mathrm{SNR} = \frac{\mathbb{E}\|\mathbf{H}\mathbf{s}\|_{2}^{2}}{\mathbb{E}\|\mathbf{n}\|_{2}^{2}}.
\end{align}
The network of the data-driven DL detector has $4$ layers and each hidden layer is equipped with the same number of neurons. Denote $\tilde{d}$ as the width of the data-driven DL detector and $\tilde{d}=100$. We consider QPSK constellation and all transmitted symbols are generated with equal probability.
For data-driven DL detector, the simulation results are evaluated under the FC and VC cases, respectively. Unless stated otherwise, we train the network of the data-driven DL detector for $1,00,000$ iterations with a batch size of $2,000$ independently generated samples and test over $10,000$ samples. The MAP detector in~\eqref{map_detec} serves as the benchmark. The traditional model-based ZF, AMP, and SD algorithms are used to test against the data-driven DL detector.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\columnwidth]{simulation_sys3.pdf}
\caption{The network structure of the SIC-Net.}
\label{figure6-3}
\end{figure}
The model-driven DL detector in simulation is based on the SIC algorithm, which is referred as to SIC-Net. Let $\hat{\mathbf{p}}_{i}^{(q)}\in\mathbb{R}^{|\mathbb{S}|^{2}}$ denote the estimated probability vector at the $i$-th transmitted antenna for $i\in\{1,\ldots,d_{t}\}$, where $q\in\{1,\ldots,Q\}$ is the iteration index for $Q\in\mathbb{N}$. The $j$-th entry of $\hat{\mathbf{p}}_{i}^{(q)}\in\mathbb{R}^{|\mathbb{S}|^{2}}$ is $\hat{p}_{i,j}^{(q)}$ that presents the estimated probability of $\mathbf{s}_{j}$ for $j\in\{1,\ldots,|\mathbb{S}|^{2}\}$. Let $\mathbf{s}_{i}^{(q)}$ be the estimated expected symbol at the $i$-th transmitted antenna for the $q$-th iteration, computed via
\begin{align}\label{unfolded2}
\mathbf{s}_{i}^{(q)}=\tau_{q}\bigg(\sum_{j=1}^{|\mathbb{S}|^{2}}\mathbf{s}_{j}\hat{p}_{i,j}^{(q)}+\xi_{q}\mathbf{s}_{i}^{(q-1)}\bigg),
\end{align}
where $\tau_{q}$ and $\xi_{q}$ are trainable variables. Other settings are the same as the SIC algorithm in~\cite{choi2000} and the corresponding network structure is illustrated in Fig.~\ref{figure6-3}. The performance of SIC-Net is evaluated in linear Gaussian MIMO channels. We train SIC-Net with a relatively small $5,000$ samples and test over $10,000$ samples. The MAP detector in~\eqref{map_detec} is used as the benchmark. Other model-driven DetNet~\cite{Samuel2019Learning} and OAMP-Net~\cite{hedetec2020} MIMO detectors are adopted for comparison.
\subsection{Linear Systems}\label{linear-simu}
In this subsection, we investigate the bit-error rate (BER) performance and convergence of the data-driven DL detector under a linear MIMO model in~\eqref{signal_model2}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\columnwidth]{simulation_fig1.pdf}
\caption{The BER performance of the data-driven DL detector versus SNR compared to other MIMO detectors over the time-invariant channel.}
\label{figure3-1}
\end{figure}
\begin{figure*}[t]
\vskip 0.2in
\centering
\subfigure[Network width]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\columnwidth]{simulation_fig8.pdf}
\end{minipage}
}
\subfigure[Sample size]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\columnwidth]{simulation_fig5.pdf}
\end{minipage}
}
\caption{The BER performance of the data-driven DL detector versus $\tilde{d}$ and $|\mathcal{Z}|$ over the time-invariant channel, respectively.}
\label{figure2}
\vskip -0.2in
\end{figure*}
\subsubsection{Time-invariant Channel}\label{FC Model}
Fig.~\ref{figure3-1}(a) compares the BER performance of the model-based ZF, AMP, SD, MAP and the data-driven DL detectors versus SNR where a $4\times 4$ time-invariant correlated channel is generated according to the one-ring model in~\cite{shiu2000}. We assume that perfect CSI is available for the ZF, AMP, and SD detectors while data-driven DL detector has no CSI. Moreover, the MAP detector is evaluated under both perfect and imperfect CSI, respectively. As shown in Fig.~\ref{figure3-1}, the data-driven DL detector can well approximate the MAP detector of perfect CSI and significantly outperforms other model-based detectors, which confirms that $p_{\boldsymbol{\theta}_{\mathcal{Z}}}(\mathbf{u}|\mathbf{x})\approx p_{o}(\mathbf{u}|\mathbf{x})$ in Corollary~\ref{corollary}. Moreover, the MAP detector with imperfect CSI suffers from serious performance degradation. Nevertheless, the data-driven DL detector is immune to CSI uncertainty since it requires no channel information for training.
Fig.~\ref{figure2}(a) shows the BER performance of the data-driven DL detector versus the network width, $\tilde{d}$, under fixed SNRs with the $2\times 2$ Gaussian channel. We train the network for $400,000$ independently generated samples. The BERs of the MAP detectors derived at the same SNRs are used as the benchmark. The approximation error determines the BER performance of the DL estimator since $|\mathcal{Z}|$ is sufficiently large. When $\tilde{d}$ is small, the dimension of the parameter space $\Theta_{R}$ is not big enough to fit with $p_{o}(\mathbf{u}|\mathbf{x})$. Hence, the BERs of the data-driven DL detector is significantly larger than those of the MAP detector. As $\tilde{d}$ increases, the dimension of the parameter space of $\Theta_{R}$ is enlarged and the approximation error decreases until both BERs converge, which verifies Theorem~\ref{theorem1}.
Fig.~\ref{figure2}(b) shows the BER performance of the data-driven DL estimator versus the size of training samples, $|\mathcal{Z}|$, over the $2\times 2$ Gaussian channel. The SNRs are fixed and $\tilde{d}=100$. As in Fig.~\ref{figure2}(a), the BERs of the MAP detector are used as the benchmark. Similarly, the generalization error is the main factor that affects the BER performance of the DL estimator under large $\tilde{d}$. When $|\mathcal{Z}|$ is small, the BERs of the data-driven DL detector do not converge and are significantly larger than those of the MAP estimator. As $|\mathcal{Z}|$ increases, the BERs of the data-driven DL detector gradually approach to those of the MAP detector, which verifies Theorem~\ref{theorem2}.
\subsubsection{Time-varying Channel} Fig.~\ref{figure6-1}(a) compares the BER performance of the model-based ZF, AMP, SD, MAP and the data-driven DL detectors versus SNR over the $2\times 2$ Gaussian channel. We assume that the ZF, AMP, and SD detectors have perfect CSI while both the DL and MAP detectors are evaluated under perfect and imperfect CSI, respectively. As illustrated in Fig.~\ref{figure5}(a), the data-driven DL detector manages to achieve the Bayes-optimal BER performance by incorporating the perfect $\mathbf{H}$ in time-varying channels and outperforms the other model-based detectors substantially. Fig.~\ref{figure5}(a) demonstrates that the data-driven DL detector can learn properly over time-varying channels. However, the BER performance of the data-driven DL detector is severely deteriorated by imperfect CSI in time-varying channels and can only approach to the MAP detector with imperfect CSI.
\begin{figure*}[t]
\vskip 0.2in
\centering
\subfigure[Gaussian channel]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\columnwidth]{simulation_fig3.pdf}
\end{minipage}
}
\subfigure[Correlated channel]{
\begin{minipage}[t]{0.46\linewidth}
\centering
\includegraphics[width=\columnwidth]{simulation_fig4.pdf}
\end{minipage}
}
\caption{The BER performance of data-driven DL detector versus SNR compared to other MIMO detectors over the time-varying channel.}
\label{figure6-1}
\vskip -0.2in
\end{figure*}
Fig~\ref{figure6-1}(b) compares the BER performance of the model-based ZF, AMP, SD, MAP and the data-driven DL detectors versus SNR over the $2\times 2$ correlated channel generated according to~\cite{shiu2000}. We assume that perfect CSI is available at the receiver. Fig~\ref{figure6-1}(b) shows that the BER performance of the data-driven DL detector coincides with that of the MAP detector and substantially outperforms other model-based algorithms, demonstrating its ability to achieve optimal accuracy in complex environments. Fig~\ref{figure5}(b) also indicates that the data-driven DL detector is model independent and manages to learn MAP comparable detection mapping over various channel models. Hence, the data-driven DL detector can keep the performance comparable with the MAP detector in various scenarios.
\subsection{Nonlinear Systems}
In this subsection, we evaluate the BER performance of the data-driven DL detector under a nonlinear MIMO system. We will demonstrate that the data-driven DL detector is applicable to a broader range of scenarios than the traditional MIMO detection algorithms. Furthermore, we only compare the data-driven DL detector to the ZF and MAP detectors.
Consider a MIMO system corrupted by the quantization error of ADC. We assume that each element of the channel output undergoes an entry-wise $B$ bit uniform quantizer $Q_{c}$. The channel input-output model is represented by~\eqref{nonlinear1} and can be rewritten as
\begin{align}
\mathbf{x}=Q_{c}(\mathbf{H}\mathbf{s}+\mathbf{n}).
\end{align}
Each real-valued input of $Q_{c}$ is mapped to one of $2^{B}$ bins, which are defined by the set of $2^{B}-1$ thresholds $[r_{1},r_{2},\ldots,r_{2^{B}-1}]$ such that $-\infty<r-{1}<r_{2}<\cdots<r_{2^{B}-1}<\infty$. Specifically, we define $r_{0}=-\infty$ and $r_{2^{B}}=\infty$. The threshold $r_{b}$ is given by
\begin{align}
r_{b}=\sqrt{B}(-2^{B-1}+b)2^{-B},\,\mathrm{for}\ b=1,\ldots,2^{B}-1,
\end{align}
where the quantization output of $Q_{c}$ is $r_{b}-\frac{\Delta}{2}$ when the input falls in the interval $(r_{b-1},r_{b}]$.\footnote{If $b=2^{B}$, the output of $Q_{c}$ is $\sqrt{B}(-2^{B-1}-2^{-1})2^{-B}$.}
In Fig~\ref{figure5}, we compare the BER performance of the data-driven DL detector with the ZF and MAP detectors versus SNR over the quantized $2\times 2$ time-varying Gaussian channel. The quantization bits are set to be $4$ and $8$, respectively. The perfect CSI is assumed to be available at the receiver. In Fig.~\ref{figure5}, the BER performance of the data-driven DL detector is close to that of the MAP detector under both $5$-bit and $10$-bit quantizers while the BER performance of the ZF detector degrades significantly. Hence, data-driven DL detector is also able to provide MAP comparable BER performance in quantized Gaussian channels. Fig.~\ref{figure5} also verifies model independence of the data-driven DL detector for nonlinear systems.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\columnwidth]{simulation_fig6-2.pdf}
\caption{The BER performance of data-driven DL detector versus SNR compared to other MIMO detectors over the quantized time-varying channel.}
\label{figure5}
\end{figure}
\subsection{Model-driven DL Detector}
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\columnwidth]{simulation_fig6-1.pdf}
\caption{The BER performance of SIC-Net versus SNR compared to other MIMO detectors over the time-varying channel.}
\label{figure6}
\end{figure}
In Fig.~\ref{figure6}, we evaluate the BER performance of the DetNet~\cite{Samuel2019Learning}, OAMP-Net~\cite{hedetec2020}, SIC~\cite{choi2000}, SIC-Net and MAP detectors over the $8\times 8$ time-varying Gaussian channel. We assume that perfect CSI is available at the receiver. Fig.~\ref{figure6} shows that the BER performance of the SIC-Net detector is significantly better than those of the DetNet and OAMP-Net detectors but there still exits a gap between the SIC-Net and the MAP detectors, especially when SNR is high. Hence, the SIC-Net detector is suboptimal compared to the data-driven DL and MAP detectors. In Fig.~\ref{figure6}, the BER performance of the SIC detector is close to that of the SIC-Net detector. Theorem~\ref{theorem3} demonstrates that the SIC detector is equivalent to the optimal model-driven DL detector with the minimum $J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}})$. Therefore, the performance of SIC-Net detector is determined by the SIC algorithm and the selection of detection algorithms is more important than trainable variables in improving the BER performance of the model-driven DL detectors.
\section{Conclusions}\label{conclusion-sec}
In this paper, we have made the first attempt on interpreting DL-based MIMO detection with two different deep architectures: DNN embedded data-driven DL detector and iterative model-driven DL detectors. We have showed that the data-driven DL detector can converge to the MAP detector in various scenarios under suitably configured structure and sufficiently large training set. Specifically, the data-driven DL detector is robust to CSI uncertainty in time-invariant channels and suffers from imperfect CSI in time-varying channels. Moreover, the data-driven DL detector is ineffective in large-sized MIMO systems due to its requirement on a large number of training samples. On the other hand, the model-driven DL detector successfully addresses this problem by exploiting model expert knowledge and achieves relatively good performance with only a small training set since its parameter space is with a small size. However, the model-driven DL detector is suboptimal compared to the MAP detector. The strengths and weaknesses of DL-based MIMO detection should be carefully balanced when deployed in different environments.
\section{Acknowledgement}
Thanks for suggestions and comments from Dr. Shenglong Zhou of Imperial College London.
\begin{appendices}
\section{Proof for Theorem~\ref{theorem1}}\label{apex-a}
From~\eqref{posterior}, we have
\begin{align}\label{theo1-kl}
J(p_{o})-J(p_{\boldsymbol{\theta}})&=\sum_{i=1}^{d_{l+1}}\int\ln\frac{ p_{o}(\tilde{\mathbf{u}}_{i}|\mathbf{x})}{ p_{\boldsymbol{\theta},i}(\mathbf{x})}f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})d\mathbf{x}=\sum_{i=1}^{d_{l+1}}\int g_{\boldsymbol{\theta},i}(\mathbf{x})f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})d\mathbf{x}
\end{align}
for any $\boldsymbol{\theta}\in\Theta_{R}$, where
\begin{align}\label{theo1-p}
p_{\boldsymbol{\theta},i}(\mathbf{x})=\frac{\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))}{\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},j}(\mathbf{x}))},
\end{align}
$f_{\boldsymbol{\theta},i}(\mathbf{x})$ is the $i$-th entry of $\mathbf{f}_{\boldsymbol{\theta}}(\mathbf{x})$, $g_{\boldsymbol{\theta},i}(\mathbf{x})=\ln\frac{ p_{o}(\tilde{\mathbf{u}}_{i}|\mathbf{x})}{ p_{\boldsymbol{\theta},i}(\mathbf{x})}$, and $f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})$ is joint probability density function.
Since $J(p_{o})-J(p_{\boldsymbol{\theta}})\geq 0$, we obtain
\begin{align}\label{theo1-inequal}
J(p_{o})-J(p_{\boldsymbol{\theta}})&=|J(p_{o})-J(p_{\boldsymbol{\theta}})|\leq\sum_{i=1}^{d_{l+1}}\int |g_{\boldsymbol{\theta},i}(\mathbf{x})|f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})d\mathbf{x}.
\end{align}
Substituting~\eqref{theo1-p} into $|g_{\boldsymbol{\theta},i}(\mathbf{x})|$ yields
\begin{align}\label{theo1_epsilon2}
|g_{\boldsymbol{\theta},i}(\mathbf{x})|=\bigg|\ln p_{o}(\tilde{\mathbf{u}}_{i}|\mathbf{x})-f_{\boldsymbol{\theta},i}(\mathbf{x})+\ln\bigg[\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},j}(\mathbf{x}))\bigg]\bigg|.
\end{align}
From~\eqref{network}, $f_{\boldsymbol{\theta},i}(x)$ is an $\mathbb{R}^{d_{0}}\rightarrow\mathbb{R}$ piecewise linear function. According to~\cite[Theorem 2.1]{arora2018understanding}, any $\mathbb{R}^{d_{0}}\rightarrow\mathbb{R}$ piecewise linear function can be represented by a ReLU DNN with no more than $\lceil\log_{2}(d_{0}+1)\rceil$ hidden layers.
Since $\ln p_{o}(\mathbf{u}|\mathbf{x})$ has finite 2-norm for every possible $\mathbf{u}$, each element of $\{\ln p_{o}(\tilde{\mathbf{u}}_{1}|\mathbf{x}),\ldots,\ln p_{o}(\tilde{\mathbf{u}}_{d_{l+1}}|\mathbf{x})\}$ can be approximated by a ReLU DNN with at most $\lceil\log_{2}(d_{0}+1)\rceil$ hidden layers~\cite{arora2018understanding,qiang2020channel}. We simply put these ReLU DNNs in parallel and combine their outputs together to compose a single ReLU DNN. As a result, there exits a DL detector with $\boldsymbol{\theta}_{\varepsilon}\in\Theta_{R}$ and at most
$\lceil\log_{2}(d_{0}+1)\rceil$ hidden layers
such that
\begin{align}\label{theo1_epsilon}
|f_{\boldsymbol{\theta}_\varepsilon,i}(\mathbf{x})-\ln p_{o}(\tilde{\mathbf{u}}_{i}|\mathbf{x})|\leq\varepsilon
\end{align}
for any $\varepsilon>0$ and $i\in\{1,\ldots,d_{l+1}\}$.
From~\eqref{theo1_epsilon} and~\eqref{theo1_epsilon2}, $|g_{\boldsymbol{\theta},i}(\mathbf{x})|$ satisfies
\begin{align}\label{theo1_in}
|g_{\boldsymbol{\theta}_\varepsilon,i}(\mathbf{x})|&\leq\big|\ln p_{o}(\tilde{\mathbf{u}}_{i}|\mathbf{x})-f_{\boldsymbol{\theta}_\varepsilon,i}(\mathbf{x})\big|+\bigg|\ln\bigg[\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta}_\varepsilon,j}(\mathbf{x}))\bigg]\bigg|\leq\varepsilon+\bigg|\ln\bigg[\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta}_\varepsilon,j}(\mathbf{x}))\bigg]\bigg|.
\end{align}
Specifically, $\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta}_\varepsilon,j}(\mathbf{x}))$ in~\eqref{theo1_in} is upper bounded by
\begin{align}
\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta}_\varepsilon,j}(\mathbf{x}))\leq e^{\varepsilon}\sum_{j=1}^{d_{l+1}}p_{o}(\tilde{\mathbf{u}}_{j}|\mathbf{x})=e^{\varepsilon}
\end{align}
and is lower bounded by
\begin{align}
\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta}_\varepsilon,j}(\mathbf{x}))\geq e^{-\varepsilon}.
\end{align}
Then,
\begin{align}
\bigg|\ln
\bigg[\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta}_\varepsilon,j}(\mathbf{x}))\bigg]\bigg|\leq \varepsilon
\end{align}
and
\begin{align}\label{theo1_g}
|g_{\boldsymbol{\theta}_\varepsilon,i}(\mathbf{x})|\leq 2\varepsilon
\end{align}
hold. Substituting~\eqref{theo1_g} into~\eqref{theo1-inequal} yields
\begin{align}
J(p_{o})-J(p_{\boldsymbol{\theta}_\varepsilon})\leq 2\varepsilon\sum_{i=1}^{d_{l+1}}\int f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})d\mathbf{x}=2\varepsilon.
\end{align}
From~\eqref{mini1}, $p_{\boldsymbol{\theta}_{o}}(\mathbf{u}|\mathbf{x})$ has the lowest KL information for all $\boldsymbol{\theta}\in\Theta_{R}$ and we have
\begin{align}\label{theo1-2}
J(p_{o})-J(p_{\boldsymbol{\theta}_{o}})\leq J(p_{o})-J(p_{\boldsymbol{\theta}_\varepsilon})\leq 2\varepsilon.
\end{align}
It is then easy to derive~\eqref{theo1} from~\eqref{theo1-2} since $\varepsilon$ is an arbitrary positive value, which completes the proof.
\section{Proof for Lemma~\ref{lemma-1}}\label{apex-b}
According to Symmetrization Lemma in~\cite[Chapter II.3]{pollard2012convergence}, the inequality in~\eqref{lemma1} holds if
\begin{align}\label{lemma1-1}
\mathbf{P}(|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J(p_{\boldsymbol{\theta}})|>\frac{\varepsilon}{2})\leq\frac{1}{2}
\end{align}
for all $\boldsymbol{\theta}\in\Theta_{R}$. Let $\sigma^{2}(p_{\boldsymbol{\theta}})$ be the variance of $\ln p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})$. Using Chebyshev's inequality~\cite{mendenhall2012introduction} yields
\begin{align}\label{chebshev}
\mathbf{P}\big(|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J(p_{\boldsymbol{\theta}})|\geq\frac{\varepsilon}{2}\big)\leq\frac{4\sigma^{2}(p_{\boldsymbol{\theta}})}{|\mathcal{Z}|\varepsilon^{2}}
\end{align}
for all $\boldsymbol{\theta}\in\Theta_{R}$. Specifically, $\sigma^{2}(p_{\boldsymbol{\theta}})$ satisfies
\begin{align}\label{app-theo4-3}
\sigma^{2}(p_{\boldsymbol{\theta}})&=\mathbb{E}\big\{[\ln p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})]^{2}\big\}-J(p_{\boldsymbol{\theta}})^{2}\leq\mathbb{E}\big\{[\ln p_{\boldsymbol{\theta}}(\mathbf{u}|\mathbf{x})]^{2}\big\}\nonumber\\
&\leq\sum_{i=1}^{d_{l+1}}\int\big[|f_{\boldsymbol{\theta},i}(\mathbf{x})|+\big|\ln(\sum_{j=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},j}(\mathbf{x})))\big|\big]^{2}f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})d\mathbf{x}.
\end{align}
Assume that $\mathcal{X}$ follows the partition in~\eqref{partition} and use the fact~\eqref{network}. The triangle inequality assures that
\begin{align}\label{lemma1-3}
|f_{\boldsymbol{\theta},i}(\mathbf{x})|&=\|\mathbf{w}_{\mathcal{X}_{k},i}\mathbf{x}+b_{\mathcal{X}_{k},i}\|_{2}\leq\|\mathbf{w}_{\mathcal{X}_{k},i}\|_{2}\|\mathbf{x}\|_{2}+|b_{\mathcal{X}_{k},i}|
\end{align}
for $\mathbf{x}\in\mathcal{X}_{k}$, where $\mathbf{w}_{\mathcal{X}_{k},i}$ and $b_{\mathcal{X}_{k},i}$ are the $i$-th row and the $i$-th entry of $\mathbf{W}_{\mathcal{X}_{k}}$ and $\mathbf{b}_{\mathcal{X}_{k}}$, respectively.
From~\eqref{hyperplane} and~\eqref{network}, $\|\mathbf{w}_{\mathcal{X}_{k},i}\|_{2}$ and $\|b_{\mathcal{X}_{k},i}\|_{2}$ in~\eqref{lemma1-3} are upper bounded by
\begin{align}\label{weight}
\|\mathbf{w}_{\mathcal{X}_{k},i}\|_{2}&=\big\|\tilde{\mathbf{w}}_{l,i}\prod_{j=0}^{l-1}\tilde{\mathbf{W}}_{j}\big\|_{2}=\big\|\mathbf{w}_{l,i}\mathbf{\Lambda}_{l}\prod_{j=0}^{l-1}\mathbf{W}_{j}\mathbf{\Lambda}_{j}\big\|_{2}\leq\|\mathbf{w}_{l,i}\mathbf{\Lambda}_{l}\|_{2}\prod_{j=0}^{l-1} \big\|\mathbf{W}_{j}\mathbf{\Lambda}_{j}\big\|_{2}\leq\|\mathbf{w}_{l,i}\|_{2}\prod_{j=0}^{l-1} \big\|\mathbf{W}_{j}\big\|_{2}
\end{align}
and
\begin{align}\label{bias}
|b_{\mathcal{X}_{k},i}|&=\Big|\sum_{j=0}^{l-1}\tilde{\mathbf{w}}_{l,i}\bigg(\prod_{q=0}^{j-1}\tilde{\mathbf{W}}_{l-1-q}\bigg)\mathbf{b}_{l-1-j}+b_{l,i}\Big|\leq\sum_{j=0}^{l-1}\|\tilde{\mathbf{w}}_{l,i}\|_{2}\big\|\prod_{q=0}^{j-1}\tilde{\mathbf{W}}_{l-1-q}\big\|_{2}\|\mathbf{b}_{l-1-j}\|_{2}+|b_{l,i}|\nonumber\\
&\leq\sum_{j=0}^{l-1}\|\mathbf{w}_{l,i}\|_{2}\bigg(\prod_{q=0}^{j-1}\|\mathbf{W}_{l-1-q}\|_{2}\bigg)\|\mathbf{b}_{l-1-j}\|_{2}+|b_{l,i}|,
\end{align}
respectively, where $\tilde{\mathbf{w}}_{l,i}$ and $\mathbf{w}_{l,i}$ are the $i$-th rows of $\tilde{\mathbf{W}}_{l}$ and $\mathbf{W}_{l}$ and $b_{l,i}$ is the $i$-th entry of $\mathbf{b}_{l}$. Since $\|\mathbf{W}_{j}\|_{2}\leq R\|\mathbf{d}\|_{\infty}=\alpha$, $\|\mathbf{b}_{j}\|_{2}\leq\sqrt{R}\|\mathbf{d}\|_{\infty}\leq\alpha$, $\|\mathbf{w}_{l,i}\|_{2}\leq\sqrt{R}\|\mathbf{d}\|_{\infty}\leq\alpha$, and $|b_{l,i}|\leq\|\mathbf{d}\|_{\infty}\leq\alpha$ for $j\in\{0,1,\ldots,l-1\}$,~\eqref{weight} and~\eqref{bias} can be further bounded by
\begin{align}\label{lemma1-weight1}
\|\mathbf{w}_{\mathcal{X}_{k},i}\|_{2}\leq\alpha^{l+1}
\end{align}
and
\begin{align}\label{lemma1-bias1}
&|b_{\mathcal{X}_{k},i}|\leq\bigg(\sum_{i=0}^{l-1}\alpha^{i+1}\bigg)\alpha+\alpha=\frac{\alpha^{l+2}-\alpha}{\alpha-1}\leq\beta(\alpha^{l+1}-1),
\end{align}
respectively. Substituting~\eqref{lemma1-3},~\eqref{lemma1-weight1}, and~\eqref{lemma1-bias1} into~\eqref{lemma1-3} yields
\begin{align}\label{lemma1-4}
|f_{\boldsymbol{\theta},i}(\mathbf{x})|&\leq \alpha^{l+1}(\|\mathbf{x}\|_{2}+\beta)-\beta.
\end{align}
Using~\eqref{lemma1-4}, $\big|\ln[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))]\big|$ in~\eqref{app-theo4-3} is upper bounded by
\begin{align}\label{lemma1-5}
\bigg|\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))\bigg]\bigg|\leq \ln d_{l+1}\big[\alpha^{l+1}(\|\mathbf{x}\|_{2}+\beta)-\beta\big].
\end{align}
Combining~\eqref{lemma1-4} and~\eqref{lemma1-5}, we have
\begin{align}\label{lemma1-2}
\sigma^{2}(p_{\boldsymbol{\theta}})&\leq\sum_{i=1}^{d_{l+1}}\int\big[(\ln d_{l+1}+1)(\alpha^{l+1}(\|\mathbf{x}\|_{2}+\beta)-\beta)\big]^{2}f_{de}(\tilde{\mathbf{u}}_{i},\mathbf{x})d\mathbf{x}
\nonumber\\
&=\mathbb{E}\big\{\big[(\ln d_{l+1}+1)(\alpha^{l+1}(\|\mathbf{x}\|_{2}+\beta)-\beta)\big]^{2}\big\}=\nu.
\end{align}
Replacing $\sigma^{2}(p_{\boldsymbol{\theta}})$ in~\eqref{chebshev} by its bound in~\eqref{lemma1-2} and letting $|\mathcal{Z}|\geq 4\nu/\varepsilon^{2}$, we obtain the inequalities in~\eqref{lemma1} and~\eqref{lemma1-1},
which completes the proof.
\section{Proof for Lemma~\ref{lemma2}}\label{apex-c}
Let $\mathcal{Z}_{i}$ denote the index set of training samples in~$\mathcal{Z}$ where
\begin{align}
\mathbf{u}_{j}=\tilde{\mathbf{u}}_{i},\ \forall j\in\mathcal{Z}_{i}
\end{align}
for $i\in\{1,\ldots,d_{l+1}\}$. Since samples in $\mathcal{Z}_{i}$ are generated according to $\tilde{\mathbf{u}}_{i}$ and $(\tilde{\mathbf{u}}_{1},\ldots,\tilde{\mathbf{u}}_{d_{l+1}})$ is an i.i.d. multinomial random variable, $(|\mathcal{Z}_{1}|,\ldots,|\mathcal{Z}_{d_{l+1}}|)$ is also an i.i.d. multinomial random variable with probability $(p(\tilde{\mathbf{u}}_{1}),\ldots,p(\tilde{\mathbf{u}}_{d_{l+1}}))$ and $\sum_{i=1}^{d_{l+1}}|\mathcal{Z}_{i}|=|\mathcal{Z}|$. Then, $|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{\boldsymbol{\lambda}})|$ is upper bounded by
\begin{align}\label{lemma2-1}
&|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{\boldsymbol{\lambda}})|\leq\frac{1}{|\mathcal{Z}|}\sum_{i=1}^{d_{l+1}}\sum_{j\in\mathcal{Z}_{i}}|\ln p_{\boldsymbol{\theta}}(\mathbf{u}_{j}|\mathbf{x}_{j})-\ln p_{\boldsymbol{\lambda}}(\mathbf{u}_{j}|\mathbf{x}_{j})|\nonumber\\
&\leq\frac{1}{|\mathcal{Z}|}\sum_{i=1}^{d_{l+1}}\sum_{j\in\mathcal{Z}_{i}}\bigg|f_{\boldsymbol{\theta},i}(\mathbf{x}_{j})-f_{\boldsymbol{\lambda},i}(\mathbf{x}_{j})-\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}_{j}))\bigg]+\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\lambda},i}(\mathbf{x}_{j}))\bigg]\bigg|\nonumber\\
&\leq \frac{1}{|\mathcal{Z}|}\sum_{i=1}^{d_{l+1}}\sum_{j\in\mathcal{Z}_{i}}\big|f_{\boldsymbol{\theta},i}(\mathbf{x}_{j})-f_{\boldsymbol{\lambda},i}(\mathbf{x}_{j})\big|+\bigg|\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}_{j}))\bigg]-\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\lambda},i}(\mathbf{x}_{j}))\bigg]\bigg|.
\end{align}
Denote $\mathbf{V}_{i}$ and $\boldsymbol{\upsilon}_{i}$ as the weight and the bias of the $i$-th layer of $\mathcal{D}$ corresponding to $\boldsymbol{\lambda}$ for $i\in\{0,\ldots,l\}$. Let
\begin{align}
\boldsymbol{\theta}_{j}=(\mathrm{vec}(\mathbf{W}_{0}),\mathbf{b}_{0},\ldots,\mathrm{vec}(\mathbf{W}_{j-1}),\mathbf{b}_{j-1})
\end{align}
and
\begin{align}
\boldsymbol{\lambda}_{j}=(\mathrm{vec}(\mathbf{V}_{0}),\boldsymbol{\upsilon}_{0},\ldots,\mathrm{vec}(\mathbf{V}_{j-1}),\boldsymbol{\upsilon}_{j-1}),
\end{align}
be the partial parameters up to the $j$-th layer for $j\in\{1,\ldots,l+1\}$. Hence, the corresponding partial network outputs are denoted by $\mathbf{f}_{\boldsymbol{\theta}_{j}}(\mathbf{x})$ and $\mathbf{f}_{\boldsymbol{\lambda}_{j}}(\mathbf{x})$, respectively.
Let $e_{j}=\|\mathbf{f}_{\boldsymbol{\theta}_{j}}(\mathbf{x})-\mathbf{f}_{\boldsymbol{\lambda}_{j}}(\mathbf{x})\|_{2}$
be the partial error for $j\in\{1,\ldots,l+1\}$ and $e_{l+1}$ represents $\big|f_{\boldsymbol{\theta},i}(\mathbf{x}_{m})-f_{\boldsymbol{\lambda},i}(\mathbf{x}_{m})\big|$ in~\eqref{lemma2-1}. According to~\cite[Lemma~2]{qiang2020channel} and Lemma~\ref{lemma-1}, the upper bound on $e_{j+1}$ is given by
\begin{align}\label{lemma2-2}
e_{j+1}\leq r\|\mathbf{d}\|_{\infty}\big[(3\alpha)^{j}(\|\mathbf{x}\|_{2}+1)+\sum_{q=0}^{j-1}(3\alpha)^{q}(y_{j-q}+1)\big]
\end{align}
for $j\in\{1,\ldots,l\}$, where $r=\|\boldsymbol{\theta}-\boldsymbol{\lambda}\|_{\infty}\leq 2R$ and
$y_{j}=\alpha^{j}(\|\mathbf{x}\|_{2}+\beta)-\beta$ is the upper bound on $\|\mathbf{f}_{\boldsymbol{\theta}_{j}}(\mathbf{x})\|_{2}$ and $\|\mathbf{f}_{\boldsymbol{\lambda}_{j}}(\mathbf{x})\|_{2}$.
From~\eqref{lemma2-2}, we know $|f_{\boldsymbol{\theta},i}(\mathbf{x})-f_{\boldsymbol{\lambda},i}(\mathbf{x})|$ is upper bounded by
\begin{align}\label{lemma2-4}
|f_{\boldsymbol{\theta},i}(\mathbf{x})-f_{\boldsymbol{\lambda},i}(\mathbf{x})|&\leq r\|\mathbf{d}\|_{\infty}\bigg[\sum_{q=0}^{l} 3^{q}\alpha^{l}\|\mathbf{x}\|_{2}+(3\alpha)^{l}+\sum_{q=0}^{l-1}\big[3^{q}\alpha^{l}\beta-(3\alpha)^{q}(\beta-1)\big]\bigg]\nonumber\\
&= r\|\mathbf{d}\|_{\infty}\bigg[ \frac{3^{l+1}-1}{2}\alpha^{l}\|\mathbf{x}\|_{2}+(3\alpha)^{l}+\frac{3^{l}-1}{2}\alpha^{l}\beta -\frac{(3\alpha)^{l}-1}{3\alpha-1}(\beta-1)\bigg]\nonumber\\
&\leq \frac{3}{2}r\|\mathbf{d}\|_{\infty}(3\alpha)^{l}(\|\mathbf{x}\|_{2}+\beta)=\xi r,
\end{align}
where $\xi =\frac{3}{2}\|\mathbf{d}\|_{\infty}(3\alpha)^{l}(\|\mathbf{x}\|_{2}+\beta)$.
From~\eqref{lemma2-4}, the upper bound on $\ln(\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x})))$ can be expressed as
\begin{align}
\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))\bigg]&\leq\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\lambda},i}(\mathbf{x})+\xi r)\bigg]=\xi r+\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\lambda},i}(\mathbf{x})\bigg].
\end{align}
Similarly, $\ln\big[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))\big]$ is lower bounded by
\begin{align}
\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))\bigg]\geq\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\lambda},i}(\mathbf{x}))\bigg]-\xi r.
\end{align}
Therefore, we have
\begin{align}\label{lemma2-6}
\bigg|\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\theta},i}(\mathbf{x}))\bigg]-\ln\bigg[\sum_{i=1}^{d_{l+1}}\exp(f_{\boldsymbol{\lambda},i}(\mathbf{x}))\bigg]\bigg|\leq\xi r.
\end{align}
Using~\eqref{lemma2-4} and~\eqref{lemma2-6}, we derive the upper bound on $|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{\boldsymbol{\lambda}})|$ as
\begin{align}
|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{\boldsymbol{\lambda}})|&\leq \frac{1}{|\mathcal{Z}|}\sum_{m=1}^{|\mathcal{Z}|}3r\|\mathbf{d}\|_{\infty}(3\alpha)^{l}(\|\mathbf{x}_{m}\|_{2}+\beta)\leq 3^{l+1}\|\mathbf{d}\|_{\infty}\alpha^{l}(\delta+\beta)\|\boldsymbol{\theta}-\boldsymbol{\lambda}\|_{\infty},
\end{align}
which completes the proof.
\section{Proof for Theorem~\ref{theorem2}}\label{apex-d}
From~\eqref{mini1} and~\eqref{min2}, we can bound $J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})$ by
\begin{align}\label{theorem2-1}
0&\leq J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})=[J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]-[J_{\mathcal{Z}}(p_{\boldsymbol{\theta}_{o}})-J_{\mathcal{Z}}(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]+[J_{\mathcal{Z}}(p_{\boldsymbol{\theta}_{o}})-J_{\mathcal{Z}}(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]\nonumber\\
&\leq [J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]-[J_{\mathcal{Z}}(p_{\boldsymbol{\theta}_{o}})-J_{\mathcal{Z}}(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]\leq 2\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J(p_{\boldsymbol{\theta}})|.
\end{align}
According to~\eqref{theorem2-1} and Lemma~\ref{lemma-1}, we know the following inequalities
\begin{align}\label{theorem2-7}
\mathbf{P}([J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]>\varepsilon)&\leq \mathbf{P}\big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J(p_{\boldsymbol{\theta}})|>\frac{\varepsilon}{2}\big)\leq 4\mathbf{P}\big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(p_{\boldsymbol{\theta}})|>\frac{\varepsilon}{8}\big),
\end{align}
holds if $|\mathcal{Z}|\geq 16\nu/\varepsilon^{2}$.
Assume that $\mathcal{Z}$ is fixed with $\frac{1}{|\mathcal{Z}|}\sum_{m=1}^{|\mathcal{Z}|}\|\mathbf{x}_{m}\|_{2}^{2}\leq\delta^{2}$. Let $C=C(\varepsilon/16,\Theta_{R})$ and choose a collection of functions $\mathbf{p}_{1}(\mathbf{x}),\ldots,\mathbf{p}_{C}(\mathbf{x})\in\mathcal{D}$ such that
\begin{align}\label{theorem2-2}
|J_{\mathcal{Z}}(p_{\boldsymbol{\theta}})-J_{\mathcal{Z}}(p_{j})|\leq\frac{\varepsilon}{16},\ \forall j\in\{1,\ldots,C\},
\end{align}
for any $p_{\boldsymbol{\theta}}(\mathbf{x})\in\mathcal{D}$ and $\boldsymbol{\theta}\in\Theta_{R}$. According to~\cite[Theorem~2]{qiang2020channel}, the following inequality
\begin{align}\label{theorem2-3}
\mathbf{P}\Big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(\mathbf{f}_{\boldsymbol{\theta}})|>\frac{\varepsilon}{8}|\mathcal{Z}\Big)\leq\sum_{j=1}^{C} 2\mathrm{exp}\bigg[-2(\frac{|\mathcal{Z}|}{16}\varepsilon)^{2}/\sum_{m=1}^{|\mathcal{Z}|}(2\ln p_{j}(\mathbf{u}_{m}|\mathbf{x}_{m}))^{2}\bigg]
\end{align}
holds. From~\eqref{app-theo4-3} and~\eqref{lemma1-2}, we know $\sum_{m=1}^{|\mathcal{Z}|}(\ln p(\mathbf{u}_{m}|\mathbf{x}_{m}))^{2}$ is upper bounded by
\begin{align}\label{theorem2-5}
\sum_{m=1}^{|\mathcal{Z}|}(\ln p(\mathbf{u}_{m}|\mathbf{x}_{m}))^{2}&\leq \sum_{m=1}^{|\mathcal{Z}|}\big[(\ln d_{l+1}+1)(\alpha^{l+1}(\|\mathbf{x}_{m}\|_{2}+\beta)-\beta)\big]^{2} \nonumber\\
&\leq |\mathcal{Z}|\big[(\ln d_{l+1}+1)(\alpha^{l+1}\delta+\beta)-\beta)\big]^{2}=|\mathcal{Z}|\delta_{\mathrm{1}}.
\end{align}
Replacing $\sum_{m=1}^{|\mathcal{Z}|}(\ln p(\mathbf{u}_{m}|\mathbf{x}_{m}))^{2}$ in~\eqref{theorem2-3} by the upper bound in~\eqref{theorem2-5} yields
\begin{align}\label{theorem2-6}
\mathbf{P}\Big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(\mathbf{f}_{\boldsymbol{\theta}})|>\frac{\varepsilon}{8}|\mathcal{Z}\Big)\leq 2\mathrm{exp}\Big(\ln C-\frac{|\mathcal{Z}|\varepsilon^{2}}{512\delta_{\mathrm{1}}}\Big).
\end{align}
According to~\eqref{covering}, $\ln C$ is upper bounded by
\begin{align}
&\ln C =\ln C(\varepsilon/16,\Theta_{R})\leq d_{s}\ln\Big[\frac{3^{l+1}2^{6}\alpha^{l}(\delta+\beta)}{\varepsilon}\Big]\leq d_{s}\ln\frac{\delta_{2}}{\varepsilon}.
\end{align}
If $|\mathcal{Z}|\geq (1024\delta_{1}d_{s}\ln\frac{\delta_{2}}{\varepsilon})/\varepsilon^{2}$,
then we have $
\ln C\leq |\mathcal{Z}|\varepsilon^{2}/(1024\delta_{1})$
and
\begin{align}\label{theorem2-11}
\mathbf{P}\Big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(p_{\boldsymbol{\theta}})|>\frac{\varepsilon}{8}|\mathcal{Z}\Big)\leq 2\mathrm{exp}\Big(-\frac{|\mathcal{Z}|\varepsilon^{2}}{1024\delta_{\mathrm{1}}}\Big).
\end{align}
Integrating out $\mathbf{P}(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(p_{\boldsymbol{\theta}})|>\frac{\varepsilon}{8}|\mathcal{Z})$ over $\mathcal{Z}$ in~\eqref{theorem2-11} produces
\begin{align}\label{theorem2-8}
&\mathbf{P}\Big(\sup_{\boldsymbol{\theta}\in\Theta_{R}}|J_{\mathcal{Z}}^{\circ}(\mathbf{f}_{\boldsymbol{\theta}})|>\frac{\varepsilon}{8}\Big)\leq 2\mathrm{exp}
\Big(-\frac{|\mathcal{Z}|\varepsilon^{2}}{1024\delta_{\mathrm{1}}}\Big)+\mathbf{P}_{\mathcal{Z}},
\end{align}
where
\begin{align}
\mathbf{P}_{\mathcal{Z}}=\mathbf{P}\Big(\frac{1}{|\mathcal{Z}|}\sum_{m=1}^{|\mathcal{Z}|}\|\mathbf{x}_{m}\|_{2}^{2}\geq\delta^{2}\Big).
\end{align}
If $\delta^{2}>\mu$, then Chebyshev's inequality~\cite{mendenhall2012introduction} assures that
\begin{align} \mathbf{P}_{\mathcal{Z}}&\leq\mathbf{P}\big[|\frac{1}{|\mathcal{Z}|}\sum_{m=1}^{|\mathcal{Z}|}\|\mathbf{x}_{m}\|_{2}^{2}-\mu|\geq(\delta^{2}-\mu)\big]\leq\frac{\sigma^{2}}{|\mathcal{Z}|(\delta^{2}-\mu)^{2}}.
\end{align}
Combining~\eqref{theorem2-7} and~\eqref{theorem2-8}, we obtain
\begin{align}\label{theorem2-9}
\mathbf{P}([J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]>\varepsilon)\leq 8\mathrm{exp}\Big(-\frac{|\mathcal{Z}|\varepsilon^{2}}{1024\delta_{\mathrm{1}}}\Big)+\frac{4\sigma^{2}}{|\mathcal{Z}|(\delta^{2}-\mu)^{2}},
\end{align}
which completes the proof.
\section{Proof for Corollary~\ref{corollary}}\label{apex-e}
According to~\eqref{decom}, $D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}_{\mathcal{Z}}})$ is decomposed into the approximation and generalization errors.
Theorem~\ref{theorem1} demonstrates that there exits an optimized data-driven DL estimator $\mathbf{p}_{\boldsymbol{\theta}_{o}}(\mathbf{x})$ with
at most $
\lceil\log_{2}(d_{0}+1)\rceil
$ hidden layers and sufficiently large $R$ such that the approximation error $
J(p)-J(p_{\boldsymbol{\theta}_{o}})\leq\varepsilon$
for any $\varepsilon>0$.
Moreover, Theorem~\ref{theorem2} implies that the generalization error $J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})$ satisfies
\begin{align}
\lim_{|\mathcal{Z}|\rightarrow+\infty}\mathbf{P}\big([J(p_{\boldsymbol{\theta}_{o}})-J(p_{\boldsymbol{\theta}_{\mathcal{Z}}})]>\varepsilon\big)=0
\end{align}
for any $\varepsilon>0$. Combining Theorems~\ref{theorem1} and Theorem~\ref{theorem2}, we have
\begin{align}
\lim_{|\mathcal{Z}|\rightarrow+\infty}\mathbf{P}\big(D_{\mathrm{KL}}(p_{o},p_{\boldsymbol{\theta}_{\mathcal{Z}}})>\varepsilon\big)=0
\end{align}
for any $\varepsilon>0$, which completes the proof.
\section{Proof for Theorem~\ref{theorem3}}\label{apex-f}
Assume that $\Omega$ is fixed with $\sum_{m=1}^{|\Omega|}[J(\mathbf{f}_{\boldsymbol{\vartheta}}(\mathbf{x}_{m},\mathbf{H}_{m}))]^{2}\geq|\Omega|\delta_{u}$ for $\delta_{u}>0$, and then choose a collection of functions $\mathbf{f}_{1}(\mathbf{x},\mathbf{H}),\ldots,\mathbf{f}_{C_{u}}(\mathbf{x},\mathbf{H})\in\mathcal{D}_{u}$ such that $C_{u}=C_{u}(\varepsilon/16,\Theta_{u})$.
According to Hoeffding's Inequality~\cite{mendenhall2012introduction} and Theorem~\ref{theorem2}, we have
\begin{align}
\mathbf{P}_{u}([J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_\Omega})-J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_o})]>\varepsilon
|\,\Omega)&\leq\sum_{j=1}^{C_{u}}8\exp\bigg[-2(\frac{|\Omega|}{16}\varepsilon)^{2}/\sum_{m=1}^{|\Omega|}(2J(\mathbf{f}_{j}(\mathbf{x}_{m},\mathbf{H}_{m})))^{2}\bigg]\nonumber\\
&\leq8\exp\bigg(\ln C_{u}-\frac{|\Omega|\varepsilon^{2}}{512\delta_{u}}\bigg).
\end{align}
Integrating out $\mathbf{P}_{u}([J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_\Omega})-J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_o})]>\varepsilon
|\,\Omega)$ over $\Omega$ yields
\begin{align}
\mathbf{P}_{u}\big([J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_\Omega})-J_{u}(\mathbf{f}_{\boldsymbol{\vartheta}_o})]>\varepsilon
\big)\leq8\exp\left(\ln C_{u}-\frac{|\Omega|\varepsilon^{2}}{512\delta_{u}}\right)+\mathbf{P}_{\Omega},
\end{align}
which completes the proof.
\end{appendices}
\linespread{1.083}
|
\section{Introduction}
Traversable wormholes are spacetime geometries emerging as solutions of
Einstein equations. Conventionally they are thought as fascinating
configurations that might have formed in remote regions of the Universe.
Nowadays such a perspective has changed: The understanding of the physics
governing wormholes is instrumental for the concrete realization of laboratory
devices for interstellar travels \cite{White03}.
The current efforts in such a research field are focused on the
conditions of traversability of wormholes. In practice, the wormhole throat
can be stable only if there is a source of energy able to balance the
gravitational pull. Contrary to the case of ordinary stars, such a balance implies the source to be of
\textit{exotic} type. The latter is a term used to indicate a non standard
matter that violates the null energy conditions (NEC), namely the positiveness
of the energy momentum tensor, $T_{\mu\nu}k^{\mu}k^{\nu}\geq0$ for any null
vector $k^{\mu}$. Wormholes necessitate such a violation of NEC since, according to the Raychaudhuri equation \cite{Ray55}, the expansion of timelike congruence becomes negative at the throat while remains positive elsewhere. In other words, a set of world lines undergoes a contraction at the throat, indicating an inevitable collapse into a singularity unless exotic matter locally counteracts it. Although the pressure of ordinary matter can, in general, counteract a collapse, it would not be enough high to reestablish the positiveness of the congruence expansion on the other side of the throat \cite{MoT88}. As far as we know, the Casimir energy represents the only
artificial source of exotic matter realizable in a laboratory \cite{White03,Garattini:1999mh,Garattini:2000kd,Garattini19,Garattini20}. Alternatively, a local violation of energy conditions can occur due to the quantum fluctuation of the graviton. This fact propelled
the idea of \textit{self-sustained traversable wormholes}, that have been introduced
in \cite{Garattini05,Garattini07c,Garattini:2010dn}. To study such wormholes one considers the semiclassical Einstein equations,
\begin{equation}
G_{\mu\nu}=\kappa \left<T_{\mu\nu}\right>^\mathrm{ren}, \quad \left<T_{\mu\nu}\right>^\mathrm{ren}=-\frac{1}{\kappa}\left<\Delta G_{\mu\nu}\left(\bar{g}_{\mu\nu}, h_{\mu\nu}\right)\right>^\mathrm{ren}
\label{eq:semclEE}
\end{equation}
where the source term is the expectation value of the renormalized quantum stress tensor of the metric perturbation $h_{\mu\nu}$, with $g_{\mu\nu}=\bar{g}_{\mu\nu}+h_{\mu\nu}$. Eq. \eqref{eq:semclEE} simplifies by focusing on the energy components, namely by a projection on a constant time spacelike hypersurface $\Sigma$. In such a way, one obtains Hamiltonian and energy densities, that, after integration, give the equation for the stability of the wormhole:%
\begin{equation}
H_{\Sigma}^{(0)}=-E^{\bot}.\label{SS}%
\end{equation}
Here $E^{\bot}$ is the total regularized graviton one loop energy coming from the quantized stress tensor and
$H_{\Sigma}^{(0)}$ is the classical term, coming from the Einstein tensor. Only the transverse traceless (TT)
component of the graviton contribute to the energy $E^{\bot}$ and for this
reason we introduced the superscript $^{\bot}$.
For a
spherically symmetric line element of the form%
\begin{equation}
ds^{2}=-\exp\left( -2\Phi\left( r\right) \right) dt^{2}+\frac{dr^{2}%
}{1-\frac{b\left( r\right) }{r}}+r^{2}d\Omega^{2},\label{metric}%
\end{equation}
where $\Phi\left( r\right) $ is the redshift function, $b\left( r\right) $
is the shape function and $d\Omega^{2}=d\theta^{2}+\sin^{2}\theta d\phi^{2}$
is the line element of the unit sphere, the classical term reduces
to%
\begin{align}
H_{\Sigma}^{(0)} & =\int_{\Sigma}\,d^{3}x\left[ \,\left( 16\pi G\right)
G_{ijkm}\pi^{ij}\pi^{km}-\frac{\sqrt{g}}{16\pi G}\!{}\!\,\ ^3R\right]
\nonumber\\
& =-\frac{1}{16\pi G}\int_{\Sigma}\,d^{3}x\,\sqrt{g}\,\ ^3R\,=-\frac{1}{2G}%
\int_{r_{0}}^{\infty}\,\frac{dr\,r^{2}}{\sqrt{1-b(r)/r}}\,\frac{b^{\prime}%
(r)}{r^{2}}\,.
\end{align}
Here we have expressed the three dimensional scalar
curvature $^3R$ in terms of $b(r)$. The symbol
$G_{ijkm}$ denotes the super-metric and $\pi^{ij}$ the super-momentum. Due to static conditions the kinetic term $G_{ijkm}\pi^{ij}\pi^{km}$ disappears. For a full derivation see \cite{GaL17}.
At the one-loop
approximation level, the energy $E^{\bot}$ is identified as a Casimir like energy
against the fixed background. Its evaluation requires functional integration methods including the solution of
an appropriate eigenvalue
equation in terms of a modified Lichnerowicz operator (see \cite{GaL17} for the full derivation).
The conventional Lichnerowicz operator can be defined through its
action on a tensor $h_{ij}$ as
\begin{align}
& \left( \bigtriangleup_{L}h\right) _{ij}=\bigtriangleup h_{ij}%
-2R_{ikjm}h^{km}+R_{ik}h_{j}^{k}+R_{jk}h_{i}^{k}\nonumber\\
& \bigtriangleup=-\nabla^{a}\nabla_{a},\label{DeltaL}%
\end{align}
where Latin indexes run from $1$ to $3$ and $\nabla^a$ is the covariant derivative with respect to the $3$-metric, $\bar{g}_{ij}$. For ease of notation we employ, in the following equations, the symbol $g_{ij}$ for the background
without superscript $\bar{{}}$.
As said above, we have to
consider the TT
component of the field $h_{ij}$ describing a spin 2 particle, namely
\begin{equation}
g^{ij}h_{ij}^{\bot}=0,\qquad\nabla^{i}h_{ij}^{\bot}=0.
\end{equation}
Thus, the problem turns into the
determination of the eigenvalues of the following modified Lichnerowicz
operator,
\begin{equation}
\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}=\left(
\bigtriangleup_{L\!}\!{}h^{\bot}\right) _{ij}-4R{}_{i}^{k}\!{}\ h_{kj}^{\bot
}+\text{ }^{3}R{}\!{}\ h_{ij}^{\bot}.\label{M Lichn}%
\end{equation}
namely
\begin{equation}
\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}=\lambda
\!{}h_{ij}^{\bot},\label{Eq}%
\end{equation}
describing the energy spectrum of $h_{ij}^{\bot}$ resulting from \eqref{eq:semclEE}.
To have a well-posed equation in \eqref{Eq}, one has, however, to make sure that the
above operator does not change the TT properties of $h_{ij}^{\bot}$. This is, in general, a major issue because the l.h.s. is not a TT tensor for some kind of backgrounds \cite{Delay07}.
The present paper aims to circumvent such difficulties and pave the way of a consistent study of wormhole
stability within this formalism.
\section{The Lichnerowicz operator for TT tensors}
\label{appT}
It is straightforward to see that the standard Lichnerowicz operator, $\left(
\bigtriangleup_{L}h^{\bot}\right) _{ij}$, is traceless. This is, however, not
enough to conclude that $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$ is traceless too.
Indeed, one has
\begin{equation}
\mathrm{Tr}\left[ \left( \bigtriangleup_{L}h^{\bot}\right) _{ij}%
-4R_{ik}h_{j}^{{\bot}\ k}+\text{ }^{3}Rh_{ij}^{\bot}\right] =-4R_{k}^{m}%
h_{m}^{{\bot}\ k}.\label{Tr}%
\end{equation}
As a first step, we aim to write such an operator in terms of a trace free part
and a term determining the trace. For the line
element $\left( \ref{metric}\right) $, the mixed Ricci tensor $R_{j\text{ }%
}^{i}$ is:
\begin{equation}
R_{j}^{i}=\left\{ \frac{b^{\prime}\left( r\right) }{r^{2}}-\frac{b\left(
r\right) }{r^{3}},\frac{b^{\prime}\left( r\right) }{2r^{2}}+\frac{b\left(
r\right) }{2r^{3}},\frac{b^{\prime}\left( r\right) }{2r^{2}}+\frac{b\left(
r\right) }{2r^{3}}\right\} .\label{eq:riccimixed}%
\end{equation}
Therefore, we can write
\[
R_{k}^{i}h_{j}^{k\ \bot}=\underset{\mathrm{traceless}}{\underbrace{f\left(
r\right) h_{i}^{\bot\ j}}}+\left( R_{1}^{1}-f\left( r\right) \right)
\delta_{j}^{1}\delta_{1}^{i}h_{i}^{\bot\ j}%
\]
where we have used the properties $R_{2}^{2}=R_{3}^{3}\equiv f\left(
r\right) $ and $h_{1}^{\bot\ 1}+h_{2}^{\bot\ 2}+h_{3}^{\bot\ 3}=0$. As a
result, one finds
\begin{equation}
\left( \tilde{\bigtriangleup}_{L\!}\!{}\ h^{\bot}\right) _{i}^{j}=\left(
\tilde{\bigtriangleup}_{L\!}\!{}\ h^{\bot}\right) _{i}^{j\ \mathrm{T}}%
+
\left\{
\frac{1}{2}\left[ \frac{3b\left( r\right) }{r^{3}}-\frac{b^{\prime}\left(
r\right) }{r^{2}}
\right]
\delta_{j}^{1}\delta_{1}^{i}
\right\}
h_{i}^{\bot
\ j}\label{eq:trace}%
\end{equation}
with the trace free part defined as
\begin{equation}
\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{i}^{j\ \mathrm{T}%
}\equiv\left( \bigtriangleup_{L}h^{\bot}\right) _{i}^{j}-4f\left( r\right)
h_{i}^{\bot\ j}+\text{ }^{3}Rh_{i}^{\bot\ j}.
\label{eq:tracefreepart}
\end{equation}
Now we compute the divergence of the above trace free part,
namely \cite{Delay07}
\begin{align}
& \nabla^{m}\left( \tilde{\bigtriangleup}_{L\!}\!{}\ h^{\bot}\right)
_{m}^{j\ \mathrm{T}}=\nabla^{m}\left( \bigtriangleup_{L}h^{\bot}\right)
_{m}^{j}-\nabla^{m}\left( 4f\left( r\right) h_{m}^{{\bot}\,j}-\text{ }%
^{3}Rh_{m}^{{\bot}\,j}\right) \nonumber\\
& =\bigtriangleup\left( \nabla^{m}h_{m}^{{\bot}\,j}\right) +R_{k}^{j}\left(
\nabla^{m}h_{m}^{{\bot}\,k}\right) +\left( \nabla^{j}R_{m}^{k}\right)
h_{k}^{m\,{\bot}}-\nabla^{m}\left( 4f\left( r\right) h_{m}^{{\bot}%
\,j}-\text{ }^{3}Rh_{m}^{{\bot}\,j}\right) \nonumber\\
& =\left( \nabla^{j}R_{m}^{k}\right) h_{k}^{\bot m}-4\left( \nabla
^{m}f\left( r\right) \right) h_{m}^{\bot\,j}+\left( \nabla^{m}\text{ }%
^{3}R\right) h_{m}^{\bot\,j}.\label{eq:divtracefree}%
\end{align}
where we have used the transverse property $\nabla^{m}h_{m}^{\bot\,j}=0$ and
$$\nabla^{m}\left(\bigtriangleup_{L\!}\!{}\ h^{\bot}\right)_{m}^{j\ \mathrm{T}}=\bigtriangleup\left( \nabla^{m}h_{m}^{{\bot}\,j}\right) +R_{k}^{j}\left(
\nabla^{m}h_{m}^{{\bot}\,k}\right) +\left( \nabla^{j}R_{m}^{k}\right)
h_{k}^{m\,{\bot}}.$$ In
summary one finds:%
\begin{equation}
\nabla^{m}\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right)
_{m}^{j\ \mathrm{T}}=\left\{-\frac{1}{2}\ g^{11}\partial_{1}\left[ \frac{7b\left(
r\right) }{r^{3}}-\frac{b^{\prime}\left( r\right) }{r^{2}}\right]
\delta_{j}^{1}\delta_{1}^{m}\right\} h_{m}^{j\,\bot}.\label{eq:tracefreeLop}%
\end{equation}
The r.h.s. of the above equation vanishes provided
\[
\frac{7b\left( r\right) }{r^{3}}-\frac{b^{\prime}\left( r\right) }{r^{2}%
}=\mathrm{constant}.
\]
Thus, one obtains
\begin{equation}
b(r)=Ar^{3}
\label{eq:br}
\end{equation}
where $A$ is a constant coefficient. Accordingly, one has $A>0$ for de Sitter
space, $A<0$ for Anti-de Sitter space and $A=0$ for Minkowski space. In such
cases, one finds that the trace in \eqref{eq:trace} vanishes. As a result, one
can conclude, from the condition \eqref{eq:br}, that the
operator, $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$, is a TT tensor in case of constant or vanishing curvature.
In the presence of a gravitational source, the curvature is, in general, not
constant. The operator $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$ can, however, satisfy the
transverse condition up to negligible terms if the curvature variation is
small. The reference scale in such a case is the Planck mass cubed,
$M_{\mathrm{P}}^{3}$. The vanishing of the trace in \eqref{eq:trace} requires
the curvature itself to be small with respect to $M_{\mathrm{P}}^{2}$, namely
\begin{align}
\label{eq:curvcond}\frac{b}{r}\ll\frac{r^{2}}{L_{\mathrm{P}}^{2}},\quad
b^{\prime}\ll\frac{r^{2}}{L_{\mathrm{P}}^{2}},
\end{align}
where $L_{\mathrm{P}}=1/M_{\mathrm{P}}$. As a result, the trace freedom is a
stronger condition with respect to the small variation of the curvature. Both
conditions are easily met in the large distance limit. It is sufficient to
assume $r\gg L_{\mathrm{P}}$ for having the TT condition fulfilled, provided
$b^{\prime}$ is bounded.
At short distance, namely at the wormhole throat, the r.h.s. of
\eqref{eq:tracefreeLop} vanishes because of the presence of the metric
coefficient $g^{11}$. The conditions \eqref{eq:curvcond} are fulfilled for a
wormhole throat $r_\mathrm{t}$ such that $r_\mathrm{t}\gg L_{\mathrm{P}}$,
being $b^\prime(r_\mathrm{t})<1$ according to the flaring out condition.
Up to now, the presented analysis has been focused on differential
conditions.
To calculate the sought eigenvalues, one has, however,
to consider matrix elements, expressed in terms of the following integral%
\begin{equation}
\int_{\Sigma}d^{3}x\sqrt{g}\ h_{j}^{\bot\,i}\left( \tilde{\bigtriangleup}%
_{L}h^{\bot}\right) _{i}^{j},\label{eq:matrixel}%
\end{equation}
resulting from the one loop Hamiltonian at the r.h.s of \eqref{eq:semclEE},
\begin{align}
& H_{\Sigma}^{\bot}=\frac{1}{4V}%
\int_{\Sigma}d^{3}x\sqrt{\bar{g}}\ G^{ijkm}\left[ \left( 2\kappa\right)
K^{-1\bot}\left( x,x\right) _{ijkm
+\frac{1}{\left( 2\kappa\right) }\!{}\left( \tilde
{\bigtriangleup}_{L\!}\right) _{j}^{a}K^{\bot}\left( x,x\right)
_{iakm}\right], \label{p22}%
\end{align}
where
\begin{equation}
K^{\bot}\left( \overrightarrow{x},\overrightarrow{y}\right) _{iakm}%
=\sum_{\tau}\frac{h_{ia}^{\left( \tau\right) \bot}\left( \overrightarrow
{x}\right) h_{km}^{\left( \tau\right) \bot}\left( \overrightarrow
{y}\right) }{2\ell\left( \tau\right) },\label{proptt}%
\end{equation}
is the graviton propagator, $\ell(\tau)$ a set of variational parameters to be determined by minimizing \eqref{p22} and $\tau$ denotes a complete set of indexes (see \cite{GaL17}).
One can try to circumvent the problem of the TT nature of the operator $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$, since integral relations generally demand
softer conditions than differential relations. To this purpose we observe that
\begin{align}
\int_{\Sigma}d^{3}x\sqrt{g}\ h_{j}^{\bot\,i}\left( \tilde{\bigtriangleup}%
_{L}h^{\bot}\right) _{i}^{j} & =\int_{\Sigma}d^{3}x\sqrt{g}\ h_{j}^{\bot
\,i}\left[ \left( \tilde{\bigtriangleup}_{L}h^{\bot}\right) _{i}^{j}%
+\frac{4}{3}\ \delta_{i}^{\ j}\left( R_{m}^{\ k}\ h_{k}^{\bot\ m}\right)
\right] =\nonumber\\
& =\int_{\Sigma}d^{3}x\sqrt{g}\ h_{j}^{\bot\,i}\left( \tilde{\bigtriangleup
}_{L\!}\!{}\ h^{\bot}\right) _{i}^{j\ \mathrm{T}}.\label{Ip}%
\end{align}
The integral of the term $R_{m}^{\ k}\ h_{k}^{\bot\ m}$ at the r.h.s.
identically vanishes, being $\mathrm{Tr}\ h_{k}^{\bot\ m}=h_{k}^{\bot\ k}=0 $.
Due to \eqref{Tr} and the generic relation
\begin{equation}
\left( T_{ij}\right) ^{\mathrm{T}}=T_{ij}-\frac{1}{3}g_{ij}\left[
\mathrm{Tr}\ T_{km}\right] .
\end{equation}
we obtained the trace free part. The
operator $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$ is not traceless, but its integral is equivalent to that of its trace
free part. This property is instrumental to prove that, at level of integral relations, the transverse property is
satisfied too. After gauging the trace away by means of \eqref{Ip}, the transverse property can be analyzed by considering just the integral of the trace free part \eqref{eq:tracefreepart}. The latter can be written as
\begin{equation}
\left( \tilde
{\bigtriangleup}_{L}h^{\bot}\right) _{ij}^{\mathrm{T}}\equiv \left( \bigtriangleup
_{L}h^{\bot}\right) _{ij}-4\left( R_{ik}h_{j}^{{\bot}\ k}-\frac{1}%
{3}\ g_{ij}\left( R_{m}^{\ k}h_{k}^{\bot\ m}\right) \right) +\text{ }%
^{3}Rh_{ij}^{\bot}.
\end{equation}
We can further gauge the integral of $\left( \tilde{\bigtriangleup}%
_{L}h^{\bot}\right) _{i}^{j\ \mathrm{T}}$ by adding a vanishing contribution whose integrand is traceless in order not to alter the trace free property, namely
\begin{equation}
\label{eq:inttransversegauge}
\int_{\Sigma}d^{3}x\sqrt{g}\ h_{j}^{\bot\,i}\left( \tilde{\bigtriangleup}%
_{L}h^{\bot}\right) _{i}^{j\ \mathrm{T}}=\int_{\Sigma}d^{3}x\sqrt{g}%
\ h_{j}^{\bot\,i}\left[ \left( \tilde{\bigtriangleup}_{L}h^{\bot}\right)
_{i}^{j\ \mathrm{T}}+\left( LM\right) _{i}^{j}\right] ,
\end{equation}
where
\begin{equation}
\left( LM\right) _{ij}=\nabla_{i}M_{j}+\nabla_{j}M_{i}-\frac{2}{3}%
g_{ij}\left( \nabla_k M^k\right).
\label{eq:LM}
\end{equation}
One can check that the integral of $\left( LM\right) _{ij}$ is vanishing by
integrating by parts the first two terms $\nabla_{i}M_{j}$ and using the
transverse condition $\nabla_{i}h_{j}^{\bot\,i}=0$. The third term vanishes
because $h_{j}^{\bot\,i}$ is trace free.
At this point, one can suitably select $\left( LM\right) _{ij}$
to get the tranverse condition, namely%
\begin{eqnarray}
&& 0 =\nabla^{i}\left\{ \left[ \left( \bigtriangleup_{L}h^{\bot
}\right) _{ij}-4\left( R_{im}h_{j}^{{\bot}\ m}-\frac{1}{3}\ g_{ij}\left(
R_{i}^{\ m}\ h_{m}^{\bot\ i}\right) \right) +\text{ }^{3}Rh_{ij}^{\bot
}\right]+\left( L M\right) _{ij}\right\} =\nonumber\\
&& =\left( \nabla_{j}R_{i}^{m}\right) h_{m}^{\bot i} - \nabla^{i}\left[4 R_{im}h_{j}^{{\bot}\ m}-\frac{4}{3}\ g_{ij}\left(
R_{i}^{\ m}\ h_{m}^{\bot\ i}\right)-\text{ }^{3}Rh_{ij}^{\bot
}-\left( L M\right) _{ij}\right] =\nonumber
\\
&&=\left( \nabla_{j}R_{i}^{m}\right) h_{m}^{\bot i}-4\left( \nabla
^{i}f\left( r\right) \right) h_{ij}^{\bot}+\left( \nabla^{i}\text{ }%
^{3}R\right) h_{ij}^{\bot}+\nabla^{i}\left( LM\right) _{ij}.\label{Mel}%
\end{eqnarray}
provided \eqref{Mel} has solutions.
Alternatively one can consider, in place of $(LM)_{ij}$ in \eqref{eq:inttransversegauge}, an antisymmetric term of the kind
\begin{equation}
\left( L N\right) _{ij}=\nabla_{i}N_{j}-\nabla_{j}N_{i},
\end{equation}
that is trace free and has a vanishing integral. Its covariant derivative is
formally equivalent to the four current of the electromagnetic field tensor,
namely
\begin{equation}
\nabla^{i}\left( L N\right) _{i}^{j}=S^{j}.
\end{equation}
In such a case, one can select $S^{j}$ provided the equivalent of \eqref{Mel} for $(LN)_{ij}$ has solutions.
In conclusion,
even if the operator is not a TT tensor
for arbitrary backgrounds, its integral \eqref{eq:matrixel} is
equivalent to the integral of an operator \eqref{eq:inttransversegauge} that display TT properties.
\section{Final remarks}
In this paper we have presented a solution to an open issue in the literature,
namely the calculation of graviton energies at the one-loop
approximation associated to a Lichnerowicz operator.
In case of
spherically symmetric spacetimes, such energies
come from the TT component of the perturbation, namely
\begin{equation}
E^{\bot}=\frac{1}{3}\sum_\tau\left[\sqrt{\lambda_1^2(\tau)}+\sqrt{\lambda_2^2(\tau)}\right]
\end{equation}
where the eingenvalues correspond to the two graviton polarizations.
Unfortunately, the operator $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$ is not, in general, a TT tensor,
a fact that deprives the
formalism of its predictive power, apart from the case of specific spacetimes
where the eigenvalue equations can be solved.
Against this background,
we have shown that $E^{\bot}$ can be calculated in terms of another
operator that exhibits TT properties. Such an operator is obtained by a
suitable ``gauge'' of the original operator $\left( \tilde{\bigtriangleup}_{L\!}\!{}h^{\bot}\right) _{ij}$ in the integral relations \eqref{eq:matrixel} and \eqref{p22}.
The proposed results can pave the way to further studies based on the
Lichnerowicz operator to scrutinize the conditions of stability of traversable
wormholes spacetimes.
\subsection*{Acknowledgements}
This work has been supported by the project ``Traversable Wormholes: A Road To Interstellar Exploration,'' an Interstellar Initiatives Grant award funded by the Limitless Space Institute.
The work of P.N. has partially
been supported by GNFM, the Italian National Group for Mathematical Physics.
|
\section{Introduction}
Magnetic resonance techniques, such as nuclear magnetic resonance (NMR) and electron spin resonance (ESR), provide exquisite information about local chemical environments.
NMR spectroscopy is routinely used in chemical synthesis for structural analysis of small molecules.
The nitrogen-vacancy (NV) center is an excellent candidate for nanoscale magnetic resonance as its spin state can be optically initialized and readout, has demonstrated long coherence times, and is highly sensitive to external magnetic fields.~\cite{Gruber1997, Jelezko2004,Childress2006, Takahashi2008, Bauch2018, Wolf2015, Schmitt2017}
The NV center's unique properties have enabled both nanoscale NMR and ESR with sensitivity down to the level of a single spin.~\cite{Muller2014,Shi2018,Shi2013,Sushkov2014,Abeywardana2016}
NV-detected NMR is now widely used at low magnetic fields ($<0.1$ T), such as for NV depth estimation, liquid state NMR, two-dimensional NMR, hyperpolarized NMR, nanodiamond based NMR, and even for selective spin manipulation in a 10-qubit quantum register.~\cite{Pham2016,Kehayias2017,Smits2019,Holzgrafe2020,Bradley2019}
NMR at high magnetic fields greatly increases the spectral resolution and improves sensitivity.
The increase in field strength increases the frequency difference between closely related chemical species and enables resolution of small chemical shifts.
High field NMR offers new insights into molecules with many similar nuclei, low gyromagnetic ratios, and low natural abundance, such as for $^{17}O$ NMR in pharmaceutical compounds and biomacromolecules.~\cite{Kong2013,Keeler2019}
Commercial NMR magnets operating at 28.2 T (proton Larmor frequency of 1.2 GHz) have recently become available, with hybrid magnets at fields of 35.2 T (corresponding to 1.5 GHz proton NMR) being available in user facilities.~\cite{Luchinat2021, Gan2017}
Implementation of NV-detected NMR at a high magnetic field is highly desirable.
However, there have only been a handful of studies on NV-based sensing at high magnetic fields due to technological challenges involved with combining a NV ODMR system with a high magnetic field ESR system.~\cite{Stepanov2015,Fortman2020,Aslam2015,Aslam2017}
In this paper, we discuss the implementation of NV-detected NMR at a high magnetic field.
NV-detected NMR can be achieved by hyperfine spectroscopic techniques.
There are three primary pulsed ESR hyperfine spectroscopic techniques: electron spin echo envelope modulation (ESEEM), electron-nuclear double resonance (ENDOR), and electron-electron double resonance detected NMR (EDNMR).~\cite{schweiger2001principles, Mims1965,Schosseler1994}
Most NV-detected NMR spectroscopy performed at a low magnetic field is based on ESEEM where the hyperfine coupling between the NV center and nuclear spins mixes the spin-state and results in periodic revivals of the echo intensity.~\cite{Childress2006}
This technique functions very efficiently only when the energies of the hyperfine coupling and the nuclear Larmor frequency are comparable.
Therefore, it works well at a low magnetic field, but becomes unfeasible at a high magnetic field.~\cite{schweiger2001principles}
In ENDOR and EDNMR techniques, the population difference of an ESR transition is monitored via a detection scheme while either pulsed RF (ENDOR) or off resonance MW (EDNMR) radiation is applied to drive polarization transfer.
Nuclear identification and determination of hyperfine coupling is performed based on the frequency of polarization transfer.
EDNMR uses a high turning angle (HTA) pulse to drive population transfer on forbidden transitions.
At higher fields, the Zeeman interaction more completely dominates over the hyperfine interaction, reducing state mixing and consequently, transition probability.
Therefore, higher magnetic fields require stronger or longer HTA pulses to induce polarization transfer.
Both EDNMR and ENDOR are promising for NV-based sensing, as they are applicable to single and ensemble NV systems and are limited by the longitudinal relaxation time, $T_1$, instead of the transverse relaxation time, $T_2$.
The $T_1$ relaxation time for NV ensembles has been shown to extend dramatically (up to minutes) at low temperature.~\cite{Takahashi2008,Jarmola2012}
More recently, EDNMR has emerged as a promising technique due to its higher sensitivity and resiliency against RF related artifacts.~\cite{Cox2017}
EDNMR has an additional advantage over ENDOR in that it does not require an additional RF power amplifier or tuned RF circuit and can thus be readily implemented over a large frequency range for the detection of nuclei with a wide range of gyromagnetic ratios.
Within this work, we demonstrate optically detected magnetic resonance (ODMR) on the NV center at the highest field and frequency to date, 8.3 T, corresponding to the NV's Larmor frequency of 230 GHz (proton Larmor frequency of 350 MHz).
We successfully implement EDNMR using ensemble NV centers and detect $^{13}C$ nuclear bath spins in the diamond crystal.
Since the EDNMR technique is limited by $T_1$, not $T_2$, NV-detected NMR based on EDNMR can take advantage of the NV center's long $T_1$ to perform measurements with a long HTA pulse.
With development of suitable pulse capabilities, the described NV-detected NMR technique will be advantageous for the development of NV-detected NMR at higher fields and frequencies where the microwave power is often limited.~\cite{Morley2008, Takahashi12, Fortman2020}
\section{Methods and Materials}
\begin{figure}
\centering
\includegraphics[width=3.3in]{Fig1.eps}
\caption{Overview of the experimental setup.
The transmission (Tx) setup consists of two independently controllable frequency sources (MW1 and MW2) that pass through PIN switches to a frequency multiplication chain.
High frequency MW excitation is propagated through quasioptics and a corrugated waveguide to the sample stage within a 12.1 Tesla variable field magnet.
Pulsed laser excitation is directed through an acousto-optic modulator (AOM) and an optical fiber to a system of lenses, a fast steering mirror, and the sample stage.
At the sample stage, a microscope objective directs laser intensity and collects sample fluorescence.
The fluorescence is redirected through a dichroic mirror to a photodiode where it is integrated using either gated boxcar integrators or a fast oscilloscope.
The MW components, laser, and boxcar integrators are all controlled through a central computer equipped with a fast TTL logic board and digital to analog converter (DAQ).
The magnetic field ($B_0$) is aligned with the optical axis.
}
\label{fig:Setup}
\end{figure}
A home-built, high field (HF) ODMR spectrometer operating in the band of 215-240 GHz was used.
An overview of the experimental setup is shown in Fig.~\ref{fig:Setup}.
The diamond sample was mounted at the center of a variable field 12.1 T superconducting magnet (Cryogenic Limited).
Microwave (MW) excitation was produced by a solid state source (Virginia Diodes) and directed through quasioptics to the sample stage.
The output power of both channels from the source was 115 mW at 230 GHz.
Laser excitation was produced from a solid-state single mode laser (Crystalaser) and directed through an acousto-optic modulator (Isomet), single mode fiber (Thorlabs), and microscope objective (Zeiss100X, NA=0.8) before reaching the sample stage.
The excitation beam position was controlled using a fast steering mirror (Newport) and a system of lenses below the microscope objective.
Fluorescence (FL) collected at the objective was directed back through a dichroic mirror and fluorescence filters (Omega Optics) before being detected using a photodiode (Thorlabs 130A2).
The typical excitation spot size was a few $\mu$m$^2$.
Typical laser excitation of $\sim 4$ mW at the sample stage resulted in 1-2 $\mu$W of detected FL.
The output of the photodiode was directed to a signal integrator.
Integration was performed using either a pair of analog boxcar integrators (Stanford Research Systems SR250) or a fast digitizing oscilloscope (Tektronix MSO64B). The analog output of the boxcar integrators was digitized using a fast DAQ (National Instruments PCIe-6321).
Gate timing was controlled using a gated TTL logic board (SpinCore Technologies PB-500).
Additional details of the HF-ESR/ODMR spectrometer have been described previously.~\cite{Cho2014,Cho2015, Stepanov2015, Fortman2020}
For this study, two samples were used. Sample 1 was a 2.0 $\times$ 2.0 $\times$ 0.3 mm$^3$ size, (111)-cut high pressure, high temperature type Ib diamond from Sumitomo Electric Industries. Sample 2 was a hexagonal 4.4 $\times$ 3.9 $\times$ 0.5 mm$^3$ size, (111)-cut high pressure high temperature type-Ib diamond obtained from Element Six.
Both diamonds had previously been subjected to high energy (4 MeV) electron beam irradiation and were exposed to a total fluence of $1.2 \times 10^{18}$ e$^-/$cm$^2$ followed by an annealing process at 1000 $^o$C. This treatment produced a NV concentration greater than 1 ppm.~\cite{Fortman2020}
\section{Discussion}
We begin by performing pulsed ODMR on ensemble NV centers.
\begin{figure}
\centering
\includegraphics[width=3.3in]{Fig2.eps}
\caption{Ensemble ODMR at 230 GHz.
(a) Pulsed ODMR data. For all ODMR measurements, laser pulses of $20$ $\mu$s and $15$ $\mu$s were used for initialization (Init) and readout (RO), respectively.
After initialization, a MW1 pulse ($t_p$) of $1.9$ $\mu$s was applied and varied in frequency.
Clear reductions in FL intensity were resolved at $229.953$ GHz and $235.687$ GHz, corresponding to the lower ($\ket{m_S=0} \leftrightarrow \ket{m_S=-1}$) and upper ($\ket{m_S=0} \leftrightarrow \ket{m_S=+1}$) transitions of the NV center.
The magnetic field was found to be $8.306$ T with a polar angle of $1.50 \pm 0.02^\circ$.
Fitting was performed using nonlinear least squares regression and the NV center Hamiltonian ($S = 1$, $D = 2870$ MHz, $g = 2.0028$).~\cite{Takahashi2008}
(b) Measurement of Rabi oscillations.
The frequency of MW1 was set at the lower resonance and the pulse length was varied.
From the observed oscillations, a $\pi$ pulse length of $1.9$ $\mu$s was found.
(c) Measurement of $T_1$ relaxation. Measurements were performed with (Sig1) and without (Sig2) a $\pi$ pulse.
The difference (Sig2-Sig1) was normalized and then fit to a single exponential decay.~\cite{Jarmola2012}
Data was collected using (a) 10 scans, (b) 18 scans, and (c) Sig1 and Sig2 were measured sequentially with 5 scans each.
}
\label{fig:NVCharacterization}
\end{figure}
For pulsed ODMR, the relative FL intensity was monitored while a MW pulse was varied in frequency.
As seen in Fig.~\ref{fig:NVCharacterization}(a), clear reductions in FL intensity were resolved at $229.953$ GHz and $235.687$ GHz, corresponding to the lower ($\ket{m_S=0} \leftrightarrow \ket{m_S=-1}$) and upper ($\ket{m_S=0} \leftrightarrow \ket{m_S=+1}$) transitions of a [111] oriented NV with a polar offset angle of $1.50 \pm 0.02$ degrees.
Next, Rabi oscillations were recorded by fixing the frequency of MW1 at $229.953$ GHz ($\ket{m_S=0} \leftrightarrow \ket{m_S=-1}$ transition) and varying the pulse length as seen in Fig.~\ref{fig:NVCharacterization}(b).
From these measurements, damped oscillations and a $\pi$ pulse length of $1.9$ $\mu$s was observed.
Next, the NV ensemble's spin-lattice relaxation time, $T_1$, was recorded.
For this measurement, the duration between the laser initialization and readout pulse ($\tau$) was varied (see Fig.~\ref{fig:NVCharacterization} (c).
Two sequential measurements were performed by varying the spacing between initialization and readout with and without a MW $\pi$ pulse before normalization. A $T_1$ time of $3.9 \pm 0.2$ ms was found by fitting to a single exponential decay.
Next we perform EDNMR using the NV center.
As shown in Fig.~\ref{fig:HFNVEDNMR}(a), EDNMR is a form of high field hyperfine spectroscopy that utilizes two microwave frequencies, MW1 ($\nu_0$) and MW2 ($\nu_1$).
EDNMR measurements vary the frequency ($\nu_1$) of a HTA MW2 pulse, while MW1 applies a detection pulse sequence, such as Hahn echo, at $\nu_0$ to measure the spin polarization of an ESR transition.~\cite{Schosseler1994}
As the frequency of $\nu_1$ is swept, the frequency shifts on resonance with transitions below the central transition ($\nu_1<\nu_0$) due to weakly coupled hyperfine nuclei, as seen in Fig.~\ref{fig:HFNVEDNMR}(b).
These transitions are generally forbidden as they involve a flip of both the electron and nuclear spin ($\Delta m_S = 1, \Delta m_I = 1$).
The forbidden transitions become weakly allowed with partial state mixing, leading to polarization transfer and a reduction in the ESR signal intensity.
This change is detected as an EDNMR signal.
Application of a long HTA pulse improves the likelihood of population transfer, but the total length of the HTA pulse must be short relative to $T_1$ in order to maximize the observable contrast.
As $\nu_1$ approaches the central allowed transition ($\Delta m_S = 1,\Delta m_I = 0$) there is significant population transfer leading to a highly intense change and the so-called "central blind spot".
Since the central blind spot highly distorts EDNMR signal, in practice the measurement is performed at a frequency range outside of the central blind spot.
After passing the central blind spot, $\nu_1$ then induces forbidden transitions from hyperfine coupled nuclei with a positive frequency offset ($\nu_1>\nu_0$) relative to the central transition.
For NV detected EDNMR, the spin population can be directly detected via optical spin state readout, eliminating the need for an echo detection sequence.
EDNMR with the NV center has an advantage over conventional EDNMR, as optical initialization of the NV center ensures high spin polarization and improves EDNMR sensitivity.
The usage of optical initialization shortens the measurement time by eliminating the need for long cycle delays between subsequent experiments (typically $\gg$ $T_1$).
As shown in Fig.~\ref{fig:HFNVEDNMR}(a), we perform the experiment by applying an initialization laser pulse, MW2 HTA pulse at frequency $\nu_1$, MW1 $\pi$ pulse at frequency $\nu_0$, and laser readout pulse.
During the experiment $\nu_1$ is varied while $\nu_0$ is fixed at the lower NV resonance.
When the HTA pulse drives a transition, the population of the $\ket{m_S = 0}$ spin state is reduced before the MW1 $\pi$ pulse transfers the population to the $\ket{ms=-1}$ state.
Therefore, when the HTA pulse is in resonance with a transition, an increase in the FL intensity is observed.
For the present experiment, a HTA pulse length of $500$ $\mu$s was chosen.
In principle, longer length pulses, up to $T_1$, can be applied.
Figure~\ref{fig:HFNVEDNMR}(c) shows the result of the experiment and we observe signals at $\pm88$, $-64$, $-30$, $+28$, and $+65$ MHz.
The strong change in the FL intensity at $0$ MHz corresponds to the central blind spot.
The signals at $-64$ and $+65$ MHz give the hyperfine coupling constant of 129 MHz, consistent with nearest neighbor $^{13}C$ hyperfine interaction ($126$-$130$ MHz) splitting the allowed ESR transition.~\cite{Nizovtsev2010,Smeltzer2011,Jarmola2016}
The reduced intensity relative to the central transition corresponds to the low natural abundance of $^{13}C$ ($\sim$1.1\%) and low probability of nearest neighbor locality.
\begin{figure*}
\centering
\includegraphics[width=6in]{Fig3.eps}
\caption{NV detected EDNMR at high field.
(a) Pulse sequence used in the NV-detected EDNMR experiment.
In the experiment, a HTA pulse was applied with MW2 at frequency $\nu_1$ before a $\pi$ pulse was applied with MW1 at frequency $\nu_0$.
The frequency of $\nu_0$ was set to match the lower transition.
The application of a $\pi$ pulse increases the sensitivity by isolating the FL of [111] oriented NV centers from non axial orientations.
(b) Energy level diagram.
Nuclei coupled via weak hyperfine interaction are represented by $m_I = +$ and $m_I = -$.
During the experiment, the frequency of the HTA pulse is swept from below ($\nu_1<\nu_0$) to above ($\nu_1>\nu_0$) the central ESR resonance.
Population is transferred when the HTA pulse is in resonance with the difference between coupled states, resulting in an increase in the observed FL.
Due to the length of the HTA pulse and state mixing induced by the hyperfine interaction, this occurs for both allowed and forbidden transitions.
The intensity of the central blind spot is due to the allowed transitions.
(c) Experimental spectra.
The data are shown with reference to the MW frequency offset ($\nu_1 -\nu_0$) and normalized to the intensity of the central blind spot.
In the present case, $\nu_0 = 229.9528$ GHz.
A 500 $\mu$s HTA pulse and 1.9 $\mu$s $\pi$ pulse were used.
The length of the HTA pulse was chosen to minimize the influence of $T_1$ relaxation after population transfer.
EDNMR signals due to forbidden transitions involving $^{14}N$ and $^{13}C$ are indicated.
Grey stars are used to indicate peaks due to allowed transitions from nearest neighbor $^{13}C$ lattice sites.
For (c), data was collected using 20 scans over a period of 11 hours.
}
\label{fig:HFNVEDNMR}
\end{figure*}
Next we discuss signals at $\pm 88$ MHz.
In order to understand the signals we discuss the following Hamiltonian:
\begin{equation}
H_{NV} = \mu_B g_{NV} \Vec{B_0} \cdot \Vec{S} + D \Vec{S_z}^2 +H_{N} + H_{C},
\label{eq:NVHamil}
\end{equation}
where $D = 2.87$ GHz, $g_{NV} =2.0028$, and $\Vec{S}$ is the electronic spin operator.~\cite{Takahashi2008} $H_{N}$ and $H_{C}$ represent the Hamiltonians of hyperfine coupled nitrogen in the NV center and surrounding $^{13}C$ bath spins.
The nuclear spin Hamiltonians may be written as:
\begin{subequations}
\begin{align}
H_{N} = -\gamma_{^{14}N} \Vec{B_0} \cdot \Vec{I_1} +\Vec{S} \cdot \Vec{A}_{^{14}N} \cdot \Vec{I_1} + P I_{1z}^2 \label{H14N},\\
H_{C} = -\gamma_{^{13}C} \Vec{B_0} \cdot \Vec{I_2} +\Vec{S} \cdot {A}_{^{13}C} \cdot \Vec{I_2} \label{H13C},
\end{align}
\end{subequations}
where $\gamma_{nuc}$ represents the gyromagnetic ratios ($3.077$ and $10.708$ MHz/T for $^{14}N$ and $^{13}C$, respectively), $\Vec{I_1}$ ($\Vec{I_2}$) is the $^{14}N$ ($^{13}C$) nuclear spin operator, $\Vec{A}_{nuc}$ is the hyperfine interaction ($^{14}N$: $A_{\perp} = -2.14$ MHz, $A_{\parallel} = -2.70$ MHz), and P represents the nuclear quadrupole interaction ($-5.0$ MHz).~\cite{Felton2009}
We focus our study on weakly coupled $^{13}C$ nuclear bath spins.
Using Eq.~\ref{eq:NVHamil}, we determine all eigenvalues based on the observed magnetic field.
The observed states and energies are listed in Table~\ref{tb:NVEigenVals}.
\begin{table}
\caption{State identification and energy values determined from Eq.~\ref{eq:NVHamil} based on a magnetic field of 8.306 Tesla with an offset angle of 1.5 degrees and $A_{^{13}C} = 1$ kHz. The nuclear magnetic spin value of $^{14}N$ ($^{13}C$) is shown as $m_{I1}$ ($m_{I2}$).
}
\begin{ruledtabular}
\begin{tabular}{cc}
\textbf{State $\ket{m_S,m_{I1},m_{I2}}$} & \textbf{Energy [MHz]} \\ \hline
{-1, +1, +1/2} & -230023.7 \\
{-1, 0, +1/2} & -229995.3 \\
{-1, -1, +1/2} & -229976.9 \\
{-1, +1, -1/2} & -229934.8 \\
{-1, 0, -1/2} & -229906.3 \\
{-1, -1, -1/2} & -229887.9 \\
{0, +1, +1/2} & -73.1 \\
{0, 0, +1/2} & -42.5 \\
{0, -1, +1/2} & -21.9 \\
{0, +1, -1/2} & 15.9 \\
{0, 0, -1/2} & 46.4 \\
{0, -1, -1/2} & 67.0
\end{tabular}
\label{tb:NVEigenVals}
\end{ruledtabular}
\end{table}
From Table~\ref{tb:NVEigenVals}, it is seen that the $m_S = 0$ states are not evenly spaced around zero.
This spacing is induced by partial field misalignment and nuclear quadrupole interaction that mixes the states and results in twelve non degenerate energy levels.
We next calculate allowed transitions ($\Delta m_S = 1$, $\Delta m_I = 0$) and double quantum transitions ($\Delta m_S = 1$, $\Delta m_I = 1$) involving a simultaneous electron and nuclear spin flip.
We tabulate the allowed transitions and double quantum transitions involving $^{13}C$ and $^{14}N$ spin flips in Table~\ref{tb:NVTransitions}.
\begin{table}
\caption{Simulated transition energies calculated from Table~\ref{tb:NVEigenVals}.
The states involved in the transition are listed in the left and central columns, while the calculated difference is shown in the right column.
For clarity, the transition relative to the central transition ($\nu_{sim.}-\nu_{obs.}$) was tabulated ($\nu_{obs.} = 229.9528$ GHz).
Allowed transitions ($\Delta m_S = 1$, $\Delta m_{I} = 0$) are shown in the top panel.
The middle and bottom panel show double quantum transitions ($\Delta m_S = 1$, $\Delta m_{I} = 1$) involving a simultaneous electron and nuclear spin flip. The middle panel shows transitions involving $^{14}N$ and the bottom panel shows transitions involving $^{13}C$.
}
\begin{ruledtabular}
\begin{tabular}{ccc}
$\bf \ket{0, m_{I1}, m_{I2}}$ & {$\bf \ket{-1, m_{I1}, m_{I2}}$} & \textbf{$\Delta$ E [MHz]} \\ \hline
{ +1, +1/2} & { +1, +1/2} & -2.1 \\
{ +1, -1/2} & {+1, -1/2} & \\
{ 0, +1/2} & { 0, +1/2} & 0.0 \\
{ 0, -1/2} & { 0, -1/2} & \\
{-1, +1/2} & {-1, +1/2} & 2.1 \\
{-1, -1/2} & {-1, -1/2} & \\ \hline
{+1, +1/2} & { 0, +1/2} & -30.6 \\
{+1, -1/2} & { 0, -1/2} & \\
{ 0, +1/2} & {-1, +1/2} & -18.4 \\
{ 0, -1/2} & {-1, -1/2} & \\
{-1, +1/2} & { 0, +1/2} & 20.6 \\
{-1, -1/2} & { 0, -1/2} & \\
{ 0, +1/2} & {+1, +1/2} & 28.4 \\
{ 0, -1/2} & {+1, -1/2} & \\ \hline
{+1, +1/2} & {+1, -1/2} & -91.1 \\
{ 0, +1/2} & { 0, -1/2} & -88.9 \\
{-1, +1/2} & {-1, -1/2} & -86.8 \\
{+1, -1/2} & {+1, +1/2} & 86.8 \\
{ 0, -1/2} & { 0, +1/2} & 88.9 \\
{-1, -1/2} & {-1, +1/2} & 91.1
\end{tabular}
\end{ruledtabular}
\label{tb:NVTransitions}
\end{table}
As seen in Table~\ref{tb:NVTransitions}, the allowed ESR transitions are spaced by the axial hyperfine coupling to $^{14}N$, contributing to the central blind spot.
As shown in the inset of Fig. \ref{fig:C13EDNMR}(a), the signals at $-30$ and $+28$ MHz are in excellent agreement with the predicted peak positions for $^{14}N$ predicted in Table~\ref{tb:NVTransitions}.
The proximity to the central spot makes identification of the peaks at $-18$ and $20$ MHz difficult, but a dip at $-18$ MHz is in agreement with the expected peak position.
The polarity inversion of the signals is under further investigation. The observed signals are not symmetric due to the nuclear quadrupole interaction.
The main graph of Fig. \ref{fig:C13EDNMR}(a) shows the signals at $\pm88$ MHz in excellent agreement with double quantum transitions for $^{13}C$ bath spins and are well spaced from the central blind spot.
We next repeat the measurements on additional locations to confirm the observed signals.
We adjust the mirror to position 2, $\sim 50$ $\mu m$ from position 1, and repeat EDNMR measurements.
We also include data from a separate experimental run as position 3.
For position 3, the sample was removed from the setup and replaced, resulting in a different sample location.
Fig. \ref{fig:C13EDNMR}(a) shows $^{13}C$ EDNMR signals at $\pm 88$ MHz for all positions in excellent agreement with the simulation.
\begin{figure*}
\centering
\includegraphics[width=6.3in]{Fig4.eps}
\caption{
NV detected EDNMR at 8.3 Tesla
(a) NV detected EDNMR from sample 1.
EDNMR detection of $^{13}C$ is shown in the main graph and EDNMR detection of $^{14}N$ is shown in the inset.
Data is offset for clarity. The presented data is from three areas: positions 1 and 2 were spaced $\sim 50$ $\mu$m apart, position 3 was taken after removing and replacing the sample. The data for position 3 was integrated with boxcar integrators, all other measurements were integrated using the fast oscilloscope.
For position 3, variations in the experimental setup resulted in slightly different parameters: the magnetic field was 8.298 T with a polar offset angle of $1.9 \pm 0.1^o$. Rabi oscillations showed a $\pi$ pulse length of $1.6$ $\mu$s. The change in magnetic field resulted in a small ($\sim 0.1$ MHz) shift in the transition frequencies.
The stick spectrum shows double quantum transitions from Table \ref{tb:NVTransitions}.
A simulation based upon $^{13}C$ coupling to the NV center is shown in red.
The red line shows a simulation of $L(\omega;\Delta \omega,\omega_i) = A/\pi \sum_{\omega_i} {\Delta \omega}/({\Delta \omega^2+4(\omega - \omega_i)^2})$ where $A$ is an amplitude and the sum runs over the resonance positions ($\omega_i$).
A nitrogen spin concentration of 70 ppm ($\Delta \omega = 3.2$ MHz) was used in agreement with sample properties.
Nonlinear regression of $L(\omega;\Delta \omega,\omega_i)$ was used to determine $\Delta \omega$ from the experimental data (fits not shown). $\Delta \omega$ was measured to be $2.3 \pm 0.3 $, $2.9 \pm 0.4 $, and $3.7 \pm 0.8 $ MHz for positions 1, 2, and 3 respectively.
(b) NV detected EDNMR from sample 2.
EDNMR detection of $^{13}C$ is shown in the main graph and EDNMR detection of $^{14}N$ is shown in the inset. Data is offset for clarity.
The presented data is from three areas: position 1, 2 and 3 were spaced $\sim 50$ $\mu$m apart from each other. The stick spectrum shows the position of double quantum transitions.
For sample 2, $\Delta \omega$ was measured to be $2.7 \pm 0.3 $, $2.9 \pm 0.3 $, and $2.5 \pm 0.3 $ MHz for positions 1, 2, and 3 respectively.
}
\label{fig:C13EDNMR}
\end{figure*}
We next investigate the linewidth of the $^{13}C$ signals in more detail.
We plot the signals related to double quantum transitions in Fig.~\ref{fig:C13EDNMR} and show the transitions as a stick spectrum.
The calculated three transition frequencies are ranged by $4.3$ MHz, which is comparable to the observed linewidth.
In general, the EDNMR linewidth is dependent on a variety of factors, including both intrinsic properties, such as spin relaxation times, and experimental parameters, such as the HTA pulse length and intensity.\cite{Cox2017}
In the present case, the observed linewidth was observed to be constant when HTA pulse lengths from $300 - 1000$ $\mu$s were used, suggesting that the linewidths are broadened by internal dynamics.
Therefore, we focus our discussion on magnetic dipole coupling from surrounding spins which can contribute to the observed linewidth.
In general, the magnetic field at an "A" spin fluctuates due to the interaction with random spin flips of dipolar-coupled "B" spins.
When the concentration of "B" spins is sufficiently dilute, this interaction broadens the linewidth of the "A" spin by inducing a distribution of Larmor frequencies.
In this case, the Larmor frequency fluctuations ($\Delta \omega$), at an "A" spin from the j-th dipolar coupled "B" spins may be written as:
\begin{equation}
\Delta \omega_j = \gamma_a \delta b_j = \frac{\mu_0 \gamma_a \gamma_b \hbar}{ 4 \pi} \frac{(1-3 \cos^2\theta_j)m_j}{r^3_j},
\label{eq:BVariance}
\end{equation}
where $\gamma_a$ ($\gamma_b$) is the gyromagnetic ratio of the "A" ("B") spin, $\mu_0$ is the permeability of free space, and $\hbar$ is the reduced planck constant.
The spin state of the j-th spin is given by $m_j$ ($m_j = \pm 1/2$ for an $S = 1/2$ spin) with $\theta_j$ representing the angle between the vector joining the spins, $r_j$, and the applied magnetic field.
Now by considering that "B" spins are randomly distributed and the populations of the up- and down-states of "B" spins are equal, we can average $\Delta \omega_j$ by considering the probability of finding a spin at the j-th position and integrating over possible angles and spin states.~\cite{Mims1968, Klauder1962, Abragam1961}
The integral gives the full-width at the half-maximum of the Lorentzian function as a linewidth ($\Delta \omega$), which may be written as:
\begin{equation}
\Delta \omega = \sum_j \Delta \omega_j = \frac{ 2 \pi \mu_0\hbar}{ 9 \sqrt{3}}\gamma_a \gamma_b n,
\label{eq:LinewidthC13}
\end{equation}
where n is the concentration of "B" spins in units of spins per cubic meter.
In the case of the present EDNMR study, "A" spin is the NV center and "B" spins are surrounding paramagnetic spins such as P1 centers and $^{13}C$ nuclear spins.
The concentration of nitrogen in the present sample was estimated to be $\sim$ 70 ppm from a 230 GHz pulsed ESR measurement of the P1 center's $T_2$ ($T_2 = 1.07 \pm 0.01$ $\mu$s; data not shown).\cite{Stepanov2016}
Using 70 ppm for the concentration of P1 centers, we obtained $\Delta \omega = 3.2$ MHz.
As shown in Fig.~\ref{fig:C13EDNMR}, the simulated peaks with the three resonance frequencies and $\Delta
\omega$
gives excellent agreement with the observed data.
Furthermore, the observed linewidth is in excellent agreement with the linewidth of the lower NV resonance (Fig.~\ref{fig:NVCharacterization} (a)) and with previous work on type-Ib diamonds.~\cite{Stepanov2016}
The use of high purity, isotopically purified diamonds with low concentrations of paramagnetic spins can be used to further improve the spectral resolution and is the subject of current work.
For example, we note that Eq.~\ref{eq:LinewidthC13} predicts $\Delta \omega \sim 0.2$ MHz from dipolar broadening due to natural abundance $^{13}C$.
We next discuss measurements on sample 2. All measurements previously discussed were repeated on sample 2. From measurement of both the lower and upper ODMR transitions, the magnetic field was determined to be 8.306 T with a polar offset angle of $1.88 \pm 0.03^o$. Rabi oscillations showed a $\pi$ pulse length of $1.6$ $\mu$s and the $T_1$ relaxation time was measured as $3.8 \pm 0.3$ ms (data not shown).
As seen in Fig. \ref{fig:C13EDNMR}(b), EDNMR was measured at three different locations on sample 2, with EDNMR signals from $^{13}$C resolved at $\pm 88$ MHz in each location. The observed signals are in excellent agreement with the expected peak positions. The slight variation in the observed height and width from sample 1 indicates small sample to sample variation.
The inset shows EDNMR signals resolved from $^{14}$N. Clear signals are resolved at $-31$ and $+28$ MHz in excellent agreement with the simulated peak positions and sample 1.
In summary, we have demonstrated pulsed ODMR on an ensemble system of NV centers at 8.3 Tesla and 230 GHz.
Ensemble NV centers were utilized to perform pulsed EDNMR with optical readout of the spin population.
EDNMR signals were resolved from $^{13}C$ bath spins with the linewidth limited by the concentration of paramagnetic impurities.
This work provides a clear demonstration of NV center detected EDNMR, and establishes groundwork for the implementation of NV-detected NMR at higher magnetic fields, with shallow NV centers, and for the study of nuclei with a variety of gyromagnetic ratios.
EDNMR can resolve spins whose gyromagnetic ratios shift the resonance from the central blind spot.
Nuclei with large gyromagnetic ratios, such as $^{1}H$ and $^{19}F$, are excellent candidates for future research.
Signals from bath $^{13}C$ spins were resolved in this work.
From previous measurements, it is known that weakly coupled $^{13}C$ hyperfine interaction is on the order of $10-100$ kHz. \cite{Zopes2018,Bradley2019}
Based on a dipolar calculation, a hyperfine coupling of more than 10 kHz is expected for surface protons within 8 nm of NVs.
With the fabrication of NVs with $T_1$ times of a few ms, stable photoluminescence, and a depth of at least 8 nm,\cite{Ishiwata2017,Osterkamp2019,Loretz2014} NV-NMR of protons at the diamond surface will be detectable with the presented EDNMR technique.
Chemical functionalization techniques can be used to bring spins of interest within close proximity of shallow NV centers.~\cite{Romanova2013, Akiel2016}
Furthermore, the described technique is limited by the comparative length of the HTA pulse relative to $T_1$ relaxation.
As $T_1$ can be extended up to several seconds at cryogenic temperatures, this technique can utilize a long HTA pulse to perform measurements at higher fields and frequencies where microwave power is often limited.~\cite{Takahashi2008, Jarmola2012}
With the development of suitable pulsing techniques, this method will enable measurements in higher magnetic fields, such as those in the National High Magnetic Field Laboratory.~\cite{Zvyagin2004,Stoll2011}
\section{Acknowledgements}
This work was supported by the National Science Foundation (CHE-2004252 with partial co-funding from the Quantum Information Science program in the Division of Physics), the USC Anton B. Burg Foundation, and the Searle scholars program (ST). L.M.S. thanks support from USC Dornsife Chemical Biology Training Program.
Zaili Peng is acknowledged for helpful discussions regarding EDNMR.
\smallskip
\noindent{\bf DATA STATEMENT}\newline
The data that support the findings of this study are available from the corresponding author upon reasonable request.
|
\section{Introduction}
\label{sec:introduction}
Humor, a way of giving entertainment and provoking laughter, is an exceptionally savvy communicative activity. In recent years, quite some attention has been devoted to humor production and perception. What makes us laugh when reading a funny sentence? Mostly it depends on the cultural area background of a person. Humor cannot be defined in standards as it varies with nation to nation, area to area, people to people. Many types of humor require substantial external knowledge such as irony, wordplay, metaphor and sarcasm. These factors make the task of humor recognition difficult.
Recently, with the advance of deep learning that allows end-to-end training with big data without human intervention of feature selection, humor recognition becomes promising. However detecting humor in news headlines from which one word is chosen to be replaced by another word to make the micro-edited sentences look funny is quite challenging. To address the challenges of humor detection in edited news headlines, ~\cite{SemEval2020Task7} proposed detection of funniness level in dataset named as Humicroedit\footnote{Dataset formed for this task is named as humicroedit in the article ~\cite{hossain-etal-2019-president}.}, Task 7 at SemEval-2020.
They focus on two related subtasks. Task 1 defines regression problem where a system needs to predict how much funny is the edited news headlines (humicroedit) from range 0 to 3. Whereas Task 2 defines the classification problem where a system need to predict the funniest sentence between two edited headlines.
In this paper, we only focus on the Humicroedit data for both Tasks 1 and 2. No extra data, or other sources is used to get more data. The rest of the paper is structured as follows: \textbf{Section~\ref{sec:related work}} provides a brief overview of prior research. In \textbf{Section~\ref{sec:proposed model}}, we introduce our proposed neural network model. \textbf{Section~\ref{sec:ExperimentsEvaluations}} includes experiments and evaluations as well as the analysis of our proposed method. Some concluded remarks and future directions of our work are described in \textbf{Section~\ref{sec:Conclusion}}.
\section{Related Research}
\label{sec:related work}
\blfootnote{
%
%
\hspace{-0.65cm}
%
%
%
}
Humor is omnipresent, universal, elusive event which exists in all societies and cultures and fulfills a range of social, cognitive and emotional functions. The task of automatic humor recognition refers to deciding whether a given sentence expresses a certain degree of humor. In previous researches and studies, mostly the problems related to humor were based on binary classification, or based on selection on linguistic features. Taylor and Mazlack analyzed on a specific type of humor. The methodology which they used was based on the extraction of structural patterns and peculiar structure of jokes newcite~\cite{Taylor2000}.
Purandare and Litman used humorous spoken conversations as data from a classic comedy television. To identify humorous speech in the conversation, they used standard supervised learning classifiers~\cite{purandare2006humor}.
Besides, a work has been done to discover latent semantic structure behind humor from four perspectives: incongruity, ambiguity, interpersonal effect and phonetic style. For that Yang created a computational model. He also formulated a classifier to distinguish between humorous and non-humorous instances~\cite{Yang2015}.
In other researches, with the development of artificial neural networks, many studies utilize the methods for humor recognition. Luke de Oliveira and Alfredo applied recurrent neural network (RNN) to humor detection from reviews in Yelp dataset. They also applied convolutional neural networks (CNNs) to train a model and the results from model trained with CNNs have more accurate humor recognition~\cite{de2015humor}.
Recently, Bertero and Fung proposed a first-ever attempt to employ a Long ShortTerm memory(LSTM) based framework to predict humor in dialogues. They showed how the LSTM effectively models the setup-punchline relation reducing the number of false positives and increasing the recall~\cite{Bertero2016a}.
In a recent work, Chen and Lee predicted audience’s laughter also using convolutional neural network. Their work gets higher detection accuracy and is able to learn essential feature automatically~\cite{Chen2018a}.
Among several prominent works, Peng-Yu Chen and Von-Wun Soo build the humor recognizer by using CNNs with extensive filter size and number, and the result shows higher accuracy from previous CNNs models by conducting experiments on two different datasets.~\cite{Chen2018}.
\section{Proposed Model}
\label{sec:proposed model}
In this section, we describe the details of our proposed neural network model. The goal of our proposed approach is to predict how much funny the edited news headlines (humicroedit) is. We tried to train our model after data processing. Figure 1 depicts an overview of our proposed model.
\begin{figure}[h!]
\begin{center}
\centerline{\includegraphics[width=1\linewidth]{fig1.pdf}}
\caption{Proposed Model} \label{Figure 1}
\end{center}
\end{figure}
At first, in a dataset, original sentences were changes by using edit words. Then on edited sentences, pre-trained word embedding model was implemented to obtain the high-quality distributed vector representations for our datasets. Next, we apply the Bidirectional LSTMs (BiLSTMs) models to extract the higher level feature sequences with sequential information from the edited news headlines embedding. We employ an encoder pre-trained by Google News word2vec model to encode each word into 300-dimensional feature vector. These feature vectors are then sent to a Bidirectional LSTM module. Finally, the generated output feature sequences from Bidirectional LSTMs fed into the fully-connected prediction module to determine the prediction. Next, we describe each component elaborately.
\subsection{Embedding Layer}
\label{ref:embeddingLayer}
The Embedding layer in the leftmost part of Figure1 is initialized with random weights and will learn an embedding for all of the words in the training dataset. This flexible layer is defined as the first hidden layer of a network.
A pre-trained model used in embedding layer is nothing more than a file containing tokens and their associated word vectors. The pre-trained Google word2vec model was trained on Google news data (about 100 billion words); it contains 3 million words and phrases and was fit using 300-dimensional word vectors with the size of 1.53 GB.
In our proposed framework, we utilize a pre-trained word embedding model based on google news to obtain the high-quality distributed vector representations of headlines. The dimensionality of the
Embedding matrix will be $L \times D$, where $L$ is the sentence length, and $D$ is the word-vector dimension.
\subsection{Bidirectional LSTMs}
\label{ref:BiLstmLayer}
Bidirectional Long Short Term Memory (BiLSTM) in the central part of Figure1 is a bidirectional version of LSTM. BiLSTM combines the forward hidden layer with the backward hidden layer, which can access both the preceding and succeeding contexts.
BiLSTM neural network is resorted to attain the vector representation of the input sentence, which can effectively capture the semantics of the sentence.
The final result from output layer of BiLSTM is generated through combining the results produced by both RNN hidden layers i.e. forward layer and backward layer. Following equations show the mathematical formulation behind setting up the bidirectional hidden layer, where $V$ is for vocabulary, $b$ as bias term of weight matrix $W$. The only difference between these two relationships is in the direction of recursing through the corpus.
$$\overrightarrow{h} _t=f(\overrightarrow{W} x_t+\overrightarrow{V}\overrightarrow{h}_{t-1}+\overrightarrow{b})$$
\label{bilstmeq}
$$\overleftarrow{h} _t=f(\overleftarrow{W} x_t+\overleftarrow{V}\overleftarrow{h}_{t-1}+\overleftarrow{b})$$
\label{bilstmeqq}
\subsection{Batch Normalization Layer}
For training very deep neural networks, a technique named as Batch normalization is used that standardizes the inputs to a layer for each mini-batch. The effect of this method is to stabilize the learning process and
to dramatically reduce the number of training epochs required to train deep networks.
The batch normalization method is used to accelerate the training of deep learning neural networks~\cite{ioffe2015batch}.
Batchnorm for short, is proposed as a technique to help coordinate the update of multiple layers in the model. It provides an elegant way to redesign or reparametrizing almost every deep network. The reparametrization eliminates the issue of organizing updates over several layers.
Essentially, when applying this layer with BiLSTM layer, the non-linear activations of the LSTM are removed but not the gate activations. To increase the stability of a neural network, batch normalization normalizes the output of a previous activation layer by subtracting the batch mean and dividing by the batch standard deviation.
\subsection{Prediction module}
In the end a linear activation layer is used as our task is a regression type problem. The final layer of the neural network will have one neuron and the value it returns is a continuous numerical value. Using RELU function as activation layer can be a good option as the output of our task is non-negative real numbers. For loss function, we employ mean square error (MSE), which is defined as follows:
$$ MSE = \frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2 $$,
where
$n$ is the number of samples, $y_{i}$ is the true value of $i$-th data, and $\hat{y_{i}}$ is the predicted value of $i$-th data.
For optimization, we adopt RMSprop (Room Mean Square propagation), as RMSprop is adaptive and fast. In RMSprop, instead of taking cumulative sum of squared gradients it takes the exponential moving average of gradients \cite{Medium}.
\section{Experiments and Evaluations}
\label{sec:ExperimentsEvaluations}
\subsection{Dataset Collection}
\label{ref:datasetCollection}
For Sub-task-1, two training datasets named Humicroedit ~\cite{hossain-etal-2019-president} and Funlines ~\cite{hossain-etal-2020-stimulating} were provided. But only Humicroedits data was used. The training, validation, and test set of the news headline dataset contains the 9653, 2420 and 2025 humicroedits, respectively.
\subsection{Model configuration}
\label{ref:modelConfiguration}
In the following, we describe the set of parameters
that we have used to design our proposed
neural network model. The framework which we used to designed our model was based on TensorFlow~\cite{abadi2016tensorflow} and training of our model is done on a GPU~\cite{owens2008gpu} to capture the benefit from the efficiency
of parallel computation of tensors.
We used the 300-dimensional word2vec Google news pre-trained model, where the maximum length of sentence in given data was 20.
We trained all models with 100 epochs with a batch size of 128 and an initial
learning rate 0.001 by RMSprop optimizer. Unless
otherwise stated, default settings were used for the
other parameters.
\subsection{Evaluation Results}
\label{ref:evaluationMeasure}
To evaluate the performance of the system, the organizers
used different strategies and metrics for
the Sub-tasks 1 and 2~\cite{SemEval2020Task7}. For the Sub-task
1, Root Mean Square (RMSE), RMSE@10, RMSE@20 and RMSE@30 were applied
to estimate the performance of a system. Among three evaluation measures, RMSE was considered as the primary evaluation
measure for this task.
\begin{figure}[h!]
\begin{center}
\centerline{\includegraphics[width=0.6\linewidth]{fig2.pdf}}
\caption{\small Our approach for Sub-task 2.} \label{Figure 2}
\end{center}
\end{figure}
Figure 2 illustrates our approach used for Sub-task 2. In our approach, given two edited headlines, we apply KdeHumor Model from Sub-task 1 to each edited headline, and decides which edited headline is funnier by comparing the predicted values. In Sub-task 2, model Accuracy is used as the evaluation measure. Another auxiliary metric called the Reward is used. For a larger funniness difference between the two headlines in a pair, Reward is higher for a correct classification.
\subsection{Experimental Results }
We now evaluate the performance of our proposed
method in this section.
The summarized results for Sub-task 1 and 2
are presented in Tables 1 and 2, respectively.
\begin{table}[h]
\begin{scriptsize}
\begin{center}
\begin{tabular}{lrl}
\hline \bf Team Name & \bf RMSE & \bf RMSE@10 \\
\hline KdeHumor & 0.6164 & 1.0175 \\ \hline
Farah & 0.5339 &0.9091 \\
HonoMi & 0.4972 & 0.7791 \\
Heidy & 0.6833 & 1.0496 \\
frietz58& 0.7225 & 1.0634 \\
zxchen & 0.5288 & 0.8636 \\
Mjason & 0.5351 & 0.8748 \\
\hline
\end{tabular}
\end{center}
\caption{\label{task-1} (Sub-task 1) KdeHumor results with other selected team. }
\end{scriptsize}
\end{table}
At first, we presented the performance of our
proposed method denoted by team name KdeHumor as well as presenting the performance of randomly
chosen top-ranked participated systems.
\begin{table}[h]
\begin{scriptsize}
\begin{center}
\begin{tabular}{lrl}
\hline \bf Team Name & \bf Accuracy & \bf Reward \\
\hline KdeHumor &0.5190 & 0.0271 \\ \hline
Farah & 0.6088 &0.1840 \\
HonoMi & 0.6742 & 0.2987 \\
HumorAAC & 0.3204 & -0.2177 \\
Heidy& 0.4197 & -.0990 \\
\hline
\end{tabular}
\end{center}
\caption{\label{task-2} (Sub-task 2) KdeHumor results with other selected team. }
\end{scriptsize}
\end{table}
\section{Conclusion}
\label{sec:Conclusion}
In this paper, we presented our approach for
SemEval-2020 Task 7: Assessing Humor in Edited News Headlines. Detection of humor is not an easy task. We tried to tackle the problem by employing deep learning techniques.
We conducted some experiments using different models for example we implemented simple regression model, multilayer perceptron model but the model that gave better result was KdeHumor model.
In future, we are planning to incorporate BERT model for further improvement.
\section*{Acknowledgements}
This work was partially supported by JSPS KAKENHI Grant Number 17H01746.
\bibliographystyle{semeval2020}
|
\section{Introduction}
Stochastic differential equations (SDEs) are widely used to model stochastic systems in different branches of science and industry. Stability and boundedness of the solution are the most popular topics in the area of stochastic systems and control. We refer the reader to \cite{BHB, FP1, CFAF} and references therein. Dynamic systems may not only depend on present states but also the past states. Stochastic delay differential equations (SDDEs) and pantograph stochastic delay differential equations (PSDDEs) are often used to model these systems, whose systems depend on the past state $x(t-\tau)$ and $x(\theta t)$ respectively. The form of these equations are as follows:
\begin{align*}
{\mathord{{\rm d}}} x(t)= f( x(t), x(t-\tau), t){\mathord{{\rm d}}} t +g( x(t),x(t-\tau),t){\mathord{{\rm d}}} B(t),
\end{align*}
and
\begin{align*}
{\mathord{{\rm d}}} x(t)= f( x(t), x(\theta t), t){\mathord{{\rm d}}} t +g( x(t),x(\theta t),t){\mathord{{\rm d}}} B(t),
\end{align*}
where $\tau ,\theta $ are two constants satisfying $\tau>0, 0<\theta <1.$ We here only mention \cite{Pengsg, PSG, KKMM, RX1, LR111}, to name a few. However, there are many practical systems whose future state depends on the states over the whole time interval $[t-\tau, t]$ rather than at times $t-\tau$ and $t.$ Stochastic functional differential equations (SFDEs) have therefore been developed to describe such systems. Generally speaking, SFDEs have the form:
\begin{align*}
&{\mathord{{\rm d}}} x(t)= f( x_t, t){\mathord{{\rm d}}} t +g( x_t, t){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty), \quad x_{t_{0}}=\xi,
\end{align*}
where $x_{t}=\{x( t-\theta), 0\leq \theta \leq \tau)\},$ $\tau>0$ is a constant. As is well known, many scholars realized that numerous system in our real world may experience abrupt changes in their structure and parameters caused by phenomena such as component failures or repairs, changing subsystem interconnections and abrupt environmental disturbances. Hybrid systems driven by continuous-time Markov chains have been used to cope with such situation. Markov chains play the role of stabilizing factor in the stability of hybrid systems. That means, when some subsystems are unstable, but others are stable, then the overall system could be stable because of switching between the subsystems. Since then, the literature on the topic of stability for stochastic differential equations with Markovian switching (SDEswMS) bloomed, both in the direction of obtaining qualitative and quantitative results for the generalized emerging equations and on developing applications which aim to population ecology, network, heat exchanges, etc. For example, \cite{BBG} studied the stability of semi-linear SDEswMS, and \cite{ES}
investigated the following general nonlinear SDEswMS:
\begin{align*}
{\mathord{{\rm d}}} x(t)= f( x(t), t, r(t)){\mathord{{\rm d}}} t +g( x(t),t,r(t)){\mathord{{\rm d}}} B(t),
\end{align*}
where $r(t)$ is a Markov chain taking values in $S=\{1,2,\cdots, N \}.$
Moreover, \cite{MMXX4} applied SDEswMS to solving a control problems, \cite{MMXX5} investigated some complex-valued coupled oscillators, \cite{MMX1} studied the stability of regime-switching jump diffusion systems,
\cite{MMXX6} analyzed asymptotic stability in distribution for such type of equations.
Later, the study of stochastic functional differential equations with Markovian switching (SFDEswMS) (including stochastic delay differential equations with Markovian switching) and PSDDEs with Markovian switching have been also developed rapidly.
Many scholars have enthusiastically studied the stability of such equations and given some applications. For example, \cite{SFM} investigated the exponential stability of highly nonlinear neutral pantograph stochastic differential equations, \cite{SR} built Razumikhin-type theorems on neutral SFDEs, \cite{BJ} studied stability of neutral SFDEswMS driven by G-Brownian motion, \cite{HJsS} analyzed asymptotic stability and boundedness of SFDESwMS. More related work can be seen in \cite{OPZ, glhj, LLM, LR10, AAA1, LR111, MMXX8, MMX21}.
To the best of our knowledge, so far there is little study on hybrid pantograph stochastic functional differential equations (HPSFDEs), in which the $\theta$ changes in interval (0,1], while the $\theta$ is a constant in pantograph stochastic delay differential equations. Inspired by the works of the above articles, we aim in this paper to study several moment properties and sample properties of the solutions such as the moment exponential stability, almost surely exponential stability and almost sure polynomial stability, etc. for HPSFDEs.
We close this part by giving our organization in this article. In Section 2, we introduce some necessary notations. In Section 3, we give our main results on the moment properties and sample properties of analytical solution. Several examples are also given to illustrate the theory.
\section{Preliminaries}
\subsection{Notations}Throughout this paper, Let $(\Omega, {\mathscr F}, \{{\mathscr F}_{t}\}_{t\geq 0}, P)$ be a complete probability space satisfying the usual conditions(i.e., it is increasing and right continuous with ${\mathscr F}_{0}$ contains all $P$-null sets) taking along
a standard $d$-Brownian motion process $B(t).$ For $x, y \in {\mathbb R}^{n},$ we use $|x |$ to denote the Euclidean norm of
$x,$ and use $\langle x, y\rangle$ or $x^{T}y$ to denote the Euclidean inner product. If $A$ is a matrix, $A^{T}$ is the transpose of $A$ and $|A |$ represents $\sqrt{\mathrm{Tr} (AA^{T}).}$ Let $\lfloor a \rfloor$ be the integer parts of $a.$ Moreover, for $0<\underline{\theta} <1,$ denote by $\mathscr{C} :=\mathscr{C}([\underline{\theta}, 1]; {\mathbb R}^{n})$ the family of all continuous ${\mathbb R}^{n}-$valued functions $\varphi$ defined on $[\underline{\theta}, 1]$ with the norm $\| \varphi\|=\sup_{\underline{\theta} \leq t\leq 1}|\varphi(t)|.$ Let $t_0>0$ and $h: [t_0, \infty)\to {\mathbb R}^{n}$ be a continuous function, for $t \ge t_0$ denote $h_t(\theta)=h(\theta t), \underline{\theta} \le \theta \le 1.$ One can see that $h_t(\cdot)\in \mathscr{C}.$
Let $r(t)$ be a continuous-time Markov chain taking values in $S=\{1,2,\cdots, N\}$ with the generator $\Gamma=(\gamma_{ij})_{N\times N}$ such that
\begin{align*}
P\{r(t+\delta)=j|r(t)=i\}=
\begin{cases}
\gamma_{ij}\delta + o(\delta), & i\neq j, \\
1+\gamma_{ii}\delta + o(\delta), & i=j,
\end{cases}
\end{align*}
where $\delta > 0.$ Here $\gamma_{ij} $ is the transition rates from $i$ to $j$ and $\gamma_{ij}\ge 0$ if $i\neq j$ while $\gamma_{ii}=-\sum_{j\neq i}\gamma_{ij}.$ It is well know that almost every sample path of $r(t)$ is a right-continuous step functions with finite number of sample jumps in any finite subinterval of ${\mathbb R}_{+}=[0, \infty).$ Assume that Markov chain $r(t)$ is independent of Brownian motion.
Denote by $ C^{1,2}([t_0,+\infty)\times R^{n} \times S; [0,+\infty))$ the
family of all continuous nonnegative functions $V(t,x, i)$ defined on $[t_0,+\infty)\times R^{n} \times S $,
such that for each $i\in S$,
they are continuously once differentiable in $t$ and twice in $x$.
\section{Main Results}
Consider the following HPSFDE:
\begin{equation}\label{eq1}
\begin{cases}
&{\mathord{{\rm d}}} x(t)= f( x_t, t,r(t)){\mathord{{\rm d}}} t +g( x_t, t, r(t)){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty),\\
& x(t)=\xi(t), t\in [\underline{\theta} t_{0}, t_{0}],
\end{cases}
\end{equation}
where $x_{t}=\{x(\theta t), \underline{\theta}\leq \theta \leq 1)\}$ and $0< \underline{\theta}<1$ is a constant. We would like to point out that $x_{t}\in {\mathscr{C}}$ is a segment process and $x_t(\theta)=x(\theta t)$ while $x(t)\in R^n$ is a point.
Given $V\in C^{1,2}( R^{n} \times [t_0,+\infty)\times S; [0,+\infty)), \varphi\in {\mathscr{C}}, i\in S$,
we define an operator $LV: \mathscr{C}\times [t_{0}, \infty)\times S \rightarrow {\mathbb R} $ by
\begin{align*}
LV(\varphi,t,i)&=V_{t}(\varphi(1),t,i)+ V_{x}(\varphi(1),t,i)f(\varphi,t,i)
&+\frac{1}{2}{\rm trace}(g^{T}(\varphi,t,i)V_{xx}(\varphi(1),t,i)g(\varphi,t,i))\\
&+\sum^{N}_{l=1}\gamma_{il}V(\varphi(1),t,l),
\end{align*}
where
\begin{align*}
V_t(x, t, i)=\left(\frac{\partial V(x, t, i)}{\partial t}\right), \, V_x(x, t, i)=\left(\frac{\partial V(x, t, i)}{\partial x_1}, \ldots, \frac{\partial V(x, t, i)}{\partial x_n}\right)
\end{align*}
and
\begin{align*}
V_{xx}(x, t, i)=\left(\frac{\partial^2 V(x, t, i)}{\partial x_k \partial x_l}\right)_{kl}.
\end{align*}
We have the following the corresponding It\^{o}'s formula for hybrid system \eqref{eq1}:
\\
\begin{align*}
&V(x(t),t,r(t))
=V(x(0),0,r(0))+ \int^{t}_{t_0}LV(x_s, s,r(s)){\mathord{{\rm d}}} s\\
&+ \int^{t}_{t_0}V_{x}(x(s),s,r(s))g(x_{s},s,r(s)){\mathord{{\rm d}}} B(s).
\end{align*}
The following assumptions are needed.
\begin{itemize}
\item[(H1)] For any $\varphi, \varphi' \in C([\underline{\theta}, 1]; {\mathbb R}^{n})$ satisfying $\|\varphi\|\vee \|\varphi'\|\leq R,$ there exists a positive constant $C_{R}$ such that
\begin{align*}
|f( \varphi, t,i) - f( \varphi^{\prime},t,i)|\vee|g( \varphi, t,i) - g( \varphi^{\prime},t,i)|
\leq C_{R}\|\varphi- \varphi^{\prime}\|.
\end{align*}
\item[(H2)] There exist functions $V\in C^{2,1}({\mathbb R}^{n}\times [t_{0}, \infty)\times S; {\mathbb R}_{+}),$ $U_{0}, U_{k}\in C^{2,1}({\mathbb R}^{n}\times [t_{0}, \infty); {\mathbb R}_{+}),$ and probability measures $\nu_{k}$ on $[\underline{\theta}, 1]$, and non-negative constants $a_{0}, a_{k}, b_{kl} ,$ $ k=1,2,\cdots, M,$ $ l=1,2,\cdots, l_{k} $ such that
\begin{align}
\lim_{|x|\rightarrow \infty}\inf_{t_{0}\leq t< \infty}U_{0}(x,t)=\infty,
\end{align}
\begin{align}
U_{0}(x,t)\leq V(x,t,i)\leq U_{1}(x,t), \forall (x,t,i)\in {\mathbb R}^{n}\times {\mathbb R}^{+}\times S,
\end{align}
\begin{align}\label{L0}
LV(\varphi,t,i)&\leq a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(\varphi(1), t)\nonumber\\
&+ \sum^{l_{k}}_{l=1}b_{kl}\int^{1}_{\underline{\theta}} e^{-\int^{t}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u}U_{k}^{\alpha_{kl}}(\varphi(1), t) U_{k}^{1-\alpha_{kl}}(\varphi(\theta), \theta t){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.,
\end{align}
where function $\lambda(\cdot, \cdot):[\underline{\theta}, 1]\times {\mathbb R}_{+}\rightarrow {\mathbb R}_{+}$ satisfying $\inf_{0 \leq s < \infty}\lambda (\theta, s)\geq \beta (1-\theta)$ and $\alpha_{kl}, \beta$ are constants satisfying $0\leq\alpha_{kl}\leq 1, 0<\beta<a_{1}$.
\end{itemize}
\subsection{Existence and Uniqueness}
In the same way as in \cite{LR1211}, we can show that \eqref{eq1} has a unique local solution $x(t), t\in [t_0, \sigma_{\infty})$ under $(\mathrm{H1}),$ where $\sigma_{\infty}$ is the explosion time. The following condition (H2') will guarantee a global solution to \eqref{eq1}, that is
\begin{itemize}
\item[(H2')] Assume that (H2) holds, but \eqref{L0} is replaced by
\begin{align}
LV(\varphi,t,i)&\leq a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(\varphi(1), t)\nonumber\\
&+ \sum^{l_{k}}_{l=1}b_{kl}\int^{1}_{\underline{\theta}} U_{k}^{\alpha_{kl}}(\varphi(1), t) U_{k}^{1-\alpha_{kl}}(\varphi(\theta), \theta t){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.,
\end{align}
where $0\leq \alpha_{kl}\leq 1$ are constants .
\end{itemize}
\begin{thm}
Assume that (\rm{H1}) and (\rm{H2'}) hold. If
\begin{align}\label{eu1}
-a_{k}+ \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}
+\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\le 0, \, k=1,2,\cdots, M,
\end{align}
then the equation \eqref{eq1} has a unique global solution.
\end{thm}
\begin{proof}
Let $x(t), t\in [t_0, \sigma_{\infty})$ be the unique local solution and $\sigma_{a}=\inf\{t\geq t_0: |x(t)|\geq a\}.$
Using It\^{o}'s formula and taking the expectation, we have
\begin{align}
&{\mathbb E}[V(x(t\wedge \sigma_a), t\wedge \sigma_{a}, r(t\wedge \sigma_a))]={\mathbb E}[V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}LV(x_{s},s,r(s)){\mathord{{\rm d}}} s \nonumber\\
& \leq {\mathbb E}[V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)\nonumber\\
& \quad \quad \quad \quad \quad \quad\quad \quad \quad+ \sum^{l_{k}}_{l=1}b_{kl}\int^{1}_{\underline{\theta}} U_{k}^{\alpha_{kl}}(x(s), s) U_{k}^{1-\alpha_{kl}}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s\nonumber\\
&\leq {\mathbb E}[V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)\nonumber\\
& \quad \quad + \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}U_{k}(x(s), s)+ \sum^{l_{k}}_{l=1}b_{kl}(1-\alpha_{kl})\int^{1}_{\underline{\theta}} U_{k}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s.
\end{align}
Noting that
\begin{equation*}
\begin{split}
&\int^{t\wedge \sigma_a}_{t_{0}}\int^{1}_{\underline{\theta}}U_k(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_k(\theta){\mathord{{\rm d}}} s =\int^{1}_{\underline{\theta}}\int^{t\wedge \sigma_a}_{t_{0}}U_k(x(\theta s), \theta s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_k(\theta)\\
&\le \frac{1}{{\underline{\theta}}}\int^{1}_{\underline{\theta}}\int^{\theta(t\wedge \sigma_a)}_{\theta t_{0}}U_k(x( s), s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_k(\theta)\\
&\le \frac{1}{{\underline{\theta}}}\int^{1}_{\underline{\theta}}\int^{t\wedge \sigma_a}_{\underline{\theta} t_{0}}U_k(x( s), s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_k(\theta)\\
&\le \frac{1}{{\underline{\theta}}}\int^{1}_{\underline{\theta}}\int^{t\wedge \sigma_a}_{t_{0}}U_k(x( s), s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_k(\theta)+ \frac{1}{{\underline{\theta}}}\int^{1}_{\underline{\theta}}\int^{t_0}_{\underline{\theta} t_{0}}U_k(x( s), s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_k(\theta)\\
&\le \frac{1}{{\underline{\theta}}}\int^{t\wedge \sigma_a}_{t_{0}}U_k(x( s), s){\mathord{{\rm d}}} s + \frac{1}{{\underline{\theta}}}\int^{t_0}_{\underline{\theta} t_{0}}U_k(x( s), s){\mathord{{\rm d}}} s,
\end{split}
\end{equation*}
and \eqref{eu1}, one can see that
\begin{align}\label{3.8}
&{\mathbb E}[V(x(t\wedge \sigma_a), t\wedge \sigma_{a}, r(t\wedge \sigma_a))]\nonumber\\
&\leq {\mathbb E}[V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)\nonumber\\
& \quad \quad \quad\quad \quad + \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl} U_{k}(x(s), s)+ \sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})U_{k}(x( s), s) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s\nonumber\\
& + \sum^{M}_{k=1}\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\int^{t_{0}}_{\underline{\theta}t_{0}}U_{k}(x( s), s){\mathord{{\rm d}}} s \nonumber\\
&\leq {\mathbb E}[V(x(t_0), t_0, r(t_0))] \nonumber\\
&+ {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg(\bigg.-a_{k}+ \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}
+\sum_{k=1}^{l_{k}} b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\bigg)\bigg.U_{k}(x(s), s)\bigg\}\bigg.{\mathord{{\rm d}}} s\nonumber\\
& + \sum^{M}_{k=1}\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\int^{t_{0}}_{\underline{\theta}t_{0}}{\mathbb E} U_{k}(\xi( s), s){\mathord{{\rm d}}} s \nonumber\\
&\leq c_{0} +a_{0}t,
\end{align}
where $c_{0}={\mathbb E}[V(x(t_0), t_0, r(t_0))]+ \sum^{M}_{k=1} \sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\int^{t_{0}}_{\underline{\theta}t_{0}}{\mathbb E} U_{k}(\xi( s), s){\mathord{{\rm d}}} s .$
Setting $\mu_{a}= \inf_{|x|\geq a, t_{0}\leq t< \infty}U_{0}(x, t),$ we then have
\begin{align*}
{\mathbb E}[U_{0}(x(t\wedge \sigma_a), t\wedge \sigma_a)]\geq {\mathbb E}[U_{0}(x( \sigma_a), \sigma_a)1_{\sigma_{a}\leq t}]\geq \mu_{a}P(\sigma_{a}\leq t).
\end{align*}
This immediately implies
\begin{align*}
P(\sigma_{\infty}\leq t)=\lim_{a\rightarrow \infty} P(\sigma_{a}\leq t)\leq \lim_{a\rightarrow \infty} \frac{{\mathbb E}[U_{0}(x(t\wedge \sigma_a), t\wedge \sigma_{a}]}{\mu_{a}}=\lim_{a\rightarrow \infty}\frac{c_{0} +a_{0}t}{\mu_{a}}=0.
\end{align*}
Therefore, $\sigma_{\infty}=\infty,\, a.s.,$ and there exists unique global solution $x(t)$ on $[t_{0}, \infty).$
\end{proof}
\subsection{Exponential Stability}
In this subsection, we will investigate several moment properties and sample properties of the solutions to the equations such as the moment exponential stability, almost sure exponential stability, etc. Before studying the stability of the solution to E.q.\eqref{eq1}, we present a semi-martingale convergence theorem which can be found in \cite{LSLS}.
\begin{lem}
Let $A_{1}(t), A_{2}(t)$ be two continuous adapted increasing processes on $t\geq 0$ with $A_{1}(0)=0, A_{2}(0)=0,$ a.s., M(t) a real-valued continuous local martingale with $M(0)=0,$ a.s., $\xi$ a nonnegative ${\mathscr F}_{0}-$measurable random variable such that ${\mathbb E}[\xi]< \infty.$ Set $X(t)=\xi+A_{1}(t)-A_{2}(t)+M(t), t\geq 0.$ If $X(t)$ is nonnegative , then we have the following results:
$$\{\lim_{t\rightarrow \infty}A_{1}(t)<\infty\}\subset \{\lim_{t\rightarrow \infty}A_{2}(t)<\infty\}\cap \{\lim_{t\rightarrow \infty}X(t)<\infty\},\,a.s.,$$
where $C\subset D,\, a.s.$ means $P(C\cap D^{c})=0.$ In particular, if $\lim_{t\rightarrow \infty}A_{1}(t)<\infty,\,a.s.,$ then, with probability one,
$$\lim_{t\rightarrow \infty}A_{2}(t)<\infty, \, \lim_{t\rightarrow \infty}X(t)<\infty, \, -\infty <\lim_{t\rightarrow \infty}M(t)<\infty ,\,a.s.$$
\end{lem}
\begin{thm}\label{T3.3}
Assume that $(\mathrm{\mathrm{H}1})-(\mathrm{H2})$ hold with $$-a_{k}+ \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}
+\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})<0, k=1,2,\cdots, M.$$ We then have the following results:
\begin{itemize}
\item[{\rm (i)}] $\limsup_{t\rightarrow \infty}{\mathbb E}[U_{0}(x(t_{0},t,\xi,i_{0}), t)]\leq \frac{a_{0}}{\varepsilon},$
where $0<\varepsilon \leq \beta$ is a constant satisfying $$a_{1}-\varepsilon- \sum^{l_{1}}_{l=1}b_{1l}\alpha_{1l}- \sum^{l_{1}}_{l=1}b_{1l}\frac{1}{\underline{\theta}}(1-\alpha_{1l})>0.$$\\
\item[{\rm (ii)}] \begin{align*}
&\limsup_{t\rightarrow \infty}\frac{1}{t}\int^{t}_{t_{0}}{\mathbb E}[U_{k}(x(t_{0},s,\xi,i_{0}), s)]{\mathord{{\rm d}}} s \\
&\leq \frac{a_{0}}{a_{k}-\sum^{l_{k}}_{l=1}b_{kl}e^{-\beta(1-\underline{\theta})t_{0}}\alpha_{kl}
-\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}e^{-\beta(1-\underline{\theta})t_{0}}(1-\alpha_{kl})}, \quad
k=1,2,\cdots M.
\end{align*}
\item[{\rm (iii)}] If $a_{0}=0, $ then the global solution $x(t_{0},t,\xi,i_{0})$ is exponentially stable in moment and almost surely exponential stable, i.e.
\begin{align}\label{3c}
\limsup_{t\rightarrow \infty}\frac{1}{t}\log ({\mathbb E}[U_{0}(x(t_{0},t,\xi,i_{0}), t)]) \leq -\varepsilon.
\end{align}
\begin{align}\label{3c1}
\limsup_{t\rightarrow \infty}\frac{1}{t}\log (U_{0}(x(t_{0},t,\xi,i_{0}), t)) \leq -\varepsilon,\,a.s.,
\end{align}
where $\varepsilon$ satisfies the condition in (i).
\end{itemize}
\end{thm}
\begin{proof}
(i) Using It\^{o}'s formula to $e^{\varepsilon t}V(x(t), t, r(t))$, we have
\begin{equation}\label{f1}
\begin{split}
&{\mathbb E}[e^{\varepsilon (t\wedge \sigma_{a})}V(x(t\wedge \sigma_a), t\wedge \sigma_a, r(t\wedge \sigma_a))]\\
&={\mathbb E}[e^{\varepsilon t_{0}}V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}e^{\varepsilon s}(\varepsilon V(x_{s},s,r(s)) +LV(x_{s},s,r(s))){\mathord{{\rm d}}} s \\
& \leq {\mathbb E}[e^{\varepsilon t_{0}}V(x(t_0), t_0, r(t_0))] \\
& \quad \quad + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}e^{\varepsilon s}\bigg\{\bigg. a_{0}-(a_{1}-\varepsilon)U_{1}(x(s), s)+\bigg[\bigg.-\sum^{M}_{k=2}a_{k}U_{k}(x(s), s)\\
& \quad \quad+\sum^{M}_{k=1} \sum^{l_{k}}_{l=1}b_{kl}\int^{1}_{\underline{\theta}} e^{-\int^{s}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u}U_{k}^{\alpha_{kl}}(x(s), s) U_{k}^{1-\alpha_{kl}}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s.
\end{split}
\end{equation}
Now, we compute
\begin{align}\label{3.13}
&b_{kl}\int^{t\wedge \sigma_{a}}_{t_{0}}\int^{1}_{\underline{\theta}} e^{\varepsilon s-\int^{s}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u}U_{k}^{\alpha_{kl}}(x(s), s) U_{k}^{1-\alpha_{kl}}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta){\mathord{{\rm d}}} s\nonumber\\
& \leq b_{kl}\alpha_{kl} \int^{t\wedge \sigma_{a}}_{t_{0}}\int^{1}_{\underline{\theta}}e^{\varepsilon s-\int^{s}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u}U_{k}(x(s), s) {\mathord{{\rm d}}} \nu_{k}(\theta) {\mathord{{\rm d}}} s\nonumber\\
&+b_{kl}(1-\alpha_{kl})\int^{t\wedge \sigma_{a}}_{t_{0}}\int^{1}_{\underline{\theta}} e^{\varepsilon s-\int^{s}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u} U_{k}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) {\mathord{{\rm d}}} s\nonumber\\
& \leq b_{kl}\alpha_{kl}\int^{1}_{\underline{\theta}}e^{-\beta(1-\theta)t_{0}}{\mathord{{\rm d}}} \nu_{k}(\theta) \int^{t\wedge \sigma_{a}}_{t_{0}}e^{\varepsilon s}U_{k}(x(s), s) {\mathord{{\rm d}}} s\nonumber\\
&+b_{kl}(1-\alpha_{kl})\int^{t\wedge \sigma_{a}}_{t_{0}}\int^{1}_{\underline{\theta}} e^{\varepsilon s-\int^{s}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u} U_{k}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) {\mathord{{\rm d}}} s\nonumber\\
&\leq b_{kl}\alpha_{kl}\int^{t\wedge \sigma_{a}}_{t_{0}}e^{\varepsilon s}U_{k}(x(s), s) {\mathord{{\rm d}}} s+b_{kl}(1-\alpha_{kl})\int^{1}_{\underline{\theta}}\int^{t\wedge \sigma_{a}}_{t_{0}} e^{\varepsilon s-\int^{s}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u} U_{k}(x(\theta s), \theta s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_{k}(\theta) \nonumber\\
&\leq b_{kl}\alpha_{kl}\int^{t\wedge \sigma_{a}}_{t_{0}}e^{\varepsilon s}U_{k}(x(s), s) {\mathord{{\rm d}}} s+b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\int^{1}_{\underline{\theta}}\int^{t\wedge \sigma_{a}}_{\underline{\theta}t_{0}} e^{\frac{\varepsilon }{\theta}s-\int^{\frac{s}{\theta}}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u} U_{k}(x(s), s){\mathord{{\rm d}}} s{\mathord{{\rm d}}} \nu_{k}(\theta) \nonumber\\
&\leq b_{kl}\alpha_{kl}\int^{t\wedge \sigma_{a}}_{t_{0}}e^{\varepsilon s}U_{k}(x(s), s) {\mathord{{\rm d}}} s+ b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\int^{t\wedge \sigma_{a}}_{t_{0}} e^{\varepsilon s} U_{k}(x( s), s){\mathord{{\rm d}}} s \nonumber\\
& + b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\int^{t_{0}}_{\underline{\theta}t_{0}} e^{\varepsilon s} U_{k}(x( s), s){\mathord{{\rm d}}} s.
\end{align}
This, together with \eqref{f1}, yields that
\begin{equation}\label{f2}
\begin{split}
&{\mathbb E}[e^{\varepsilon (t\wedge \sigma_{a})}V(x(t\wedge \sigma_a), t\wedge \sigma_a, r(t\wedge \sigma_a))]\\
& \leq \bar{c}_{0}
+ {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}e^{\varepsilon s}\bigg\{\bigg. a_{0}-\bigg(\bigg.a_{1}-\varepsilon- \sum^{l_{1}}_{l=1}b_{1l}\alpha_{1l}- \sum^{l_{1}}_{l=1}b_{1l}\frac{1}{\underline{\theta}}(1-\alpha_{1l})\bigg)\bigg.U_{1}(x(s), s)\\
&\quad \quad + \sum^{M}_{k=2}\bigg(\bigg.-a_{k}+\sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}+\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})\bigg)\bigg.U_{k}(x(s), s)\bigg\}\bigg. {\mathord{{\rm d}}} s \\
& \leq \bar{c}_{0} + \frac{a_{0}}{\varepsilon}e^{\varepsilon t},
\end{split}
\end{equation}
where $\bar{c}_{0}={\mathbb E}[e^{\varepsilon t_{0}}V(x(t_0), t_0, r(t_0))]+ \sum^{M}_{k=1}\sum^{l_{k}}_{l=1}{\mathbb E}\int^{t_0}_{\underline{\theta}t_{0}}e^{\varepsilon s}\frac{1}{\underline{\theta}}b_{kl}(1-\alpha_{kl})U_{k}(\xi(s), s) {\mathord{{\rm d}}} s.$
Letting $a \rightarrow \infty,$ it leads to
\begin{align*}
{\mathbb E}[e^{\varepsilon t}U_{0}(x(t), t, r(t))]
\leq \bar{c}_{0} + \frac{a_{0}}{\varepsilon}e^{\varepsilon t}.
\end{align*}
The assertion (i) follows by letting $t \rightarrow \infty$.
(ii) Similar to the proofs of \eqref{3.8} and \eqref{3.13}, we can show that
\begin{equation}
\begin{split}
&{\mathbb E}[V(x(t\wedge \sigma_a), t\wedge \sigma_{a}, r(t\wedge \sigma_a))]={\mathbb E}[V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}LV(x_{s},s,r(s)){\mathord{{\rm d}}} s \\
& \leq {\mathbb E}[V(x(t_0), t_0, r(t_0))] + {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)\\
& \quad \quad \quad+ \sum^{l_{k}}_{l=1}b_{kl}\int^{1}_{\underline{\theta}} e^{-\int^{t}_{0}\lambda(\theta, u){\mathord{{\rm d}}} u}U_{k}^{\alpha_{kl}}(x(s), s) U_{k}^{1-\alpha_{kl}}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s\\
&\leq {\mathbb E}[V(x(t_0), t_0, r(t_0))] \\
&+ {\mathbb E}\int^{t\wedge \sigma_a}_{t_{0}}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg(\bigg.-a_{k}+ \sum^{l_{k}}_{l=1}b_{kl}e^{-\beta(1-\underline{\theta})t_{0}}\alpha_{kl}\\
& \quad \quad \quad \quad\quad \quad \quad \quad\quad \quad \quad \quad+\sum_{k=1}^{l_{k}} b_{kl}\frac{1}{\underline{\theta}}e^{-\beta(1-\underline{\theta})t_{0}}(1-\alpha_{kl})\bigg)\bigg.U_{k}(x(s), s)\bigg\}\bigg.{\mathord{{\rm d}}} s\\
& + \sum^{M}_{k=1}\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}e^{-\beta(1-\underline{\theta})t_{0}}(1-\alpha_{kl})\int^{t_{0}}_{\underline{\theta}t_{0}}U_{k}(\xi( s), s){\mathord{{\rm d}}} s \\
&\leq c_{0} +a_{0}t.
\end{split}
\end{equation}
Letting $a\rightarrow \infty,$ we obtain
\begin{align}\label{3.16}
&\sum^{M}_{k=1}\bigg(\bigg.a_{k}-\sum^{l_{k}}_{l=1}b_{kl}e^{-\beta(1-\underline{\theta})t_{0}}\alpha_{kl}
-\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}e^{-\beta(1-\underline{\theta})t_{0}}(1-\alpha_{kl})\bigg)\bigg.{\mathbb E}\int^{t}_{t_0}U_{k}(x(s), s){\mathord{{\rm d}}} s\nonumber\\
&\leq c_{0} +a_{0}t.
\end{align}
This means that assertion (ii) holds.
(iii) Since $a_0=0,$ we derive from \eqref{f2} that
\begin{align*}
{\mathbb E}[e^{\varepsilon t}U_{0}(x(t), t\wedge a, r(t))]
\leq \bar{c}_{0}.
\end{align*}
This implies that \eqref{3c} holds.
Using the similar method in (i) without taking the expectation, we can show that
\begin{align}
e^{\varepsilon t}U_{0}(x(t), t)\leq \bar{c}_{0} + M(t),
\end{align}
where $M(t)= \int^{t}_{0}e^{\varepsilon t}V_{x}(x(s),s, r(s))g(x_{s},t, r(s)){\mathord{{\rm d}}} B(s).$
Due to Lemma 3.2, it follows that
$$\limsup_{t\rightarrow \infty}e^{\varepsilon t}U_{0}(x(t), t)< \infty,\,a.s.$$
Thus, there exists a finite positive random variable $\eta$ such that
$$ \sup_{t_{0}\leq t <\infty}e^{\varepsilon t}U_{0}(x(t), t)<\eta,\,a.s.$$
Thus, the proof of (\ref{3c1}) is complete.
\end{proof}
We now illustrate the theoretical results in Theorem \ref{T3.3} by the following example.
\begin{exa} {\rm
Let $\nu(\cdot)$ be a probability measure on $[\underline{\theta}, 1]$. Set $ S=\{1,2\}, \beta=0.5, \lambda(\theta, t)=0.5(1-\theta), \underline{\theta}=0.5$, $d=1$. Let $r(t)$ be a Markov chain with generator
$$
\Gamma=\left(
\begin{array}{ccc}
-1 & 1 \\
2 & -2 \\
\end{array}
\right).
$$
Consider the following equation:
\begin{align}
&{\mathord{{\rm d}}} x(t)= f( x_t, t,r(t)){\mathord{{\rm d}}} t +g( x_t, t, r(t)){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty)\nonumber\\
& x(t)=\xi(t), t\in [\underline{\theta} t_{0}, t_{0}],
\end{align}
where for $\varphi \in {\mathscr{C}}$
\begin{align*}
f( \varphi, t,i)=
\begin{cases}
-5(\varphi(1)+\varphi^{3}(1)+\varphi^{5}(1)) +0.5\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta), i=1,\\
0.05\varphi(1) +0.05\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta), i=2;
\end{cases}
\end{align*}
and
\begin{align*}
g(\varphi, t,i)
=\begin{cases}
0.5\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(1)|^{2}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta), i=1,\\
0.2\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta), i=2.
\end{cases}
\end{align*}
Define
\begin{align*}
V( x, t,i)
=\begin{cases}
x^{2}, \quad i=1,\\
2(x^{2}+x^{6}), \quad i=2.
\end{cases}
\end{align*}
When $i=1, $ it follows that
\begin{align*}
&LV(\varphi, t,1)=2\varphi(1) f(\varphi, t,1)+|g(\varphi, t,1)|^{2}+ \sum^{2}_{j=1}\gamma_{1j}V(\varphi(1), t,j)\\
&\leq 2\varphi(1)[-5(\varphi(1)+\varphi^{3}(1)+\varphi^{5}(1))] +\varphi(1)\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta)\\
&\quad \quad \quad+0.25\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(1)|^{4}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)-|\varphi(1)|^{2}+2(|\varphi(1)|^{2}+|\varphi(1)|^{6})\\
&\leq -9|\varphi(1)|^{2}-10|\varphi(1)|^{4}-8|\varphi(1)|^{6}+\varphi(1)\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta)\\
&\quad \quad \quad+0.25\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(1)|^{4}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta).
\end{align*}
When $i=2, $ we have
\begin{align*}
&LV(\varphi, t,2)=(4\varphi(1)+12\varphi^{5}(1))f(\varphi, t,2)+0.02(4+60\varphi^{4}(1))\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)\\
&+ \sum^{2}_{j=1}\gamma_{2j}V(\varphi(1), t,j)\\
& \leq 0.2|\varphi(1)|^{2}+0.6|\varphi(1)|^{6}+0.2\varphi(1)\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta)\\
&+0.6\varphi^{5}(1)\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta)+0.08\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)\\
&+1.2\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(t)|^{4}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)+2|\varphi(1)|^{2}-4(|\varphi(1)|^{2}+|\varphi(1)|^{6})\\
& \leq -1.8|\varphi(1)|^{2}-3.4|\varphi(1)|^{6}\\
&+0.2\varphi(1)\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta)+0.6\varphi^{5}(1)\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu (\theta)\\
&+0.08\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)+1.2\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}|\varphi(1)|^{4}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)\\
& \leq -1.8|\varphi(1)|^{2}-3.4|\varphi(1)|^{6}\\
&+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}0.2(|\varphi(1)|^{2})^{\frac{1}{2}}(|\varphi(\theta )|^{2})^{\frac{1}{2}}{\mathord{{\rm d}}} \nu (\theta)+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}0.6(|\varphi(1)|^{6})^{\frac{5}{6}}(|\varphi(\theta )|^{6})^{\frac{1}{6}}{\mathord{{\rm d}}} \nu (\theta)\\
&+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}0.08|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu (\theta)+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}1.2(|\varphi(1)|^{6})^{\frac{4}{6}}(|\varphi(\theta )|^{6})^{\frac{2}{6}}{\mathord{{\rm d}}} \nu (\theta)\\
& \leq -1.8|\varphi(1)|^{2}-3.4|\varphi(1)|^{6}\\
&+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}[0.2(|\varphi(1)|^{2})^{\frac{1}{2}}(|\varphi(\theta )|^{2})^{\frac{1}{2}}+0.08|\varphi(\theta )|^{2}]{\mathord{{\rm d}}} \nu (\theta)\\
&+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}[0.6(|\varphi(1)|^{6})^{\frac{5}{6}}(|\varphi(\theta )|^{6})^{\frac{1}{6}}+1.2(|\varphi(1)|^{6})^{\frac{4}{6}}(|\varphi(\theta )|^{6})^{\frac{2}{6}}]{\mathord{{\rm d}}} \nu (\theta).
\end{align*}
Then, we have
\begin{align*}
LV(\varphi, t,i) &\leq -1.8|\varphi(1)|^{2}-3.4|\varphi(1)|^{6}\\
&+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}[(|\varphi(1)|^{2})^{\frac{1}{2}}(|\varphi(\theta )|^{2})^{\frac{1}{2}}+0.08|\varphi(\theta )|^{2}]{\mathord{{\rm d}}} \nu (\theta)\\
&+\int^{1}_{\frac{1}{2}}e^{-0.5(1-\theta) t}[0.6(|\varphi(1)|^{6})^{\frac{5}{6}}(|\varphi(\theta )|^{6})^{\frac{1}{6}}+1.2(|\varphi(1)|^{6})^{\frac{4}{6}}(|\varphi(\theta )|^{6})^{\frac{2}{6}}]{\mathord{{\rm d}}} \nu (\theta).
\end{align*}
Obviously, we can choose
\begin{align*}
&U_{0}(x,t)=|x|^{2},U_{1}(x,t)=|x|^{2}, U_{2}(x,t)=|x|^{6}, a_{0}=0, a_{1}=1.8,\\
& a_{2}=3.4,, b_{11}=1, b_{12}=0.08,
b_{21}=0.6, b_{22}=1.2.
\end{align*}
From Theorem 3.3, we could know that the following results hold.
\begin{enumerate}
\item[(i)] $$\limsup_{t\rightarrow \infty}{\mathbb E}[|x(t_{0},t,\xi,i_{0})|^{2}]=0,$$
\item[(ii)] $$\limsup_{t\rightarrow \infty}\frac{1}{t}\int^{t}_{t_{0}}{\mathbb E}[|x(t_{0},s,\xi,i_{0})|^{2}]{\mathord{{\rm d}}} s =0,$$
$$\limsup_{t\rightarrow \infty}\frac{1}{t}\int^{t}_{t_{0}}{\mathbb E}[|x(t_{0},s,\xi,i_{0})|^{6}]{\mathord{{\rm d}}} s =0,$$
\item[(iii)] \begin{align*}
\limsup_{t\rightarrow \infty}\frac{1}{t}\log ({\mathbb E}[|x(t_{0},t,\xi,i_{0})|^{2}]) \leq -0.05.\nonumber
\end{align*}
\begin{align*}
\limsup_{t\rightarrow \infty}\frac{1}{t}\log (|x(t_{0},t,\xi,i_{0})|^{2}) \leq -0.05,\,a.s.
\end{align*}
\end{enumerate}
}
\end{exa}
Now, we give the second example.
\begin{exa} {\rm
Let $\nu_{1}(\cdot)$ be a probability measure on $[\underline{\theta}, 1]$ and $\nu_{2}(\cdot)=\delta_{1}(\cdot).$ Set $S=\{1,2\}, \beta=0.6, \lambda(\theta, t)=0.6(1-\theta), \underline{\theta}=0.7, d=1$. Let $r(t)$ be a Markov chain with generator
$$
\Gamma=\left(
\begin{array}{ccc}
-1 & 1 \\
3 & -3 \\
\end{array}
\right).
$$
Consider the following equation:
\begin{align}
&{\mathord{{\rm d}}} x(t)= f( x_t, t,r(t)){\mathord{{\rm d}}} t +g( x_t, t, r(t)){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty)\nonumber\\
& x(t)=\xi(t), t\in [\underline{\theta} t_{0}, t_{0}],
\end{align}
where for $\varphi\in {\mathscr{C}}$
\begin{align*}
f( \varphi, t,i)=
\begin{cases}
-6(\varphi(1)+\varphi^{3}(1)+\varphi^{7}(1)) +\int^{1}_{0.7}e^{-0.6(1-\theta) t}\varphi(\theta ){\mathord{{\rm d}}} \nu_{1} (\theta), i=1,\\
0.04\varphi(1) +0.04\int^{1}_{0.7}e^{-0.6(1-\theta) t}\varphi(\theta ){\mathord{{\rm d}}} \nu_{2} (\theta), i=2;
\end{cases}
\end{align*}
and
\begin{align*}
g( \varphi, t,i)
=\begin{cases}
0.5\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(1)|^{2}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu_{1} (\theta), i=1,\\
0.1\int^{1}_{0.7}e^{-0.6(1-\theta) t}\varphi(\theta ){\mathord{{\rm d}}} \nu_{2} (\theta), i=2.
\end{cases}
\end{align*}
From above equation, when $r(t)=2, $ by the definition of $\nu_2$ the equation becomes
\begin{align}
&{\mathord{{\rm d}}} x(t)= 0.08x(t){\mathord{{\rm d}}} t +0.1x(t){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty)\nonumber\\
& x(t)=\xi(t), t\in [\underline{\theta} t_{0}, t_{0}],
\end{align}
Obviously, the solution of the above equation will blow up. But in the following, we will show that the overall system is stable.
Set
\begin{align*}
V( x, t,i)
=\begin{cases}
x^{2}, i=1,\\
2x^{2}+3x^{8}, i=2.
\end{cases}
\end{align*}
When $i=1, $ it follows that
\begin{align*}
&LV(\varphi, t,1)=2\varphi(1)f(\varphi, t,1)+|g(\varphi, t,1)|^{2}+ \sum^{2}_{j=1}\gamma_{1j}V(\varphi(1), t,j)\\
&\leq 2\varphi(1)[-6(\varphi(1)+\varphi^{3}(1)+\varphi^{7}(1))] +2|\varphi(1)|\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&\quad \quad \quad+0.25\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(1)|^{4}|\varphi(\theta )|^{4}{\mathord{{\rm d}}} \nu_{1} (\theta)-|\varphi(1)|^{2}+2|\varphi(1)|^{2}+3|\varphi(1)|^{8}\\
&\leq -11|\varphi(1)|^{2}-12|\varphi(1)|^{4}-9|\varphi(1)|^{8}+2|\varphi(1)|\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&\quad \quad \quad+0.25\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(1)|^{4}|\varphi(\theta )|^{4}{\mathord{{\rm d}}} \nu_{1} (\theta).
\end{align*}
When $i=2, $ we have
\begin{align*}
&LV(\varphi, t,2)=(4\varphi(1)+24\varphi^{7}(1))f(\varphi, t,2)+0.005(4+168\varphi^{6}(1))\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu_{2} (\theta)\\
&+ \sum^{2}_{j=1}\gamma_{2j}V(\varphi(1), t,j)\\
& \leq 0.08(4\varphi(1)+24\varphi^{7}(1))\varphi(1)+0.01(2+84\varphi^{6}(1))|\varphi(1)|^{2}+3\varphi^{2}(1)-3(2\varphi^{2}(1)+3\varphi^{8}(1))\\
& \leq -2.64|\varphi(1)|^{2}-6.24|\varphi(1)|^{8}.
\end{align*}
Then, we have
\begin{align*}
LV(\varphi, t,i) &\leq -2.66|\varphi(1)|^{2}-6.24|\varphi(1)|^{8}\\
&+2\varphi(1)\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{1} (\theta)
+0.25\int^{1}_{0.7}e^{-0.6(1-\theta) t}|\varphi(1)|^{4}|\varphi(\theta )|^{4}{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&\leq -2.64|\varphi(1)|^{2}-6.24|\varphi(1)|^{8}\\
&+\int^{1}_{0.7}e^{-0.6(1-\theta) t}2(|\varphi(1)|^{2})^{\frac{1}{2}}(|\varphi(\theta )|^{2})^{\frac{1}{2}}{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&+\int^{1}_{0.7}e^{-0.6(1-\theta) t}0.25(|\varphi(1)|^{8})^{\frac{1}{2}}(|\varphi(\theta )|^{8})^{\frac{1}{2}}{\mathord{{\rm d}}} \nu_{1} (\theta).
\end{align*}
Obviously, we can choose
\begin{align*}
&U_{0}(x,t)=|x|^{2},U_{1}(x,t)=|x|^{2}, U_{2}(x,t)=|x|^{8}, a_{0}=0, a_{1}=2.64,\\
& a_{2}=6.24,, b_{11}=2,
b_{21}=0.25.
\end{align*}
From Theorem 3.3, we have the following results:
\begin{itemize}
\item[(i)] $$\limsup_{t\rightarrow \infty}{\mathbb E}[|x(t_{0},t,\xi,i_{0})|^{2}]=0,$$
\item[(ii)] $$\limsup_{t\rightarrow \infty}\frac{1}{t}\int^{t}_{t_{0}}{\mathbb E}[|x(t_{0},s,\xi,i_{0})|^{2}]{\mathord{{\rm d}}} s =0,$$
$$\limsup_{t\rightarrow \infty}\frac{1}{t}\int^{t}_{t_{0}}{\mathbb E}[|x(t_{0},s,\xi,i_{0})|^{8}]{\mathord{{\rm d}}} s =0,$$
\item[(iii)] \begin{align*}
\limsup_{t\rightarrow \infty}\frac{1}{t}\log ({\mathbb E}[|x(t_{0},t,\xi,i_{0})|^{2}]) \leq -0.1.\nonumber
\end{align*}
\begin{align*}
\limsup_{t\rightarrow \infty}\frac{1}{t}\log (|x(t_{0},t,\xi,i_{0})|^{2}) \leq -0.1,\,a.s.
\end{align*}
\end{itemize}
}
\end{exa}
\subsection{Polynomial Stability}
In this subsection, we will investigate the polynomial stability of the solution for HPSFEDs \eqref{eq1}.
\begin{thm}
Assume (H1), and let (H2') hold with $a_0=0$. If
\begin{align}
-a_{k}+\sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}
+\sum^{l_{k}}_{l=1}b_{kl}\frac{1}{\underline{\theta}}(1-\alpha_{kl})<0, \, k=1, 2, \cdots, M,
\end{align}
then
the global solution $x(t_{0},t,\xi,i_{0})$ has almost surely polynomial stability, i.e.
\begin{align}\label{24}
\limsup_{t\rightarrow \infty }\frac{\log U_{0}(x(t),t)}{\log(1+t)}\leq -\varepsilon,
\end{align}
where $\varepsilon$ is a positive constant satisfying $$-a_{k}+\sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}
+\sum^{l_{k}}_{l=1}b_{kl}\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{kl})<0, k=2,3,\cdots, M,$$
and
$$
\varepsilon -a_{1}+\sum^{l_{1}}_{l=1}b_{1l}\alpha_{1l}
+\sum^{l_{1}}_{l=1}b_{1l}\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{1l})<0.
$$
\end{thm}
\begin{proof}
Define
$\sigma_{a}=\inf\{t\geq t_0: |x(t)|\geq a\}$
as before. Set
\begin{align*}
M(t)=\int^{t}_{t_{0}}(1+s)^{\varepsilon}V_{x}(x(s), s, r(s))g(x_{s}, s,r(s)){\mathord{{\rm d}}} B(s).
\end{align*}
Using It\^{o}'s formula and taking the expectation, we have
\begin{align}
&(1+t\wedge \sigma_{a})^{\varepsilon}V(x(t\wedge \sigma_a), t\wedge \sigma_a, r(t\wedge \sigma_a))]\nonumber\\
&=(1+t_{0})^{\varepsilon}V(x(t_0), t_0, r(t_0))+ \int^{t\wedge \sigma_a}_{t_{0}}\varepsilon (1+s)^{\varepsilon-1}V(x_{s},s,r(s)){\mathord{{\rm d}}} s\nonumber\\
& \quad \quad\quad+ \int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}LV(x_{s},s,r(s)){\mathord{{\rm d}}} s +M(t\wedge \sigma_{a}) \nonumber\\
& \leq (1+t_{0})^{\varepsilon}V(x(t_0), t_0, r(t_0)) +\int^{t\wedge \sigma_a}_{t_{0}}\varepsilon (1+s)^{\varepsilon}V(x_{s},s,r(s)){\mathord{{\rm d}}} s \nonumber\\
& \quad \quad\quad+ \int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)\nonumber\\
&\quad \quad \quad+ \sum^{l_{k}}_{l=1}b_{kl}\int^{1}_{\underline{\theta}} U_{k}^{\alpha_{kl}}(x(s), s) U_{k}^{1-\alpha_{kl}}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s
+M(t\wedge \sigma_{a})\nonumber\\
&\leq (1+t_{0})^{\varepsilon}V(x(t_0), t_0, r(t_0))+\int^{t\wedge \sigma_a}_{t_{0}}\varepsilon (1+s)^{\varepsilon}V(x_{s},s,r(s)){\mathord{{\rm d}}} s \nonumber\\
& \quad \quad\quad+ \int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)\nonumber\\
& \quad \quad\quad +\sum^{l_{k}}_{l=1} b_{kl}\int^{1}_{\underline{\theta}} U_{k}^{\alpha_{kl}}(x(s), s) U_{k}^{1-\alpha_{kl}}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s+M(t\wedge \sigma_{a})\nonumber\\
&\leq (1+t_{0})^{\varepsilon}V(x(t_0), t_0, r(t_0)) +\int^{t\wedge \sigma_a}_{t_{0}}\varepsilon (1+s)^{\varepsilon}V(x_{s},s,r(s)){\mathord{{\rm d}}} s\nonumber\\
& \quad \quad\quad + \int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}\bigg\{\bigg. a_{0}+\sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)+ \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}U_{k}(x(s), s)\nonumber\\
& \quad \quad\quad +\sum^{l_{k}}_{l=1}b_{kl}(1-\alpha_{kl})\int^{1}_{\underline{\theta}} U_{k}(x(\theta s), \theta s){\mathord{{\rm d}}} \nu_{k}(\theta) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s+M(t\wedge \sigma_{a})\nonumber\\
&\leq \tilde{c}_{0} +\int^{t\wedge \sigma_a}_{t_{0}}\varepsilon (1+s)^{\varepsilon}V(x_{s},s,r(s)){\mathord{{\rm d}}} s\nonumber\\
& \quad \quad\quad + \int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}\bigg\{\bigg. \sum^{M}_{k=1}\bigg[\bigg.-a_{k}U_{k}(x(s), s)+ \sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}U_{k}(x(s), s)\nonumber\\
&\quad\quad\quad+\sum^{l_{k}}_{l=1}b_{kl}\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{kl})U_{k}(\varphi( s), s) \bigg]\bigg.\bigg\}\bigg.{\mathord{{\rm d}}} s+M(t\wedge \sigma_{a}) \nonumber\\
&\leq \tilde{c}_{0} +\int^{t_{0}}_{\underline{\theta}t_{0}}\sum^{M}_{k=1}\sum^{l_{k}}_{l=1}b_{kl}
\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{kl})U_{k}(\xi( s), s){\mathord{{\rm d}}} s\nonumber\\
&+ \int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}\bigg(\bigg.\varepsilon -a_{1}+\sum^{l_{1}}_{l=1}b_{1l}\alpha_{1l}
+\sum^{l_{1}}_{l=1}b_{1l}\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{1l})\bigg)\bigg.U_{1}(x(s), s)\nonumber\\
&+\int^{t\wedge \sigma_a}_{t_{0}}(1+s)^{\varepsilon}\bigg\{\bigg. \sum^{M}_{k=2}\bigg(\bigg.-a_{k}+\sum^{l_{k}}_{l=1}b_{kl}\alpha_{kl}
+\sum^{l_{k}}_{l=1}b_{kl}\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{kl})\bigg)\bigg.U_{k}(x(s), s)\bigg\}\bigg.{\mathord{{\rm d}}} s\nonumber\\
&\leq \tilde{c}_{0}+M(t\wedge \sigma_{a}),
\end{align}
where $\tilde{c}_{0}=(1+t_{0})^{\varepsilon}V(x(t_0), t_0, r(t_0))+\int^{t_{0}}_{\underline{\theta}t_{0}}\sum^{M}_{k=1}\sum^{l_{k}}_{l=1}b_{kl}
\underline{\theta}^{-(1+\varepsilon)}(1-\alpha_{kl})U_{k}(\xi( s), s){\mathord{{\rm d}}} s.$\\
By virtue of the conditions in the theorem, we have
\begin{align*}
(1+t)^{\varepsilon}U_{0}(x(t), t)
\leq \tilde{c}_{0}+M(t).
\end{align*}
Then,
\begin{align*}
\limsup_{t\rightarrow \infty}(1+t)^{\varepsilon}U_{0}(x(t), t)
< \infty.
\end{align*}
This implies the required assertion (\ref{24}) immediately.
\end{proof}
The following example illustrates the theory of polynomial stability.
\begin{exa} {\rm
Let $\nu_{1}(\cdot)$ be a probability measure on $[\underline{\theta}, 1]$ and $\nu_{2}(\cdot)=\delta_{1}(\cdot).$ Set $S=\{1,2\}, \underline{\theta}=0.75, d=1$. Let $r(t)$ be a Markov chain with generator
$$
\Gamma=\left(
\begin{array}{ccc}
-1 & 1 \\
4 & -4 \\
\end{array}
\right).
$$
Consider the following HPSFDE:
\begin{align}
&{\mathord{{\rm d}}} x(t)= f( x_t, t,r(t)){\mathord{{\rm d}}} t +g( x_t, t, r(t)){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty)\nonumber\\
& x(t)=\xi(t), t\in [\underline{\theta} t_{0}, t_{0}],
\end{align}
where for $\varphi \in {\mathscr{C}}$
\begin{align*}
f( \varphi, t,i)=
\begin{cases}
-6(\varphi(1)+\varphi^{3}(1)+\varphi^{7}(1)) +0.5\int^{1}_{0.75}\varphi(\theta ){\mathord{{\rm d}}} \nu_{1} (\theta), r(t)=1,\\
0.04\varphi(1) +0.03\int^{1}_{0.75}\varphi(\theta ){\mathord{{\rm d}}} \nu_{2} (\theta), r(t)=2;
\end{cases}
\end{align*}
and
\begin{align*}
g( \varphi, t,i)
=\begin{cases}
0.2\int^{1}_{0.75}|\varphi(1)|^{1.5}|\varphi(\theta )|^{2.5}{\mathord{{\rm d}}} \nu_{1} (\theta), r(t)=1,\\
0.1\int^{1}_{0.75}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{2} (\theta), r(t)=2.
\end{cases}
\end{align*}
From above equation, when $r(t)=2, $ we can see that the equation is
\begin{align}
&{\mathord{{\rm d}}} x(t)= 0.07x(t){\mathord{{\rm d}}} t +0.1x(t){\mathord{{\rm d}}} B(t), t\in [t_{0}, \infty)\nonumber\\
& x(t)=\xi(t), t\in [\underline{\theta} t_{0}, t_{0}],
\end{align}
obviously, the solution of the above equation will blow up. But in the following, we will show that the overall system is polynomial stable.
Set
\begin{align*}
V( x, t,i)
=\begin{cases}
x^{4}, i=1,\\
2x^{4}+3x^{10}, i=2.
\end{cases}
\end{align*}
When $i=1, $ it follows that
\begin{align*}
&LV(\varphi, t,1)=4\varphi^{3}(1)f(\varphi, t,1)+0.24\varphi^{2}(1)|g(\varphi, t,1)|^{2}+ \sum^{2}_{j=1}\gamma_{1j}V(\varphi(1), t,j)\\
&\leq 4\varphi^{3}(1)[-6(\varphi(1)+\varphi^{3}(1)+\varphi^{7}(1))] +2\varphi^{3}(1)\int^{1}_{0.75}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&\quad \quad \quad \quad\quad \quad \quad+0.24\int^{1}_{0.75}|\varphi(1)|^{5}|\varphi(\theta )|^{5}{\mathord{{\rm d}}} \nu_{1} (\theta)-|\varphi(1)|^{4}+2|\varphi(1)|^{4}+3|\varphi(1)|^{10}\\
&\leq -23|\varphi(1)|^{4}-24|\varphi(1)|^{6}-21|\varphi(1)|^{10}+2|\varphi(1)|^{3}\int^{1}_{0.75}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&\quad \quad \quad \quad\quad \quad \quad+0.24\int^{1}_{0.75}|x(t)|^{5}|\varphi(\theta )|^{5}{\mathord{{\rm d}}} \nu_{1} (\theta).
\end{align*}
When $i=2, $ we have
\begin{align*}
&LV(\varphi, t,2)=(8\varphi^{3}(1)+30\varphi^{9}(1))f(\varphi, t,2)+0.005(24\varphi^{2}(1)+270\varphi^{8}(1))\int^{1}_{0.75}|\varphi(\theta )|^{2}{\mathord{{\rm d}}} \nu_{2} (\theta)\\
&+ \sum^{2}_{j=1}\gamma_{2j}V(x(t), t,j)\\
& \leq (8\varphi^{3}(1)+30\varphi^{9}(1))0.07\varphi(1)+0.01(12\varphi^{2}(1)+135\varphi^{8}(1))\varphi^{2}(1)\\
&+4\varphi^{4}(1)-4(2\varphi^{4}(1)+3\varphi^{10}(1))\\
& \leq -3.32|\varphi(1)|^{4}-8.55|\varphi(1)|^{10}.
\end{align*}
Then, we have
\begin{align*}
LV(\varphi, t,i) &\leq -3.32|\varphi(1)|^{4}-8.55|\varphi(1)|^{10}\\
&+2|\varphi(1)|^{3}\int^{1}_{0.75}|\varphi(\theta )|{\mathord{{\rm d}}} \nu_{1} (\theta)
+0.24\int^{1}_{0.75}|x(t)|^{5}|\varphi(\theta )|^{5}{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&\leq -3.32|\varphi(1)|^{2}-8.25|\varphi(1)|^{8}\\
&+\int^{1}_{0.75}2(|\varphi(1)|^{4})^{\frac{3}{4}}(|\varphi(\theta )|^{4})^{\frac{1}{4}}{\mathord{{\rm d}}} \nu_{1} (\theta)\\
&+\int^{1}_{0.75}0.24(|\varphi(1)|^{10})^{\frac{1}{2}}(|\varphi(\theta )|^{10})^{\frac{1}{2}}{\mathord{{\rm d}}} \nu_{1} (\theta)
\end{align*}
Obviously, we can choose
\begin{align*}
&U_{0}(x,t)=|x|^{4},U_{1}(x,t)=|x|^{4}, U_{2}(x,t)=|x|^{10}, a_{0}=0, a_{1}=3.32,\\
& a_{2}=8.55,, b_{11}=2,
b_{21}=0.24.
\end{align*}
From Theorem 3.6, we conclude that the overall system is polynomial stable.
}
Obviously, this example is similar to example 3.5, but $f,g$ in this example satisfy (H2)' while $f,g$ in example 3.5 satisfy (H2). This difference leads to different stable properties of the solution.
\end{exa}
\section*{Acknowledgements}
This research was supported by the National Natural Science Foundation of China (Grant no.61876192, 11626236), the Fundamental Research Funds for the Central Universities of South-Central University for Nationalities (Grant nos. CZY15017, KTZ20051, CZT20020)
|
\section{Introduction}
Parton distribution functions (PDFs) play a crucial role in the prediction of Standard Model observables at the Large Hadron Collider (LHC) and beyond~\cite{Forte:2020yip, Ethier:2020way}. PDF uncertainties are already a limiting factor in these predictions~\cite{Azzi:2019yne}. This means that a precise and accurate knowledge of PDFs is essential for full exploitation of the physics at LHC. Besides the familiar experimental uncertainties in the data that are input to a global PDF fit, it is becoming increasingly necessary to also consider theoretical uncertainties: missing higher order uncertainties (MHOUs) due to the truncation of perturbative expansions in the hard cross-sections and parton evolution, uncertainties due to the use of nuclear targets, uncertainties due to missing higher twist, uncertainties due to external parameters such as quark masses, showering and hadronization uncertainties in final state Monte Carlos, and so on. In the past the technique used to estimate the effect of theoretical uncertainties was to compare the PDFs produced with and without various theoretical corrections, rather than to incorporate the uncertainties into the fit itself. However the limitations of such a procedure are clear.
Recently a new approach to estimating the impact of theoretical uncertainties on global PDF fits has been developed, through the construction of a `theory covariance matrix' \cite{Ball:2018lag}, analogous to the experimental covariance matrix used in global PDF fits. By adding the theory covariance matrix to the experimental covariance matrix as an additional source of uncertainty, theoretical uncertainties can be incorporated directly in the fit, where they impact not only the overall level of PDF uncertainty, but also the relative weight of different datasets. This novel approach has so far been applied to uncertainties associated with nuclear effects \cite{Ball:2018twp,Ball:2020xqw}, and to the estimation of MHOUs by scale variation \cite{Pearson:2018tim, AbdulKhalek:2019bux, AbdulKhalek:2019ihb} (though other methods~\cite{Cacciari:2011ze, Bagnaschi:2014wea, Bonvini:2020xeo} of estimating MHOU are under development). Factorization scale variations estimate the MHOUs in parton evolution (correlated across all observables), and renormalization scale variations estimate the MHOUs in fixed order calculations of process-dependent hard cross-section (correlated only across a given class of processes). The first global PDFs including MHOUs estimated in this way were presented in \cite{AbdulKhalek:2019ihb}.
When making predictions for hadronic observables there are again two sources of MHOU: uncertainties in the PDF evolution, which can also be estimated by factorization scale variation, and uncertainties in the hard cross-section, again estimated by renormalization scale variation. These arise on top of the MHOUs incorporated in the determination of the PDFs themselves, manifested as a part of the PDF uncertainty. So there will be correlations between the MHOUs in the PDFs and the additional MHOUs in the predictions. The PDFs themselves contain a wealth of data from a wide range of processes. When making a prediction for one of these processes (for example in a new kinematic regime), there will inevitably be correlations between the renormalization scale variation in the PDF determination, and that in the prediction. Even if the process is a new one, for example Higgs production, correlations due to factorization scale variation will still be present: all processes dependent on PDFs have a MHOU due to the need to evolve the PDFs to the scale of the process.
The potential importance of these correlations can be exposed by considering a simple situation in which a PDF is determined from data at a given scale on a single observable (such as a nonsinglet structure function), and used to make predictions of another observable at the same scale \cite{Harland-Lang:2018bxd}. In this special case the PDF can be eliminated altogether, and the predicted observable determined directly from the measured observable. Including the MHOU twice (once in the calculation of the measured observable, then again in making the predicted observable), whilst ignoring their possible correlation, would amount to ``double counting", and thus an overestimate of the uncertainty. Of course in a global PDF fit, involving many different processes, this particular formulation is no longer applicable. However the fact remains that there will still be some residual correlation between the MHOU in the PDF determination and the MHOU in the prediction, and to neglect it may lead to an overestimation of the MHOU.
In Ref.~\cite{AbdulKhalek:2019ihb}, it was seen that in a realistic NLO global fit, the effect of MHOU on the PDF uncertainty is
quite small, the main consequence being a rebalancing of the impact of different datasets depending on their relative MHOU. The MHOU in the PDF is consequently rather smaller than the MHOU in the prediction, and if they are combined in quadrature, the effect of missing correlations will likely also be small. It was further argued that combining the PDF uncertainties and theoretical uncertainties in quadrature is conservative, since it can only lead to an overestimate of uncertainties, and thus better than neglecting MHOUs altogether. Nevertheless, since the correlation between the uncertainties has not been computed explicitly, there remains the intriguing possibility that in some circumstances including the correlation may yield more precise, and perhaps even more accurate, predictions.
In this paper we show explicitly how to propagate theoretical uncertainties in the determination of global PDFs into the predictions made using these PDFs, taking account of all correlations between theoretical uncertainties. As this is a complicated problem, we proceed step by step. In Sec.~1 we show how a single source of theoretical uncertainty can be reformulated in terms of a nuisance parameter, which holds the key to the propagation of uncertainties. We explore the interplay between the theoretical uncertainties and the experimental uncertainties in the data in two idealised contexts: the first in which there are no fitted parameters, the second in which there as many parameters as data, so the experimental data can be fitted exactly. This allows us to identify three distinct effects of the correlation of theoretical uncertainties. Then in Sec.~2 we consider a more realistic, but still simplified, situation where we fit the data using a single parameter, still in a theory with a single source of theoretical uncertainty, and find that all three of these correlation effects are still present, and can be computed. In Sec.~3 these results are extended to multiple theoretical uncertainties, in a multiparameter fit, and then less trivially to a PDF fit where the PDFs are continuous functions, with a functional uncertainty. Finally in Sec.~4 we present numerical results, made in the context of the NNPDF3.1 NLO global fit with MHOU presented in Ref.~\cite{AbdulKhalek:2019ihb}, and make predictions with MHOU for repetitions of the experiments included in the fit (so-called `autopredictions'), and for genuine predictions (top and Higgs production). We are able to compute all correlations explicitly, calculating corrections to central values and changes in PDF and theoretical uncertainties for both autopredictions and genuine predictions, and confirm that the correlated predictions can be both more accurate and more precise than the conservative prescription. A summary is provided in Sec.~5.
\section{Predictions with Correlated Theoretical Uncertainties}
\label{sec:generic}
We showed in Refs.\cite{Ball:2018lag, Ball:2018twp} that when we fit $N$ experimental data points $D_i$ to theoretical predictions $T_i$, $i = 1,\ldots,N$, then the uncertainties in the theoretical predictions can can be incorporated into the fit simply by adding a theoretical covariance matrix $S_{ij}$ to the experimental covariance matrix $C_{ij}$. The only assumptions made in deriving this result are that all uncertainties, both experimental and theoretical, can be treated as Gaussian, and that the theoretical uncertainties are independent of the experimental data. We used this result in
Refs.~\cite{Ball:2018twp,Ball:2020xqw} to include nuclear uncertainties in a PDF fit, and in
Refs.~\cite{AbdulKhalek:2019bux,AbdulKhalek:2019ihb} to incorporate missing higher order uncertainties in global PDF fits.
The result of Sec.~2 of Ref.\cite{Ball:2018twp} may be summarised in terms of the Bayesian probability
\begin{equation}
\label{eq:PTD}
P(T|D)\propto \exp\left(-\smallfrac{1}{2}(T-D)^T(C+S)^{-1}(T-D)\right)
\end{equation}
where for simplicity we adopt a matrix notation: $C$ and $S$ are real symmetric matrices, with $C$ strictly positive definite (thus invertible), and $S$ positive semi-definite. In practice theoretical uncertainties are highly correlated, so $S$ may (and generally will) have zero eigenvalues. We determine $T$ from $D$ by maximizing $P(T|D)$: this is equivalent to minimizing
\begin{equation}
\label{eq:chisq}
\chi^2 = (T-D)^T(C+S)^{-1}(T-D)
\end{equation}
with respect to free parameters characterizing the theoretical prediction.
In this section we will also assume that there is only a single source of fully correlated theoretical uncertainty, so that the theory covariance matrix can be written as
\begin{equation}
\label{eq:Sdef}
S = \beta\beta^T
\end{equation}
for some real nonzero vector $\beta$ (i.e. that $S_{ij} = \beta_i\beta_j)$. Then all eigenvalues of $S$ except one are zero. We will develop a nuisance parameter formalism to propagate this theoretical uncertainty (Sec.~\ref{subsec:nuisance1}), and then show how to apply it in two extreme cases: firstly to the situation in which the theory contains no free parameters, and thus where there is no fitting (Sec.~\ref{subsec:puretheory}), and then to the situation where there as many free parameters as data points, so that we can achieve a perfect fit (Sec.~\ref{subsec:phenomenology}).
\subsection{Nuisance Parameters}
\label{subsec:nuisance1}
It will be useful in what follows to introduce a nuisance parameter $\lambda$ for the correlated theoretical uncertainty. Following the notation in Ref.~\cite{Ball:2012wy}, we model the theoretical uncertainty as a fully correlated shift in the theoretical prediction: $T\to T+\lambda\beta$. The nuisance parameter $\lambda$ then gives the size of the shift. For a given shift, with Gaussian experimental uncertainties
\begin{equation}
\label{eq:PTDl}
P(T|D\lambda)\propto \exp\left(-\smallfrac{1}{2}(T+\lambda\beta-D)^TC^{-1}(T+\lambda\beta-D)\right).
\end{equation}
Now using Bayes' Theorem,
\begin{equation}
\label{eq:bayes}
P(T|D\lambda)P(\lambda|D) = P(\lambda|TD)P(T|D).
\end{equation}
To determine $P(T|D)$ we need to first fix the prior distribution of $\lambda$. Since this is a theoretical uncertainty, it is reasonable to assume that the prior is independent of the data, thus that $P(\lambda|D)=P(\lambda)$. Then, marginalizing Eq.~(\ref{eq:prior}) over $\lambda$,
\begin{equation}
\label{eq:marginalise1}
P(T|D) = \int\! d\lambda\, P(T|D\lambda)P(\lambda).
\end{equation}
Taking $P(\lambda)$ as a Gaussian, centred on zero (so the theoretical uncertainty is unbiased), with unit width (fixing the overall normalization of the prior theoretical uncertainty), we choose
\begin{equation}
\label{eq:prior}
P(\lambda) \propto \exp\left(-\smallfrac{1}{2}\lambda^2\right).
\end{equation}
The integration over the nuisance parameter is now Gaussian,
\begin{equation}
\label{eq:marginalise_gauss}
P(T|D) \propto\int d\lambda\, \exp\left(-\smallfrac{1}{2}[(T+\lambda\beta-D)^TC^{-1}(T+\lambda\beta-D)+\lambda^2]\right)\, ,
\end{equation}
and can be performed in the usual way by completing the square:
\begin{eqnarray}
\label{eq:completesquare}
(T+\lambda\beta-D)^TC^{-1}(T+\lambda\beta-D)+\lambda^2
&=&
Z^{-1}\left(\lambda
+Z\beta^TC^{-1}(T-D)\right)^2\\ && + (T-D)^TC^{-1}(T-D) - Z(\beta^TC^{-1}(T-D))^2,\nonumber
\end{eqnarray}
where we have defined
\begin{equation}
\label{eq:Zdef}
Z = (1+\beta^TC^{-1}\beta)^{-1} = 1-\beta^T(C+S)^{-1}\beta.
\end{equation}
The second expression was obtained by noting that
\begin{equation}
\label{eq:Zalgebra}
(1+\beta^T C^{-1}\beta)(1-\beta^T(C+S)^{-1}\beta)=1
\end{equation}
using Eq.~(\ref{eq:Sdef}).
Since
\begin{equation}
\label{eq:numa}
(\beta^TC^{-1}(T-D))^2 = (T-D)^TC^{-1}\beta\beta^TC^{-1}(T-D),
\end{equation}
we can combine the two terms on the second line of Eq.~(\ref{eq:completesquare})
to give
\begin{equation}
\label{eq:chisqtoo}
(T-D)^T(C^{-1}-ZC^{-1}\beta\beta^TC^{-1})(T-D) = (T-D)^T(C+\beta\beta^T)^{-1}(T-D),
\end{equation}
since
\begin{equation}
\label{eq:inverse}
(C+\beta\beta^T)(C^{-1}-ZC^{-1}\beta\beta^TC^{-1}) =1 + \beta\beta^TC^{-1}- Z\beta\beta^TC^{-1} - Z\beta\beta^TC^{-1}\beta\beta^TC^{-1} = 1,
\end{equation}
by substituting (from Eq.~(\ref{eq:Zdef})) $\beta^TC^{-1}\beta = Z^{-1}-1$ in the last term. Using the definition Eq.~(\ref{eq:Sdef}), we recognise Eq.~(\ref{eq:chisqtoo}) as the $\chi^2$, Eq.~(\ref{eq:chisq}). Defining
\begin{equation}
\label{eq:lambdabar}
\overline{\lambda}(T,D) = Z\beta^TC^{-1}(D-T)=\beta^T(C+S)^{-1}(D-T),
\end{equation}
we can thus write Eq.~(\ref{eq:marginalise_gauss}) as
\begin{equation}
\label{eq:integration}
P(T|D)\propto\int d\lambda\, \exp\left(-\smallfrac{1}{2} Z^{-1}(\lambda-\overline{\lambda})^2 - \smallfrac{1}{2}\chi^2\right) \propto \exp(-\smallfrac{1}{2}\chi^2),
\end{equation}
since the integration over $\lambda$ yields a factor $(2\pi Z)^{1/2}$ which
is independent of $T$ and $D$.
The nuisance parameter formalism thus produces our original result Eq.~(\ref{eq:PTD}). However it is useful because it allows us to also determine the posterior distribution of the nuisance parameter: inverting Eq.~(\ref{eq:bayes})
\begin{equation}
\label{eq:posterior}
P(\lambda|TD)\propto \exp\left(-\smallfrac{1}{2} Z^{-1}(\lambda-\overline{\lambda}(T,D))^2\right),
\end{equation}
so once we use the information on $T$ and $D$, the prior distribution
Eq.~(\ref{eq:prior}) is modified. In particular the peak of the distribution is shifted away from zero to $\overline{\lambda}$, while the width is now given by $Z$. It is easy to see from the definition Eq.~(\ref{eq:Zdef}) that
\begin{equation}
\label{zbounds}
0 < Z < 1,
\end{equation}
so the width of the theoretical uncertainty is generally reduced by the addition of new information.
\subsection{Predictions and Autopredictions without Fits}
\label{subsec:puretheory}
In order to explore the implications of the formalism described in the previous section we first consider a `pure' theory $T=T_0$, one where there are no unknown parameters to be fitted, but which nevertheless has a theoretical uncertainty. Despite the fact that this theory cannot be fitted to the data, so in general $T_0\neq D$, the data can still inform the theoretical predictions.
Computing expectation values of functions of $\lambda$ using the probability distribution $P(\lambda|T_0 D)$,
\begin{equation}
\label{eq:Elampt}
{\rm E}[\lambda] = {\cal N}_\lambda \int d\lambda\;\lambda\; P(\lambda|T_0 D) = \overline{\lambda}(T_0,D),
\end{equation}
the normalization ${\cal N}_\lambda$ being chosen such that ${\rm E}[1] =1$, while
\begin{equation}
\label{eq:varlampt}
{\rm Var}[\lambda] \equiv {\rm E}[(\lambda -{\rm E}[\lambda])^2] = Z.
\end{equation}
Now in the nuisance parameter formalism, the theoretical predictions are
\begin{equation}
\label{eq:Tlambda}
T(\lambda)=T_0 +\lambda\beta.
\end{equation}
If we made these predictions before comparing to any data, we would use the prior distribution Eq.~(\ref{eq:prior}) for $\lambda$. Since this is a unit gaussian centred on zero, we would then find that ${\rm E}[T(\lambda)] = T_0$, while ${\rm Cov}[T(\lambda)] = \beta\beta^T=S$ as expected.
However if instead we first compare $T$ to the data $D$, then we can instead compute expectation values using $P(\lambda|TD)$. We then find (using Eqs.~(\ref{eq:Elampt},\ref{eq:lambdabar}))
\begin{equation}
\label{eq:ETlampt}
{\rm E}[T(\lambda)] = T_0+\overline{\lambda}(T_0,D)\beta = T_0 + \beta\beta^T(C+S)^{-1}(D-T_0),
\end{equation}
while (using Eq.~(\ref{eq:varlampt}))
\begin{equation}
\label{eq:varTlampt}
{\rm Cov}[T(\lambda)] \equiv {\rm E}[(T(\lambda) -{\rm E}[T(\lambda)])(T(\lambda) -E[T(\lambda)])^T] ={\rm Var}[\lambda]\beta\beta^T =ZS.
\end{equation}
One can consider this as an `autoprediction': first the theory is compared to the data, and then using this information one makes new theoretical predictions for precise repetitions of the same experiments. The original theoretical predictions $T_0$ are then shifted by an amount
\begin{equation}
\label{eq:delTpt}
\delta T = -S(C+S)^{-1}(T_0-D)
\end{equation}
and their theoretical uncertainties are reduced by a factor of $\sqrt{Z}$, since the data add new information: the covariance matrix of the autopredictions is
\begin{equation}
\label{eq:ZS}
ZS = S - S(C+S)^{-1}S = C(C+S)^{-1}S = S(C+S)^{-1}C.
\end{equation}
This is a simple example of Bayesian learning: the theory `learns' from the data, within the constraints imposed by the prior theoretical uncertainty.
It is interesting to compare the experimental $\chi^2$ of the original predictions
\begin{equation}
\label{eq:expchisq}
\chi^2_{\rm exp} = (T_0-D)^TC^{-1}(T_0-D)
\end{equation}
to that obtained with the autopredictions
\begin{eqnarray}
\chi^2_{\rm auto} &=& (T_0+\delta T-D)^T C^{-1}(T_0+\delta T-D)\nonumber\\
&=& (T_0-D)^T (C+S)^{-1}C(C+S)^{-1}(T_0-D),
\label{eq:chisqauto}
\end{eqnarray}
since
\begin{equation}
\label{eq:TplusdelTminD}
T_0+\delta T-D = C(C+S)^{-1}(T_0-D).
\end{equation}
It is easy to see that $\chi^2_{\rm auto}\leq\chi^2_{\rm exp}$ since $C$ is positive definite (i.e. $C>0$) and $S$ semi-positive definite (i.e. $S\geq 0$), so $2S+SC^{-1}S\geq 0$, whence $(C+S)C^{-1}(C+S)\geq C$, and $(C+S)^{-1}C(C+S)^{-1}\leq C^{-1}$. So the data induced shifts are always such as to improve the quality of the fit to the data, by exploiting the theoretical uncertainty.
To make this more explicit, consider a simple model in which the experimental uncertainties are uncorrelated, and the same for each data point: then we can write
\begin{equation}
\label{eq:modelCS}
C = \sigma^2 1,\qquad S = s^2 e e^T,
\end{equation}
with $e$ a unit vector, $e^Te=1$, and $\beta = s e$ so $S=\beta\beta^T$. Thus $\sigma$ is the experimental uncertainty on each data point, and $s/\sqrt{N}$ is the size of the correlated theoretical uncertainty. Then it is easy to see that
\begin{equation}
\label{eq:modelCplusSinv}
(C+S)^{-1} = \frac{1}{\sigma^2}\left(1-\frac{s^2}{\sigma^2+s^2}e e^T\right),
\end{equation}
and (using Eq.~(\ref{eq:Zdef}))
\begin{equation}
\label{eq:modelZS}
Z = (1+s^2/\sigma^2)^{-1}.
\end{equation}
So the reduction in theoretical uncertainties depends on the ratio $s^2/\sigma^2$: when $s^2\ll\sigma^2$ the influence of the data on the theoretical uncertainty is very small,
while when $s^2\gg\sigma^2$ the size of the theoretical uncertainty is reduced from $s$ to $\sigma$, since in the limit $s^2/\sigma^2\to \infty$
\begin{equation}
\label{eq:modelZ}
ZS = \frac{\sigma^2 s^2}{\sigma^2+s^2}ee^T \to \sigma^2 ee^T.
\end{equation}
Note that when the theoretical uncertainties are comparable to the experimental, $s^2/N\sim\sigma^2$, and $Z \sim 1/(N+1)$: if there are a large number of independent data points, the reduction of the theoretical uncertainty can be very substantial.
In this model the shifts in the theoretical predictions is (using Eq.~(\ref{eq:delTpt}))
\begin{equation}
\label{eq:modeldelTpt}
\delta T = - \frac{s^2}{\sigma^2+s^2} (e^T(T_0-D)) e,
\end{equation}
as expected in the direction $e$ of the theoretical uncertainty. When $s^2/\sigma^2\to \infty$, $e^T(T_0+\delta T)\to e^TD$, so in this direction the autopredictions coincide precisely with the data. The experimental $\chi^2$ for the autopredictions is (using Eq.~(\ref{eq:chisqauto}))
\begin{equation}
\label{eq:modelchisqauto}
\chi^2_{\rm auto} = (T_0-D)^T \frac{1}{\sigma^2}\Big(1 - \frac{s^2(s^2+2\sigma^2)}{(s^2+\sigma^2)^2}ee^T\Big)(T_0-D),
\end{equation}
So $N-1$ contributions to the $\chi^2$ orthogonal to $e$ are unchanged, while the contribution along $e$ is reduced by a factor $Z^2$. The autopredictions will in general have a $\chi^2$ of size $N-1$, rather than the $N$ of the original predictions, as naively expected since the nuisance parameter is effectively fitted.
Of course we can also consider genuine predictions $\widetilde{T}_I$, $I=1,\ldots,\widetilde{N}$, which again have no free parameters, but have a theoretical uncertainty which is correlated to that of the observables $T_i$, $i=1,\ldots, N$ for which we have the data $D_i$. The theoretical predictions including the theoretical uncertainty may be written
\begin{equation}
\label{eq:Ttillambda}
\widetilde{T}(\widetilde{\lambda}) = \widetilde{T} + \widetilde{\lambda}\widetilde{\beta},
\end{equation}
where the vector $\widetilde{\beta}_I$ gives the size and direction of the theoretical uncertainty in $\widetilde{T}_I$. Now if the nuisance parameters $\widetilde{\lambda}$ are independent of the parameters $\lambda$, the theoretical uncertainties in $\widetilde{T}$ are uncorrelated with those in $T$, and the theory covariance matrix for the prediction $\widetilde{T}$ is given by
\begin{equation}
\widetilde{S} = \widetilde{\beta}\betatil^T,\label{eq:Stildef}
\end{equation}
However if they are correlated, $\widetilde{\lambda}=\lambda$, and then taking advantage of the data $D$ for $T$, we can compute expectation values for the predictions using $P(\lambda|TD)$. We then find that
\begin{equation}
\label{eq:ETtillampt}
{\rm E}[\widetilde{T}(\lambda)] = \widetilde{T}+\overline{\lambda}(T,D)\widetilde{\beta} = \widetilde{T} + \widetilde{\beta}\beta^T(C+S)^{-1}(D-T_0),
\end{equation}
so the predictions are shifted by
\begin{equation}
\label{eq:delTtilpt}
\delta \widetilde{T} = -\widehat{S}(C+S)^{-1}(T_0-D),
\end{equation}
where
\begin{equation}
\widehat{S} = \widetilde{\beta}\beta^T,\label{eq:Shatdef}
\end{equation}
is the cross-covariance matrix between observables $T$ for which we have data $D$, and the predictions $\widetilde{T}$. Likewise
\begin{equation}
\label{eq:varTtillampt}
{\rm Cov}[\widetilde{T}(\lambda)] \equiv {\rm E}[(\widetilde{T}(\lambda) -{\rm E}[\widetilde{T}(\lambda)])(\widetilde{T}(\lambda) -E[\widetilde{T}(\lambda)])^T] ={\rm Var}[\lambda]\widetilde{\beta}\betatil^T = Z \widetilde{S}.
\end{equation}
so the covariance matrix of the predictions is reduced by the same factor $Z$ as that for the autopredictions.
Thus the data $D$ can lead to more precise (and if the theory is correct, also more accurate) predictions for observables that are not yet measured, through the correlation of theoretical uncertainties. The reduction in the size of the covariance matrix is through the same factor $Z$ as for the autopredictions, while the size of the shift is proportional to the cross-covariance between the theoretical uncertainties.
Note that if, having made the predictions $\widetilde{T}+\delta\widetilde{T}$, an experimentalist made measurements to produce independent data $\widetilde{D}$, with experimental covariance matrix $\widetilde{C}$, and we wanted to combine the datasets $\{D,\widetilde{D}\}$ into a single dataset, then the combined $(N+\widetilde{N})\times(N+\widetilde{N})$ theoretical covariance matrix for $\{T,\widetilde{T}\}$ to be used to compare to the data would be
\begin{equation}
\left(\begin{array}{cc}
S&\widehat{S}^T\\
\widehat{S} &\widetilde{S}\end{array}\right)
=
\left(\begin{array}{cc}
\beta\beta^T&\beta\widetilde{\beta}^T\\
\widetilde{\beta}\beta^T&\widetilde{\beta}\betatil^T\end{array}\right).
\label{eq:covmatglobal}
\end{equation}
While we might hope that the shifted predictions $\widetilde{T}+\delta\widetilde{T}$ give a better $\chi^2$ to the new data $\widetilde{D}$, this is no longer guaranteed, since the shifts are driven by the old data $D$, and it is possible that $\widetilde{D}$ are inconsistent with them.
\subsection{Autopredictions in Perfect Fits}
\label{subsec:phenomenology}
In the previous section we considered the situation of a theory $T$ which was not fitted to the data $D$. Now consider what is in some sense the opposite situation: a `perfect' fit, where the theoretical predictions $T$ have sufficient flexibility to fit the data $D$ exactly. For a perfect fit, $P(T|D)$ is always maximized when $T=D$, and thus (Eq.~(\ref{eq:chisq})) $\chi^2=0$. We compute expectation values of functions of $T$ using the probability distribution $P(T|D)$ in Eq.~(\ref{eq:PTD}), thus
\begin{equation}
\label{eq:ETpheno}
{\rm E}[T] = {\cal N}_T \int dT\; T\; P(T|D) = D,
\end{equation}
the normalization ${\cal N}_T$ being chosen such that ${\rm E}[1] =1$, while
\begin{equation}
\label{eq:varTpheno}
{\rm Cov}[T] \equiv {\rm E}[(T -{\rm E}[T])(T -{\rm E}[T])^T] = C+S.
\end{equation}
These are the expectation value and covariance matrix of the variables $T$ fitted to the data $D$.
To compute the autopredictions, we need to consider
$T(\lambda)=T +\lambda\beta$ (Eq.~(\ref{eq:Tlambda})), and this is more subtle since, as we saw in the previous section, expectation values of $\lambda$ involve $T$. We thus need to generalise the definition Eq.~(\ref{eq:Elampt}) of expectation values of functions of $\lambda$ using the probability distribution $P(\lambda|TD)$, to also include the subsequent integration over $T$: thus we define
\begin{equation}
\label{eq:Edef}
{\rm E}[f(T,\lambda)] \equiv {\cal N}_T \int dT\; \Big({\cal N}_\lambda \int d\lambda\;f(T,\lambda)\; P(\lambda|TD)\Big)P(T|D),
\end{equation}
for any function $f(\lambda,T)$ of $\lambda$ and $T$. There are several things to note about this procedure:
\begin{itemize}
\item We always perform the integration over $\lambda$, weighted with the probability distribution $P(\lambda|TD)$, before we perform the integration over $T$ using $P(T|D)$: this is because while $P(\lambda|TD)$ depends on $T$, $P(T|D)$ does not depend on $\lambda$, because it has been marginalised as in Eq.~(\ref{eq:integration}).
\item The data $D$ are always held fixed throughout: both $P(\lambda|TD)$ and $P(T|D)$ are conditional on $D$.
\item For functions $f(\lambda,T)$ which only depend on $T$, the integration over $\lambda$ is trivial and and we recover for example the results
Eqs.~(\ref{eq:ETpheno},\ref{eq:varTpheno}).
\item For the pure theory discussed in the previous section, the theory $T$ was held fixed, so the $T$ integration was trivial and we recover the results Eqs.~(\ref{eq:Elampt},\ref{eq:varlampt}).
\end{itemize}
Thus in the case of a perfect fit, the expectation value of the nuisance parameter is
\begin{equation}
\label{eq:Elampheno}
{\rm E}[\lambda] = E[\overline{\lambda}(T,D)]=\beta^T(C+S)^{-1}E[D-T] = 0,
\end{equation}
using Eq.~(\ref{eq:lambdabar}) and then finally Eq.~(\ref{eq:ETpheno}). The calculation of the variance requires some care, to ensure that the average over $\lambda$ is separated out from the average over $T$. This is most easily accomplished by adding and subtracting $\overline{\lambda}(T,D)$, so
\begin{eqnarray}
{\rm Var}[\lambda] &=& {\rm E}[(\lambda -{\rm E}[\lambda])^2]= {\rm E}[(\lambda -\overline{\lambda}(T,D)+ \overline{\lambda}(T,D))^2]\nonumber\\
&=& {\rm E}[(\lambda -\overline{\lambda}(T,D))^2]+ E[(\overline{\lambda}(T,D))^2]\nonumber\\
&=& Z + \beta^T(C+S)^{-1}E[(T-D)(T-D)^T](C+S)^{-1}\beta\nonumber\\
&=& Z + \beta^T(C+S)^{-1}{\rm Cov}[T] (C+S)^{-1}\beta\label{eq:varlamgen}\\
&=& 1 - \beta^T(C+S)^{-1}\beta + \beta^T(C+S)^{-1}\beta = 1.\label{eq:varlampheno}
\end{eqnarray}
where in the second line we note that the cross term vanishes, while in the last line we used Eq.~(\ref{eq:Zdef}) for $Z$, and Eqs.~(\ref{eq:varTpheno}) to simplify the second term. We thus find that in a perfect fit, the probability distribution of the nuisance parameters after fitting the data is the same as the prior distribution Eq.~(\ref{eq:prior}): we learn nothing from the data about the theoretical uncertainty because all the information in the data is absorbed in the fitted parameters. The calculation of the variance is particularly instructive: the reduction by the factor $Z$ found in the pure theory, Eq.~(\ref{eq:varlampt}) is now precisely cancelled by the fluctuation of $\overline{\lambda}(T,D)$ due to the covariance Eq.~(\ref{eq:varTpheno}) of $T$.
Turning to the autopredictions Eq.~(\ref{eq:Tlambda}), we have
\begin{equation}
\label{eq:ETlampheno}
{\rm E}[T(\lambda)] = E[T+\lambda\beta]=D,
\end{equation}
using Eqs.~(\ref{eq:ETpheno},\ref{eq:Elampheno}): as expected in a perfect fit, the autopredictions simply return the original data. Their covariance is more interesting, as we now have to take account of the correlation between fitted theory $T$ and the nuisance parameter $\lambda$: proceeding as in Eq.~(\ref{eq:varlampheno})
\begin{eqnarray}
\label{eq:covTlampheno}
{\rm Cov}[T(\lambda)] &=& {\rm E}[(T(\lambda) -{\rm E}[T(\lambda)])(T(\lambda) -{\rm E}[T(\lambda)])^T]\nonumber\\
&=& {\rm E}[(T -D +\lambda\beta)(T-D+\lambda\beta)^T]\nonumber\\
&=& {\rm E}[(T-D)(T-D)^T] + {\rm E}[\lambda \beta(T-D)^T] + {\rm E}[(T-D)\lambda \beta^T]+ {\rm E}[\lambda^2]\beta\beta^T.
\end{eqnarray}
Now the first term is just ${\rm Cov}[T]$, while the last is just ${\rm Var}[\lambda]S$, while the cross-terms can be evaluated using Eq.~(\ref{eq:lambdabar}):
\begin{eqnarray}
\label{eq:covTlamphenocross}
{\rm E}[\lambda \beta(T-D)^T] &=& {\rm E}[ \beta\overline{\lambda}(T,D)(T-D)^T] = -S(C+S)^{-1}{\rm E}[(T-D)(T-D)^T] \nonumber\\
&=& -S(C+S)^{-1}{\rm Cov}[T].
\end{eqnarray}
We thus find that
\begin{eqnarray}
{\rm Cov}[T(\lambda)] &=& {\rm Cov}[T]-S(C+S)^{-1}{\rm Cov}[T]-{\rm Cov}[T](C+S)^{-1}S+{\rm Var}[\lambda]S \label{eq:covTlamgen}\\
&=& (C+S) - S -S + S = C,\label{eq:covTlamphenoC}
\end{eqnarray}
where in the last line we used Eq.~(\ref{eq:varTpheno}) for ${\rm Cov}[T]$, and Eq.~(\ref{eq:varlampheno}) for ${\rm Var}[\lambda]$. Thus the covariance of the autopredictions in a perfect fit is simply the covariance of the data. The way in which this arises is that the theory covariance arising from the fit (the first term in Eq.(\ref{eq:covTlamgen}) and the theory covariance arising in the autoprediction (the last term in Eq.(\ref{eq:covTlamgen}) are each cancelled by the cross-covariance between fitting and prediction, just as was argued in Ref.~\cite{Harland-Lang:2018bxd}. When we have a perfect fit, there is really no distinction between the autoprediction and the data, and the theory uncertainty thus becomes irrelevant. So in a sense this model is pure phenomenology: the only nontrivial information is in the data. Indeed in this model as it stands there is no possibility of making genuine predictions, since a prediction of the form Eq.~(\ref{eq:Ttillambda}) is useless unless we can determine $\widetilde{T}$, presumably as functions of $T$, and for this we need a genuine theory.
\section{Correlated Theory Uncertainties in One Parameter Fits}
\label{sec:oneparameter}
In the previous section, we considered two simple but unrealistic models: the first in which the theory $T$ is fixed, with no free parameters to be fitted to the data (pure theory), and the second in which the theory $T$ is so flexible that we could achieve a perfect fit, $T=D$ (pure phenomenology). These exercises were useful, in that they gave us some practice in the use of nuisance parameters to propagate theoretical uncertainties. However we now need to consider the more realistic situation in which the theory has parameters that can be constrained by data, but is still sufficiently restrictive that it can be considered a theory. The fit to the data is then not perfect, but the theory is sufficiently constraining that it can be used to predict new observables, $\widetilde T$, for which we as yet have no data. We will find that the interesting features of the pure theory and pure phenomenology models (the shifts, the reduction in uncertainties due to Bayesian learning, and the correlations between theory uncertainties in the fitting and theory uncertainties in predictions) are also found in the more realistic theories.
In this section we consider a theory with only one fitted parameter: in Sec.~\ref{subsec:single} we explain for the fitting is performed using replicas, in Sec.~\ref{subsec:autoprediction} we consider autopredictions in such a theory, and in Sec.~\ref{subsec:prediction} we consider general predictions. Generalization to many fitting parameters will be considered in the following section.
\subsection{Fitting a Theory with a Single Parameter}
\label{subsec:single}
We can model this situation by considering theoretical predictions $T(\theta)$ which depend on a single parameter $\theta$, so that $\chi^2(\theta)$ is minimized for some choice of this parameter, $\theta=\theta_0$, with some variance ${\rm Var}[\theta]$. Other observables $\widetilde{T}(\theta)$ are then predicted to be $\widetilde{T}(\theta_0)$, with an associated uncertainty proportional to ${\rm Var}[\theta]$. We assume as before that uncertainties are Gaussian, which means that we can linearize $T(\theta)$ about $T(\theta_0)\equiv T_0$:
\begin{equation}
\label{eq:Tlin}
T(\theta) = T_0 + (\theta-\theta_0)\dot{T}_0.
\end{equation}
This model has the advantage that while it captures the essence of the fitting problem, it is sufficiently simple that we can solve it exactly.
In order to determine the uncertainty on $\theta$, we will need to propagate the experimental uncertainties in the data $D$ and the theoretical uncertainties in the the predictions $T(\theta)$ into $\theta$. This can be done most easily by generating $N_{\rm rep}$ pseudodata replicas $D^{(r)}$ distributed according to a Gaussian distribution centred on the actual data $D$, with covariance $C+S$: defining the replica average
\begin{equation}
\label{eq:repav}
\langle F(D^{(r)})\rangle = \lim_{N_{\rm rep}\to \infty}\smallfrac{1}{N_{\rm rep}}\sum_{r=1}^{N_{\rm rep}}F(D^{(r)})
\end{equation}
for any function $F$ of the replicas, the replicas are chosen such that
\begin{equation}
\label{eq:repavD}
\langle D^{(r)}\rangle \equiv D, \qquad \langle (D^{(r)}-D)(D^{(r)}-D)^T\rangle = C+S.
\end{equation}
A parameter replica $\theta^{(r)}$ is then fitted to each pseudodata replica $D^{(r)}$
by maximizing $P(T(\theta)|D^{(r)})$ as given by Eq.~(\ref{eq:PTD}), and thus by
minimizing
\begin{equation}
\label{eq:chi2rep}
\chi_r^2[\theta] = (T(\theta)-D^{(r)})^T(C+S)^{-1}(T(\theta)-D^{(r)}),
\end{equation}
with respect to $\theta$, replica by replica. Using Eq.~(\ref{eq:Tlin}), minimization of the quadratic gives
\begin{equation}
\label{eq:arep}
\theta^{(r)} - \theta_0 = \frac{\dot{T}_0^T(C+S)^{-1}(D^{(r)}-T_0)}{\dot{T}_0^T(C+S)^{-1}\dot{T}_0}.
\end{equation}
Using the replica averages Eq.~(\ref{eq:repavD}), and choosing $\theta_0 = \langle \theta^{(r)}\rangle$, we find for consistency
\begin{equation}
\label{eq:consistency}
\dot{T}_0^T(C+S)^{-1}(D-T_0)=0,
\end{equation}
and thus we can rewrite Eq.~(\ref{eq:arep}) as
\begin{equation}
\label{eq:arep2}
\theta^{(r)} - \theta_0 = \frac{\dot{T}_0^T(C+S)^{-1}(D^{(r)}-D)}{\dot{T}_0^T(C+S)^{-1}\dot{T}_0}.
\end{equation}
Then since $C$ and $S$ are symmetric matrices,
\begin{eqnarray}
{\rm Var}[\theta] &=& \langle(\theta^{(r)}-\theta_0)^2\rangle\nonumber\\
&=& \frac{\dot{T}_0^T(C+S)^{-1}\langle(D^{(r)}-D)(D^{(r)}-D)^T\rangle (C+S)^{-1}\dot{T}_0}{(\dot{T}_0^T(C+S)^{-1}\dot{T}_0)^2}\nonumber\\
&=& (\dot{T}_0^T(C+S)^{-1}\dot{T}_0)^{-1}.
\label{eq:vara}
\end{eqnarray}
Note the way the double reciprocation in this expression works: data points with a relatively large dependence on $\theta$ (i.e. large $\dot{T}_0$) contribute more than those with small dependence, however directions with large uncertainty (i.e. projections of $C+S$) contribute less than those with small uncertainty.
Now that we understand the uncertainty of the fitted parameter $\theta$, we can use it to predict the uncertainties on $T(\theta)$:
\begin{equation}
\label{eq:repET}
E[T] \equiv \langle T(\theta^{(r)})\rangle = T(\theta_0) = T_0,
\end{equation}
so, writing $T^{(r)} = T(\theta^{(r)})$,
\begin{eqnarray}
X\equiv{\rm Cov}[T(\theta)] &=& \langle(T^{(r)}-T_0)(T^{(r)}-T_0)^T\rangle\label{eq:Xdef}\\
&=& \dot{T}_0\langle(\theta^{(r)}-\theta_0)^2\rangle\dot{T}_0^T
= \dot{T}_0(\dot{T}_0^T(C+S)^{-1}\dot{T}_0)^{-1}\dot{T}_0^T\label{eq:Xdef2}\\
&=& n(n^T(C+S)^{-1}n)^{-1}n^T,
\label{eq:Xdef3}
\end{eqnarray}
where $n$ is a unit vector in the direction of $\dot{T}_0$, $n^Tn=1$. This shows that $X$ depends only on $n$, and
not on $|\dot{T}_0|$.
The singular matrix $X$ will play an important role in what follows: it is the covariance matrix of $T$ due to the experimental and theoretical uncertainties in the fitting of the parameter $\theta$ --- the `fitting uncertainty'. When the fitted parameter minimizes the $\chi^2$, and is thus given by Eq.~(\ref{eq:arep2}), $X$ satisfies the projective relation
\begin{equation}
\label{eq:XsqeqX}
X = X(C+S)^{-1}X.
\end{equation}
Using Eq.~(\ref{eq:arep2}) in Eq.~(\ref{eq:Tlin}), we see that $X(C+S)^{-1}$ projects the data replicas onto the theory replicas:
\begin{equation}
T^{(r)}-T_0 = X(C+S)^{-1}(D^{(r)}-D).
\label{eq:projection}
\end{equation}
Because this relation is projective, some information is lost whenever we perform the fit: Eq.~(\ref{eq:projection}) cannot be inverted to obtain data replicas from theory replicas. This is an inevitable consequence of describing $N$ data (assuming $N>1$) with only a single parameter $\theta$.
To make $X$ more explicit, consider the simple model Eq.~(\ref{eq:modelCS}) for $C$ and $S$. Then using Eq.~(\ref{eq:modelCplusSinv}),
\begin{equation}
\label{eq:denom}
n^T(C+S)^{-1}n= \frac{\sigma^2+s^2\sin^2\phi}{\sigma^2(\sigma^2+s^2)},
\end{equation}
where $\cos\phi = n^Te$, and thus if we project $X$ onto $n$
(projections orthogonal to $n$ give zero)
\begin{equation}
\label{eq:nXnmod}
n^TXn = \frac{\sigma^2(\sigma^2+s^2)}{(\sigma^2+s^2\sin^2\phi)}.
\end{equation}
The contribution of the theory uncertainty $s$ thus depends on how well aligned $e$ is to the direction $n$ of the parameter dependence: if $\phi=0$ we have complete alignment, and the variance of $T$ in this direction is $\sigma^2+s^2$ as expected, while if $\phi=\frac{\pi}{2}$ we have orthogonality, and the variance of $T$ is $\sigma^2$ --- the theory uncertainty is then irrelevant to the fitting.
\subsection{Autopredictions in Single Parameter Fits}
\label{subsec:autoprediction}
We can now consider the evaluation of the mean and covariance of the `autopredictions'
\begin{equation}
\label{eq:autopred}
T(\theta,\lambda)=T(\theta)+\lambda\beta
\end{equation}
in this one parameter model. Just as in Sec.~\ref{subsec:phenomenology}, we do this by first computing expectation values over $\lambda$, using $P(\lambda|TD)$, which depend on $T$, and then evaluate the expectation values over $T$, according to the probability distribution $P(T|D)$, now performed by averaging over theory replicas $T^{(r)} = T(\theta^{(r)})$. It is important to note that both these averages are performed holding the data $D$ fixed, as both probabilities are conditional on the data: the data replicas $D^{(r)}$ employed in Sec.~\ref{subsec:single} are only a device to generate the theory replicas $T^{(r)}$, and are not to be averaged over when determining expectation values. Accordingly, Eq.~(\ref{eq:Edef}) now becomes
\begin{equation}
\label{eq:Edef}
{\rm E}[f(T,\lambda)] =\Big\langle \Big({\cal N}_\lambda \int d\lambda\;f(T^{(r)},\lambda)\; P(\lambda|T^{(r)}D)\Big)\Big\rangle,
\end{equation}
where the angled brackets denote the replica average Eq.~(\ref{eq:repav}).
Following the same steps as in the perfect fit in Sec.~\ref{subsec:phenomenology}, but now using the theory replicas $T^{(r)} = T(\theta^{(r)})$ determined in the one parameter fit Sec.~\ref{subsec:single}, we find
\begin{equation}
\label{eq:explam}
E[\lambda] \equiv \langle \overline\lambda(T(\theta^{(r)}),D)\rangle = \beta^T(C+S)^{-1}(D-T_0)\equiv\overline\lambda_0,
\end{equation}
Unlike in the perfect fit Eq.~(\ref{eq:Elampheno}), but just as in the pure theory Eq.~(\ref{eq:Elampt}) the nuisance parameters can thus now have nonzero expectation values, since the one parameter fit no longer fits the data exactly. These in turn give nontrivial shifts in the theoretical predictions:
\begin{equation}
\label{eq:ET}
E[T(\theta,\lambda)] = \langle T^{(r)}+\overline\lambda(T^{(r)},D)\beta\rangle
= T_0+\overline\lambda_0\beta = T_0+\beta\beta^T(C+S)^{-1}(D-T_0).
\end{equation}
So again the data give us information, inducing shifts in the autopredictions:
\begin{equation}
\label{eq:shift}
\delta T = -S(C+S)^{-1}(T_0-D).
\end{equation}
Note however that since (from Eq.~(\ref{eq:consistency})) $n^T(C+S)^{-1}(T_0-D)=0$, these shifts will only be nonzero when $n$ and $e$ (the data and the theory) point in different directions: when they are parallel ($\phi=0$), the theoretical uncertainty is simply absorbed by the fit, just as it was in the perfect fit in Sec.~\ref{subsec:phenomenology}. We can use the same argument as in Sec.~\ref{subsec:puretheory}, Eq.~(\ref{eq:chisqauto}), to show that the shifts will always improve the fit to the experimental data.
For the uncertainties, consider first the variance of $\lambda$: following the same argument that led to Eq.~(\ref{eq:varlampheno}) in Sec.~{subsec:phenomenology}, we now
find
\begin{eqnarray}
{\rm Var}[\lambda] &=& {\rm E}[(\lambda -{\rm E}[\lambda])^2]= {\rm E}[(\lambda -\overline{\lambda}(T,D)+ \overline{\lambda}(T,D)-\overline\lambda_0)^2]\nonumber\\
&=& {\rm E}[(\lambda -\overline{\lambda}(T,D))^2]+ \langle(\overline{\lambda}(T^{(r)},D)-\overline\lambda_0)^2\rangle\nonumber\\
&=& Z + \beta^T(C+S)^{-1}\langle(T^{(r)}-T_0)(T^{(r)}-T_0)^T\rangle(C+S)^{-1}\beta\nonumber\\
&=& 1 - \beta^T(C+S)^{-1}\beta + \beta^T(C+S)^{-1}X(C+S)^{-1}\beta\equiv \overline{Z} .\label{eq:Zbardef}
\end{eqnarray}
where in the second line we turned the expectation value over $T$ in to a replica average, and in the last line we used Eq.~(\ref{eq:Zdef}) for $Z$, and Eq.~(\ref{eq:Xdef}) for ${\rm Cov}[T]$. We thus find that in the more restrictive environment of the one parameter fit, the last two terms no longer cancel: the information in the data can no longer be entirely absorbed in the single fitted parameter, and so it can still inform the nuisance parameter.
It is easy to see that $\overline{Z}\geq Z$ because $(C+S)^{-1}X(C+S)^{-1}$ is positive semi-definite, while $\overline{Z}\leq 1$ since $X(C+S)^{-1}$
is projective, Eq.~(\ref{eq:XsqeqX}, so its eigenvalues are either zero or one. So in place of Eq.~(\ref{zbounds}) we now have
\begin{equation}
\label{Zbarbounds}
0<Z\leq\overline{Z}\leq 1.
\end{equation}
The information on theoretical uncertainties extracted from the data is thus less in the one parameter fit than it was in the pure theory of Sec.~\ref{subsec:puretheory}, due to the extra uncertainty arising in the fit itself, but unlike in the perfect fit Sec.~\ref{subsec:phenomenology}, the data will still constrain the theoretical uncertainties provided the parameter and theoretical uncertainty act in different directions.
In the model Eq.~(\ref{eq:modelCS}) for $C$ and $S$, and using Eq.~(\ref{eq:nXnmod}) for $X$, Eq.~(\ref{eq:Zbardef}) becomes
\begin{equation}
\label{eq:Zbarmod}
\overline{Z} = \frac{\sigma^2}{\sigma^2+s^2\sin^2\phi}.
\end{equation}
Comparing with the corresponding expression for $Z$, Eq.~(\ref{eq:modelZ}), we see that indeed $\overline{Z}=1$ when $\phi=0$, thus when $n=e$ and the parameter variation and theoretical uncertainty are aligned, while $\overline{Z}=Z$ only if $\phi=\pi/2$, so when $n$ and $e$ are orthogonal, the data have the greatest influence on the uncertainty.
For the covariance of the autopredictions Eq.~(\ref{eq:autopred}) we also have to take account of the correlation between the fitted theory and the nuisance parameter: proceeding as in Eq.~(\ref{eq:covTlampheno})
\begin{eqnarray}
{\rm Cov}[T(\theta,\lambda)] &=& {\rm E}[(T(\theta,\lambda) -{\rm E}[T(\theta,\lambda)])(T(\theta,\lambda) -{\rm E}[T(\theta,\lambda)])^T]\nonumber\\
&=& {\rm E}[(T -T_0 +(\lambda-\overline\lambda_0)\beta)(T-T_0+(\lambda-\overline\lambda_0)\beta)^T]\nonumber\\
&=& \langle(T^{(r)}-T_0)(T^{(r)}-T_0)^T\rangle + {\rm E}[(\lambda-\overline\lambda_0) \beta(T-T_0)^T] \nonumber\\
&&\qquad + {\rm E}[(T-T_0)(\lambda-\overline\lambda_0) \beta^T]+ {\rm E}[(\lambda -\overline\lambda_0)^2]\beta\beta^T.\label{eq:covTlamonex}
\end{eqnarray}
Then again the first term is ${\rm Cov}[T]=X$, Eq.~(\ref{eq:Xdef}), while the last is ${\rm Var}[\lambda]S$, Eq.~(\ref{eq:Zbardef}), while the cross-terms can be evaluated using Eq.~(\ref{eq:lambdabar}):
\begin{eqnarray}
\label{eq:covTlamphenocross}
{\rm E}[(\lambda-\overline\lambda_0) \beta(T-T_0)^T] &=& \langle \beta(\overline{\lambda}(T^{(r)},D)-\overline{\lambda}(T_0,D)(T^{(r)}-T_0)^T\rangle \nonumber\\
&=& -S(C+S)^{-1}\langle(T^{(r)}-T_0)(T^{(r)}-T_0)^T\rangle \nonumber\\
&=& -S(C+S)^{-1}{\rm Cov}[T].
\end{eqnarray}
We thus find that the result is simply
\begin{equation}
{\rm Cov}[T(\lambda)] = X-S(C+S)^{-1}X-X(C+S)^{-1}S+\overline{Z} S. \label{eq:covTlamone}
\end{equation}
The meaning of the four terms is easy to understand: the first is the `fitting uncertainty' (which includes contributions from both experimental and theoretical uncertainties), the last the `theory uncertainty', reduced through exposure to the data, and the middle two terms are due to the correlations between the two sources of theoretical uncertainty.
We can simplify it by using Eq.~(\ref{eq:Zbardef}) to show that
\begin{equation}
\label{eq:Zbaralgebra}
\overline{Z} S = S(C+S)^{-1}X(C+S)^{-1}S + ZS,
\end{equation}
and then some straightforward algebra to write
\begin{equation}
\label{eq:Xalgebra}
X-S(C+S)^{-1}X-X(C+S)^{-1}S+S(C+S)^{-1}X(C+S)^{-1}S = C(C+S)^{-1}X(C+S)^{-1}C.
\end{equation}
We thus find finally
\begin{equation}
{\rm Cov}[T(\lambda)] = C(C+S)^{-1}X(C+S)^{-1}C + ZS.\label{eq:covTlamfin}
\end{equation}
Note that if we write $X=C+S$ (as in the perfect fit model in Sec.~\ref{subsec:phenomenology}), this result reduces to $C$, as it should. The cancellations noted in Eq.~(\ref{eq:covTlamphenoC}) between the cross-terms and the covariances of of $T$ and $\lambda$ are no longer exact in Eq.~(\ref{eq:covTlamfin}) , because ${\rm Cov}[T]$ is no longer $C+S$, but rather the smaller matrix $X$ (which is in a sense $C+S$ restricted to the space of variation of the parameter $\theta$ as in Eq.~~(\ref{eq:Xdef})). Thus the result is no longer the experimental covariance matrix $C$, but rather the sum in quadrature of the `fitting uncertainty', $X$, and the `theory uncertainty' $S$, each reduced to some extent by the correlation of the theoretical uncertainties in fit and prediction.
In the model Eq.~(\ref{eq:modelCS}) for $C$ and $S$, and using Eq.~(\ref{eq:nXnmod}) for $X$, we now find
\begin{equation}
{\rm Cov}[T(\lambda)] = \frac{\sigma^2(\sigma^2+s^2)}{(\sigma^2+s^2\sin^2\phi)}
\Big(nn^T - \frac{s^2}{\sigma^2+s^2}\cos\phi(en^T+ne^T) + \frac{s^2}{\sigma^2+s^2}ee^T\Big) .\label{eq:covTlamonemod}
\end{equation}
The first term is just $X$, the off-diagonal term in the middle is the correlation term, and the last is $\overline{Z} S$. If $\phi=0$, so $n=e$, the three terms combine to give simply $\sigma^2nn^T$: in the direction of the fitted parameter, the uncertainty in the autoprediction is the experimental uncertainty, just as in the perfect fit described in Sec.~\ref{subsec:phenomenology}. On the other hand, if $\phi=\pi/2$, so $n$ and $e$ are orthogonal, the correlation term disappears, and the result reduces to $X+ZS$: we add the uncertainties in quadrature, since they are in orthogonal directions, and the theoretical uncertainty is reduced just as in the pure theory described in Sec.~\ref{subsec:puretheory}. The one parameter fit thus interpolates smoothly between these two extremes. Note that we can write Eq.~(\ref{eq:covTlamonemod}) as
\begin{equation}
{\rm Cov}[T(\lambda)] = \frac{\sigma^2(\sigma^2+s^2)}{(\sigma^2+s^2\sin^2\phi)}
\Big(n - \frac{s^2\cos\phi}{\sigma^2+s^2} e\Big)\Big(n^T - \frac{s^2\cos\phi}{\sigma^2+s^2}e^T\Big) + \frac{s^2\sigma^2}{\sigma^2+s^2}ee^T .\label{eq:covTlamonemodx}
\end{equation}
Here the last term is just $ZS$, while the first is $X$, but with the vector $n$ given an additional component in the direction $e$ due to the theory correlation: besides changing its direction, this reduces the size of the fitting uncertainty by a factor $\sqrt{\sin^2\phi + Z^2\cos^2\phi}$. However it is easy to see that the size of the correlated fitting uncertainty is still larger than it would be if the theory uncertainty had not been included in the fit.
\subsection{Correlated Predictions in One Parameter Fits}
\label{subsec:prediction}
We now consider predictions $\widetilde{T}_I(\theta)$, $I=1,\ldots,\widetilde{N}$, which depend on the same parameter $\theta$ as the fitted predictions $T_i(\theta)$, $i=1,\ldots, N$. There are two distinct sources of uncertainty in $\widetilde{T}_I(\theta)$: uncertainties in the determination of $\theta$ due to the experimental uncertainties in the data $D_i$ and theoretical uncertainties in the theory $T_i(\theta)$ used in its determination; and theoretical uncertainties in the predictions $\widetilde{T}_I(\theta)$ themselves.
The first uncertainty is expressed through Eq.~(\ref{eq:arep2}), which gives the variance Eq.~(\ref{eq:vara}). In analogy with Eq.~(\ref{eq:Tlin}) the linearised dependence of the predictions $\widetilde{T}(\theta)$ may be written
\begin{equation}
\label{eq:Tlin2}
\widetilde{T}(\theta) = \widetilde{T}_0 + (\theta-\theta_0)\skew6\dot\Ttil_0,
\end{equation}
with $\widetilde{T}_0=\widetilde{T}(\theta_0)$. Then the covariance of $\widetilde{T}_I(\theta)$ due to the uncertainty in the parameter $\theta$ is derived just as in Eqs.~(\ref{eq:Xdef}-\ref{eq:Xdef3}) for the autopredictions: writing $\widetilde{T}^{(r)}\equiv\widetilde{T}(\theta^{(r)})$
\begin{eqnarray}
\widetilde{X}\equiv{\rm Cov}[\widetilde{T}(\theta)] &=& \langle(\widetilde{T}^{(r)}-\widetilde{T}_0)(\widetilde{T}^{(r)}-\widetilde{T}_0)^T\rangle\label{eq:Xtildef}\\
&=& \skew6\dot\Ttil_0\langle(\theta^{(r)}-\theta_0)^2\rangle\skew6\dot\Ttil_0^T
= \skew6\dot\Ttil_0(\dot{T}_0^T(C+S)^{-1}\dot{T}_0)^{-1}\skew6\dot\Ttil_0^T\label{eq:Xtildef2}.
\end{eqnarray}
The second uncertainty --- the theoretical uncertainty in the predictions $\widetilde{T}_I(\theta)$ --- may again be either correlated or uncorrelated with the theoretical uncertainty in $T_i(\theta)$. Consider first the simpler situation when it is uncorrelated: this might be the case if, for example, the observable $\widetilde{T}(\theta)$ was a different type of observable to the $T(\theta)$ used to determine $\theta$. Then
introducing a nuisance parameter $\widetilde{\lambda}$, Gaussian distributed about zero with unit variance, and uncorrelated with $\lambda$, we can write (as in the pure theory model Eq.~(\ref{eq:Ttillambda}))
\begin{equation}
\widetilde{T}(\theta,\widetilde{\lambda}) = \widetilde{T}(\theta) + \widetilde{\lambda}\widetilde{\beta},\label{eq:uncor}
\end{equation}
where the vector $\widetilde{\beta}_I$ gives the size of the theoretical uncertainties
in $\widetilde{T}_I(\theta)$.
Since $\theta$ and $\widetilde{\lambda}$ are uncorrelated, we then have
\begin{eqnarray}
E[\widetilde{T}(\theta,\widetilde{\lambda})] &=& \widetilde{T}(\theta_0),\label{eq:uncorpred}\\
{\rm Cov}[\widetilde{T}(\theta,\widetilde{\lambda})]&=& {\rm Cov}[\widetilde{T}(\theta)]+{\rm Var}[\widetilde{\lambda}]\widetilde{\beta}\betatil^T
= \widetilde{X} + \widetilde{S},\label{eq:uncorcov}
\end{eqnarray}
where $\widetilde{S} = \widetilde{\beta}\betatil^T$
is the theory covariance matrix for the prediction $\widetilde{T}(\theta)$ (compare Eq.~(\ref{eq:Stildef}) in the pure theory). Thus when the theoretical uncertainty is uncorrelated we simply add it in quadrature to the uncertainty due to that in the parameter $\theta$ derived from the fit.
Now consider the more interesting case in which the theoretical uncertainty in
$\widetilde{T}_I(\theta)$ is fully correlated to that in the $T_i(\theta)$ used in the fit to determine $\theta$: then $\widetilde{\lambda}=\lambda$, which has already been determined in the fit to have nonzero expectation value and variance Eqs.~(\ref{eq:explam},\ref{eq:Zbardef}).
Then writing $\widetilde{T}(\theta,\lambda) = \widetilde{T}(\theta)+\lambda\widetilde{\beta}$,
\begin{equation}
\label{eq:corpredmean}
E[\widetilde{T}(\theta,\lambda)] = \widetilde{T}_0 + \overline\lambda(T_0,D)\widetilde{\beta},
\end{equation}
so the correlation induces a similar shift in the predictions to that in the autopredictions Eq.~(\ref{eq:shift}): using Eq.~(\ref{eq:lambdabar})
\begin{equation}
\label{eq:shiftpred}
\delta \widetilde{T}(\theta_0) = \widetilde{\beta}\beta^T(C+S)^{-1}(D-T_0) = -\widehat{S} (C+S)^{-1}(T_0-D).
\end{equation}
where $\widehat{S} = \widetilde{\beta}\beta^T$, Eq.~(\ref{eq:Shatdef}) is the matrix of cross-correlations between observables $T(\theta)$ used in the fit and the predictions $\widetilde{T}(\theta)$ .
Likewise using the same arguments as were used to derive ${\rm Cov}[T(\theta,\lambda)]$, Eq.~(\ref{eq:covTlamone})
\begin{eqnarray}
{\rm Cov}[\widetilde{T}(\theta,\lambda)] &=& {\rm E}[(\widetilde{T}(\theta,\lambda) -{\rm E}[\widetilde{T}(\theta,\lambda)])(\widetilde{T}(\theta,\lambda) -{\rm E}[\widetilde{T}(\theta,\lambda)])^T]\nonumber\\
&=& \langle\widetilde{T}^{(r)}-\widetilde{T}_0)(\widetilde{T}^{(r)}-\widetilde{T}_0)^T\rangle + {\rm E}[(\lambda-\overline\lambda_0) \widetilde{\beta}(\widetilde{T}-\widetilde{T}_0)^T] \nonumber\\
&&\qquad + {\rm E}[(\widetilde{T}-\widetilde{T}_0)(\lambda-\overline\lambda_0) \widetilde{\beta}^T]+ {\rm E}[(\lambda -\overline\lambda_0)^2]\widetilde{\beta}\betatil^T.\label{eq:covTtillamonex}
\end{eqnarray}
Then again the first term is ${\rm Cov}[\widetilde{T}]=\widetilde{X}$, Eq.~(\ref{eq:Xtildef}), while the last is ${\rm Var}[\lambda]\widetilde{S}$, Eqs.~(\ref{eq:Zbardef},\ref{eq:Stildef}), while the cross-terms can be evaluated using Eq.~(\ref{eq:lambdabar}):
\begin{eqnarray}
\label{eq:covTtillamphenocross}
{\rm E}[(\lambda-\overline\lambda_0) \widetilde{\beta}(\widetilde{T}-\widetilde{T}_0)^T] &=& \langle \widetilde{\beta}(\overline{\lambda}(T^{(r)},D)
-\overline{\lambda}(T_0,D))(\widetilde{T}^{(r)}-\widetilde{T}_0)^T\rangle \nonumber\\
&=& -\widehat{S}(C+S)^{-1}\langle(T^{(r)}-T_0)(\widetilde{T}^{(r)}-\widetilde{T}_0)^T\rangle\nonumber\\ &=& -\widehat{S}(C+S)^{-1}\widehat{X}^T,
\end{eqnarray}
where $\widehat{S}$ is given by Eq.~(\ref{eq:Shatdef}), and in analogy to Eq.~(\ref{eq:Xdef}) and Eq.~(\ref{eq:Xtildef}) we define the cross-covariance between $\widetilde{T}$ and $T$
\begin{eqnarray}
\widehat{X} &\equiv& \langle(\widetilde{T}^{(r)}-\widetilde{T}_0)(T^{(r)}-T_0)^T\rangle\label{eq:Xhatdef}\\
&=& \skew6\dot\Ttil_0\langle(\theta^{(r)}-\theta_0)^2\rangle\dot{T}_0^T=\skew6\dot\Ttil_0(\dot{T}_0^T(C+S)^{-1}\dot{T}_0)^{-1}\dot{T}_0^T\label{eq:Xhatdef2}.
\end{eqnarray}
We thus find that
\begin{equation}
{\rm Cov}[\widetilde{T}(\theta,\lambda)] = \widetilde{X}-\widehat{S}(C+S)^{-1}\widehat{X}^T-\widehat{X}(C+S)^{-1}\widehat{S}^T+\overline{Z} \widetilde{S}. \label{eq:covTtillamone}
\end{equation}
Using Eq.~(\ref{eq:Zbardef}), we can write the last term as
\begin{eqnarray}
\overline{Z} \widetilde{S} &=& Z\widetilde{S} + \widehat{S}(C+S)^{-1}X(C+S)^{-1}\widehat{S}^T,\label{eq:ZbarStil}\\
Z\widetilde{S} &=&\widetilde{S} - \widehat{S}(C+S)^{-1}\widehat{S}^T.
\label{eq:ZStil}
\end{eqnarray}
Note that the coefficients $Z$ and $\overline{Z}$ are the same as for the autopredictions, and thus satisfy the bounds Eq.~(\ref{Zbarbounds}), we must have (for positive definite $C$ and positive semi-definite $S$, i.e. $C>0$, $S\geq 0$)
\begin{equation}
\label{eq:possemdeftil}
0\leq \widehat{S}(C+S)^{-1}X(C+S)^{-1}\widehat{S}^T\leq \widehat{S}(C+S)^{-1}\widehat{S}^T \leq \widetilde{S},
\end{equation}
so in particular the subtraction (the last term in Eq.~(\ref{eq:ZStil}) can never be so large that it makes the entire covariance matrix negative.
In summary, comparing Eqs.~(\ref{eq:shiftpred},\ref{eq:covTtillamone}) with
Eqs.~(\ref{eq:uncorpred},\ref{eq:uncorcov}), we see that including the correlations between the theoretical uncertainties in the fit and the prediction results in three effects: a shift in the central value of the prediction, a reduction in the theoretical uncertainty, and a reduction in the fitting uncertainty due to the correlations. Performing the fit gives us information (from the data) about the theory, which results in more precise, and hopefully more accurate, predictions.
\section{Correlated MHOU in PDF fits}
\label{sec:corrlnpdffits}
We now repeat the above analysis, but instead of the toy model we consider the more realistic situation in which the theoretical expressions $T_i[f]$ depend on PDFs $f$, determined in a global fit to $N$ data $D_i$, with experimental covariance matrix $C_{ij}$, and then used to make $\widetilde{N}$ predictions $\widetilde{T}_I[f]$. There are then many sources of theoretical uncertainty in the relation between the theoretical calculations and the PDFs: here we consider the most generic, the missing higher order uncertainty (MHOU), computed using scale variations according to one of the prescriptions set out in Ref.~\cite{AbdulKhalek:2019bux,AbdulKhalek:2019ihb}. The theory covariance matrices $S_{ij}$ and $\widetilde{S}_{IJ}$ associated with the MHOU will then have many non-zero eigenvalues, and thus there will be $n$ nuisance parameters $\lambda_\alpha$, $\alpha=1,\ldots, n$ to take into account. There is in principle no limit on $n$, though in practice $n\ll N$. As in the toy model, the fit to the PDFs will determine the mean and covariance of the nuisance parameters, which will then translate into systematic shifts and changes in the uncertainties of the theoretical predictions.
\subsection{Expectation and Covariance of Multiple Nuisance Parameters}
\label{subsec:multiplenuisance}
The nuisance parameters $\lambda_\alpha$ correspond to shifts in the theoretical predictions: $T_i[f]\to T_i[f] + \lambda_\alpha\beta_{i,\alpha}[f]$, where we adopt the summation convention for the index $\alpha$. The shift vectors $\beta_{i,\alpha}$ are not necessarily orthogonal to each other. We again assume Gaussian uncertainties, so that in place of Eq.~(\ref{eq:PTDl}) we now have
\begin{equation}
\label{eq:PTDlf}
P(T|D\lambda)\propto \exp\big(-\smallfrac{1}{2}(T[f]+\lambda_\alpha\beta_\alpha-D)^TC^{-1}(T[f]+\lambda_\alpha\beta_\alpha-D)\big),
\end{equation}
and assume that each nuisance parameter has a prior which is Gaussian distributed with unit variance, centred on zero, the distributions being independent both of each other and of the data, so that
\begin{equation}
\label{eq:priorf}
P(\lambda|D)=P(\lambda) \propto \exp\big(-\smallfrac{1}{2}\lambda_\alpha\lambda_\alpha\big).
\end{equation}
We now marginalize over $\lambda_\alpha$, as in Eq.~(\ref{eq:marginalise1})
\begin{equation}
\label{eq:marginalise2}
P(T|D) \propto\int d^n\lambda\, \exp\left(-\smallfrac{1}{2}[(T[f]+\lambda_\alpha\beta_\alpha-D)^TC^{-1}(T[f]+\lambda_\beta\beta_\beta-D)+\delta_{\alpha\beta}\lambda_\alpha\lambda_\beta]\right)\, ,
\end{equation}
by first completing the square: the details are messy, but the result is very similar to Eq.~(\ref{eq:integration}), namely
\begin{equation}
\label{eq:integrationf}
P(T|D)\propto\int d^n\lambda\, \exp\left(-\smallfrac{1}{2}(\lambda_\alpha-\overline{\lambda}_\alpha) Z_{\alpha\beta}^{-1}(\lambda_\beta-\overline{\lambda}_\beta) - \smallfrac{1}{2}\chi^2\right) \propto \exp(-\smallfrac{1}{2}\chi^2),
\end{equation}
where now
\begin{equation}
\label{eq:Zdefmat}
Z_{\alpha\beta} = (\delta_{\alpha\beta}+\beta_\alpha^TC^{-1}\beta_\beta)^{-1},
\end{equation}
the inverse on the right hand side being the matrix inverse with respect to the indices $\alpha$ and $\beta$,
\begin{equation}
\label{eq:lambdabarf}
\overline{\lambda}_\alpha(T,D) = Z_{\alpha\beta}\beta_\beta^TC^{-1}(D-T),
\end{equation}
and $\chi^2$ is once again given by Eq.~(\ref{eq:chisq}), but now with in place of Eq.~(\ref{eq:Sdef})
\begin{equation}
\label{eq:Sdeff}
S = \beta_\alpha\beta^T_\alpha,
\end{equation}
as expected. The Gaussian integration in Eq.~(\ref{eq:integrationf}) is now trivial, taking us back again to Eq.~(\ref{eq:PTD}) up to a factor $(2\pi)^{n/2}({\rm det}Z)^{1/2}$, which we can ignore as it does not depend on $T$ or $D$, while Bayes' Theorem Eq.~(\ref{eq:bayes}) gives us the posterior distribution of the nuisance parameters:
\begin{equation}
\label{eq:posteriorf}
P(\lambda|TD)\propto \exp\big(-\smallfrac{1}{2}(\lambda_\alpha-\overline{\lambda}_\alpha) Z_{\alpha\beta}^{-1}(\lambda_\beta-\overline{\lambda}_\beta)\big),
\end{equation}
whence we see that
\begin{equation}
\label{eq:meanvarlamf}
E[\lambda_\alpha] =\overline{\lambda}_\alpha,\qquad E[(\lambda_\alpha-\overline{\lambda}_\alpha)(\lambda_\beta-\overline{\lambda}_\beta)] = Z_{\alpha\beta}.
\end{equation}
It is easy to see from the definition Eq.~(\ref{eq:Zdefmat}) that if $e_\alpha$ is a unit eigenvector of $Z_{\alpha\beta}$, and $\beta=e_\alpha\beta_\alpha$, then the corresponding eigenvalue of $Z_{\alpha\beta}$ is $z = (1+\beta^TC^{-1}\beta)^{-1}$, so $0<z<1$, and (in analogy to the bounds Eq.~(\ref{zbounds}) $Z_{\alpha\beta}$ is positive definite (thus invertible) and $\delta_{\alpha\beta}-Z_{\alpha\beta}$ is also positive definite (because the eigenvalues $z$ are all less than one). We can summarise this by writing, in place of Eq.~(\ref{zbounds})
\begin{equation}
0 < Z_{\alpha\beta} < \delta_{\alpha\beta}.
\label{eq:Zabbounds}
\end{equation}
We can express $Z_{\alpha\beta}$ in terms of the inverse of $C+S$:
\begin{equation}
\label{eq:Zdefab}
Z_{\alpha\beta} = \delta_{\alpha\beta}-\beta_\alpha^T(C+S)^{-1}\beta_\beta,
\end{equation}
since
\begin{eqnarray}
\label{eq:invalgebra}
&&(\delta_{\alpha\gamma}+\beta_\alpha^TC^{-1}\beta_\gamma)(\delta_{\gamma\beta}-\beta_\gamma^T(C+S)^{-1}\beta_\beta)\nonumber\\
&&\qquad\qquad\qquad=\delta_{\alpha\beta} + \beta_\alpha^T(C^{-1}-(C+S)^{-1}-C^{-1}S(C+S)^{-1})\beta_\beta = \delta_{\alpha\beta}.
\end{eqnarray}
Combining Eq.~(\ref{eq:Zdefab}) with Eq.~(\ref{eq:lambdabarf}), we then have
\begin{equation}
\label{eq:lambdabarfx}
\overline{\lambda}_\alpha = \beta_\alpha^T(C+S)^{-1}(D-T),
\end{equation}
since $(1-(C+S)^{-1}S)C^{-1} = (C+S)^{-1}$.
\subsection{Fitting the PDFs}
\label{subsec:pdfexactfit}
We now proceed to apply the above results in the context of a PDF fit incorporating MHOU~\cite{AbdulKhalek:2019bux,AbdulKhalek:2019ihb}. We consider first a fixed parametrization: the PDFs $f(\theta)$ will then depend on $m$ parameters $\theta_p$, $p = 1,\ldots,m$, with $m<N$ so the data $D$ are sufficient to determine all the parameters through minimization of the $\chi^2$
Eq.~(\ref{eq:chisq}).
We can then follow the same procedure as in Sec.\ref{sec:oneparameter}, the only difference being that now we fit the $m$ parameters $\theta_p$ rather than just the single parameter $\theta$. Writing the theoretical predictions $T[f(\theta)]\equiv T(\theta)$, the linearization relation Eq.~(\ref{eq:Tlin2}) becomes
\begin{equation}
\label{eq:Tlinf}
T(\theta) = T_0 + (\theta_p-\theta_p^0)T_p,
\end{equation}
where $f(\theta^0)\equiv f_0$ is the PDF that minimizes the $\chi^2$, $T_0\equiv T(\theta^0)$, $T_p\equiv \partial T(\theta^0)/\partial\theta_p^0$, and we use the summation convention for indices $p$. Using the data replicas Eq.~(\ref{eq:repav},\ref{eq:repavD}), minimizing Eq.~(\ref{eq:chi2rep}) with respect to $\theta_p$ rather than $\theta$, we find in place of Eq.~(\ref{eq:arep2}) that the
fluctuations of the PDF replica parameters are given by
\begin{equation}
\label{eq:arep2f}
\theta^{(r)}_p - \theta^0_p = (T_p^T(C+S)^{-1}T_q)^{-1}T_q^T(C+S)^{-1}(D^{(r)}-D),
\end{equation}
where the matrix inverse in the first factor on the right hand side is with respect to the $p,q$ indices. It follows that instead of Eq.~(\ref{eq:vara}) we now have the covariance matrix
\begin{eqnarray}
{\rm Cov}_{pq}[\theta] &=& \langle(\theta^{(r)}_p-\theta_p^0)(\theta^{(r)}_q-\theta_q^0)\rangle\nonumber\\
&=& (T_p^T(C+S)^{-1}T_q)^{-1},
\label{eq:varaf}
\end{eqnarray}
while the expression Eq.~(\ref{eq:Xdef3}) for the covariance of the predictions $T[f]$ becomes, on writing $T_p = |T_p|n_p$, where $n_p$ are unit vectors (which are however not necessarily orthogonal)
\begin{equation}
X = n_p(n_p^T(C+S)^{-1}n_q)^{-1}n_q^T.
\label{eq:Xdeffpq}
\end{equation}
so the projective relation Eq.~(\ref{eq:XsqeqX}) still holds, and $X(C+S)^{-1}$ projects data replicas onto theory replicas as in Eq.~(\ref{eq:projection}).
It is now easy to see that the results for the autopredictions in Sec.\ref{subsec:autoprediction} continue to hold, and that in particular that since the central values of the nuisance parameters
$\overline{\lambda}_\alpha$ are given by
\begin{equation}
\label{eq:Elambdaf}
{\rm E}[\lambda_\alpha] = -\beta_\alpha^T(C+S)^{-1}(\langle T^{(r)}\rangle-D),
\end{equation}
the shifts Eq.~(\ref{eq:shift}) are now
\begin{equation}
\label{eq:shiftmult}
\delta T[f] = \beta_\alpha\beta_\alpha^T(C+S)^{-1}(D-T[f_0]) = -S(C+S)^{-1}(T[f_0]-D).
\end{equation}
These shifts will improve the $\chi^2$ to the experimental data, in just the same way as in Sec.~\ref{subsec:puretheory}.
Likewise for the uncertainties: Eq.~(\ref{eq:Zbardef}) for the variance of the nuisance parameter becomes an equation for the covariance matrix of the nuisance parameters in the context of the PDF fit,
\begin{eqnarray}
{\rm Cov}_{\alpha\beta}[\lambda] &\equiv& {\rm E}[(\lambda_\alpha - {\rm E}[\lambda_\alpha])(\lambda_\beta - {\rm E}[\lambda_\beta])]\nonumber\\
&=&\delta_{\alpha\beta} -\beta_\alpha^T(C+S)^{-1}\beta_\beta-\beta_\alpha^T(C+S)^{-1}X(C+S)^{-1}\beta_\beta\equiv \overline{Z}_{\alpha\beta},
\label{eq:Zbardefab}\end{eqnarray}
using the projective relation Eq.~(\ref{eq:XsqeqX}). Again, both $\overline{Z}_{\alpha\beta}-Z_{\alpha\beta}$ and $\delta_{\alpha\beta}-\overline{Z}_{\alpha\beta}$ are positive semi-definite, so
\begin{equation}
0 < Z_{\alpha\beta} \leq \overline{Z}_{\alpha\beta} \leq\delta_{\alpha\beta}.
\label{eq:zbarabbounds}
\end{equation}
The covariance matrix of the theoretical autopredictions $T(f,\lambda)\equiv T[f]+\lambda_\alpha\beta_\alpha$, Eq.~(\ref{eq:covTlamone},\ref{eq:covTlamfin}) then become
\begin{eqnarray}
{{\rm Cov}}[T(f,\lambda)] &=& X - S(C+S)^{-1}X-X(C+S)^{-1}S+\beta_\alpha\overline{Z}_{\alpha\beta}\beta_\beta^T \label{eq:covTlamf}\\
&=& C(C+S)^{-1}X(C+S)^{-1}C + S - S(C+S)^{-1}S,\label{eq:covTfitf}
\end{eqnarray}
the second expression being identical to the one we found in Sec.~\ref{subsec:autoprediction}.
The same holds true of course for correlated predictions:
the shifts Eq.~(\ref{eq:shiftpred}) are now
\begin{equation}
\label{eq:shiftpredf}
\delta \widetilde{T}[f] = \widetilde{\beta}_\alpha\beta_\alpha^T(C+S)^{-1}(D-T[f_0]) = \widehat{S} (C+S)^{-1}(D-T[f_0]),
\end{equation}
where $\widehat{S} = \widetilde{\beta}_\alpha\beta_\alpha^T$, while if $\widetilde{T}(f,\lambda)=\widetilde{T}[f]+\lambda_\alpha\widetilde{\beta}_\alpha$, Eq.~(\ref{eq:covTtillamone} becomes
\begin{equation}
{{\rm Cov}}[\widetilde{T}(f,\lambda)]
= \widetilde{X} - \widehat{S}(C+S)^{-1}\widehat{X}^T-\widehat{X}(C+S)^{-1}\widehat{S}^T+ \widetilde{\beta}_\alpha\overline{Z}_{\alpha\beta}\widetilde{\beta}_\beta^T \label{eq:covTtillamf},
\end{equation}
where $\widetilde{S} = \widetilde{\beta}_\alpha\widetilde{\beta}_\alpha^T$, and
\begin{eqnarray}
\widetilde{X}&=& \widetilde{T}_p(T_p^T(C+S)^{-1}T_q)^{-1}\widetilde{T}_q^T,\label{eq:Xtildeff}\\
\widehat{X}&=& \widetilde{T}_p(T_p^T(C+S)^{-1}T_q)^{-1}T_q^T.\label{eq:Xhatdeff}
\end{eqnarray}
Using Eq.~(\ref{eq:Zbardefab}), we can write the last term as
\begin{eqnarray}
\widetilde{\beta}_\alpha\overline{Z}_{\alpha\beta}\widetilde{\beta}_\beta^T &=& \widetilde{\beta}_\alpha Z_{\alpha\beta}\widetilde{\beta}_\beta^T + \widehat{S}(C+S)^{-1}X(C+S)^{-1}\widehat{S}^T,\label{eq:ZbarStilab}\\
\widetilde{\beta}_\alpha Z_{\alpha\beta}\widetilde{\beta}_\beta^T &=&\widetilde{S} - \widehat{S}(C+S)^{-1}\widehat{S}^T.
\label{eq:ZStilab}
\end{eqnarray}
The final result thus again has exactly the same for as that found in Sec.~\ref{subsec:prediction}: once the nuisance parameters are all eliminated, the only changes are in the expressions for the covariances $X$, $\widetilde{X}$ and $\widehat{X}$, generalizing the previous one parameter expressions to many parameters.
\subsection{Fitting NNPDFs}
\label{subsec:fittingnnpdfs}
PDFs in an NNPDF fit are parametrized by a neural network, with a very large number of parameters. The fitting procedure differs from that using a fixed parametrization, since we want to avoid fitting noise. In practice this is achieved using a cross--validation procedure. It follows that when we fit to each data replica $D^{(r)}$, the neural net parameters, and thus the PDF replicas $f^{(r)}$, are not precisely determined through exact minimization of the $\chi^2$, but rather include some random noise, which is responsible for the `functional uncertainty' inherent in the fit \cite{Ball:2014uwa}. It is not easy to describe this analytically: all we can say is that while all the general results in Sec.\ref{subsec:multiplenuisance} remain valid, relations such as Eq.~(\ref{eq:arep2f}) for the fitted parameters, and thus the subsequent results Eqs.~(\ref{eq:varaf},\ref{eq:Xdeffpq}), no longer hold. However we can still use Eq.~(\ref{eq:lambdabarfx}) to compute the expectation and covariance of the nuisance parameters, and obtain the same results Eqs.~(\ref{eq:varaf},\ref{eq:Zbardefab}), provided we define $T^{(r)}\equiv T[f^{(r)}]$, and $T^{(0)}\equiv \langle T^{(r)}\rangle$
\begin{equation}
X\equiv{\rm Cov}[T[f]] = \langle (T^{(r)} - T^{(0)}) (T^{(r)} - T^{(0)})^T\rangle
\label{eq:XdefNN}
\end{equation}
as averages over the PDF replicas. This matrix gives the PDF uncertainties (and correlations) for the observables $T[f]$, which includes both the experimental uncertainties in the data and the theoretical uncertainties in extracting the PDFs from the data.
Note that in an NNPDF fit $X$ no longer satisfies the projective relation Eq.~(\ref{eq:XsqeqX}), and indeed $X(C+S)^{-1}$ no longer projects data replicas directly onto theory replicas as in Eq.~(\ref{eq:projection}). It can confirm this for a given set of PDF replicas by computing the cross-covariance matrix
\begin{equation}
Y \equiv {\rm Cov}[T,D]= \langle (T^{(r)} - T^{(0)}) (D^{(r)} - D)^T\rangle.
\label{eq:YdefNN}
\end{equation}
For a fixed parametrization, we can use Eq.~(\ref{eq:projection}) and Eq.~(\ref{eq:repavD}) to show that then $Y=X=Y^T$. However it is easy to check by explicit computation that in an NNPDF fit $Y$ is generally considerably smaller than $X$: the fluctuations of the theory replicas are not very well correlated to the fluctuations of the data replicas due to the functional uncertainty. So although many of the eigenvalues of $X(C+S)^{-1}$ will still be zero (because $m<N$), the nonzero eigenvalues will differ from one, and many will be somewhat larger than one due to the functional uncertainty. This means that while Eq.~(\ref{eq:Zabbounds}) still holds, the upper bound on $\overline{Z}_{\alpha\beta}$, Eq.(\ref{eq:zbarabbounds}), does not: the covariance of the nuisance parameters can be larger than the prior when the functional uncertainty is large.
Note that the fact that $X$ is not invertible is not in any sense a technical limitation: the mapping of a global dataset into a set of PDFs cannot in principle be invertible (except possibly in certain special cases, such as data from a single process taken at a single scale \cite{Harland-Lang:2018bxd}), since it is impossible to recover the data solely from the PDFs. This is in part because the PDFs are only functions of $x$, while the data also depend on a scale: when we determine PDFs, all the data are effectively projected onto a common scale. But it is also because PDFs are by definition universal, i.e. process independent, so given a set of PDFs it is impossible in principle to say even which processes were used to determine them.
We can now derive general results for the expectation and covariance of autopredictions from the three matrices $C$, $S$ and $X$, following the procedure set out in Sec.~\ref{subsec:autoprediction}. The shifts in the autopredictions are given by a similar expression to Eq.~(\ref{eq:shiftmult}),
\begin{equation}
\label{eq:shiftNN}
\delta T[f] = -S(C+S)^{-1}(T^{(0)}-D),
\end{equation}
and will reduce the experimental $\chi^2$ as explained already in Sec.~\ref{subsec:puretheory}. The covariance matrix is still given by Eq.~(\ref{eq:covTfitf}):
\begin{equation}
P\equiv{{\rm Cov}}[T(f,\lambda)] = C(C+S)^{-1}X(C+S)^{-1}C + (S - S(C+S)^{-1}S).\label{eq:PNN}
\end{equation}
If the theory uncertainty $S$ is much smaller than the experimental uncertainty $C$, $P$ approaches the result
\begin{equation}
P_{\rm con}= X+S;\label{eq:PconNN}
\end{equation}
the fitting uncertainty and theoretical uncertainty can be combined in quadrature. So when the experimental uncertainties dominate there is almost complete decorrelation of the theoretical uncertainties, and the `conservative' prescription recommended in Ref.~\cite{AbdulKhalek:2019ihb} is a useful approximation.
Returning to the more general correlated result Eq.~(\ref{eq:PNN}), both contributions to $P$ (which we may call the correlated PDF uncertainty and the correlated theory uncertainty) are also positive semi-definite, and combine in quadrature to give the total uncertainty. Moreover the correlated theory uncertainty bounded above by the corresponding uncorrelated theory uncertainty:
\begin{equation}
0\leq S - S(C+S)^{-1}S = C(C+S)^{-1}S \leq S,\label{eq:Sbound}
\end{equation}
It is tempting to also think that the correlated PDF uncertainty will also be bounded above by the uncorrelated PDF uncertainty $X$, because since $C$ is positive definite, and $S$ positive semi-definite, $C\leq C+S$, so $C(C+S)^{-1}\leq 1$, and $C(C+S)^{-1}X(C+S)^{-1}C \leq X$. This argument is wrong, however, and the correlated PDF uncertainty can sometimes exceed the uncorrelated. Writing
\begin{equation}
C(C+S)^{-1}X(C+S)^{-1}C = X - S(C+S)^{-1}X- X(C+S)^{-1}S+S(C+S)^{-1}X(C+S)^{-1}S,
\label{eq:Xalgebra2}
\end{equation}
in some circumstances the sum of the last three terms can be positive. For this reason it seems impossible to prove in general that $P\leq P_{\rm con}$, though in all practical applications we have tested so far this seems to be the case.
For genuine predictions, with theory uncertainties correlated to those in the fit, shifts are given by Eq.~(\ref{eq:shiftpredf}),
\begin{equation}
\label{eq:shifttilNN}
\delta \widetilde{T}[f] = -\widehat{S}(C+S)^{-1}(T^{(0)}-D),
\end{equation}
while Eq.~(\ref{eq:covTtillamf}) is most usefully written in the form
\begin{eqnarray}
\widetilde{P}&\equiv&{{\rm Cov}}[\widetilde{T}(f,\lambda)]\nonumber\\
&=&\widetilde{X} - \widehat{X}(C+S)^{-1}\widehat{S}^T - \widehat{S}(C+S)^{-1}\widehat{X}^T+\widehat{S}(C+S)^{-1}X(C+S)^{-1}\widehat{S}^T\nonumber\\
&&\qquad\qquad\qquad\qquad + (\widetilde{S} - \widehat{S}(C+S)^{-1}\widehat{S}^T) \label{eq:PtilNN},
\end{eqnarray}
where now besides the matrix $X$ Eq.~(\ref{eq:XdefNN}) we must now also evaluate
\begin{eqnarray}
\widetilde{X}&\equiv&{\rm Cov}[\widetilde{T}[f,\lambda]] = \langle (\widetilde{T}^{(r)} -\widetilde{T}^{(0)}) (\widetilde{T}^{(r)} - \widetilde{T}^{(0)})^T\rangle\label{eq:XtildefNN},\\
\widehat{X}&\equiv&{\rm Cov}[\widetilde{T}[f,\lambda],T[f,\lambda]] = \langle (\widetilde{T}^{(r)} -\widetilde{T}^{(0)}) (T^{(r)} - T^{(0)})^T\rangle\label{eq:XhatdefNN}.
\end{eqnarray}
Again the covariance Eq.~(\ref{eq:PtilNN}) separates into the sum in quadrature of a correlated PDF uncertainty (the first line) and a correlated theory uncertainty (the second line), When the cross-covariance $\widehat{S}$ is very small, we obtain the conservative result
\begin{equation}
\widetilde{P}_{\rm con} = \widetilde{X}+\widetilde{S}, \label{eq:PtilconNN}
\end{equation}
proposed in Ref.~\cite{AbdulKhalek:2019ihb}. This will typically be the case for predictions of new processes where the dominant MHOU is in the hard cross-section. However for processes already included in the fit, the situation is more complex, since $\widetilde{S}$ and $\widehat{S}$ may be large even if $S$ is small.
Note that since the inclusion of MHOU in the PDF determination leads to only a small increase in the uncertainties of the PDFs~\cite{AbdulKhalek:2019ihb}, the conservative results Eq.~(\ref{eq:PconNN},\ref{eq:PtilconNN}) give uncertainties very close to the conventional prescription, in which the PDF uncertainty (without MHOU) is combined in quadrature with the MHOU in the prediction. This will be particularly true when the MHOU in the prediction is larger than the PDF uncertainty.
\section{Numerical Results}
\label{sec:numeric}
\begin{table}[b!]
\centering
\renewcommand*{\arraystretch}{1.3}
\begin{tabular}{|c|c|}
\hline
Process Type & Datasets \\
\hline
DIS NC & SLAC, BCDMS, NMC, HERA NC \\
DIS CC & NuTeV, CHORUS, HERA CC \\
DY & CDF, D0, ATLAS, CMS, LHCb (dileptons, $W$ and $Z$ diff xsecs) \\
JET & ATLAS, CMS inclusive jets \\
TOP & ATLAS, CMS total \& diff xsecs \\
\hline
\end{tabular}
\caption{\label{tab:expclassification}
Classification of datasets into process types.
}
\end{table}
In Sec.~\ref{subsec:fittingnnpdfs} we saw that in a realistic global PDF fit that we still use the same analytic expressions Eqs.~(\ref{eq:shiftNN},\ref{eq:PNN},\ref{eq:shifttilNN},\ref{eq:PtilNN}) for the shifts and reduction in uncertainties induced by the correlations between the theoretical uncertainties in fit and prediction as we would use in a fit with a fixed parametrization Sec.~\ref{subsec:pdfexactfit}. This is despite the fact that PDFs are smooth functions, which cannot be determined uniquely from a finite set of discrete data, but necessarily have an additional `functional uncertainty', so the PDF parameters are not fixed uniquely by the fit. All that is necessary is to evaluate the matrices $X$, $\widetilde{X}$ and $\widehat{X}$ Eqs.~(\ref{eq:XdefNN},\ref{eq:XtildefNN},\ref{eq:XhatdefNN}) as ensemble averages over the PDF replicas determined in the fit. In this section we will compute these matrices in a realistic global PDF fit with theory uncertainties, and use them to evaluate autopredictions and genuine predictions including the effect of the correlated theoretical uncertainties.
We will carry out these studies in the context of the NNPDF3.1 NLO global fit with MHOU presented in Ref.~\cite{AbdulKhalek:2019ihb}.
This in turn employed the same experimental data and theory calculations in NNPDF3.1~\cite{Ball:2017nwa} with two minor differences:
the value of the lower kinematic
cut was increased from
$Q_{\rm min}^2=2.69$~GeV$^2$ to
$13.96$~GeV$^2$, and the HERA $F_2^b$, fixed-target Drell-Yan cross-sections, and some LHC inclusive jet data were removed, for technical
reasons. This left a total of $N_{\rm dat}=2819$ data points. The complete list of data included in the fit may be found in Tab.~6.3 of Ref.~\cite{AbdulKhalek:2019ihb}. These data were divided into five classes, depending on the type of process involved,
as summarized in Tab.~\ref{tab:expclassification}. The MHOU covariance matrix $S_{ij}$ was constructed using renormalization and factorization scale variations, by a factor of two either side. The factorization scale variations (estimating the MHOU in the NLO parton evolution) are correlated across all processes, but the renormalization scale variations (estimating the MHOU in the NLO hard cross-sections peculiar to each process) while correlated within data belonging to the same process, are uncorrelated between different processes. These variations were then combined to give $S_{ij}$ using the 9pt scheme, as explained in Ref.~\cite{AbdulKhalek:2019ihb}. The matrices $C_{ij}$ and $S_{ij}$ computed in Ref.~\cite{AbdulKhalek:2019ihb} are reproduced in Fig.~\ref{fig:CnS} as heat maps.
\begin{figure}[h!]
\begin{center}
\makebox{\includegraphics[width=0.48\columnwidth]{plots/C.png}}
\makebox{\includegraphics[width=0.48\columnwidth]{plots/S.png}}
\end{center}
\vspace{-0.55cm}
\caption{The experimental covariance matrix, $C_{ij}$, normalized to the theoretical predictions $T^{(0)}_i$ (left), and the corresponding theory covariance matrix for MHOU, $S_{ij}$ (right). The datasets are arranged in the order given in Fig.~\ref{fig:chi2auto} below: so SLAC data are in the top left corner, and LHC top data in the lower right corner.}
\label{fig:CnS}
\end{figure}
\subsection{Covariance of PDF uncertainties $X$}
\label{subsec:XnY}
We begin by computing the covariance matrix $X_{ij}$, Eq.~(\ref{eq:XdefNN}), shown in Fig.~\ref{fig:X} as a heat map alongside the corresponding correlation matrix.
It can be seen that the off-diagonal elements of $X_{ij}$ are almost as large as the diagonal elements: this is confirmed by examination of the correlation matrix. This is because theoretical predictions are often very strongly correlated, not only for nearby bins within the same experiment, but also for different processes at nearby scales, due primarily to the smoothness of the underlying PDFs, both in $x$ and in $Q^2$, but also due to the highly correlated theoretical uncertainties included in the fit.
\begin{figure}[h!]
\begin{center}
\makebox{\includegraphics[width=0.48\columnwidth]{plots/X_covmat_cutscale.png}}
\makebox{\includegraphics[width=0.48\columnwidth]{plots/Xcorrmat.png}}
\end{center}
\vspace{-0.55cm}
\caption{The covariance matrix of PDF uncertainties, $X_{ij}$, normalized to the theoretical predictions $T^{(0)}_i$ (left), and the corresponding correlation matrix $X_{ij}/\sqrt{X_{ii}X_{jj}}$ (right). The datasets are arranged in the order given in Fig.~\ref{fig:chi2auto} below: so SLAC data are in the top left corner, and LHC top data in the lower right corner.}
\label{fig:X}
\end{figure}
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.99\columnwidth]{plots/CXS.png}}
\end{center}
\vspace{-0.55cm}
\caption{The square root of the diagonal elements of the matrices $X$ (in orange), $C$ (in green) and $S$ (in purple) normalized to the theoretical predictions $T^{(0)}_i$, with those for $C$ and $S$ the same as in Ref.~\cite{AbdulKhalek:2019ihb}. The datasets are arranged in the order given in Fig.~\ref{fig:chi2auto} below.}
\label{fig:CXS}
\end{figure}
We compare the PDF uncertainties to the experimental and theoretical uncertainties by looking at the per-point uncertainty (Fig.~\ref{fig:CXS}). Recall (Eqn.~(\ref{eq:repavD})) that $C+S$ is the covariance of the data replicas to which the PDF replicas are fitted. It can be seen that at NLO the relative size of the experimental uncertainties $C_{ii}$ and the theoretical uncertainties $S_{ii}$ varies considerably between different datasets: for the fixed target DIS data $S_{ii}$ is generally below $C_{ii}$, except at large $x$, whereas for the HERA NC data $S_{ii}$ is much less than $C_{ii}$ at large $x$, but the other way around at small $x$ where the theoretical uncertainty dominates. For CHORUS, the experimental uncertainty also dominates, while for most DY datasets the theoretical uncertainty dominates. In contrast the PDF uncertainties $X_{ii}$ are generally less than either $C_{ii}$ or $S_{ii}$, because the combination of data within given datasets and information from other datasets in the fit conspire to reduce the uncertainty. This is especially evident for DIS CC, DY and JETS. However for some data sets, particularly cross-section ratios with very small theory uncertainty (such as NMC $d/p$, the asymmetries, and the differential top data), $X_{ii}$ lies above $S_{ii}$, though still below $C_{ii}$.
\subsection{Nuisance Parameters}
\label{subsec:nuisance2}
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.7\columnwidth]{plots/lambdas.png}}
\end{center}
\vspace{-0.55cm}
\caption{The $28$ positive eigenvalues $s^\alpha$ of the theory uncertainty matrix $S_{ij}$ (above), shown in descending order, and 28 nuisance parameters $\lambda_\alpha$ corresponding to the $28$ eigenvectors $\beta_\alpha$ (below), as given by Eq.~(\ref{eq:Elambdaf}).The uncertainties in the nuisance parameters are shown in total (square roots of the diagonal entries of Eq.~(\ref{eq:Zbardefab}), and broken down into the contribution from scale uncertainties alone (square roots of the diagonal entries of Eq.~(\ref{eq:Zdefab}) and from PDF uncertainties (square roots of the diagonal entries of the last term in Eq.~(\ref{eq:Zbardefab}). The yellow bands highlight the region between $\pm 1$.}
\label{fig:nuisancediag}
\begin{center}
\makebox{\includegraphics[width=0.6\columnwidth]{plots/lambdas_phys.png}}
\end{center}
\vspace{-0.55cm}
\caption{Nuisance parameters $\lambda$ for directions in the space of scale variations corresponding to up/down changes in factorization scale, and in renormalization scale for the five types of processes in the determination of the $9$-pt theory covariance matrix for MHOU. The uncertainties in the nuisance parameters are shown in total, and broken down into the contribution from scale uncertainties alone and from PDF uncertainties, just as in Fig.~\ref{fig:nuisancediag}. The yellow bands highlight the region between $\pm 1$.}
\label{fig:nuisancephys}
\end{figure}
Having computed $X$, we next calculate the nuisance parameters $\lambda_\alpha$ of the theory covariance matrix $S_{ij}$ for the MHOU in the NNPDF3.1 NLO global fit. The posterior distributions of these nuisance parameters give us information on which directions of MHOU are constrained by the fit. We showed in Ref.~\cite{AbdulKhalek:2019ihb} that when there are five different
processes, there are 28 nonzero eigenvalues. Thus we have 28 nuisance parameters, in one-to-one correspondence with the 28 eigenvectors with nonzero eigenvalues. The eigenvalues are shown in descending order in Fig.~\ref{fig:nuisancediag}, with the nuisance parameters below them. The expectation values of the nuisance parameters are computed using Eq.~(\ref{eq:Elambdaf}), and their uncertainties using Eq.~(\ref{eq:Zbardefab}). The nuisance parameters are normalized so that their prior is a unit gaussian centred on zero, as in Eq.~(\ref{eq:priorf}). It can be seen that after fitting, the uncertainty in the nuisance parameters associated with the largest nine or so eigenvalues has been substantially reduced from one, indicating that exposure to the data has reduced the MHOUs. For those corresponding to the smaller eigenvalues there is very little reduction, showing that the data do not much constrain these directions in the space of MHOU. The central values for the three largest eigenvalue nuisance parameters remain close to zero within uncertainties, showing that the prior choices (mainly overall normalizations) were reasonable, while the next three or four show significant deviations from zero: for these the data seem to carry significant information about the MHOU. For the remaining (smaller) eigenvalues the central values of nuisance parameters are all consistent with zero, and clearly for the very small ones the data have no effect at all, the posterior distributions being the same as the prior. This shows that only the eigenvectors corresponding to the larger eigenvalues are actually relevant for the PDF determination: the remainder correspond to such small changes in theoretical uncertainty that the fit ignores them.
We can understand these features better by separating out the two contributions to the total uncertainty in the nuisance parameters: that due purely to the impact of the fit of a given replica on the MHOU (given by Eq.~(\ref{eq:Zdefab})), and that due to the additional PDF uncertainty when the fits to all the replicas are averaged over PDF replicas, given by the last term in Eq.~(\ref{eq:Zbardefab}), also shown in Fig.~\ref{fig:nuisancediag}. We see that when fitting a single replica, the uncertainties in the nuisance parameters corresponding to the larger eigenvalues are indeed very substantially reduced: the MHOU along the eigenvectors corresponding to these larger eigenvalues is learnt in the fit to the data, just as we saw in the simple models in Sec.~\ref{subsec:puretheory}. Again, very little information is retained about the smaller eigenvalues. The uncertainty contributed by averaging over the PDF replicas is also small for the largest eigenvalue nuisance parameters, but becomes the dominant contribution after the first three. For the smallest it is very small again: for these the data have no effect.
We can learn a little more about which MHOUs are learnt most by choosing different directions for the shift vectors $\beta_\alpha$ in Eq.~(\ref{eq:PTDlf}) than for the eigenvectors of $S_{ij}$. Specifically, we can choose the $\beta_\alpha$ to correspond to factorization scale variations (up or down), or renormalization scale variations (up or down, but now separately for each process). The results are shown in Fig.~\ref{fig:nuisancephys}, where we again show the total uncertainty, scale uncertainty, and PDF uncertainty. The central values fluctuate about zero, but all remain within the band $\pm 1$, showing that the effect of fitting the experimental data on the nuisance parameters is rather mild: this is reassuring, as it confirms the choice of central scales used to make the predictions, and the choice of the range of the scale variations (implicit in the choice of the prior for the nuisance parameters, Eq.~(\ref{eq:priorf})). We also see that the uncertainties in the nuisance parameters corresponding to factorization scale variations (estimating the MHOUs in parton evolution), are reduced the most when fitting to any given replica, as we would expect since MHOUs in parton evolution are common to all data included in the fit. A little is also learnt about the renormalization scales for DIS. However, the PDF uncertainties partially wash out these effects. This suggests that the significant shifts in the nuisance parameters seen in
Fig.~\ref{fig:nuisancediag} are due to global tensions between different processes, rather than problems with the choice of scales for particular processes.
Already we see that information from the data in the fit significantly updates the priors for the nuisance parameter distribution. From this it is likely that there will be an effect at the level of autopredictions, which is the subject of the next section.
\subsection{Autopredictions}
\label{subsec:autopredictions}
We now present results for the `autopredictions'. As we explained already in Sec~\ref{subsec:puretheory}, these are the theoretical predictions we make for all the datasets included in the PDF fit, including theoretical uncertainties, after the fitting of the PDFs (with these same theoretical uncertainties). They can thus be thought of as `postdictions', or predictions for the results of experiments run in exactly the same way, with the same equipment, as the original experiment, but taking account of the original global dataset. They thus form an ideal theoretical laboratory for testing the extent of the decorrelation between the theoretical uncertainties in the PDF fit, and those in the (auto)predictions.
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.95\columnwidth]{plots/shifts.png}}
\end{center}
\vspace{-0.55cm}
\caption{The shifts $\delta T_i$, Eq.~(\ref{eq:shiftNN}) (in blue) compared to the differences between theory and data, $D_i-T^{(0)}_i$ (in green), both normalized to $T^{(0)}_i$.}
\label{fig:shifts}
\begin{center}
\makebox{\includegraphics[width=0.99\columnwidth]{plots/chi2_auto.png}}
\end{center}
\vspace{-0.55cm}
\caption{The experimental $\chi^2$ for each data set, comparing the original result of the NLO fit with no theory uncertainties to the fit with theory uncertainties, and then including the correlated shift in the autopredictions.}
\label{fig:chi2auto}
\end{figure}
We begin by computing the shifts $\delta T_i$, Eq.~(\ref{eq:shiftNN}), in the autopredictions, due to the correlation in theoretical uncertainty: these are shown in Fig.~\ref{fig:shifts}, normalized to the orginal theoretical prediction $T^{(0)}_i$. We also show for comparison the differences $D_i - T^{(0)}_i $. It can be seen from the plot that these shifts are generally much smaller than the difference between data and theory, particularly for DIS NC and DY. However for some datasets (in particular CHORUS and inclusive jets), there seems to be an overall shift in central value of the same order as the difference between experiment and theory. However it is difficult to draw any further conclusions from these observations, since the shifts are very correlated within datasets.
\begin{table}[b!]
\centering
\begin{tabular}{|l||ccccc|c|}
\hline
& \textbf{DIS NC} & \textbf{DIS CC} & \textbf{DY} & \textbf{JETS} & \textbf{TOP} & \textbf{Total} \\
\hline
No th uncs & 1.13 & 0.98 & 1.56 & 0.88 & 1.20 & 1.17 \\
Uncorr th uncs & 1.15 & 1.06 & 1.53 & 0.90 & 1.27 & 1.19 \\
Correlated th uncs & 1.09 & 0.91 & 1.47 & 0.83 & 0.97 & 1.10 \\
\hline
\end{tabular}
\vspace{1cm}
\caption{The experimental $\chi^2$ per data point for each process, comparing the original result of the NLO fit with no theory uncertainties to the fit with theory uncertainties, and then including the shift in the autopredictions.}
\label{tab:chisq}
\end{table}
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.48\columnwidth]{plots/P.png}}
\makebox{\includegraphics[width=0.48\columnwidth]{plots/P_corrmat.png}}
\end{center}
\vspace{-0.55cm}
\caption{The autoprediction covariance matrix $P_{ij}$ Eq.~(\ref{eq:PNN}), normalized to the theoretical predictions $T^{(0)}_i$ (left), and the corresponding corrrelation matrix $P_{ij}/\sqrt{P_{ii}P_{jj}}$ (right).}
\label{fig:P}
\begin{center}
\makebox{\includegraphics[width=0.99\columnwidth]{plots/XPPcon.png}}
\end{center}
\vspace{-0.55cm}
\caption{The percentage uncertainties of the autopredictions $\sqrt{P_{ii}}$ Eq.~(\ref{eq:PNN}) (blue) compared to the PDF uncertainty $\sqrt{X_{ii}}$ (orange), and the conservative result, $\sqrt{P^{\rm con}_{ii}}$ Eq.~(\ref{eq:PconNN}) (cyan), all normalised to the theoretical predictions $T^{(0)}_i$.}
\label{fig:Pdiag}
\end{figure}
To see whether the shifts actually improve the autopredictions, in Fig.~\ref{fig:chi2auto} we thus show the experimental $\chi^2$ for the original data, computed using the autopredictions for the fit with no theory uncertainties, those when the theory uncertainties are included in the fit, and then when the autoprediction includes the shift. Needless to say all three results are generally very close, and including the theory uncertainties in the fit has mixed results, some predictions getting better, but at the expense of others getting worse, since the main effect of the theory uncertainties is to rebalance the datasets in the fit \cite{AbdulKhalek:2019ihb} . Nevertheless when the correlated shifts are included, the fit to most datasets improves, sometimes quite substantially, just as anticipated in the very simple `pure theory' model in Sec.~\ref{subsec:puretheory}, and confirmed for the more realistic models in Sec.~\ref{subsec:autoprediction} and Sec.~\ref{subsec:pdfexactfit}. The numbers broken down by process are shown in Tab.~\ref{tab:chisq}. When including theory uncertainties the total $\chi^2$ increases just a little, from $1.17$ to $1.19$, but when the correlated shift is added to the theoretical predictions, we see a significant improvement to $1.10$. This improvement is seen across all the processes.
Although these autopredictions are in some sense artificial --- in practice experiments are never repeated using exactly the same equipment and settings --- the implications of this exercise for the learning of theoretical uncertainties are nevertheless rather general. This is because in a global fit of the size of that performed here, with 2819 data points from 35 datasets, involving five different processes, removing any one of the smaller datasets has very little impact on the PDFs, and removing any dataset has the effect of increasing PDF uncertainties, while theoretical uncertainties for the remaining data remain unchanged. Consequently if we were to perform the PDF fit without a given dataset, and repeat the analysis, so that the autoprediction becomes a genuine prediction (or more properly `postdiction'), the result for this genuine prediction would be very close to the autoprediction. So we expect the shifts in central values that we see in the autopredictions to be also give improvements in such predictions: the shifts should improve the accuracy of the such predictions.
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.9\columnwidth]{plots/split_th.png}}
\end{center}
\vspace{-0.55cm}
\caption{The contributions to the diagonal elements of the correlated theory uncertainty normalised to diagonal elements of S: $(S-S(C+S)^{-1}S)_{ii}/S_{ii}$ (pink), and $(S-S(C+S)^{-1}S+S(C+S)^{-1}X(C+S)^{-1}S )_{ii}/S_{ii}$ (black).}
\label{fig:Scpts}
\begin{center}
\makebox{\includegraphics[width=0.9\columnwidth]{plots/split_pdf.png}}
\end{center}
\vspace{-0.55cm}
\caption{The contributions to the diagonal elements of the correlated PDF uncertainty normalised to diagonal elements of X: $(X-S(C+S)^{-1}X-X(C+S)^{-1}S)_{ii}/X_{ii}$ (lilac), and $(C(C+S)^{-1}X(C+S)^{-1}C )_{ii}/X_{ii}$ (see Eq.~(\ref{eq:Xalgebra2}) (green).}
\label{fig:Xcpts}
\end{figure}
To see whether we can also increase the precision, we consider the uncertainties in the autopredictions. In Fig.~\ref{fig:P} we show the full covariance matrix $P_{ij}$, Eq.~(\ref{eq:PNN}), again normalized to the theoretical predictions, and corresponding correlation matrix. The matrix $P_{ij}$ is the sum of the PDF uncertainty (derived from data uncertainties and theoretical uncertainties combined) and the theoretical uncertainty in the autoprediction, each reduced in size to account for the learning of the theoretical uncertainties and the correlation between the two sources of theoretical uncertainty. As might be expected there are very large correlations in the autopredictions within datasets, particularly for nearby kinematic points, but there are also smaller correlations, and anticorrelations, between datasets. They are due not only to the correlations of experimental uncertainties within datasets, but also to the use of a common set of smooth underlying PDFs, and the correlations of the theory uncertainties. The correlations within each process are generally larger than those between processes. This suggests that the combined effects of the correlations due to the use of a common factorization scale and the correlations induced by the smoothness of the PDFs is small compared to the correlations from the renormalization scale.
In Fig.~\ref{fig:Pdiag} we show the percentage uncertainties of the autopredictions: $\sqrt{P_{ii}}$, compared to the purely PDF uncertainties $\sqrt{X_{ii}}$ to aid comparison. The correlated autoprediction uncertainties are generally of similar size to the PDF uncertainties; they are
rather larger for some of the DY datasets and JETS, but are actually smaller for most of the DIS NC data (most remarkably for the HERA data at small $x$), and some DY data. So the full autopredictions are not only more accurate: they are also more precise. This increase in precision must be taken with a pinch of salt, since it depends to some extent on the assumptions made in modelling the prior MHOU~\cite{AbdulKhalek:2019ihb}, in particular the choice of independent scales, and the scheme through which they are combined into the theory covariance matrix $S$. In particular the aggressive reduction in the small $x$ uncertainties for the HERA NC autopredictions seen in Fig.~\ref{fig:Pdiag} may be due to the adoption of the same factorization scale for both singlet and nonsinglet, which overconstrains the singlet evolution at small $x$ \cite{Harland-Lang:2018bxd}. We leave the relaxation of these kinds of assumptions for future work.
As expected all of the autopredictions uncertainties are smaller than would be obtained by the standard prescription of adding PDF uncertainties and theory uncertainties in quadrature~\cite{AbdulKhalek:2019ihb}, which ignores both learning and correlation. However the conservative approach overestimates the correlated uncertainty for almost all datapoints, typically by a factor of two or more, particularly those for which the theoretical uncertainty is larger than the PDF uncertainty. The only data for which the conservative prescription works well are ratio data (for example the NMC $d/p$ data), for which theoretical uncertainties are very small.
To understand better how these changes in uncertainty arise, we show in Fig.~\ref{fig:Scpts} the contributions to the diagonal elements of the correlated theory uncertainty (the second term in Eq.~(\ref{eq:PNN}), normalised to the uncorrelated elements $S_{ii}$. The `learning' of the theoretical uncertainty, given by the contribution $-S(C+S)^{-1}S$ (note that $S-S(C+S)^{-1}S$ is equal in the one parameter example described in Sec.~\ref{subsec:autoprediction} to $ZS$) is very significant, reducing the prior uncertainty $S$ almost to zero for NC DIS and DY (where there is considerable data), and by an order of magnitude for DIS CC, JETS and TOP. Probably more flexibility is required in the modelling of the prior for this uncertainty. However the PDF fluctuations $S(C+S)^{-1}X(C+S)^{-1}S$ (note that $S-S(C+S)^{-1}S+S(C+S)^{-1}X(C+S)^{-1}S$ is equal in the one parameter example described in Sec.~\ref{subsec:autoprediction} to $\overline{Z}S$) undo much of this learning, though for all data points this effect is insufficient to take the ratio to $S$ above one.
A similar breakdown of the contributions to the diagonal elements of the correlated PDF uncertainty (the first term in Eq.~(\ref{eq:PNN}), which can be expanded as in Eq.~(\ref{eq:Xalgebra2})), normalised to the uncorrelated elements $X_{ii}$, is shown in Fig.~\ref{fig:Xcpts}. The correlation terms $-S(C+S)^{-1}X-X(C+S)^{-1}S$ are indeed very large, as anticipated in Ref.\cite{Harland-Lang:2018bxd}, in particular for data with relatively large theoretical uncertainty (such as HERA NC at small $x$, or JETS), sufficient there to overwhelm $X$ and give a negative result. However they can also be positive for some data (such as JETS). In any event, the addition of the PDF fluctuation term $S(C+S)^{-1}X(C+S)^{-1}S$ (remember the decomposition Eq.~(\ref{eq:Xalgebra2}) of the total correlated PDF uncertainty) is always sufficient to restore positivity of the correlated PDF uncertainty, and can in some situations (where $S$ is large, in particular for JETS, but also some DIS CC and DY data) take the total correlated PDF uncertainty above the uncorrelated result $X$. Thus the correlations, while generally reducing uncertainties, can in some circumstances increase them, in contrast to learning which always reduces them.
For autopredictions we expect high levels of learning and correlation because we are making predictions for exact repeats of experiments already in the fit. As we noted for the shifts however, removing a smaller dataset from the fit has little effect in the PDFs, so we might expect similar effects for genuine predictions of processes already included in the fit, particular if they are in a similar kinematic region.
\subsection{Predictions for Top}
\label{subsec:topnhiggs}
\begin{figure}[b!]
\begin{center}
\makebox{\includegraphics[width=0.46\columnwidth]{plots/topdilepton.png}}
\makebox{\includegraphics[width=0.46\columnwidth]{plots/toplj.png}}
\end{center}
\vspace{-0.55cm}
\caption{The upper two panels show predictions for $t\bar{t}$ unnormalized rapidity distribution data taken at 13~TeV by CMS, the dilepton rapidity distribution \cite{Sirunyan:2018ucr} (left) and the lepton+jets distribution \cite{Sirunyan:2018wem} (right). The four predictions show: the NLO fit with no MHOUs, PDF error only; the combined PDF and MHOU fit, ignoring correlations (thus $\sqrt{P_{II}^{\rm con}}$); the correlated result including the shift, and uncertainty computed using the simplified result (thus $\sqrt{P_{II}^{\rm sim}}$); the result with the same shift, but with the correlations included exactly (thus $P_{II}$), and the NNLO result with no MHOU. In the middle panels the same is shown, but normalized to the uncorrelated result. In the lower panels we show the fractional reduction in the PDF uncertainty and the theory uncertainty due to the inclusion of the correlations.}
\label{fig:CMSttbar}
\end{figure}
We now consider genuine predictions, for experiments not used in the PDF fit. These are of two kinds: those for datasets obtained through processes already contained in the fit, and those for completely new processes. For the former we consider the ${\rm t}\bar{\rm t}$ production rapidity distributions (dilepton and lepton+jets) measured by CMS at 13 TeV \cite{Sirunyan:2018ucr,Sirunyan:2018wem}. We chose these datasets for two reasons: firstly, the MHOU is large compared to the experimental uncertainty; secondly, the fit in Ref.~\cite{AbdulKhalek:2019ihb} contains the total ${\rm t}\bar{\rm t}$ cross-sections at 7, 8 and 13 TeV, and the normalized rapidity distributions at 8 TeV, all from both ATLAS and CMS. Both these factors mean that we expect to see some of the largest possible effects due to correlations between the theoretical uncertainties of the data in the PDFs and the theoretical uncertainties of the 13 TeV rapidity distributions.
To make genuine predictions, including all correlations, we need to first calculate the covariance matrix for the MHOU of the predictions, $\widetilde{S}_{IJ}$, and its cross-covariance with the MHOU of the theoretical predictions for the data used in the PDF fit,
$\widehat{S}_{Ij}$ (the indices $I,J$ running over the predicted data points, while $i,j$ run over the data included in the PDF fit): these are in fact the same as if we were planning to include data for the new process in a PDF fit, since the complete covariance matrix for the MHOU would be then of the form Eq.~(\ref{eq:covmatglobal}). Similarly we need to compute, using the fitted PDF replicas the covariance matrix $\widetilde{X}_{IJ}$ of the PDF uncertainty for the new predictions, and the cross-covariance $\widehat{X}_{Ij}$ with the PDF uncertainty of the observables used in the fit. All of these matrices are required for an exact calculation of the correlated shifts in the theoretical predictions $\delta\widetilde{T}_I$, Eq.~(\ref{eq:shifttilNN}) and the covariance matrix $\widetilde{P}_{IJ}$ of their combined PDF and correlated theoretical uncertainties Eq.~(\ref{eq:PtilNN}).
Predictions for the CMS 13 TeV ${\rm t}\bar{\rm t}$ production rapidity distributions were computed using the same tool chain as in Ref.~\cite{Ball:2017nwa} for the 8 TeV distributions: NLO theoretical predictions were generated with {\tt Sherpa}~\cite{Gleisberg:2008ta}, in a format compliant with {\tt APPLgrid}~\cite{Carli:2010rw}, using the {\tt MCgrid}
code~\cite{DelDebbio:2013kxa} and the {\tt Rivet}~\cite{Buckley:2010ar} analysis package, with {\tt OpenLoops}~\cite{Cascioli:2011va} for the NLO
matrix elements. Renormalization and factorization scales have been chosen based on the recommendation of
Ref.~\cite{Czakon:2016dgf} as $H_T/4$.
The results of these calculations are shown in Fig.~\ref{fig:CMSttbar}. The prior theoretical uncertainty in the original prediction is around 10\%, considerably greater than the PDF uncertainty, as expected since the hard cross-sections are only computed at NLO. The correlated shift is sizeable, around 5\%, and almost fully correlated across all the rapidity distributions. This is because these are unnormalized distributions, and thus have an overall theoretical normalization uncertainty which is strongly correlated to the measurements of the total ${\rm t}\bar{\rm t}$ cross-sections at 7, 8 and 13 TeV by ATLAS and CMS included in the PDF fit. This is confirmed by breaking down the contributions to the shift Eq.~(\ref{eq:shiftpredf}) from the various data points included in the fit: the results are shown in Tab.~\ref{tab:deltilcons}. All of the shift comes from the six total cross-section measurements, while the 8 TeV normalized rapidity distributions push it down again by around 25\%. The remaining data make almost no contribution. Nevertheless, the shift is still rather less than the theoretical uncertainty in the original prediction, as expected from the shifts in the nuisance parameters for the renormalization scale variation for top processes shown in Fig.~\ref{fig:nuisancephys}.
\begin{table}[t!]
\centering
\scriptsize
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{|llll|llll|l|}
\hline
{\bf ATLAS} &&&& {\bf CMS} &&&& {\bf Other}\\
{\it tot} &&&{\it diff} &{\it tot}&&&{\it diff} & \\
7 TeV & 8 TeV & 13 TeV & 8 TeV &7 TeV & 8 TeV &13 TeV & 8 TeV & \\
\hline
0.37 & 0.11 & 0.24 & -0.21 & 0.26 & 0.21 & 0.07 & -0.04 & -0.01 \\
\hline
\end{tabular}
\caption{The fractional contributions of different data sets included in the fit to the
shifts in the top rapidity distributions, averaged over all 21 data points.}
\label{tab:deltilcons}
\end{table}
\begin{table}[t!]
\centering
\begin{tabular}{|l||cc|c|}
\hline
& \textbf{dilepton} & \textbf{lepton+jet} & \textbf{combined} \\
\hline
No th uncs & 0.55 & 0.37 & 0.46 \\
Uncorr th uncs & 0.57 & 0.42 & 0.49 \\
Correlated th uncs & 0.49 & 0.37 & 0.43 \\
NNLO, no th uncs & 0.45 & 0.39 & 0.42 \\
\hline
\end{tabular}
\caption{The experimental $\chi^2$ per data point for the CMS 13 TeV top dilepton and lepton+jet rapidity distributions, comparing the original result of the NLO fit with no theory uncertainties to the fit with theory uncertainties, and then including the correlated shift in the autopredictions. Also shown for comparison is the result in a NNLO fit with no theory uncertainties, only PDF uncertainties.}
\label{tab:topchisq}
\end{table}
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.4\columnwidth]{plots/Xheattop.png}}
\makebox{\includegraphics[width=0.4\columnwidth]{plots/Pheattop.png}}
\end{center}
\vspace{-0.55cm}
\caption{The left hand plot shows the correlation matrix $\widetilde{X}_{IJ}/\sqrt{\widetilde{X}_{II}\widetilde{X}_{JJ}}$ of the contribution of the PDF uncertainties to the predictions for the 13~TeV rapidity distributions by CMS: the right hand plot shows the correlation matrix $\widetilde{P}_{IJ}/\sqrt{\widetilde{P}_{II}\widetilde{P}_{JJ}}$ of the total uncertainties including the correlated theoretical uncertainties. Note the expanded scales on the heat maps, different in each plot.}
\label{fig:CMSttbarcorrlns}
\end{figure}
We can compare the shift due to correlations to that from going from NLO to NNLO. We thus show in Fig.~\ref{fig:CMSttbar} the results of a complete NNLO calculation (without theory uncertainties): the NNLO corrections also increase predictions by 5-8\%. It is very interesting that the shift, driven by the data for the ${\rm t}\bar{\rm t}$ total cross-sections at 7, 8 and 13 TeV, largely accounts for the NNLO correction: the data know that the NLO theoretical predictions are on the low side, and this information is carried over into the prediction for the 13 TeV rapidity distributions. Indeed, we compare the experimental $\chi^2$ for these data in the various calculations in Tab.~\ref{tab:topchisq}, while the theory uncertainties increase the $\chi^2$ (due presumably to the other top data being deweighted in the fit), the shift gives a significant improvement both for dileptons and lepton+jets, comparable to that obtained with the complete NNLO corrections. So the shifts provide a new method for using experimental data to make improved theoretical predictions through the learning of theoretical uncertainties. The method should be particularly effective when there is substantial data on the process to be predicted already included in the PDF fit, as is the case here.
The middle panels of Fig.~\ref{fig:CMSttbar} shows the same points as the top panel, but as a ratio to the conservative result, making the uncertainties more visible. Comparing the uncertainties, the difference between the uncorrelated and correlated uncertainty is striking; the correlated uncertainties are much smaller than the uncorrelated. The correlated uncertainties are however still larger than the purely PDF uncertainties, but the very large theoretical uncertainty has been substantially reduced. So not only are the correlated predictions more accurate, they are also more precise. Despite this significant shrinking of uncertainties, the correlated predictions are still compatible with the NNLO result, thanks to the shift in central values. While the conservative prescription is also compatible with the NNLO result, it is immediately clear from the plot that it is inferior to the correlated prediction.
The breakdown of the reduction in uncertainties due to the correlations is shown in the lower panels of Fig.~\ref{fig:CMSttbar}. The correlated theory uncertainty is substantially reduced (uniformly across the rapidities), due to the learning of the normalization from the data already included in the fit, while the correlated PDF uncertainty is reduced rather less: as much as a factor of two when the differential cross-section is small, but hardly at all when it is large. So here the dominant effect is clearly the learning of the theoretical uncertainty in the overall normalization.
The theoretical uncertainties in the theoretical predictions are strongly correlated amongst themselves, and between the two rapidity distributions: in Fig.~\ref{fig:CMSttbarcorrlns} we show the correlation matrices for the PDF uncertainties, $\widetilde{X}_{IJ}$, and that of the correlated prediction $\widetilde{P}_{IJ}$. We see that while the predictions are more than 50\% correlated across the range of rapidities by the PDF, when the correlated theoretical uncertainties are also included all points are rather more correlated, to more than 70\%. The pattern of correlations reflects the symmetry in the dilepton distribution and asymmetry in the lepton+jet distribution: the least correlated points are those with the greatest rapidity separation.
\subsection{Predictions for Higgs}
\label{subsec:higgs}
\begin{figure}[t!]
\begin{center}
\makebox{\includegraphics[width=0.6\columnwidth]{plots/higgs.png}}
\end{center}
\vspace{-0.55cm}
\caption{Predictions for the Higgs total cross-section at 14 TeV, made using a variety of approximations. All results use NLO PDFs, while the Higgs total cross-section is computed at NLO (left panel), NNLO (centre panel) and N3LO (right panel). In each panel, we then have, from left to right: MHOU included only in the PDF determination in the 9pt scheme; the same but with the factorization scale uncertainty (MHOU in PDF evolution) included in quadrature; the same but with instead the renormalization scale uncertainty (MHOU in the Higgs cross-section); the total PDF uncertainty and 9pt MHOU combined in quadrature, as recommended in Ref.~\cite{AbdulKhalek:2019ihb}; the total PDF plus 9pt MHOU, but now including also the shift and the correlation between theoretical uncertainties. In the centre panel we also show the NNLO prediction with NNLO PDFs (but no theoretical uncertainties), as a dashed line. }
\label{fig:Higgs}
\end{figure}
As an example of a process not included in the PDF fit we consider the prediction for the total cross-section for Higgs production in gluon fusion at 14 TeV. For the calculation of the cross-section we performed calculations using
{\tt ggHiggs}~\cite{Ball:2013bra,Bonvini:2014jma,Bonvini:2016frm}.
Renormalization and factorization scales are set to $m_h/2$, and the computation is performed using rescaled effective theory.
Our results are shown in in Fig.~\ref{fig:Higgs}. The PDFs are still the NLO PDFs with MHOU from Ref.~\cite{AbdulKhalek:2019ihb}, but the Higgs total cross-sections are computed at NLO, NNLO and N3LO. At NLO the MHOU in the Higgs cross-section, estimated by varying the renormalization scale, completely dominates all other uncertainties (the PDF uncertainty, the theoretical uncertainty in the PDFs, and the scale uncertainty in the PDF evolution, estimated by factorization scale variation), so the effect of correlations in the MHOU is completely negligible. However when the cross-section is computed at NNLO, the renormalization scale uncertainty is rather smaller, and at N3LO it becomes more comparable to the other sources of uncertainty. The shift due to the correlation between MHOUs is always small compared to the overall uncertainty, and becomes smaller still as the perturbative order of the cross-section is increased, as expected. This is because, unlike in the top predictions, data for this process are not included in the fit, so the renormalization scale uncertainty is completely uncorrelated. It is interesting to note however that the small shift due to the correlation in factorization scale takes the NNLO prediction very close to the result computed with NNLO PDFs (though the coincidence is surely accidental).
Unlike for the top predictions, the effect of the correlation on the size of the overall uncertainty is small. Again, this is because the PDF contains far less information about Higgs production than it does about top production. We saw in Fig.~\ref{fig:P} that the information propagated through to the correlated uncertainties was primarily through the renormalization scales, and that correlation due to the factorization scale was a lot weaker. Higgs production, being a new process, is therefore only impacted through the weak factorization scale correlations, and so the reduction in uncertainties is small. Overall, the uncorrelated conservative prescription~\cite{AbdulKhalek:2019ihb} is comparable in this case to the fully correlated one. Note that if we were to perform these studies with NNLO (or indeed N3LO) PDFs which include MHOU, the MHOU in the PDF determination would have presumably been rather smaller than at NLO, and thus the effect of correlations in the MHOU, in particular the shift, but also the effect on the size of the uncertainty would be even smaller than the small corrections we see here.
From these examples of autopredictions, and genuine predictions for top and Higgs, we have seen that the extent of the shift and correlation can vary quite significantly, depending on the type of prediction being made and what information is already contained in the PDFs. The conservative prescription recommended in Ref.~\cite{AbdulKhalek:2019ihb} is certainly not appropriate in general, as the full inclusion of correlations can be quite substantially reduce uncertainties, as we saw both for the autopredictions and top predictions. However, when predicting a new process for which the PDF contains little information about correlated theoretical uncertainties, unsurprisingly the impact of correlations is small and the conservative prescription is quite sufficient.
\section{Summary}
In this paper we studied in detail the correlation between theoretical uncertainties in the calculations used in the determination of PDFs in a global fit, as formulated in Ref.~\cite{AbdulKhalek:2019ihb}, and the theoretical uncertainties in the predictions made using these PDFs. We began by recasting the theoretical uncertainties using nuisance parameters, determined replica by replica, which carry all the information about the effect of the experimental data on the theoretical uncertainties. Using increasingly realistic models of the fitting procedure, we produced analytical formulae for computing fully correlated predictions. In the process we identified three distinct but related effects, each of which has a significant impact on the final theoretical predictions:
\begin{itemize}
\item {\bf Shifts in central values.} These are an effect of Bayesian learning: just as we can use experimental data to determine PDFs, so we can also use it to identify theoretical corrections that improve the agreement between data and theory, while remaining within theoretical uncertainties. The correlations between theoretical uncertainties in the fit and those in predictions then lead to more accurate predictions. This effect was first identified in Sec.~\ref{subsec:puretheory}.
\item {\bf Learning of theoretical uncertainties.} A second consequence of Bayesian learning is a reduction in theoretical uncertainty, due to the information provided by the data in the PDF fit, which through correlation of theoretical uncertainties can lead to a corresponding reduction in the theoretical uncertainties in predictions. This effect was also identified in Sec.~\ref{subsec:puretheory}, and is complementary to the shift in central values.
\item {\bf Correlation in theoretical uncertainties.} The third effect is that the correlation between the theoretical uncertainties in the fit and the theoretical uncertainties in the predictions lead to a change in the PDF uncertainties in the prediction, even in situations where there is no shift, thus avoiding any `double counting' of the theoretical uncertainty. The existence of this effect was first noted in Ref.\cite{Harland-Lang:2018bxd}, and identified as an effect distinct from Bayesian learning in Sec.~\ref{subsec:phenomenology}.
\end{itemize}
While these three effects were first identified in the simple models of Sec.~\ref{sec:generic}, we showed that they are all present in the one parameter fits of Sec.~\ref{sec:oneparameter} and the more realistic fits with multiple parameters in Sec.~\ref{sec:corrlnpdffits}. Using the NNPDF3.1 NLO global fits with MHOU~\cite{AbdulKhalek:2019ihb}, we demonstrated in Sec.~\ref{sec:numeric} that the shifts can give sensible estimates of NNLO corrections, and thereby reduce the $\chi^2$ to the experimental data. We also showed that while the uncertainty in NLO predictions is still a sum in quadrature of the theoretical uncertainty and the PDF uncertainty (which also includes a theory uncertainty), this sum can be significantly reduced, depending on the relative size of the theoretical and experimental uncertainties. Consequently the `conservative' prescription of Ref.~\cite{AbdulKhalek:2019ihb}, where the theory uncertainty in the prediction is combined in quadrature with the PDF uncertainty, is indeed conservative. We expect these conclusions to also hold in global PDF fits with fixed parametrization and tolerance \cite{Bailey:2020ooq,Hou:2019efy}, if these were to include MHOU in the PDF fit.
The degree of correlation is highly dependent on the type of prediction being made. For the autopredictions (predictions for new measurements of the same data points as those included in the fit), Sec.~\ref{subsec:autopredictions}, where there is maximal correspondence between the data in the fit and the predictions being made, the correlation is very high, leading to shifts that improve the quality of the fit to the data, together with a significant reduction in uncertainties, in some cases down to a small fraction of the uncorrelated values. For genuine predictions for new measurements of processes already included in the PDF fit, such as the new measurements of differential top production discussed in Sec.~\ref{subsec:topnhiggs}, we observe that the shift takes the correlated NLO predictions very close to the NNLO prediction, with a significant reduction in uncertainties: the prediction is both more accurate and more precise. For Higgs production, discussed in Sec.~\ref{subsec:higgs}, a process not included in the PDF fit, the level of correlation is much smaller, since the dominant uncertainty (the MHOU in the hard cross-section) is uncorrelated with the MHOU of the fitted processes. In this case the shift is well within uncertainties, and the reduction in uncertainty very modest, so here the use of the conservative prescription Ref.~\cite{AbdulKhalek:2019ihb} is entirely appropriate. We expect this to be true of predictions for any new process.
Thus our main conclusion is that when using PDFs which include MHOUs, taking account of the correlations between the MHOU included in the determination of the PDFs and the MHOU in the prediction can result in a significant improvement in both accuracy and precision. This is especially true in the case where the predicted process is among those included in the fit. However the correlated predictions must be treated with care, since their reliability relies to some extent on the generality of the prior estimation of the MHOU: if unjustified assumptions are made in the choice of prior, the uncertainty estimates in the correlated predictions may be too aggressive. For these reasons the conservative prescription, as an upper bound on the overall uncertainty, may sometimes be preferable, especially for predictions of new processes.
In order to calculate fully correlated predictions and uncertainties, one requires besides the PDF replicas some additional information: the cross-correlations between the theoretical uncertainties in the prediction and those in the theoretical calculations used to determine the PDFs, $\widehat{S}_{Ij}$; and the cross-correlations between the PDF uncertainties in the prediction and all the calculations included in the fit, $\widehat{X}_{Ij}$. In the future, it may be possible to present this information in separate NNPDF deliverables to facilitate the calculation of the correlation effects.
Although we presented our numerical study of correlations in the context of MHOUs, we would expect similar results for other kinds of theoretical uncertainty, such as nuclear uncertainties, higher twist uncertainties, or indeed parametric uncertainties: once the theory covariance matrix has been computed, the linear algebra has no concern for the type of theoretical uncertainty it contains. This suggests a new technique for determining external parameters in PDF fits, such as quark masses or electroweak parameters, taking full account of all correlations with the PDFs and theoretical uncertainties. We hope to explore this possibility in the near future.
\paragraph{Acknowledgments.}
We would like to offer our warmest thanks to Emanuele Nocera for his assistance in the numerical computations. We would also like to thank an anonymous referee for comments which significantly improved the final draft of the paper. R.D.B. and R.L.P. are supported by the UK Science and Technology Facility Council through grants ST/P000630/1 and ST/R504737/1.
|
\section{Introduction}
At present, with the rapid development of computer technology, image recognition technology does not only stay in the laboratory, but has been widely used in various fields in the society. And the urgent problem of sea surface target recognition is one of them, as an important part in the development of ship intelligent navigation technology. In recent years, there are many recognition methods for sea surface targets, most of which are remote sensing images of sea surface obtained by synthetic aperture radar. Guo et al. improved the network structure of CenterNet to enhance the recognition of small targets of ships \cite{1}, Li et al. proposed a new two-branch regression network to improve the localization of ships \cite{2}, and Fu et al. used an anchor-free frame approach with feature balancing and refinement networks to improve the detection of ships in complex scenes \cite{3}. The above-mentioned articles all identify sea surface ship targets in vertical view and need to be photographed by aircraft, satellites and other aerial vehicles, while there are fewer studies related to identification of surrounding sea surface targets by horizontal view on board ships. Not only that, most of the studies have discussed the problem of small target identification, which takes into account the relatively long field of view in wide waters, but there are also situations where ships are sailing in narrow waters, when there are more ships around, and it is very easy for them to block each other in the observation field of view, so the problem of quickly and accurately locating targets with different degrees of overlap for identification becomes one of the key points.
There have been some research results on the problem of recognition of obscured targets. Wan et al. recovered local facial features through generative adversarial networks to achieve face recognition under occlusion \cite{4}, Chowdhury et al. combined progressive expansion algorithms with graph attention networks to improve license plate recognition under street congestion \cite{5}, and Liu et al. used coupled networks to improve recognition accuracy for small targets or occluded pedestrians \cite{6}. They improved the recognition of obscured targets by optimizing the neural network, which more or less increases the computational complexity and reduces the detection speed of target recognition due to the enhancements made to the network structure.
At this stage, target detection methods are mainly divided into two types, the first one is a two-stage detection method represented by the R-CNN series \cite{7,8,9,10}, which first extracts the region of interest using the region proposal network, and then traverses the test images from top to bottom and from left to right for recognition and detection based on the size of the region of interest. The second one is the one-stage detection algorithm represented by SSD \cite{11} and YOLO series \cite{12,13,14,15}, which does not rely on the region proposal network, but directly uses the anchor frame, multiple square regions into which the test image is segmented, and recognizes and detects each region separately. In contrast, the one-stage method is more accurate but generates a certain amount of redundant computations. The two-stage method greatly improves the detection speed at the expense of a portion of accuracy, which is beneficial to meet the needs of real-time monitoring.
The Yolov4 algorithm \cite{15} will be used as the basic model for the experiments, and although it has been proposed for a relatively short period of time, it has already been used in agriculture \cite{16}, construction \cite{17}, medicine \cite{18} and other fields. The experiments in the aforementioned literature were conducted using servers deployed on land for training and testing; the situation is different in that the communication signal at sea is relatively unstable and the size and energy consumption of the recognition system needs to be minimized due to the complex sea conditions, so the detection system needs to be considered for offline use. Therefore, the Yolov4-tiny lightweight algorithm is chosen as the main experimental model in this paper.
Based on the above analysis, an improved mosaic data enhancement method is proposed in this paper, and the main contributions can be summarized as follows.
(1) Data enhancement is performed on the images in the training dataset to enhance the recognition capability of the target recognition algorithm for overlapping targets of ships without changing the network structure.
(2) For the small mobile platform in offline state, a suitable lightweight algorithm is selected to improve the recognition accuracy while maintaining its recognition speed, and to reduce the impact of input images with different resolutions on the recognition performance of the algorithm.
(3) By conducting real-time detection tests in different sea areas of the ship dataset, the results show that the algorithm trained by the improved mosaic method is more accurate in the detection of overlapping targets of ships, which proves the effectiveness and robustness of the improved method.
The remainder of this paper is organized as follows: Section II describes in detail the main structures of the improved moasic method and the Yolov4-tiny algorithm. Section III describes the simulation experiment comparison and the results of the real test experiment. Finally, conclusions are drawn in Section IV.
\section{Methodology}
The Yolo \cite{12} algorithm has received a lot of attention since it was first proposed in 2016, and the subsequently proposed Yolov2 \cite{13} and Yolov3 \cite{14} algorithms have made a series of improvements to the model structure, data preprocessing methods, and loss function calculation methods based on it, which greatly improve the speed and accuracy of target detection. And the Yolov4 algorithm adds many optimization techniques to the calculation method of Yolov3 algorithm, which improves the recognition accuracy with the same recognition speed.
\subsection{Overview of the Yolov4-tiny algorithm}
The Yolov4-tiny network is based on the Yolov4 network simplified by reducing the number of parameters by a factor of 10 at the expense of some recognition accuracy. It is reduced from about 60 million parameters in the Yolov4 network to about 6 million parameters in Yolov4-tiny. Its network structure is shown in Figure 1.
\begin{figure}
\centering\includegraphics[width=3.3in]{figure1}
\caption{Characteristic structure of Yolov4-tiny network}
\end{figure}
Among them, the convolutional block in the backbone network consists of convolutional layers, batch normalization layers \cite{19}, and leaky Relu \cite{20} activation function. And the residual block is the CSPDarknet53-Tiny networks \cite{21}, and its structure is shown in Figure 2.
\begin{figure}
\centering\includegraphics[width=3.3in]{figure2}
\caption{Structure of residual blocks in CSPDarknet53-Tiny networks}
\end{figure}
It consists of a dense layer and a connected layer, which first performs a convolution operation on the output $ x_{U-1} $ of the previous convolution layer to generate a new convolution layer, and divides its output $ x_0 $ = $ [x_{0^{'}},x_{0^{''}}] $ into two parts $ x_{0^{'}} $ and $ x_{0^{''}} $ before and after for the forward propagation. In the network structure of Yolov4-tiny, the second part is taken first for forward propagation, and then the first part is directly concatenated with the second part to the end of the stage, skipping the dense layer. Waiting until the $ x_{0^{''}} $ of the second part finishes the forward calculation after feature concatenation with $ x_0 $ in the transition layer, the output $ x_T $ is obtained, which undergoes max pooling to produce the output $ x_U $ of the residual block. The process of forward propagation and backward propagation of the residual block is shown in Eqs. (1) to (2).
\begin{align} \begin{split}
x_T &= w_T \cdot \left[ x_{0^{''}} , x_1 \right] \\
x_U &= w_U \cdot \left[ x_0 , x_1 \right]
\end{split} \end{align}
\begin{align} \begin{split}
\omega ^{'}_{T} &= f_T \left( \omega _T \cdot \left\{ g_{0}^{''} , g_1 \right\} \right) \\
\omega ^{'}_{U} &= f_U \left( \omega _U \cdot \left\{ g_0 , g_T \right\} \right)
\end{split} \end{align}
\noindent where $\omega_{i}$ and $\omega^{'}_{i}$ are the weights during forward and backward propagation, $ f_{i} $ denotes the function of weight update, and $ g_{i} $ denotes the gradient propagated to the $ i^{th} $ layer. $ i $ equals T or U, representing the output of the connected layer or residual block, respectively. Therefore, using the structure of CSPNet in back propagation, the gradients on different channels can be integrated separately, for example, when the gradient information passes through the dense layer, it will only change the weights on the $ x_{0^{''}} $ channel but will not affect $ x_{0^{'}} $. This reduces the excessive and repetitive gradient information while retaining the feature values at different depths, reduces the memory overhead and improves the network computation speed without affecting the network feature extraction effect.
After the backbone network, the network features are optimized using the feature pyramid structure \cite{22} , and the implementation of a small feature pyramid for the Yolov4-tiny network is shown in Figure 3.
\begin{figure}
\centering\includegraphics[width=2.3in]{figure3}
\caption{Structure of a small feature pyramid}
\end{figure}
The intermediate features are the output of the fourth convolutional layer in the third residual block of the backbone network, which is concatenated with the up-sampled features from the output of the fourth convolutional block in the network. As can be seen in Figure 1, the backbone network contains only the first three CBL layers, so the output generated by the concatenation operation performed in Figure 3 represents the fusion of the shallow network features and the deep network features. Due to the top-down feature extraction by multi-layer convolution, the deep network retains most of the feature values of large targets, and few or even zero of the feature values of small targets are preserved. Therefore, the feature pyramid structure is used to extract the features of several different layers of the network, and after up-sampling and amplification, they are stitched together from the bottom up to achieve the feature fusion of multiple layers, which improves the recognition ability of the network for different size targets at multiple separate rates.
After that, the two outputs of the small feature pyramid are plugged into the head network for calculation, and two sets of images containing different perceptual fields are generated, which are adjusted to the prior frame contained in themselves respectively. The non-maximum suppression method is used to identify and detect targets of different sizes in the original image and improve the overall detection capability of the neural network for multi-scale targets.
\subsection{Data enhancement methods}
The mosaic method is an extension of the CutMix \cite{23} method to generate a new data enhancement algorithm, which differs from the two-image overlay fusion of the CutMix method; instead, it uses four images for cropping and stitching to form a new image. This method can better enrich the background of the target and prevent the degradation of the network generalization ability due to the similar background of the training set.
The output image of Yolov4-tiny algorithm contains two different perceptual fields, while the output image of Yolov4 algorithm has three different perceptual fields, so the recognition ability of Yolov4-tiny algorithm for multi-scale targets will be relatively weak. Therefore, it is especially important to improve the data enhancement method to enhance the generalization ability of the network.
The improved mosaic data enhancement method is shown in Figure 4. The original mosaic method uses the top and middle orange channels in Figure 2 for feature enhancement, while the improved mosaic method adds the bottom gold channel to the original one and uses three channels for feature enhancement. The output of the third golden channel is obtained by increasing the number of images arranged in each row and column compared to the above two. For the sake of illustration, the newly generated nine-in-one image with a specification of $ 3*3 $ will be called $ m9 $, the four-in-one image with a specification of $ 2*2 $ will be called $ m4 $, and the image generated without merging with a specification of $ 1*1 $ will be called $ m1 $. The ratio of $ m1 $, $ m4 $ and $ m9 $ is $ o:p:q $. This combination, to some extent, makes the scale variation characteristics of the training dataset more diverse, thus further attenuating the interference of the background on the target features.
\begin{figure}
\centering\includegraphics[width=2.3in]{figure4}
\caption{The improved mosaic data enhancement method}
\end{figure}
The nine-in-one image $ m9 $ is generated as shown in Fig. 5, which is mainly divided into three stages \textbf{A}, \textbf{B}, and \textbf{C}. In stage \textbf{A}, the width and height ($ W $, $ H $) of the input image are used as the boundary values, and the image is first scaled with the scaling multipliers $ t_{X} $ and $ t_{Y} $ for the X and Y axes, as shown in Eqs. (3)-(4).
\begin{align}
{t_X} &= {f_{rand}}({t_W},{t_W} + \Delta {t_W})\\
{t_{\mathop{\rm Y}\nolimits} } &= {f_{rand}}({t_H},{t_H} + \Delta {t_H})
\end{align}
\noindent where $ t_{W} $ and $ t_{H} $ are the minimum values of the wide and high scaling multipliers, respectively, and $ \Delta t_{W} $ and $ \Delta t_{H} $ are the lengths of the random intervals of the wide and high scaling multipliers, respectively, both of which are hyperparameters. $ f_{rand}() $ denotes the random value function.
\begin{figure}
\centering\includegraphics[width=3.3in]{figure5}
\caption{Flowchart of $ m9 $ image generation}
\end{figure}
The coordinates of the top-left and bottom-right corners of the image after scaling are $ [(a_{i}, b_{i}), (c_{i}, d_{i})] $, which are obtained from Eqs. (5)-(8).
\begin{align}
a_i = \begin{cases}
0, & i = 1,2,3 \\
W \cdot r_1, & i = 4,5,6 \\
W \cdot r_2, & i = 7,8,9
\end{cases} \\
b_i = \begin{cases}
0, & i = 1,4,7 \\
H \cdot r_3, & i = 2,5,8 \\
H \cdot r_4, & i = 3,6,9
\end{cases}
\end{align}
\begin{align}
{c_i} &= {a_i} + W \cdot {t_W}\\
{d_i} &= {b_i} + H \cdot {t_H}
\end{align}
\noindent Among them, $ r_{1} $ and $ r_{2} $ are the ratio of the distance between the upper left coordinate point and the 0 point of the two sets of images except the 0 point of the X-axis to the total width, respectively, and $ r_{3} $ and $ r_{4} $ are the ratio of the distance between the upper left coordinate point and the 0 point of the two sets of images except the 0 point of the Y-axis to the total height, respectively, and both are also hyperparameters. And the short black lines in the gray area are the scale bars, each small segment represents a one-tenth of the width or height. Using the scale bars, we can see that the scaling of the images from the 2nd to the 9th sheet is the same as that of the 1st sheet, and the width and height are both $ t_{W} $ and $ t_{H} $ times of the original.
In stage \textbf{B}, the nine images cropped in the previous stage need to be stitched together and the part of the overflowing bounding box cropped off. It can be seen that there is a certain degree of overlap in the merged images, so each small area needs to be divided. From the schematic diagram of stage \textbf{A}, it can be seen that when the scaled images are placed at the specified position according to the coordinates, there will be an overflowing border. At this time, the overflowing part needs to be cropped, as shown in Eqs. (9)-(10).
\begin{align}
c_{i}^{'} = \begin{cases}
c_i, & \text{if } c_1 < W \\
W, & \text{if } c_1 \ge W
\end{cases} \\
c_{i}^{'} = \begin{cases}
d_i, & \text{if } d_1 < H \\
H, & \text{if } d_1 \ge H
\end{cases}
\end{align}
After edge cropping, the four square regions enclosed by eight two-by-two parallel dotted lines are used as random intervals of the split lines. Where the value of $ r_{i}=(r_{1}, r_{2}, r_{3}, r_{4}) $ is equal to the ratio of the distance between the coordinates of the split line and the point 0 to the length of the boundary, and $ \Delta i $ is the length of the random interval of the split line.
In stage \textbf{C}, a second cut will be made to the internal overlapping part, whose split line coordinates $ s_{i} $ can be obtained from Eq. (11)
\begin{equation}
{s_i} = {f_{rand}}({r_i},{r_i} + \Delta {r_i})\;\;\;i = 1,2,3,4
\end{equation}
After cropping, the stitched $ m9 $ image is obtained. Since the original image is partially missing in the scaling and stitching process, it is possible that the targets at the edges of the original image are partially or completely cut off during the operation. Therefore, it is also necessary to crop or even reject the real frame corresponding to these targets to meet the needs of target detection.
\begin{figure}
\centering\includegraphics[width=3.3in]{figure6}
\caption{Flowchart of $ m4 $ image generation}
\end{figure}
The method of generating the $ m4 $ image is similar to that of generating $ m9 $, as shown in Figure 6. The upper left corner coordinates $ (a^{*}_{i}, b^{*}_{i}) $ and the split line coordinates $ s^{*}_{i} $ are shown by Eqs. (12)-(14).
\begin{align}
a^{*}_{i} = \begin{cases}
0 , & i = 1,2 \\
W \cdot r^{*}_{1} , & i = 3,4
\end{cases} \\
b^{*}_{i} = \begin{cases}
0 , & i = 1,3 \\
H \cdot r^{*}_{2} , & i = 2,4
\end{cases}
\end{align}
\begin{equation}
s^{*}_{i} = f_{rand} \left( r^{*}_{i} , r^{*}_{i} + \Delta r^{*}_{i} \right) , i = 1,2
\end{equation}
\noindent where $ ^{*} $ is used as the distinguishing symbol between the $ m4 $ image and the $ m9 $ image. The scaling multipliers $ t^{*}_{X} $, $ t^{*}_{Y} $, $ \Delta t^{*}_{W} $ and $ \Delta t^{*}_{H} $ and the lower right corner coordinates $ (c^{*}_{i}, d^{*}_{i}) $ are calculated in the same way as for the $ m9 $ image. Since there is no out-of-border overflow, only the internal overlap part needs to be segmented and cropped.
The $ m1 $ image only requires feature enhancement by conventional methods such as flip and color gamut change before entering the network because there is no stitching of multiple images.
\subsection{Loss function}
The Yolov4-tiny algorithm as a whole is roughly the same as its modified previous version, and its loss function contains three components: loss of confidence in the target ($ Loss_{\rm{conf}} $), loss of classification ($ Loss_{\rm{cls}} $), and loss of position ($ Loss_{\rm{loc}} $). As shown in Eqs. (15)-(19).
\begin{gather}
\begin{split}
Loss = \lambda_{\rm{conf}} \cdot Loss_{\rm{conf}} + \lambda _{\rm{cls}} \cdot Loss_{\rm{cls}} \\
+ \lambda _{\rm{loc}} \cdot Loss_{\rm{loc}}
\end{split} \\
\begin{split}
Loss_{\rm{conf}} = \sum\limits_{i = 0}^{K \times K} \sum\limits_{j = 0}^M I_{ij}^{\rm{obj}} Loss_{\rm{BCE}} \left( \hat C_i , C_i \right) \\
- \sum\limits_{i = 0}^{K \times K} \sum\limits_{j = 0}^M I_{ij}^{\rm{noobj}} Loss_{\rm{BCE}} \left( \hat C_i , C_i \right)
\end{split} \\
Loss_{\rm{cls}} = \sum\limits_{i = 0}^{K \times K} I_{ij}^{\rm{obj}} \sum\limits_{k \in \rm{classes}}^{K \times K} Loss_{\rm{BCE}} \left( \hat p_i(k) , p_i(k) \right) \\
Loss_{\rm{BCE}} ( \hat N , N) = \hat N \log (N) + (1 - \hat N) \log (1 - N) \\
Loss_{\rm{loc}} = \sum\limits_{i = 0}^{K \times K} \sum\limits_{j = 0}^M I_{ij}^{\rm{obj}} \cdot loss_{\rm{CIoU}}
\end{gather}
\noindent where $ \lambda_{conf} $, $ \lambda_{cls} $ and $ \lambda_{loc} $ represent the weights of three different categories of loss in the loss function, respectively, the Yolov4-tiny network divides each input image into K × K cells first, and each grid produces M anchor boxes. After each anchor is subjected to the network's antecedent computation, an adjusted bounding box is obtained, and the total number of anchors is K × K × M. $ I^{obj}_{ij} $ and $ I^{noobj}_{ij} $ are used to determine whether the center coordinates of the target are in the $ j^{th} $ anchor box in the $ i^{th} $ grid, if yes the former is equal to 1 and the latter is equal to 0, otherwise the opposite. $ C_{i} $ is the confidence of the true box in the $ i^{th} $ cell and $ {\hat C_i} $ is the confidence of the prediction box in the $ i^{th} $ cell. $ p_{i}(k) $ denotes the conditional probability that the true box in the $ i^{th} $ cell contains the $ k^{th} $ type of target and $ {\hat p_i}(k) $ denotes the conditional probability that the prediction box in the $ i^{th} $ cell contains the $ k^{th} $ type of target.
Unlike the Yolov3 algorithm, the Yolov4-tiny algorithm uses CIoU loss in the calculation of the location loss function \cite{24} instead of the cross-entropy loss used in the confidence loss and classification loss, which enables a more accurate description of the location information. The CIoU loss is calculated as shown in Eqs. (20)-(24).
\begin{gather}
loss_{\rm{CIoU}} = 1 - IoU + R_{\rm{CIoU}} \left( B , B ^{gt} \right) \\
IoU = \frac{ \left| B \cap B^{gt} \right| }{ \left| B \cup B^{gt} \right| } \\
R_{\rm{CIoU}} \left( \rm B ,\rm B^{gt} \right) = \frac{\rho \left( \rm b , \rm b^{gt} \right)}{c^2} + \alpha v \\
\alpha = \frac{v}{\left( 1 - IoU \right) + v} \\
v = \frac{4}{\pi ^2} \left( \arctan \frac{w^{gt}}{h^{gt}} - \arctan \frac{w}{h} \right)^2
\end{gather}
\noindent where $ IoU $ is the intersection ratio, the prediction box $ B = (x, y, w, h) $, and the true box $ B^{gt} = (x^{gt}, y^{gt}, w^{gt}, h^{gt}) $, which consist of x, y coordinates indicating the location of the center point and w, h coordinates indicating the width and height length. $ R_{CIoU}(B, B^{gt}) $ is the penalty term between the prediction box $ B $ and the real box $ B^{gt} $, $ b $ and $ b^{gt} $ represent the centroids of $ B $ and $ B^{gt} $, $ \rho(\bullet) $ denotes the Euclidean distance, and $ c $ is the diagonal distance of the smallest box that can contain both the prediction box and the real box. $ \alpha $ is a positive trade-off parameter and $ v $ is a parameter that measures the consistency of the aspect ratio, which gives a higher priority to factors in the region where the predicted box overlaps with the true box relative to the non-overlapping part in the regression calculation.
\subsection{Network optimization methods}
In order to better combine the characteristics of the ship dataset, the Yolov4-tiny algorithm uses the K-mean clustering algorithm to divide the real frames of different sizes in the training set into m classes before starting the training, and the boxes represented by the center points of the real boxes in each class are used as anchor boxes, so that anchor boxes can be obtained that are more suitable for detecting ship targets. In this paper m=6, these anchor boxes will be divided into 2 groups of 3 boxes each according to the size to detect target objects of different scales.
In the convolution block, the data extracted by the convolution layer, after batch normalization, is activated using the Leaky Relu activation function, which does not set all negative values to 0 as in the Relu function, but sets a non-zero slope, as shown in Eq. (25).
\begin{align}
f_{\rm{Leaky Relu}}(n_i) = \begin{cases}
n_i, & \text{if } n_i \ge 0 \\
\varphi \cdot n_i, & \text{if } n_i < 0, a_i \in (0,1)
\end{cases}
\end{align}
\noindent where $ \varphi $ is the slope when the input value is less than 0 and is the hyperparameter.
In the early stage of training, using a large learning rate can make the network converge quickly, while in the later stage of training, using a small learning rate is more also beneficial for the network to converge to the optimal value. Therefore, the exponential decay strategy of learning rate is utilized for training, and the learning rate $ \gamma $ is calculated as shown in equation (26).
\begin{equation}
\gamma = \varepsilon ^\tau \gamma_0
\end{equation}
\noindent where $ \gamma_0 $ denotes the initial learning rate, $ \varepsilon $ is the decay rate, and $ \tau $ is the number of iterations of the training network.
The Yolov4-tiny model is built based on convolutional neural networks, so its features extracted at different depth levels are not the same. So the network model is first trained in a large dataset, and when it has the ability to extract basic and abstract features, it is then fine-tuned using transfer learning methods to transfer the weights and biases after training to the network in the new training environment. Since the types of targets are different in various training sets, the weights and deviations of the last layer of the network model are structured differently, except that they can be transferred.
\section{Experimental simulation and testing}
\subsection{Planning of data sets}
The widely used datasets, such as VOC dataset \cite{25} and COCO dataset \cite{26}, contain ships that are classified into only 1 category, which exist with random image size and low resolution (no more than 640*360). In contrast, the ship dataset \cite{27} used for training and testing in this paper contains a total of 7000 images with a resolution of 1920*1080, which are intercepted from video clips taken by surveillance cameras that belong to a sea surface surveillance system deployed along the coastline, which includes 156 cameras in 50 different locations. The dataset contains six different types of ships, the number and categories of which are shown in Table 1.
\begin{table}[]
\caption{Ship data set target object category and number}
\centering
\begin{tabular}{cc}
\hline
\textbf{Vessel type} & \textbf{Quantity(pcs)} \\ \hline
Ore carrier & 2084 \\
Bulk cargo carrier & 1811 \\
General cargo ship & 1426 \\
Container ship & 898 \\
Fishing boat & 1539 \\
Passenger ship & 455 \\ \hline
\end{tabular}
\end{table}
The ships in these images have different lighting conditions, observation angles, distance and proximity scales, and overlap levels, making the dataset much more complex and increasing the difficulty of target detection algorithm recognition.
\subsection{Training and test results}
The algorithms in this paper were implemented on the open source neural network framework Pytorch (3.8.5). The computational workstation configuration consists of a GPU (GeForce RTX 3090), CPU (AMD Ryzen 9 3950x 16 Core/ 3.5 GHz/72 M), and 128 G RAM. It consists of a 1080p camera module, power supply module, display output module, and control module, as shown in Figure 7.
\begin{figure}
\centering\includegraphics[width=3.0in]{figure7}
\caption{Small mobile test platform}
\end{figure}
Before the training starts, the dataset is classified. 6000 target images of ships with almost no occlusion are selected as the training set, and 1000 images of ships with different severity of overlap occur as the test set. By optimizing the target recognition algorithm, the target ships with different degrees of overlap and occlusion can be captured more quickly and accurately, reducing and improving the recognition accuracy.The parameters of the mosaic method in the experiment are shown in Table 2, and the parameters in the network optimization method are shown in Table 3.
\begin{table}[]
\caption{Experimental parameters in mosaic data enhancement methods}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{cccccc}
\hline
\textbf{Parameter} & \textbf{Numeric} & \textbf{Parameter} & \textbf{Numeric} & \textbf{Parameter} & \textbf{Numeric} \\ \hline
$ W $ & 608 & $ \Delta r3 $ & 0.05 & $ \Delta r^{*}_{1} $ & 0.2 \\
$ H $ & 608 & $ \Delta r4 $ & 0.05 & $ \Delta r^{*}_{2} $ & 0.2 \\
$ r_{1} $ & 0.3 & $ t_{W} $ & 0.4 & $ t^{*}_{W} $ & 0.4 \\
$ r_{2} $ & 0.65 & $ t_{H} $ & 0.4 & $ t^{*}_{H} $ & 0.4 \\
$ r_{3} $ & 0.3 & $ \Delta t_{W} $ & 0.05 & $ \Delta t^{*}_{W} $ & 0.2 \\
$ r_{4} $ & 0.65 & $ \Delta t_{H} $ & 0.05 & $ \Delta t^{*}_{H} $ & 0.2 \\
$ \Delta r_{1} $ & 0.05 & $ r^{*}_{1} $ & 0.4 & & \\
$ \Delta r_{2} $ & 0.05 & $ r^{*}_{2} $ & 0.4 & & \\ \hline
\end{tabular}
}
\end{table}
\begin{table}
\caption{Parameters in network optimization methods}
\centering
\begin{tabular}{cc}
\hline
\textbf{Parameter} & \textbf{Numeric} \\ \hline
$ \varphi $ & 0.1 \\
$ \gamma _{0} $ & $ 1 \times 10^{-3} $ \\
$ \varepsilon $ & 0.95 \\ \hline
\end{tabular}
\end{table}
The images used for training and validation in the training dataset are randomly segmented in a ratio of 9 to 1. After training starts, the network stops after 100 iterations, and Figure 8 represents the recognition accuracy of the Yolov4-tiny algorithm at different numbers of iterations. The values of the legend in the figure are the values of o:p:q, and the accuracy is represented by mAP.
As can be seen in Figure 8, the values of the yellow curve are slightly higher than the values of the green curve during the last twenty iterations that tend to be smooth, and the recognition results of the Yolov4-tiny algorithm improve slightly after using the mosaic method, while the improved mosaic method indicated by the blue curve greatly improves the recognition accuracy, even higher than that of the Yolov4 using the original mosaic method algorithm using the original mosaic method. So the improved mosaic method not only improves the recognition accuracy of the Yolov4-tiny algorithm, but also greatly improves the detection speed of ship recognition compared to the Yolov4 algorithm. The experimental procedure of data enhancement using m4 and m9 methods is shown in Figure 9.
\begin{figure*}
\centering\includegraphics[width=0.7\linewidth]{figure8}
\caption{mAPs of Yolov4-tiny algorithm with different number of iterations}
\end{figure*}
\begin{figure}
\centering
\subfigure[]{
\begin{minipage}[b]{0.4\linewidth}
\includegraphics[width=1\textwidth]{figure9a}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[b]{0.4\linewidth}
\centering
\includegraphics[width=1\textwidth]{figure9b}
\end{minipage}
}
\centering
\caption{Experimental procedure of $ m4 $ and $ m9 $ methods, (a) $ m4 $, (b) $ m9 $}
\end{figure}
The curve with the highest recognition accuracy in Figure 8 was trained using the modified mosaic method, which uses the $ m1 $, $ m4 $ and $ m9 $ methods as inputs with probability according to the ratio $ o:p:q $ = 2:2:1 . In order to investigate the effect of the different ratios on the recognition accuracy, several sets of experiments were conducted in this paper to compare the results as shown in Table 4.
\begin{table*}
\caption{Effect of mosaic data enhancement methods on target recognition accuracy at different scales}
\centering
\begin{tabular}{ccccc}
\hline
\textbf{Algorithm Type} & \textbf{o:p:q} & \textbf{Highest mAP in 100 iterations} & \textbf{Average mAP in the last 20 iterations} & \textbf{Ratio relative to the original mosaic (1:1:0)} \\ \hline
\multirow{9}{*}{Yolov4-tiny} & 1:0:0 & 62.28\% & 61.54\% & 99.27\% \\
& 1:1:0 & 62.56\% & 61.99\% & 100.00\% \\
& 1:1:1 & 62.39\% & 60.53\% & 97.64\% \\
& 1:2:1 & 61.88\% & 60.10\% & 96.95\% \\
& 2:1:1 & 63.11\% & 61.44\% & 99.11\% \\
& \textbf{2:2:1} & \textbf{65.06\%} & \textbf{64.09\%} & \textbf{103.39\%} \\
& 3:2:1 & 63.78\% & 62.49\% & 100.81\% \\
& 4:2:1 & 63.31\% & 62.86\% & 101.40\% \\
& 4:3:2 & 62.98\% & 62.25\% & 100.42\% \\ \hline
Yolov4 & 1:1:0 & 63.90\% & 62.56\% & 100.92\% \\ \hline
\end{tabular}
\end{table*}
In Table 4, it can be seen that relative to the original Mosaic (1:1:0) method, the recognition algorithm will achieve better results when $ m4 $ is satisfied with double $ m9 $ and $ m1 $ is greater than or equal to $ m4 $, and its recognition accuracy will be higher than the Yolov4 algorithm applying the original Mosaic method when $ o:p:q $ = 2:2:1 and 4:2:1. In addition, not all the improved methods at all ratios are superior, which indicates that the dataset should be focused on $ m1 $ data, so that the network learns the overall features of the target well, and on this basis, $ m4 $ and $ m9 $ are used to enhance the learning of local features, respectively, to improve the generalization ability of the network.
To further test the recognition capability of the network for overlapping targets, a small mobile testbed is used for real-time inspection of the sea surface, located at Gulangyu Island (Xiamen, China). A real-time video clip of two fishing boats overlapping during travel was used as the experiment to calculate the recognition capability of the network at each moment. The video clip had a resolution of 1080P, a duration of 38 seconds, 24 frames per second, and a total of 912 frames, and the test results are shown in Figure 10. The weight file selected in the actual test experiment is the weight corresponding to the highest mAP in 100 iterations.
\begin{figure*}
\centering\includegraphics[width=0.8\linewidth]{figure10}
\caption{Overlap detection experiment comparison chart}
\end{figure*}
In Figure 10, it can be seen that in frame 691, when the two boats are already severely overlapped, the Yolov4-tiny algorithm using the improved mosaic method can still identify both boats at the same time, while the Yolov4-tiny algorithm using the original mosaic method and the one not using the mosaic method can only identify one of the boats. In frame 756 when the two ships are getting rid of the overlap, the Yolov4-tiny algorithm with the improved mosaic method and the original mosaic method can already recognize both ships at the same time, while the Yolov4-tiny algorithm without the mosaic method can still recognize only one of the ships. This further illustrates the effectiveness of the improved mosaic method in improving the detection capability of overlapping targets.
Due to the limited arithmetic power of small mobile devices, the recognition speed is slow at 1080p resolution, so the detection of ship targets in different resolution videos is tested and the recognition speed of the network is shown in Table 5.
\begin{table*}[]
\caption{Comparison of recognition speed at different resolutions}
\centering
\begin{tabular}{ccccccc}
\hline
\multirow{2}{*}{\textbf{Network Type}} & \multicolumn{4}{c}{\textbf{Recognition speed /fps}} & \multirow{2}{*}{\textbf{Number of network weights/pcs}} & \multirow{2}{*}{\textbf{Network weight size /MB}} \\
& \textbf{1080p} & \textbf{720p} & \textbf{480p} & \textbf{360p} & & \\ \hline
Yolov4-tiny & 6.34 & 8.68 & 10.48 & 11.38 & 5,885,666 & 22.45 \\
Yolov4 & 2.03 & 2.27 & 2.41 & 2.46 & 63,964,611 & 244.01 \\ \hline
\end{tabular}
\end{table*}
\begin{table*}
\caption{Comparison of the recognition ability of three mosaic methods for overlapping problems at different resolutions}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{lllllll}
\hline
\textbf{Video Resolution} & \textbf{o:p:q} & \textbf{\begin{tabular}[c]{@{}l@{}}Unable to separate \\ the starting frames\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Unable to separate \\ end frames\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Number of frames that \\ cannot be separated/pcs\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Unable to separate \\ time/sec\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Proportion of time notseparable\\ relative to the original \\ Mosaic method (1:1:0)\end{tabular}} \\ \hline
\multirow{3}{*}{1080p} & 1:0:0 & 654 & 804 & 151 & 6.29 & 196\% \\
& 1:1:0 & 676 & 752 & 77 & 3.21 & 100\% \\
& \textbf{2:2:1} & \textbf{692} & \textbf{755} & \textbf{64} & \textbf{2.67} & \textbf{83\%} \\ \hline
\multirow{3}{*}{720p} & 1:0:0 & 649 & 810 & 162 & 6.75 & 210\% \\
& 1:1:0 & 673 & 752 & 80 & 3.33 & 104\% \\
& \textbf{2:2:1} & \textbf{691} & \textbf{755} & \textbf{65} & \textbf{2.71} & \textbf{84\%} \\ \hline
\multirow{3}{*}{480p} & 1:0:0 & 650 & 912 & 263 & 10.96 & 342\% \\
& 1:1:0 & 658 & 753 & 96 & 4.00 & 125\% \\
& \textbf{2:2:1} & \textbf{692} & \textbf{755} & \textbf{64} & \textbf{2.67} & \textbf{83\%} \\ \hline
\multirow{3}{*}{360p} & 1:0:0 & 649 & 912 & 264 & 11.00 & 343\% \\
& 1:1:0 & 656 & 754 & 99 & 4.13 & 129\% \\
& \textbf{2:2:1} & \textbf{692} & \textbf{756} & \textbf{65} & \textbf{2.71} & \textbf{84\%} \\ \hline
\end{tabular}
}
\end{table*}
\begin{figure*}
\centering\includegraphics[width=0.7\linewidth]{figure11}
\caption{Time growth rate of three mosaic methods that cannot be separated at different resolutions}
\end{figure*}
As can be seen from Table 5, the recognition speed at all four resolutions is low because the Yolov4 network structure is more complex and has more weight parameters, which leads to its higher arithmetic power required for recognition. The recognition speed of Yolov4-tiny algorithm at 360p video resolution is 179.50\% of its recognition speed at 1080p video resolution, and 462.60\% of the recognition speed of Yolov4 algorithm at 360p video resolution.
Since the recognition speed of different algorithms varies at different resolutions, the importance of improving the recognition ability of mosaic at different resolutions is self-evident. Thus, the recognition ability of the overlap problem is judged by comparing the number of frames in the interval from the beginning of the overlap to the end of the overlap in which the Yolov4-tiny algorithm with different mosaic methods correctly identifies and separates the two ships. The specific experimental results are shown in Table 6.
From Table 6, it can be seen that with the decrease of resolution, each algorithm has different degrees of weakening for overlapping target recognition speed, and the degree of weakening can be expressed by the growth rate of the time when it is impossible to separate, as shown in Figure 11. It can be seen that the Yolov4-tiny algorithm using the improved mosaic method achieves a greater advantage at all four different resolutions. Not only the time to fail to separate the overlapping targets is shorter when the ship targets overlap, but also the high recognition rate can be maintained when the resolution degradation occurs.
The experiments prove that the improved mosaic data enhancement method has a certain improvement on the recognition effect of overlapping targets of ships, and the applied Yolov4-tiny algorithm can be deployed on small mobile devices and can be flexibly assembled on various platforms, thus realizing the real-time monitoring of ship targets on the sea surface in the offline state.
\section{Conclusion}
In this paper, an improved mosaic data enhancement method is proposed based on the Yolov4-tiny algorithm as the experimental basis, and the different ratios of mosaic filling are tested for comparison to investigate its effect on the detection accuracy of ship overlapping targets. In the simulation experiments, the ship dataset is analyzed, and the ship overlapping targets in the dataset are used as detection objects. Compared with the original algorithm, the improved method enhances the recognition accuracy of the test dataset by 2.5\% and achieves the same detection effect as the original Yolov4 algorithm, which improves the accuracy and reduces the arithmetic power consumption at the same time. In the real test experiments, the algorithm was deployed on a small mobile testbed for real tests, and the improved method achieved a 17\% reduction in target loss time and a 27.01\% improvement in recognition stability at different video resolutions compared to the original algorithm. Therefore, the target recognition algorithm can improve its ability to recognize overlapping targets of ships after being trained by the improved mosaic data enhancement method.
\section*{Acknowledgment}
This work was supported in part by the National Natural Science Foundation of China under Grant No.51879118, in part by the Natural Science Foundation of Fujian Province No.2020J01688, in part by the Transportation Industry High-Level Technical Talent Training Project No.2019-014, in part by t the Science and Technology Support Project of Fujian Province No. B19101, in part by the Young Talent of Jimei University No. ZR2019006.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Anecdotally speaking, fluent bilingual speakers rarely face trouble translating a task learned in one language to another. For example, a bilingual speaker who is taught a math problem in English will trivially generalize to other known languages. Furthermore there is a large collection of evidence in linguistics arguing that although separate lexicons exist in multilingual speakers the core representations of concepts and theories are shared in memory \citep{altarriba1992representation, mitchel2005bilinguals, bentin1985event}. The fundamental question we're interested in answering is on the learnability of these shared representations within a statistical framework.
We approached this problem from a linguistics perspective. Languages have vastly varying syntactic features and rules. \emph{Linguistic Relativity} studies the impact of these syntactic variations on the formations of concepts and theories \citep{au1983chinese}. Within this framework of study, the two schools of thoughts are linguistic determinism and weak linguistic influence. \emph{Linguistic determinism} argues that language entirely forms the range of cognitive processes, including the creation of various concepts, but is generally agreed to be false \citep{hoijer1954sapir, au1983chinese}. Although there exists some weak linguistic influence, it is by no means fundamental \citep{ahearn2016living}. The superfluous nature of syntactic variations across languages brings forward the argument of \emph{principles and parameters} (PnP) which hypothesizes the existence of a small distributed parameter representation that captures the syntactic variance between languages denoted by parameters (e.g. head-first or head-final syntax), as well as common principles shared across all languages \citep{culicover1997principles}. \emph{Universal Grammar} (UG) is the study of principles and the parameters that are universal across languages \citep{montague1970universal}.
The ability to learn these universalities would allow us to learn representations of language that are fundamentally agnostic of the specific language itself. Doing so would allow us to learn a task in one language and reap the benefits of all other languages without needing multilingual datasets. Our attempt to learn these representations begins by taking inspiration from linguistics and formalizing UG as an optimization problem.
We train downstream models using language agnostic universal representations on a set of tasks and show the ability for the downstream models to generalize to languages that we did not train on.
\section{Related Work}
Our work attempts to unite universal (task agnostic) representations with multilingual (language agnostic) representations \citep{elmo,mccann2017learned}. The recent trend in universal representations has been moving away from context-less unsupervised word embeddings to context-rich representations. Deep contextualized word representations (ELMo) trains an unsupervised language model on a large corpus of data and applies it to a large set of auxiliary tasks \citep{elmo}. These unsupervised representations boosted the performance of models on a wide array of tasks. Along the same lines \cite{mccann2017learned} showed the power of using latent representations of translation models as features across other non-translation tasks. In general, initializing models with pre-trained language models shows promise against the standard initialization with word embeddings. Even further, \cite{radford2017learning} show that an unsupervised language model trained on a large corpus will contain a neuron that strongly correlates with sentiment without ever training on a sentiment task implying that unsupervised language models maybe picking up informative and structured signals.
In the field of multilingual representations, a fair bit of work has been done on multilingual word embeddings. \cite{ammar2016massively} explored the possibility of training massive amounts of word embeddings utilizing either parallel data or bilingual dictionaries via the SkipGram paradigm. Later on an unsupervised approach to multilingual word representations was proposed by \cite{chen2018unsupervised} which utilized an adversarial training regimen to place word embeddings into a shared latent space. Although word embeddings show great utility, they fall behind methods which exploit sentence structure as well as words. Less work has been done on multilingual sentence representations. Most notably both \cite{schwenk2017learning} and \cite{artetxe2017unsupervised} propose a way to learn multilingual sentence representation through a translation task.
We propose learning language agnostic representations through constrained language modeling to capture the power of both multilingual and universal representations. By decoupling language from our representations we can train downstream models on monolingual data and automatically apply the models to other languages.
\section{Universal Grammar as an Optimization Problem}
Statistical language models approximate the probability distribution of a series of words by predicting the next word given a sequence of previous words.
\begin{equation*}
p(w_0,...,w_n) = \prod_{i=1}^n p(w_i \mid w_0,...,w_{i-1})
\end{equation*}
where $w_i$ are indices representing words in an arbitrary vocabulary.
Learning grammar is equivalent to language modeling, as the support of $p$ will represent the set of all grammatically correct sentences. Furthermore, let $p_j(\cdot)$ represent the language model for the j\textsuperscript{th} language and $w^j$ represents a word from the jth language. Let $k_j$ represent a distributed representation of a specific language along the lines of the PnP argument \citep{culicover1997principles}. UG, through the lens of statistical language modeling, hypothesizes the existence of a factorization of $p_j(\cdot)$ containing a language agnostic segment.
The factorization used throughout this paper is the following:
\begin{align}
b &= u \circ e_j(w^j_0,...,w^j_i) \label{ug:2}\\
p_j(w_i \mid w_0,...,w_{i-1}) &= e_j^{-1}(h(b,k_j)) \label{ug:3}\\
&s.t. \quad d(\mathbf{p}(b\mid j_\alpha) \mid\mid \mathbf{p}(b\mid j_\beta)) \leq \epsilon
\end{align}
\begin{figure}[h]
\centering
\includegraphics[height=8.8cm]{ppt/UGWGAN_1.pdf}
\caption{Architecture of UG-WGAN. The amount of languages can be trivially increased by increasing the number of language agnostic segments $k_j$ and $e_j$.}
\end{figure}
The distribution matching constraint $d$, insures that the representations across languages are common as hypothesized by the UG argument.
Function $e_j: \mathbb{N}^i \rightarrow \mathbb{R}^{i \times d}$ is a language specific function which takes an ordered set of integers representing tokens and outputs a vector of size $d$ per token. Function $u: \mathbb{R}^{i \times d} \rightarrow \mathbb{R}^{i \times d}$ takes the language specific representation and attempts to embed into a language agnostic representation. Function $h: (\mathbb{R}^{i \times d}, \mathbb{R}^{f}) \rightarrow \mathbb{R}^{i \times d}$ takes the universal representation as well as a distributed representation of the language of size $f$ and returns a language specific decoded representation. $e^{-1}$ maps our decoded representation back to the token space.
For the purposes of distribution matching we utilize the GAN framework. Following recent successes we use Wasserstein-1 as our distance function $d$ \citep{arjovsky2017wasserstein}.
Given two languages $j_\alpha$ and $j_\beta$ the distribution of the universal representations should be within $\epsilon$ with respect to the $W_1$ of each other. Using the Kantarovich-Rubenstein duality we define
\begin{equation}
d(\mathbf{p}(b\mid j_\alpha) \mid\mid \mathbf{p}(b\mid j_\beta)) =
\sup_{||f_{\alpha,\beta}||_L \leq 1} \mathbb{E}_{x\sim\mathbf{p}(b\mid j_\alpha)}\left[f_{\alpha,\beta}(x)\right] - \mathbb{E}_{x\sim\mathbf{p}(b\mid j_\beta)}\left[f_{\alpha,\beta}(x)\right]
\end{equation}
where $L$ is the Lipschitz constant of $f$. Throughout this paper we satisfy the Lipschitz constraint by clamping the parameters to a compact space, as done in the original WGAN paper \citep{arjovsky2017wasserstein}. Therefore the complete loss function for $m$ languages each containing $N$ documents becomes:
\begin{equation}
\max_{\theta} \sum_{\alpha=0}^m \sum_{i=0}^N \log p_{j_\alpha}(w_{i,0}^\alpha,...,w_{i,n}^\alpha; \theta)\nonumber - \frac{\lambda}{m^2}\sum_{\alpha=0}^m \sum_{\beta=0}^m d(\mathbf{p}(b\mid j_\alpha) \mid\mid \mathbf{p}(b\mid j_\beta))
\end{equation}
$\lambda$ is a scaling factor for the distribution constraint loss.
\section{UG-WGAN}\label{UG-WGAN}
Our specific implementation of this optimization problem we denote as UG-WGAN.
Each function described in the previous section we implement using neural networks. For $e_j$ in equation~\ref{ug:2} we use a language specific embedding table followed by a LSTM \citep{hochreiter1997long}. Function $u$ in equation~\ref{ug:2} is simply stacked LSTM's. Function $h$ in equation~\ref{ug:3} takes input from $u$ as well as a PnP representation of the language via an embedding table. Calculating the real inverse of $e^{-1}$ is non trivial therefore we use another language specific LSTM whose outputs we multiply by the transpose of the embedding table of $e$ to obtain token probabilities. For regularization we utilized dropout and locked dropout where appropriate \citep{gal2016theoretically}.
The critic, adopting the terminology from \cite{arjovsky2017wasserstein}, takes the input from $u$, feeds it through a stacked LSTM, aggregates the hidden states using linear sequence attention as described in DrQA \citep{chen2017reading}. Once we have the aggregated state we map to a $m \times m$ matrix from where we can compute the total Wasserstein loss. A Batch Normalization layer is appended to the end of the critic \citep{ioffe2015batch}. The $\alpha, \beta$th index in the matrix correspond to the function output of $f$ in calculating $W_1(\mathbf{p}(b\mid j_\alpha) \mid\mid \mathbf{p}(b\mid j_\beta))$.
We trained UG-WGAN with a variety of languages depending on the downstream task. For each language we utilized the respective Wikipedia dump. From the wikipedia dump we extract all pages using the wiki2text\footnote{\url{https://github.com/rspeer/wiki2text}} utility and build language specific vocabularies consisting of 16k BPE tokens \citep{sennrich2015neural}. During each batch we sample documents from our set of languages which are approximately the same length. We train our language model via BPTT where the truncation length progressively grows from 15 to 50 throughout training. The critic is updated $10$ times for every update of the language model. We trained each language model for 14 days on a NVidia Titan X. For each language model we would do a sweep over $\lambda$, but in general we have found that $\lambda=0.1$ works sufficiently well for minimizing both perplexity and Wasserstein distance.
\subsection{Exploration}
A couple of interesting questions arise from the described training procedure. Is the distribution matching constraint necessary or will simple joint language model training exhibit the properties we're interested in? Can this optimization process fundamentally learn individual languages grammar while being constrained by a universal channel? What commonalities between languages can we learn and are they informative enough to be exploited?
We can test out the usefulness of the distribution matching constraint by running an ablation study on the $\lambda$ hyper-parameter. We trained UG-WGAN on English, Spanish and Arabic wikidumps following the procedure described above. We kept all the hyper-parameters consistent apart for augmenting $\lambda$ from 0 to 10. The results are shown in Figure~\ref{fig:ablation}. Without any weight on the distribution matching term the critic trivially learns to separate the various languages and no further training reduces the wasserstein distance. The joint language model internally learns individual language models who are partitioned in the latent space. We can see this by running a t-SNE plot on the universal ($u(\cdot)$) representation of our model and seeing existence of clusters of the same language as we did in Figure~\ref{fig:tsne} \citep{maaten2008visualizing}. An universal model satisfying the distribution matching constrain would mix all languages uniformly within it's latent space.
\begin{figure}
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[height=4.4cm]{vis/wass_ablation.pdf}
\caption{Wasserstein Estimate}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[height=4.4cm]{vis/perp_ablation.pdf}
\caption{Language Model Perplexity}
\end{subfigure}
\caption{Ablation study of $\lambda$. Both Wasserstein and Perplexity estimates were done on a held out test set of documents.}
\label{fig:ablation}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[height=4.4cm]{vis/tsne_L_0_0.pdf}
\caption{$\lambda=0.0$}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[height=4.4cm]{vis/tsne_L_0_1.pdf}
\caption{$\lambda=0.1$}
\end{subfigure}
\caption{T-SNE Visualization of $u(\cdot)$. Same colored dots represent the same language.}
\label{fig:tsne}
\end{figure}
To test the universality of UG-WGAN representations we will apply them to a set of orthogonal NLP tasks. We will leave the discussion on the learnability of grammar to the Discussion section of this paper.
\section{Experiments}
By introducing a universal channel in our language model we reduced a representations dependence on a single language. Therefore we can utilize an arbitrary set of languages in training an auxiliary task over UG encodings. For example we can train a downstream model only on one languages data and transfer the model trivially to any other language that UG-WGAN was trained on.
\subsection{Sentiment Analysis}
To test this hypothesis we first trained UG-WGAN in English, Chinese and German following the procedure described in Section~\ref{UG-WGAN}. The embedding size of the table was $300$ and the internal LSTM hidden size was 512. A dropout rate of $0.1$ was used and trained with the ADAM optimization method \citep{kingma2014adam}. Since we are interested in the zero-shot capabilities of our representation, we trained our sentiment analysis model only on the english IMDB Large Movie Review dataset and tested it on the chinese ChnSentiCorp dataset and german SB-10K \citep{Sentiment, tan2008empirical}. We binarize the label's for all the datasets.
Our sentiment analysis model ran a bi-directional LSTM on top of fixed UG representations from where we took the last hidden state and computed a logistic regression. This was trained using standard SGD with momentum.
\begin{table}[h]
\centering
\begin{tabular}{@{}lrrr@{}}
\toprule
Method & IMDB & ChnSentiCorp & SB-10K \\ \midrule
NMT + Logistic \citep{schwenk2017learning} & 12.44\% & 20.12\% & 22.92\% \\
FullUnlabeledBow \citep{Sentiment} & 11.11\% & * & * \\
NB-SVM TRIGRAM \citep{mesnil2014ensemble} & 8.54\% & 18.20\% & 19.40\% \\
\bf{UG-WGAN} $\lambda=0.1$ + Logistic (Ours) & 8.01\% & 15.40\% & 17.32\% \\
UG-WGAN $\lambda=0.0$ + Logistic (Ours) & 7.80\% & 53.00\% & 49.38\% \\
Sentiment Neuron \cite{radford2017learning} & 7.70\% & * & * \\
SA-LSTM \citep{dai2015semi} & 7.24\% & * & * \\ \bottomrule
\end{tabular}
\caption{Zero-shot capability of UG and OpenNMT representation from English training. For all other methods we trained on the available training data. Table shows error of sentiment model.}
\end{table}
We also compare against encodings learned as a by-product of multi-encoder and decoder neural machine translation as a baseline \citep{opennmt}. We see that UG representations are useful in situations when there is a lack of data in an specific language. The language agnostics properties of UG embeddings allows us to do successful zero-shot learning without needing any parallel corpus, furthermore the ability to generalize from language modeling to sentiment attests for the universal properties of these representations. Although we aren't able to improve over the state of the art in a single language we are able to learn a model that does surprisingly well on a set of languages without multilingual data.
\subsection{NLI}
A natural language inference task consists of two sentences; a premise and a hypothesis which are either contradictions, entailments or neutral. Learning a NLI task takes a certain nuanced understanding of language. Therefore it is of interest whether or not UG-WGAN captures the necessary linguistic features. For this task we use the Stanford NLI (sNLI) dataset as our training data in english \citep{bowman2015large}. To test the zero-shot learning capabilities we created a russian sNLI test set by random sampling 400 sNLI test samples and having a native russian speaker translate both premise and hypothesis to russian. The label was kept the same.
For this experiment we trained UG-WGAN on the English and Russian language following the procedure described in Section~\ref{UG-WGAN}. We kept the hyper-parameters equivalent to the Sentiment Analysis experiment. All of the NLI model tested were run over the fixed UG embeddings. We trained two different models from literature, Densely-Connected Recurrent and Co-Attentive Network by \cite{kim2018semantic} and Multiway Attention Network by \cite{tan2018multiway}. Please refer to this papers for further implementation details.
\begin{table}[h]
\small
\centering
\begin{tabular}{@{}p{10.5cm}rr@{}}
\toprule
Method & sNLI(en) & sNLI (ru) \\ \midrule
Densely-Connected Recurrent and Co-Attentive Network Ensemble \citep{kim2018semantic} & \bf 9.90\% & * \\
\bf UG-WGAN ($\lambda=0.1$) + Densely-Connected Recurrent and Co-Attentive Network \citep{kim2018semantic} & 12.25\% & \bf 21.00\% \\
UG-WGAN ($\lambda=0.1$) + Multiway Attention Network \citep{tan2018multiway} & 21.50\% & 34.25\% \\
UG-WGAN ($\lambda=0.0$) + Multiway Attention Network \citep{tan2018multiway} & 13.50\% & 65.25\% \\
UG-WGAN ($\lambda=0.0$) + Densely-Connected Recurrent and Co-Attentive Network \citep{kim2018semantic} & 11.50\% & 68.25\% \\
Unlexicalized features + Unigram + Bigram features \citep{bowman2015large} & 21.80\% & 55.00\% \\ \bottomrule
\end{tabular}
\caption{Error in terms of accuracy for the following methods. For \textit{Unlexicalized features + Unigram + Bigram features} we trained on 200 out of the 400 Russian samples and tested on the other 200 as a baseline.}
\end{table}
UG representations contain enough information to non-trivially generalize the NLI task to unseen languages. That being said, we do see a relatively large drop in performance moving across languages which hints that either our calculation of the Wasserstein distance may not be sufficiently accurate or the universal representations are biased toward specific languages or tasks.
One hypothesis might be that as we increase $\lambda$ the cross lingual generalization gap (difference in test error on a task across languages) will vanish. To test this hypothesis we conducted the same experiment where UG-WGAN was trained with a $\lambda$ ranging from $0$ to $10$. From each of the experiments we picked the model epoch which showed the best perplexity. The NLI specific model was the Densely-Connected Recurrent and Co-Attentive Network.
\begin{figure}[h]
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[height=4.4cm]{vis/cross_lingual_error.pdf}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[height=4.4cm]{vis/cross_lingual.pdf}
\end{subfigure}
\caption{Cross-Lingual Generalization gap and performance}
\label{fig:crosslingualgap}
\end{figure}
Increasing $\lambda$ doesn't seem to have a significant impact on the generalization gap but has a large impact on test error. Our hypothesis is that a large $\lambda$ doesn't provide the model with enough freedom to learn useful representations since the optimizations focus would largely be on minimizing the Wasserstein distance, while a small $\lambda$ permits this freedom. One reason we might be seeing this generalization gap might be due to the way we satisfy the Lipschitz constraint. It's been shown that there are better constraints than clipping parameters to a compact space such as a gradient penalty \citep{gulrajani2017improved}. This is a future direction that can be explored.
\section{Discussion}
Universal Grammar also comments on the learnability of grammar, stating that statistical information alone is not enough to learn grammar and some form of native language faculty must exist, sometimes titled the poverty of stimulus (POS) argument \citep{chomsky2010poverty, lewis2001learnability}. From a machine learning perspective, we're interested in extracting informative features and not necessarily a completely grammatical language model. That being said it is of interest to what extent language models capture grammar and furthermore the extent to which models trained toward the universal grammar objective learn grammar.
\begin{figure}[h]
\centering
\includegraphics[height=4.4cm]{vis/lang_ablation.pdf}
\caption{Perplexity calculations on a held out test set for UG-WGAN trained on a varying number of languages.}
\label{fig:varylanguages}
\end{figure}
One way to measure universality is by studying perplexity of our multi-lingual language model as we increase the number of languages. To do so we trained 6 UG-WGAN models on the following languages: English, Russian, Arabic, Chinese, German, Spanish, French. We maintain the same procedure as described above. The hidden size of the language model was increased to 1024 with 16K BPE tokens being used. The first model was trained on English Russian, second was trained on English Russian Arabic and so on. For arabic we still trained from left to right even though naturally the language is read from right to left. We report the results in Figure~\ref{fig:varylanguages}. As the number of languages increases the gap between a UG-WGAN without any distribution matching and one with diminishes. This implies that the efficiency and representative power of UG-WGAN grows as we increase the number of languages it has to model.
We see from Figure~\ref{fig:ablation} that perplexity worsens proportional to $\lambda$. We explore the differences by sampling sentences from an unconstrained language model and $\lambda=0.1$ language model trained towards English and Spanish in Table~\ref{table:samples}. In general there is a very small difference between a language model trained with a Universal Grammar objective and one without. The Universal Grammar model tends to make more gender mistakes and mistakes due to Plural-Singular Form in Spanish. In English we saw virtually no fundamental differences between the language models. This seems to hint the existence of an universal set of representations for languages, as hypothesized by Universal Grammar. And although completely learning grammar from statistical signals might be improbable, we can still extract useful information.
\begin{table}[t]
\small
\begin{tabular}{@{}lp{6.6cm}p{6.6cm}@{}}
\toprule
& $\lambda=0.0$ & $\lambda=0.1$ \\ \midrule
en & earth's oxide is a monopoly that occurs towing of the carbon-booed trunks, resulting in a beam containing of oxygen through the soil, salt, warm waters, and the different proteins. & the practice of epimatic behaviours may be required in many ways of all non-traditional entities. \\\midrule
& the groove and the products are numeric because they are called "pressibility" (ms) nutrients containing specific different principles that are available from the root of their family, including a wide variety of molecular and biochemical elements. & a state line is a self-government environment for statistical cooperation, which is affected by the monks of canada, the east midland of the united kingdom. \\\midrule
& however, compared to the listing of special definitions, it has evolved to be congruent with structural introductions, allowing to form the chemical form. & the vernacular concept of physical law is not as an objection (the whis) but as a universal school. \\ \midrule \midrule
es & la revista más reciente varió el manuscrito originalmente por primera vez en la revista publicada en 1994. & en el municipio real se localiza al mar del norte y su entorno en escajáríos alto, con mayor variedad de cíclica población en forma de cerca de 1070 km2. \\\midrule
& de hecho la primera canción de "blebe cantas", pahka zanjiwtryinvined cot de entre clases de fanáticas, apareció en el ornitólogo sello triusion, jr., en la famosa publicación playboy de john allen. & fue el último habitantes de suecia, con tres hijos, atasaurus y aminkinano (nuestra). \\\midrule
& The names of large predators in charlesosaurus include bird turtles hibernated by aerial fighters and ignored fish. & jaime en veracruz fue llamado papa del conde mayor de valdechio, hijo de diego de zúñiga. \\ \bottomrule
\end{tabular}
\caption{Example of samples from UG-WGAN with $\lambda=0.0$ and $\lambda=0.1$}
\label{table:samples}
\end{table}
\section{Conclusion}
In this paper we introduced an unsupervised approach toward learning language agnostic universal representations by formalizing Universal Grammar as an optimization problem. We showed that we can use these representations to learn tasks in one language and automatically transfer them to others with no additional training. Furthermore we studied the importance of the Wasserstein constraint through the $\lambda$ hyper-parameter. And lastly we explored the difference between a standard multi-lingual language model and UG-WGAN by studying the generated outputs of the respective language models as well as the perplexity gap growth with respect to the number of languages.
|
\section{Introduction}
Given a dataset $\mathcal{S}\subset\mathbb{R}^d$ containing $n$ vectors (also called items) and a query $q\in\mathbb{R}^d$, maximum inner product search (MIPS) finds the vector in $\mathcal{S}$ that has the maximum inner product with $q$,
\begin{equation}\label{equ:mips}
p=\arg \max_{x\in\mathcal{S}}{q^{\top}x}.
\end{equation}
MIPS may require items with the top $k$ inner products and it usually suffices to return approximate results (i.e., items with inner products close to the maximum). MIPS has many important applications including recommendation based on user and item embeddings obtained from matrix factorization~\citep{koren:mf}, multi-class classification with linear classifier~\citep{dean:obj}, filtering in computer vision~\citep{felzens:obj}, etc.
MIPS is a challenging problem as modern datasets often have high dimensionality and large cardinality. Initially, tree-based methods~\citep{ram:cone, koenigstein:retrival} were proposed for MIPS, which use the idea of branch and bound similar to k-d tree~\citep{friedman:tree}. However, these methods suffer from the curse of dimensionality and their performance can be even worse than linear scan when feature dimension is as low as 20~\citep{weber:performance}. Shrivastava and Li proposed {\sc l{\footnotesize 2}-alsh}~\citeyearpar{shrivastava:alsh}, which attains the first provable sub-linear query time complexity for approximate MIPS that is independent of dimensionality. {\sc l{\footnotesize 2}-alsh} applies an asymmetric transformation~\footnote{Asymmetric transformation means that the transformations for the queries and the items are different, while symmetric transformation means the same transformation is applied to the items and queries.} to transform MIPS into $L_2$ similarity search, which can be solved with well-known LSH functions. Following the idea of {\sc l{\footnotesize 2}-alsh}, Shrivastava and Li formulated another pair of asymmetric transformations called {\sc sign-alsh}~\citeyearpar{shrivastava:signalsh} to transform MIPS into angular similarity search and obtained lower query time complexity than that of {\sc l{\footnotesize 2}-alsh}.
Neyshabur and Srebro showed that asymmetry is not necessary when queries are normalized and items have bounded 2-norm~\citeyearpar{neyshabur:simple-lsh}. They proposed {\sc simple-lsh}, which adopts a symmetric transformation and transforms MIPS into angular similarity search similar to {\sc sign-alsh}. However, they proved that {\sc simple-lsh} is a universal LSH for MIPS, while {\sc l{\footnotesize 2}-alsh} and {\sc sign-alsh} are not. {\sc simple-lsh} is also parameter-free and avoids the parameter tuning of {\sc l{\footnotesize 2}-alsh} and {\sc sign-alsh}. Most importantly, {\sc simple-lsh} achieves superior performance over {\sc l{\footnotesize 2}-alsh} and {\sc sign-alsh} in both theory and practice, and thus is the state-of-the-art hashing based algorithm for MIPS.
{\sc simple-lsh} requires the 2-norms of the items to be bounded, which is achieved by normalizing the items with the largest 2-norm in the dataset. However, real datasets often have long tails in the distribution of 2-norm, meaning that the largest 2-norm can be much larger than the majority of the items. As we will show in Section~\ref{sec:problem}, the excessive normalization of {\sc simple-lsh} makes the maximum inner product between the query and the items small, which harms the performance of {\sc simple-lsh} in both theory and practice.
To solve this problem, we propose {\sc norm-ranging lsh}. The idea is to partition the dataset into multiple sub-datasets according to the percentiles of the 2-norm distribution. For each sub-dataset, {\sc norm-ranging lsh} uses {\sc simple-lsh} as a subroutine to build an index independent of other sub-datasets. As each sub-dataset is normalized by its own maximum 2-norm, which is usually significantly smaller than the maximum 2-norm in the entire dataset, {\sc norm-ranging lsh} achieves lower query time complexity than {\sc simple-lsh}. To support efficient query processing, we also formulate a novel similarity metric which defines a probing order for buckets from different sub-datasets. We compare {\sc norm-ranging lsh} with {\sc simple-lsh} and {\sc l{\footnotesize 2}-alsh} on three real datasets and show empirically that {\sc norm-ranging lsh} offers an order of magnitude speedup.
\section{Locality Sensitive Hashing for MIPS}\label{sec:lsh}
\subsection{Locality Sensitive Hashing}
A definition of locality sensitive hashing (LSH)~\citep{indyk:lsh, andoni:approximate} is given as follows:
\begin{lsh}\label{def:lsh}
(LSH) A family $\mathcal{H}$ is called $(S_0, cS_0, p_1, p_2)$-sensitive if, for any two vectors $x, y\in\mathbb{R}^d$:
\begin{itemize}
\item if $sim(x,y)\geq S_0$, then $\mathbb{P}_{\mathcal{H}}\left[h(x)=h(y) \right]\geq p_1$,
\item if $sim(x,y)\leq cS_0$, then $\mathbb{P}_{\mathcal{H}}\left[h(x)=h(y) \right]\leq p_2$.
\end{itemize}
\end{lsh}
Note the original LSH is defined for distance function, we adopt a formalization adapted for similarity function~\citep{shrivastava:alsh}, which is more suitable for MIPS. For a family of LSH to be useful, it is required that $p_1>p_2$ and $0<c<1$. Given a family of $(S_0, cS_0, p_1, p_2)$-LSH, a query for $c$-approximate nearest neighbor search~\footnote{$c$-approximate nearest neighbor search solves the following problem: given parameters $S_0>0$ and $\delta>0$, if there exists an $S_0$-near neighbor of $q$ in $\mathcal{S}$, return some $cS_0$-near neighbor in $\mathcal{S}$ with probability at least $1-\delta$.} can be processed with a time complexity of $\mathcal{O}(n^\rho \log n)$, where $\rho=\frac{\log p_1}{\log p_2}$. For $L_2$ distance, there exists a well-known family of LSH defined as follows:
\begin{equation}\label{equ:l2lsh}
h_{a,b}^{L_2}(x)=\left \lfloor \frac{a^{\top}x+b}{r} \right \rfloor,
\end{equation}
where $\left \lfloor \right \rfloor$ is the floor operation, $a$ is a random vector whose entries follow i.i.d. standard normal distribution and $b$ is generated by a uniform distribution over $[0, r]$. When a hash function is drawn randomly and independently for each pair of vectors~\citep{wang:lsh}, the collision probability of~\eqref{equ:l2lsh} is given as:
\begin{equation}\label{equ:l2lsh_prob2}
\mathbb{P}_{\mathcal{H}}\left[h_{a,b}^{L_2}(x)=h_{a,b}^{L_2}(y)\right]=F_r(d)=1-2\Phi(-\frac{r}{d})-\frac{2d}{\sqrt{2\pi}r}(1-e^ {-(r/d)^2/2}),
\end{equation}
in which $\Phi(x)$ is the cumulative density function of standard normal distribution and $d=\Vert x-y \Vert$ is the $L_2$ distance between $x$ and $y$. For angular similarity, sign random projection is an LSH. Its expression and collision probability can be given as~\citep{goemans:sign}:
\begin{equation}\label{equ:sign}
h_a(x)=\mathrm{sign} (a^{\top}x), \mathbb{P}_{\mathcal{H}}\left[h_a(x)=h_a(y)\right]=1-\frac{1}{\pi} \cos^{-1} \left( \frac{x^{\top} y}{\Vert x \Vert \Vert y \Vert} \right),
\end{equation}
where the entries of $a$ follow i.i.d. standard normal distribution.
\subsection{L2-ALSH}
Shrivastava and Li proved that there exists no symmetric LSH for MIPS if the domain of the item $x$ and query $q$ are both $\mathbb{R}^d$~\citeyearpar{shrivastava:alsh}. They applied a pair of asymmetric transformations, $P(x)$ and $Q(q)$, to the items and the query, respectively.
\begin{equation}\label{equ:asy}
\begin{aligned}
&P(x)=[Ux;\Vert Ux \Vert^2;\Vert Ux \Vert^4;...;\Vert Ux \Vert^{2^m} ];
&Q(q)=[q;1/2;1/2;...;1/2]
\end{aligned}
\end{equation}
The scaling factor $U$ should ensure that $\Vert Ux \Vert<1$ for all $x\in \mathcal{S}$ and the query is normalized to unit 2-norm before the transformation. After the transformation, we have:
\begin{equation}\label{equ:res}
\Vert P(x)-Q(q) \Vert^2=1+\frac{m}{4}-2Ux^{\top}q+\Vert Ux \Vert^{2^{m+1}}.
\end{equation}
As the scaling factor $U$ is common for all items and the last term vanishes with sufficiently large $m$ because $\Vert Ux \Vert<1$, ~(\ref{equ:res}) shows that the problem of MIPS is transformed into finding the nearest neighbor of $Q(q)$ in terms of $L_2$ distance, which can be solved using the hash function in ~(\ref{equ:l2lsh}). Given $S_0$ and $c$, a query time complexity of $\mathcal{O}(n^\rho \log n)$ can be obtained for $c$-approximate MIPS with:
\begin{equation}\label{equ:alsh_rho}
\rho=\frac{\log F_r(\sqrt{1+m/4-2US_0+(US_0)^{2^{m+1}})}}{\log F_r(\sqrt{1+m/4-2cUS_0)}}.
\end{equation}
It is suggested to use a grid search to find the parameters ($m$, $U$ and $r$) that minimize $\rho$.
\subsection{SIMPLE-LSH}
Neyshabur and Srebro proved that {\sc l{\footnotesize 2}-alsh} is not a universal LSH for MIPS, that is~\citeyearpar{neyshabur:simple-lsh}, for any setting of $m$, $U$ and $r$, there always exists a pair of $S_0$ and $c$ such that $x^{\top} q=S_0$ and $y^{\top} q=cS_0$ but $\mathbb{P}_{\mathcal{H}}[h_{a,b}^{L_2}(P(x))=h_{a,b}^{L_2}(Q(q))]<\mathbb{P}_{\mathcal{H}}[h_{a,b}^{L_2}(P(y))=h_{a,b}^{L_2}(Q(q))]$. Moreover, they showed that asymmetry is not necessary if the items have bounded 2-norm and the query is normalized, which is exactly the assumption of {\sc l{\footnotesize 2}-alsh}. They proposed a symmetric transformation to transform MIPS into angular similarity search as follows:
\begin{equation}\label{equ:simpleLSH}
P(x)=[x;\sqrt{1-\Vert x \Vert^2}]; P(q)^{\top}P(x)=[q;0]^{\top}[x;\sqrt{1-\Vert x \Vert^2}]=q^{\top}x.
\end{equation}
They apply the sign random projection in~(\ref{equ:sign}) to $P(x)$ and $P(q)$ to obtain an LSH for $c$-approximate MIPS with a query time complexity $\mathcal{O}(n^\rho \log n)$ and $\rho$ is given as:
\begin{equation}\label{equ:simprho}
\rho=G(c,S_0)=\frac{\log (1-\frac{\mathrm{cos}^{-1}(S_0)}{\pi})}{\log (1-\frac{\mathrm{cos}^{-1}(cS_0)}{\pi})}.
\end{equation}
They called their algorithm {\sc simple-lsh} as it avoids the parameter tuning process of {\sc l{\footnotesize 2}-alsh}. Moreover, {\sc simple-lsh} is proved to be a universal LSH for MIPS under any valid configuration of $S_0$ and $c$. {\sc simple-lsh} also obtains better (lower) $\rho$ values than {\sc l{\footnotesize 2}-alsh} and {\sc sign-alsh} in theory and outperforms both of them empirically on real datasets~\citep{shrivastava:signalsh}.
\begin{figure}
\centering
\begin{minipage}[b]{0.25\columnwidth}
\label{fig:a}
\centering
\includegraphics[width=\linewidth]{./rho-crop.pdf}
\vspace{-1em}
\captionof*{figure}{(a)}
\end{minipage
\begin{minipage}[b]{0.25\columnwidth}
\centering
\includegraphics[width=1.03\linewidth]{./2NormImageNet-crop.pdf}
\captionof*{figure}{(b)}
\end{minipage
\begin{minipage}[b]{0.25\columnwidth}
\centering
\includegraphics[width=\linewidth]{./MaxmumInnerProductImageNet-crop.pdf}
\captionof*{figure}{(c)}
\end{minipage
\begin{minipage}[b]{0.25\columnwidth}
\label{fig:trans:b}
\centering
\includegraphics[width=\linewidth]{./MaxmumInnerProductImageNetMultilevel-crop.pdf}
\captionof*{figure}{(d)}
\end{minipage
\vspace{-0.5em}
\caption{(a) The relation between $\rho$ and $S_0$; (b) 2-norm distribution of the SIFT descriptors in the ImageNet dataset (maximum 2-norm scaled to 1); (c) The distribution of the maximum inner product of the queries after the normalization process of {\sc simple-lsh}; (d) The distribution of the maximum inner product of the queries after the normalization process of {\sc range-lsh} (32 sub-datasets).}
\label{fig:fig:rho and 2-norm}
\end{figure}
\section{Norm-Ranging LSH}\label{sec:multi-level}
In this section, we first motivate norm-ranging LSH by showing the problem of {\sc simple-lsh} on real datasets, then introduce how norm-ranging LSH (or {\sc range-lsh} for short) solves the problem.
\subsection{SIMPLE-LSH on real datasets}\label{sec:problem}
We plot the relation between $\rho$ and $S_0$ for {\sc simple-lsh} in Figure~\ref{fig:fig:rho and 2-norm}(a). Recall that the query time complexity of {\sc simple-lsh} is $\mathcal{O}(n^\rho \log n)$ and observe that $\rho$ is a decreasing function of $S_0$. As $\rho$ is large when $S_0$ is small, {\sc simple-lsh} suffers from poor query performance when the maximum inner product between a query and the items is small. Before applying the transformation in~(\ref{equ:simpleLSH}), {\sc simple-lsh} requires the 2-norm of the items to be bounded by 1, which is achieved by normalizing the items with the maximum 2-norm $U=\max_{x\in\mathcal{S}}\Vert x \Vert$. Assuming $q^{\top}x=S$ for item vector $x$, we have $q^{\top}x=S/U$ after normalization. If $U$ is significantly larger than $\Vert x \Vert$, the inner product will be scaled to a small value, and small inner product leads to high query complexity.
We plot the distribution of the 2-norm of a real dataset in Figure~\ref{fig:fig:rho and 2-norm}(b). The distribution has a long tail and the maximum 2-norm is much larger than the majority of the items. We also plot in Figure~\ref{fig:fig:rho and 2-norm}(c) the distribution of the maximum inner product of the queries after the normalization process of {\sc simple-lsh}. The results show that for the majority of the queries, the maximum inner product is small, which translates into a large $\rho$ and poor theoretical performance.
The long tail in 2-norm distribution also harms the performance of {\sc simple-lsh} in practice. If $\Vert x \Vert$ is small after normalization, the $\sqrt{1-\Vert x \Vert^2}$ term, which is irrelevant to the inner product between $x$ and $q$, will be dominant in $P(x)=[x;\sqrt{1-\Vert x \Vert^2}]$. In this case, the result of sign random projection in~(\ref{equ:sign}) will be largely determined by the last entry of $a$, causing many items to be gathered in the same bucket. In our sample run of {\sc simple-lsh} on the ImageNet dataset~\citep{deng:imagenet} with a code length of 32, there are only 60,000 buckets and the largest bucket holds about 200,000 items. Considering that the ImageNet dataset contains roughly 2 million items and 32-bit code offers approximately $4\times 10^9$ buckets, these statistics show that the large $\sqrt{1-\Vert x \Vert^2}$ term severely degrades bucket balance in {\sc simple-lsh}. Bucket balance is important for the performance of binary hashing algorithms such as {\sc simple-lsh} because they use Hamming distance to determine the probing order of the buckets~\citep{cai:revisit, gong:itq}. If the number of buckets is small or some buckets contain too many items, Hamming distance cannot define a good probing order for the items, which results in poor query performance.
\subsection{Norm-Ranging LSH}
\begin{algorithm}[tb]
\caption{Norm-Ranging LSH: Index Building}
\label{alg:mlevel hashing}
\begin{algorithmic}[1]
\STATE {\bfseries Input:} Dataset $\mathcal{S}$, dataset size $n$, number of sub-datasets $m$
\STATE {\bfseries Output:} A hash index $\mathcal{I}_j$ for each of the $m$ sub-datasets
\STATE Rank the items in $\mathcal{S}$ according to their 2-norms;
\STATE Partition $\mathcal{S}$ into $m$ sub-datasets \{${\mathcal{S}_1,\mathcal{S}_2,...,\mathcal{S}_m}$\} such that $\mathcal{S}_j$ holds items whose 2-norms ranked in the range $[\frac{(j-1)n}{m}, \frac{jn}{m}]$;
\FOR{every sub-dataset $\mathcal{S}_j$}
\STATE Use $U_j=\max_{x\in\mathcal{S}_j}\Vert x \Vert$ to normalize $\mathcal{S}_j$;
\STATE Apply {\sc simple-lsh} to build index $\mathcal{I}_j$ for $\mathcal{S}_j$;
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[tb]
\caption{Norm-Ranging LSH: Query Processing}
\label{alg:query}
\begin{algorithmic}[1]
\STATE {\bfseries Input:} Hash indexes \{${\mathcal{I}_1,\mathcal{I}_2,...,\mathcal{I}_m}$\} for the sub-datasets, query $q$
\STATE {\bfseries Output:} A $c$-approximate MIPS $x^{\star}$ to $q$
\FOR{every hash index $\mathcal{I}_j$}
\STATE Conduct MIPS with $q$ to get $x^{\star}_j$;
\ENDFOR
\STATE Select the item in \{$x^{\star}_1$, $x^{\star}_2$, ..., $x^{\star}_m$\} that has the maximum inner product with $q$ as the answer $x^{\star}$;
\end{algorithmic}
\end{algorithm}
The index building and query processing procedures of {\sc range-lsh} are presented in Algorithm~\ref{alg:mlevel hashing} and Algorithm~\ref{alg:query}, respectively. To solve the excessive normalization problem of {\sc simple-lsh}, {\sc range-lsh} partitions the items into $m$ sub-datasets according to the percentiles of the 2-norm distribution so that each sub-dataset contains items with similar 2-norms. Note that ties are broken arbitrarily in the ranking process of Algorithm~\ref{alg:mlevel hashing} to ensure that the percentiles based partitioning works even when many items have the same 2-norm. Instead of using $U$, i.e., the maximum 2-norm in the entire dataset, {\sc simple-lsh} uses the local maximum 2-norm $U_j=\max_{x\in\mathcal{S}_j}\Vert x \Vert$ in each sub-dataset for normalization, so as to keep the inner products of the queries large. In Figure~\ref{fig:fig:rho and 2-norm}(d), we plot the maximum inner product of the queries after the normalization process of {\sc range-lsh}. Comparing with Figure~\ref{fig:fig:rho and 2-norm}(c), the values of the inner product are significantly larger. As a result, the $\rho_j$ of sub-dataset $\mathcal{S}_j$ becomes $\rho_j=G(c, S_0/U_j)$, which is smaller than $\rho=G(c, S_0/U)$ if $U_j<U$. The smaller $\rho$ values translate into better query performance. The idea of dataset partitioning is also used in \citep{andoni:optimalhashing} for $L_2$ similarity search, where the partitioning is conducted in a pseudo-random manner. In the following, we prove that {\sc range-lsh} achieves a lower query time complexity bound than {\sc simple-lsh} under mild conditions.
\begin{theorem}\label{theoerm:comp}
{\sc range-lsh} attains lower query time complexity upper bound than that of {\sc simple-lsh} for $c$-approximate MIPS with sufficiently large $n$, if the dataset is partitioned into $m=n^{\alpha}$ sub-datasets and there are at most $n^{\beta}$ sub-datasets with $U_j=U$, where $0<\alpha<\mathrm{min}\{\rho, \frac{\rho-\rho^{\star}}{1-\rho^{\star}}\}$, $0<\beta<\alpha \rho$, $\rho^{\star}=\max_{\rho_j<\rho} \rho_j$, $\rho_j=G(c, S_0/U_j)$ and $\rho=G(c, S_0/U)$.
\end{theorem}
\begin{proof}
Firstly, we prove the correctness of {\sc range-lsh}, that is, it indeed returns a $cS_0$ approximate answer with probability at least $1-\delta$. Note that $S_0$ is a pre-specified parameter common to all sub-datasets rather than the actual maximum inner product in each sub-dataset. If there is an item $x^{\star}$ having an inner product of $S_0$ with $q$ in the original dataset, it is certainly contained in one of the sub-datasets. When we conduct MIPS on all the sub-datasets, the sub-dataset containing $x^{\star}$ will return an item having inner product $cS_0$ with $q$ with probability at least $1-\delta$ according to the guarantee of {\sc simple-lsh}. The final query result is obtained by selecting the optimal one (the one having the largest inner product with $q$) from the query answers of all sub-dataset according to Algorithm~\ref{alg:query}, which is guaranteed to be no less than $cS_0$ with probability at least $1-\delta$.
Now we analyze the query time complexity of {\sc range-lsh}. For each sub-dataset $\mathcal{S}_j$, it contains $n^{1-\alpha}$ items and the query time complexity upper bound of $c$-approximate MIPS is $\mathcal{O}(n^{(1-\alpha)\rho_j}\log n^{1-\alpha})$ with $\rho_j=G(c, S_0/U_j)$. As there are $m=n^{\alpha}$ sub-datasets, the time complexity of selecting the optimal one from the answers of all sub-datasets is $\mathcal{O}(n^{\alpha})$. Considering $\rho_j$ is an increasing function of $U_j$ and there are at most $n^{\beta}$ sub-datasets with $U_j=U$, the query time complexity of {\sc range-lsh} can be expressed as:
\begin{equation}
\begin{aligned}\label{equ:complex}
f(n&)\!=\!n^{\alpha}+\sum_{j=1}^{n^{\alpha}} n^{(1-\alpha)\rho_j} \log n^{1-\alpha}<\!n^{\alpha}+\sum_{j=1}^{n^{\alpha}} n^{(1-\alpha)\rho_j} \log n\\
&\!=\!n^{\alpha}\!+\!\sum_{j=1}^{\!\!n^{\alpha}-n^{\beta}}n^{(1-\alpha)\rho_j} \log n\!+\!n^{\beta}n^{(1-\alpha)\rho} \log n\\
&\!<\!n^{\alpha}+n^{\alpha}n^{(1-\alpha)\rho^{\star}} \log n+n^{\beta}n^{(1-\alpha)\rho} \log n
\end{aligned}
\end{equation}
Strictly speaking, the equal sign in the first line of~(\ref{equ:complex}) is not rigorous as the constants and non-dominant terms in the complexity of querying each sub-dataset are ignored. However, we are interested in the order rather than the precise value of query time complexity, so the equal sign is used for the conciseness of expression. Comparing $f(n)$ with the $\mathcal{O}(n^\rho \log n)$ complexity of {\sc simple-lsh},
\begin{equation}
\begin{aligned}\label{eqa:res}
\frac{f(n)}{n^\rho \log n}&<\frac{n^{\alpha}\!+\!\big(n^{\alpha}n^{(1-\alpha)\rho^{\star}}\!+\!n^{\beta}n^{(1-\alpha)\rho}\big) \log n}{n^\rho \log n}\\
&=n^{\alpha-\rho}/\log n+n^{\alpha+(1-\alpha)\rho^{\star}-\rho}+n^{\beta-\alpha \rho}
\end{aligned}
\end{equation}
~(\ref{eqa:res}) goes to 0 with sufficiently large $n$ when $\alpha\le \rho$, $\alpha+(1-\alpha)\rho^{\star}<\rho$ and $\beta-\alpha \rho<0$, which is satisfied by $\alpha<\mathrm{min}\{\rho, \frac{\rho-\rho^{\star}}{1-\rho^{\star}}\}$ and $\beta<\alpha \rho$.
\end{proof}
Note that the conditions of Theorem~(\ref{theoerm:comp}) can be easily satisfied. Theorem~(\ref{theoerm:comp}) imposes an upper bound instead of a lower bound on the number of sub-datasets, which is favorable as we usually do not want to partition the dataset into a large number of sub-datasets. Moreover, the condition that the number of sub-datasets with $U_j=U$ is smaller than $n^{\alpha \rho}$ is easily satisfied as very often only the sub-dataset that contains the items with the largest 2-norms has $U_j=U$. The proof also shows that {\sc range-lsh} is not limited to datasets with long tail in 2-norm distribution. As long as $U>U_j$ holds for most sub-datasets, {\sc range-lsh} can provide better performance than {\sc simple-lsh}. We acknowledge that {\sc range-lsh} and {\sc simple-lsh} are equivalent when all items have the same 2-norm. However, MIPS is equivalent to angular similarity search in this case, and thus can be solved directly with sign random projection rather than using {\sc simple-lsh}. In most applications that involve MIPS, there are considerable variations in the 2-norms of the items.
The lower theoretical query time complexity of {\sc range-lsh} also translates into much better bucket balance in practice. On the ImageNet dataset, {\sc range-lsh} with 32-bit code maps the items to approximately 2 million buckets and most buckets contain only 1 item. Comparing with the statistics of {\sc simple-lsh} in Section~\ref{sec:problem}, these numbers show that {\sc range-lsh} has much better bucket balance, and thus better ability to define a good probing order for the items. This can be explained by the fact that {\sc range-lsh} uses more moderate scaling factors for each sub-dataset than {\sc simple-lsh}, thus significantly reducing the magnitude of the $\sqrt{1-\Vert x \Vert^2}$ term in $P(x)=[x;\sqrt{1-\Vert x \Vert^2}]$.
\subsection{Similarity Metric}
Although the theoretical guarantee of LSH only holds when using multiple hash tables, in practice LSH is usually used in a single-table multi-probe fashion for candidate generation for similarity search~\citep{andoni:practical, lv:multi-probe}. The buckets(items) are ranked according to the number of identical hashes they have with the query (e.g., Hamming ranking) and the top-ranked buckets are probed first. Bucket ranking is challenging for {\sc range-lsh} as different sub-datasets use different normalization constants and buckets from different sub-datasets cannot be ranked simply according to their number of identical hashes. To support multi-probe in {\sc range-lsh}, we a similarity metric for bucket ranking that is efficient to manipulate.
Combining the index building process of {\sc range-lsh} and the collision probability of sign random projection in~(\ref{equ:sign}), the probability that an item $x\in \mathcal{S}_j$ and the query collide on one bit is $p=1-\frac{1}{\pi} \cos^{-1} \left( \frac{q^{\top} x}{U_j} \right)$, where $U_j$ is the maximum 2-norm in sub-dataset $\mathcal{S}_j$. Denote the code length as $L$ and the number of identical hashes bucket $b$ has with the query as $l$, we can obtain an estimate of the collision probability $p$ as $\hat{p}=l/L$. Plug $\hat{p}$ into the collision probability, we get an estimate $\hat{s}$ of the inner product between $q$ and the items in bucket $b$ as:
\begin{equation}\label {euq:similairty metric}
\hat{s}=U_j\cos\left[\pi(1-\frac{l}{L})\right].
\end{equation}
Therefore, we can compute $\hat{s}$ for the buckets(items) and use it for ranking. When $l>L/2$, $\cos\left[\pi(1-\frac{l}{L})\right]>0$, thus larger $U_j$ indicates higher inner product while the opposite is true when $l<L/2$. Since the code length is limited and $l/L$ can diverge from the actual collision probability $p$, it is possible that a bucket has large $U_j$ and large inner product with $q$, but it happens that $l<L/2$. In this case, it will be probed late in the query process, which query performance. To alleviate this problem, we adjust the similarity indicator to $\hat{s}=U_j\cos\left[\pi(1-\epsilon)(1-\frac{l}{L})\right]$, where $0<\epsilon<1$ is a small number. For the adjusted similarity indicator, $\cos\left[\pi(1-\epsilon)(1-\frac{l}{L})\right]<0$ only when $l<L\left[\frac{1}{2}-\frac{\epsilon}{2(1-\epsilon)}\right]$, which leaves some room to accommodate the randomness in hashing.
Note that the similarity metric in~\eqref{euq:similairty metric} can be manipulated efficiently with a complexity similar to Hamming distance. We can calculate the values of $\hat{s}$ for all possible combinations of $l$ and $U_j$, and sort them during index building. Note that the sorted structure is common for all queries and does not take too much space~\footnote{$l$ can take $L+1$ values, $U_j$ can take $m$ values, so the size of the sorted structure is $mL+m$.}. When a query comes, query processing can be conducted by traversing the sorted structure in ascending order. For a pair $(U_j, l)$, $U_j$ determines the sub-dataset while $l$ is used to choose the buckets to probe in that sub-dataset with standard hash lookup. We also provide an efficient method to rank the items when code length is large and there are many empty buckets in the supplementary material.
\begin{figure*}[t]
\input{netflix-top10}
\input{yahoomusic-top10}
\input{imagenet-top10}
\caption{Probed item-recall curve for top 10 MIPS on Netflix (top row), Yahoo!Music (middle row), and ImageNet (bottom row). From left to right, the code lengths are 16, 32 and 64, respectively.}
\label{fig:recall-time}
\end{figure*}
\section{Experimental Results}\label{sec:exp}
We used three popular datasets, i.e., Netflix, Yahoo!Music and ImageNet. For the Netflix dataset and Yahoo!Music dataset, the user and item embeddings were obtained using alternating least square based matrix factorization~\citep{yun:als}, and each embedding has 300 dimensions. We used the item embeddings as dataset items and the user embeddings as queries. The ImageNet dataset contains more than 2 million SIFT descriptors of the ImageNet images, and we sampled 1000 SIFT descriptors as queries and used the rest used as dataset items. Note that the 2-norm distributions of the Netflix and Yahoo!Music embeddings do not have long tail and the maximum 2-norm is close to the median (see the supplementary material), which helps verify the robustness of {\sc range-lsh} to different 2-norm distributions. For each dataset, we report the average performance of 1,000 randomly selected queries.
We compared {\sc range-lsh} with {\sc simple-lsh} and {\sc l{\footnotesize 2}-alsh}. For {\sc l{\footnotesize 2}-alsh}, we used the parameter setting recommended by its authors, i.e., $m=3$, $U=0.83$, $r=2.5$. For {\sc range-lsh}, part of the bits in the binary code are used to encode the index of the sub-datasets and the rest are generated by hashing. For example, if the code length is 16 and the dataset is partitioned into 32 sub-datasets, the 16-bit code of {\sc range-lsh} consists of 5 bits for indexing the 32 sub-datasets, while the remaining 11 bits are generated by hashing. We partitioned the dataset into 32, 64 and 128 sub-datasets under a code length of 16, 32 and 64, respectively. For fairness of comparison, all algorithms use the same total code length~\footnote{Experiment codes \url{https://github.com/xinyandai/similarity-search/tree/mipsex}.}. Following existing researches, we mainly compare the performance of the algorithms for single-table based multi-probing. While a comparison between the multi-table single probe performance of {\sc range-lsh} and {\sc simple-lsh} can be found in the supplementary material.
We plot the probed item-recall curves in Figure~\ref{fig:recall-time}. The results show that {\sc range-lsh} probes significantly less items compared with {\sc simple-lsh} and {\sc l{\footnotesize 2}-alsh} at the same recall. Due to space limitation, we only report the performance of top 10 MIPS, the performance under more configurations can be found in the supplementary material.
\input{partition-method}
Recall that Algorithm~\ref{alg:mlevel hashing} partitions a dataset into sub-datasets according to percentiles in the 2-norm distribution. We tested an alternative partitioning scheme, which divides the domain of 2-norms into \textit{uniformly} spaced ranges and items falling in the same range are partitioned into the same sub-dataset. The results are plotted in Figure~\ref{fig:partition and subdatasetnum}(a), which shows that uniform partitioning achieves slightly better performance than percentile partitioning. This proves {\sc range-lsh} is general and robust to different partitioning methods as long as items with similar 2-norms are grouped into the same sub-dataset. We also experimented the influence of the number of sub-datasets on performance in Figure~\ref{fig:partition and subdatasetnum}(b). The results show that performance improves with the number of sub-datasets when the number of sub-datasets is still small, but stabilizes when the number of sub-datasets is sufficiently large.
\section{Extension to L2-ALSH}\label{sec:general method}
In this section, we show that the idea of {\sc range-lsh}, which partitions the original dataset into sub-datasets with similar 2-norms, can also be applied to {\sc l{\footnotesize 2}-alsh}~\citep{shrivastava:alsh} to obtain more favorable (smaller) $\rho$ values than ~(\ref{equ:alsh_rho}). Note that we get~(\ref{equ:alsh_rho}) from~(\ref{equ:res}) as we only have $0\le \Vert x \Vert\le S_0$ if the entire dataset is considered. For a sub-dataset $\mathcal{S}_j$, if we have the range of its 2-norms as $u_{j-1} < \Vert x \Vert\le u_j$ and $u_{j-1}>0$, we can obtain the $\rho_j$ of $\mathcal{S}_j$ as:
\begin{equation}\label{equ:alsh_new_rho}
\rho_j\!=\!\frac{\log F_r(\sqrt{1+m/4-2U_jS_0+(U_ju_j)^{2^{m+1})}}}{\log F_r(\sqrt{1+m/4-2cU_jS_0+(U_ju_{j-1})^{2^{m+1}})}}.
\end{equation}
As $u_j\le S_0$ and $u_{j-1}>0$, the collision probability in the numerator increases while the the collision probability in the denominator decreases if we compare~\eqref{equ:alsh_new_rho} with~\eqref{equ:alsh_rho}. Therefore, we have $\rho_j<\rho$. Moreover, partitioning the original dataset into sub-datasets allows us to use different normalization factor $U_j$ for each sub-dataset and we only need to satisfy $U_j<1/u_j$ rather than $U<1/\max_{x\in\mathcal{S}}\Vert x \Vert$, which allows more flexibility for parameter optimization. Similar to Theorem~(\ref{theoerm:comp}), it can also be proved that dividing the dataset into sub-datasets results in an algorithm with lower query time complexity than the original {\sc l{\footnotesize 2}-alsh}. We show empirically that dataset partitioning improves the performance of {\sc l{\footnotesize 2}-alsh} in the supplementary material.
\section{Conclusions}\label{sec:conc}
Maximum inner product search (MIPS) has many important applications such as collaborative filtering and computer vision. We showed that, {\sc simple-lsh}, the state-of-the-art hashing method for MIPS, has critical performance limitations due to the long tail in the 2-norm distribution of real datasets. To tackle this problem, we proposed {\sc range-lsh}, which attains provably lower query time complexity than {\sc simple-lsh} under mild conditions. In addition, we also formulated a novel similarity metric that can be processed with low complexity. The experimental results showed that {\sc range-lsh} significantly outperforms {\sc simple-lsh}, and {\sc range-lsh} is robust to the shape of 2-norm distribution and different partitioning methods. We also showed that the idea of {\sc simple-lsh} hashing is general and can be applied to boost the performance of other hashing based methods for MIPS. The superior performance of {\sc range-lsh} can benefit many applications that involve MIPS.
\subsubsection*{Acknowledgments}
We thank the reviewers for their valuable comments. This work was supported in part by Grants (CUHK 14206715 \& 14222816) from the Hong Kong RGC.
|
\section{Introduction}
\label{2:Introduction}
The study of double pair plasma (DPP) has received an enormous attention in plasma physics research community due to their
ubiquitous existence in astrophysical environments, viz., upper regions of Titan's atmosphere \cite{El-Labany2012},
D-regions ($\rm H^+, O_2^-$) and F-regions ($\rm H^+, H^-$) of the Earth's ionosphere \cite{Elwakil2010} as well as potential diverse application in
laboratory experiments, viz., Fullerene ($\rm C^+, C^-$) \cite{Sabry2008,Oohara2003a,Oohara2003b}, neutral beam sources \cite{Bacal1979}, plasma processing
reactors \cite{Gottscho1986}, and laboratory experiment ($\rm Ar^+, F^-$) \cite{Jacquinot1977,Abdelwahed2016}, etc. Many researchers
have utilized wave dynamics, namely, ion-acoustic (IA) waves (IAWs) for understanding a variety of nonlinear structures, likely, shock, soliton,
envelope soliton \cite{Chowdhury2018}, and rogue waves (RWs), in the DPP medium (DPPM) \cite{El-Labany2012,Elwakil2010,Sabry2008}.
The presence of super-thermal electrons/positrons, which move faster than their thermal speed and observed
by Freja satellite \cite{Louran1994}, in DPPM are provided a great effect to generate various kind of nonlinear
phenomena, namely, modulational instability (MI), envelope soliton \cite{Chowdhury2018}, and
gigantic waves \cite{Chowdhury2017a}, etc. These super-thermal
electrons/positrons are described appropriately by the generalized Lorengian or $\kappa$-distribution
functions \cite{Vasyliunas1968,Chowdhury2017a}. This spectral index $\kappa$ in such $\kappa$-distribution
measures the strength of plasma particles. It may be noted that the small values of $\kappa$ represent a ``hard"
spectrum with a long tail while for large values of $\kappa$ (specially $\kappa\rightarrow\infty$)
represent the usual Maxwellian distributions \cite{Vasyliunas1968,Chowdhury2017a}. Ghosh \textit{et al.} \cite{Ghosh2012}
studied IA solitary waves (IASWs) in presence of $\kappa$-distributed electrons in a three component
unmagnetized plasma medium, and observed that $\kappa$ leads to decrease in the velocity of IASWs.
Hussain \textit{et al.} \cite{Hussain2012} examined that small values of $\kappa$ enhances the nonlinearity
of the plasma medium and the amplitude of the IASWs. Saha \textit{et al.} \cite{Saha2014} studied
IAWs in presence of $\kappa$-distributed electrons and positrons in electron-positron-ion (e-p-i) plasmas
and observed that the amplitude of the periodic wave decreases with $\kappa$. Chatterjee \textit{et al.} \cite{Chatterjee2011}
investigated IASWs in e-p-i plasma with $\kappa$-distributed electrons and positrons, and
found that the super-thermality of the electrons and positrons play a significant role on the collision of IASWs.
During the last few decades, the study of MI and associated nonlinear structures (due to the empirical results support the existence of
envelope soliton \cite{Chowdhury2018} or gigantic waves \cite{Abdelwahed2016,Chowdhury2017a} in fluid dynamics,
optical fiber, and plasma physics) is one of the
eye-catching topic for researchers. Actually, the MI leads to generate a new kind of high energy and very large amplitude structures
known as RWs and this kind of waves are short-lived phenomenon that emerge from nowhere and disappear without a trace.
Recently, a number of authors have studied the MI and RWs in DPPM. Sabry \cite{Sabry2008} studied MI in pair-ion plasma medium (PIPM) in
presence of dust impurities. Elwakil \textit{et al.} \cite{Elwakil2010} reported the propagation of the IAWs
in PIPM, and found that the non-thermality of the electrons decreases stable domain of the IAWs.
Abdelwahed \textit{et al.} \cite{Abdelwahed2016} studied propagation of IAWs in PIPM, and observed
that ratio of ion mass plays a vital role to manifest the IA RWs (IARWs) in the presence of
$\kappa$-distributed electrons. El-Labany \textit{et al.} \cite{El-Labany2012} investigated IAWs in three components PIPM in presence
of iso-thermal electrons, and found that the negative ion number density enhances the amplitude of the IARWs.
To the best of our knowledge, the effects of the $\kappa$-distributed electrons and positrons on the MI of the
IAWs and IARWs in a four component (comprising $\kappa$-distributed electrons and positrons, adiabatic positive
and negative ions) DPPM have not been investigated. Therefore, it is a practical interest to
examine the effects of $\kappa$-distributed electrons/positrons on the MI and IARWs in a four component DPPM.
The rest of the paper is organized in the following fashion: The model equations for the
IAWs in a DPPM with super-thermal electrons and positrons is presented in Sec. \ref{2:Governing equation}.
The stability of the IAWs is shown in Sec. \ref{1:Stability analysis}.
The discussion is provided in Sec. \ref{1:Discussion}.
\section{Model Equations}
\label{2:Governing equation}
We consider a collisionless unmagnetized four component plasma medium having inertialess
$\kappa$-distributed electrons (mass $m_e$; charge $q_e=-e$) and positrons (mass $m_p$; charge $q_p=e$),
inertial adiabatic negative ions (mass $m_-$; charge $q_-=-eZ_-$) as well as
adiabatic positive ion (mass $m_+$; charge $q_+=eZ_+$).
Here, $Z_-$ ($Z_+$) is the charge state of negative (positive) ion. The overall charge
neutrality condition at equilibrium can be expressed as $Z_+ n_{+0}+n_{p0}=n_{e0}+Z_-n_{-0}$,
where $n_{+0}$, $n_{p0}$, $n_{-0}$, and $n_{e0}$ are the equilibrium number densities of
positive ions, super-thermal positrons, negative ions, and super-thermal electrons, respectively.
Now, the dynamics of the plasma system can be expressed by the dimensionless form of the basic equations as:
\begin{eqnarray}
&&\hspace*{-1.3cm}\frac{\partial n_-}{\partial t} + \frac{\partial}{\partial x} (n_- u_-)=0,
\label{2eq:1}\\
&&\hspace*{-1.3cm}\frac{\partial u_-}{\partial t} + u_-\frac{\partial u_- }{\partial x}+3 \sigma_1 n_- \frac{\partial n_-}{\partial x}=\frac{\partial \phi}{\partial x},
\label{2eq:2}\\
&&\hspace*{-1.3cm}\frac{\partial n_+}{\partial t} + \frac{\partial}{\partial x} (n_+ u_+)=0,
\label{2eq:3}\\
&&\hspace*{-1.3cm}\frac{\partial u_+}{\partial t} + u_+\frac{\partial u_+ }{\partial x}+3 \sigma_2 n_+ \frac{\partial n_+}{\partial x}=-\alpha \frac{\partial \phi}{\partial x},
\label{2eq:4}\\
&&\hspace*{-1.3cm}\frac{\partial^2 \phi}{\partial x^2}=(\mu+\mu_p -1)n_e-\mu_p n_p + n_- -\mu n_+,
\label{2eq:5}\
\end{eqnarray}
The normalizing and associated parameters are represented as: $n_-=N_-/n_{-0}$, $n_+=N_+/n_{+0}$,
$n_e=N_e/n_{e0}$, $n_p=N_p/n_{p0}$, $u_-=U_-/C_-$, $u_+=U_+/C_-$, $x=X/\lambda_{D-}$, $t=T\omega_{p-}$,
$\phi=e\tilde{\Phi}/k_BT_e$, $C_-=\sqrt{Z_- k_BT_e/m_-}$, $\omega_{p-}=\sqrt{4 \pi e^2 Z_-^2 n_{-0}/m_-}$,
$\lambda_{D-}=\sqrt{k_BT_e/4 \pi e^2 Z_- n_{-0}}$, $P_-=P_{-0}(N_-/n_{-0})^\gamma$, $P_{-0}=n_{-0}k_BT_-$, $P_+=P_{+0}(N_+/n_{+0})^\gamma$,
$P_{+0}=n_{+0}k_BT_+$, $\gamma=(N+2)/N$, $\sigma_1=T_-/Z_- T_e$, $\sigma_2=m_- T_+/Z_- m_+ T_e$, $\mu=Z_+n_{+0}/Z_- n_{-0}$, $\mu_p=n_{p0}/Z_- n_{-0}$,
$\alpha=Z_+ m_-/Z_- m_+$; where $n_-$, $n_+$, $n_e$, $n_p$, $u_-$, $u_+$, $x$, $t$, $\tilde{\Phi}$, $C_-$, $\omega_{p-}$,
$\lambda_{D-}$, $T_-$, $T_+$, $T_e$, $T_p$, $P_{-0}$, and $P_{+0}$ is the number densities of negative ion, positive ion, electron, positron, negative ion fluid speed, positive ion fluid speed, space co-ordinate, time co-ordinate, electro-static wave potential, sound speed of negative ion, negative ion plasma frequency, negative ion Debye length, negative ion temperature, positive ion temperature, electrons temperature, positrons temperature, the equilibrium adiabatic pressure of the negative ions, and the equilibrium adiabatic pressure of the positive ions, respectively. Here, $N$ is the number of degrees of freedom and $N=1$ stands for adiabatic one-dimensional case. In addition, it is important to note that we consider for our numerical analysis $m_+ > m_-$, $n_{+0}>n_{-0}$, and $T_e$, $T_p$ $>$ $T_-$, $T_+$. The normalized number density of the $\kappa$-distributed
electrons is given by \cite{Vasyliunas1968,Chowdhury2017a}
\begin{eqnarray}
&&\hspace*{-1.3cm}n_e =\Big[1-\frac{\phi}{\kappa-3/2}\Big]^{-\kappa+\frac{1}{2}}
\nonumber\\
&&\hspace*{-0.8cm}=1+n_1 \phi +n_2 \phi^2 +n_3 \phi^3 +\ldots,
\label{2eq:6}
\end{eqnarray}
where
\begin{eqnarray}
&&\hspace*{-1.3cm}n_1 =\frac{(\kappa-1/2)}{(\kappa-3/2)},
\nonumber\\
&&\hspace*{-1.3cm}n_2 =\frac{(\kappa-1/2) (\kappa+1/2)}{2 (\kappa-3/2)^2},
\nonumber\\
&&\hspace*{-1.3cm}n_3 =\frac{(\kappa-1/2) (\kappa+1/2)(\kappa+3/2)}{6 (\kappa-3/2)^3}.
\nonumber\
\end{eqnarray}
It is essential to specify that small values of $\kappa$ represent strong super-thermality and
for a physically acceptable distribution, $\kappa>3/2$ is required. However, in the limit
$\kappa\rightarrow\infty$, the difference amongst kappa and Maxwellian distribution is negligible.
The normalized number density of the $\kappa$-distributed positrons is given by \cite{Vasyliunas1968,Chowdhury2017a}
\begin{eqnarray}
&&\hspace*{-1.3cm}n_p =\Big[1+\frac{\delta\phi}{\kappa-3/2}\Big]^{-\kappa+\frac{1}{2}}
\nonumber\\
&&\hspace*{-0.8cm}=1-n_1 \delta\phi +n_2 \delta^2\phi^2 -n_3 \delta^3\phi^3 +\ldots,
\label{2eq:7}
\end{eqnarray}
where $\delta=T_e/T_p$. Now, by substituting \eqref{2eq:6} and \eqref{2eq:7} into \eqref{2eq:5}, and expanding the equation up to third order, we get
\begin{eqnarray}
&&\hspace*{-1.3cm}\frac{\partial^2 \phi}{\partial x^2}+1+\mu n_+=\mu+n_- +\gamma_1 \phi+\gamma_2\phi^2 +\gamma_3 \phi^3 +\ldots,
\label{2eq:8}\
\end{eqnarray}
where
\begin{eqnarray}
&&\hspace*{-1.3cm}\gamma_1= n_1 (\mu+\mu_p-1+\mu_p\delta),
\nonumber\\
&&\hspace*{-1.3cm}\gamma_2= n_2 (\mu+\mu_p-1-\mu_p\delta^2),
\nonumber\\
&&\hspace*{-1.3cm}\gamma_3= n_3 (\mu+\mu_p-1+\mu_p\delta^3).
\nonumber\
\end{eqnarray}
In order to analyze the one-dimensional electrostatic perturbations propagating in our DPPM, we will
derive the nonlinear Schr\"{o}dinger equation (NLSE) by employing the reductive perturbation method (RPM).
The independent variables are stretched as $\xi=\epsilon (x-v_gt)$ and $\tau=\epsilon^2t$, where
$\epsilon$ is a small expansion parameter and $v_g$ is the group velocity of the IAWs. All dependent
variables can be expressed in a power series of $\epsilon$ as:
\begin{eqnarray}
&&\hspace*{-1.3cm}\Lambda (x.t)=\Lambda_0+\sum_{m=1}^\infty\epsilon ^{(m)}\sum_{l=-\infty}^\infty \Lambda_{l}^{(m)}(\xi,\tau)~\mbox{exp}(il\Upsilon),
\label{2eq:9}\
\end{eqnarray}
where $\Lambda_{l}^{(m)}=[n_{-l}^m$, $u_{-l}^m$, $n_{+l}^m$, $u_{-l}^m$, $\phi_l^m$], $\Lambda_0=[1, 0, 1, 0, 0]^T$, and
$\Upsilon=kx-wt$. Here, the carrier wave number $k$ and frequency $\omega$ are real variables.
We are going to parallel mathematical steps as Chowdhury \textit{et al.} \cite{Chowdhury2018} have done in their work to find successively
the IAWs dispersion relation, group velocity, and NLSE. The IAWs dispersion relation
\begin{eqnarray}
&&\hspace*{-1.3cm}\omega^2=\frac{A \pm \sqrt{ A^2 - 4 BD}}{2B}.
\label{2eq:10}
\end{eqnarray}
where
$A=k^2(\lambda_1 k^2 +\gamma_1\lambda_1+\lambda_2 k^2+\gamma_1\lambda_2+1+\alpha \mu)$,
$B=(k^2 +\gamma_1)$, $D= k^4(\lambda_1 \lambda_2 k^2 + \gamma_1\lambda_1 \lambda_2+\lambda_1\alpha \mu +\lambda_2)$,
$\lambda_1=3 \sigma_1$, and $\lambda_2= 3 \sigma_2$.
Equation \eqref{2eq:10} describes both the fast (for positive sign) and slow (for negative sign) IA modes denoted by $\omega_f$ and $\omega_s$, respectively.
It is worth mentioning that the condition $A^2 >4 BD$ should be verified to get real and positive values of $\omega$.
The group velocity $v_g$ of IAWs can be expressed as
\begin{eqnarray}
&&\hspace*{-1.3cm}v_g= \frac{S^2 F_1+ M^2(\lambda_1 k^2 +\omega^2 -S)}{2\omega k(M^2 + \alpha \mu S^2)},
\label{2eq:11}\
\end{eqnarray}
where $F_1=(\alpha \mu \omega^2 -2 M^2 +\alpha \mu M + k^2 \lambda_2\alpha \mu)$, $S=k^2 \lambda_1 -\omega^2$, and $M= \omega^2 - k^2 \lambda_2$. Finally, one can obtain the following NLSE:
\begin{eqnarray}
&&\hspace*{-1.3cm}i \frac{\partial \Phi}{\partial \tau} + P \frac{\partial^2 \Phi}{\partial \xi^2}+ Q |\Phi|^2 \Phi=0,
\label{2eq:12}\
\end{eqnarray}
where $\Phi=\phi_1^{(1)}$ for simplicity. The dispersion coefficient $P$ and the nonlinear coefficient $Q$ are, respectively, given by
\begin{eqnarray}
&&\hspace*{-1.3cm}P= \frac{F_2-S^2 M^2}{2\omega k^2(\alpha \mu S^2 +M^2)},
\label{2eq:13}\\
&&\hspace*{-1.3cm}Q=\frac{F_3}{2 \omega k^2 (\alpha \mu S^2 +M^2)},
\label{2eq:14}\
\end{eqnarray}
where
\begin{figure}[!tbp]
\centering
\subfloat[]{\includegraphics[width=70mm, height=40mm]{F1.eps}\label{2Fig:F1}}
\hfill
\subfloat[]{\includegraphics[width=70mm, height=40mm]{F2.eps}\label{2Fig:F2}}
\hfill
\subfloat[]{\includegraphics[width=70mm, height=40mm]{F3.eps}\label{2Fig:F3}}
\hfill
\caption{The variation of $P/Q$ with $k$ for different values of (a) $\alpha$ (when $\omega_f$ and $\kappa=2.0$); (b) $\alpha$ (when $\omega_s$ and $\kappa=2.0$); (c) $\kappa$ (when $\omega_f$ and $\alpha=0.5$); along with $\delta=1.3$, $\mu=1.2$, $\mu_p=0.3$, $\sigma_1=0.005$, and $\sigma_2=0.04$.}
\end{figure}
\begin{eqnarray}
&&\hspace*{-1.3cm}F_2=\frac{M^2}{S} [(v_g \omega k-\lambda_1 k^2)( \lambda_1 k^2-2v_g\omega k+\omega^2-S)
\nonumber\\
&&\hspace*{-0.6cm}+(v_g k-\omega)(\omega \lambda_1 k^2 -2v_g k\omega^2 +\omega^3-v_g S k)]
\nonumber\\
&&\hspace*{-0.6cm}-\frac{\alpha \mu S^2}{M}[(v_g \omega k -\lambda_2 k^2)(\lambda_2 k^2 - 2v_g \omega k +M+\omega^2)
\nonumber\\
&&\hspace*{-0.6cm}+(v_g k - \omega)(\omega \lambda_2 k^2 + v_g M k + \omega^3 - 2 v_g k \omega^2)],
\nonumber\\
&&\hspace*{-1.3cm}F_3=3 \gamma_3 M^2 S^2-\alpha \mu k^2 S^2 (\lambda_2 k^2+\omega^2)(C_3 +C_8)
\nonumber\\
&&\hspace*{-0.6cm}+2 \gamma_2 M^2 S^2 (C_5 +C_{10})-2 \alpha \mu \omega S^2 k^3(C_4 +C_9)
\nonumber\\
&&\hspace*{-0.6cm}-k^2 M^2(\lambda_1 k^2+ \omega^2)(C_1 +C_6)-2 \omega M^2 k^3(C_2 +C_7).
\nonumber\
\end{eqnarray}
\begin{eqnarray}
&&\hspace*{-1.3cm}C_1=\frac{k^2 ( 2C_5 S^2 -3 \omega^2 k^2 -\lambda_1 k^4)}{2 S^3},
\nonumber\\
&&\hspace*{-1.3cm}C_2=\Big(\frac{C_1 \omega}{k} -\frac{\omega k^3}{S^2}\Big),
\nonumber\\
&&\hspace*{-1.3cm}C_3=\frac{\alpha k^2 (2 C_5 M^2 + 3 \alpha \omega^2 k^2 +\alpha \lambda_2 k^4 )}{2 M^3},
\nonumber\\
&&\hspace*{-1.3cm}C_4=\frac{\omega (C_3 M^2 -\alpha^2 k^4)}{k M^2},
\nonumber\\
&&\hspace*{-1.3cm}C_5=\frac{ F_4- 2\gamma_2 S^3 M^3}{2 S^3 [M^3 (4k^2+\gamma_1)-\alpha \mu k^2 M^2]+ 2S^2 k^2 M^3},
\nonumber\\
&&\hspace*{-1.3cm} F_4=M^3 k^4 (3 \omega^2 + \lambda_1 k^2)+ \mu S^3 k^4(3 \alpha^2 \omega^2 + \lambda_2 \alpha^2 k^2),
\nonumber\\
&&\hspace*{-1.3cm}C_6=\frac{k^2(\lambda_1 k^2+2 v_g \omega k+\omega^2)-C_{10}S^2}{S^2 (v_g^2 -\lambda_1)},
\nonumber\\
&&\hspace*{-1.3cm}C_7=\frac{C_6 v_g S^2 - 2 \omega k^3}{S^2},
\nonumber\\
&&\hspace*{-1.3cm}C_8=\frac{\alpha^2 k^2(2 v_g \omega k+ \lambda_2 k^2 + \omega^2)+ C_{10} \alpha M^2}{M^2 (v_g^2-\lambda_2)},
\nonumber\\
&&\hspace*{-1.3cm}C_9=\frac{C_8 v_g M^2 - 2 \omega \alpha^2 k^3}{M^2},
\nonumber\\
&&\hspace*{-1.3cm}C_{10}=\frac{2\gamma_2 S^2 M^2 (v_g^2 -\lambda_1)(v_g^2-\lambda_2)+F_5}{S^2 M^2 F_6},
\nonumber\\
&&\hspace*{-1.3cm}F_5= M^2 k^2 (2 v_g k w +\lambda_1 k^2 + \omega^2)(v_g^2-\lambda_2)
\nonumber\\
&&\hspace*{-0.6cm}-\mu S^2 k^2(2v_g \omega k \alpha^2+\lambda_2 \alpha^2 k^2+\alpha^2 \omega^2)(v_g^2-\lambda_1),
\nonumber\\
&&\hspace*{-1.3cm}F_6=[v_g^2 (\alpha \mu +1)-(\alpha \mu \lambda_1+\lambda_2)-\gamma_1(v_g^2-\lambda_1)(v_g^2-\lambda_2)].
\nonumber\
\end{eqnarray}
\section{Stability analysis}
\label{1:Stability analysis}
The sign of the ratio $P/Q$ can recognize the stable/unstable domain for the IAWs in presence of the external perturbations.
The positive (negative) sign of the ratio $P/Q$ defines, respectively, unstable (stable) domain for the IAWs \cite{Sultana2011,Fedele2002,Kourakis2005}.
The intersecting point, in which the stable and unstable domain can be obtain for IAWs, of the $P/Q$ curve
with $k-$axis in $P/Q$ versus $k$ graph is known as critical/threshold wave number ($k_c$).
We have investigated the stable/unstable domain for the IAWs by depicting $P/Q$ versus $k$ graph for different
values of $\alpha$ and $\kappa$ in Figs. \ref{2Fig:F1}, \ref{2Fig:F2}, and \ref{2Fig:F3} for $\omega_f$, $\omega_s$, and $\omega_f$, respectively.
It can be seen from these figures that (a) the IAWs remain stable for small $k$ ($k<k_c$) and the MI sets in for large values of $k$ ($k>k_c$);
(b) the $k_c$ lies almost in the range of $0.49$ to $0.61$ depending upon the value of $\alpha$ (for fast IA mode) and
the $\alpha$ reduces the stable domain of the IAWs (see Fig. \ref{2Fig:F1}). The variation of $k_c$ with $\alpha$ for slow mode,
exactly an opposite trend is observed with respect to the fast IA mode, can be seen from Fig. \ref{2Fig:F2}.
In this case the $k_c$ bears a value around $3.3$ to $3.6$ (see Fig. \ref{2Fig:F2}).
So, the negative ion mass reduces (enhances) the stable domain of the IAWs corresponds to fast (slow) IA
modes for constant values of positive ion mass ($m_+$), charge state of the positive ($Z_+$) and negative ($Z_-$) ions (via $\alpha$).
\begin{figure}[!tbp]
\centering
\subfloat[]{\includegraphics[width=70mm, height=40mm]{F4.eps}\label{2Fig:F4}}
\hfill
\subfloat[]{\includegraphics[width=70mm, height=40mm]{F5.eps}\label{2Fig:F5}}
\hfill
\subfloat[]{\includegraphics[width=70mm, height=40mm]{F6.eps}\label{2Fig:F6}}
\hfill
\caption{The variation of $|\Phi|$ with $\xi$ for different values of (a) $\mu$ (when $\delta=1.3$ and $\mu_p=0.3$);
(b) $\mu_p$ (when $\delta=1.3$ and $\mu=1.2$); (c) $\delta$ (when $\mu=1.2$ and $\mu_p=0.3$);
along with $\alpha=0.5$, $\sigma_1=0.005$, $\sigma_2=0.04$, $\tau=0$, $k=0.7$, and $\omega_f$.}
\end{figure}
The effect of the super-thermality on the stability domain of the IAWs can be observed from
Fig. \ref{2Fig:F3}. This figure shows that by decreasing $\kappa$ leads to an
increase in the stable domain of IAWs and this result agrees with the result
of Alinejad \textit{et al.} \cite{Alinejad2014} and Gharaee \textit{et al.} \cite{Gharaee2011} works.
In the MI region, a random perturbation in the oscillating ambient background causes
the exponential growth of IAWs amplitude, then rapidly decay without leaving any trace.
This mechanism can be expressed by the first order rational solution or rogue waves, which
developed by the Darboux Transformation Scheme, solution of the NLSE (\ref{2eq:12})
and can be expressed as \cite{Akhmediev2009,Ankiewiez2009}
\begin{eqnarray}
&&\hspace*{-1.3cm}\Phi(\xi,\tau)=\sqrt{\frac{2P}{Q}}\Big[\frac{4+16 i\tau P}{1+4 \xi^2 + 16\tau^2 P^2}-1\Big] \mbox{exp} (2i\tau P),
\label{2eq:15}
\end{eqnarray}
The solution reveals that a significant amount of IAWs energy is concentrated
into a comparatively small region in DPPM. We are interested to investigate the
effect of various plasma parameters on the features of IARWs [obtained numerically
by using \eqref{2eq:15}]. The effects of the adiabatic positive and negative ion
population, in fact their charge state, on the IARWs can be observed from Fig. \ref{2Fig:F4}
and it is obvious that (a) as we increase the value of $\mu$, the height of the
IARWs decreases; (b) the wave potential increases with
increase in the value of negative ion population ($n_{-0}$), but decreases with
increase of the positive ion population ($n_{+0}$) when $Z_+$ and $Z_-$ remain constant.
Similar fashion is observed from Fig. \ref{2Fig:F5}. In this case, the number density of the
positron (negative ion) minimizes (maximizes) the nonlinearity of the DPPM,
and decreases (increases) the height of the IARWs for constant value of the $Z_-$ (via $\mu_p$).
So, it is clear from both Figs. \ref{2Fig:F4} and \ref{2Fig:F5} that the negative ion population enhances the nonlinearity of the
DPPM in the unstable domain of the IAWs and this is a good agreement with the result of El-Labany \textit{et al.} \cite{El-Labany2012} work.
Figure \ref{2Fig:F6} discloses the effect of electron and positron temperature on the IARWs (via $\delta$).
The electron temperature reduces the nonlinearity of the DPPM, i.e., the height of the IARWs
decreases, but increases with the increase positron temperature, i.e., the height of the IARWs increases.
\section{Discussion}
\label{1:Discussion}
In summary, a NLSE has been successfully derived to examine and numerically analyzed the MI of IAWs in a DPPM
composed of $\kappa-$distributed electrons and positrons, negatively and positively charged adiabatic ions.
The core results from our present investigation can be summarized as follows:
\begin{enumerate}
\item{The $k_c$, which separates the stable domain from the unstable region, totally depends on the super-thermality of the electrons and positrons, and negative ion mass.}
\item{The negative ion population enhances the nonlinearity of the DPPM in the unstable domain of the IAWs, and increases the height of the IARWs.}
\item{The population of the positron minimizes the nonlinearity of the plasma medium by depicting smaller IARWs.}
\item{The electron (positron) temperature reduces (increases) the height of the IARWs.}
\end{enumerate}
The implications of our results should be useful to understand the nonlinear phenomena (MI and IARWs)
in astrophysical environments, namely, upper regions of Titan's atmosphere \cite{El-Labany2012},
D-region ($\rm H^+, O_2^-$) and F-region ($\rm H^+, H^-$) of the Earth's ionosphere \cite{Elwakil2010} as well as in
laboratory experiments, namely, pair-ion Fullerene ($\rm C^+, C^-$) \cite{Sabry2008,Oohara2003a,Oohara2003b}.
\section*{References}
|
\section{Introduction}
Causal inference has been given rise to extensive attention and applied in several areas including statistics, neuroscience and sociology in recent years. An efficient approach for causal discovery is to conduct randomized controlled experiments. These experiments, however, are usually very expensive and sometimes practically infeasible. Therefore, causal inference methods using passive observational data take center stage, and many of them have been proposed, especially in the past ten years.
Existing causal inference methods that use passive observational data can be roughly categorized into two classes according to their objectives. One class of methods aim at identifying the variable that is the cause of the other in a variable pair \citep{hoyer2009nonlinear,zhang2009identifiability,janzing2012information, chen2014causal}, which is often termed a cause-effect pair. Most of the methods in this class first model the relation between the cause and the effect using a functional model with certain assumptions. Then they derive a certain property which only holds in the causal direction and is violated in the anticausal direction to infer the true causal direction. This kind of widely used property is often termed cause-effect asymmetry. For example, the additive noise model (ANM) \citep{hoyer2009nonlinear} represents the effect as a function of the cause with an additive independent noise: $Y = f(X) + E_{Y}$. The authors showed that there is no model of the form $X = g(Y) + E_{X}$ that admits an ANM in the anticausal direction for most combinations $\left(f, p(X), p(E_{Y})\right)$. Therefore, the inference of ANM is done by finding the direction that fits ANM better. Similar methods include postnonlinear model (PNL) \citep{zhang2009identifiability} and information geometric causal inference (IGCI) \citep{janzing2012information}. Recently, a kernel-based, EMD (or abbreviation for EMbeDding) \citep{chen2014causal} using the framework of IGCI is proposed. EMD differs from the previous methods in the sense that it does not assume any specific functional model, but it still resorts to find the cause-effect asymmetry.
The other class of methods aims at recovering the structure of causal graphs. Constraint-based methods \citep{spirtes1991algorithm, spirtes1991probability, perlcausality, spirtes2000causation, Cheng200243}, which belong to this class, exploit the causal Markov condition and have been widely used in the social sciences, medical science, and bioinformatics. However, these methods allow one only to obtain the Markov equivalent class of the graph and are of high computational cost. In 2006, a linear nongaussian acyclic model (LiNGAM) \citep{shimizu-et-al:lingam} which exploits the nongaussian property of the noise, was showed to be able to recover the full causal structure by using independent component analysis (ICA) \citep{comon1994independent, hyvarinen2000independent}. To avoid the problem that ICA may result in a solution of local optima, different methods \citep{shimizu2011directlingam, hyvarinen2013pairwise} were proposed to guarantee the correctness of the causal order of variables in the causal graph.
Both classes of existing methods are based on the assumption that all observations are sampled from a fixed causal model. By ``fixed causal model,'' we mean that the (joint) distribution of variables and the mechanism mapping cause(s) to effect(s) are unchanged during the data collecting process. For example in an ANM $Y = f(X) + E_{Y}$, both the distribution of the cause $p(X)$ and the causal mechanism $f$ are assumed to be fixed. Although some of these methods do achieve inspiring results and provide valuable insights for subsequent research, data generated from a varying causal model are much more common in practice and existing methods based on a fixed causal model would come across some problems when applied to varying causal models \citep{zhang2015discovery}. Therefore, we consider causal models where distributions of variables and causal mechanisms vary across domains or over different time periods and call these models \emph{non-stationary causal models}. An example is the model of daily returns of different stocks. The distribution of the return of each stock varies with the financial status, and the causal mechanisms between different stocks also vary according to the relations between these companies. Recently, a method called Enhanced Constraint-based Procedure (ECBP) was proposed for causal inference of non-stationary causal models \citep{zhang2015discovery}. The authors resorted to an index variable $C$ to quantify the nonstationarity and proposed ECBP, which is built on constraint-based methods to recover the skeleton of the augmented graph, which consists of both observed variables $\mathbf{V}$ and some unobserved quantities determined by $C$. They also showed that it is possible to infer the parent nodes of variables adjacent to $C$ (termed \emph{$C$-specific variables}) and proposed a measure to infer the causal direction between each $C$-specific variable and its parents. However, their method fails to ensure the recovery of the full causal structure, which is due to the limitation of methods that rely on conditional independence test. In contrast, our method, which is proposed originally for cause-effect pairs inference, is also extended to infer the complete causal structure of two kinds of graphs by transforming the nonstationarity into a LiNGAM model.
In this paper, we introduce a nonstationary causal model and develop algorithms, which we call \emph{embedding-based nonstationary causal model inference} (ENCI) for inferring the complete causal relations of the model. Our model assumes that the underlying causal relations (i.e. the causal direction of a cause-effect pair or the causal structure of a graph) are fixed, whereas the distributions of variables and the causal mechanisms (i.e. the conditional distribution of the effect given the cause(s)) change across domains or over different time periods. To infer the nonstationary causal model, ENCI reformulates the relation among variables into a linear model in the Reproducing Kernel Hilbert Space (RKHS) and leverages the identifiability of the linear causal model to tackle the original complicated problem. Specifically, for a cause-effect pair, we embed the variation of the density of each variable into an RKHS to transform the original unknown causal model to a linear nongaussian additive model \citep{kano2003causal} based on the independence between the mechanism generating the cause and the mechanism mapping the cause to the effect. Then we infer the causal direction by exploiting the causal asymmetry of the obtained linear model. We also extend our approach to discover the complete causal structure of two kinds of causal graphs in which the distribution of each variable and the causal mechanism mapping cause(s) to effect(s) vary and the causal mechanism could be nonlinear.
This paper is organized as follows. In section 2, we formally define our model and objective of causal inference. In section 3, some preliminary knowledge of reproducing kernel Hilbert space embedding is introduced. In section 4, we elaborate our methods for cause-effect pairs. In section 5, we extend our methods to two kinds of causal graphs. In section 6, we report experimental results on both synthetic and real-world data to show the advantage of our approach over existing ones.
\section{Problem Description}
In this section we formalize the nonstationary causal model and the objective of our causal inference task. For a pair of variable $X$ and $Y$, we consider the case where $X$ is the cause and $Y$ is the effect without loss of generality throughout this paper.
\subsection{Non-stationary Causal Model}
We assume the data generating process of a cause-effect pair fulfills the following properties:
\begin{itemize}
\item The causal direction between $X$ and $Y$ stays the same throughout the process.
\item Observations are collected from $N$ different domains. The density of the cause $\left(p(X)\right)$ and the conditional density of the effect given the cause $\left(p(Y|X)\right)$ are fixed within each domain.
\item $p(X)$ and $p(Y|X)$ vary in different domains.
\end{itemize}
We call this a \emph{nonstationary causal model} due to the variation in distributions over domains. The data-generating process is illustrated in Figure~\ref{fig:model}.
\begin{figure}[hbtp]
\label{model}
\begin{center}
\includegraphics[width = 0.8\linewidth]{model2.png}
\end{center}
\caption{Data generating process of non-stationary causal model}
\label{fig:model}
\end{figure}
The collection of data obtained from each domain is called a data group $G_{i}$, and the entire data set is denoted by $\mathbf{G} = \lbrace G_{1}, G_{2}, \dots, G_{N} \rbrace$. This nonstationarity over groups is common in the real world, as the observations we obtained are usually collected over different time periods or from different sources (e.g. different geographical regions or experimental settings).
\subsection{Objective of Non-stationary Causal Model Inference}
Our goal of nonstationary causal model inference is, by exploiting the variation of distributions in different groups, to accurately estimate the causal direction between $X$ and $Y$. We also extend, our approach to learn the full causal structure of two kinds of causal graphs by transforming their relationship among groups into a LiNGAM model. For clarity, we list some of the notations we use in the following sections in Table~\ref{notation}.
\begin{table}[!ht]
\caption{Notations}
\label{notation}
\begin{center}
\begin{tabular}{ll}
\toprule
Symbol & Description \\
\midrule
$p^{(i)}(X), p^{(i)}(Y)$ & Density of $X$, $Y$ in group $i$ \\
$\overline{p}(X)$ & Base of the density of $X$ \\
$\Delta p^{(i)}(X)$ & Variation of the density of $X$ in group $i$ \\
$p^{(i)}(Y|X)$ & Conditional density of $Y$ given $X$ in group $i$ \\
$\overline{p}(Y|X)$ & Base of the conditional density of $Y$ given $X$ \\
$\Delta p^{(i)}(Y|X)$ & Variation of the conditional density of $Y$ given $X$ in group $i$ \\
$\mathcal{X}, \mathcal{Y}$ & domain of variable $X$, $Y$ \\
$\mu^{(i)}_{\otimes X}$, $\mu^{(i)}_{\otimes Y}$ & Mean embedding of $p^{(i)}(X)$ in $\mathcal{X} \otimes \mathcal{X}$, $p^{(i)}(Y)$ in $\mathcal{Y} \otimes \mathcal{Y}$ \\
$\overline{\mu}_{\otimes X}$, $\overline{\mu}_{\otimes Y}$ & Mean embedding of $\overline{p}(X)$ in $\mathcal{X} \otimes \mathcal{X}$, $\overline{p}(Y)$ in $\mathcal{Y} \otimes \mathcal{Y}$ \\
$\Delta \mu^{(i)}_{\otimes X}$, $\Delta \mu^{(i)}_{\otimes Y}$ & Mean embedding of $\Delta p^{(i)}(X)$ in $\mathcal{X} \otimes \mathcal{X}$, $\Delta p^{(i)}(Y)$ in $\mathcal{Y} \otimes \mathcal{Y}$ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\section{Hilbert Space Embedding of Distributions}
Kernel embedding-based approaches represent probability distributions by elements in a reproducing kernel Hilbert space (RKHS) and it serves as the main tool in this letter to characterize distributions.
An RKHS $\mathcal{F}$ over $\mathcal{X}$ with a kernel $k$ is a Hilbert space of functions $f:\mathcal{X} \rightarrow \mathbb{R}$. Denoting its inner product by $\langle \cdot, \cdot \rangle_{\mathcal{F}}$, RKHS $\mathcal{F}$ fulfills the reproducing property $\langle f(\cdot), k(x,\cdot) \rangle_{\mathcal{F}} = f(x)$. People often regard $\phi(x) \coloneqq k(x,\cdot)$ as a feature map of $x$. Kernel embedding of a marginal density $p(X)$ \citep{smola2007hilbert} is defined as the expectation of its feature map:
\begin{equation}
\mu_{X} \coloneqq \mathbb{E}_{X}[\phi(X)] = \int _{ \mathcal{X} }^{ }{ \phi(x)p(x)dx },
\label{knl:1}
\end{equation}
where $\mathbb{E}_{X}[\phi(X)]$ is the expectation of $\phi(X)$. It has been shown that $\mu_{X}$ is guaranteed to be an element in RKHS if $\mathbb{E}_{X}[k(X,X)]<\infty$ is satisfied. It is also generalized to joint distribution using tensor product feature spaces. The kernel embedding of a joint density $p(X, Y)$ is defined as
\begin{align}
\mathcal{C}_{XY} & \coloneqq \mathbb{E}_{XY}[\phi(X) \otimes \phi(Y)] = \int_{\mathcal{X} \times \mathcal{Y}}{ \phi(x) \otimes \phi(y)p(x, y)dxdy}.
\label{knl:2}
\end{align}
Similarly, we have that $\mathcal{C}_{XX} \coloneqq \mathbb{E}_{X}[\phi(X) \otimes \phi(X)]$. The embedding of conditional densities is viewed as an operator that maps from $\mathcal{F}$ to $\mathcal{G}$ which is an RKHS over $\mathcal{Y}$ \citep{song2009hilbert}. Imposing that the conditional embedding satisfies the following two properties:
\begin{align}
& \mu_{Y|x} \coloneqq \mathbb{E}_{Y|x}[\phi(Y)|x] = \mathcal{U}_{Y|X}k(x, \cdot), \label{req:1}\\
& \mathbb{E}_{Y|x}[g(Y)|x]=\langle g, \mu_{Y|x} \rangle_{\mathcal{G}}, \label{req:2}
\end{align}
where $g \in \mathcal{G}$ and $\mu_{Y|x}$ is kernel embedding of marginal density $p(Y|X=x)$, \cite{song2009hilbert} showed that conditional embedding can be defined as $\mathcal{U}_{Y|X} \coloneqq \mathcal{C}_{YX} \mathcal{C}_{XX}^{-1}$ to fulfill equations \ref{req:1} and \ref{req:2}. In the following sections, we follow the definition of kernel mean embedding and embed distributions in a tensor product space to represent distribution of each group.
\section{Embedding-based Nonstationary Causal Model Inference}
In this section we introduce our proposed approach to infer the causal structure of nonstationary causal models.
\subsection{Basic Idea}
Currently, the most widely used idea of inferring causal direction is to quantify the independence between the mechanism generating the cause and the mechanism mapping the cause to the effect. One way to interpret the independence between these two mechanisms is to measure the independence between the cause and the noise. ANM and PNL lie in this field and LiNGAM methods could also be interpreted from this viewpoint \citep{hyvarinen2013pairwise}. We adopt a different interpretation which uses the independence between the marginal distribution of the cause and the conditional distribution of the effect given the cause to capture the independence between these two mechanisms and further exploit causal asymmetry. This kind of independence has also been used in many existing causal inference methods \citep{janzing2009telling, janzing2012information, chen2014causal}. We formalize this independence in postulate 1:
\begin{post}\label{post1}
The mechanism generating the cause and the mechanism mapping the cause to the effect are two independent natural processes.
\end{post}
\cite{janzing2012information} proposed this postulate and developed information geometry causal inference (IGCI). IGCI uses the density of the cause to characterize the first mechanism and the derivative of the function mapping the cause to the effect to characterize the second. In our approach, the variation of the marginal density of the cause is used to characterize the first mechanism, which is similar to IGCI. What differs from IGCI is that we use the variation of the conditional density to characterize the second mechanism. In subsequent sections, we introduce how we obtain the variation of densities and how we infer the causal direction based on the independence between them.
\subsection{Decomposition of Distributions}
Given the entire data set $\mathbf{G} = \{G_{1}, G_{2}, \dots, G_{N}\}$, which consists of $N$ groups, we make use of the variation of densities in each group. To obtain the variation, we first compute the mean of marginal densities and conditional densities of all groups as
\begin{align}
\overline{p}(X) = \frac{1}{N}\sum_{i=1}^{N} p^{(i)}(X), \quad \quad \overline{p}(Y|X) = \frac{1}{N}\sum_{i=1}^{N} p^{(i)}(Y|X), \label{1_1}
\end{align}
where $p^{(i)}(X)$ is the density of $X$ and $p^{(i)}(Y|X)$ is the conditional density of $Y$ given $X$ in group $i$. We call $\overline{p}(X)$ and $\overline{p}(Y|X)$ the base of marginal and conditional densities, respectively. Then the variation of density of each group is given by:
\begin{dfnt}[Variation of density]\label{def1}
For any $G_{i} \in \mathbf{G}$, we decompose $p^{(i)}(X)$ and $p^{(i)}(Y|X)$ into two parts: one is the base of the (conditional) density and the other is a varying part, i.e. $p^{(i)}(X) = \overline{p}(X) + \Delta p^{(i)}(X)$ and $p^{(i)}(Y|X) = \overline{p}(Y|X) + \Delta p^{(i)}(Y|X)$.
We call $\Delta p^{(i)}(X)$ and $\Delta p^{(i)}(Y|X)$ the variation of the marginal and conditional density of group $i$, respectively.
\end{dfnt}
Since the base of densities is the mean of densities of all groups, $\Delta p^{(i)}(X)$ and $\Delta p^{(i)}(Y|X)$ fulfill the following properties.
\begin{align}
\frac{1}{N}\sum_{i=1}^{N}\Delta p^{(i)}(X) \equiv 0, \quad \frac{1}{N}\sum_{i=1}^{N}\Delta p^{(i)}(Y|X) \equiv 0. \label{1_2}
\end{align}
Making use of the decomposition of distributions defined in definition \ref{def1}, we are able to analyze densities of each group with some components fixed, which finally guides us to a fixed linear causal model. We take group $i$ as an example to provides some insights before elaborating the derivations. The marginal density of effect $Y$ is given by
\begin{align}
p^{(i)}(Y)=\int{ \left(\overline{p}(x) + \Delta p^{(i)}(x) \right) \left(\overline{p}(Y|x) + \Delta p^{(i)}(Y|x)\right) dx }, \label{1_3}
\end{align}
where $\overline{p}(X)$ and $\overline{p}(Y|X)$ are the same in all groups. Therefore, we would obtain a fixed term $\int{ \overline{p}(x) \overline{p}(Y|x) dx }$ which does not change over $i$ in the expansion of equation \ref{1_3}. Although $\int{ \Delta p^{(i)}(x) \overline{p}(Y|x) dx }$ and $\int{ \overline{p}(x) \Delta p^{(i)}(Y|x) dx }$ vary over groups, they also consist of $\overline{p}(X)$ and $\overline{p}(Y|X)$ which allows us to use the invariant to formulate the relation between them into a fixed causal model. In subsequent sections, we adopt kernel embedding to transform these kinds of invariant into a linear model to infer the causal direction.
\subsection{Kernel Embedding of Distributions in Tensor Product Space}
We resort to kernel embedding to represent distributions. The marginal distributions of $X$ and $Y$ of each group are embedded in tensor product space $\cal{X}\otimes\cal{X}$ and $\cal{Y}\otimes\cal{Y}$, respectively. For simplicity, we use $\mathcal{H}$ to represent the tensor product space $\cal{X}\otimes\cal{X}$ and $\mathcal{G}$ to represent $\cal{Y}\otimes\cal{Y}$ in subsequent sections. Following the definition of kernel mean embedding, we define the mean embeddings of $X$ and $Y$ of group $i$ in $\mathcal{H}$ and $\mathcal{G}$ as:
\begin{dfnt}[tensor mean embedding]\label{def2}
\begin{align}
\mu^{(i)}_{\otimes X} \coloneqq \int{ \phi(x)\otimes\phi(x)p^{(i)}(x)dx }, \quad
\mu^{(i)}_{\otimes Y} \coloneqq \int{ \phi(y)\otimes\phi(y)p^{(i)}(y)dy }.
\label{1_4}
\end{align}
where $\phi(x)$ is the feature map of $x$ and $p^{(i)}(x)$ is the density of $x$ in group $i$. Similar notations go for $y$.
\end{dfnt}
Definition \ref{def2} is the embedding of marginal densities of each group. Since our analysis is conducted on the base and variation of density of each group, we further define the tensor mean embedding of the base and variation of densities:
\begin{dfnt}[tensor mean embedding of the base and variation of distributions]\label{def3}
\begin{align}
\overline{\mu}_{\otimes X} &\coloneqq \int{ \phi(x)\otimes\phi(x)\overline{p}(x)dx }, \\
\Delta \mu^{(i)}_{\otimes X} &\coloneqq \int{ \phi(x)\otimes\phi(x)\Delta p^{(i)}(x)dx }.
\label{1_5}
\end{align}
\end{dfnt}
$\overline{\mu}_{\otimes X}$ is the same in all groups and we have $\mu^{(i)}_{\otimes X} = \overline{\mu}_{\otimes X} + \Delta \mu^{(i)}_{\otimes X}$ from definitions \ref{def2} and \ref{def3}. Similarly, there is $\mu^{(i)}_{\otimes Y} = \overline{\mu}_{\otimes Y} + \Delta \mu^{(i)}_{\otimes Y}$. Definition \ref{def2} and \ref{def3} together state how marginal distributions are embedded in the tensor product space after decomposition. Next, we show how we make use of these tensor mean embeddings to infer the causal direction between $X$ and $Y$. To avoid analyzing probability densities directly, we substitute equation \ref{1_3} into definition \ref{def2} to conduct analysis on their embeddings:
\begin{align}
& ~\mu^{(i)}_{\otimes Y} \nonumber \\
= & \int{ \phi(y)\otimes\phi(y) \left[ \int{ (\overline{p}(y|x) + \Delta p^{(i)}(y|x) ) (\overline{p}(x) + \Delta p^{(i)}(x)) dx} \right] dy } \nonumber \\
= & \int{ \left[ \int{ \phi(y)\otimes\phi(y) \left(\overline{p}(y|x) + \Delta p^{(i)}(y|x) \right) dy} \right] \left(\overline{p}(x) + \Delta p^{(i)}(x) \right) dx } \nonumber \\
\approx & \int{ \left[ \int{ \phi(y)\otimes\phi(y) \overline{p}(y|x) dy} \right] \overline{p}(x) dx } + \int{ \left[ \int{ \phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x) dy} \right] \overline{p}(x) dx } \nonumber \\
& + \int{ \left[ \int{ \phi(y)\otimes\phi(y) \overline{p}(y|x) dy} \right] \Delta p^{(i)}(x) dx },
\label{1_6}
\end{align}
where we omit the term $\int{ \left( \int{ \phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x) dy} \right) \Delta p^{(i)}(x) dx }$. Since the ranges of variables are usually bounded and distributions usually change smoothly instead of drastically in real-world situations, we consider it reasonable to omit the one with two variation terms. Although there exits sets of densities in which the omitted term of certain group would have magnitude comparable to the sum of the remaining three terms when the distribution shifts drastically, we deem it less likely to occur in real situations. Note that this claim is close in spirit to an assumption in \cite{zhang2015discovery} in which the authors assume the nonstationarity can be written as smooth functions of time or domain index. With this claim, we have the tensor mean embedding of the base of distributions as:
\begin{align}
& ~\overline{\mu}_{\otimes Y} \nonumber \\
= & \int{ \phi(y)\otimes \phi(y) \overline{p}(y)dy } \nonumber \\
= & \int{ \phi(y)\otimes \phi(y) \left[ \frac{1}{N} \sum_{j=1}^{N}p^{(j)}(y) \right] dy } \nonumber \\
= & \int{ \phi(y)\otimes \phi(y) \left[\int \frac{1}{N} \sum_{j=1}^{N} \left(\overline{p}(y|x) + \Delta p^{(j)}(y|x) \right) \left( \overline{p}(x) + \Delta p^{(j)}(x) \right) dx\right] dy } \nonumber \\
\approx & \int{ \phi(y)\otimes \phi(y) \left[\int \frac{1}{N} \sum_{j=1}^{N} \left(\overline{p}(y|x)\overline{p}(x) + \overline{p}(y|x)\Delta p^{(j)}(x) + \Delta p^{(j)}(y|x)\overline{p}(x) \right) dx\right] dy } \nonumber \\
= &\int{ \phi(y)\otimes \phi(y) \left[ \int \overline{p}(y|x)\overline{p}(x)dx\right] dy },
\end{align}
where the approximately equal mark is again derived by omitting the one with two variation terms and the last equality is directly derived from the property shown in equation \ref{1_2}. Then we have the tensor mean embedding of the variation of distributions as:
\begin{align}
\Delta\mu^{(i)}_{\otimes Y} = \mu^{(i)}_{\otimes Y} - \overline{\mu}_{\otimes Y} \approx & \int{ \left( \int{ \phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x) dy} \right) \overline{p}(x) dx } \nonumber \\
&+ \int{ \left( \int{ \phi(y)\otimes\phi(y) \overline{p}(y|x) dy} \right) \Delta p^{(i)}(x) dx } ,
\label{1_delta}
\end{align}
which shows the relation between the tensor mean embedding of the variation of the effect and cause.
$\int{ \phi(y)\otimes\phi(y) \overline{p}(y|x) dy}$ and $\int{ \phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x) dy}$ are matrices of functions of $X$. In addition, they are both symmetric and positive definite so they admit decomposition:
\begin{align}
&\int{ \phi(y)\otimes\phi(y) \overline{p}(y|x) dy} = V(X)V^{T}(X) = \sum_{j=1}^{N_{H}} {v_{j}(X) v^{T}_{j}(X) }, \\
&\int{ \phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x) dy} = \Delta U(X) \Delta U^{T}(X) = \sum_{j=1}^{N_{H}} {\Delta u^{(i)}_{j}(X) {\Delta u^{(i)}}^{T}_{j}(X) },
\label{1_7}
\end{align}
where $V(X)$ and $\Delta U(X)$ are lower triangular matrices, $v_{j}(X)$ and $\Delta u^{(i)}_{j}(X)$ denote the $j$-th column of $V(X)$ and $\Delta U(X)$, respectively; and $N_{H}$ denotes the dimension of $V(X)$. The symbol $\Delta$ indicates the corresponding relation of $\Delta U(X)$ to the variation of densities. By assuming that $v_{j}(X)$ and $\Delta u^{(i)}_{j}(X), j = 1, \dots, N_{H}$ lie in the space of $\phi(X)$, we have $v_{j}(X) = \mathcal{A}_{j} \phi(X)$ and $\Delta u^{(i)}_{j}(X) = \Delta \mathcal{B}^{(i)}_{j} \phi(X)$. $\mathcal{A}_{j}$ and $\Delta \mathcal{B}^{(i)}_{j}$ are matrices containing coefficient mapping from $\phi(X)$ to $v_{j}(X)$ and $\Delta u^{(i)}_{j}(X)$, respectively. Then we have
\begin{align}
&\int{ \phi(y)\otimes\phi(y) \overline{p}(y|x) dy} = \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \phi(X) \otimes \phi(X) \mathcal{A}^{T}_{j}, \label{uyx} \\
&\int{ \phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x) dy} = \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \phi(X) \otimes \phi(X) {\Delta \mathcal{B}^{(i)}_{j}}^{T}. \label{duyix}
\end{align}
By substituting equation \ref{uyx} and \ref{duyix} into equation \ref{1_delta}, we further obtain
\begin{align}
\Delta \mu^{(i)}_{\otimes Y} & \approx \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \Delta \mu^{(i)}_{\otimes X} \mathcal{A}^{T}_{j} + \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \overline{\mu}_{\otimes X} {\Delta \mathcal{B}^{(i)}_{j}}^{T},
\label{1_8}
\end{align}
where $\Delta \mu^{(i)}_{\otimes X}$ and $\overline{\mu}_{\otimes X}$ are substituted in according to definition \ref{def2} and \ref{def3}.
\subsection{Inferring Causal Directions}
In this section, we discuss how we infer the causal direction using the kernel embedding of decomposed densities. Note again that we consider the case $X \to Y$ without loss of generality throughout this letter.
We start by taking normalized trace $\tau$ on both sides of equation \ref{1_8},
\begin{align}
\tau \left( \Delta\mu^{(i)}_{\otimes Y} \right) &\approx \tau \left( \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \Delta \mu^{(i)}_{\otimes X} \mathcal{A}^{T}_{j} \right) + \tau \left( \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \overline{\mu}_{\otimes X} {\Delta \mathcal{B}^{(i)}_{j}}^{T} \right) \nonumber \\
&= \tau \left( \sum_{j=1}^{N_{H}} \mathcal{A}^{T}_{j} \mathcal{A}_{j} \Delta \mu^{(i)}_{\otimes X} \right) + \tau \left( \sum_{j=1}^{N_{H}} {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j} \overline{\mu}_{\otimes X} \right) \nonumber \\
&= \tau \left( \mathcal{A} \Delta \mu^{(i)}_{\otimes X} \right) + \tau \left( {\Delta \mathcal{B}^{(i)}} \overline{\mu}_{\otimes X} \right),
\label{1_10}
\end{align}
where $\tau(A) = \tr(A) / l_{A}$ is called the normalized trace of $A$, $l_{A}$ is the size of $A$, $\mathcal{A} = \sum_{j=1}^{N_{H}} \mathcal{A}^{T}_{j} \mathcal{A}_{j}$ and $\Delta \mathcal{B}^{(i)} = \sum_{j=1}^{N_{H}} {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j}$. Since the independence of the two mechanisms in Postulate \ref{post1} is difficult to quantify, we consider to use the density of the cause and the conditional density of the effect given the cause to represent the two mechanisms and adopt the independence between the base and variation of these two densities to infer the causal direction. The independence we rely on is based on the concept of free independence \citep{voiculescu1992free, voiculescu1997free}.
\begin{dfnt}[Free independence]\label{def4}
\citep{voiculescu1992free, voiculescu1997free}. Let $\mathcal{D}$ be an algebra and $\psi: \mathcal{D} \rightarrow \mathbb{R}$ a linear functional on $\mathcal{D}$ with $\psi(1) = 1$. Then $A$ and $B$ are called free if
\begin{align}
\psi \left( p_{1}(A)q_{1}(B)p_{2}(A)q_{2}(B)\cdots \right) = 0,
\end{align}
for polynomials $p_{i}$, $q_{i}$, whenever $p_{i}(A) = q_{i}(B) = 0$.
\end{dfnt}
It is straightforward from definition \ref{def4} that if $A$ and $B$ are free independent, it holds that $\psi(AB) = \psi(A)\psi(B)$ \citep{voiculescu1992free, voiculescu1997free}. Then we have the following two assumptions to characterize the independence in postulate \ref{post1}:
\begin{assp}\label{assump1}
We assume that the tensor mean embedding of the variation of marginal density of the cause ($\Delta \mu^{(i)}_{\otimes X}, i=1, \dots, N$) and $\mathcal{A}$ is free independent, and the tensor mean embedding of the base of marginal density of the cause ($\overline{\mu}_{\otimes X}$) and $\Delta \mathcal{B}^{(i)}, i=1, \dots, N$, is free independent, that is,
\begin{align}
\tau\left( \mathcal{A} \Delta \mu^{(i)}_{\otimes X}\right) = \tau\left( \mathcal{A} \right) \tau\left( \Delta \mu^{(i)}_{\otimes X} \right), i = 1, \dots, N, \label{assp_eq1} \\
\tau\left( \Delta \mathcal{B}^{(i)} \overline{\mu}_{\otimes X} \right) = \tau\left( \Delta \mathcal{B}^{(i)} \right) \tau\left(\overline{\mu}_{\otimes X} \right), i = 1, \dots, N, \label{assp_eq2}
\end{align}
where $N$ is the number of groups.
\end{assp}
Assumption \ref{assump1} captures the independence between the mechanism generating the cause and the mechanism mapping the cause to the effect. In equation \ref{assp_eq1}, $\mathcal{A}$ depends only on the base of the conditional densities $\overline{p}(Y|X)$ which corresponds to the second mechanism, and $\Delta \mu^{(i)}_{\otimes X}$ depends only on the variation of the marginal densities of the cause $\Delta p^{(i)}(X)$, which corresponds to the first mechanism. Therefore, the free independence between them characterizes the independence in postulate 1. Similarly, we have assumptions shown in equation \ref{assp_eq3}.
\begin{assp}\label{assump2}
Regarding the normalized trace of the tensor mean embedding of variation of marginal densities of the cause in each group as a realization of a random variable $\tau_{\Delta \mu_{\otimes X}}$ and each $\tau \left(\Delta \mathcal{B}^{(i)} \right)$ as a realization of another random variable $\tau_{\Delta \mathcal{B}}$, we assume that these two random variables are independent, i.e.
\begin{align}
\tau_{\Delta \mu_{\otimes X}} \perp\!\!\!\perp \tau_{\Delta \mathcal{B}}.
\label{assp_eq3}
\end{align}
\end{assp}
Assumption \ref{assump2} is also motivated by the independence in postulate \ref{post1}. Specifically, $\tau_{\Delta \mu_{\otimes X}}$ captures the information of the variation of marginal densities of the cause, and $\tau_{\Delta \mathcal{B}}$ captures the information of the variation of conditional densities. We interpret postulate \ref{post1} as the independence between the marginal and conditional. Therefore, this independence between their variations of densities (approximately) holds. With assumption \ref{assump1}, equation \ref{1_10} becomes
\begin{align}
\tau\left(\Delta\mu^{(i)}_{\otimes Y}\right) \approx \tau\left( \mathcal{A} \right) \tau\left( \Delta \mu^{(i)}_{\otimes X} \right) + \tau\left( \Delta \mathcal{B}^{(i)} \right) \tau\left(\overline{\mu}_{\otimes X} \right).
\label{1_11}
\end{align}
Since $\overline{p}(x)$ and $\overline{p}(y|x)$ are fixed given $\mathbf{G}$, $\tau\left(\overline{\mu}_{\otimes X}\right)$ and $\tau\left(\mathcal{A}\right)$ are the same in all groups. We introduce the following notations for simplicity:
\noindent
\textbf{Notation 1.} For any $G_{i} \in \mathbf{G}$, we use $\tau_{x}^{(i)}$ and $\tau_{y}^{(i)}$
to represent $\tau\left(\Delta\mu^{(i)}_{\otimes X}\right)$ and $\tau\left(\Delta\mu^{(i)}_{\otimes Y}\right)$, respectively. $\epsilon_{x\rightarrow y}^{(i)}$ denotes $\tau\left(\Delta \mathcal{B}^{(i)} \overline{\mu}_{\otimes X} \right)$, which is the corresponding noise term. $c_{y|x}$ denotes $\tau\left(\mathcal{A}\right)$. We view each $\tau_{x}^{(i)}$ as a realization of a random variable $\tau_{x}$. Similarly, we have $\tau_{y}$ and $\epsilon_{x\rightarrow y}$.
\begin{prop}\label{prop1}
If the causal direction is $X\to Y$ and assumptions \ref{assump1} and \ref{assump2} hold, the normalized trace of the tensor mean embeddings of the variation of the densities of the cause $\left(\tau_{x}\right)$ and the effect $\left(\tau_{y}\right)$ fulfill the following linear nongaussian additive model \citep{kano2003causal}:
\begin{align}
\tau_{y} &\approx c_{y|x} \tau_{x}+ \epsilon_{x\rightarrow y}.
\label{lnr}
\end{align}
\end{prop}
\begin{proof}
By adopting notations in notation 1, equation \ref{1_11} becomes
\begin{align}
\tau_{y}^{(i)} \approx c_{y|x} \tau_{x}^{(i)}+ \epsilon_{x\rightarrow y}^{(i)},\quad i = 1, \dots, N.
\end{align}
We first show that $\epsilon_{x\rightarrow y}$ follows nongaussian distributions. According to assumption 1, we have
\begin{align}
\epsilon_{x\rightarrow y}^{(i)} = \tau\left( \Delta \mathcal{B}^{(i)} \right) \tau\left(\overline{\mu}_{\otimes X} \right),
\end{align}
where $\tau\left(\overline{\mu}_{\otimes X} \right)$ is fixed and thus can be viewed as a constant. From the definition of $\Delta \mathcal{B}^{(i)}$ we have
\begin{align}
\tau\left( \Delta \mathcal{B}^{(i)} \right) = \frac{1}{N_{H}} \tr\left( \sum_{j=1}^{N_{H}} {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j} \right) = \frac{1}{N_{H}} \sum_{j=1}^{N_{H}} \tr\left( {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j} \right).
\end{align}
Since $\tr\left( {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j} \right)$ are positive for all $j$, we have $\tau\left( \Delta \mathcal{B}^{(i)} \right) > 0$. Therefore, the distribution of $\epsilon_{x\rightarrow y}^{(i)}$ is not symmetric and is thus not Gaussian distributed.
Second, we have $\tau_{x}$ is independent of $\epsilon_{x\rightarrow y}$ according to the independence between $\tau_{\Delta\mu_{\otimes X}}$ and $\tau_{\Delta \mathcal{B}}$ in assumption \ref{assump2}. Then we conclude equation \ref{lnr} forms a linear non-Gaussian additive model.
\end{proof}
According to the identifiability of LiNGAM~\citep{kano2003causal,shimizu-et-al:lingam}, $\tau_{y}$ and $\epsilon_{y\rightarrow x}$ are dependent. By exploiting the cause-effect asymmetry that the cause is independent of the noise only in the causal direction, we propose the following causal inference approach: embedding-based nonstationary causal model inference (ENCI).
\noindent
\textbf{Causal Inference Approach (ENCI)}: Given data set $\mathbf{G}$, we compute $\tau^{(i)}_{x}$ and $\tau^{(i)}_{y}$ for $i = 1, \dots, N$ and conclude that $X\to Y$ if $\tau_{x} \perp\!\!\!\perp \epsilon_{x\rightarrow y}$, otherwise $Y \to X$ if $\tau_{y} \perp\!\!\!\perp \epsilon_{y\rightarrow x}$.
Hilbert Schimidt Independence Criterion (HSIC)~\citep{gretton2007kernel} is applied to measure the independence between the regressor and its corresponding noise on both hypothetical directions, and we favor the direction with less dependence in practice. The ENCI algorithm is given in algorithm~\ref{alg1}.
\begin{algorithm}[ht]
\caption{ENCI for cause-effect pairs}
\label{alg1}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{algorithmic}[1]
\REQUIRE $N$ data groups $\mathbf{G} = \{G_{1}, G_{2}, \dots, G_{N}\}$
\ENSURE The causal direction
\STATE Normalize $X$ and $Y$ in each group;
\STATE Compute $\tau_{x}^{(i)}$ and $\tau_{y}^{(i)}$ for $i=1, \dots, N$;
\STATE Compute residual $\epsilon_{x\rightarrow y}$ and $\epsilon_{y\rightarrow x}$ by conducting least square regressions;
\STATE Apply HSIC on $\tau_{x}$ and $\epsilon_{x\rightarrow y}$, denote the quotient of testStat and thresh returned by HSIC by $r_{x\rightarrow y}$. Similarly we have $r_{y\rightarrow x}$.
\IF{$ r_{x\rightarrow y} < r_{y\rightarrow x}$}
\STATE The causal direction is $x \rightarrow y$;
\ELSIF{$ r_{x\rightarrow y} > r_{y\rightarrow x}$}
\STATE The causal direction is $y \rightarrow x$;
\ELSE
\STATE No decision made.
\ENDIF
\end{algorithmic}
\end{algorithm}
\subsection{Empirical Estimations}
In this section, we show how to estimate $\tau_{x}^{(i)}$ and $\tau_{y}^{(i)}$ for $i=1, \dots, N$ based on the observations.
Let $\mathbf{\Phi}^{(i)} = \left[ \phi(x_{1}^{(i)}), \dots, \phi(x_{N_{i}}^{(i)}) \right]$ and $\mathbf{\Gamma}^{(i)} = \left[ \gamma(x_{1}^{(i)}), \dots, \gamma(x_{N_{i}}^{(i)}) \right]$ be the feature matrices of $X$ and $Y$ in group $i$, respectively, given observations in $\mathbf{G}$. We estimate the mean embedding of $p^{(i)}(X)$ in $\mathcal{X}\otimes \mathcal{X}$ as
\begin{align}
\hat{\mu}_{\otimes X}^{(i)} = \frac{1}{N_{i}} \Phi^{(i)} H \left( \Phi^{(i)} H\right)^{T},
\end{align}
where $N_{i}$ is the number of observations in $i$th group, $H = I - \frac{1}{N_{i}} \mathbf{1} \mathbf{1}^{T}$ and $\mathbf{1}$ is a column vector of all 1s. Since we have
\begin{align}
\hat{\overline{\mu}}_{\otimes X} & = \int{ \phi(x)\otimes \phi(x) \hat{\overline{p}}(x)dx } \nonumber \\
& = \int{ \phi(x)\otimes \phi(x) \left[ \frac{1}{N} \sum_{j=1}^{N}\hat{p}^{(j)}(x) \right] dx } \nonumber \\
& = \frac{1}{N} \sum_{j=1}^{N}{ \int{ \phi(x)\otimes \phi(x) \hat{p}^{(j)}(x) dx } } \nonumber \\
& = \frac{1}{N} \sum_{j=1}^{N} {\hat{\mu}_{\otimes X}^{(j)}},
\end{align}
for estimating the tensor mean embedding of the base of distributions $\overline{\mu}_{\otimes X}$, the tensor mean embedding of the variation of distributions $\Delta \mu_{\otimes X}^{(i)}$ is estimated as
\begin{align}
\Delta \hat{\mu}_{\otimes X}^{(i)} = \hat{\mu}_{\otimes X}^{(i)} - \hat{\overline{\mu}}_{\otimes X} = \hat{\mu}_{\otimes X}^{(i)} - \frac{1}{N} \sum_{j=1}^{N} {\hat{\mu}_{\otimes X}^{(j)}}.
\label{est_delta}
\end{align}
By taking the normalized trace on both sides of equation \ref{est_delta}, we have
\begin{align}
\tau_{x}^{(i)} & = \tau\left( \hat{\mu}_{\otimes X}^{(i)} \right) - \tau\left( \frac{1}{N} \sum_{i=1}^{N} {\hat{\mu}_{\otimes X}^{(i)}} \right) \nonumber \\
& \approx \tau\left( \frac{1}{N_{i}} \mathbf{\Phi}^{(i)} H \left( \mathbf{\Phi}^{(i)} H\right)^{T} \right) - \frac{1}{N} \sum_{j=1}^{N} \tau\left( \frac{1}{N_{j}} \mathbf{\Phi}^{(j)} H \left( \mathbf{\Phi}^{(j)} H\right)^{T} \right) \nonumber \\
& = \frac{1}{N_{i}^{2}} \tr\left( K^{(i)}_{x} H \right) - \frac{1}{N} \sum_{j=1}^{N} \left[ \frac{1}{N_{j}^{2}} \tr\left( K^{(j)}_{x} H \right) \right],
\end{align}
where $N$ is the total number of groups, $N_{i}$ is the number of observations in $i$th group and $K^{(i)}_{x} = \left(\mathbf{\Phi}^{(i)}\right)^{T}\mathbf{\Phi}^{(i)}$ is the kernel matrix of $X$ in $i$th group. Similarly, we have
\begin{align}
\tau_{y}^{(i)} = = \frac{1}{N_{i}^{2}} \tr\left( K^{(i)}_{y} H \right) - \frac{1}{N} \sum_{j=1}^{N} \left[ \frac{1}{N_{j}^{2}} \tr\left( K^{(j)}_{y} H \right) \right],
\end{align}
where $K^{(i)}_{y} = \left(\mathbf{\Gamma}^{(i)}\right)^{T}\mathbf{\Gamma}^{(i)}$ is the kernel matrix of $X$ in $i$th group. We can see that both $\tau_{x}^{(i)}$ and $\tau_{y}^{(i)}$ can be easily calculated from Gram matrix using kernel methods.
\section{Extending ENCI to Causal Graph Discovery}
In this section, we extend ENCI to causal discovery for two kinds of directed acyclic graphs (DAGs). One is a tree-structured graph in which each node has at most one parent node. The other is multiple-independent-parent graph in which parent nodes of each node are mutually independent. Examples of these two kinds of DAGs are shown in Figure~\ref{fig:exmp}.
\begin{figure}[!ht]
\begin{subfigure}{.5\textwidth}
\centering
\begin{tikzpicture}[scale=1.0,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, below=of x1, xshift=-1.5cm] (x2) {$x_{2}$} ; %
\node[latent, below=of x1, xshift=-0.5cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=0.5cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x1, xshift=1.5cm] (x8) {$x_{8}$} ; %
\edge {x1}{x2, x3, x5, x8} ; %
\node[latent, below=of x2] (x7) {$x_{7}$} ; %
\node[latent, below=of x3] (x4) {$x_{4}$} ; %
\node[latent, below=of x5] (x9) {$x_{9}$} ; %
\node[latent, below=of x8] (x10) {$x_{10}$} ; %
\edge {x2}{x7} ; %
\edge {x3}{x4} ; %
\edge {x5}{x9, x10} ; %
\node[latent, below=of x4] (x6) {$x_{6}$} ; %
\edge {x4}{x6} ; %
\end{tikzpicture}
\caption{}
\label{fig:exmp_tree}
\end{subfigure}
\begin{subfigure}{.5\textwidth}
\centering
\begin{tikzpicture}[scale=1.0,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, right=of x1] (x2) {$x_{2}$} ; %
\node[latent, right=of x2] (x3) {$x_{3}$} ; %
\node[latent, below=of x2, xshift=-0.75cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x2, xshift=0.75cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x4, xshift=0.75cm] (x6) {$x_{6}$} ; %
\edge{x1}{x4};
\edge{x2, x3}{x5};
\edge{x4, x5}{x6};
\end{tikzpicture}
\caption{}
\label{fig:exmp_mtp}
\end{subfigure}
\caption{Examples of (a) Tree-structured graph (b) Multiple-independent-parent graph.}
\label{fig:exmp}
\end{figure}
\subsection{Describing Causal Relationship by Directed Acyclic Graphs}
Consider a finite set of random variables $\mathbf{X} = (X_{1}, \dots, X_{p})$ with index set $\mathbf{V} \coloneqq \{1, \dots, p\}$. A graph $\mathcal{G}=(\mathbf{V}, \mathbf{E})$ consists of nodes in $\mathbf{V}$ and edges $(m, n)$ in $\mathbf{E}$ for any $m, n \in \mathbf{V}$. Then we introduce graph terminologies required for subsequent sections. Most of the definitions are from \citep{spirtes2000causation}.
Edge $(m, n)$ is a directed link from node $m$ to node $n$. Node $m$ is called a parent of $n$, and $n$ is called a child of $m$ if $(m,n)\in \mathbf{E}$. The parent set of $n$ is denoted by $pa(n)$ and its child set by $ch(n)$. Nodes $m$, $n$ are called adjacent if either $(m, n)\in \mathbf{E}$ or $(n, m)\in \mathbf{E}$. A path in $\mathcal{G}$ is a sequence of distinct vertices $m_{1}, \dots, n_{q}$ such that $m_{k}$ and $n_{k+1}$ are adjacent for all $k=1, \dots, q-1$. If $(m_{k}, m_{k+1}) \in \mathbf{E}$ for all $k$, the path is also called a directed path from $m_{1}$ to $m_{q}$. $\mathcal{G}$ is called a partially directed acyclic graph (PDAG) if there is no directed cycle, i.e., there is no pair $(m, n)$ such that there are directed paths from $m$ to $n$ and from $n$ to $m$. $\mathcal{G}$ is called a directed acyclic graph (DAG) if it is a PDAG and all edges are directed.
General causal graph discovery is very challenging, especially when the relation between a variable pair is a complicated nonlinear stochastic process. In the following section, we show how we discover the causal structure tree-structured graphs (TSG) and multiple-independent-parents graph (MIPG). Note that the causal relation between a variable and its parent node in our model not only could be complicated nonlinear functions but also varies in different groups.
\subsection{Tree-Structured Causal Graph Discovery}
In a TSG $\mathcal{G}$ with $p$ nodes, each variable $X_{m}$ and its only parent node $pa(X_m)$ fulfill the linear relation in Equation \ref{lnr}. Thus we have the following proposition for TSG:
\begin{prop}
In a TSG $\mathcal{G}$ where each variable $X_{m}$ has only one parent node, the normalized traces of the tensor mean embedding of the variation of densities of all variables $\left(\tau_{x_{m}},~ m = 1, \dots, p \right)$ fulfill a linear nongaussian acyclic model (LiNGAM) \citep{shimizu-et-al:lingam} if assumption \ref{assump1} and \ref{assump2} hold:
\begin{align}
\bm{\tau}_{x}&\approx \bm{C} \bm{\tau}_{x} + \bm{\epsilon},
\end{align}
where $\bm{\tau}_{x} = \left[\tau_{x_{1}}, \dots, \tau_{x_{p}} \right]^{T}$, coefficient matrix $\bm{C}$ whose element on $n$-th row and $m$-th column equals to $c_{x_{n}|x_{m}}$ could be permuted to a lower triangular matrix and $\bm{\epsilon} = \left[\epsilon_{pa(x_{1})\to x_{1}}, \dots, \epsilon_{pa(x_{p})\to x_{p}} \right]^{T}$ collects all noise terms $\epsilon_{pa(x_{m})\to x_{m}},~ m = 1, \dots, p$.
\end{prop}
\begin{proof}
First, $\tau_{x_{m}}$, where $m = 1, \dots, p$, could be arranged in a causal order in which no later variable is the cause of earlier ones due to the acyclicity of the graph. Note that causal order in subsequent sections also means that this condition holds for a sequence of variables. Second, the noise term $\epsilon_{pa(x_{m})\rightarrow x_{m}}$, where $m = 1, \dots, p$, follows nongaussian distributions as shown in proposition \ref{prop1}. Thirdly, assumption \ref{assump2} ensures that $\tau_{x_{m}} \perp\!\!\!\perp \epsilon_{pa(x_{m})\rightarrow x_{m}}$ for $m = 1, \dots, p$.
\end{proof}
Therefore, the graph formed by $\tau_{x_{m}}$, where $m = 1, \dots, p$, fulfills the structure of LiNGAM~\citep{shimizu-et-al:lingam} so we can apply LiNGAM on $\tau_{x}$ to infer the causal structure of the causal graph consists of $X_{1}, \dots, X_{p}$.
\subsection{Multiple-Independent-Parent Graph Discovery}
We extend ENCI to cases where each node could have more than one parent node provided that all its parent nodes are mutually independent.
Suppose a variable $Y$ in graph $\mathcal{G}$ has $q$ independent parent nodes - $X_{1}, \dots, X_{q}$. The marginal density of $Y$ in group $i$ can be obtained from
\begin{align}
p^{(i)}(Y) = \int{ p^{(i)}(Y|x_{1}, \dots, x_{q})p^{(i)}(x_{1}, \dots, x_{q})dx_{1}\cdots dx_{q} }.
\end{align}
Then by substituting $p^{(i)}(Y)$ into $\mu^{(i)}_{\otimes Y}$ with $p^{(i)}(Y|X_{1}, \dots, X_{q})$ decomposed as $p^{(i)}(Y|X_{1}, \dots, X_{q}) = \overline{p}(Y|X_{1}, \dots, X_{q}) + \Delta p^{(i)}(Y|X_{1}, \dots, X_{q})$ and integrating with respect to $Y$, we have
\begin{align}
\mu^{(i)}_{\otimes Y}=& \int \phi(y)\otimes\phi(y) \left( \int{ \overline{p}(y|x_{1}, \dots, x_{q})p^{(i)}(x_{1}, \dots, x_{q})dx_{1}\cdots dx_{q} } + \nonumber \right. \\
& \left. \int{\Delta p^{(i)}(y|x_{1}, \dots, x_{q})p^{(i)}(x_{1}, \dots, x_{q})dx_{1}\cdots dx_{q}}\right) dy \nonumber \\
= & \int{ \left( \int{\phi(y)\otimes\phi(y) \overline{p}(y|x_{1}, \dots, x_{q}) dy} \right) p^{(i)}(x_{1}, \dots, x_{q}) dx_{1}\cdots dx_{q} } + \nonumber \\
& \int{ \left( \int{\phi(y)\otimes\phi(y) \Delta p^{(i)}(y|x_{1}, \dots, x_{q}) dy} \right) p^{(i)}(x_{1}, \dots, x_{q}) dx_{1}\cdots dx_{q} }.
\label{mlt_1}
\end{align}
Following the same idea in the previous section, we conduct decomposition on both $ \int{\phi(y)\otimes\phi(y) \overline{p}(y|X_{1}, \dots, X_{q}) dy}$ and $\int{\phi(y)\otimes\phi(y) \Delta p^{(i)}(y|X_{1}, \dots, X_{q}) dy}$ and thus obtain
\begin{align}
&\int{\phi(y)\otimes\phi(y) \overline{p}(y|X_{1}, \dots, X_{q}) dy} = \sum_{j=1}^{N_{H}} {v_{j}(X_{1}, \dots, X_{q}) v^{T}_{j}(X_{1}, \dots, X_{q}) }, \\
&\int{\phi(y)\otimes\phi(y) \Delta p^{(i)}(y|X_{1}, \dots, X_{q}) dy} = \sum_{j=1}^{N_{H}} {\Delta u^{(i)}_{j}(X_{1}, \dots, X_{q}) \Delta{u^{(i)}}^{T}_{j}(X_{1}, \dots, X_{q}) },
\end{align}
where $v_{j}(X_{1}, \dots, X_{q})$ denotes the $j$-th column of $\int{\phi(y)\otimes\phi(y) \overline{p}(y|X_{1}, \dots, X_{q}) dy}$ and $\Delta u^{(i)}_{j}(X_{1}, \dots, X_{q})$ denotes the $j$-th column of $\int{\phi(y)\otimes\phi(y) \Delta p^{(i)}(y|X_{1}, \dots, X_{q}) dy}$. By assuming that both $v_{j}(X_{1}, \dots, X_{q})$ and $\Delta u^{(i)}_{j}(X_{1}, \dots, X_{q}), j = 1, \dots, N_{H}$ lie in the space of feature map $\phi(X_{1}, \dots, X_{q})$, we have $v_{j}(X_{1}, \dots, X_{q}) = \mathcal{A}_{j} \phi(X_{1}, \dots, X_{q})$ and $\Delta u^{(i)}_{j}(X_{1}, \dots, X_{q}) = \Delta \mathcal{B}^{(i)}_{j} \phi(X_{1}, \dots, X_{q})$. Then they become
\begin{small}
\begin{align}
&\int{\phi(y)\otimes\phi(y) \overline{p}(y|X_{1}, \dots, X_{q}) dy} = \sum_{j=1}^{N_{H}} { \mathcal{A}_{j} \phi(X_{1}, \dots, X_{q}) \otimes \phi(X_{1}, \dots, X_{q}) \mathcal{A}^{T}_{j}}, \label{sum_g_1} \\
&\int{\phi(y)\otimes\phi(y) \Delta p^{(i)}(y|X_{1}, \dots, X_{q}) dy} = \sum_{j=1}^{N_{H}} { \Delta \mathcal{B}^{(i)}_{j} \phi(X_{1}, \dots, X_{q}) \otimes \phi(X_{1}, \dots, X_{q}) {\Delta \mathcal{B}^{(i)}_{j}}^{T} }. \label{sum_g_2}
\end{align}
\end{small}
By plugging in equations \ref{sum_g_1} and \ref{sum_g_2}, equation \ref{mlt_1} becomes
\begin{align}
& \mu^{(i)}_{\otimes Y} \nonumber \\
= & \int{ \left( \sum_{j=1}^{N_{H}} { \mathcal{A}_{j} \phi(x_{1}, \dots, x_{q}) \otimes \phi(x_{1}, \dots, x_{q}) \mathcal{A}^{T}_{j}} \right) p^{(i)}(x_{1}, \dots, x_{q})dx_{1}\cdots dx_{q} } \nonumber \\
& + \int{ \left( \sum_{j=1}^{N_{H}} { \Delta \mathcal{B}^{(i)}_{j} \phi(x_{1}, \dots, x_{q}) \otimes \phi(x_{1}, \dots, x_{q}) {\Delta \mathcal{B}^{(i)}_{j}}^{T} } \right) p^{(i)}(x_{1}, \dots, x_{q})dx_{1}\cdots dx_{q} } \nonumber \\
= & \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \left[ \int{ \phi(x_{1}, \dots, x_{q}) \otimes \phi(x_{1}, \dots, x_{q}) p^{(i)}(x_{1}, \dots, x_{q})dx_{1}\cdots dx_{q} } \right] \mathcal{A}^{T}_{j} \nonumber \\
& + \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \left[ \int{ \phi(x_{1}, \dots, x_{q}) \otimes \phi(x_{1}, \dots, x_{q}) p^{(i)}(x_{1}, \dots, x_{q}) dx_{1}\cdots dx_{q} } \right] {\Delta \mathcal{B}^{(i)}_{j}}^{T}.
\label{mu_y_eq_3}
\end{align}
Observing that there exists a common term of integration in each term of the summation in equation \ref{mu_y_sumi}, we now analyze this integral term in square brackets. Due to mutual independence among variables $X_{k}$ for $k=1,\dots,q$, $p^{(i)}(X_{1}, \dots, X_{q})$ admits the following factorization:
\begin{align}
p^{(i)}(X_{1}, \dots, X_{q})=p^{(i)}(X_{1}) \cdots p^{(i)}(X_{q}).
\end{align}
Then we adopt Bochner's theorem~\citep{rudin2011fourier} in analyzing $\phi(X_{1}, \dots, X_{q})$. Bochner's theorem states that a continuous shift-invariant kernel $K(x,y)=k(x-y)$ is a positive-definite function if and only if $k(t)$ is the Fourier transform of a nonnegative measure $\rho(\omega)$. Let $\alpha=\int{d\rho(\omega)}$, $p_{\omega}=\rho/\alpha$, and $\omega_{1},\omega_{2}, \dots, \omega_{k}$ be independent samples from $p_{\omega}$. Then the random projection vector $\phi(X)$ can be
\begin{align}
\phi(X) = \frac{\alpha}{\sqrt[]{k}}\left[ e^{-i\omega_{1}^TX}, \dots, e^{-i\omega_{k}^TX} \right].
\end{align}
Similarly, we have
\begin{align}
\phi(X_{1}, \dots, X_{n})=\frac{\alpha}{\sqrt[]{k}}\left[ e^{-i\left(\omega_{11}^TX_{1}+\cdots + \omega_{1n}^TX_{n}\right)}, \dots, e^{-i\left(\omega_{k1}^TX_{1}+\cdots+ \omega_{kn}^TX_{n}\right)} \right],
\end{align}
which leads to
\begin{align}
\phi(X_{1}, \dots, X_{q})&=\phi(X_{1})\circ\dots\circ\phi(X_{q}),
\end{align}
where $\phi(X_{j})\circ\phi(X_{k})$ denotes the element-wise product. Since
\begin{align}
&\left(\phi(X_{1})\circ\dots\circ\phi(X_{q})\right)\otimes\left(\phi(X_{1})\circ\dots\circ\phi(X_{q})\right) \nonumber \\
= &\left(\phi(X_{1})\otimes\phi(X_{1})\right)\circ\dots\circ\left(\phi(X_{q})\otimes\phi(X_{q})\right),
\end{align}
the integration in equation \ref{mu_y_eq_3} becomes
\begin{align}
&\int{ \phi(x_{1}, \dots, x_{q}) \otimes \phi(x_{1}, \dots, x_{q}) p^{(i)}(x_{1}, \dots, x_{q}) dx_{1}\cdots dx_{q} } \nonumber \\
= & \int{ \left(\phi(x_{1})\otimes\phi(x_{1})\right)\circ\dots\circ\left(\phi(x_{q})\otimes\phi(x_{q})\right) ~ p^{(i)}(x_{1})\cdots p^{(i)}(x_{q}) ~ dx_{1}\cdots dx_{q}} \nonumber \\
= & \int{ \phi(x_{1})\otimes\phi(x_{1}) \left(p(x_{1}) + \Delta p^{(i)}(x_{1})\right)dx_{1}} \circ \dots \nonumber \\
& \dots \circ \int{ \phi(x_{q})\otimes\phi(x_{q}) \left(p(x_{q}) + \Delta p^{(i)}(x_{q})\right) dx_{q}} \nonumber \\
= & \left(\overline{\mu}_{\otimes X_{1}} + \Delta\mu_{\otimes X_{1}}^{(i)}\right) \circ\dots\circ\left(\overline{\mu}_{\otimes X_{q}} + \Delta\mu_{\otimes X_{q}}^{(i)}\right).
\label{mu_y_sumi}
\end{align}
By substituting equation \ref{mu_y_sumi} into equation \ref{mu_y_eq_3} we have
\begin{align}
\mu^{(i)}_{\otimes Y}&= \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \left[ \left(\overline{\mu}_{\otimes X_{1}} + \Delta\mu_{\otimes X_{1}}^{(i)}\right) \circ\dots\circ\left(\overline{\mu}_{\otimes X_{q}} + \Delta\mu_{\otimes X_{q}}^{(i)}\right) \right] \mathcal{A}^{T}_{j} \nonumber \\
& \quad + \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \left[ \left(\overline{\mu}_{\otimes X_{1}} + \Delta\mu_{\otimes X_{1}}^{(i)}\right) \circ\dots\circ\left(\overline{\mu}_{\otimes X_{q}} + \Delta\mu_{\otimes X_{q}}^{(i)}\right) \right] {\Delta \mathcal{B}^{(i)}_{j}}^{T} \nonumber \\
&\approx \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \left[\left( \overline{\mu}_{\otimes X_{1}}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right) + \left(\Delta \mu_{\otimes X_{1}}^{(i)}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right) + \cdots \right. \nonumber \\
& \quad + \left. \left(\overline{\mu}_{\otimes X_{1}}\circ\dots\circ \Delta\mu_{\otimes X_{q}}^{(i)} \right) \right] \mathcal{A}^{T}_{j} + \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \left[ \overline{\mu}_{\otimes X_{1}}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right] {\Delta \mathcal{B}^{(i)}_{j}}^{T},
\end{align}
where we omit terms with more than one tensor mean embedding of variation of densities . Following the same idea in equation \ref{1_delta}, we compute the variation of tensor embedding of $Y$ by
\begin{align}
\Delta \mu^{(i)}_{\otimes Y} &\approx \mu^{(i)}_{\otimes Y} - \overline{\mu}_{\otimes Y} \nonumber \\
&= \sum_{j=1}^{N_{H}} \mathcal{A}_{j} \left[\left(\Delta \mu_{\otimes X_{1}}^{(i)}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right) + \cdots + \left(\overline{\mu}_{\otimes X_{1}}\circ\dots\circ \Delta\mu_{\otimes X_{q}}^{(i)} \right) \right] \mathcal{A}^{T}_{j} \nonumber \\
& \quad + \sum_{j=1}^{N_{H}} \Delta \mathcal{B}^{(i)}_{j} \left[ \overline{\mu}_{\otimes X_{1}}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right] {\Delta \mathcal{B}^{(i)}_{j}}^{T}.
\label{bf_tau}
\end{align}
Then by taking normalized trace on both sides of equation \ref{bf_tau} we have
\begin{align}
\tau \left( \Delta \mu^{(i)}_{\otimes Y} \right) &\approx \tau \left( \sum_{j=1}^{N_{H}} \mathcal{A}^{T}_{j} \mathcal{A}_{j} \left[\left(\Delta \mu_{\otimes X_{1}}^{(i)}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right) + \cdots + \left(\overline{\mu}_{\otimes X_{1}}\circ\dots\circ \Delta\mu_{\otimes X_{q}}^{(i)} \right) \right] \right) \nonumber \\
& \quad + \tau \left( \sum_{j=1}^{N_{H}} {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j} \left[ \overline{\mu}_{\otimes X_{1}}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right] \right) \nonumber \\
&= \tau \left( \mathcal{A} \left[\left(\Delta \mu_{\otimes X_{1}}^{(i)}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right) + \cdots + \left(\overline{\mu}_{\otimes X_{1}}\circ\dots\circ \Delta\mu_{\otimes X_{q}}^{(i)} \right) \right] \right) \nonumber \\
& \quad + \tau \left( \Delta \mathcal{B}^{(i)} \left[ \overline{\mu}_{\otimes X_{1}}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right] \right) \nonumber \\
&= \tau\left(\mathcal{A}\right)\tau\left(\Delta \mu_{\otimes X_{1}}^{(i)}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right)+\dots +\tau\left(\mathcal{A}\right)\tau\left(\overline{\mu}_{\otimes X_{1}}\circ\dots\circ \Delta\mu_{\otimes X_{q}}^{(i)}\right) \nonumber \\
&\quad +\tau\left(\Delta \mathcal{B}^{(i)}\right)\tau\left(\overline{\mu}_{\otimes X_{1}}\circ\dots\circ \overline{\mu}_{\otimes X_{q}} \right),
\label{tr_2}
\end{align}
where $\mathcal{A} = \sum_{j=1}^{N_{H}} \mathcal{A}^{T}_{j} \mathcal{A}_{j}$ and $\Delta \mathcal{B}^{(i)} = \sum_{j=1}^{N_{H}} {\Delta \mathcal{B}^{(i)}_{j}}^{T} \Delta \mathcal{B}^{(i)}_{j}$. The last equality derives directly from assumption \ref{assump2}. Now we introduce another assumption for further analysis of MIPG.
\begin{lema}\label{lema}
Two high dimensional square matrices (e.g. $A$, $B$) whose elements are generated independently from two random variables fulfill the following property
\begin{align}
\tau (A \circ B) \approx \tau(A) \tau(B).
\label{lema1}
\end{align}
\end{lema}
\begin{proof}
Firstly, the elements of $A$ and $B$ can be viewed as realizations of two underlying random variables; we denote them by $X_{A}$ and $X_{B}$, respectively. The left hand side of Equation \ref{lema1} becomes
\begin{align}
\tau (A \circ B) = \frac{1}{l_{A}} \tr(A \circ B) = \frac{1}{l_{A}} \sum_{j=1}^{l_{A}}A_{jj}B_{jj} \approx \mathbb{E}\left[ X_{A}X_{B} \right],
\end{align}
where $l_{A}$ is the size of $A$ and $A_{jj}$ denotes $A$'s element on $j$th row and $j$th column. Similarly, we have $B_{jj}$. Then the right hand side of equation \ref{lema1} becomes
\begin{align}
\tau(A) \tau(B) = \left( \frac{1}{l_{A}} \sum_{j=1}^{l_{A}}A_{jj} \right) \left( \frac{1}{l_{B}} \sum_{j=1}^{l_{B}}B_{jj} \right) \approx \mathbb{E}\left[ X_{A} \right] \mathbb{E}\left[ X_{B} \right],
\end{align}
where $l_{A}$ is the size of $A$ and $l_{A} = l_{B}$. Finally, by adopting the independence between $X_{A}$ and $X_{B}$, we complete the proof.
\end{proof}
Based on Lemma \ref{lema}, we make following assumption for MIPG.
\begin{assp}\label{assump3}
We assume that the elements of tensor mean embedding of the variation of density of each parent node (e.g. $\Delta \mu_{\otimes X_{k}}^{(i)}$k = 1, \dots, q) of certain variable and that of the base of densities of other parent nodes (e.g. $\overline{\mu}_{\otimes X_{l}}, l \neq k$) are generated independently.
\end{assp}
A basic example implied by assumption \ref{assump3} is $\tau (\Delta \mu_{\otimes X_{k}}^{(i)} \circ \overline{\mu}_{\otimes X_{l}}) = \tau(\Delta \mu_{\otimes X_{k}}^{(i)}) \tau(\overline{\mu}_{\otimes X_{l}})$. $\Delta \mu_{\otimes X_{k}}^{(i)}$ depends only on $\Delta p^{(i)}(X_{k})$ and $\overline{\mu}_{\otimes X_{l}}$ depends only on $\overline{p}(X_{l})$. Based on the mutual independence among parent nodes of variables in MIPG, assumption \ref{assump3} further states that the tensor mean embedding of the variation of the density of a parent node is independent of that of the base of the density of another parent node. This can be easily extended to cases with more than two terms provided that the independence holds. Under assumption \ref{assump3}, equation \ref{tr_2} becomes
\begin{align}
\tau\left(\Delta\mu^{(i)}_{\otimes Y}\right)\approx &~\tau\left(\mathcal{A}\right)\tau\left(\Delta \mu_{\otimes X_{1}}^{(i)}\right)\tau\left(\mu_{\otimes X_{2}}\circ\dots\circ \mu_{\otimes X_{q}} \right)+\dots \nonumber \\
&\dots +\tau\left(\mathcal{A}\right)\tau\left(\mu_{\otimes X_{1}}\circ\dots\circ \mu_{\otimes X_{q-1}}\right)\tau\left(\Delta\mu_{\otimes X_{q}}^{(i)} \right) \nonumber \\
&+\tau\left(\Delta \mathcal{B}^{(i)}\right)\tau\left(\mu_{\otimes X_{1}}\circ\dots\circ \mu_{\otimes X_{q}} \right).
\label{mtp_1}
\end{align}
We introduce the following notations for simplicity:
\noindent
\textbf{Notation 2.} We denote the $k$th parent node of $Y$ by $pa_{k}(Y)$, $\tau\left(\Delta \mu_{\otimes Y}^{(i)}\right)$ by $\tau_{ y }^{(i)}$, $\tau\left(\Delta \mu_{\otimes X_{k}}^{(i)}\right)$ by $\tau_{ pa_k(y) }^{(i)}$, $\tau\left(\mathcal{A}\right)\tau\left(\mu_{\otimes X_{1}}\circ\dots\mu_{\otimes X_{k-1}} \circ \mu_{\otimes X_{k+1}} \circ \dots\circ \mu_{\otimes X_{q}} \right)$ by $c_{y|pa_{k}(y)}$ and $\tau\left(\Delta \mathcal{B}^{(i)}\right)\tau\left(\mu_{\otimes X_{1}}\circ\dots\circ \mu_{\otimes X_{q}} \right)$ by $\epsilon_{pa(y)\rightarrow y}^{(i)}$. We view each $\tau_{ y }^{(i)}$ as a realization of a random variable $\tau_{ y }$. Similarly, there are variables $\tau_{ pa_k(y) }, k=1, \dots, q$ and $\epsilon_{pa(y)\rightarrow y}$.
Then equation \ref{mtp_1} is formalized in the following proposition:
\begin{prop}\label{prop3}
In an MIPG $\mathcal{G}$ of $p$ nodes where each variable $X_{m}$ has $q_m$ independent parent nodes, if assumption \ref{assump1} to \ref{assump3} hold, the normalized traces of the tensor mean embedding of the variation of densities of all variables $\left( \tau_{x_{m}}, m = 1, \dots, p \right)$ fulfill a linear nongaussian acyclic model (LiNGAM) \citep{shimizu-et-al:lingam},
\begin{align}
\bm{\tau}_{x}&\approx \bm{C} \bm{\tau}_{x} + \bm{\epsilon},
\end{align}
where $\bm{\tau}_{x} = \left[\tau_{x_{1}}, \dots, \tau_{x_{p}} \right]^{T}$, coefficient matrix $\bm{C}$ whose element on $n$th row and $m$th column equals to $c_{x_{n}|x_{m}}$ could be permuted to a lower triangular matrix and $\bm{\epsilon} = \left[\epsilon_{pa(X_{1})\rightarrow X_{1}}, \dots, \epsilon_{pa(X_{p})\rightarrow X_{p}} \right]^{T}$.
\end{prop}
\begin{proof}
First, $\tau_{x_{m}}$, where $m = 1, \dots, p$, could be arranged in an \emph{causal order} due to the acyclicity of the graph. Second, the noise term $\epsilon_{pa(x_{m})\rightarrow x_{m}}$, where $m = 1, \dots, p$, follows nongaussian distributions as shown in proposition \ref{prop1}. Thirdly, assumption \ref{assump2} ensures that $\tau_{x_{m}} \perp\!\!\!\perp \epsilon_{pa(x_{m})\rightarrow x_{m}}$ for $m = 1, \dots, p$.
\end{proof}
\begin{algorithm}[!ht]
\caption{ENCI for causal graphs}
\label{alg2}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{algorithmic}[1]
\REQUIRE $N$ data groups $\mathbf{G} = \{G_{1}, G_{2}, \dots, G_{N}\}$
\ENSURE The estimated coefficient matrix $C_{ENCI}$ of the causal graph
\STATE Normalize $X_{m}$ in each group for $m = 1, \dots, p$;
\STATE Compute $\tau_{x_{1}}^{(i)}, \dots, \tau_{x_{p}}^{(i)}$ for $i=1, \dots, N$;
\STATE Apply LiNGAM on $\tau_{x_{1}}, \dots, \tau_{x_{p}}$ and obtain the coefficient matrix $\mathbf{C}$;
\STATE Denote the number of rows and columns with only one non-zero element by $n_{row}$ and $n_{col}$, respectively;
\IF{$n_{row} > n_{col}$}
\STATE Set elements in the rows with more than one non-zero element to be zero except for the maximal element and return the resulting matrix $C_{ENCI}$.
\ELSIF{$n_{row} < n_{col}$}
\STATE Set elements in the columns with more than one non-zero element to be zero except for the maximal element and return the resulting matrix $C_{ENCI}$.
\ELSE
\STATE Return $C$ as $C_{ENCI}$.
\ENDIF
\end{algorithmic}
\end{algorithm}
According to proposition \ref{prop3}, we can apply LiNGAM on the normalized traces of the tensor mean embedding of the variation of densities of all variables to infer the causal structure. However, the coefficient matrix $\mathbf{C}$ returned by LiNGAM needs to be further adjusted since LiNGAM is not restricted to the two kinds of causal graphs we are considering in this letter. Obviously for TSGs, each row of $\mathbf{C}$ contains at most one non-zero element. For MIPGs, each column contains at most one non-zero element since it can be obtained by reversing all directed edges of TSGs. Therefore, we first determine whether the returned coefficient matrix is more likely to be a TSG or MIPG by simply comparing the number of rows and columns with one non-zero element. Then we adjust those rows (columns) that violate the corresponding graph structure. The algorithm of extending ENCI to discover the causal structure of a graph with multiple variables are given in algorithm~\ref{alg2}.
\section{Experiment}
We conduct experiments on both synthetic and real data to verify the effectiveness of our proposed causal discovery algorithm. Unless specified, we adopt gaussian kernel with median $\left( d_{M} \right)$ as its kernel width across all subsections. The implementations of ENCI for cause-effect pairs\footnote{\url{https://github.com/amber0309/ENCI_cause-effect-pair}} and causal graphs\footnote{\url{https://github.com/amber0309/ENCI_causal-graph}} are available online.
\subsection{Synthetic Cause-effect Pairs}
We generate the cause $X$ from the following family of distributions
\begin{align}
X\sim\frac { c_{1} }{ \sqrt { 2\pi { \left( 0.3 \right) }^{ 2 } } } { e }^{ -\frac { { \left( X-1 \right) }^{ 2 } }{ 2{ \left( 0.3 \right) }^{ 2 } } } + \frac { c_{2} }{ \sqrt { 2\pi { \left( 0.3 \right) }^{ 2 } } } { e }^{ -\frac { { \left( X \right) }^{ 2 } }{ 2{ \left( 0.3 \right) }^{ 2 } } }
+ \frac { c_{3} }{ \sqrt { 2\pi { \left( 0.3 \right) }^{ 2 } } } { e }^{ -\frac { { \left( X+1 \right) }^{ 2 } }{ 2{ \left( 0.3 \right) }^{ 2 } } }, \nonumber
\end{align}
where $c_{1}$, $c_{2}$ and $c_{3}$ are randomly sampled from a uniformly distributed simplex. When generating a group of data, $c_{1}$ to $c_{3}$ are firstly sampled to determine the distribution of $X$. Then $40\sim50$ data points are sampled from the corresponding distribution to form a group, and 200 groups are generated in each experiment. The generating mechanism of $c_{1}$ to $c_{3}$ leads to the independence and difference of distributions in different groups. We conduct experiments with both an additive mechanism, $Y = f(X) + E$, and a multiplicative mechanism, $Y = f(X) \times E$. $E$ is the standard Gaussian noise. The function mapping $X$ to $Y$ of each group is randomly chosen from $f_{1}$ to $f_{7}$,
\begin{table}[H]
\centering
\begin{tabular}{llll}
$f_{1}(x)=\frac{1}{x^{2} + 1}$ & $f_{2}(x)=sign(cx) \times (cx)^{2}$ & $f_{3}(x)=\cos(cxn)$ & $f_{4}(x)=x^{2}$ \\
$f_{5}(x)=\sin(cx)$ & $f_{6}(x)=2\sin(x) + 2\cos(x)$ & $f_{7}(x)=4\sqrt{|x|}$ & \\
\end{tabular}
\end{table}
\noindent
where $c$ is a random coefficient independently and uniformly sampled from interval $[0.8, 1.2]$. Overall, $p(X)$ and function $f$ are fixed within each group, whereas they vary in different groups.
We compare ENCI with ANM, PNL, IGCI and ECBP. These existing methods are applied in two different causal inference schemes: $(1)$ on the entire dataset, which is obtained by combing all groups (ALL) and $(2)$ on each group and choose their majority estimation to be their final causal direction estimation (MV). The experimental results of each setting are shown in Table~\ref{synpair}. Note that the accuracies of ECBP are from 50 independent experiments due to its high time complexity and that of other methods are from 100 independent experiments.
\begin{table*}[!ht]
\centering
\caption{Accuracy of synthetic cause-effect pairs}
\label{synpair}
\begin{tabular}{c|ccccccccc}
\toprule
\multirow{2}{*}{Mechanism} & \multirow{2}{*}{ENCI} & \multicolumn{2}{c}{ANM} & \multicolumn{2}{c}{PNL} & \multicolumn{2}{c}{IGCI} & \multicolumn{2}{c}{ECBP}\\
& & MV & ALL & MV & ALL & MV & ALL & MV & ALL\\
\midrule
Additive & \textbf{100} & \textbf{100} & 63 & 99 & 50 & \textbf{100} & 66 & \textbf{100} & \textbf{100} \\
Multiplicative & \textbf{100} & 0 & 26 & 4 & 5 & \textbf{100} & 90 & \textbf{100} & 88\\
\bottomrule
\end{tabular}
\end{table*}
From the experimental results, we can see that ENCI, IGCI-MV, and ECBP-MV performs best compared with other cases. ANM and PNL could not make correct decision in both mechanisms at the same time, and the accuracy of IGCI-ALL is much lower than IGCI-MV, which is probably because of the influence of nonstationarity. ECBP takes non-stationarity into consideration so it achieves satisfactory accuracy in ECBP-ALL. However, we observe that its performance on multiplicative mechanism is not as good as ENCI in our experimental setting.
\subsection{Synthetic Causal Graph}
In this section, we show our experimental results of both kinds of causal graphs.
In the case of tree-structured graph, we conduct experiments on randomly generated graphs with 10 and 50 variables, respectively. First, the distributions of the root node is determined in the same way as the cause $X$ in the previous section. Then each effect is determined by a multiplicative mechanism from its parent node. The function $f$ is randomly chosen from $f_{1}$ to $f_{7}$, and all noise terms follow uniform distribution $\mathcal{U}(0, 1)$. Each time, 1000 groups of data are generated in total. Note that samples within each group are generated from a fixed causal model, but the distribution of the nodes and the mappings between them can vary in different groups.
We compare ENCI with seven existing methods. ECBP \citep{zhang2015discovery}, ICA-LiNGAM \citep{shimizu-et-al:lingam}, DirectLiNGAM \citep{shimizu2011directlingam} and pairwiseLiNGAM \citep{hyvarinen2013pairwise} are directly applied after combining all groups of data. ANM \citep{hoyer2009nonlinear}, PNL \citep{zhang2009identifiability} and IGCI \citep{janzing2012information} are applied on each pair of adjacent nodes so we only have the proportion of correctly estimated edges (recall) for these three methods. Figure~3 shows one of the estimated results of the methods which are able to recover the causal structure. In each experiment, we compute the recall (and precision) of edge from the estimation results. The mean precision (prc) and recall (rcl) are given in column TSG of Table~\ref{synnet}\footnote{Note that the precision and recall of ECBP are computed from the skeleton instead of the directed graph.}. The results of ECBP on TSG with 10 and 50 variables are the mean of 50 and 20 independent experiments, respectively, due to its high time complexity. The results of other methods are the mean of 100 independent runs.
\begin{figure*}[!ht]
\begin{subfigure}{.23\textwidth}
\centering
\begin{tikzpicture}[scale=0.7,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, below=of x1, xshift=-1.5cm] (x2) {$x_{2}$} ; %
\node[latent, below=of x1, xshift=-0.5cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=0.5cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x1, xshift=1.5cm] (x8) {$x_{8}$} ; %
\edge {x1}{x2, x3, x5, x8} ; %
\node[latent, below=of x2] (x7) {$x_{7}$} ; %
\node[latent, below=of x3, xshift=0.25cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x5, xshift=0.25cm] (x9) {$x_{9}$} ; %
\node[latent, below=of x8, xshift=0.25cm] (x10) {$x_{10}$} ; %
\edge {x2}{x7} ; %
\edge {x3}{x4} ; %
\edge {x5}{x9, x10} ; %
\edge[style=dashed] {x1}{x4};
\node[latent, below=of x4] (x6) {$x_{6}$} ; %
\edge {x4}{x6} ; %
\end{tikzpicture}
\caption{}
\label{Fig:syn_enci}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\begin{tikzpicture}[scale=0.7,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, below=of x1, yshift=0.5cm] (x2) {$x_{2}$} ; %
\node[latent, below=of x2, yshift=0.5cm] (x3) {$x_{3}$} ; %
\node[latent, right=of x3, xshift=-0.5cm] (x4) {$x_{4}$} ; %
\node[latent, right=of x4, xshift=-0.5cm] (x8) {$x_{8}$} ; %
\node[latent, right=of x2, xshift=-0.5cm] (x6) {$x_{6}$} ; %
\node[latent, right=of x6, xshift=-0.5cm] (x7) {$x_{7}$} ; %
\node[latent, right=of x1, xshift=-0.5cm] (x5) {$x_{5}$} ; %
\node[latent, right=of x5, xshift=-0.5cm] (x9) {$x_{9}$} ; %
\node[latent, above=of x5] (x10) {$x_{10}$} ; %
\edge[style=dashed] {x10}{x5} ; %
\end{tikzpicture}
\caption{}
\label{Fig:syn_ling}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\begin{tikzpicture}[scale=0.55,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, below=of x1, xshift=-3cm, yshift=0.5cm] (x2) {$x_{2}$} ; %
\node[latent, below=of x1, xshift=-1cm, yshift=-0.5cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=1cm, yshift=0.25cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x1, xshift=3cm, yshift=-0.25cm] (x8) {$x_{8}$} ; %
\node[latent, below=of x2] (x7) {$x_{7}$} ; %
\node[latent, below=of x3] (x4) {$x_{4}$} ; %
\node[latent, below=of x5] (x9) {$x_{9}$} ; %
\node[latent, below=of x8] (x10) {$x_{10}$} ; %
\node[latent, below=of x4] (x6) {$x_{6}$} ; %
\edge{x1}{x2, x3, x5, x8};
\edge[style=dashed]{x1}{x4, x6, x7, x9, x10};
\edge[style=dashed]{x2}{x6, x4, x10, x5, x3};
\edge[style=dashed]{x4}{x3};
\edge[style=dashed]{x5}{x6, x4, x3};
\edge[style=dashed]{x6}{x4, x3};
\edge[style=dashed]{x7}{x4, x3, x6, x5, x10, x2};
\edge[style=dashed]{x8}{x2, x3, x4, x5, x6, x7, x9, x10};
\edge[style=dashed]{x9}{x2, x3, x4, x5, x6, x7, x10};
\edge[style=dashed]{x10}{x3, x4, x5, x6};
\end{tikzpicture}
\caption{}
\label{Fig:dling}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\begin{tikzpicture}[scale=0.55,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, below=of x1, xshift=-3cm, yshift=0.5cm] (x2) {$x_{2}$} ; %
\node[latent, below=of x1, xshift=-1cm, yshift=-0.5cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=1cm, yshift=0.25cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x1, xshift=3cm, yshift=-0.25cm] (x8) {$x_{8}$} ; %
\node[latent, below=of x2] (x7) {$x_{7}$} ; %
\node[latent, below=of x3] (x4) {$x_{4}$} ; %
\node[latent, below=of x5] (x9) {$x_{9}$} ; %
\node[latent, below=of x8] (x10) {$x_{10}$} ; %
\node[latent, below=of x4] (x6) {$x_{6}$} ; %
\edge{x1}{x2};
\edge[style=dashed]{x1}{x6};
\edge[style=dashed]{x2}{x5, x8};
\edge[style=dashed]{x3}{x1, x2, x8, x10, x5}
\edge[style=dashed]{x4}{x1, x2, x8, x7, x10, x5, x3};
\edge[style=dashed]{x6}{x2, x3, x8, x10, x9, x7, x5};
\edge[style=dashed]{x7}{x8, x2, x9, x1, x5, x10, x3};
\edge[style=dashed]{x8}{x1, x5, x9};
\edge[style=dashed]{x9}{x2, x1, x5, x3, x4, x10};
\edge[style=dashed]{x10}{x8, x2, x5, x1};
\end{tikzpicture}
\caption{}
\label{Fig:pling}
\end{subfigure}
\caption{Examples of estimated results of (a) ENCI (b) ICA-LiNGAM (c) DirectLiNGAM (d) pairwiseLiNGAM.}
\end{figure*}
\begin{table*}[!ht]
\centering
\caption{Accuracy of synthetic cause-effect pairs}
\label{synnet}
\begin{tabular}{c|cccccc}
\toprule
\multirow{3}{*}{Methods} & \multicolumn{4}{c}{TSG} & \multicolumn{2}{c}{MIPG} \\
& \multicolumn{2}{c}{10 vars} & \multicolumn{2}{c}{50 vars} & \multicolumn{2}{c}{6 vars} \\
& prc & rcl & prc & rcl & prc & rcl \\
\midrule
ENCI & \textbf{74.55} & 91.56 & \textbf{61.36} & 89.31 & \textbf{57.17} & 96.60 \\
ECBP & 47.23 & 39.18 & 47.69 & 41.12 & 35.92 & \textbf{98.00} \\
ICA-LiNGAM & 7.41 & 0.82 & 5.76 & 0.49 & 30.60 & 91.60 \\
pairwiseLiNGAM & 16.82 & 84.11 & 3.65 & 91.16 & 13.47 & 40.40 \\
DirectLiNGAM & 7.16 & 35.78 & 0.92 & 23.10 & 0.27 & 0.80 \\
ANM & - & 24.33 & - & 26.42 & - & 6.60 \\
PNL & - & 22.44 & - & 17.76 & - & 13.20 \\
IGCI & - & \textbf{99.33} & - & \textbf{92.43} & - & 97.33 \\
\bottomrule
\end{tabular}
\end{table*}
Next we conduct experiments on graphs that allow each variable to have multiple independent parent nodes. The experimental settings are similar to tree-structured case except that we generate 2000 data groups instead of 1000 and the ground truth of the synthetic network structure is fixed to be the graph on the right hand side of Figure~\ref{fig:exmp}. The mean precision and recall are given in the MIPG column of Table~\ref{synnet}. Note again that the results of ECBP are the mean of 20 independent experiments and that of other methods are the mean of 100 independent experiments.
The experimental results show a clear advantage of ENCI over ECBP, ICA-LiNGAM, pairwiseLiNGAM and DirectLiNGAM in estimating nonstationary causal graph. In both cases, ENCI achieves the highest precisions which are far higher than that of other methods. The recall of ENCI are also much higher compared with ICA-LiNGAM and DirectLiNGAM. Although in some cases ECBP and pairwiseLiNGAM return higher recall, their small precisions indicate that they find a large number of spurious edges, which makes their estimations less reliable. Comparing the recall of ENCI with ANM, PNL and IGCI, we find that ENCI still outperforms ANM and PNL. IGCI always performs the best among these four methods. Note that ANM, PNL and IGCI are not able to estimate the network structure and the recall of ENCI is relatively close to that of IGCI.
\subsection{Real Cause-effect Pairs}
This section and the next present the experimental results on real cause-effect pairs and causal graph, respectively. Note that experiments of applying ENCI on both real cause-effect pairs and real causal graphs are conducted on subsampled groups. In other words, we sampled data groups from the raw single data set to create the non-stationarity artificially and then applied ENCI on those randomly sampled groups to evaluate the performance of ENCI on real data.
We test the performance of ENCI on real world benchmark cause-effect pairs\footnote{https://webdav.tuebingen.mpg.de/cause-effect/.}. There are 106 pairs which come from 41 different data sets. Eight data sets are excluded in our experiment because they consists of either multivariate data or categorical data\footnote{Some of the existing methods or their implementations are not applicable to these data}. The corresponding pairs are of ID 47, 52, 53, 54, 55, 70, 71, 101 and 105. ENCI are compared with ANM \citep{hoyer2009nonlinear}, PNL \citep{zhang2009identifiability}, IGCI \citep{janzing2012information} and ECBP \citep{zhang2015discovery}.
\begin{figure}[!hbtp]
\begin{center}
\includegraphics[width = 0.8\linewidth]{real_pairs_2.png}
\end{center}
\caption{Accuracy of methods on real world cause-effect pairs.}
\label{real_pairs}
\end{figure}
We repeat 100 independent experiments for each pair and compute the percentage of correct inference. Then we compute the average percentage of pairs from the same source as the accuracy of the corresponding data set. In the experiment of ENCI, we apply ENCI on 90 groups, each of which consists of 50 to 60 points randomly sampled from the raw data without replacement. Four methods are directly applied on 90 points randomly sample from raw data without replacement in each experiment. Note that ENCI and IGCI is applied using different configurations, and the best result of each pair is adopted for evaluation. For ENCI, we test kernel width $d \in \{1/10d_{M}, 1/5d_{M}, 1/4d_{M}, 1/3d_{M}, 1/2d_{M}, d_{M}, 2d_{M}, 3d_{M}, 4d_{M}, 5d_{M}, 10d_{M} \}$, where $d_{M}$ is the median distance. For IGCI, we test different reference measures (i.e. uniform and gaussian) and estimators (i.e. entropy and integral estimation).
The summary of accuracies on 33 data sets of each method is given in Figure \ref{real_pairs} with orange solid line indicating median of accuracies and green dashed line indicating mean of accuracies. It shows that the performance of ENCI is satisfactory, with both median and mean accuracy about 79\%. IGCI also performs quite well, especially in terms of median, followed by PNL. ANM and ECBP performs poorly on these real cause-effect pairs, which might be due to their model restrictions. ENCI is much more stable than IGCI although its median accuracy is slightly lower. The results on real cause-effect pairs also indicate that ENCI could achieve satisfactory accuracy when applied on subgroups sampled from original data which does not strictly follow our non-stationary model.
\subsection{Real Causal Graph}
In this section, we test ENCI on a sociological data set from a data repository, General Social Survey\footnote{http://www.norc.org/GSS+Website/}.
This dataset consists of 6 observed variables, $x_{1}$: father's occupation level, $x_{2}$: son's income, $x_{3}$: father's education, $x_{4}$: son's occupation level, $x_{5}$: son's education, $x_{6}$: and number of siblings. We use the status attainment model based on domain knowledge~\citep{duncan1972socioeconomic} as the ground truth (see Figure~\ref{Fig:housing_gt}) and compare ENCI with ICA-LiNGAM, DirectLiNGAM and ECBP.
\begin{figure}{}
\centering
\begin{tikzpicture}[scale=0.85,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, above=of x1, xshift=-1.5cm, yshift=-1cm] (x6) {$x_{6}$} ; %
\node[latent, above=of x1, xshift=1.5cm, yshift=-1cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=-0.75cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x1, xshift=0.75cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x5, xshift=-0.75cm] (x2) {$x_{2}$} ; %
\edge[] {x3}{x6};
\edge[] {x6}{x3};
\edge[] {x3}{x1};
\edge[] {x1}{x3};
\edge[] {x1}{x6};
\edge[] {x6}{x1};
\edge[] {x1, x3, x6}{x5} ;
\edge[] {x1, x5, x6}{x4} ;
\edge[] {x4, x5}{x2} ;
\end{tikzpicture}
\caption{Reference Graph of sociological dataset.}
\label{Fig:housing_gt}
\end{figure}
Before applying ENCI, we first adopt k-means++ \citep{arthur2007k} to cluster the original data into 15 clusters. In this way, we regard points within each cluster to be generated from the same causal model. Then we sample 1500 groups, which consists of 50 points sampled without replacement from each cluster with more than 50 points, and apply ENCI on these sampled groups. For ICA-LiNGAM, DirectLiNGAM and ECBP, we directly apply them on the original data set. We show one of the best results of ENCI (coefficient matrix $\mathbf{C}$ obtained from applying LiNGAM on $\tau_{x_{i}}, i=1, \dots, 6$) and the estimated graph from ICA-LiNGAM, DirectLiNGAM and ECBP in Figure~\ref{Fig:housing}.
\begin{figure*}[!ht]
\begin{subfigure}{.24\textwidth}
\centering
\begin{tikzpicture}[scale=0.85,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, above=of x1, xshift=-1.5cm, yshift=-1cm] (x6) {$x_{6}$} ; %
\node[latent, above=of x1, xshift=1.5cm, yshift=-1cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=-0.75cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x1, xshift=0.75cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x5, xshift=-0.75cm] (x2) {$x_{2}$} ; %
\edge[] {x3}{x1};
\edge[] {x1, x3, x6}{x5} ;
\edge[] {x1, x5}{x4} ;
\edge[] {x4, x5}{x2} ;
\edge[style=dashed] {x1}{x2}
\end{tikzpicture}
\caption{}
\label{Fig:housing_enci}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\begin{tikzpicture}[scale=0.85,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, above=of x1, xshift=-1.5cm, yshift=-1cm] (x6) {$x_{6}$} ; %
\node[latent, above=of x1, xshift=1.5cm, yshift=-1cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=-0.75cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x1, xshift=0.75cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x5, xshift=-0.75cm] (x2) {$x_{2}$} ; %
\edge[] {x3}{x1} ;
\edge[] {x5}{x4} ;
\edge[] {x4}{x2} ;
\edge[style=dashed] {x5}{x3, x6};
\end{tikzpicture}
\caption{}
\label{Fig:housing_ling}
\end{subfigure}
\begin{subfigure}{.25\textwidth}
\centering
\begin{tikzpicture}[scale=0.85,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, above=of x1, xshift=-1.5cm, yshift=-1cm] (x6) {$x_{6}$} ; %
\node[latent, above=of x1, xshift=1.5cm, yshift=-1cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=-0.75cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x1, xshift=0.75cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x5, xshift=-0.75cm] (x2) {$x_{2}$} ; %
\edge[] {x3}{x1} ;
\edge[] {x5}{x4} ;
\edge[] {x4}{x2} ;
\edge[] {x6}{x5} ;
\edge[style=dashed] {x5}{x3};
\end{tikzpicture}
\caption{}
\label{Fig:housing_dling}
\end{subfigure}
\begin{subfigure}{.24\textwidth}
\centering
\begin{tikzpicture}[scale=0.85,transform shape,
state/.style={circle,draw,thick,loop above,inner sep=0,minimum width=10}]
\node[latent] (x1) {$x_{1}$} ; %
\node[latent, above=of x1, xshift=-1.5cm, yshift=-1cm] (x6) {$x_{6}$} ; %
\node[latent, above=of x1, xshift=1.5cm, yshift=-1cm] (x3) {$x_{3}$} ; %
\node[latent, below=of x1, xshift=-0.75cm] (x4) {$x_{4}$} ; %
\node[latent, below=of x1, xshift=0.75cm] (x5) {$x_{5}$} ; %
\node[latent, below=of x5, xshift=-0.75cm] (x2) {$x_{2}$} ; %
\edge[] {x3}{x1};
\edge[style=dashed] {x4}{x1};
\edge[style=dashed] {x4, x2}{x5} ;
\edge[style=dashed] {x5}{x3} ;
\edge[] {x4}{x2}
\end{tikzpicture}
\caption{}
\label{Fig:housing_ecbm}
\end{subfigure}
\caption{Estimated graph of (a) ENCI (b) ICA-LiNGAM (c) DirectLiNGAM (d) ECBP.}
\label{Fig:housing}
\end{figure*}
ENCI outperforms ICA-LiNGAM, DirectLiNGAM which is consistent with our expectation since they are developed for linear stationary models. ENCI also outperforms ECBP which may be due to the lack of nonstationarity of the raw data.
There are two facets of ENCI worth noting from the results of real data experiments of both pairs and causal graphs. First, ENCI is applied on subgroups sampled from the raw data since each set of real data is a single collection of observations and does not contain the form of nonstationarity our model assumes. However, the results of ENCI on real pairs is quiet competitive and it performs much better than LiNGAM family methods in real causal graph. This gives some evidence that our model could achieve satisfactory performance with subtle nonstationarity, which may be simply generated by subsampling a single data set. Second, the reference graph in the real graph experiment does not strictly fulfill the requirements of ENCI, but we obtain acceptable estimation results, which implies that ENCI may be applicable for other kinds of causal graphs.
\section*{Conclusion}
In this paper, we introduce the nonstationary causal model and prove the asymmetry of non-stationarity between the causal direction and anti-causal direction based on certain assumptions. By exploiting this asymmetry, we propose a reproducing kernel Hilbert space embedding-based method, ENCI, to infer the causal structure of both cause-effect pairs and two kinds of causal graphs. Theoretical analysis and experiments show the advantage of ENCI over existing methods based on fixed causal models when being applied on nonstationary passive observations.
Compared with ECBP which is also for non-stationary causal model inference, the theoretical scope of application of ENCI is more restricted in the sense that we require non-stationarity in both $p(X)$ and $p(Y|X)$, whereas ECBP would also work when only one of them is nonstationary. In addition, ENCI requires nonstationarity exists in every variable of a causal graph, whereas ECBP only requires the existence of nonstationarity. However, ENCI outperforms ECBP on the experiments of both real cause-effect pairs and causal graph in which the data generating process does not strictly follow our model assumptions and the nonstationarity among artificial groups is subtle. Therefore, we deem that ENCI could be applied on a much wider scope of problems in reality and achieve satisfactory performance. In this way, ENCI is eligible to join a pool of state-of-the-art algorithms for learning general causal models.
\subsection*{Acknowledgments}
We would like to thank Biwei Huang and Kun Zhang for providing the code of Enhanced Constraint-based Procedure (ECBP).
\bibliographystyle{unsrt}
|
\section{Introduction}{\label{intro}}
Modulation instability (MI) is one of the most fundamental process of nonlinear wave dynamics in various systems \cite{zakharov2009modulation}.
The instability undergoes a spontaneous growth from small-amplitude modulated waves and leads to
stable large-amplitude localized waves, typically solitary waves, as a result of
interplay between the intrinsic nonlinearity and the dispersion.
The MI has been studied in the nonlinear dynamics
of matter waves, corresponding to Bose-Einstein condensates (BECs) in ultracold atomic
gases \cite{kevrekidis2007emergent}. The dynamics of the matter waves is described by the nonlinear Schr\"{o}dinger equation, also
known as the Gross-Pitaevskii (GP) equation, where the nonlinearity is associated with
the interatomic collisions. The remarkable feature of this system is that
the dispersion as well as nonlinearity can be controlled experimentally in a well controlled manner \cite{morsch2006dynamics,chin2010feshbach}.
In addition, there is diverse richness as the nonlinear wave system; for
example, we can consider the system of multicomponent order parameters
with various linear and nonlinear couplings between them.
The MI in scalar BECs has been firstly discussed in the context of the formation
of bright soliton trains \cite{theocharis2003modulational,salasnich2003modulational,carr2004spontaneous}.
This has been experimentally demonstrated by tuning the interaction of condensed
atoms from repulsive to attractive \cite{strecker2002formation,nguyen2017formation,everitt2017observation}.
For a uniform scalar condensate, the MI is possible only for the attractive nonlinearity.
Recently, the argument of MI has been discussed for the
BEC with long-range interactions \cite{ferrier2018onset}, where
the MI is related to the formation of quantum droplets as a
result of the beyond mean-field corrections \cite{ferrier2016observation,schmitt2016self}.
Also, the MI has been extended to the system of multicomponent BECs.
The presence of the intercomponent interaction induces the MI even for the
condensates with repulsive nonlinearity \cite{goldstein1997quasiparticle}.
When the intercomponent repulsion is stronger than the intracomponent one,
the MI induces the formation of multiple domains \cite{kasamatsu2004multiple,kasamatsu2006modulation,PhysRevA.85.043602,vidanovic2013spin,eto2016nonequilibrium}.
Recent theoretical analysis has revealed that the formation and subsequent coalesce
dynamics of condensate domains are governed by the universal scaling law
\cite{sabbatini2011phase,de2014quenched,hofmann2014coarsening,takeuchi2015phase}.
The recent papers on MI has shown that two-component BECs
with spin-orbit coupling (SOC) are always subject to the MI
for arbitrary choice of the nonlinearities \cite{Bhat:2015, Bhuvaneswari:2016,Congy:2016}.
Thus, the SOC extends the parameter region of the MI from that of the conventional
two-component BECs. The SOC can be synthesized by
the Raman laser coupling scheme between internal states of cold atoms \cite{lin2011spin}.
The static and dynamical properties of the BECs with Raman-laser induced SOC have been studied in
many papers
\cite{lin2011spin,ho2011bose,li2012quantum,chen2017collective,martone2012anisotropic,zhang2012collective,zheng2013properties,achilleos2013matter,kasamatsu2015dynamics}, but the strong nonlinear evolution caused by MI has not been studied so much.
Recently, Ye \textit{et al}. studied the domain formation through the parameter quench from the mixed phase to the
plane-wave phase of the BEC with the SOC \cite{ye2018universal}.
The objective of the present work is the analysis of nonlinear dynamics caused
by MI in the effectively one-dimensional (1D) BEC with synthetic SOC by the numerical
simulations of the GP equation.
Following the MI analysis by Bhat \textit{et al} \cite{Bhat:2015}, we study the nonlinear
dynamics starting from continuous wave (cw) states of miscible two-component BECs.
By suddenly turning on the synthetic SOC, the system is modulationally unstable and
there appears a pulse-like structure. The initial stage of the evolution is consistent
with the prediction of the MI analysis, where the dynamically unstable modulations grow spontaneously from
primary and secondary instability bands in small-$k$ and large-$k$ regions.
The subsequent nonlinear evolutions exhibit complicated dynamics of pattern formation in a
real space, while clear separation of the wave functions of the two component BECs in a momentum
space is observed due to the effect of the SOC. Our results reveals the richness in the complex dynamics exhibiting by the SOC BECs, alongside of the recent experimental and theoretical observation of the spin dynamics and the dynamical instabilities \cite{Burdick:2016, Zhu:2016,Khamehchi:2017} .
The paper is organized as follows. In Sec.~\ref{MIofBEC}, we introduce the basic formulation
of the problem and briefly review the MI in BECs with the Raman-induced SOC.
Section~\ref{nonlinearMI} presents the results of the series of numerical simulations of the MI induced nonlinear dynamics.
In Sec.~\ref{conclusion}, we devote to the conclusion.
\section{Modulation instability of BECs with synthetic SOC}\label{MIofBEC}
In this section, we first introduce the basic formulation of BECs with a synthetic SOC in our problem.
Next, we briefly review the MI analysis done by Bhat \textit{et al}. \cite{Bhat:2015} and
specify the parameter region to see the MI-induced nonlinear dynamics.
\subsection{Model}\label{model}
We consider quasi-1D two-component (psudospin-1/2) BECs
with the combined Rashba-Dresselhaus SOC induced by the Raman lasers \cite{lin2011spin}.
The single-particle hamiltonian with the synthetic SOC in a quasi-momentum frame has the $2\times 2$ matrix structure \cite{Zhu:2016}
\begin{equation}
h_0 = \frac{1}{2m} \left(p_x \sigma_0- \hbar k_\text{R} \sigma_z \right)^2
+ \frac{\hbar \delta}{2} \sigma_z+ \frac{\hbar\Omega_\text{R}}{2} \sigma_x + V_\text{tr} \sigma_0.
\label{singleho}
\end{equation}
Here, $m$ is the atomic mass, $p_x = -i \hbar \partial_x$ the quasi-momentum operator
along the $x$-direction, $\sigma_r$ for $r=x,y,z$ is one
of the Pauli matrices and $\sigma_0$ is the unit matrix. The quasi-momentum
is related with the real momentum $p_x'$ as $p_x' \sigma_0=p_x \sigma_0- \hbar k_\text{R} \sigma_z$.
The trapping potential is assumed to be a harmonic form
$V_\text{tr} = m \omega^2 x^2 / 2$.
The SOC is characterized by three parameters, $k_\text{R}$, $\Omega_\text{R}$ and $\delta$ under experimental
control \cite{lin2011spin}, where $k_\text{R}$ is the wavenumber of the Raman laser which couples the two atomic
hyperfine states, $\Omega_\text{R}$ is the Rabi frequency determined by
the intensity of the Raman laser, and $\delta$ is the detuning.
For simplicity, the detuning $\delta$ is set to be zero.
The kinetic energy term has a uniform synthetic gauge field
$ -\hbar k_\text{R} \sigma_z$ proportional
to the spin matrix $\sigma_z$, which represents the 1D SOC whose magnitude can be controlled by $k_\text{R}$.
The GP energy functional including the single-particle hamiltonian
of Eq.~(\ref{singleho}) and the atom-atom interactions is given by
\begin{align}
E = \int dx \left( \Psi^\dagger h_0 \Psi +
\frac{1}{2} \sum_{j=1,2} u_{j}|\Psi_j|^4 + u_{12} |\Psi_1|^2 |\Psi_2|^2 \right). \label{EGPSO}
\end{align}
The order parameters are represented by the 2-component spinor $\Psi = (\Psi_1,\Psi_2)^{T}$.
The parameters $u_j$ and $u_{12}$ are the coupling constants adjusted to the quasi-1D description
by incorporating the length scale along the tightly confined direction \cite{kasamatsu2006modulation},
being proportional to the $s$-wave scattering lengths of atoms.
We have set the same intracomponent coupling constant as $u_1=u_2=u$ for simplicity.
When we take the energy scale by the recoil energy $E_\text{R} = \hbar^2k_\text{R}^2/m$ and
the length scale by $k_\text{R}^{-1}$, Eq.~\eqref{singleho} is scaled as
\begin{align}
\tilde{h}_0 = \frac{1}{2} \left(- i \frac{\partial}{\partial \tilde{x}} \sigma_0 - \tilde{\gamma} \sigma_z \right)^2
+ \tilde{\Gamma} \sigma_x + \tilde{V}_{\mathrm{tr}} \sigma_0, \label{singleph}
\end{align}
where the dimensionless quantities are represented by symbols with tildes.
In our unit, although the parameter $\tilde{\gamma}$ is kept unity, we leave this notation
because this parameter is used as a quench parameter to induce the MI.
The Rabi frequency is written as $\tilde{\Gamma} = \hbar \Omega_\text{R} / (2 E_R)$ and
the trap potential as $\tilde{V}_\text{tr} = \lambda^2 \tilde{x}^2 / 2$ with
the coefficient $\lambda = (a_\text{ho} k_\text{R})^{-2}$, where
$a_\text{ho} = \sqrt{\hbar/(m\omega)}$ is the harmonic oscillator length.
Following a usual experimental condition, we use $\lambda = 0.02$ in the following calculation.
The normalization of the wave function is given by the total particle number in the 1D system
$N = \int dx \Psi^{\dagger} \Psi = \int d x (|\Psi_1|^2 + |\Psi_2|^2) = N_1 + N_2$.
By replacing the wave function as $\Psi = \sqrt{N k_\text{R}} \tilde{\psi}$,
we have $\int d \tilde{x} \tilde{\psi}^{\dagger} \tilde{\psi} = \int d \tilde{x} ( |\tilde{\psi}_1|^2 + |\tilde{\psi}_2|^2) =1$ and
define the dimensionless coupling strengths as $\tilde{g} = m N u / (\hbar^2 k_\text{R})$ and $\tilde{g}_{12} = m N u_{12} / (\hbar^2 k_\text{R})$.
The time-dependent GP equations derived from Eq.~(\ref{EGPSO})
can be written as
\begin{align}
i \frac{\partial \psi_1}{\partial t} = \left( -\frac{\partial_x^2}{2} - i \gamma \partial_x + V_\text{tr} + g |\psi_1|^2 + g_{12} |\psi_2|^2 \right) \psi_1 \nonumber \\
+ \Gamma \psi_2, \label{tdgp1} \\
i \frac{\partial \psi_2}{\partial t} = \left( -\frac{\partial_x^2}{2} + i \gamma \partial_x + V_\text{tr} + g |\psi_2|^2 + g_{12} |\psi_1|^2 \right) \psi_2 \nonumber \\
+ \Gamma \psi_1, \label{tdgp2}
\end{align}
where tildes are omitted in the notation. The unit of time is taken as $\hbar/E_\text{R}$.
%
\subsection{Ground States}
\begin{figure}[!htbp]
\includegraphics[width=4.2cm,scale=1]{ground
\includegraphics[width=4.2cm,scale=1]{ground1l} \\
\includegraphics[width=4.2cm,scale=1]{ground2l}
\includegraphics[width=4.2cm,scale=1]{ground3l
\vspace{-2.0pt}
\caption{\label{Fig1} {\footnotesize(Color online) Density profiles of the ground state of Eqs. \eqref{tdgp1} and \eqref{tdgp2} for the coupling constants $g=50$ and $g_{12}=0.95g$,
and the trap frequency $\lambda=0.02$. Figure (a) represents the profile without the SOC, $\Gamma=\gamma=0$, which corresponds to the initial states of the
time evolution shown below. Figures (b) represents the ground state for $\Gamma=0.1$, and (c) for $\Gamma=0.5$, and (d) for $\Gamma=1.5$,
corresponding to the stripe phase ($0<\Gamma \lesssim 0.2$), plane-wave phase ($0.2\lesssim \Gamma \lesssim 1$) and mixed phase ($1 \lesssim \Gamma$), respectively with $\gamma=1$.
}}
\end{figure}
It is well known that two-component BECs without SOC have two types of the ground state phases characterized by the miscible or immiscible
density profile \cite{ao1998binary,trippenbach2000structure}. This miscible-immiscible transition is associated with the relation of the coupling constant; when $g > g_{12}$ ($g < g_{12}$)
the system is miscible (immiscible). This condition is deeply related to the MI in conventional two-component BECs \cite{goldstein1997quasiparticle,kasamatsu2004multiple,kasamatsu2006modulation}.
When, the SOC is present, for a given strength of inter and intracomponent interaction, there exist three different ground state phases
depending on the Rabi frequency $\Gamma$ \cite{li2012quantum,martone2012anisotropic,chen2017collective}.
The phases are (i) the stripe (supersolid) phase, (ii) the plane-wave (polarized) phase, and (iii) the mixed (single-minimum) phase \cite{Chuanzhou:2016}.
In the following, we consider the case $g > g_{12} > 0$, which yields the miscible phase and the modulationally stable condition in the absence of SOC.
We fix the value $g=50$ and $g_{12}=0.95g$. The typical ground state solutions with and without the SOC are shown in Fig.~\ref{Fig1}.
With these coupling constants, the phase (i) appears for $0<\Gamma \lesssim 0.2$, (ii) for $0.2 \lesssim \Gamma \lesssim 1.0$, and (iii) for $1.0 \lesssim \Gamma$.
The transition between phase (ii) and (iii) is consistent with the property of the dispersion relation of the single particle Hamiltonian Eq.~\eqref{singleph} for $V_\text{tr} = 0$,
which is given by
\begin{equation}
\epsilon_{\pm} = \frac{k^2}{2} + \frac{\gamma^2}{2} \pm \sqrt{k^2 \gamma^2 + \Gamma^2} \label{sdispersion};
\end{equation}
the branch $\epsilon_{-}$ has a change between single- and double-minimum structure at $\Gamma = \gamma$ ($\gamma=1$ in our unit).
We choose the values of the Rabi coupling $\Gamma =$0.1, 0.5, and 1.5, corresponding to three different phases, to study the nonlinear dynamics caused by MI for the initial state in Fig.~\ref{Fig1}(a).
\subsection{MI in BECs with the Raman-induced SOC}
\begin{figure}[!htbp]
\includegraphics[width=0.8\linewidth,scale=1]{MIgain}
\caption{\label{FigMIcond}{\footnotesize (Color online) MI gain defined as $\xi = \text{Im}(\Omega)$ with the excitation frequency $\Omega$
in the $k$-$\Gamma$ plane for $g=50$, $g_{12}=0.95g$ and $n_0 = 0.0066$, which is taken from the Thomas-Fermi density at $x=0$ [see Fig.~\ref{Fig1}(a)].
There are four instability bands, where the two inner bands are given by the two branches $\text{Im}(\Omega_{+})$ and $\text{Im}(\Omega_{-})$ with an equivalent contribution,
while the outer two bands are given by $\text{Im}(\Omega_{-})$.
}}
\end{figure}
Here, we address the MI condition for the two-component BECs with the Raman-induced SOC.
Bhat \textit{et al.} considered the MI of the miscible cw (uniform) state of two-component BECs with respect to the general parameter sets
of the Raman-induced SOC and the interaction strengths \cite{Bhat:2015}.
Irrespective of the combinations of the interaction strengths, the cw states are always affected by the MI
in the presence of the SOC. In this work, we confine ourselves to the case $g > g_{12}$, $\gamma =1$ and $\Gamma > 0$.
The condition of MI is given by the appearance of the imaginary component in the excitation frequency $\Omega$
for the small-amplitude modulation around the initial cw state.
The details of this condition are given in Ref.~\cite{Bhat:2015} and the results are summarized in the Appendix~\ref{append2}.
Figure~\ref{FigMIcond} represents the MI gain defined by $\xi = |\text{Im}(\Omega_{\pm})|$ with respect to the
wave number $k$ and the Rabi frequency $\Gamma$.
Here, the dispersion relation has two branches, corresponding to the (in-phase) density wave excitation $\Omega_{+}$ and
the (out-of-phase) spin wave excitation $\Omega_{-}$. From Fig.~\ref{FigMIcond}, we see that the cw state is always dynamically unstable when
there is the SOC.
There are four unstable domains in the $k$-$\Gamma$ space; the two domains in the smaller-$|k|$ region, referred to as a ``primary MI band'', come from
the two branches $\Omega_{\pm}$ with a equivalent contribution, while the other two domains, referred to as a ``secondary MI band'', in the larger-$|k|$ region
come from only the branch $\Omega_{-}$. Thus, the SOC brings about a new regime of the MI for the two-component BECs
which is dynamically stable without the SOC.
%
\section{Nonlinear dynamics}\label{nonlinearMI}
In this section, we discuss the MI-induced nonlinear dynamics of the BECs with the synthetic SOC.
We solve Eqs.~(\ref{tdgp1}) and (\ref{tdgp2}) numerically by the split-step Fast-Fourier method;
Details of the numerical method are described in the Appendix \ref{append1}. To generate MI, we turn on the SOC
in the initially miscible condensates in Fig.~\ref{Fig1}(a) by introducing $\Gamma$ and $\gamma (= 1)$ suddenly at $t=0$.
The developed MI dynamics for different strengths of $\Gamma$, which gives the three different ground states in Fig.~\ref{Fig1}(b)-(d),
are as follows.
\subsubsection{$\Gamma=0.1$}
\begin{figure*}[!htbp]
\includegraphics[width=15.0cm,scale=1]{mith1a} \\
\includegraphics[width=17.0cm,scale=1]{MIl_1t}
\vspace{-2.0pt}
\caption{\label{Figa0.1}{\footnotesize (Color online) Time development of the condensate densities in the real coordinate space, (a) $n_1=|\psi_1(x)|^2$
and (b) $n_2=|\psi_2(x)|^2$, and in the quasi-momentum space, (c) $n_1 = |\phi_1(k)|^2$ and (d) $n_2=|\phi_2(k)|^2$ for $\Gamma=0.1$. The solid red and the dashed green curves represent the MI gain $\xi$ [Eq.~\eqref{eq. gain} and the cross section of Fig.~\ref{FigMIcond} at $\Gamma=0.1$] for $\Omega_{-}$ and $\Omega_{+}$, respectively.
The lower panels show the snapshots of the density profiles at different times: (e)-(i) in the coordinate space and (j)-(n) in the quasi-momentum space.
Simulation is done for the range [-200:200] with 4096 grid points in the coordinate space.}}
\end{figure*}
Figure~\ref{Figa0.1} shows the time development of the MI-induced spatial pattern formation from the cw state for $\Gamma = 0.1$.
The upper left panels (a, b) and the lower panels (e-i) represent the dynamics of the condensate densities $n_i (x)= |\psi_i(x)|^2$
in the coordinate space. The motions of both components behave similarly and keep inversion symmetry with respect to $x=0$.
First, the condensates make out-of-phase dipole motions to shift the centers of mass from each other.
As time evolves, the densities of both components are well separated and break into smaller domains. After $t=50$,
the both components fragment into the non-periodic short-wavelength domains and continue to make a chaotic oscillation.
Further insight can be seen in the dynamics of the wave function in the quasi-momentum space.
The upper left panels (c, d) and the lower panels (j-n) represent the dynamics of the densities $n_i(k_x) = |\phi_i(k_x)|^2$
of the Fourier component $\phi_i(k_x) = \int dx \psi_i(x) e^{-ik_x x}$, where in a lab frame Fig.~\ref{Figa0.1}(j) corresponds to the initially overlapped wave packets sitting at trap center with left-going spin-up momentum $-\hbar k_R$ and a right-going spin-down momentum $ \hbar k_R$.
The panels (c, d) also show the MI gain (cross section of Fig.~\ref{FigMIcond} at $\Gamma=0.1$) for clarity.
In the quasi-momentum space, the SOC $\pm i \gamma \partial_x$ in Eqs.~\eqref{tdgp1} and \eqref{tdgp2} contributes
as a linear potential $+\gamma k_x$ and $-\gamma k_x$ for $\phi_1$ and $\phi_2$, respectively.
Thus, $\phi_1$ and $\phi_2$ are basically driven to the negative and positive direction in the $k$-space, respectively.
First, since the primary MI bands exist around $k=0$, the low-energy dipole motions grow spontaneously due to the MI.
This dipole motion induces the spin-wave excitations and leads to the generation of unstable modes in the secondary bands.
The panels (c) and (d) clearly show this transfer process of the wave component from small- to large-$k$ region, where the $n_{1(2)}(k_x)$ propagates to the
negative (positive) wave number of the secondary MI band associated with $\Omega_{-}$. During the time evolution, from the panels (j-n), we can see that the waves collapse into the complicated short-wavelength domains but
show clear phase separation in the positive and negative range of the quasi-momentum space because of the constant bias of the SOC.
Since the double minima of Eq.~\eqref{sdispersion} exist at $k = \pm \sqrt{\gamma^2 - \Gamma^2}/\gamma \approx \pm 1$,
$|\phi_{1}(k_x)|^2$ and $|\phi_{2}(k_x)|^2$ eventually distribute around $k_x \sim 1$ on average.
\subsubsection{$\Gamma=0.5$}
\begin{figure*}[!htbp]
\includegraphics[width=15.0cm,scale=1]{mith2a} \\
\includegraphics[width=17.0cm,scale=1]{MIl_2t}
\vspace{-2.0pt}
\caption{\label{Figa0.5}{\footnotesize (Color online) The results similar to Fig.~\ref{Figa0.1} for $\Gamma=0.5$.}}
\end{figure*}
For the increased value of Rabi frequency, $\Gamma=0.5$, the evolution of the density in the coordinate space is shown in Fig.~\ref{Figa0.5}(a,b) and snap shots of density at different times are depicted in Fig.~\ref{Figa0.5}(e-i).
Initially a few density stripes appear in both components at the center of the condensate and grow in-phase. As time evolves, the number of stripes increases and outer density exhibits large oscillatory behaviors. Also, one can see that the initial shift of the center-of-mass shift observed for $\Gamma =0.1$ is suppressed here.
Finally, the both components are fragmenting into the non-periodic short-wavelength domains as in the case of $\Gamma =0.1$.
In the Fourier space, Figs.~\ref{Figa0.5}(c,d) and (j-n) show that the initially excited wave vector exactly matches with the analytically predicted $k_\text{max}$ of the primary MI band, as seen
in Fig.~\ref{Figa0.5}(c,d). Although a small amount of excitations appears at the secondary bands at later times, the increased distance (when compared to the that of $\Gamma=0.1$ case) between the primary the secondary bands keeps the excitations well in the primary bands.
Also, the dynamics of the components ($\phi_1$ and $\phi_2$) are not completely separated in to the positive and negative wave vector values as in the case of $\Gamma=0.1$. This is expected because of the reduction in the effective potential given by the combination of Rabi coupling and the SOC. These properties are again
consistent with the single-particle dispersion of $\epsilon_-$, where the separation of the double minima are reduced compared to the case of $\Gamma= 0.1$.
\subsubsection{$\Gamma=1.5$}
\begin{figure*}[!htbp]
\includegraphics[width=15.0cm,scale=1]{mith3a} \\
\includegraphics[width=17.0cm,scale=1]{MIl_3t}
\vspace{-2.0pt}
\caption{\label{Figa1.5}{\footnotesize (Color online) The results similar to Fig.~\ref{Figa0.1} for $\Gamma=1.5$.}}
\end{figure*}
\begin{figure}[!htbp]
\includegraphics[width=8.6cm,scale=1]{MI_time} \\
\vspace{-2.0pt}
\caption{\label{Fig_MI_time}{\footnotesize (Color online) The modulation instability time (black dot) is plotted for varying Rabi frequency $\Gamma$. The dashed line guides the eye.}}
\end{figure}
Fig.~\ref{Figa1.5}(a,b) and Fig.~\ref{Figa1.5}(e-i) show that the spatial pattern formation for $\Gamma=1.5$ which resemble that of $\Gamma=0.5$. The density modulations of the both components generate spontaneously at the center, growing to large amplitude density waves. Then, each component makes a spatial separation [Fig.~\ref{Figa1.5}(h)] and continues a complicated chaotic oscillation. In the Fourier space, Fig.~\ref{Figa1.5}(c,d) and Fig.~\ref{Figa1.5}(j-n) display that initial modulation appears at the primary and secondary bands, consistent with the MI analysis. Here, the excitations happens symmetric to $k = 0$ in the primary bands, meanwhile, in the secondary band excitations are asymmetric due to the increased value of $|k|$.
Since the dispersion $\epsilon_{-}$ of Eq.~\eqref{sdispersion} has a single minimum at $k=0$ for $\Gamma=1.5$, the separation of $\phi_i$ in the quasi-momentum space does not occur; the fragmented domains of $\phi_i$ oscillate around $k=0$ as seen in Fig.~\ref{Figa1.5}(c,d).
Fig.~\ref{Fig_MI_time} shows the plot of the critical time $t_{\text{MI}}$ at which MI induced chaotic dynamics starts. It is found that with Rabi frequency $\Gamma$, $t_{\text{MI}}$ also increases and exhibits a linear relation for $\Gamma \gtrsim 0.5$.
This is because the primary MI band are well separated from the $k \simeq 0$ region, thus, there require some time to reach the finite wave length modes
through the nonlinear mode couplings. If we input the initial noise corresponding to the wave number at the primary or secondary MI bands, the MI grows quickly after the SOC is turned on.
\section{Conclusions} \label{conclusion}
In conclusion, we investigate the nonlinear dynamics induced by MI in two-component BECs with Raman-induced synthetic SOC.
In the previous studies, the MI was predicted to occur for arbitrary choices of the intra- and intercomponent coupling constants.
We demonstrated that even for miscible two-component BECs, which is dynamically stable without SOC, the MI
can take place and cause complicated nonlinear dynamics of pattern formation.
The onset of the nonlinear evolution is consistent with the theoretical prediction of the MI for homogeneous
two-component BECs. The presence of the primary and secondary MI bands induces the characteristic two-step
nonlinear evolution of the pattern formation. At the later stages of the evolutions, the wave functions in the
quasi-momentum space undergo the separation due to the asymmetric feature of the SOC, depending on the values of the
Rabi coupling corresponding to the three ground state phases.
Although the MI yields the very complicated dynamics as shown in Figs.~\ref{Figa0.1}, \ref{Figa0.5} and \ref{Figa1.5}, the ground states
are well-defined shape as in Figs.~\ref{Fig1}(b-d). It is interesting to see the relaxation process from the
strongly nonequilibrium fragmented states to the ordered ground states by introducing some dissipation
or fluctuation effects. Also, revealing nonlinear dynamics in higher-dimensions is an interesting direction for future studies.
We hope that this work further stimulates the studies of the nonlinear dynamics in multicomponent BECs.
\section{Acknowledgements}
T. M acknowledges financial support from IBS (Project Code No. IBS-R024-D1). The work of K.K. is partly supported by KAKENHI from the Japan Society for the Promotion of Science (JSPS) Grant-in-Aid for Scientific Research (KAKENHI Grant No. 18K03472).
|
\section{Introduction}
The purpose of this paper is to present numerical evidence for a conjecture on the spectral asymptotics of the Laplacian on a surface of constant curvature presented by the second author in \cite{strichartz2016}. The conjecture was initially limited to surfaces of either zero or constant positive curvature, but we present strong evidence that a version of it should also be valid in the case of constant negative curvature. Results of Bleher \cite{bleher1994} show that it cannot be valid for surfaces of variable curvature.
We consider surfaces $S$ of finite area $A$ with boundary $\partial S$ of finite perimeter $P$
that is made up of a finite number of smooth curves meeting at angles ${\{\theta_{j}\}}$. Simple
examples are triangles and discs in either the Euclidean plane, the sphere, or hyperbolic 2-space.
We will also look at more complicated examples where $S$ is not simply connected and has
non-convex boundary. We consider the standard Laplacian $\Delta$ with either Dirichlet ($D$),
Neumann ($N$), or mixed boundary conditions ($D$ on a portion of the boundary with perimeter
${P_{D}}$, and $N$ on the remaining portion of the boundary with perimeter ${P_{N}}$). We let
${\{\lambda_{j}\}}$ be the set of eigenvalues ${-\Delta u_{j} = \lambda_{j}u_{j}}$ repeated
according to multiplicity, so the ${\lambda_{j}}$ are all nonnegative and ${\lambda_{j}\rightarrow\infty}$
as ${j\rightarrow\infty}$. The eigenvalue counting function is defined as
\begin{equation}
{N(t) = \# \{\lambda_{j} \leq t\}}\tag{1.1} \newline
\end{equation}
(Note that some references will use $\lambda_{j}\leq t^2$ instead). The well-known Weyl asymptotic formula $N(t) \sim \frac{A}{4\pi}t$ was refined by Ivvii \cite{ivrii1984precise} to
\begin{equation}
\tag{1.2} {N(t) = \frac{A}{4\pi}t + \frac{P_{N}-P_{D}}{4\pi}t^{1/2}} + O(t^{1/2}) \newline
\end{equation}
In \cite{strichartz2016} we proposed a still more refined asymptotic
\begin{equation}
\tag{1.3} {\widetilde{N}}(t) = \frac{A}{4\pi}t + \frac{P_{N}-P_{D}}{4\pi}t^{1/2} + C \newline
\end{equation}
where the constant $C$ will be explained in Definition 1.1 below. However, it is impossible to see the constant from $N(t)$ alone, since it is expected that
\begin{equation}
\tag{1.4} D(t) = N(t) - {\widetilde{N}(t)}
\end{equation}
will have at least growth $O(t^{1/4})$. Instead we consider the ordinary average error
\begin{equation}
\tag{1.5} A(t) = \frac{1}{t}\int_0^t D(s)\,ds \newline
\end{equation} We conjecture that this is bounded and decays on the order of $O(t^{-1/4})$ in the flat or negative curvature case. See Conjecture 1.2 below for a more detailed description. For this to be valid we need the correct value for the constant.
We note that a different kind of average, the trace of the heat kernel
\begin{equation}
\tag{1.6}h(t) = - \sum_j e^{-t\lambda_{j}} \newline
\end{equation} (as $t\rightarrow 0$) has been extensively studied, beginning with the famous paper of Mark Kac \cite{kac1996hear} and continuing with \cite{vandenBerg1990}, \cite{buser1992geometry}, \cite{gilkey2003asymptotic}, \cite{mckean_jr.1967}, and \cite{stewartson_waechter_1971}. A related "logarithmic Gaussian averaged error estimate" is studied by Brownell in \cite{brownell1957extended} (see also \cite{baltes1976spectra} for a discussion of this). These are smoother type averages than the ones we consider, and in particular they involve the entire spectrum. Because they are smoother averages, they effectively erase some of the interesting detail that the rougher averages see. It is straightforward to obtain the smooth average results from the rough average results, and in particular our formula (1.3) for the refined asymptotics is consistent with the earlier results. To go in the reverse direction requires using a Tauberian theorem that only yields the original Weyl asymptotics. The average $A(t)$ is a special case of Riesz means, which have also been studier, starting with H\"{o}rmander \cite{hörmander1968}. See also \cite{Fulling1999} for an extensive survey of this approach.
The method we use to numerically approximate the spectrum of the Laplacian is extremely straightforward. We use the finite element solver built into MATLAB. For surfaces in the plane we just have to give a description of the boundary. For
surfaces in the hyperbolic plane or sphere we us a conformally flat coordinate system so the surface Laplacian becomes a scalar multiple of the Euclidean Laplacian. By using the mesh refinement option we obtain better and better approximations of
smaller initial segments of the spectrum. Given the computation time constraints, this allows us to get confident approximations for only a couple hundred eigenvalues. We then use an ad hoc extrapolation method on the sequence of approximations with
increasing refinements to get a slightly improved final approximation. We were pleasantly surprised to see that this small peek at an initial segment of the spectrum already yields strong evidence for the conjecture. In other words, it appears that the asymptotic regimine kicks in very early in the game. In the case of the Euclidean disc we have a better alternative method, since there the eigenvalues are given explicitly as squares of zeroes of Bessel functions of the first kind (D) or zeroes of derivatives of Bessel functions of the first kind (N). This allows us to go higher up in the spectrum with greater accuracy, and serves as a check on the size of the error obtained by the cruder method. Another check on error size is provided by doing the computations for the few triangles where the exact spectrum is known.
We now present the details concerning the constant $C$ in (1.3)
\begin{definition}
Let $K_{2}(S)$ denote the curvature of $S$, which is assumed to be constant, and let $K_{1}$ denote the curvature function on the smooth pieces of $\partial S$ as viewed from $S$.
\end{definition}
Further define
\begin{equation}
\tag{1.7} \varphi (\theta) = \frac{1}{24}\left(\frac{\pi}{\theta}-\frac{\theta}{\pi}\right) \newline
\end{equation}Then
\begin{equation}
\tag{1.8}C = C_1 + C_2 + C_3 \newline
\end{equation}
where
\begin{equation}
\tag{1.9}C_3 = \frac{1}{12\pi}A K_2(S) \\
\end{equation}
\begin{equation}
\tag{1.10}C_2 = \frac{1}{12\pi}\int_{\partial S} K_1 \\
\end{equation}
\begin{equation}
\tag{1.11}C_1 = \sum_j \varphi(\theta_j) \\
\end{equation}
in the case of $D$ or $N$ boundary conditions throughout, or
\begin{equation}
\tag{1.12}C_1 = \sum \varphi(\theta_j') + \sum (\varphi(2\theta_j'')-\varphi(\theta_j''))
\end{equation}
for mixed boundary conditions, where the corner angles are sorted into $\{\theta_j'\}$ where the same type of boundary condition is imposed on both sides of the corner, and $\{\theta_j''\}$ where opposite type boundary conditions are imposed on the two side arcs.\par We note that in [S] we also allowed a finite number of cone point singularities on $S$ with cone angles $\{\alpha_j \}$, and these contributed an additional term
\begin{equation}
\tag{1.13}\sum 2\varphi(\alpha_j/2) \newline
\end{equation}
to $C_1$. However, we are unable to do our computations if there are cone point sungularities, so we can't test the conjecture in such cases.
\begin{conjecture}\label{conj_curv_neg}
Assume the curvature of $S$ is zero or negative. Then there exists a uniformly almost periodic function $g$ such that
\begin{equation}
\tag{1.14}A(t) = g(t^{1/2})t^{-1/4}+O(t^{-1/2})
\end{equation}
as $t\rightarrow\infty$. In the case of zero curvature the almost periodic function $g$ has mean value zero.
\end{conjecture}
\begin{conjecture}\label{conj_curv_pos}
Assume the curvature of $S$ is positive. Then there exists a uniformly almost periodic funtion $g$ of mean value zero such that
\begin{equation}
\tag{1.15}A(t) = g(t^{1/2}) + O(t^{-1/2})
\end{equation}as $t\rightarrow 0$.
\end{conjecture}
We note that in \cite{strichartz2016} we conjectured that (1.14) and (1.15) are the first terms in an asymptotic expansion, but we are unable to test this here. Indeed, we cannot test the rate of decay in (1.14) and (1.15), since we don't know what $g$ should be. So basically we will observe that $t^{1/4}A(t)$ in case the case of conjecture \ref{conj_curv_neg} and $A(t)$ in the case conjecture \ref{conj_curv_pos} appear to be bounded functions of $t^2$ with mean value zero that could reasonably be almost periodic. Since almost periodicity is a global property, there is no way to test it by examining a small portion of the graph. We will observe, however, that there is no discernable difference between examples where the almost periodicity is known to be true, and all the other examples.
This paper is organized as follows: in section \ref{sec_test_ex} we perform our experimental computations for examples where the spectrum is known exactly, two Euclidean triangles and the Euclidean disc with Dirichlet and Neumann boundary conditions. We introduce the six part graphical display of data that will be used throughout the paper (except for the spherical surfaces in section \ref{sec_sph}). The reader will be able to see at a glance both confirmation of the predicted behavior and deviations due to computational error. In section \ref{sec_flat} we examine many examples of flat surfaces, including surfaces with mixed boundary conditions, surfaces that are not convex, and surfaces that are not simply connected. In section \ref{sec_hyp} we study hyperbolic surfaces, both triangles and discs. We see here experimental evidence that the conjecture for flat surfaces carries over into this case. In section \ref{sec_sph} we study spherical surfaces. Since the conjecture is different in this case (with no decay in $A(t)$) we use a five part graphical display. We give a discussion of all our results in section \ref{sec_discus}. We also mention the interesting question of the behavior of differences of consecutive eigenvalues. We have gathered data for all the examples studied here, and present a small selection of it. At this time we are not able to propose any conjectures.
The website \cite{murray2015results} contains the complete data on all the examples discussed here, as well as many other related examples. Additionally, a zip-file of all of our code is available for download. Automated scripts to generate each set of eigenvalues for an arbitrary number of refinements are available. However, please note that as much of each experiment was done through in-console manipulations there is no one unified script or function to generate the predicted eigenvalues or graphs once the initial refinements are performed.
\section{Some Test Examples}\label{sec_test_ex}
In this section we discuss our results for a few examples of surfaces where the spectrum is known exactly.
\subsection{Euclidean right isosceles triangle with Dirichlet boundary conditions}
The set of eigenvalues is $\pi^{2}(j^{2}+k^{2})$ for all pairs $(j,k)$ of distinct positive integers. We will normalize all eigenvalues by dividing by $\pi^2$ so that we are dealing with integer values. In Table 1 we show the data for the first 10 eigenvalues (the full table is on the website \cite{murray2015results}). In the first column we show the initial MATLAB computation of $\frac{\lambda_j}{\pi^2}$. In the next 6 columns we show the same value after successive refinements of the mesh. So the initial value of $\frac{\lambda_{10}}{\pi^2}$ is 44.931704, which is quite far from the true value of 37, but by the 6$^{th}$ refinement the approximation has improved to 37.001949. The next column is our predicted value obtained by fitting the data $x_{n}$ for refinements $n = 4,5,6$ to $x_{n} = x + cr^{n}$ and taking $x$ for the prediction. In this case the prediction is 37.000001. If we look further up in the spectrum we can see eigenvalues with multiplicity 2. For example $\frac{\lambda_{133}}{\pi^{2}} = \frac{\lambda_{134}}{\pi^{2}} = 377$. At refinement 4 the two values are 380.03292 and 380.1314. Not very close to each other and far off from the true value. At refinement 5 the two values are 377.7568 and 377.7816, closer to the true value but still not too close to each other. The predicted values are 376.9999 and 376.9998. Even though the order gets switched, the error is still quite acceptable.
\begin{table}[ht]\label{tab_euc_right_iso_tri}
\centering
\begin{tabular}{| l | l | l | l | l | l | l | l | }
\hline
& Initial & 1 & 2 & 3 & 4 & 5 & Predicted \\ \hline
1 & 5.13589 & 5.03479 & 5.0088 & 5.00221 & 5.000134 & 5.000138 & 5 \\ \hline
2 & 10.5735 & 10.1448 & 10.0364 & 10.0091 & 10.0006 & 10.0006 & 10 \\ \hline
3 & 13.9042 & 13.2281 & 13.0573 & 13.0143 & 13.0009 & 13.0009 & 13 \\ \hline
4 & 18.6783 & 17.4194 & 17.1051 & 17.0263 & 17.0016 & 17.0016 & 17 \\ \hline
5 & 22.3425 & 20.5806 & 20.1451 & 20.0363 & 20.0022 & 20.0022 & 20 \\ \hline
6 & 28.5140 & 25.8760 & 25.2190 & 25.0548 & 25.0034 & 25.0034 & 25 \\ \hline
7 & 29.7992 & 26.9473 & 26.2370 & 26.0593 & 26.0037 & 26.0037 & 26 \\ \hline
8 & 33.5825 & 30.1526 & 29.2891 & 29.0724 & 29.0045 & 29.0045 & 29 \\ \hline
9 & 40.6045 & 35.6485 & 34.4114 & 34.1029 & 34.0064 & 34.0064 & 34 \\ \hline
10 & 44.9317 & 38.9934 & 37.4981 & 37.1246 & 37.0078 & 37.0077 & 37 \\ \hline
... & ... & ... & ... & ...& ... & ... & ...\\ \hline
133 & 0 & 0 & 0 & 389.656 & 377.756 & 377.756 & 376.9999 \\ \hline
134 & 0 & 0 & 0 & 398.873 & 377.781 & 377.781 & 376.9998 \\ \hline
\end{tabular}
\caption{Euclidean right isosceles triangle}
\end{table}
In Figure 1 we show the graphs of
\begin{enumerate}
\item $N(t)$
\item $D(t) = N(t) - \widetilde{N}(t)$
\item $A(t) = \frac{1}{t}\int_{0}^{t}D(s)ds$
\item $t^{\frac{1}{4}}A(t)$
\item $t^{\frac{1}{4}}A(t^{2})$
\item $\frac{1}{t-a}\int_{a}^{t}s^{\frac{1}{2}}A(s^{2})ds$ for $a = $ the highest predicted eigenvalue divided by 16, removing the first $\frac{1}{4}$ of graph 5 from figuring into graph 6 and eliminating potential early extreme values so that it converges to 0 more quickly.
\end{enumerate}
We will use this set of six graphs for each Euclidean and hyperbolic region which we analyze. The $x$-scales of the first four graphs were picked to use all predicted eigenvalues with an acceptable level of error, usually between the first 120 and 150 eigenvalues (the number used is in the $y$-axis of the first graph). When the true values are known, we use those and frequently use more than 150, as in figure 2 where the first 1000 eigenvalues are used. The scale of the $x$-axis in the fifth and sixth graphs is approximately the square root of the scale of the $x$-axis of the first four graphs.
For Figure 1 We used the exact values for the first approximately 150 eigenvalues. A quick look at these graphs yields some simple observations. The graph 1 shows that $N(t)$ grows approximately linearly, while 2 shows that $D(t)$ grows at a relatively slow rate. the graph 3 suggests that $A(t)$ is converging to 0 at a slow rate, while graph 4 confirms that $O(t^{-\frac{1}{4}})$ is a plausible decay rate. The function in graph 5 is known to be converging to an almost periodic function g(t) (see \cite{strichartz2016}), but this is not apparent from the graph. Presumably the almost periods are too large to show up in the range of data we have plotted. On the other hand, graph 6 gives strong evidence that the almost periodic function has mean value 0.
\begin{figure}[ht]\label{count_euc_right_iso_tri}
\centering
\includegraphics[width = 0.8\textwidth]{EucDRightIsoTriCountingGraphs}\\
\caption{Euclidean right isosceles triangle}
\centering
\end{figure}
\subsection{The Euclidean equilateral triangle with Dirichlet boundary conditions}
Here the eigenvalues are known to be the values $(\frac{4}{3}\pi)^{2}(j^{2} + k^{2} +jk)$ for the positive integers $j,k$. This typically produces multiplicity 1 when $j = k$ and multiplicity 2 when $j \neq k$. Here we only used 5 refinements. Table 2 and Figure 2 show the same information for this example as before. For $\lambda_{119} = \lambda_{120} = 219$ our predicted values are 219.0009455 and 219.0005973 while on the $5^{th}$ refinement they are 219.4644686 and 219.4662058. The qualitative features of Figure 2 are much the same as that of Figure 1.
\begin{figure}[ht]\label{count_euc_equ_tri}
\centering
\includegraphics[width=0.8\textwidth]{EucDEquTriCountingGraphs}\\
\caption{Euclidean equilateral triangle}
\end{figure}
\begin{table}[ht]\label{tab_euc_equ_tri}
\centering
\begin{tabular}{| l | l | l | l | l | l | l | l | }
\hline
& Initial & 1 & 2 & 3 & 4 & 5 & Predicted \\ \hline
1 & 3.08733 & 3.02224 & 3.005611 & 3.001408 & 3.000088 & 3.000088 & 3.000001 \\ \hline
2 & 7.44103 & 7.11170 & 7.028133 & 7.007054 & 7.000441 & 7.000441 & 7.000003 \\ \hline
3 & 7.49871 & 7.12432 & 7.03113 & 7.007791 & 7.000487 & 7.000487 & 7.000003 \\ \hline
4 & 13.4606 & 12.3613 & 12.09015 & 12.02254 & 12.00141 & 12.00141 & 12.00001 \\ \hline
5 & 14.6422 & 13.4084 & 13.10213 & 13.02555 & 13.0016 & 13.0016 & 13.00001 \\ \hline
6 & 14.7099 & 13.4240 & 13.10604 & 13.02654 & 13.00166 & 13.00166 & 13.00001 \\ \hline
7 & 22.3866 & 19.8479 & 19.2122 & 19.05312 & 19.00332 & 19.00332 & 19.00002 \\ \hline
8 & 22.7581 & 19.9425 & 19.23499 & 19.05874 & 19.00367 & 19.00367 & 19.00002 \\ \hline
9 & 25.3291 & 22.0593 & 21.26293 & 21.06566 & 21.0041 & 21.0041 & 21.00002 \\ \hline
10 & 25.3551 & 22.0705 & 21.26635 & 21.06656 & 21.00416 & 21.00416 & 21.00002 \\ \hline
\end{tabular}
\caption{Euclidean equilateral triangle}
\end{table}
\subsection{Euclidean disc with Dirichlet boundary conditions}
We take the radius to be one since all discs have eigenvalues that scale by the radius. In this case the eigenvalues are the squares of the zeroes of the Bessel functions $J_{k}$ for nonnegative integers $k$ with multiplicity one for $k = 0$ and multiplicity two for $k \geq 1$. It is possible to get accurate values of these zeros so we have exact values for the first 660 eigenvalues. In this example $\widetilde{N}(t) = \frac{1}{4}t - \frac{1}{2}\sqrt{t} + \frac{1}{6}$.
Figure 3 displays the same graphs as before using the exact values. We note that graph 5 is just as plausibly an asymptotic almost periodic function as the same graphs in the triangle cases where we know the function is asymptotically almost periodic. On the other hand, graph 6 shows a much slower rate of decay than in the triangle cases. It is still plausible that this gives supportive evidence that the presumed almost periodic function has mean value zero, but the evidence is not decisive.
\begin{figure}[ht]\label{count_euc_disc}
\centering
\includegraphics[width=0.8\textwidth]{EucDDiscCountingGraphs}\\
\caption{Euclidean disk (Dirichlet conditions)}
\end{figure}
\subsection{Euclidean disk with Neumann boundary conditions (again with radius one)}
In this case $\lambda = (z_{n})^{2}$ where $J'_{k}(z_{n}) = 0$. Here we were able to obtain the exact values for the first 550 eigenvalues. Figure 4 shows the corresponding data with $\widetilde{N}(t) = \frac{1}{4}t + \frac{1}{2}\sqrt{t} + \frac{1}{6}$. The qualitative features observed for the previous example are evident here as well.
\begin{figure}\label{count_euc_n_disc}
\includegraphics[width=0.8\textwidth]{EucNDiscCountingGraphs}\\
\caption{Euclidean disk (Neumann conditions)}
\end{figure}
\section{Flat Surfaces}\label{sec_flat}
In this section we discuss examples of polygonal surfaces in Euclidean space. In particular we examined examples of nonconvex surfaces, surfaces with angles exceeding $\pi$, and surfaces that are not simply connected. There are still more examples on the website \cite{murray2015results}. For each example we give the counting function $\widetilde{N}(t)$ and the analog of Figure 1
\subsection{Triangle with Dirichlet boundary conditions}
The angles are $\theta_1 =\frac{\pi}{4}, \theta_2 = \frac{\pi}{5}, \theta_3 = \frac{11\pi}{20}$ and
$$\widetilde{N}(t) = \frac{\sin{\theta_{2}}\sin{\theta_{1}}}{8\pi\sin{\theta_{3}}}t-\frac{\frac{\sin{\theta_1}}{\sin{\theta_3}}\sqrt{t}+\frac{\sin{\theta_2}}{\sin{\theta_3}}+1}{4\pi}+\frac{9}{22}$$
Here we used the first 130 calculated eigenvalues, as accuracy begins to break down after that point. The graphs in Figure 5 are analogous to those in Figure 1 and show similar behavior.
\begin{figure}[ht]\label{count_euc_tri2}
\includegraphics[width=0.8\textwidth]{EucDTri2CountingGraphs}\\
\caption{Triangle with Dirichlet boundary conditions}
\end{figure}
\subsection{Triangle with Neumann boundary conditions}
This is the same triangle as above, with
$$\widetilde{N}(t) = \frac{\sin{\theta_{2}}\sin{\theta_{1}}}{8\pi\sin{\theta_{3}}}+\frac{\frac{\sin{\theta_1}}{\sin{\theta_3}}t+\frac{\sin{\theta_2}}{\sin{\theta_3}}+1}{4\pi}\sqrt{t}+\frac{9}{22}$$
Here we used the first 150 calculated eigenvalues. The graphs in Figure 6 are analogous to Figure 5 and display the same behavior, except that in graph six of Figure 6 the graph is decreasing to 0, whereas it is increasing to zero in Figure 5. This difference is a result of the boundary conditions and is mirrored in all graphs of the same shape under Dirichlet and Neumann boundary conditions.
\begin{figure}[ht]\label{count_euc_n_tri2}
\centering
\includegraphics[width=0.8\textwidth]{EucNTri2CountingGraphs}\\
\caption{Triangle with Neumann boundary conditions}
\end{figure}
\subsection{Triangle with mixed boundary conditions}
With the same triangle, we impose Dirichlet boundary conditions on $s_1$ and $s_2$ and Neumann boundary conditions on $s_3$. Here
$$\widetilde{N}(t) = \frac{\sin{\theta_{2}}\sin{\theta_{1}}}{8\pi\sin{\theta_{3}}}t-\frac{\frac{\sin{\theta_1}}{\sin{\theta_3}}\sqrt{t}-\frac{\sin{\theta_2}}{\sin{\theta_3}}+1}{4\pi}\sqrt(t)+\frac{189}{110}$$
and the resulting graphs can be seen in Figure 7, which display similar behavior to those in Figures 5 and 6.
\begin{figure}[ht]\label{count_euc_m_tri2}
\centering
\includegraphics[width=\textwidth]{EucMixed2CountingGraphs}\\
\caption{Triangle with mixed boundary conditions}
\end{figure}
\subsection{Arrowhead with Dirichlet boundary conditions (see Figure 8)}
\begin{figure}[ht]\label{fig_arrowhead}
\centering
\includegraphics[width=0.8\textwidth]{arrowhead}\\
\caption{Arrowhead region}
\end{figure}
\iffalse We define an arrowhead to a Euclidean quadrilateral with one concave angle that is formed by joining two triangles with obtuse angles that have a pair of sides of equal length that are adjacent to the obtuse angle so that the concave angle of the resulting quadrilateral has a measure equal to the sum of the measures of the obtuse angles. The angle opposite the concave angle is the sum of the two non-obtuse angles adjacent to the sides that were joined together, and the remaining two angles are the remaining angles in each of the triangles. The sides of the arrowhead are then the sides of the triangles excluding the side they were joined on.\fi Here the sides are $s_1 = \frac{\sqrt{13}}{2}, s_2 = \frac{\sqrt{2}}{2}, s_3 = \sqrt{2}, s_4 = \sqrt{5}$. The angle $\theta_i$ joins sides $s_i$ and $s_{i-1}$; $\theta_1$ joins $s_1$ and $s_4$. The angle measures are $\theta_1 = \sin^{-1}{\frac{1}{\sqrt{13}}}+\sin^{-1}{\frac{1}{\sqrt{5}}}, \theta_2 = \cos^{-1}{(\frac{s^{2}_1+s^{2}_2-1}{2s_{1}s_{2}})}, \theta_4 = \cos^{-1}{(\frac{s^{2}_3+s^{2}_3-1}{2s_{3}s_{4}})}, \theta_3 = 2\pi-\theta_1-\theta_2-\theta_4$, and $\theta_3>\pi$. We then have
$$a = \sqrt{\frac{s_1+s_2+1}{2}(s_1+\frac{s_2}{2}+1)(\frac{s_1}{2}+s_2+1)(s_1+s_2+\frac{1}{2})}$$
$$b = \sqrt{\frac{s_3+s_4+1}{2}(s_3+\frac{s_4}{2}+1)(\frac{s_3}{2}+s_4+1)(s_3+s_4+\frac{1}{2})}$$
and
$$\widetilde{N}(t) = \frac{a+b}{4\pi}t-\frac{\sum_{i=1}^{4}{s_i}}{4\pi}\sqrt{t}+\frac{\sum_{i=1}^{4}{\theta_i}}{24}$$
Figure 9 is analogous to the graphs we have seen before, and we can see that it displays the same behavior. Note that this region is not convex and contains an angle greater than $\pi$.
\begin{figure}[ht]\label{count_euc_arrowhead}
\includegraphics[width=0.8\textwidth]{mEucDArr1CountingGraphs}\\
\caption{Arrowhead with Dirichlet boundary conditions}
\end{figure}
\subsection{Region between triangles with Dirichlet boundary conditions (see Figure 10)}
\begin{figure}[ht]\label{fig_euc_tri_in_tri}
\centering
\includegraphics[width=0.8\textwidth]{TriInTri}\\
\caption{Region between triangles}
\end{figure}
This is not simply connected. The angles of the interior triangle are viewed from the surface and hence are the exterior angles. We need to keep the vertices of the inner triangle a reasonable distance from the edges of the outer triangle in order to have reasonable accuracy in computing eigenvalues. Note that the formula $\widetilde{N}(t) = \frac{3\sqrt{3}}{64\pi}t-\frac{9}{8\pi}{\sqrt{t}}+\frac{1}{15}$ is independent of the location and orientation of the inner triangle. The graphs seen in Figure 11 are analagous to those seen before and display similar results.
\begin{figure}[ht]\label{count_euc_tri_in_tri}
\centering
\includegraphics[width=0.8\textwidth]{EucDTrinTri1CountingGraphs}\\
\caption{Region between triangles with Dirichlet boundary conditions}
\end{figure}
\subsection{Regular pentagon with Dirichlet boundary conditions}
Here $\widetilde{N}(t) = \frac{\sqrt{5(5+2\sqrt{5}}}{16\pi}t-\frac{5}{4\pi}\sqrt{t}+\frac{2}{9}$. Note that there are many eigenvalues of multiplicity two, due to the $D_5$ symmetry group. This gives us a reasonable tool for assessing the accuracy of our computations (since MATLAB does not select symmetric triangulations). The first 10 eigenvalues are displayed in Table 3.
\begin{table}[ht]\label{tab_euc_pent}
\centering
\begin{tabular}{| l | l | l | l | l | l | l | l | }
\hline
& Initial & 1 & 2 & 3 & 4 & 5 & Predicted \\ \hline
1 & 11.1479 & 11.03526 & 11.00624 & 10.99889 & 10.99704 & 10.99658 & 10.99643 \\ \hline
2 & 28.8006 & 28.04273 & 27.85074 & 27.80238 & 27.79025 & 27.78721 & 27.7862 \\ \hline
3 & 28.8171 & 28.04763 & 27.85201 & 27.8027 & 27.79033 & 27.78723 & 27.7862 \\ \hline
4 & 52.4110 & 50.06061 & 49.47096 & 49.323 & 49.28594 & 49.27667 & 49.27358 \\ \hline
5 & 52.5710 & 50.10611 & 49.48295 & 49.32606 & 49.28671 & 49.27686 & 49.27359 \\ \hline
6 & 61.5873 & 58.21567 & 57.37533 & 57.16477 & 57.11204 & 57.09885 & 57.09447 \\ \hline
7 & 84.9386 & 78.95964 & 77.47257 & 77.1007 & 77.00764 & 76.98437 & 76.97664 \\ \hline
8 & 85.0667 & 78.9798 & 77.47749 & 77.10193 & 77.00795 & 76.98445 & 76.97664 \\ \hline
9 & 99.4387 & 91.724 & 89.80689 & 89.32718 & 89.2071 & 89.17706 & 89.16708 \\ \hline
10 & 100.119 & 91.89246 & 89.84908 & 89.33775 & 89.20975 & 89.17773 & 89.16708 \\ \hline
\end{tabular}
\caption{Regular pentagon with Dirichlet boundary conditions}
\end{table}
\begin{figure}[ht]\label{count_euc_pent}
\includegraphics[width=0.8\textwidth]{EucDPentCountingGraphs}\\
\caption{Regular pentagon with Dirichlet boundary conditions}
\end{figure}
\subsection{Regular hexagon with Dirichlet boundary conditions}
Here $\widetilde{N}(t) = \frac{3\sqrt{3}}{8\pi}t-\frac{3}{\pi}\sqrt{t}+\frac{5}{24}$. Note that all Dirichlet eigenfunctions of the equilateral triangle extend by odd reflections to Dirichlet eigenfunctions of the hexagon with the same eigenvalue. In our table of eigenvalues we therefore divide by $(\frac{4}{3}\pi)^{2}$ so that these eigenvalues become integers. This gives us an accuracy check. We have a $D_{6}$ symmetry group so that most eigenvalues have multiplicity two.
\begin{table}[ht]\label{tab_euc_hex}
\centering
\begin{tabular}{| l | l | l | l | l | l | l | l | }
\hline
& Initial & 1 & 2 & 3 & 4 & 5 & Predicted \\ \hline
1 & 0.413621 & 0.409345 & 0.408200 & 0.407905 & 0.40783 & 0.40781 & 0.40781 \\ \hline
2 & 1.069612 & 1.042747 & 1.035759 & 1.03398 & 1.03353 & 1.03342 & 1.03338 \\ \hline
3 & 1.07095 & 1.043083 & 1.035843 & 1.034001 & 1.03354 & 1.03342 & 1.03338 \\ \hline
4 & 1.968578 & 1.880006 & 1.857227 & 1.851466 & 1.85002 & 1.84966 & 1.84953 \\ \hline
5 & 1.969445 & 1.880229 & 1.857298 & 1.851485 & 1.85002 & 1.84966 & 1.84953 \\ \hline
6 & 2.288059 & 2.175413 & 2.146518 & 2.139204 & 2.13737 & 2.13691 & 2.13675 \\ \hline
7 & 2.976487 & 2.782209 & 2.731673 & 2.718854 & 2.71563 & 2.71482 & 2.71455 \\ \hline
8 & 3.328669 & 3.08177 & 3.020466 & 3.005122 & 3.00128 & 3.0003 & 3 \\ \hline
9 & 3.808021 & 3.523318 & 3.450285 & 3.431787 & 3.42714 & 3.4260 & 3.42558 \\ \hline
10 & 3.816109 & 3.524926 & 3.450607 & 3.43186 & 3.42715 & 3.4260 & 3.42558 \\ \hline
\end{tabular}
\caption{Regular hexagon with Dirichlet boundary conditions}
\end{table}
\begin{figure}[ht]\label{count_euc_hex}
\centering
\includegraphics[width=0.8\textwidth]{EucDHexCountingGraphs}\\
\caption{Regular hexagon with Dirichlet boundary conditions}
\end{figure}
\subsection{6-regular star with Dirichlet boundary conditions}
Here $\widetilde{N}(t) = \frac{s\sqrt{3}}{4\pi}t -\frac{3}{\pi}\sqrt{t}+\frac{25}{48}$. As in the case of the hexagon, Dirichlet eigenfunctions of the equilateral triangle extend by odd reflection, and there is a $D_6$ symmetry group. Therefore we again divide our table by $(\frac{4}{3}\pi)^{2}$ so that these eigenvalues become integers. The other eigenvalues of the hexagon do not, however, extend to the 6-regular star.
\begin{figure}[ht]\label{count_euc_star}
\centering
\includegraphics[width=0.8\textwidth]{EucDSixStarCountingGraphs}\\
\caption{6-regular star with Dirichlet boundary conditions}
\end{figure}
\section{Hyperbolic Surfaces}\label{sec_hyp}
In this section we discuss examples of surfaces in the hyperbolic plane of constant negative curvature $-1$. We use the upper half-plane model. In this model the Laplacian is given by
\begin{equation}
\tag{4.1} \Delta = y^2(\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial y^2})
\end{equation}
so the eigenvalue problem
\begin{equation}
\tag{4.2} -\Delta u = \lambda u
\end{equation}
is transformed into
\begin{equation}
\tag{4.3} -(\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial y^2})u(x,y) = \lambda y^{2}u(x,y)
\end{equation}
and we used MATLAB to solve $(4.3)$ on the surfaces with the appropriate boundary conditions. For simplicity we restricted our attention to Dirichlet boundary conditions, and our surfaces were either disks or triangles.
To describe triangles we recall that geodesics in the upper half-plane model are either vertical half lines or half circles that intersect the x-axis perpendicularly. Without loss of generality we may take one side of the triangle to lie along the y-axis. Specifically, the triangle will have vertices $(0,y_1), (0,y_2)$ and $(x_3,y_3)$, seen in Figure 15 as points C, A, and B, respectively.
\begin{figure}[ht]\label{fig_hyp_tri}
\centering
\includegraphics[width=0.5\textwidth]{HypTri}\\
\caption{Hyperbolic Triangle}
\end{figure}
The two boundary circles are $y^2 + (x-a_j)^2 = r_{j}^2$ for $j = 1,2$, and $(x_3,y_3)$ lies at the intersection of these circles, so $x_3 = \frac{r_2^2-r_1^2-a_2^2+a_1^2}{s(a_1-a_2)}, y_3 = \sqrt{r_1^2-(x_3-a_1)^2}$, and also $y_j = \sqrt{r_j^2-a_j^2}$ for $j = 1,2$.
Since the model is conformal, the angles are the same as the Euclidean angles, so we have $\alpha_1 = \frac{\pi}{2}-\tan^{-1}{(\frac{a_1}{y_1})}, \alpha_2 = \frac{\pi}{2} - \tan^{-1}{(\frac{-a_2}{y_2})}$ and \newline{} $\alpha_3 = \tan^{-1}{(\frac{a_1-x_3}{y_3})}+\tan^{-1}{(\frac{x_3-a_a}{y_3})}$. The lengths of the opposite sides are $L_j = \frac{1}{2}log(\frac{r_j+a_j}{r_j-a_j})-\frac{1}{2}log(\frac{r_j-x_x+a_j}{r_j+x_3+a_j})$ for $j = 1,2$ and $L_3 = log\frac{y_2}{y_1}$. The area of the triangle is
$$A=\int\!\!\!\int_T \frac{dxdy}{y^2}=\frac{1}{r_2}(\cos^{-1}(\frac{-a_2}{r_2})-\cos^{-1}(\frac{x_3-a_2}{r_2}))+\frac{1}{r_1}(\cos^{-1}(\frac{x_3-a_1}{r_1})-\cos^{-1}(\frac{-a_1}{r_1}))$$ Thus we have
\begin{equation}
\tag{4.4} \widetilde{N}(t)=\frac{1}{4\pi}At-\frac{1}{4\pi}(L_1+L_2+L_3)t^{\frac{1}{2}}+C
\end{equation}
for
\begin{equation}
\tag{4.5} C=-\frac{1}{12\pi}A+\frac{1}{24}\sum_{j=0}^{2}(\frac{\pi}{\alpha_j}-\frac{\alpha_j}{\pi})
\end{equation}
Of course everything may be expressed entirely in terms of the angles, since the angles determine the triangle. Thus the hyperbolic law of cosines yields
\begin{equation}
\tag{4.6} L_i=\cosh^{-1}(\frac{\cos\alpha_j\cos\alpha_k+\cos\alpha_i}{\sin\alpha_j\sin\alpha_k})
\end{equation}
for $(i,j,k)$ any permutation of $(1,2,3)$, and the angle defect formula yields
\begin{equation}
\tag{4.7} A=\pi-(\alpha_1+\alpha_2+\alpha_3)
\end{equation}
\subsection{Hyperbolic Equilateral Triangles with Dirichlet boundary conditions}
We take $\alpha_1 = \alpha_2 = \alpha_3 = \frac{\pi}{k}$ for $k$ an integer, $k \geq 4$. These triangles tesselate the hyperbolic plane. When $k$ is even we may take odd reflections of the Dirichlet eigenfunctions to see that we are generating a subset of the collection of eigenfunctions on the hyperbolic closed manifolds $\Gamma \backslash SL(2,\mathbb{R})/SO(2)$ for the appropriate discrete subgroup $\Gamma$.
We show the results for $k = 4,6$ in Figures 16 and 17, respectively. Already for $k = 6$ the accuracy of our approximations begins to degrade. The website shows complete data for $k = 4,5,6,7$.
\begin{figure}[ht]\label{count_hyp_equ_tri4}
\centering
\includegraphics[width=0.8\textwidth]{HypDEquTri4CountingGraphs}\\
\caption{Hyperbolic equilateral triangle with $k = 4$}
\end{figure}
\begin{figure}[ht]\label{count_hyp_equ_tri6}
\centering
\includegraphics[width=\textwidth]{HypDEquTri6CountingGraphs}\\
\caption{Hyperbolic equilateral triangle with $k = 6$}
\end{figure}
\subsection{General Hyperbolic Triangles with Dirichlet boundary conditions}
We present two hyperbolic triangles here with arbitrary measurements. Triangles are specified by a label $(u,v,s)$ which correspond to the measurements in Figure 15. Figure 18 corresponds to $(5,10,11)$ and Figures 19 corresponds to $(3,4,6)$. Note that in the fifth and sixth counting graphs in figures 18 and 19 we begin to lose accuracy more quickly than we do in the Euclidean results. This is not unique to the arbitrary triangles, as it is present in both the hyperbolic equilateral triangles and hyperbolic discs, but it is especially noticeable here.
The complete results for more arbitrary hyperbolic triangles are shown on the website.
\begin{figure}[ht]\label{count_hyp_tri1}
\centering
\includegraphics[width=0.8\textwidth]{HypDTri1CountingGraphs}\\
\caption{First hyperbolic triangle}
\end{figure}
\begin{figure}[ht]\label{count_hyp_tri2}
\centering
\includegraphics[width=0.8\textwidth]{HypDTri2CountingGraphs}\\
\caption{Second hyperbolic triangle}
\end{figure}
\subsection{Hyperbolic Discs with Dirichlet boundary conditions}
We are able to calculate eigenvalues on a disc of hyperbolic radius $ R$ by calculating eigenvalues for a Euclidean disc of radius $r=\frac{e^{2R}}{2}$ centered at $\frac{e^{2R}}{2}+1$. The resulting disc has area $A=4\pi\sinh^{2}(\frac{R}{2})$ and circumference $C = 2\pi\sinh(R)$ We then have $\widetilde{N}(t)=\frac{A}{4\pi}t-\frac{C}{r\pi}\sqrt{t}+\frac{1}{6}$. We can see from figures 20 and 21 that this $\widetilde{N}(t)$ appears to be strongly supported, though as is the case of the hyperbolic triangles, we begin to lose accuracy in the predicted eigenvalues more quickly here than in the Euclidean case. This becomes a particular issue for the MATLAB PDE solver in the case of discs however, as the radius of the Euclidean disc we solve for eigenvalues on grows exponentially with the hyperbolic radius, leading to longer computation times as mesh with an exponentially growing number of points is needed to estimate the values. The discs in Figures 20 and 21 have R = 1, 1/2, respectively.
\begin{figure}[ht]\label{count_hyp_disc1}
\centering
\includegraphics[width=0.8\textwidth]{HypDDisc1CountingGraphs}\\
\caption{Hyperbolic disk with $R = 1$}
\end{figure}
\begin{figure}[ht]\label{count_hyp_disc_half}
\centering
\includegraphics[width=0.8\textwidth]{HypDDiscHalfCountingGraphs}\\
\caption{Hyperbolic disk with $R = \frac{1}{2}$}
\end{figure}
\section{Spherical Surfaces}\label{sec_sph}
In this section we discuss examples of surfaces in the unit sphere (curvature +1). We use stereographic projection, placing the center of the sphere at $(0,0,1)$ and projecting from $(0,0,2)$ onto the $(u,v)$ plane by $u = \frac{2x}{2-z}, v = \frac{2y}{2-z}$. The equator is mapped to the circle $u^2+v^2=4$, great circles through the poles are mapped to the lines through the origin, and other great circles are mapped to circles intersecting $u^2+v^2=4$ at two antipodal points. If we fix parameters to $t>0$ and $\beta$ then these circles are given by $(u-t\sin{\beta})^2+(v+t\cos{\beta})^2=t^2+4$ (intersecting $u^2+v^2=4$ at $\pm (2\cos{\beta},2\sin{\beta})$).
We will consider triangles with vertices $(u_1,0), (u_2,0)$ and $(u_3,v_3)$, with one edge along the $u$-axis and two edges being arcs of circles $(u-t_j\sin{\beta_j})^2+(v+t_j\cos{\beta_j})^2=t_j^2+4$ for $j=1,2$. The angles of the triangle are given by
$$\alpha_1={\tan}^{-1}{(\frac{-u_1+t_1\sin{\beta_1}}{t_1\cos{\beta_1}})}$$
$$\alpha_2=\pi-{\tan}^{-1}{(\frac{-u_2+t_2\sin{\beta_2}}{t_2\cos{\beta_2}})}$$
$$\alpha_3={\tan}^{-1}{(\frac{-u_3+t_2\sin{\beta_2}}{v_3+t_1\cos{\beta_2}})-{\tan}^{-1}{(\frac{-u_3+t_1\sin{\beta_1}}{v_3+t_1\cos{\beta_1}})}}$$
The angles completely determine the triangle. The lengths of the sides are given by the spherical law of cosines\\
\begin{equation}
\tag{5.1} L_i= \cos^{-1}{(\frac{\cos{\alpha_i}+\cos{\alpha_j}\cos{\alpha_k}}{\sin{\alpha_j}\sin{\alpha_k}})}\\
\end{equation}
for $(i,j,k)$ a permutation of $(1,2,3)$, and the area is given by the angle defect\\
\begin{equation}
\tag{5.2} A = (\alpha_1+\alpha_2+\alpha_3) - \pi
\end{equation}
The Laplacian is given by
\begin{equation}
\tag{5.3} \Delta = (\frac{u^2+v^2+4}{4})^2(\frac{\partial^2}{\partial u^2}+\frac{\partial^2}{\partial v^2})
\end{equation}
\subsection{Spherical Equilateral Right Triangle with Dirichlet boundary conditions}
This triangle serves as our main accuracy check for our calculated eigenvalues in spherical space. This is because it is a region for which the eigenvalue spectrum is known: the $i^{th}$ distinct eigenvalue is equal to $4i^2+6i+2$ and has multiplicity $i$. We can see the first eigenvalues in Table 5 below.
\begin{table}[ht]\label{tab_sph_equ_right_tri}
\centering
\begin{tabular}{| l | l | l | l | l | l | l | l | l | }
\hline
& Initial & 1 & 2 & 3 & 4 & 5 & Predict & True\\ \hline
1 & 12.1683 & 12.0426 & 12.0107 & 12.0027 & 12.0007 & 12.0002 & 12 & 12 \\ \hline
2 & 30.9285 & 30.2355 & 30.0593 & 30.0148 & 30.0037 & 30.0009 & 30 & 30 \\ \hline
3 & 31.1082 & 30.2803 & 30.0704 & 30.0176 & 30.0044 & 30.0011 & 30 & 30 \\ \hline
4 & 58.8956 & 56.7339 & 56.1845 & 56.0462 & 56.0116 & 56.0029 & 56 & 56 \\ \hline
5 & 59.6055 & 56.9108 & 56.2287 & 56.0573 & 56.0143 & 56.0036 & 56.0000 & 56 \\ \hline
6 & 59.8717 & 56.9775 & 56.2454 & 56.0615 & 56.0154 & 56.0038 & 56.0000 & 56 \\ \hline
7 & 97.1655 & 91.8111 & 90.4552 & 90.1140 & 90.0285 & 90.0071 & 90.0000 & 90 \\ \hline
8 & 98.8889 & 92.2347 & 90.5603 & 90.1402 & 90.0351 & 90.0088 & 90.0000 & 90 \\ \hline
9 & 99.7452 & 92.4530 & 90.6150 & 90.1539 & 90.0385 & 90.0097 & 90.0000 & 90 \\ \hline
10 & 100.1208 & 92.5527 & 90.6404 & 90.1603 & 90.0401 & 90.0100 & 90.0000 & 90 \\ \hline
\end{tabular}
\caption{Spherical equilateral right triangle}
\end{table}
In the graphical data displayed in Figure 22 (and subsequent figures in this section) we show
\begin{enumerate}
\item $N(t)$
\item $D(t) = N(t) - \widetilde{N}(t)$
\item $A(t) = \frac{1}{t}\int_{0}^{t}D(s)ds$
\item $A(t^{2})$
\item $\frac{1}{t-a}\int_{a}^{t}s^{\frac{1}{2}}A(s^{2})ds$ for $a = $ the highest predicted eigenvalue divided by 16, removing the first $\frac{1}{4}$ of graph 4 from figuring into graph 5 and eliminating potential early extreme values so that it converges to 0 more quickly.
\end{enumerate}
The scales for each of these graphical displays are the same as the scales in the analagous set of six graphs we used for Euclidean and hyperbolic regions.
\begin{figure}[ht]\label{count_sph_equ_right_tri}
\centering
\includegraphics[width=0.8\textwidth]{SphEquRightDTriCountingGraphs}\\
\caption{Spherical equilateral right triangle}
\end{figure}
\subsection{General Spherical Triangle with Dirichlet boundary conditions}
We will now present a spherical triangle with arbitrary measurements. Spherical triangles are specified by a label $(t_1,\beta_1,t_2,\beta_2)$ which correspond to the measurements described above. Here we have $\widetilde{N}(t)=\frac{Area}{4\pi}t-\frac{\Sigma_{i=1}^3{L_i}}{4\pi}\sqrt{t}+(\frac{Area}{12\pi}+\frac{1}{24}\Sigma_{i=1}^3{(\frac{\pi}{\alpha_i}-\frac{\alpha_i}{\pi})})$ The triangle corresponding to Figure 23 is $(-1.5,\frac{\pi}{4},-2,-\frac{\pi}{6})$. Note in the figure that while the accuracy of our calculated eigenvalues suffers some decay, it does so at a slower rate than in the Hyperbolic surfaces.
\begin{figure}[ht]\label{count_sph_tri2}
\centering
\includegraphics[width=0.8\textwidth]{SphDTri2CountingGraphs}\\
\caption{General spherical triangle}
\end{figure}
\subsection{Spherical Disc with Dirichlet boundary conditions}
The final case we wish to present is that of the spherical disc. As with the spherical triangles, we use stereographic projection to solve for eigenvalues on a Euclidean disc using (5.3) as the Laplacian. A Euclidean disc centered at the origin with radius $r$ corresponds to a spherical disc with radius $R=\frac{2\sin{r}}{1+\cos{r}}$. The spherical disc then has area $A=2\pi(1-\cos{r})$ and circumference $C=2\pi\sin{r}$. This gives us $\widetilde{N}(t)=\frac{A}{4\pi}t-\frac{C}{4\pi}\sqrt{t}+\frac{1}{6}$. Here the hemisphere, $r=\frac{\pi}{2}$, serves as a test case for the accuracy of our predicted eigenvalues: it has a known eigenvalue spectrum such that the $n^{th}$ unique eigenvalue is $n(n+1)$ and has multiplicity $n$. Table 6 and Figures 24 strongly support the accuracy of $\widetilde{N}(t)$. Figure 25 is the corresponding graphs for the disc with $r=\frac{\pi}{4}$, and likewise strongly supports our predicted eigenvalues. Unfortunately, as can be seen in graphs 3, 4, and 5 of 25, our accuracy again begins to drop after a point, but they are accurate enough to support our calculation of $\widetilde{N}(t)$.
\begin{table}[ht]\label{tab_sph_hemisphere}
\centering
\begin{tabular}{| l | l | l | l | l | l | l | l | l | }
\hline
& Initial & 1 & 2 & 3 & 4 & 5 & Predict & True\\ \hline
1 & 2.03144 & 2.00787 & 2.00197 & 2.00049 & 2.00012 & 2.00003 & 2 & 2 \\ \hline
2 & 6.19516 & 6.04907 & 6.0123 & 6.00308 & 6.00077 & 6.00019 & 6 & 6 \\ \hline
3 & 6.20475 & 6.05152 & 6.0129 & 6.00324 & 6.00080 & 6.00020 & 6 & 6 \\ \hline
4 & 12.7272 & 12.1822 & 12.0457 & 12.0114 & 12.0029 & 12.0007 & 12.0000 & 12 \\ \hline
5 & 12.7330 & 12.1830 & 12.0458 & 12.0115 & 12.0029 & 12.0007 & 12.0000 & 12 \\ \hline
6 & 12.8645 & 12.2143 & 12.0535 & 12.0134 & 12.0034 & 12.0008 & 12.0000 & 12 \\ \hline
7 & 21.8055 & 20.4515 & 20.1130 & 20.0283 & 20.0071 & 20.0018 & 20.0000 & 20 \\ \hline
8 & 21.8512 & 20.4617 & 20.1155 & 20.0289 & 20.0072 & 20.0018 & 20.0000 & 20 \\ \hline
9 & 22.2046 & 20.5453 & 20.1362 & 20.0341 & 20.0085 & 20.0021 & 20.0000 & 20 \\ \hline
10 & 22.3183 & 20.5798 & 20.1451 & 20.0363 & 20.0091 & 20.0023 & 20.0000 & 20 \\ \hline
252 & 0 & 0 & 0 & 530.180 & 512.023 & 507.504 & 505.985 & 506 \\ \hline
253 & 0 & 0 & 0 & 530.425 & 512.092 & 507.522 & 505.986 & 506 \\ \hline
\end{tabular}
\caption{Hemisphere}
\end{table}
\begin{figure}[ht]\label{count_sph_hemisphere}
\centering
\includegraphics[width=0.8\textwidth]{SphDisc_50CountingGraphs}\\
\caption{Hemisphere}
\end{figure}
\begin{figure}[ht]\label{count_sph_disc2}
\centering
\includegraphics[width=0.8\textwidth]{SphDisc_25CountingGraphs}\\
\caption{Spherical disk with radius $r = \frac{\pi}{4}$}
\end{figure}
\section{Discussion}\label{sec_discus}
We have presented strong experimental evidence for the conjectures, although certain statements in the conjectures that refer to almost periodicity are not candidates for verification by examining a bottom segment of the spectrum. Indeed, since the conjectures concern asymptotic behavior, it could have happened that the bottom segment would not have given a clue to the ultimate asymptotics. As it turned out we were lucky, and the conjectured asymptotic statements kicked in very early in the game. We are inclined to believe that this is not just luck, but that there is some paradigm at work here, to the effect that qualitative asymptotic statements about spectra can be refined to quantitative error estimates that would imply "early onset." We invite the reader to speculate about this possibility.
The conjectures, first put forth in \cite{strichartz2016}, were based on examining a collection of examples for which it is possible to compute the spectrum exactly. All of these examples exhibit a high degree of symmetry. It is always risky to jump to conclusions based on symmetric examples to the general case. Having now provided experimental evidence for surfaces that are not symmetric, convex, or even simply connected, we have a much firmer platform to support the conjectures. It should be kept in mind that the role of conjecture in mathematics is not always to light the way to future truth so much as to stimulate research on interesting problems. Even conjectures that eventually turned out to be incorrect have played an important role in the development of mathematics.
We note that in \cite{strichartz2016}, Conjecture \ref{conj_curv_neg} was stated only for flat surfaces. Indeed, there are no examples of hyperbolic surfaces for which it is possible to compute the spectrum exactly. Nevertheless, there is a great deal of interest in the spectra of hyperbolic surfaces, so we are pleased that our experiments support extending the conjecture to hyperbolic surfaces.
A very interesting question, which has not yet been explored in the literature, is the behavior of the differences of consecutive eigenvalues. Note that there is an immediate difference between the nature of differences on very symmetric surfaces and "generic" surfaces. Indeed, if the surface has a nonabelian symmetry group, then there will be eigenvalues with multiplicities greater than 1, so zero will be a difference that occurs often. In the generic case we do not expect any multiplicities greater than one. So we cannot expect a general answer that applies to all examples. We have gathered data on the differences for all our examples, and this may be found on the website \cite{murray2015results}, so it is possible that we have not been able to compute enough eigenvalues with enough accuracy to make a general pattern clear. We invite the reader to consider Figures 26-34 while thinking about this challenging question. In these figures, the $x$-axis is the distance between successive eigenvalues and the curve in the first figure if the number of differences less than or equal to $x$, while the second figure is a histogram of the successive differences. Note that eigenvalue spectra containing several values with multiplicities greater than 1 have the first curve start with a high proportion of the differences at 0 (in particular see Figure 33).
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{EucDRightIsoTriDifferenceGraphs}\\
\caption{Euclidean Right Isosceles Triangle with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{EucDDiscDifferenceGraphs}\\
\caption{Euclidean Disc with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{EucDTri2DifferenceGraphs}\\
\caption{General Euclidean Triangle with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{EucDTrinTri1DifferenceGraphs}\\
\caption{Euclidean Region between Triangles with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{HypDEquTri4DifferenceGraphs}\\
\caption{Hyperbolic Equilateral Triangle ($\theta =\frac{\pi}{4}$) with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{HypDTri1DifferenceGraphs}\\
\caption{General Hyperbolic Triangle with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{HypDDiscHalfDifferenceGraphs}\\
\caption{Hyperbolic Disc ($R = \frac{1}{2}$) with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{SphEquRightDTriDifferenceGraphs}\\
\caption{Spherical Right Equilateral Triangle with Dirichlet boundary conditions}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.95\textwidth]{SphDTri2DifferenceGraphs}\\
\caption{Spherical General Triangle with Dirichlet boundary conditions}
\end{figure}
\clearpage
\bibliographystyle{AIMS}
|
\section{Introduction}
The adversarial vulnerability \cite{Szegedy:2013} of Deep Neural Nets (DNNs) threaten their applicability in security critical tasks, e.g., autonomous cars \cite{Akhtar:2018}, robotics \cite{Giusti:2016Drones}, DNN-based malware detection systems \cite{PapernotSecurity:2016,PapernotMalware:2016}. Since the pioneering work by Szegedy et al. \cite{Szegedy:2013}, many advanced adversarial attacks have been devised to generate imperceptible perturbations to fool the DNN \cite{Goodfellow:2014AdversarialTraining,PapernotAttack:2016,CWAttack:2016,Wu:2018,Ilyas:2018,Athalye:2018B,Dong_2018_CVPR}. Not only are adversarial attacks successful in white-box attacks, i.e., when the adversary has access to the DNN parameters, but they are also successful in black-box attacks, i.e., without access to network parameters. Adversarial attacks are transferable in the sense that a perturbed image meant to be misclassified by one DNN also has a significant chance to be misclassified by another DNN \cite{DBLP:journals/corr/PapernotMG16}. Due to this transferability, adversaries can attack DNN without knowing the network parameters (i.e. blackbox) \cite{LiuYanpei:2016,Brendel:2017}. There even exist universal perturbations that can imperceptibly perturb any image and cause misclassification for any given network \cite{Moosavi-Dezfooli_2017_CVPR}. And recently, there has been much work on defending against these universal perturbations \cite{Akhtar_2018_CVPR}.
In this work, we defend against adversarial attacks by replacing the commonly used output activation of DNN with a manifold-interpolating function. Together with the Projected Gradient Descent (PGD) adversarial training \cite{Madry:2018}, Total Variation Minimization (TVM), and training data augmentation, we show state-of-the-art results for adversarial defense on the CIFAR10 benchmark.
\subsection{Related Work}
Defensive distillation was recently proposed to increase the robustness of DNN \cite{PapernotDistillationDefense:2016}, and a related approach \cite{tramer2018ensemble} cleverly modifies the training data to increase robustness against black-box attacks and adversarial attacks in general. To counter adversarial perturbations, Guo et al. \cite{ChuanGuo:2018}, proposed to use image transformation, e.g., bit-depth reduction, JPEG compression, TVM, and image quilting. A similar idea of denoising the input was later explored in \cite{DBLP:journals/corr/abs-1802-06806}, where the authors divide the input into patches, denoise each patch, and then reconstruct the image. These input transformations are intended to be non-differentiable, thus making adversarial attacks more difficult, especially for gradient-based attacks. Another denoising approach is introduced by Liao et al. \cite{Liao_2018_CVPR}, where they proposed a High-level Representation Guided (HGD) denoiser -- the idea is that while perturbations seem small in the original and adversarial images, these perturbations are amplified in higher representations. Transformation-based defenses have also been proposed by Xie et al. \cite{ie:2018}, and Luo et al. \cite{DBLP:journals/corr/LuoBRPZ15}. Song et al. \cite{Song:2018}, noticed that small adversarial perturbations shift the distribution of adversarial images far from the distribution of clean images. Therefore, they proposed to purify the adversarial images by PixelDefend. And Prakash et al. \cite{Prakash_2018_CVPR}, also seek to examine image statistics in order to construct an adversarial defense -- in their work, they introduce Pixel Deflection where they force images to match statistics of natural images. Lee et al. \cite{2018arXiv180703888L}, have also used the distribution of images to detect adversarial examples. Adversarial training is another family of defense methods to enhance the stability of DNN \cite{Goodfellow:2014AdversarialTraining,Madry:2018,Na:2018}. In particular, the PGD adversarially trained DNN achieves state-of-the-art resistance to the available attacks \cite{Madry:2018}. GANs are also employed for adversarial defense \cite{Samangouei:2018}. In \cite{Athalye:2018}, the authors proposed an approximated gradient to attack the defenses that are based on the obfuscated gradient.
Instead of using the softmax function as DNN's output activation, Wang et al. \cite{BaoWang:2018NIPS,BaoWang:2019}, utilized a class of non-parametric interpolating functions. This is a combination of both deep and manifold learning which causes the DNN to utilize the geometric information of the training data sufficiently. The authors show a significant amount of generalization accuracy improvement, and the results are more stable when one only has a limited amount of training data. Recently, Wang et al. \cite{wang2019resnets} modeled ResNet as a transport equation, and they proposed an Feynman-Kac formalism principled adversarial robust DNN.
\subsection{Organization}
We organize this paper as follows: In section~\ref{section:DNN:Interpolation}, we overview the DNN with a graph Laplacian-based high dimensional interpolating activation function. In section~\ref{section:Attacks}, we present a few adversarial attacks that will be used as benchmarks for this work. In section~\ref{section:Interpolate:TVM}, we elaborate on adversarial defense via interpolating activation together with TVM. In section~\ref{section:PGD}, we further study the robustness of PGD adversarially trained DNN with interpolating activation. This paper ends up with concluding remarks.
\section{DNN with Data-Dependent Activation}\label{section:DNN:Interpolation}
In this section, we summarize the architecture, training, and testing procedures of the DNN with the data-dependent activation \cite{BaoWang:2018NIPS}. For the standard DNN with softmax activation, the training and testing are shown in Fig.~\ref{fig:WNLL-DNN-Structure} (a) and (b), respectively. In the $k$th iteration of training, given a mini-batch of training data $\mathbf{X}, \mathbf{Y}$, we perform:
{\it Forward propagation:} Transform $\mathbf{X}$ into features by the DNN block (a combination of convolutional layers, nonlinearities, etc.), and then feed the output into the softmax activation to obtain the predictions $\tilde{\mathbf{Y}}$, i.e.,
$$
\tilde{\mathbf{Y}} = {\rm Softmax}({\rm DNN}(\mathbf{X}, \Theta^{k-1}), \mathbf{W}^{k-1}).
$$
}
Then the loss is computed (e.g., cross entropy) between $\mathbf{Y}$ and $\tilde{\mathbf{Y}}$:
$\mathcal{L} \doteq \mathcal{L}^{\rm Linear} = {\rm Loss}(\mathbf{Y}, \tilde{\mathbf{Y}})$.
{\it Backpropagation:} Update weights ($\Theta^{k-1}$, $\mathbf{W}^{k-1}$) by gradient descent with learning rate $\gamma$
$$\mathbf{W}^{k} = \mathbf{W}^{k-1} - \gamma \frac{\partial \mathcal{L}}{\partial \tilde{\mathbf{Y}}}\cdot \frac{\partial \tilde{\mathbf{Y}}}{\partial \mathbf{W}},
$$
$$\Theta^{k} = \Theta^{k-1} - \gamma \frac{\partial \mathcal{L}}{\partial \tilde{\mathbf{Y}}}\cdot \frac{\partial \tilde{\mathbf{Y}}}{\partial \tilde{\mathbf{X}}}\cdot \frac{\partial \tilde{\mathbf{X}}}{\partial \Theta}.$$
}
\begin{figure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.42\columnwidth]{DNN1_training.png}&
\includegraphics[width=0.42\columnwidth]{DNN1_testing.png}\\
(a)&(b)\\
\includegraphics[width=0.51\columnwidth]{WNLL_DNN1_training.png}&
\includegraphics[width=0.41\columnwidth]{WNLL_DNN1_testing.png}\\
(c)&(d)\\
\end{tabular}
\caption{Training and testing procedures of the DNN with softmax and WNLL functions as the output activation layer. (a) and (b) show the training and testing steps for the standard DNN, respectively; (c) and (d) illustrate the training and testing procedure of the WNLL activated DNN, respectively.}
\label{fig:WNLL-DNN-Structure}
\end{figure}
Once the model is optimized, with optimal parameters being $(\Theta, \mathbf{W})$,
the predicted labels for testing data $\mathbf{X}$ are
$$
\tilde{\mathbf{Y}} = {\rm Softmax}({\rm DNN}(\mathbf{X}, \Theta), \mathbf{W}).
$$
Wang et al \cite{BaoWang:2018NIPS} proposed to replace the data-agnostic softmax by an interpolating function defined below.
\subsection{Manifold Interpolation - A Harmonic Extension Approach}
Let $\mathbf{X} = \{\mathbf{x}_1, \mathbf{x}_2, \cdots, \mathbf{x}_n\}$ be a set of points on a high dimensional manifold $\mathcal{M}\subset\mathbb{R}^d$ and $\mathbf{X}^{\rm te} = \{\mathbf{x}^{\rm te}_1, \mathbf{x}^{\rm te}_2, \cdots, \mathbf{x}^{\rm te}_m\}$ (``te" for template) be a subset of $\mathbf{X}$ which are labeled with label function $g(\mathbf{x})$ \footnote{
The minimum requirement is that the template data needs to cover all classes. In \cite{BaoWang:2018NIPS}, we show that for an image classification task with $m$ number of different classes, the size of the template needs to be at least $m\log m$. In practice, the size of the template set will not affect the performance much as long as the template set size is more than 1K for CIFAR10 and CIFAR100.}. We want to interpolate a function $u$ that is defined on $\mathcal{M}$ and can be used to label the entire dataset $\mathbf{X}$. The harmonic extension is a natural approach to find such an interpolating function, which is defined by minimizing the following Dirichlet energy functional
\begin{equation}
\label{DirichletEnergy}
\mathcal{E}(u)=\frac{1}{2}\sum_{\mathbf{x}, \mathbf{y}\in \mathbf{X}} w(\mathbf{x}, \mathbf{y})\left(u(\mathbf{x})-u(\mathbf{y})\right)^2,
\end{equation}
}
with the boundary condition
$$
u(\mathbf{x})=g(\mathbf{x}), \ \mathbf{x}\in \mathbf{X}^{\rm te},
$$
}
where $w(\mathbf{x}, \mathbf{y})$ is a weight function, chosen to be Gaussian: $w(\mathbf{x}, \mathbf{y})=\exp(-\frac{||\mathbf{x}-\mathbf{y}||^2}{\sigma^2})$ with $\sigma$ being a scaling parameter. The Euler-Lagrange equation for Eq.~(\ref{DirichletEnergy}) is
\begin{equation}
\label{EL-Equation}
\begin{cases}
\sum_{\mathbf{y}\in \mathbf{X}} \left(w(\mathbf{x}, \mathbf{y})+w(\mathbf{y}, \mathbf{x})\right)\left(u(\mathbf{x})-u(\mathbf{y})\right)=0 & \hskip -0.2cm \mathbf{x}\in \mathbf{X}/\mathbf{X}^{\rm te}\\
u(\mathbf{x})=g(\mathbf{x}) & \hskip -0.3cm \mathbf{x}\in \mathbf{X}^{\rm te}.
\end{cases}
\end{equation}
}
By solving the linear system Eq.~(\ref{EL-Equation}), we obtain labels $u(\mathbf{x})$ for the unlabeled data $\mathbf{x}\in \mathbf{X}/\mathbf{X}^{\rm te}$. This interpolation becomes invalid when the labeled data is tiny, i.e., $|\mathbf{X}^{\rm te}|\ll |\mathbf{X}
/\mathbf{X}^{\rm te}|$. To resolve this issue, the weights of the labeled data is increased in the Euler-Lagrange equation, which gives
\begin{equation}
\label{WNLL}
\begin{cases}
\sum_{\mathbf{y}\in \mathbf{X}} \left(w(\mathbf{x}, \mathbf{y})+w(\mathbf{y}, \mathbf{x})\right)\left(u(\mathbf{x})-u(\mathbf{y})\right)+\\
\left(\frac{|\mathbf{X}|}{|\mathbf{X}^{\rm te}|}-1\right)\sum_{\mathbf{y}\in \mathbf{X}^{\rm te}}w(\mathbf{y}, \mathbf{x})\left(u(\mathbf{x})-u(\mathbf{y})\right)=0 & \hskip -0.2cm \mathbf{x}\in \mathbf{X}/\mathbf{X}^{\rm te}\\
u(\mathbf{x})=g(\mathbf{x}) &\hskip -0.2cm \mathbf{x}\in \mathbf{X}^{\rm te}.
\end{cases}
\end{equation}
}
The solution to Eq.~(\ref{WNLL}) is named weighted nonlocal Laplacian (WNLL), denoted as ${\rm WNLL}(\mathbf{X}, \mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te})$. Shi et al. \cite{WNLL:2018}, showed that WNLL converges to the solution of the high dimensional Laplace-Beltrami equation. For classification, $g(\mathbf{x})$ is the one-hot label for the example $\mathbf{x}$.
\subsection{Training and Testing of the DNN with Data-Dependent Activation Function}
For a standard DNN, we denote the WNLL activated one as DNN-WNLL, e.g., the WNLL activated ResNet20 is denoted as ResNet20-WNLL. In both training and testing of the DNN-WNLL, we need to reserve a small portion of data/label pairs, denoted as $(\mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te})$, to interpolate the label for new data. We name the reserved data $(\mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te})$ as the template. Directly replacing softmax by WNLL has difficulties in back propagation, namely the true gradients $\frac{\partial \mathcal{L}}{\partial \Theta}$ and $\frac{\partial \mathcal{L}}{\partial \mathbf{W}_B}$ (here $\mathcal{L}\doteq \mathcal{L}^{WNLL} = {\rm Loss}(\hat{Y}, Y)$, as shown in Fig.~\ref{fig:WNLL-DNN-Structure}(c)) are difficult to compute since WNLL defines an implicit function. Instead, to train the DNN-WNLL, a proxy via an auxiliary DNN (Fig.~\ref{fig:WNLL-DNN-Structure}(c)) is employed. On top of the original DNN, we add a buffer block (a fully connected layer followed by a ReLU), and followed by two parallel branches, WNLL and linear (fully connected) layers. The auxiliary DNN can be trained by alternating between training the DNN with linear and WNLL activation functions, respectively. When training DNN with WNLL activation function, the training loss of the WNLL activation is backpropped via a straight-through gradient estimator \cite{Athalye:2018,BengioStraightTE:2013}, e.g., in the $k$th iteration, we use the following approximated gradient descent (Eq.~(\ref{ApproxBP})) to update $\mathbf{W}_B$ only (when backpropagating the training loss $\mathcal{L}^{\rm WNLL}$ we freeze the remaining part except for the buffer block, and the other parameters will be updated in training DNN with linear activation function),
\begin{eqnarray}
\label{ApproxBP}
\mathbf{W}_B^{k} = \mathbf{W}_B^{k-1} - \gamma \frac{\partial \mathcal{L}^{\rm WNLL}}{\partial \hat{\mathbf{Y}}}\cdot \frac{\partial \hat{\mathbf{Y}}}{\partial \hat{\mathbf{X}}}\cdot \frac{\partial \hat{\mathbf{X}}}{\partial \mathbf{W}_B} \\ \nonumber
\approx \mathbf{W}_B^{k-1} - \gamma \frac{\partial \mathcal{L}^{\rm Linear}}{\partial \tilde{\mathbf{Y}}}\cdot \frac{\partial \tilde{\mathbf{Y}}}{\partial \hat{\mathbf{X}}}\cdot \frac{\partial \hat{\mathbf{X}}}{\partial \mathbf{W}_B},
\label{eq:bp-wnll}
\end{eqnarray}
}
where $\frac{\partial \mathcal{L}^{\rm Linear}}{\partial \tilde{\mathbf{Y}}}$ and $\frac{\partial \mathcal{L}^{\rm WNLL}}{\partial \hat{\mathbf{Y}}}$ are the gradients computed through
two different activation functions. In the approximation of Eq.~(\ref{ApproxBP}), we simply replace the value of $\mathcal{L}^{\rm Linear}$ with that of $\mathcal{L}^{\rm WNLL}$, which allows us to compute the value of $\frac{\partial \mathcal{L}^{\rm WNLL}}{\partial \hat{\mathbf{Y}}}$ by leveraging the computational graph of DNN with linear activation. The detailed training procedure can be found in \cite{BaoWang:2018NIPS}.
At test time, we remove the linear activation from the neural net and use the DNN and buffer blocks together with WNLL to classify new data (Fig. \ref{fig:WNLL-DNN-Structure} (d)). Here for simplicity, we merge the buffer block to the DNN block. For a given set of testing data $\mathbf{X}$, and the labeled template $\{(\mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te})\}$, the predicted labels for $\mathbf{X}$ is given by
$$
\tilde{\mathbf{Y}} = {\rm WNLL}({\rm DNN}(\mathbf{X}, \mathbf{X}^{\rm te}, \Theta), \mathbf{Y}^{\rm te}).
$$
}
\subsection{Computational Complexity of DNN with Data-Dependent Activation}
Using WNLL activation will lead to some extra computational overhead, which comes from the nearest neighbor searching and solving a system of linear equations. We following the same training procedure as that used in \cite{BaoWang:2018NIPS} to train ResNet20.
In Table~\ref{Numerical-Compare-Complexity}, we list the training and test time on a single Titan Xp GPU for ResNet20 on CIFAR10.
\begin{table}[!ht]
\renewcommand{\arraystretch}{1.3}
\centering
\caption{Running time and GPU memory for ResNet20 with two different activation functions.}
\label{Numerical-Compare-Complexity}
\centering
\begin{tabular}{cccc}
\hline
& Training time & Testing time & Memory\\
\hline
ResNet20 & 3925.6 (s) & 0.657 (s) & 1007 (MB)\\
ResNet20-WNLL & 7378.4 (s) & 14.09 (s) & 1563 (MB)\\
\hline
\end{tabular}
\end{table}
\section{Adversarial Attacks}\label{section:Attacks}
We consider three benchmark attacks: the Fast Gradient Sign Method (FGSM) \cite{Goodfellow:2014AdversarialTraining}, Iterative FGSM (IFGSM) \cite{Kurakin:2016}, and Carlini-Wagner's $L_2$ (CW-L2) \cite{CWAttack:2016} attack. We denote the classifier defined by the DNN as $\tilde{y} = f(\theta, \mathbf{x})$ for a given instance ($\mathbf{x}$, $y$). FGSM searchs the adversarial image $\mathbf{x}'$ with a bounded perturbation by maximizing the loss $\mathcal{L}(\mathbf{x}', y) \doteq \mathcal{L}(f(\theta, \mathbf{x}'), y)$, subject to the $l_\infty$ perturbation constraint $||\mathbf{x}'-\mathbf{x}||_\infty \leq \epsilon$ with $\epsilon$ being the attack strength. We can approximately solve this constrained optimization problem by using the first order approximation of the loss function i.e., $\mathcal{L}(\mathbf{x}', y) \approx \mathcal{L}(\mathbf{x}, y) + \nabla_\mathbf{x}\mathcal{L}(\mathbf{x}, y)^T \cdot (\mathbf{x}'-\mathbf{x})$. Under this approximation, the optimal adversarial image is
\begin{equation}
\label{FGSM}
\mathbf{x}'=\mathbf{x} + \epsilon \, {\rm sign} \cdot \left( \nabla_\mathbf{x}\mathcal{L}(\mathbf{x}, y) \right).
\end{equation}
IFGSM iterates FGSM to generate the enhanced attack, i.e.,
\begin{equation}
\label{IFGSM-2}
\mathbf{x}^{(m)} = {\rm Clip}_{\mathbf{x}, \epsilon}\left\{\mathbf{x}^{(m-1)} + \alpha \cdot {\rm sign} \left( \nabla_{\mathbf{x}} \mathcal{L}(\mathbf{x}^{(m-1)}, y) \right)\right\},
\end{equation}
where $m=1, \cdots, M$, $\mathbf{x}^{(0)}=\mathbf{x}$ and $\mathbf{x}'=\mathbf{x}^{(M)}$, with $M$ be the number of iterations. $\alpha$ is the step size used in each iteration, and ${\rm Clip}_{\mathbf{x}, \epsilon}$ clips the update to be within an $\epsilon$-ball centered at $\mathbf{x}$ in $l_\infty$-norm.
Moreover, we consider the attack due to Carlini and Wagner. For a given image-label pair $(\mathbf{x}, y)$, and $\forall t\neq y$, CW-L2 searches the adversarial image that will be classified to class $t$ by solving the optimization problem
\begin{equation}
\label{cwl2-eq1}
\min_{\delta} ||\delta||_2^2,
\end{equation}
subject to
$$
f(\mathbf{x}+\delta) = t, \; \mathbf{x}+\delta \in [0, 1]^n,
$$
where $\delta$ is the adversarial perturbation (for simplicity, we ignore the dependence of $\theta$ in $f$).
The equality constraint in Eq.~(\ref{cwl2-eq1}) is hard to handle, so Carlini et al. considered the surrogate
\begin{equation}
\label{cwl2-eq2}
g(\mathbf{x}) = \max\left(\max_{i\neq t}(Z(\mathbf{x})_i) - Z(\mathbf{x})_t , 0\right),
\end{equation}
where $Z(\mathbf{x})$ is the logit vector for an input $\mathbf{x}$, i.e., output of DNN before the output layer. $Z(\mathbf{x})_i$ is the logit value corresponding to class $i$. It is easy to see that $f(\mathbf{x}+\delta)=t$ is equivalent to $g(\mathbf{x}+\delta)\leq 0$. Therefore, the problem in Eq.~(\ref{cwl2-eq1}) can be reformulated as
\begin{equation}
\label{cwl2-eq4}
\min_{\delta} ||\delta||_2^2 + c \cdot g(\mathbf{x}+\delta),
\end{equation}
subject to
$$
\mathbf{x}+\delta \in [0, 1]^n,
$$
where $c\geq 0$ is the Lagrangian multiplier.
By letting $\delta = \frac{1}{2}\left(\tanh(\mathbf{w})+1\right)-\mathbf{x}$, Eq.~(\ref{cwl2-eq4}) can be written as an unconstrained optimization problem. Moreover, Carlini et al. introduce the confidence parameter $\kappa$ into the above formulation. Above all, the CW-L2 attack seeks the adversarial image by solving the following problem
\begin{align}
\label{CWL2}
&\min_{\mathbf{w}} ||\frac{1}{2}\left(\tanh(\mathbf{w}) + 1\right) - \mathbf{x} ||_2^2 + c\cdot\\ \nonumber
&\max\left\{-\kappa, \max_{i\neq t}(Z(\frac{1}{2}(\tanh(\mathbf{w}))+1)_i)
- Z(\frac{1}{2}(\tanh(\mathbf{w}))+1)_t \right\}.
\end{align}
}
The Adam optimizer \cite{Kingma:2014Adam} can solve this unconstrained optimization problem efficiently. All three attacks clip the values of the adversarial image $\mathbf{x}'$ to between 0 and 1.
\begin{figure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\columnwidth]{Img_2_New.png}&
\includegraphics[width=0.45\columnwidth]{Img_8_New.png}\\
(a)&(b)\\
\includegraphics[width=0.45\columnwidth]{ImgTVM_2_New.png}&
\includegraphics[width=0.45\columnwidth]{ImgTVM_8_New.png}\\
(c)&(d)\\
\end{tabular}
\caption{Samples from CIFAR10. Panel (a): from the top to the last rows show the original, adversarial images by attacking ResNet56 with FGSM and IFGSM ($\epsilon=0.02$); and by attacking ResNet56-WNLL. Panel (b) corresponding to those in panel (a) with $\epsilon=0.08$. Charts (c) and (d) corresponding to the TV minimized images in (a) and (b), respectively.}
\label{fig:Adv-images}
\end{figure}
\subsection{Attack the DNN with WNLL Activation}
For a given mini-batch of testing images $(\mathbf{X}, \mathbf{Y})$ and template $(\mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te})$, we denote the DNN-WNLL as $\tilde{\mathbf{Y}} = {\rm WNLL}(Z(\{\mathbf{X}, \mathbf{X}^{\rm te}\} ), \mathbf{Y}^{\rm te})$, where $Z(\{\mathbf{X}, \mathbf{X}^{\rm te}\} )$ is the composition of the DNN and buffer blocks as shown in Fig.~\ref{fig:WNLL-DNN-Structure}(c). By ignoring dependence of the loss function on the parameters, the loss function for DNN-WNLL can be written as $\tilde{\mathcal{L}}(\mathbf{X}, \mathbf{Y}, \mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te}) \doteq {\rm Loss}(\hat{\mathbf{Y}}, \mathbf{Y})$. The above attacks for DNN-WNLL are formulated below.
\begin{itemize}
\item {\bf FGSM}
\begin{align}
\label{FGSM-WNLL}
\mathbf{X}' = \mathbf{X} + \epsilon \cdot {\rm sign}\left( \nabla_\mathbf{X} \tilde{\mathcal{L}}(\mathbf{X}, \mathbf{Y}, \mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te}) \right).
\end{align}
\item {\bf IFGSM}
\begin{align}
\label{IFGSM-WNLL}
\mathbf{X}^{(m)} = {\rm Clip}_{\mathbf{X}, \epsilon}[\mathbf{X}^{(m-1)} +
\alpha \cdot {\rm sign} \left( \nabla_{\mathbf{X}} \tilde{\mathcal{L}}(\mathbf{X}^{(m-1)}, \mathbf{Y}, \mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te}) \right)],
\end{align}
where $m=1, 2, \cdots, M$; $\mathbf{X}^{(0)} = \mathbf{X}$ and $\mathbf{X}'=\mathbf{X}^{(M)}$.
\item {\bf CW-L2}
\begin{align}\label{CWL2-WNLL}
&\min_{\mathbf{W}} ||\frac{1}{2}\left(\tanh(\mathbf{W}) + 1\right) - \mathbf{X} ||_2^2 + c\cdot \max[-\kappa, \\ \nonumber
& \max_{\mathbf{i}\neq \mathbf{t} }( Z(\frac{1}{2}(\tanh(\mathbf{W}))+1)_\mathbf{i}) - Z(\frac{1}{2}(\tanh(\mathbf{W}))+1)_\mathbf{t} ],
\end{align}
}
where $\mathbf{i}$ are the logit values of the input images $\mathbf{X}$, $\mathbf{t}$ are the target labels.
\end{itemize}
In the above attacks, $\nabla_{\mathbf{X}}\tilde{\mathcal{L}}$ is required to generate the adversarial images. In the DNN-WNLL, this gradient is difficult to compute. As shown in Fig.~\ref{fig:WNLL-DNN-Structure} (c), we approximate $\nabla_{\mathbf{X}}\tilde{\mathcal{L}}$ in the following way
\begin{eqnarray}
\label{GradientApprox}
\nabla_{\mathbf{X}}\tilde{\mathcal{L}} = \frac{\partial \mathcal{L}^{\rm WNLL}}{\partial\hat{\mathbf{Y}}}\cdot \frac{\partial\hat{\mathbf{Y}}}{\partial\hat{\mathbf{X}}} \cdot \frac{\partial\hat{\mathbf{X}}}{\partial\tilde{\mathbf{X}}} \cdot \frac{\partial\tilde{\mathbf{X}}}{\partial\mathbf{X}}
\approx
\frac{\partial \mathcal{L}^{\rm Linear}}{\partial\tilde{\mathbf{Y}}}\cdot \frac{\partial\tilde{\mathbf{Y}}}{\partial\hat{\mathbf{X}}} \cdot \frac{\partial\hat{\mathbf{X}}}{\partial\tilde{\mathbf{X}}} \cdot \frac{\partial\tilde{\mathbf{X}}}{\partial\mathbf{X}},
\end{eqnarray}
}
again, in the above approximation, we set the value of $\mathcal{L}^{\rm Linear}$ to the value of $\tilde{\mathcal{L}}$.
Based on our numerical experiments, the batch size of $\mathbf{X}$ has minimal influence on the adversarial attack and defense. In all of our experiments, we choose the size of both mini-batches $\mathbf{X}$ and the template to be $500$.
\section{Defense by Interpolating Function, TVM, and Training Data Augmentation}\label{section:Interpolate:TVM}
To defend against adversarial attacks, we first combine the data-dependent activation with input transformation and with training data augmentation. We train ResNet56 \cite{ResNet} and ResNet56-WNLL, respectively, on the original training data, the TV minimized training data, and a combination of the previous two.
Moreover, in testing, we apply the TVM \cite{ROF:1992} used by \cite{ChuanGuo:2018}, with the same setting, to transform the adversarial images to boost classification performance. The basic idea of TVM is to reconstruct the simplest image $\mathbf{z}$ from the sub-sampled image, $X\odot \mathbf{x}$ with $X$ the mask filled by a Bernoulli binary random variable, by solving
\begin{equation*}
\min_\mathbf{z}||(1-X)\odot (\mathbf{z} - \mathbf{x}) ||_2 + \lambda_{TV} \cdot TV_2(\mathbf{z}),
\end{equation*}
where $\lambda_{TV} > 0$ is the regularization constant.
We apply the three attack schemes mentioned above to attack ResNet56 and ResNet56-WNLL. For IFGSM, we run 10 iterations of Eqs.~(\ref{IFGSM-2}) and (\ref{IFGSM-WNLL}) with $\epsilon=0.1$ to attack the DNN with two different output activations, respectively. For the CW-L2 attack (Eqs.~(\ref{CWL2}, \ref{CWL2-WNLL})), in both scenarios we set the parameters $c=10$ and $\kappa=0$, and run 10 iterations of the Adam optimizer with learning rate 0.01.
Figure~\ref{fig:Adv-images} depicts three randomly selected images (horse, automobile, airplane) from the CIFAR10 dataset, as well as the perturbed images from applying different attacks on ResNet56 and ResNet56-WNLL, and the TV minimized ones. All attacks successfully fool the classifiers to classify any of them correctly. Figure~\ref{fig:Adv-images} (a) shows that the perturbations resulted from FGSM attack with $\epsilon=0.02$ is almost imperceptible.
However, both FGSM and IFGSM attacks are powerful in fooling DNNs. Figure~\ref{fig:Adv-images} (b) shows the corresponding images of (a) with a stronger attack, $\epsilon=0.08$. With a larger $\epsilon$, the adversarial images become more noisy. The TV minimized images of Fig.~\ref{fig:Adv-images} (a) and (b) are shown in Fig.~\ref{fig:Adv-images} (c) and (d), respectively. TVM removes a significant amount of information from the original and the adversarial images. Meanwhile, it also makes it harder for humans to classify them.
\subsection{Numerical Results}
In this subsection, we first discuss the transferability of adversarial examples generated by attacking DNNs with softmax and WNLL activation functions. The transferability of adversarial examples is often used for black-box adversarial attacks. Adversarial examples of a robust DNN typically have good transferability. Next, we numerically verify the efficacy of adversarial defense by leveraging DNN with the WNLL activation function and TVM. Finally, we explain the adversarial robustness by considering the deep learning features learned by DNN with different activation functions.
\subsubsection{Transferability of the Adversarial Images}
Consider the transferability of adversarial examples crafted by using the above adversarial attacks to attack ResNet56 with either softmax or WNLL activation. We utilize the training strategy used in \cite{BaoWang:2018NIPS} to train the DNNs. To test the transferability, we classify the adversarial images by using ResNet56 with the opponent activation (the opponent activation of WNLL is softmax, and vice versa). We list the mutual classification accuracy (the accuracy of DNN with one specific activation to classify adversarial images crafted by attacking DNN with the other activation) on adversarial images resulting from using FGSM or IFGSM in Table.~\ref{Table1:Mutual-Classification}. The adversarial images crafted by attacking ResNet56 with two types of activation functions are both transferable, as the mutual classification accuracy on adversarial images ($\epsilon\neq 0$) is significantly lower than testing on the clean images ($\epsilon=0$). For both FGSM and IFGSM, the stronger attack (in the sense of bigger $\epsilon$) is adapted to the opponent activation function, as the mutual classification accuracy decreases dramatically as $\epsilon$ increases. IFGSM not only fools its underlying model completely, but also significantly decreases the accuracy of the opponent DNN. The mutual classification results for the CW-L2 attack is shown in Table.~\ref{Table1:Mutual-Classification-WNLL}, where Exp-I denotes classifying adversarial images resulted from attacking ResNet56-WNLL by ResNet56, and Exp-II denotes the opposite. Training data augmentation can defend CW-L2 attack very effectively.
\begin{table}[!ht]
\centering
\fontsize{10.0}{10}\selectfont
\begin{threeparttable}
\caption{Mutual classification accuracy on the adversarial images crafted by using FGSM and IFGSM to attack ResNet56 and ResNet56-WNLL. (Unit: $\%$)}\label{Table1:Mutual-Classification}
\begin{tabular}{cccccccc}
\toprule
Attack &Training data & $\epsilon=0$ & $\epsilon=0.02$ & $\epsilon=0.04$ & $\epsilon=0.06$ & $\epsilon=0.08$ & $\epsilon=0.1$ \cr
\midrule
\multicolumn{8}{c}{Accuracy of ResNet56 on adversarial images crafted by attacking ResNet56-WNLL}\cr
\midrule
FGSM & Original data &93.0 &69.8 &56.9 &44.6 &34.6 &28.3 \cr
FGSM & TVM data &88.3 &51.5 &37.9 &30.1 &24.7 &20.9 \cr
FGSM & Original + TVM &93.1 &78.5 &70.9 &64.6 &59.8 &55.8 \cr
\midrule
IFGSM & Original data &93.0 &5.22 &5.73 &6.73 &7.55 &8.55 \cr
IFGSM & TVM data &88.3 &7.00 &6.82 &8.30 &9.28 &10.7 \cr
IFGSM & Original + TVM&93.1 &27.3 &28.6 &29.5 &29.1 &29.4 \cr
\toprule
\multicolumn{7}{c}{Accuracy of ResNet56-WNLL on adversarial images crafted by attacking ResNet56}\cr
\midrule
FGSM & Original data &94.5 &65.2 &49.0 &39.3 &32.8 &28.3 \cr
FGSM & TVM data &90.6 &45.9 &30.9 &22.2 &16.9 &13.8 \cr
FGSM & Original + TVM data &94.7 &78.3 &68.2 &61.1 &56.5 &52.5 \cr
\midrule
IFGSM & Original data &94.5 &3.37 &3.71 &3.54 &4.69 &6.41 \cr
IFGSM & TVM data &90.6 &7.88 &7.51 &7.58 &8.07 &9.67 \cr
IFGSM & Original + TVM data&94.7 &34.3 &33.4 &33.1 &34.6 &35.8 \cr
\bottomrule
\end{tabular}
\end{threeparttable}
\end{table}
\begin{table}[!ht]
\centering
\fontsize{10.0}{10}\selectfont
\begin{threeparttable}
\caption{Mutual classification accuracy on the adversarial images crafted by using CW-L2 to attack ResNet56 and ResNet56-WNLL. (Unit: $\%$)}\label{Table1:Mutual-Classification-WNLL}
\begin{tabular}{cccc}
\toprule
\ \ \ \ \ Training data \ \ \ \ &\ \ \ \ Original data\ \ \ \ & \ \ \ \ TVM data\ \ \ \ & \ \ \ Original + TVM data \cr
\midrule
Exp-I & 52.1 & 43.2 & 80.0\\
\midrule
Exp-II & 59.7 & 41.1 & 80.1 \\
\bottomrule
\end{tabular}
\end{threeparttable}
\end{table}
\begin{table}[!ht]
\centering
\fontsize{10.0}{10}\selectfont
\begin{threeparttable}
\caption{Testing accuracy on the adversarial/TVM adversarial CIFAR10 dataset. The testing accuracy with no defense is in red italic; and the results with all three defenses are in boldface. (Unit: $\%$)}\label{Table1:adversarialdat-WNLL}
\begin{tabular}{cccc}
\toprule
\ \ \ \ \ Training data\ \ \ \ \ &\ \ \ Original data\ \ \ &\ \ \ TVM data\ \ \ & \ \ \ Original + TVM data\ \ \ \cr
\midrule
ResNet56 & \textcolor{red}{\it 4.94}/32.2 & 11.8/54.0 & 15.1/52.4\\
\midrule
ResNet56-WNLL & 18.3/35.2 & 15.0/53.9 & 28/{\bf 54.5} \\
\bottomrule
\end{tabular}
\end{threeparttable}
\end{table}
\begin{table}[!ht]
\centering
\fontsize{10.0}{10}\selectfont
\begin{threeparttable}
\caption{Testing accuracy on the adversarial/TVM adversarial CIFAR10 dataset. The testing accuracy with no defense is in red italic; and the results with all three defenses are in boldface. (Unit: $\%$)}\label{Table1:adversarialdata}
\begin{tabular}{cccccccc}
\toprule
Attack &Training data & $\epsilon=0$ & $\epsilon=0.02$ & $\epsilon=0.04$ & $\epsilon=0.06$ & $\epsilon=0.08$ & $\epsilon=0.1$ \cr
\midrule
\multicolumn{8}{c}{ResNet56}\cr
\midrule
FGSM & Original data &93.0 &\textcolor{red}{\it 36.9}/19.4 &\textcolor{red}{\it 29.6}/18.9 &\textcolor{red}{\it 26.1}/18.4 &\textcolor{red}{\it 23.1}/17.9 &\textcolor{red}{\it 20.5}/17.1 \cr
FGSM & TVM data &88.3 &27.4/50.4 &19.1/47.2 &16.6/43.7 &15.0/38.9 &13.7/35.0 \cr
FGSM & Original + TVM &93.1 &48.6/51.1 &42.0/47.6 &39.1/44.2 &37.1/41.8 &35.6/39.1 \cr
\midrule
IFGSM & Original data &93.0 &\textcolor{red}{\it 0}/16.6 &\textcolor{red}{\it 0}/16.1 &\textcolor{red}{\it 0.02}/15.9 &\textcolor{red}{\it 0.1}/15.5 &\textcolor{red}{\it 0.25}/16.1 \cr
IFGSM & TVM data &88.3 &0.01/43.4 &0/42.5 &0.02/42.4 &0.18/42.7 &0.49/42.4 \cr
IFGSM & Original + TVM &93.1 &0.1/38.4 &0.09/37.9 &0.36/37.9 &0.84/37.6&1.04/37.9 \cr
\toprule
\multicolumn{8}{c}{ResNet56-WNLL}\cr
\midrule
FGSM & Original data &94.5 &58.5/26.0 &50.1/25.4 &42.3/25.5 &35.7/24.9 &29.2/22.9 \cr
FGSM & TVM data &90.6 &31.5/52.6 &24.5/49.6 &20.2/45.3 &17.3/41.6 &14.4/37.5 \cr
FGSM & Original + TVM &94.7 &{\bf60.5}/ 55.4 &{\bf56.7}/52.0 &{\bf55.3}/48.6 &{\bf53.2}/45.9 &{\bf50.1}/43.7 \cr
\midrule
IFGSM & Original data &94.5 &0.49/16.7 &0.14/17.3 &0.3/16.9 &1.01/16.6 &0.94/16.5 \cr
IFGSM & TVM data &90.6 &0.61/37.3 &0.43/36.3 &0.63/35.9 &0.87/35.9 &1.19/35.5 \cr
IFGSM & Original + TVM &94.7 &0.19/{\bf38.5 } &0.3/{\bf39.4 } &0.63/{\bf 40.1} &1.26/{\bf 38.9} &1.72/{\bf 39.1} \cr
\bottomrule
\end{tabular}
\end{threeparttable}
\end{table}
\subsubsection{Adversarial Defense}
Figure~\ref{fig:Defense-Res} plots the results of adversarial defense by combining the WNLL activation, TVM, and training data augmentation. Panels (a) and (b) show the testing accuracy of ResNet56 with and without defense on CIFAR10 data for FGSM and IFGSM, respectively. It is seen that as $\epsilon$ increases, the testing accuracy decreases rapidly. FGSM is a relatively weak attack, and the accuracy remains above 20.5$\%$ even with the most potent attack ($\epsilon=0.1$). Meanwhile, the defense raises the accuracy to 43.7$\%$. Figure~\ref{fig:Defense-Res} (b) shows that IFGSM fools ResNet56 near completely even with $\epsilon=0.02$. The defense maintains the accuracy above 38.5$\%$, 54.5$\%$ under the CW-L2 and IFGSM attacks, respectively (see Tables.~\ref{Table1:adversarialdat-WNLL} and \ref{Table1:adversarialdata}). Compared to the state-of-the-art defensive methods on CIFAR10, PixelDefend, our approach is much simpler and faster. Without adversarial training, we have shown our defense is more robust to FGSM and IFGSM attacks under the strongest attack than PixelDefend \cite{Song:2018}. Moreover, our defense strategy is additive to adversarial training and many other defenses including PixelDefend.
\begin{figure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.46\columnwidth]{Defend_FGSM.png}&
\includegraphics[width=0.46\columnwidth]{Defend_iFGSM.png}\\
(a)&(b)\\
\end{tabular}
\caption{$\epsilon$ v.s. accuracy without defense, and defending by WNLL activation, TVM and augmented training. (a) and (b) plot results for FGSM and IFGSM attack, respectively.}
\label{fig:Defense-Res}
\end{figure}
To analyze the contribution from each component of the defensive strategy, we separate the three parts and list the testing accuracy in Tables.~\ref{Table1:adversarialdat-WNLL} and \ref{Table1:adversarialdata}. Performing TVM on the adversarial images cannot defend FGSM attacks except when the training data contains the TV minimized images. For instance, when we attack the model by FGSM with $\epsilon=0.02$, the accuracy on the adversarial images for ResNet56 and ResNet56-WNLL are 36.9$\%$ and 58.5$\%$, respectively, provided the models are trained on the original training data. The accuracy reduces to 19.4$\%$ and 26.0$\%$ when testing on the TV minimized adversarial images. For ResNet56, the accuracy raises to 50.4$\%$ and 51.1$\%$ when the model is trained on the TVM and augmented data, respectively. For ResNet-WNLL, the accuracy increases to 52.6$\%$ and 55.4$\%$, respectively.
The WNLL activation improves testing accuracy of adversarial attacks significantly and persistently. Augmented training can also improve the stability consistently.
IFGSM fools the ResNet56-WNLL near completely, as the accuracy is always less than or close to 1$\%$. These results verify the efficacy of using the approximated gradient, i.e., Eq.~(\ref{GradientApprox}), in attacking the neural nets.
\begin{algorithm}[!ht]
\caption{PGD Adversarial Training of the DNN-WNLL}\label{PGD-WNLL}
\begin{algorithmic}
\State \textbf{Input: } Training set: (data, label) pairs $(\mathbf{X}, \mathbf{Y})$. $N_1$: the number of epochs in training DNN $+$ Linear blocks. $N_{\rm IFGSM}$: the number of iterations of IFGSM attack.
\State \textbf{Output: } An optimized DNN-WNLL, denoted as ${\rm DNN}_{\rm WNLL}$.
\For {${\rm iter} = 1, \dots, N$ (where $N$ is the number of alternating steps.)}
\State //{\it PGD adversarial training of the left branch: DNN with linear activation}.
\State Train DNN $+$ Linear blocks, and denote the learned model as ${\rm DNN}_{\rm Linear}$.
\State Partition the training data into $M_1$ mini-batches, i.e., $(\mathbf{X}, \mathbf{Y}) = \bigcup_{i=1}^{M_1} (\mathbf{X}_i, \mathbf{Y}_i)$.
\For{${\rm epoch_1} = 1, $\dots$, N_1$}
\For{$i = 1, $\dots$, M_1$}
\State //{\it Attack the input images by IFGSM.}
\For{${\rm iter_1} = 1, \dots, N_{\rm \footnotesize IFGSM}$}
\State Update the training image $\mathbf{X}_i = \mathbf{X}_i + \epsilon\cdot {\rm sign}\left(\nabla_{\mathbf{X}_i} \mathcal{L} \right)$, \\
\hskip 2.1cm where $\mathcal{L}$ is the loss of the prediction by DNN $+$ Linear blocks w.r.t\\
\hskip 2.1cm the ground truth labels $\mathbf{Y}_i$.
\EndFor
\State Backpropagate the classification error of the adversarial images.
\EndFor
\EndFor
\State //{\it PGD adversarial training of the right branch: DNN with WNLL activation.}
\State Split $(\mathbf{X}, \mathbf{Y})$ into training data and template, i.e.,\\
\hskip 1.1cm $(\mathbf{X}, \mathbf{Y}) \doteq (\mathbf{X}^{\rm tr}, \mathbf{Y}^{\rm tr}) \bigcup (\mathbf{X}^{\rm te}, \mathbf{Y}^{\rm te})$.
\State Partition the training data into $M_2$ mini-batches, i.e.,\\
\hskip 1.1cm $(\mathbf{X}^{\rm tr}, \mathbf{Y}^{\rm tr}) = \bigcup_{i=1}^{M_2} (\mathbf{X}_i^{\rm tr}, \mathbf{Y}_i^{\rm tr})$.
\For{${\rm epoch_2} = 1, $\dots$, N_2$}
\For{$i=1, \dots, M_2$}
\State //{\it Attack the input training images by IFGSM.}
\For{${\rm iter_1} = 1, \dots, N_{\rm \footnotesize IFGSM}$}
\State Update the training image $\mathbf{X}^{\rm tr}_i = \mathbf{X}^{\rm tr}_i + \epsilon\cdot {\rm sign}\left(\nabla_{\mathbf{X}^{\rm tr}_i} \tilde{\mathcal{L}} \right)$, \\
\hskip 2.1cm where $\tilde{\mathcal{L}}$ is the loss of the prediction by DNN with WNLL\\
\hskip 2.1cm activation w.r.t the ground truth labels $\mathbf{Y}^{\rm tr}_i$.\\
\EndFor
\State Backpropagate the classification error of the adversarial images.
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\subsubsection{Analysis of the Geometry of Features}
We consider features' geometry of the original and adversarial images.
We randomly select 1000 training and 100 testing images from the airplane and automobile classes, respectively.
We apply two visualization strategies for ResNet56: (1) Apply the principle component analysis (PCA) to reduce the 64D features from the layer before the softmax to 2D,
and (2) we add a 2 by 2 fully connected (FC) layer before the softmax to learn 2D features.
We verify that the newly added layer does not change the performance of ResNet56, as shown in Fig.~\ref{fig:Add2x2}, and the training and testing performance remains essentially the same.
\begin{figure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.46\columnwidth]{Acc-Original.png} &
\includegraphics[width=0.46\columnwidth]{Acc-Add2x2.png}\\
(a)&(b)\\
\end{tabular}
\caption{Epochs v.s. accuracy of ResNet56 on CIFAR10. (a): without the additional FC layer; (b): with the additional FC layer.}
\label{fig:Add2x2}
\end{figure}
Figure~\ref{fig:FeaturesGeometry} (a) and (b) show the 2D features generated by ResNet56 with the additional FC layer for the original and adversarial testing images, respectively, where we generate the adversarial images by using FGSM ($\epsilon=0.02$). Before adversarial perturbation (Fig.~\ref{fig:FeaturesGeometry} (a)), there is a line that can separate the two classes very well. The small perturbation mixes the features and there is no linear classifier that can easily separate these two classes (Fig.~\ref{fig:FeaturesGeometry} (b)). The first two principle components (PCs) of the 64D features of the clean and adversarial images are shown in Fig.~\ref{fig:FeaturesGeometry} (c) and (d), respectively. Again, the PCs are well separated for clean images, while adversarial images causes overlap.
The bottom charts of Fig.~\ref{fig:FeaturesGeometry} depict the first two PCs of the 64D features output from the layer before the WNLL. The distributions of the unperturbed training and testing data are the same, as illustrated in panels (e) and (f). The new features are better separated which indicates that DNN-WNLL are more accurate and more robust to small random perturbation. Panels (g) and (h) plot the features of the adversarial and TV minimized adversarial images in the test set. The adversarial attacks make the features move towards each other and TVM helps to eliminate the outliers. Based on our computation, the interpolating function on features shown in panels (g) and (h) are significantly more accurate than the softmax classifier as shown in panel (d).
The fact that the adversarial perturbations change the features' distribution was also noticed in \cite{Song:2018}, and \cite{Liao_2018_CVPR}.
\begin{figure}[!ht]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.22\columnwidth]{Feature_Net_2D.png}&
\includegraphics[width=0.22\columnwidth]{Features_Net_FGSM_2D.png}&
\includegraphics[width=0.22\columnwidth]{Features_Net_64D.png}&
\includegraphics[width=0.22\columnwidth]{Features_Net_FGSM_64D.png}\\
(a)&(b)&(c)&(d)\\
\includegraphics[width=0.22\columnwidth]{Features_WNLLTrain.png}&
\includegraphics[width=0.22\columnwidth]{Features_WNLLTesting.png}&
\includegraphics[width=0.22\columnwidth]{Features_WNLLNet_FGSM_Testing.png}&
\includegraphics[width=0.22\columnwidth]{Features_WNLLNet_FGSM_TVM_Testing.png}\\
(e)&(f)&(g)&(h)\\
\end{tabular}
\caption{Visualization of the features learned by DNN with softmax ((a), (b), (c), (d)) and WNLL ((e), (f), (g), (h)) activation functions. (a) and (b) plot the 2D features of the original and adversarial testing images; (c) and (d) are the first two principle components of the 64D features for the original and adversarial testing images, respectively. Charts (e), (f) plot the first two components of the training and testing features learned by ResNet56-WNLL; (g) and (h) show the two principle components of the adversarial and TV minimized adversarial images for the test set.
}
\label{fig:FeaturesGeometry}
\end{figure}
\section{PGD Adversarial Training with Data-Dependent Activation Function}\label{section:PGD}
Image transformation based adversarial defense has been broken recently by circumventing the obfuscated gradient \cite{Athalye:2018}. To train a DNN that is most resistant to adversarial attacks, Madry et al. \cite{Madry:2018}, incorporate the adversarial perturbation into the empirical risk function $\mathbb{E}_{(\mathbf{x}, y)\sim \mathcal{D}}\left[\mathcal{L}(\mathbf{x}, y, \theta)\right]$, where $\mathcal{D}$
is the collection of the pairs of training images and labels, and $\theta$ represents the parameters of the neural nets. The idea of PGD adversarial training is that instead of feeding samples from $\mathcal{D}$ directly into the loss $\mathcal{L}$, we use the adversary to perturb the input first, and then we end up with the following saddle point problem
\begin{equation}
\label{saddle-PGD}
\min_\theta \rho(\theta) = \min_\theta \mathbb{E}_{(\mathbf{x}, y)\sim \mathcal{D}}\left[\max_{\delta\in S}\mathcal{L}(\theta, \mathbf{x}+\delta, y)\right],
\end{equation}
where $\delta$ is the adversarial perturbation. To make the problem (Eq.~(\ref{saddle-PGD})) solvable, the inner maximization problem is relaxed to a strong adversarial attack, say IFGSM. It is argued in \cite{Athalye:2018}, that PGD adversarial training achieves the best resistance to adversarial attacks for CIFAR10 classification. We extend the PGD adversarial training to DNN-WNLL by applying the approximated gradient, Eq.~(\ref{GradientApprox}), to approximately resolve the interior maximization problem. We summarize the PGD adversarial training of DNN-WNLL in Algorithm~\ref{PGD-WNLL}.
\subsection{Numerical Results}
We consider PGD adversarial training, respectively, for the ResNet20 and ResNet20-WNLL. Again, we train the ResNet20 with two types of activation, where we follow the strategy used in \cite{BaoWang:2018NIPS}, and where all the hyper-parameters in Algorithm~\ref{PGD-WNLL} are referred. To approximate $\max_{\delta\in S}\mathcal{L}(\theta, \mathbf{x}+\delta, y)$, we apply the IFGSM attack with $\alpha=8/255$ in Eqs.~(\ref{IFGSM-2}, \ref{IFGSM-WNLL}).
First, we fixed the attack strength $\epsilon=1/255$ and vary the number of IFGSM iterations. As shown in Fig.~\ref{fig:Defense-Res2} (a), the accuracy of ResNet20 with both activations decreases as the number of iteration increases. The vanilla ResNet20's accuracy decays much faster than the ResNet20-WNLL. The difference is $\sim 23\%$ when 10 iterations of IFGSM is applied. Second, we fixed the IFGSM iteration to be 10 and vary $\epsilon$ from $0$ to $8/255$ with step size $1/255$. As shown in Fig.~\ref{fig:Defense-Res2} (b), for different nonzero attack strengths, PGD adversarial training of the ResNet20-WNLL has $\sim 23\%$ higher accuracy than the vanilla one consistently.
\begin{figure}[!ht]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\columnwidth]{Acc-Iters.png}&
\includegraphics[width=0.45\columnwidth]{PGD_Defense.png}\\
(a) & (b)\\
\end{tabular}
\caption{(a): $\#$IFGSM iterations v.s. accuracy for the ResNet20 and the ResNet20-WNLL trained with PGD adversarial training. (b):$\epsilon$ v.s. accuracy for the ResNet20 and the ResNet20-WNLL trained with PGD adversarial training.}
\label{fig:Defense-Res2}
\end{figure}
\section{Concluding Remarks}
In this paper, by analyzing the influence of adversarial perturbations on the geometric structure of the DNN features, we propose to defend against adversarial attacks by using a data-dependent activation function. We further show our defenses are additive to other defenses, namely total variation minimization, training data augmentation, and projected gradient descent adversarial training. Results on ResNet20 and ResNet56 with CIFAR10 benchmark reveal that these defenses improve robustness to adversarial perturbation significantly. Total variation minimization simplifies the adversarial images, which is very useful in removing adversarial perturbation. The data-dependent activation framework raises the accuracy of PGD adversarial training around $23\%$ under different attack strengths. An interesting direction to explore is to combine these methods with other denoising methods to remove adversarial perturbation.
\section*{Acknowledgments} This material is based on research sponsored by the National Science Foundation under grant number DMS-1924935 and DMS-1554564 (STROBE). The Air Force Research Laboratory under grant numbers FA9550-18-0167 and MURI FA9550-18-1-0502, the Office of Naval Research under grant number N00014-18-1-2527. ALB is partially supported by the Simons Math $+$ X award.
|
\section{Introduction}
Let $[n]$ be the standard $n$-element set. For an integer $0\leq k\leq n$, denote ${[n]\choose k}$ as the family of all $k$-element subsets of $[n]$. A family $\mathcal{F}$ is said to be $intersecting$ if $A\cap B\neq \emptyset$, for any $A,B\in\mathcal{F}$. The celebrated Erd\H{o}s--Ko--Rado theorem~\cite{EKR} says that if $\mathcal{F}\subseteq {[n]\choose k}$ is an intersecting family with $1\leq k\leq \frac{n}{2}$, then
\begin{equation*}
|\mathcal{F}|\leq {{n-1}\choose {k-1}},
\end{equation*}
and if $n>2k$, the equality holds if and only if every subset in $\mathcal{F}$ contains a fixed element.
Because of its fundamental status in extremal set theory, the Erd\H{o}s--Ko--Rado theorem has numerous extensions in different ways. One of the directions is to study $cross$-$t$-$intersecting$ families: Denote $2^{[n]}$ as the $power~set$ of $[n]$, let $\mathcal{A}_i\subseteq 2^{[n]}$ for each $1\leq i\leq m$, $\mathcal{A}_1,\mathcal{A}_2,\ldots,\mathcal{A}_m$ are said to be cross-t-intersecting, if $|A\cap B|\geq t$ for any $A\in\mathcal{A}_i$ and $B\in\mathcal{A}_j$, $i\neq j$. Especially, we say $\mathcal{A}_1,\mathcal{A}_2,\ldots,\mathcal{A}_m$ are cross-intersecting if $t=1$.
Hilton~\cite{H77} investigated the cross-intersecting families in ${[n]\choose k}$, and proved the following inequality:
\begin{thm}\emph{(\cite{H77})}\label{h77}
Let $\mathcal{A}_1,\mathcal{A}_2,\ldots, \mathcal{A}_m$ be cross-intersecting families in ${[n]\choose k}$ with $n\geq 2k$. Then
\begin{flalign*}
\sum\limits_{i=1}^{m}|\mathcal{A}_i|\leq\left\{\begin{array}{ll}{n\choose k},&~~~\text{if}~m\leq\frac{n}{k};\\
m\cdot{{n-1}\choose {k-1}},&~~~\text{if}~m\geq\frac{n}{k}.\end{array}\right.
\end{flalign*}
\end{thm}
In the same paper, Hilton also determined the structures of $\mathcal{A}_i$'s when the equality holds. Since then, there have been many extensions about Theorem~\ref{h77}. Borg~\cite{B09} gave a simple proof of Theorem~\ref{h77}, and generalized it to labeled sets~\cite{B08}, signed sets~\cite{BL10} and permutations~\cite{B10}. Using the results of the independent number about vertex-transitive graphs, Wang and Zhang~\cite{WZ11} extended this theorem to general symmetric systems, which comprise finite sets, finite vector spaces and permutations, etc.
Hilton and Milner~\cite{HM67} also dealt with pairs of cross-intersecting families in ${[n]\choose k}$ when neither of the two families is empty:
\begin{thm}\emph{(\cite{HM67})}\label{hm67}
Let $\mathcal{A},\mathcal{B}\subseteq {[n]\choose k}$ be non-empty cross-intersecting families with $n\geq 2k$. Then $|\mathcal{A}|+|\mathcal{B}|\leq {n\choose k}-{{n-k}\choose k}+1$.
\end{thm}
This result was generalized by Frankl and Tokushige~\cite{FT92} to the case when $\mathcal{A}$ and $\mathcal{B}$ are not necessarily in the same $k$-uniform subfamily of $2^{[n]}$:
\begin{thm}\emph{(\cite{FT92})}\label{FT92}
Let $\mathcal{A}\subseteq {[n]\choose a}$ and $\mathcal{B}\subseteq {[n]\choose b}$ be non-empty cross-intersecting families with $n\geq a+b$, $a\leq b$. Then $|\mathcal{A}|+|\mathcal{B}|\leq {n\choose b}-{{n-a}\choose b}+1$.
\end{thm}
In~\cite{WZ13}, Wang and Zhang generalized Theorem~\ref{FT92} to cross-t-intersecting families. Recently, using shifting techniques, Frankl and Kupavskii~\cite{FK17} gave another proof of the result of Wang and Zhang for the case when $\mathcal{A},\mathcal{B}\subseteq{[n]\choose k}$.
As another direction, the multi-part extension of the Erd\H{o}s--Ko--Rado problem was introduced by Frankl~\cite{F96}, in connection with a similar result of Sali~\cite{Sali92}. For an integer $p\geq 1$ and positive integers $n_1,\ldots, n_p$, take $[\sum_{i\in[p]}n_i]$ to be the ground set. Then this ground set can be viewed as the disjoint union of $p$ parts $\bigsqcup_{i\in[p]}S_i$, where $S_1=[n_1]$ and $S_i=\{1+\sum_{j\in[i-1]}n_j,\ldots,\sum_{j\in[i]}n_j\}$ for $2\leq i\leq p$. More generally, denote $2^{S_i}$ as the $power~ set$ of $S_i$, for sets $A_1\in 2^{S_1},\ldots, A_p\in 2^{S_p}$, let $\bigsqcup_{i\in[p]} A_i$ be the subset of $\bigsqcup_{i\in[p]}S_i$ with $A_i$ in the $i$-th part, and for families $\mathcal{F}_1\subseteq 2^{S_1},\ldots,\mathcal{F}_p\subseteq 2^{S_p}$, let $\prod_{i\in[p]}\mathcal{F}_i=\{\bigsqcup_{i\in[p]} A_i:~A_i\in \mathcal{F}_i\}$. Then consider $k_1\in[n_1],\ldots, k_p\in[n_p]$, we denote $\prod_{i\in[p]}{[n_i]\choose k_i}$ as the family of all subsets of $\bigsqcup_{i\in[p]}S_i$ which have exactly $k_i$ elements in the $i$-th part. Therefore, families of the form $\mathcal{F}\subseteq \prod_{i\in [p]}{[n_i]\choose k_i}$ can be viewed as the natural generalization of $k$-uniform families to the multi-part setting. Similarly, a multi-part family is intersecting if any two sets of this family intersect in at least one of the $p$ parts.
Frankl proved that for any integer $p\geq 1$, any positive integers $n_1,\ldots,n_p$ and $k_1,\ldots,k_p$ satisfying $\frac{k_1}{n_1}\leq\ldots\leq\frac{k_p}{n_p}\leq\frac{1}{2}$, if $\mathcal{F}\subseteq\prod_{i\in[p]}{[n_i]\choose k_i}$ is a multi-part intersecting family, then
\begin{equation*}
|\mathcal{F}|\leq\frac{k_p}{n_p}\cdot\prod_{i\in[p]}{n_i\choose k_i}.
\end{equation*}
This bound is sharp, for example, it is attained by the following family:
\begin{equation*}
\{A\in{[n_p]\choose k_p}:~i\in A,\text{~for~some~}i\in[n_p]\}\times\prod_{i\in[p-1]}{[n_i]\choose k_i}.
\end{equation*}
Recall that the $Kneser~graph$ $KG_{n,k}$ is the graph on the vertex set ${[n]\choose k}$, with $A,B\in{[n]\choose k}$ forming an edge if and only if $A\cap B=\emptyset$. And an intersecting subfamily of ${[n]\choose k}$ corresponds to an independent set in $KG_{n,k}$. Hence an intersecting subfamily of $\prod_{i\in[p]}{[n_i]\choose k_i}$ corresponds to an independent set in the graph (direct) product $KG_{n_1,k_1}\times\ldots\times KG_{n_p,k_p}$. Therefore, Frankl's result can be viewed as a consequence of the general fact that $\alpha(G\times H)=\max\{\alpha(G)|H|,\alpha(H)|G|\}$ for vertex-transitive graphs $G$ and $H$, which was proved by Zhang in~\cite{zhang2012independent}.
Recently, Kwan, Sudakov and Vieira~\cite{KSV18} considered a stability version of the Erd\H{o}s--Ko--Rado theorem in the multi-part setting. They determined the maximum size of the non-trivially intersecting multi-part family when all the $n_i$'s are sufficiently large. This disproved a conjecture proposed by Alon and Katona, which was also mentioned in~\cite{Katona2017}.
In this paper, we extend Theorem~\ref{h77} and Theorem~\ref{FT92} to the $multi$-$part$ version. For any subset $S\subseteq[n]$, denote $\bar{S}$ as the complementary set of $S$ in $[n]$. Moreover, let $\mathcal{F}$ be a family of subsets of $[n]$, for any subfamily $\mathcal{A}\subseteq\mathcal{F}$, denote ${\mathcal{A}}_{\mathcal{F}}=\{B\in\mathcal{F}:A\cap B=\emptyset\text{ for some }A\in\mathcal{A}\}$.
Our main results are as follows.
\begin{thm}\label{main0}
Given a positive integer $p$, let $n_1,n_2,\ldots,n_p$ and $k_1,k_2,\ldots,k_p$ be positive integers satisfying $n_i\geq 2k_i$ for all $i\in[p]$. Let $X=\prod_{i\in[p]}{[n_i]\choose k_i}$ and $\mathcal{A}_1,\mathcal{A}_2,\ldots,\mathcal{A}_m$ be cross-intersecting families over $X$ with $\mathcal{A}_1\neq \emptyset$. Then
\begin{flalign}\label{bound1}
\sum\limits_{i=1}^{m}|\mathcal{A}_i|\leq\left\{\begin{array}{ll}|X|,&~~~\text{if}~m\leq\min\limits_{i\in[p]}\frac{n_i}{k_i};\\m\cdot M,&~~~\text{if}~m\ge\min\limits_{i\in[p]}\frac{n_i}{k_i},\end{array}\right.
\end{flalign}
where $M=\max_{i\in[p]}{n_i-1\choose k_i-1}\prod_{j\ne i}{n_j\choose k_j}$. Furthermore, the bound is attained if and only if one of the following holds:
\begin{itemize}
{\item[(i)] $m<\min_{i\in[p]}\frac{n_i}{k_i}$ and $\mathcal{A}_1=X$, $\mathcal{A}_2=\cdots=\mathcal{A}_m=\emptyset$;}
{\item[(ii)] $m>\min_{i\in[p]}\frac{n_i}{k_i}$ and $\mathcal{A}_1=\cdots=\mathcal{A}_m=I$, where $I$ is a maximum intersecting family in $X$;}
{\item[(iii)] $m=\min_{i\in[p]}\frac{n_i}{k_i}$ and $\mathcal{A}_1,\ldots,\mathcal{A}_m$ are as in $(i)$ or $(ii)$, or there exists a non-empty set $S_1\subseteq\{s\in[p]:~\frac{n_s}{k_s}=2\}$ and $\mathcal{F}=\prod_{s\in S_1}{[n_s]\choose k_s}$ such that
\begin{flalign}\label{formula}
\begin{array}{l}
\mathcal{A}_1=(\mathcal{A}\sqcup(\mathcal{E}\cup \mathcal{E}_{\mathcal{F}}))\times\prod\limits_{s\in [p]\setminus S_1}{[n_s]\choose k_s}~\text{and}~
\mathcal{A}_2=(\mathcal{A}\sqcup(\mathcal{E}'\cup\mathcal{E}'_{\mathcal{F}}))\times\prod\limits_{s\in [p]\setminus S_1}{[n_s]\choose k_s}
\end{array}
\end{flalign}
for some $\mathcal{A}$, $\mathcal{E}$, $\mathcal{E}'\subseteq\mathcal{F}$, where $\mathcal{A}=\{A_1,\ldots,A_{w_0}\}$ satisfying $2w_0< |\mathcal{F}|$ and $A_i\ne \bar{A}_j$ for all $i\ne j\in [w_0]$, $\mathcal{E}\sqcup\mathcal{E}'=\{E_1,\ldots,E_v\}$ and $\mathcal{E}_\mathcal{F}\sqcup\mathcal{E}'_\mathcal{F}=\{\bar{E}_1,\ldots,\bar{E}_v\}$ satisfying $2(v+w_0)=\prod_{s\in S_1}{n_s\choose k_s}$ and $\sqcup_{j=1}^v\{E_j,\bar{E}_j\}=\mathcal{F}\setminus(\mathcal{A}\sqcup \mathcal{A}_\mathcal{F})$.
}
\end{itemize}
\end{thm}
\begin{rem}\label{rem1}
In \cite{WZ11}, the authors proved a similar result (Theorem 2.5 in \cite{WZ11}) for general connected symmetric systems. Actually, Theorem \ref{main0} can be viewed as an application of the method involved to obtain Theorem 2.5 in \cite{WZ11}. But different from the general case, Theorem \ref{main0} determines all the exact structures when the bound in (\ref{bound1}) is attained.
\end{rem}
\begin{thm}\label{main1}
For any $p\geq 2$, let $n_i, t_i, s_i$ be positive integers satisfying $n_i\geq s_i+t_i+1$, $2\leq s_i,t_i\leq \frac{n_i}{2}$ for every $i\in[p]$ and $n_i\leq\frac{7}{4}n_j$ for all distinct $i,j\in [p]$. If $\prod_{i\in [p]}{n_i\choose s_i}\geq \prod_{i\in [p]}{n_i\choose t_i}$ and $\mathcal{A}\subseteq\prod_{i\in [p]}{[n_i]\choose t_i},~\mathcal{B}\subseteq\prod_{i\in [p]}{[n_i]\choose s_i}$ are non-empty cross-intersecting families, then
\begin{equation}\label{eq02}
|\mathcal{A}|+|\mathcal{B}|\leq \prod_{i\in [p]}{n_i\choose s_i}-\prod_{i\in [p]}{{n_i-t_i}\choose s_i}+1,
\end{equation}
and the bound is attained if and only if the following holds:
\begin{itemize}
{\item[(i)] $\prod_{i\in [p]}{n_i\choose s_i}\geq \prod_{i\in [p]}{n_i\choose t_i}$, $\mathcal{A}=\{A\}$ and $\mathcal{B}=\{B\in \prod_{i\in [p]}{[n_i]\choose s_i}:B\cap A\neq\emptyset\}$ for some $A\in \prod_{i\in [p]}{[n_i]\choose t_i}$;}
{\item[(ii)] $\prod_{i\in [p]}{n_i\choose s_i}= \prod_{i\in [p]}{n_i\choose t_i}$, $\mathcal{B}=\{B\}$ and $\mathcal{A}=\{A\in \prod_{i\in [p]}{[n_i]\choose t_i}:B\cap A\neq \emptyset\}$ for some $B\in \prod_{i\in [p]}{[n_i]\choose s_i}$.}
\end{itemize}
\end{thm}
\begin{rem}\label{condition_of_main1}
The restrictions $s_i, t_i\leq \frac{n_i}{2}$ for every $i\in[p]$ and $n_i\leq\frac{7}{4}n_j$ for all distinct $i,j\in [p]$ in Theorem~\ref{main1} are necessary.
When $s_i, t_i\leq \frac{n_i}{2}$ is violated, for example, taking $n_1=n_2=18$, $(s_1,t_1)=(15,2)$ and $(s_2,t_2)=(2,3)$, we have $s_1>\frac{n_1}{2}$ and ${n_1\choose s_1}\cdot{n_2\choose s_2}={n_1\choose t_1}\cdot{n_2\choose t_2}$. Set $\mathcal{A}=\{A_1\}\times{[n_2]\choose t_2}$ for some 2-subset $A_1\subseteq [n_1]$ and $\mathcal{B}=\{B_1\in {[n_1]\choose s_1}:B_1\cap A_1\neq\emptyset\}\times{[n_2]\choose s_2}$. Then we have $|\mathcal{A}|+|\mathcal{B}|>{n_1\choose s_1}\cdot{n_2\choose s_2}-{{n_1-t_1}\choose s_1}\cdot{{n_2-t_2}\choose s_2}+1$.
As for the restriction $n_i\leq\frac{7}{4}n_j$, the constant $\frac{7}{4}$ here might not be tight, but the quantities of $n_i,n_j$ for distinct $i,j\in[p]$ need to be very close. For example, taking $n_1=5,n_2=12$ and $(s_1,t_1)=(s_2,t_2)=(2,2)$, we have $n_2>\frac{7}{4}n_1$ and ${n_1\choose s_1}\cdot{n_2\choose s_2}={n_1\choose t_1}\cdot{n_2\choose t_2}$. Similarly, set $\mathcal{A}=\{A_1\}\times{[n_2]\choose t_2}$ for some 2-subset $A_1\subseteq [n_1]$ and $\mathcal{B}=\{B_1\in {[n_1]\choose s_1}:B_1\cap A_1\neq\emptyset\}\times{[n_2]\choose s_2}$. Then we have $|\mathcal{A}|+|\mathcal{B}|>{n_1\choose s_1}\cdot{n_2\choose s_2}-{{n_1-t_1}\choose s_1}\cdot{{n_2-t_2}\choose s_2}+1$.
The families $\mathcal{A}$ and $\mathcal{B}$ we constructed here are closely related to imprimitive subsets of ${[n_1]\choose t_1}\times{[n_2]\choose t_2}$, which we will discuss later in Section 2.2 and Section 4.
\end{rem}
We shall introduce some results about the independent sets of vertex-transitive graphs and their direct products in the next section, and prove Theorem~\ref{main0} in Section 3, Theorem~\ref{main1} in Section 4. In Section 5, we will conclude the paper and discuss some remaining problems. For the convenience of the proof, if there is no confusion, we will denote $\prod_{i\in[p]}A_i$ as the subset $\bigsqcup_{i\in[p]}A_i\subseteq \bigsqcup_{i\in[p]}S_i$ in the rest of the paper.
\section{Preliminary results}
\subsection{Independent sets of vertex-transitive graphs}
Given a finite set $X$, for every $A\subseteq X$, denote $\bar{A}=X\setminus A$. For a simple graph $G=G(V,E)$, denote $\alpha(G)$ as the independent number of $G$ and $I(G)$ as the set of all maximum independent sets of $G$. For $v\in V(G)$, define the neighborhood $N_G(v)=\{u\in V(G):(u,v)\in E(G)\}$. For a subset $A\subseteq V(G)$, write $N_G(A)=\{b\in V(G):(a,b)\in E(G)$ for some $a\in A\}$ and $N_G[A]=A\cup N_G(A)$, if there is no confusion, we denote them as $N(A)$ and $N[A]$ for short respectively.
A graph $G$ is said to be vertex-transitive if its automorphism group $\Gamma(G)$ acts transitively upon its vertices. As a corollary of the ``No-Homomorphism'' lemma for vertex-transitive graphs in~\cite{albertson1985homomorphisms}, Cameron and Ku~\cite{CK03} proved the following theorem.
\begin{thm}\label{CK}\emph{(\cite{CK03})}
Let $G$ be a vertex-transitive graph and $B$ a subset of $V(G)$. Then any independent set $S$ in $G$ satisfies that $\frac{|S|}{|V(G)|}\leq\frac{\alpha(G[B])}{|B|}$, equality implies that $|S\cap B|=\alpha(G[B])$.
\end{thm}
Using the above theorem, Zhang~\cite{zhang2011primitivity} proved the following result.
\begin{lem}\emph{(\cite{zhang2011primitivity})}\label{subgraph independent set}
Let $G$ be a vertex-transitive graph, and $A$ be an independent set of $G$, then $\frac{|A|}{|N_G[A]|}\leq\frac{\alpha(G)}{|G|}$. Equality implies that $|S\cap N_G[A]|=|A|$ for every $S\in I(G)$, and in particular $A\subseteq S$ for some $S\in I(G)$.
\end{lem}
An independent set $A$ in $G$ is said to be imprimitive if $|A|<\alpha(G)$ and $\frac{|A|}{|N[A]|}=\frac{\alpha(G)}{|G|}$, and $G$ is called IS-imprimitive if $G$ has an imprimitive independent set. Otherwise, $G$ is called IS-primitive. Note that a disconnected vertex-transitive graph $G$ is always IS-imprimitive. Hence IS-primitive vertex-transitive graphs are all connected.
The following inequality about the size of an independent set $A$ and its non-neighbors $\bar{N}[A]$ is crucial for the proof of Theorem~\ref{main0}.
\begin{lem}\label{lemmaimprimitive} Let $G$ be a vertex transitive graph, and let $A$ be an independent set of $G$. Then
\begin{flalign}
|A|+\frac{\alpha(G)}{|G|}|\bar{N}[A]|\leq\alpha(G).\nonumber
\end{flalign}
Equality holds if and only if $A=\emptyset$ or $|A|=\alpha(G)$ or $A$ is an imprimitive independent set.
\end{lem}
For the integrity of the paper, we include the proof here. In~\cite{WZ11}, Wang and Zhang proved the same inequality for a more generalized combinatorial structure called $symmetric$ $system$ (see~\cite{WZ11}, Corollary 2.4).
\begin{proof}
If $A=\emptyset$ or $A=\alpha(G)$, the equality trivially holds. Suppose $0<|A|<\alpha(G)$, and let $B$ be a maximal independent set in $\bar{N}[A]$, then $|B|=\alpha(\bar{N}(A))$. Clearly, $A\cup B$ is also an independent set of $G$, thus we have $|A|+|B|\leq\alpha(G)$. By Theorem $\ref{CK}$, we obtain that $\frac{|B|}{|\bar{N}[A]|}\ge\frac{\alpha(G)}{|G|}$. Therefore,
\begin{flalign}
|A|+\frac{\alpha(G)}{|G|}|\bar{N}[A]|\leq|A|+|B|\leq\alpha(G),\nonumber
\end{flalign}
the equality holds when $\alpha(G)=|A|+\frac{\alpha(G)}{|G|}|\bar{N}[A]|=|A|+\frac{\alpha(G)}{|G|}(|G|-|N[A]|)$, which leads to $\frac{|A|}{|N[A]|}=\frac{\alpha(G)}{|G|}$, i.e., $A$ is an imprimitive independent set.
\end{proof}
Let $X$ be a finite set, and $\Gamma$ a group acting transitively on $X$. Then $\Gamma$ is said to be primitive on $X$ if it preserves no nontrivial partition of $X$. A vertex-transitive graph $G$ is called primitive if the automorphism $\text{Aut}(G)$ is primitive on $V(G)$. To show the connection between the primitivity and the IS-primitivity of a vertex-transitive graph $G$, Zhang (see Proposition 2.4 in \cite{zhang2011primitivity}) proved that if $G$ is primitive, then it must be IS-primitive. As a consequence of this result, Wang and Zhang~\cite{WZ11} derived the IS-primitivity of the Kneser graph.
\begin{prop}\emph{(\cite{WZ11})}\label{imprimitive Kneser graph}
The Kneser graph $KG_{n,k}$ is IS-primitive except for $n=2k\ge 4$.
\end{prop}
In order to deal with the multi-part case, we also need the results about the independent sets in direct products of vertex-transitive graphs. Let $G$ and $H$ be two graphs, the direct product $G\times H$ of $G$ and $H$ is defined by
\begin{flalign*}
V(G\times H)=V(G)\times V(H),
\end{flalign*}
and
\begin{flalign*}
E(G\times H)=\{[(u_1,v_1),(u_2,v_2)]:(u_1,u_2)\in E(G)\text{ and }(v_1,v_2)\in E(H)\}.
\end{flalign*}
Clearly, $G\times H$ is a graph with $\text{Aut}(G)\times\text{Aut}(H)$ as its automorphism group. And, if $G,H$ are vertex-transitive, then $G\times H$ is also vertex-transitive under the actions of $\text{Aut}(G)\times\text{Aut}(H)$. We say the direct product $G\times H$ is MIS-normal (maximum-independent-set-normal) if every maximum independent set of $G\times H$ is a preimage of an independent set of one factor under projections.
In \cite{zhang2012independent}, Zhang obtained the exact structure of the maximal independent set of $G\times H$.
\begin{thm}\emph{(\cite{zhang2012independent})}\label{tensor product} Let $G$ and $H$ be two vertex-transitive graphs with $\frac{\alpha(G)}{|G|}\ge\frac{\alpha(H)}{|H|}$. Then
\begin{flalign}
\alpha(G\times H)=\alpha(G)|H|,\nonumber
\end{flalign}
and exactly one of the following holds:
\begin{itemize}
\item[(i)] $G\times H$ is MIS-normal;
\item[(ii)] $\frac{\alpha(G)}{|G|}=\frac{\alpha(H)}{|H|}$ and one of $G$ or $H$ is IS-imprimitive;
\item[(iii)] $\frac{\alpha(G)}{|G|}>\frac{\alpha(H)}{|H|}$ and $H$ is disconnected.
\end{itemize}
\end{thm}
In fact if $\frac{\alpha(G)}{|G|}=\frac{\alpha(H)}{|H|}$ and $A$ is an imprimitive independent set of $G$, then for every $I\in I(H)$, it is easy to see that $S=(A\times V(H))\cup (\bar{N}[A]\times I)$ is an independent set of $G\times H$ with size $\alpha(G)|H|$.
Zhang \cite{zhang2011primitivity} also investigated the relationship between the graph primitivity and the structures of the maximum independent sets in direct products of vertex-transitive graphs.
\begin{thm}\emph{(\cite{zhang2011primitivity})}\label{normal-imprimitive} Suppose $G\times H$ is MIS-normal and $\frac{\alpha(H)}{|H|}\leq\frac{\alpha(G)}{|G|}$. If $G\times H$ is IS-imprimitive, then one of the following two possible cases holds:
\begin{itemize}
\item[(i)] $\frac{\alpha(G)}{|G|}=\frac{\alpha(H)}{|H|}$ and one of them is IS-imprimitive or both $G$ and $H$ are bipartite;
\item[(ii)] $\frac{\alpha(G)}{|G|}>\frac{\alpha(H)}{|H|}$ and $G$ is IS-imprimitive.
\end{itemize}
\end{thm}
As an application of Theorem~\ref{tensor product} and Theorem~\ref{normal-imprimitive}, Geng et al.~\cite{geng2012structure} showed the MIS-normality of the direct products of Kneser graphs.
\begin{thm}\emph{(\cite{geng2012structure})}\label{imprimitive product kneser graph} Given a positive integer $p$, let $n_1,n_2,\ldots,n_p$ and $k_1,k_2,\ldots,k_p$ be $2p$ positive integers with $n_i\ge 2k_i$ for $1\leq i\leq p$. Then the direct product of the Kneser graph $$\text{KG}_{n_1,k_1}\times\text{KG}_{n_2,k_2}\times\cdots\times\text{KG}_{n_p,k_p}$$ is MIS-normal except that there exist $i,j$ and $\ell$ with $n_i=2k_i\ge 4$ and $n_j=2k_j$, or $n_i=n_j=n_{\ell}=2$.
\end{thm}
\subsection{Nontrivial independent sets of part-transitive bipartite graphs}
For a bipartite graph $G(X,Y)$ with two parts $X$ and $Y$, an independent set $A$ is said to be non-trivial if $A\nsubseteq X$ and $A\nsubseteq Y$. $G(X,Y)$ is said to be part-transitive if there is a group $\Gamma$ acting transitively upon each part and preserving its adjacency relations. Clearly, if $G(X,Y)$ is part-transitive, then every vertex of $X~(Y)$ has the same degree, written as $d(X)~(d(Y))$. We use $\alpha(X,Y)$ and $I(X,Y)$ to denote the size and the set of the maximum-sized nontrivial independent sets of $G(X,Y)$, respectively.
Let $G(X,Y)$ be a non-complete bipartite graph and let $A\cup B$ be a nontrivial independent set of $G(X,Y)$, where $A\subseteq X$ and $B\subseteq Y$. Then $A\subseteq X\setminus N(B)$ and $B\subseteq Y\setminus N(A)$, which implies
\begin{equation*}
|A|+|B|\leq \max{\{|A|+|Y|-|N(A)|, |B|+|X|-|N(B)|\}}.
\end{equation*}
So we have
\begin{equation}\label{eq03}
\alpha(X,Y)=\max{\{|Y|-\epsilon(X), |X|-\epsilon(Y)\}},
\end{equation}
where $\epsilon(X)=\min\{|N(A)|-|A|: A\subseteq X, N(A)\neq Y\}$ and $\epsilon(Y)=\min\{|N(B)|-|B|: B\subseteq Y, N(B)\neq X\}$.
We call $A\subseteq X$ a fragment of $G(X,Y)$ in $X$ if $N(A)\neq Y$ and $|N(A)|-|A|=\epsilon(X)$, and we denote $\mathcal{F}(X)$ as the set of all fragments in $X$. Similarly, we can define $\mathcal{F}(Y)$. Furthermore, denoting $\mathcal{F}(X,Y)=\mathcal{F}(X)\cup\mathcal{F}(Y)$, we call an element $A\in \mathcal{F}(X,Y)$ a $k$-fragment if $|A|=k$. And we call a fragment $A\in\mathcal{F}(X)$ trivial if $|A|=1$ or $A=X\setminus N(b)$ for some $b\in Y$. Since for each $A\in \mathcal{F}(X)$, $Y\setminus N(A)$ is a fragment in $\mathcal{F}(Y)$. Hence, once we know $\mathcal{F}(X)$, $\mathcal{F}(Y)$ can also be determined.
Let $X$ be a finite set, and $\Gamma$ a group acting transitively on $X$. If $\Gamma$ is imprimitive on $X$, then it preserves a nontrivial partition of $X$, called a block system, each element of which is called a block. Clearly, if $\Gamma$ is both transitive and imprimitive, there must be a subset $B\subseteq X$ such that $1<|B|<|X|$ and $\gamma(B)\cap B=B$ or $\emptyset$ for every $\gamma \in \Gamma$. In this case, $B$ is called an imprimitive set in $X$. Furthermore, a subset $B\subseteq X$ is said to be $semi$-$imprimitive$ if $1<|B|<|X|$ and for each $\gamma\in \Gamma$ we have $\gamma(B)\cap B=B$, $\emptyset$ or $\{b\}$ for some $b\in B$.
The following theorem (cf. \cite[Theorem~1.12]{J85}) is a classical result on the primitivity of group actions.
\begin{thm}\emph{(\cite{J85})}\label{primitivity}
Suppose that a group $\Gamma$ transitively acts on $X$. Then $\Gamma$ is primitive on $X$ if and only if for each $a\in X$, $\Gamma_a$ is a maximal subgroup of $\Gamma$. Here $\Gamma_a=\{\gamma\in\Gamma:\gamma(a)=a\}$, the stabilizer of $a\in X$.
\end{thm}
Noticing the similarities about families that are cross-t-intersecting or cross-Sperner, Wang and Zhang \cite{WZ13} proved the following theorem about $\alpha(G(X,Y))$ and $I(X,Y)$ of a special kind of part-transitive bipartite graphs.
\begin{thm}\emph{(\cite{WZ13})}\label{key00}
Let $G(X,Y)$ be a non-complete bipartite graph with $|X|\leq|Y|$. If $G(X,Y)$ is part-transitive and every fragment of $G(X,Y)$ is primitive under the action of a group $\Gamma$. Then $\alpha(X,Y)=|Y|-d(X)+1$. Moreover,
\begin{itemize}
{\item[(1)] If $|X|<|Y|$, then $X$ has only 1-fragments;}
{\item[(2)] If $|X|=|Y|$, then each fragment in $X$ has size 1 or $|X|-d(X)$ unless there is a semi-imprimitive fragment in $X$ or $Y$.}
\end{itemize}
\end{thm}
To deal with multi-part cross-intersecting families, we introduce the following variation of Theorem \ref{key00}.
\begin{thm}\label{key01}
Let $G(X,Y)$ be a non-complete bipartite graph with $|X|\leq|Y|$. If $G(X,Y)$ is part-transitive under the action of a group $\Gamma$. Then
\begin{equation}\label{eq04}
\alpha(X,Y)=\max{\{|Y|-d(X)+1, |A'|+|Y|-|N(A')|, |B'|+|X|-|N(B')|\}},
\end{equation}
where $A'$ and $B'$ are minimum imprimitive subsets of $X$ and $Y$ respectively. By minimum, here we mean that $|N(A')|-|A'|=\min{\{|N(A)|-|A|: A\in X~(\text{or~} Y) \text{~is~imprimitive}\}}$.
\end{thm}
For the proof of Theorem~\ref{key01}, we need the following two lemmas from \cite{WZ13}.
\begin{lem}\emph{(\cite{WZ13})}\label{lem01}
Let $G(X,Y)$ be a non-complete bipartite graph. Then, $|Y|-\epsilon(X)=|X|-\epsilon(Y)$, and
\begin{itemize}
{\rm\item[(i)] $A\in \mathcal{F}(X)$ if and only if $(Y\setminus N(A))\in \mathcal{F}(Y)$ and $N(Y\setminus N(A))=X\setminus A$;}
{\rm\item[(ii)] $A\cap B$ and $A\cup B$ are both in $\mathcal{F}(X)$ if $A$, $B\in \mathcal{F}(X)$, $A\cap B\neq \emptyset$ and $N(A\cup B)\neq Y$.}
\end{itemize}
\end{lem}
\begin{lem}\emph{(\cite{WZ13})}\label{lem02}
Let $G(X,Y)$ be a non-complete and part-transitive bipartite graph under the action of a group $\Gamma$. Suppose that $A\in\mathcal{F}(X,Y)$ such that $\emptyset\neq\gamma(A)\cap A\neq A$ for some $\gamma\in\Gamma$. Define $\phi: \mathcal{F}(X,Y)\rightarrow\mathcal{F}(X,Y)$,
\begin{flalign*}
\phi (A)=\left\{\begin{array}{ll}Y\setminus N(A),&~~~\text{if}~A\in\mathcal{F}(X);\\X\setminus N(A),&~~~\text{if}~A\in\mathcal{F}(Y).\end{array}\right.
\end{flalign*}
If $|A|\leq|\phi(A)|$, then $A\cup\gamma(A)$ and $A\cap\gamma(A)$ are both in $\mathcal{F}(X,Y)$.
\end{lem}
\begin{rem}\label{balanced}
As a direct consequence of Lemma~\ref{lem01}, a maximum-sized nontrivial independent set in $G(X,Y)$ is of the form $A\sqcup(Y\setminus N(A))$ for some $A\in \mathcal{F}(X)$ or $B\sqcup(X\setminus N(B))$ for some $B\in \mathcal{F}(Y)$. Therefore, in order to address our problems, it suffices to determine $\mathcal{F}(X)$ $(\text{or~}\mathcal{F}(Y))$.
Meanwhile, for the mapping $\phi$ in Lemma~\ref{lem02}, we have $\phi^{-1}=\phi$ and $|A|+|\phi(A)|=\alpha(X,Y)$. When $|A|=|\phi(A)|$, we call the fragment $A$ balanced. Thus, all balanced fragments have size $\frac{1}{2}\alpha(X,Y)$.
\end{rem}
\begin{proof}[\bf{Proof of Theorem~\ref{key01}}]
The same as the original proof of Theorem \ref{key00} in \cite{WZ13}, we apply Lemma~\ref{lem02} repeatedly. For any $A_0\in\mathcal{F}(X,Y)$ satisfying $|A_0|\leq|\phi(A_0)|$, if there exists $\gamma\in\Gamma$ such that $\emptyset\neq\gamma(A_0)\cap A_0\neq A_0$, then by Lemma~\ref{lem02} we have: (1) $A_0\cap\gamma(A_0)\in\mathcal{F}(X,Y)$ or (2) $\gamma(A_0)\cap A_0=\emptyset$ or $\gamma(A_0)\cap A_0=A_0$ for any $\gamma\in\Gamma$.
For case (1), denote
\begin{flalign*}
A_1=\left\{\begin{array}{ll}A_0\cap\gamma(A_0), &~\text{if}~|A_0\cap\gamma(A_0)|\leq|\phi(A_0\cap\gamma(A_0))|;
\\ \phi(A_0\cap\gamma(A_0)), &~\text{~~~~~~~~~~~~~~~otherwise};\end{array}\right.
\end{flalign*}
and consider the primitivity of $A_1$, i.e., whether there is a $\gamma'\in\Gamma$ such that $\emptyset\neq\gamma'(A_1)\cap A_1\neq A_1$ or not.
For case (2), if $|A_0|\neq1$, according to the definition, $A_0$ is an imprimitive set of $X$ (or $Y$). Otherwise, $|A_0|=1$, which means $\mathcal{F}(X,Y)$ contains a singleton.
By doing these procedures repeatedly, after $r$ $(0\leq r\leq |A_0|-1)$ steps, we have a fragment $A_r\in\mathcal{F}(X,Y)$ such that $A_r$ is either a singleton or an imprimitive set. Hence, we have
$$\alpha(X,Y)=\max{\{|Y|-d(X)+1, |X|-d(Y)+1, |A'|+|Y|-|N(A')|, |B'|+|X|-|N(B')|\}},$$
where $A'$ and $B'$ are minimum imprimitive subsets of $X$ and $Y$ respectively. Noticing that $|Y|\geq|X|$ and $d(X)|X|=d(Y)|Y|$, we have $d(X)=d(Y)|Y|/|X|\geq d(Y)$. Therefore,
$$|Y|-|X|=d(X)|X|/d(Y)-|X|=(d(X)-d(Y))|X|/d(Y)\geq d(X)-d(Y),$$
which implies that $|X|-d(Y)+1\leq|Y|-d(X)+1$. Finally we have
$$\alpha(X,Y)=\max{\{|Y|-d(X)+1, |A'|+|Y|-|N(A')|, |B'|+|X|-|N(B')|\}}.$$
\end{proof}
\section{Proof of Theorem~\ref{main0}}
Throughout this section, for any nonempty subset $S\subseteq[p]$ and $A=\prod_{i\in S}A_i\in\prod_{i\in S}{[n_i]\choose k_i}$, denote $\bar{A}=\prod_{i\in S}\bar{A}_i$. Before we start the proof of Theorem~\ref{main0}, we introduce the following proposition about the direct product of Kneser graphs.
\begin{prop}\label{note1}
Given a positive integer $p$, let $n_1,n_2,\ldots,n_p$ and $k_1,k_2,\ldots,k_p$ be positive integers with $n_i\ge 2k_i$ for $1\leq i\leq p$. Let $G=\prod_{i\in[p]}{KG_{n_i,k_i}}$. Then $G$ is IS-imprimitive if and only if there exists an $i\in[p]$ such that $n_i=2k_i\ge 4$ or there exist distinct $i,j\in[p]$ such that $n_i=n_j=2$ and $k_i=k_j=1$.
\end{prop}
\begin{proof}
Note that if the Kneser graph $KG_{n,k}$ is disconnected, then $n=2k\geq 4$ and $KG_{n,k}$ is bipartite. Thus by Proposition \ref{imprimitive Kneser graph}, $KG_{2k,k}$ is IS-imprimitive for all $k\geq 2$. Moreover, since $\chi({KG_{n,k}})=n-2k+2$ for all $n\geq 2k$ (Lov\'{a}sz-Kneser Theorem, see \cite{Lovasz78}), we know that if $KG_{n,k}$ is bipartite, then $n=2k\geq 2$. Now we use induction on the number of factors $p$.
If $p=2$, let $G_1=KG_{n_1,k_1}$, $G_2=KG_{n_2,k_2}$, and $G=G_1\times G_2$. W.l.o.g., assume that $\frac{\alpha(G_1)}{|G_1|}\ge\frac{\alpha(G_2)}{|G_2|}$. Then, by Theorem \ref{tensor product}, (i) $G_1\times G_2$ is MIS-normal, or (ii) $\frac{\alpha(G_1)}{|G_1|}=\frac{\alpha(G_2)}{|G_2|}$ and one of $G_1$ and $G_2$ is IS-imprimitive, or (iii) $\frac{\alpha(G_1)}{|G_1|}>\frac{\alpha(G_2)}{|G_2|}$ and $G_2$ is disconnected. For case (i), by Theorem \ref{normal-imprimitive}, at least one factor of $G$ is IS-imprimitive or both $G_1$ and $G_2$ are bipartite. Noticed that $KG_{2,1}$ is IS-primitive, therefore, either there exists an $i\in [2]$ such that $n_i=2k_i\ge 4$ or there exist distinct $i,j\in[2]$ such that $n_i=n_j= 2k_i=2k_j=2$. For cases (ii) and (iii), since $G$ is not MIS-normal, by Theorem \ref{imprimitive product kneser graph}, at least one of $G_1$ and $G_2$ is IS-imprimitive. Thus the proposition holds when $p=2$.
Suppose the proposition holds when the number of factors is $p-1$. Set $G'_1=\prod_{i=1}^{p-1}KG_{n_i,k_i}$ and $G'_2=KG_{n_p,k_p}$, by Theorem \ref{normal-imprimitive}, at least one factor of $G'_1$ and $G'_2$ is IS-imprimitive or both $G'_1$ and $G'_2$ are bipartite. If $G_1'$ is IS-imprimitive, by the induction hypothesis, there exists an $i'\in[p-1]$ such that $n_{i'}=2k_{i'}\ge 4$ or there exist distinct $i',j'\in[p-1]$ such that $n_{i'}=n_{j'}=2k_{i'}=2k_{j'}=2$. If $G_2'$ is IS-imprimitive, then $n_p=2k_p\geq 4$. Otherwise, both $G'_1$ and $G'_2$ are IS-primitive and bipartite. Thus, for $G'_2$, we have $n_p=2k_p=2$. For $G'_1$, since $\chi(G'_1)\cdot\alpha(G'_1)\ge |V(G'_1)|$, we know that there exists $i'\in[p-1]$ such that $n_{i'}=2k_{i'}=2$ by Lemma \ref{tensor product}. This completes the proof.
\end{proof}
The idea of the proof for Theorem~\ref{main0} is similar to that for general connected symmetric systems in \cite{WZ11}. Since $\prod_{i=1}^{p}\text{KG}_{n_i,k_i}$ is a vertex transitive graph, by Lemma \ref{lemmaimprimitive}, we can prove the bound (\ref{bound1}). Then, through a careful analysis, we can obtain the structure of all imprimitive independent sets of this graph. This leads to the unique structures of $\mathcal{A}_1$ and $\mathcal{A}_2$ in $(\ref{formula})$.
\begin{proof}[\bf{Proof of Theorem~\ref{main0}}]
Define a graph $G$ on the vertex set $X=\prod_{s\in[p]}{[n_s]\choose k_s}$ with $A,B\in X$ forming an edge in $G$ if and only if $A\cap B=\emptyset$. Therefore, $G$ is the direct product of Kneser graphs $\text{KG}_{n_1,k_1}\times\cdots\times\text{KG}_{n_p,k_p}$.
Assume that $2\leq\frac{n_1}{k_1}\leq\frac{n_2}{k_2}\leq\ldots\leq\frac{n_p}{k_p}$, then $\frac{|G|}{\alpha(G)}=\frac{n_1}{k_1}$ by Theorem~\ref{tensor product}. Following the notations of Borg in \cite{B09,B10,BL10}, write $\mathcal{A}^\ast_i=\{A\in\mathcal{A}_i|A\cap B\ne\emptyset\text{ for any }B\in\mathcal{A}_i\}$, $\hat{\mathcal{A}}_i=\mathcal{A}_i\setminus\mathcal{A}_i^\ast$, $\mathcal{A}^\ast=\bigcup_{i=1}^m\mathcal{A}_i^{\ast}$, $\hat{\mathcal{A}}=\bigcup_{i=1}^m\hat{\mathcal{A}}_i$. Note that $\bar{N}_{G}[\mathcal{A}]=\{B\in X|A\cap B\ne\emptyset,\text{ for any }A\in\mathcal{A}\}$ for $\mathcal{A}\subseteq X$, it is easy to show that $\mathcal{A}^\ast$ is an intersecting family and $\hat{\mathcal{A}}\subseteq\bar{N}_{G}[\mathcal{A}^\ast]$. It follows that $\mathcal{A}_i\cap\mathcal{A}_j\subseteq\mathcal{A}_i^\ast\cap\mathcal{A}_j^\ast$ from the definition, therefore $\hat{\mathcal{A}}_i\cap\hat{\mathcal{A}}_j=\emptyset$ for $i\ne j$, and $|\hat{\mathcal{A}}|=\sum_{i=1}^m|\hat{\mathcal{A}}_i|$. Thus by Lemma~\ref{lemmaimprimitive} we have
\begin{align*}\label{imprimitive equation}
\sum\limits_{i=1}^m|\mathcal{A}_i|&=\sum\limits_{i=1}^m|\hat{\mathcal{A}}_i|+\sum\limits_{i=1}^m|\mathcal{A}_i^\ast|\leq|\hat{\mathcal{A}}|+m|\mathcal{A}^\ast|\leq|\bar{N}_{G}[\mathcal{A}^\ast]|+m|\mathcal{A}^\ast|\nonumber\\
&=\frac{|G|}{\alpha(G)}(\frac{\alpha(G)}{|G|}|\bar{N}_{G}[\mathcal{A}^\ast]|+|\mathcal{A}^\ast|)+(m-\frac{|G|}{\alpha(G)})|\mathcal{A}^\ast|\\
&\leq |G|+(m-\frac{|G|}{\alpha(G)})|\mathcal{A}^\ast|=|G|+(m-\frac{n_1}{k_1})|\mathcal{A}^\ast|.\nonumber
\end{align*}
If $m<\frac{n_1}{k_1}$, then $\sum_{i=1}^m|\mathcal{A}_i|\leq|G|$, and the equality implies $\mathcal{A}^\ast=\emptyset$. Thus $\mathcal{A}_i=\hat{\mathcal{A}_i}$ for every $i\in[m]$, and this yields that the graph $G$ is a disjoint union of the induced subgraph $G[\mathcal{A}_i]'s$. And by the cross-intersecting property, each $G[\mathcal{A}_i]$ is a connected component of $G$. Since $G$ is connected when $\frac{n_s}{k_s}>2$ for all $s\in[p]$ and $m\geq 2$, we know that one of $\mathcal{A}_i$ is $X$ and the rest are empty sets, as case (i).
If $m>\frac{n_1}{k_1}$, then $\sum_{i=1}^m|\mathcal{A}_i|\leq m\alpha(G)$, and the equality implies that $\mathcal{A}_1^\ast=\ldots=\mathcal{A}_m^\ast=\mathcal{A}^\ast$, $|\mathcal{A}^\ast|=\alpha(G)$, as case (ii).
If $m=\frac{n_1}{k_1}$, then $\sum_{i=1}^m|\mathcal{A}_i|\leq|X|$, and the equality implies that $\mathcal{A}_1^\ast=\ldots=\mathcal{A}_m^\ast=\mathcal{A}^{\ast}$ and $\frac{\alpha(G)}{|G|}|\bar{N}_{G}[\mathcal{A}^\ast]|+|\mathcal{A}^\ast|=\alpha(G)$. By Lemma \ref{lemmaimprimitive}, we know that $|\mathcal{A}^\ast|=0$, or $|\mathcal{A}^\ast|=\alpha(G)$, or $\mathcal{A}^\ast$ is an imprimitive independent set of $G$. In the last case, $\hat{\mathcal{A}}_1,\ldots,\hat{\mathcal{A}}_m$ are cross-intersecting families and form a partition of $\bar{N}_{G}[\mathcal{A}^\ast]$. In order to determine the structures of the maximum-sized cross-intersecting families in this case, we shall characterize the imprimitive independent set of $G$.
\begin{claim}\label{imprimitive independent set}
Let $\mathcal{F}=\prod_{s\in S}{[n_s]\choose k_s}$ and $X'=\prod_{s\in[p]\setminus S}{[n_s]\choose k_s}$, where $S=\{s\in [p]:~\frac{n_s}{k_s}=2\}$. If $\mathcal{A}^\ast$ is an imprimitive independent set of $G$, then $\mathcal{A}^\ast=\mathcal{A}\times X'$, where $\mathcal{A}\subseteq\mathcal{F}$ is a non-maximum intersecting family.
\end{claim}
According to Proposition \ref{note1}, $G$ is IS-imprimitive if and only if there exists an $i\in S$ such that $n_i=2k_i\ge 4$ or there exist distinct $i,j\in S$ such that $n_i=n_j=2$ and $k_i=k_j=1$. Thus, with the assumptions in this claim, $S\neq \emptyset$ and $S=\{i_0\}$ if and only if $n_{i_0}=2k_{i_0}\geq 4$ for some $i_0\in[p]$. W.l.o.g., assume that $S=[s_0]$, where $s_0=|S|$. Under this circumstance, $m=\frac{n_1}{k_1}=2$.
Divide $\mathcal{A}^\ast$ into $u$ disjoint parts $\{C_i\times \mathcal{D}_i\}_{i=1}^{u}$, where $C_i=C_{i,1}\times\ldots\times C_{i,s_0}\in\mathcal{F}$, $\mathcal{D}_i\subseteq X'$ for all $i\in [u]$ and $C_i\ne C_j$ for any $i\ne j\in[u]$. Since $N_G(C_i\times \mathcal{D}_i)=\bar{C_i}\times\mathcal{D}_i'$, where $\mathcal{D}'_i=\{A\in X': A\cap D_i=\emptyset\text{~for~some~} D_i\in\mathcal{D}_i\}$, we know that $N_G[C_i\times \mathcal{D}_i]\cap N_G[C_j\times \mathcal{D}_j]=\emptyset$ for all $i\ne j\in[u]$. Meanwhile, $C_i\times \mathcal{D}_i\cap N_G(C_j\times \mathcal{D}_j)=\emptyset$ for all $i\ne j\in[u]$. Otherwise, assume that there exists $T_1\times T_2\in C_i\times\mathcal{D}_i\cap N_G(C_j\times\mathcal{D}_j)$, for some $T_1\in\mathcal{F}$ and $T_2\in X'$. Thus we have $T_1\times T_2\cap C_j\times D_j=\emptyset$, for some $D_j\in\mathcal{D}_j$, which contradicts the fact that $\mathcal{A}^\ast$ is an intersecting family.
By projecting $G$ onto the last $p-s_0$ factors, we obtain a graph $G'$ with vertex set $X'$ such that $A,B\in X'$ form an edge in $G'$ if and only if $A,B$ are disjoint. Consider the cross-intersecting families $\{\mathcal{D}_i,\bar{N}_{G'}(\mathcal{D}_i)\}$ in $X'$, since $|\{\mathcal{D}_i,\bar{N}_{G'}(\mathcal{D}_i)\}|=2<\frac{n_{s_0+1}}{k_{s_0+1}}$, by case (i), we know that
\begin{flalign}
|\mathcal{D}_i|+|\bar{N}_{G'}(\mathcal{D}_i)|=|\mathcal{D}_i|+|X'|-|N_{G'}(\mathcal{D}_i)|\leq|X'|,\nonumber
\end{flalign}
thus we have $|\mathcal{D}_i|\leq|N_{G'}(\mathcal{D}_i)|$, and $|C_i\times \mathcal{D}_i|=|\mathcal{D}_i|\leq |N_{G'}(\mathcal{D}_i)|=|N_G(C_i\times \mathcal{D}_i)|$. Therefore
\begin{flalign*}
\frac{|\mathcal{A}^\ast|}{|N_G[\mathcal{A}^\ast]|}=\frac{\sum_{i\in[u]}|C_i\times\mathcal{D}_i|}{\sum_{i\in[u]}|N_G[C_i\times\mathcal{D}_i]|}\leq\frac{1}{2}=\frac{\alpha(G)}{|G|}=\frac{k_1}{n_1},
\end{flalign*}
and the equality holds if and only if for all $i\in[u]$, $\mathcal{D}_i=X'$ or $\bar{N}_{G'}(\mathcal{D}_i)=X'$. Since $\mathcal{D}_i\ne\emptyset$, we have $\mathcal{A}^\ast=\bigsqcup_{i=1}^{u}C_i\times X'=\mathcal{A}\times X'$. Recall that $\frac{n_s}{k_s}>2$ for all $s>s_0$, hence $C_i\cap C_j\neq \emptyset$ for any $i\neq j\in[u]$. Therefore, by the imprimitivity of $\mathcal{A}^\ast$, $\mathcal{A}^\ast$ is a non-maximum independent set of $G$, thus $\mathcal{A}\subseteq\mathcal{F}$ is a non-maximal intersecting family and the claim holds.
For every intersecting family $\mathcal{A}\subseteq\mathcal{F}$, since $\frac{n_s}{k_s}=2$ for all $s\in {S}$, then $\mathcal{A}=\{A_1,A_2,\ldots,A_w\}\times\prod_{s\in S\setminus S'}{[n_s]\choose k_s}$ for some nonempty subset $S'\subseteq S$, where $\{A_1,\ldots,A_w\}\subseteq\prod_{s\in S'}{[n_s]\choose k_s}$ satisfying $A_i\neq \bar{A}_j$ for all $i\neq j\in [w]$. In particular, if $\mathcal{A}$ is a maximum intersecting family, we can obtain that $\bigsqcup_{j=1}^{w}\{A_j,\bar{A}_j\}=\prod_{s\in S'}{[n_s]\choose k_s}$ and $2w=\prod_{s\in S'}{n_s\choose k_s}$.
Therefore, $\mathcal{A}^\ast=\{A_1,A_2,\ldots,A_{w_0}\}\times\prod_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X'$ and $N_G(\mathcal{A}^\ast)=\{\bar{A}_1,\bar{A}_2,\ldots,\bar{A}_{w_0}\}\times\prod_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X'$, for some positive integer $w_0<\frac{\prod_{s\in S_1}{n_s\choose k_s}}{2}$ and nonempty subset $S_1\subseteq S$.
From the structure of the imprimitive independent set $\mathcal{A}^\ast$, we know that
\begin{flalign*}
\bar{N}_G[\mathcal{A}^\ast]=\{E_1,\bar{E}_1,E_2,\bar{E}_2,\ldots,E_v,\bar{E}_v\}\times\prod\limits_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X',
\end{flalign*}
where $\emptyset\neq\{E_1,\ldots,E_v\}\subseteq\prod_{s\in S_1}{[n_s]\choose k_s}$, and $\bigsqcup_{j=1}^{w_0}\{A_j,\bar{A}_j\}\sqcup\bigsqcup_{j=1}^v\{E_j,\bar{E}_j\}=\prod_{s\in S_1}{[n_s]\choose k_s}$.
Since $E_j\times\prod_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X'$ and $\bar{E}_j\times\prod_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X'$ must be contained in the same one of $\hat{\mathcal{A}_1}$, $\hat{\mathcal{A}_2}$, we have
\begin{flalign}
\hat{\mathcal{A}}_1&=(\mathcal{E}\cup\tilde{\mathcal{E}})\times \prod\limits_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X',\nonumber\\
\hat{\mathcal{A}}_2&=(\mathcal{E}'\cup\tilde{\mathcal{E}}')\times \prod\limits_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X',\nonumber
\end{flalign}
where $\mathcal{E}\sqcup\mathcal{E}'=\{E_1,\ldots,E_v\}$ and $\tilde{\mathcal{E}}\sqcup\tilde{\mathcal{E}}'=\{\bar{E}_1,\ldots,\bar{E}_v\}$. Here we denote $\tilde{\mathcal{E}}=\{\bar{E}_{i_1},\ldots,\bar{E}_{i_l}\}$ if $\mathcal{E}=\{E_{i_1},\ldots,E_{i_l}\}\subseteq\prod_{s\in S_1}{[n_s]\choose k_s}$, for some subset $\{i_1,\ldots,i_l\}\subseteq[v]$.
Finally, to sum up,
\begin{flalign}
\mathcal{A}_1&=\mathcal{A}^\ast\sqcup\hat{\mathcal{A}}_1=(\mathcal{A}\times X')\sqcup ((\mathcal{E}\cup\tilde{\mathcal{E}})\times \prod\limits_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X')\nonumber,\\
\mathcal{A}_2&=\mathcal{A}^\ast\sqcup\hat{\mathcal{A}}_2=(\mathcal{A}\times X')\sqcup ((\mathcal{E}'\cup\tilde{\mathcal{E}}')\times \prod\limits_{s\in S\setminus S_1}{[n_s]\choose k_s}\times X')\nonumber.
\end{flalign}
\end{proof}
\section{Proof of Theorem \ref{main1}}
Throughout this section, we denote $S_n$ as the symmetric group on $[n]$ and $S_C$ as the symmetric group on $C$ for $C\subseteq [n]$. For each $i\in[p]$, let $X_i$ be a finite set, then for each family $\mathcal{A}\subseteq \prod_{i\in[p]}X_i$, we denote $\mathcal{A}|_i$ as the projection of $\mathcal{A}$ onto the $i$-th factor.
For the proof of Theorem~\ref{main1}, we need the following proposition obtained by Wang and Zhang in \cite{WZ13}.
\begin{prop}\emph{(\cite{WZ13})}\label{fragment2}
Let $G(X,Y)$ be a non-complete bipartite graph with $|X|=|Y|$ and $\epsilon(X)=d(X)-1$, and let $\Gamma$ be a group part-transitively acting on $G(X,Y)$. If each fragment of $G(X,Y)$ is primitive and there are no $2$-fragments in $\mathcal{F}(X,Y)$, then every nontrivial fragment $A\in \mathcal{F}(X)$ (if there exists) is balanced (see Remark~\ref{balanced}), and for each $a\in A$, there is a unique nontrivial fragment $B$ such that $A\cap B=\{a\}$.
\end{prop}
The proof of Theorem~\ref{main1} is divided into two parts: Firstly, we prove the bound (\ref{eq02}). Consider a non-complete bipartite graph defined by the multi-part cross-intersecting family. Through discussions about the primitivity of group $\prod_{i=1}^{p}S_{n_i}$ and careful evaluations about $|\mathcal{A}|+|\mathcal{Y}|-|N(\mathcal{A})|$, the bound (\ref{eq02}) follows from Theorem \ref{key01}. Secondly, based on a characterization of all nontrivial fragments in this bipartite graph, we determine all the structures of $\mathcal{A}$ and $\mathcal{B}$ when the bound (\ref{eq02}) is attained.
\begin{proof}[\bf{Proof of Theorem~\ref{main1}}]
With the assumptions in the theorem, we define a bipartite graph $G(\mathcal{X},\mathcal{Y})$ with $\mathcal{X}=\prod_{i=1}^{p}{[n_i]\choose t_i}$ and $\mathcal{Y}=\prod_{i=1}^{p}{[n_i]\choose s_i}$. For $A=\prod_{i=1}^{p}{A_i}\in \mathcal{X}$ and $B=\prod_{i=1}^{p}{B_i}\in \mathcal{Y}$ ($A_i\in {[n_i]\choose t_i}$ and $B_i\in {[n_i]\choose s_i}$, for every $1\leq i\leq p$), $(A,B)$ forms an edge in $G(\mathcal{X},\mathcal{Y})$ if and only if $A\cap B=\emptyset$, i.e., $A_i\cap B_i=\emptyset$ for each $1\leq i\leq p$.
It can be easily verified that $\prod_{i=1}^{p}S_{n_i}$ acts transitively on $\mathcal{X}$ and $\mathcal{Y}$, respectively, and preserves the property of cross-intersecting. Thus we have $d(\mathcal{X})=|N(A)|$ for each $A\in \mathcal{X}$, and $d(\mathcal{Y})=|N(B)|$ for each $B\in \mathcal{Y}$. Since, for each $A=\prod_{i=1}^{p}{A_i}\in \mathcal{X}$,
\begin{equation*}
N(A)=\{B=\prod_{i=1}^{p}{B_i}\in \mathcal{Y}:~A_i\cap B_i=\emptyset\text{ for each }1\leq i\leq p\}=\prod_{i=1}^{p}{[n_i]\setminus A_i\choose s_i},
\end{equation*}
we have $d(\mathcal{X})=|N(A)|=\prod_{i=1}^{p}{{n_i-t_i}\choose s_i}$. Similarly, $d(\mathcal{Y})=|N(B)|=\prod_{i=1}^{p}{{n_i-s_i}\choose t_i}$.
By Theorem~\ref{key01}, we obtain that
\begin{equation*}
\alpha(\mathcal{X},\mathcal{Y})=\max{\{|\mathcal{Y}|-d(\mathcal{X})+1, |\mathcal{A}'|+|\mathcal{Y}|-|N(\mathcal{A}')|, |\mathcal{B}'|+|\mathcal{X}|-|N(\mathcal{B}')|\}},
\end{equation*}
where $\mathcal{A}'$ and $\mathcal{B}'$ are minimum imprimitive subsets of $\mathcal{X}$ and $\mathcal{Y}$ respectively. Therefore, in order to estimate $\alpha(\mathcal{X},\mathcal{Y})$ accurately, more discussions about the sizes and the structures of the imprimitive subsets of $\mathcal{X}$ and $\mathcal{Y}$ are necessary.
\begin{claim}\label{imprimitive subset}
Let $\mathcal{A}$ and $\mathcal{B}$ be imprimitive subsets of $\mathcal{X}$ and $\mathcal{Y}$ respectively, then
\begin{align*}
\mathcal{A}&=\prod_{i\in T_1}{\{A_i,\bar{A_i}\}}\times\prod_{i\in T_2}{\{A_i\}}\times\prod_{i\in [p]\setminus(T_1\sqcup T_2)}{[n_i]\choose t_i},~\text{for some disjoint $T_1, T_2\subseteq [p]$},\\
\mathcal{B}&=\prod_{i\in R_1}{\{B_i,\bar{B_i}\}}\times\prod_{i\in R_2}{\{B_i\}}\times\prod_{i\in [p]\setminus(R_1\sqcup R_2)}{[n_i]\choose s_i},~\text{for some disjoint $R_1, R_2\subseteq [p]$},
\end{align*}
where $A_i\in {[n_i]\choose t_i}$, $B_i\in {[n_i]\choose s_i}$, $T_1\sqcup T_2\neq\emptyset$, $R_1\sqcup R_2\neq\emptyset$ and $T_2,R_2\neq[p]$. Furthermore, for each $i\in T_1$, $n_i=2t_i$ and for each $i\in R_1$, $n_i=2s_i$.
\end{claim}
If $\Gamma=\prod_{i=1}^{p}S_{n_i}$ is imprimitive on $\mathcal{X}$, then from the definition we know that $\Gamma$ preserves a nontrivial partition $\{\mathcal{X}_j\}_{j=1}^{L}$ of $\mathcal{X}$. By projecting $\mathcal{X}_j$ to the $i$-th factor, we can obtain that $\bigsqcup_{j=1}^{L}(\mathcal{X}_j|_{i})=\mathcal{X}|_i={[n_i]\choose t_i}$ and $\Gamma|_{i}=S_{n_i}$ preserving this partition of $[n_i]\choose t_i$.
It is well known that for each $A_i\in {[n_i]\choose t_i}$, the stabilizer of $A_i$ is isomorphic to $S_{t_i}\times S_{n_i-t_i}$, which is a maximal subgroup of $S_{n_i}$ if $2t_i\neq n_i$ (see e.g. \cite{NB06}). Then by Theorem~\ref{primitivity}, we obtain that $S_{n_i}$ is primitive on $[n_i]\choose t_i$ unless $2t_i=n_i$, which means for the factors with $2t_i\neq n_i$ the partition $\bigsqcup_{j=1}^{L}(\mathcal{X}_j|_{i})$ of ${[n_i]\choose t_i}$ must be a trivial partition. Thus for each $j\in L$, $\mathcal{X}_j|_{i}$ is either a singleton in ${[n_i]\choose t_i}$, or $\mathcal{X}_j|_{i}={[n_i]\choose t_i}$.
When $2t_i=n_i$, it can be easily verified that the only imprimitive subset of $[n_i]\choose t_i$ has the form $\{A_i,\bar{A_i}\}$. Therefore, for the factors with $2t_i=n_i$, the partition $\bigsqcup_{j=1}^{L}(\mathcal{X}_j|_{i})$ of ${[n_i]\choose t_i}$ is either a trivial partition, or each partition block has the form $\mathcal{X}_j|_{i}=\{A_{i,j},\bar{A}_{i,j}\}$ for some $A_{i,j}\in {[n_i]\choose t_i}$.
Since each imprimitive subset of $\mathcal{X}$ can be seen as a block of a nontrivial partition of $\mathcal{X}$, we have $\mathcal{A}=\mathcal{X}_j$ for some $j\in [L]$. From the analysis above, we know that $\mathcal{A}|_i=\{A_i\}$ or $\{A_i,\bar{A_i}\}$ for some $A_i\in {[n_i]\choose t_i}$, or $\mathcal{A}|_i={[n_i]\choose t_i}$. Therefore, set $T_1\subseteq [p]$ such that for all $i\in T_1$, $2t_i=n_i$ and $\mathcal{A}|_i=\{A_i,\bar{A_i}\}$ for some $A_i\in {[n_i]\choose t_i}$; set $T_2\subseteq [p]$ such that for all $i\in T_1$, $\mathcal{A}|_i$ is a singleton, finally, we have
\begin{equation*}
\mathcal{A}=\prod_{i\in T_1}{\{A_i,\bar{A_i}\}}\times\prod_{i\in T_2}{\{A_i\}}\times\prod_{i\in [p]\setminus(T_1\sqcup T_2)}{[n_i]\choose t_i}.
\end{equation*}
The proof for the imprimitive subsets of $\mathcal{Y}$ is the same as that of $\mathcal{X}$. Thus, the claim holds.
By Claim~\ref{imprimitive subset}, we know that for the imprimitive subsets $\mathcal{A}$ and $\mathcal{B}$ above
\begin{equation*}
|\mathcal{A}|=2^{|T_1|}\cdot\prod_{i\in [p]\setminus(T_1\sqcup T_2)}{n_i \choose t_i}~\text{and}~|\mathcal{B}|=2^{|R_1|}\cdot\prod_{i\in [p]\setminus(R_1\sqcup R_2)}{n_i \choose s_i}.
\end{equation*}
And since
\begin{align*}
N(\mathcal{A})&=\{B\in \mathcal{Y}:~A\cap B=\emptyset\text{ for some }A\in\mathcal{A}\}\\
&=\prod_{i\in T_1}{({A_i\choose s_i}\sqcup{\bar{A}_i\choose s_i})}\times\prod_{i\in T_2}{{[n_i]\setminus A_i}\choose s_i}\times\prod_{i\in [p]\setminus(T_1\sqcup T_2)}{[n_i]\choose s_i},\\
N(\mathcal{B})&=\{A\in \mathcal{X}:~A\cap B=\emptyset\text{ for some }B\in\mathcal{B}\}\\
&=\prod_{i\in R_1}{({B_i\choose t_i}\sqcup{\bar{B}_i\choose t_i})}\times\prod_{i\in R_2}{{[n_i]\setminus B_i}\choose t_i}\times\prod_{i\in [p]\setminus(R_1\sqcup R_2)}{[n_i]\choose t_i},\\
\end{align*}
we have
\begin{align*}
|N(\mathcal{A})|=2^{|T_1|}\cdot\prod_{i\in T_1}{{\frac{n_i}{2}}\choose s_i}\cdot\prod_{i\in T_2}{{n_i-t_i}\choose s_i}\cdot\prod_{i\in [p]\setminus(T_1\sqcup T_2)}{n_i\choose s_i},\\
|N(\mathcal{B})|=2^{|R_1|}\cdot\prod_{i\in R_1}{{\frac{n_i}{2}}\choose t_i}\cdot\prod_{i\in R_2}{{n_i-s_i}\choose t_i}\cdot\prod_{i\in [p]\setminus(R_1\sqcup R_2)}{n_i\choose t_i}.
\end{align*}
Now we can estimate quantities $|\mathcal{A}'|+|\mathcal{Y}|-|N(\mathcal{A}')|$ and $|\mathcal{B}'|+|\mathcal{X}|-|N(\mathcal{B}')|$.
\begin{claim}\label{size estamitae}
With the assumptions in the theorem, for all imprimitive subsets $\mathcal{A}\subseteq\mathcal{X}$ and $\mathcal{B}\subseteq\mathcal{Y}$, $|\mathcal{Y}|-d(\mathcal{X})+1> |\mathcal{A}|+|\mathcal{Y}|-|N(\mathcal{A})|$, and $|\mathcal{Y}|-d(\mathcal{X})+1> |\mathcal{B}|+|\mathcal{X}|-|N(\mathcal{B})|$.
\end{claim}
We prove the claim by estimating the difference directly. Denote
\begin{align*}
&~D_1=|N(\mathcal{A})|-|\mathcal{A}|-d(\mathcal{X})+1\text{~and}\\
D_2&=|\mathcal{Y}|-|\mathcal{X}|+|N(\mathcal{B})|-|\mathcal{B}|-d(\mathcal{X})+1
\end{align*}
to be the differences between $|\mathcal{Y}|-d(\mathcal{X})+1$ and, respectively, $|\mathcal{A}|+|\mathcal{Y}|-|N(\mathcal{A})|$ and $|\mathcal{B}|+|\mathcal{X}|-|N(\mathcal{B})|$.
Set $d_1=\frac{D_1}{|N(\mathcal{A})|}$, $d_2=\frac{D_2}{|\mathcal{X}|}$.
Then, we have $d_1=1-\beta_1-\beta_2+\theta$, $d_2=\delta+\eta_0\cdot(1-\eta_1-\eta_2)+\theta'$, where $\theta=|N(\mathcal{A})|^{-1}$, $\delta=\frac{|\mathcal{Y}|-|\mathcal{X}|}{|\mathcal{X}|}$, $\eta_0=\frac{|N(\mathcal{B})|}{|\mathcal{X}|}$, $\theta'=|\mathcal{X}|^{-1}$, $\beta_1=\frac{|\mathcal{A}|}{|N(\mathcal{A})|}$, $\beta_2=\frac{d(\mathcal{X})}{|N(\mathcal{A})|}$, $\eta_1=\frac{|\mathcal{B}|}{|N(\mathcal{B})|}$, and $\eta_2=\frac{d(\mathcal{X})}{|N(\mathcal{B})|}$.
Since ${n_i\choose t_i}\cdot{{n_i-t_i}\choose s_i}={n_i\choose s_i}\cdot{{n_i-s_i}\choose t_i}$ for each $i\in [p]$, we have $1/{{n_i-t_i}\choose s_i}={n_i\choose t_i}/({n_i\choose s_i}\cdot{{n_i-s_i}\choose t_i})$ for each $i\in[p]$. This yields that
\begin{align*}
&~~~~\beta_1=\prod_{i\in [p]}\frac{{n_i \choose t_i}}{{n_i\choose s_i}}\cdot\prod_{i\in T_1\sqcup T_2}\frac{1}{{{n_i-s_i}\choose t_i}},~~\beta_2=\frac{1}{2^{|T_1|}}\cdot\prod_{i\in [p]\setminus(T_1\sqcup T_2)}\prod_{j=0}^{s_i-1}(1-\frac{t_i}{n_i-j}),\\
\eta_1&=\prod_{i\in [p]}\frac{{n_i \choose s_i}}{{n_i\choose t_i}}\cdot\prod_{i\in R_1\sqcup R_2}\frac{1}{{{n_i-t_i}\choose s_i}},~~\eta_2=\prod_{i\in [p]}\frac{{n_i \choose s_i}}{{n_i\choose t_i}}\cdot\frac{1}{2^{|R_1|}}\cdot\prod_{i\in [p]\setminus(R_1\sqcup R_2)}\prod_{j=0}^{t_i-1}(1-\frac{s_i}{n_i-j}).
\end{align*}
By the assumptions, we know that $n_i\geq s_i+t_i+1\geq5$, $\prod_{i\in [p]}\frac{{n_i \choose t_i}}{{n_i\choose s_i}}\leq 1$ and ${{n_i-s_i}\choose t_i}\geq {\lceil\frac{n_i}{2}\rceil\choose t_i}\geq \frac{n_i}{2}$. Since $T_1\sqcup T_2\neq\emptyset$, $R_1\sqcup R_2\neq\emptyset$ and $T_2,R_2\neq[p]$, we can obtain
\begin{align*}
\beta_1&\leq\prod_{i\in T_1\sqcup T_2}\frac{1}{{{n_i-s_i}\choose t_i}}\leq\max\limits_{i\in (T_1\sqcup T_2)}{\{(\frac{2}{n_i+2})^{|T_1|}\cdot(\frac{2}{n_i})^{|T_2|}\}},\\
&~\beta_2\leq\frac{1}{2^{|T_1|}}\cdot\max\limits_{i\in[p]\setminus (T_1\sqcup T_2)}\{(1-\frac{4n_i-6}{n_i(n_i-1)})^{p-(|T_1|+|T_2|)}\},
\end{align*}
and
\begin{align*}
\eta_1&\leq(1+\delta)\cdot\prod_{i\in R_1\sqcup R_2}\frac{1}{{{n_i-t_i}\choose s_i}}\leq(1+\delta)\cdot\max\limits_{i\in (R_1\sqcup R_2)}{\{(\frac{2}{n_i+2})^{|R_1|}\cdot(\frac{2}{n_i})^{|R_2|}\}},\\
&~~~~~~~~~\eta_2\leq(1+\delta)\cdot\frac{1}{2^{|R_1|}}\cdot\max\limits_{i\in[p]\setminus (R_1\sqcup R_2)}\{(1-\frac{4n_i-6}{n_i(n_i-1)})^{p-(|R_1|+|R_2|)}\}.
\end{align*}
This leads to
\begin{flalign*}
\beta_1+\beta_2\leq\left\{\begin{array}{ll}1-\min\limits_{i\neq j\in[p]}\{\frac{6}{n_i}-\frac{2}{n_i-1}-\frac{2}{n_j}\}, &~\text{if}~T_2\neq\emptyset;\\
\frac{1}{2}-\min\limits_{i\neq j\in[p]}\{\frac{3}{n_i}-\frac{1}{n_i-1}-\frac{2}{n_j+2}\}, &~\text{otherwise};\end{array}\right.
\end{flalign*}
and
\begin{flalign*}
\frac{\eta_1+\eta_2}{1+\delta}\leq\left\{\begin{array}{ll}1-\min\limits_{i\neq j\in[p]}\{\frac{6}{n_i}-\frac{2}{n_i-1}-\frac{2}{n_j}\}, &~\text{if}~R_2\neq\emptyset;\\
\frac{1}{2}-\min\limits_{i\neq j\in[p]}\{\frac{3}{n_i}-\frac{1}{n_i-1}-\frac{2}{n_j+2}\}, &~\text{otherwise}.\end{array}\right.
\end{flalign*}
Since $5\leq n_i\leq\frac{7}{4} n_j$ for all distinct $i,j\in[p]$, thus we have $\beta_1+\beta_2,\frac{\eta_1+\eta_2}{1+\delta}\leq 1$. Therefore,
\begin{align*}
&~~~~~~~~~~~~~~~~~~~~~~~~~d_1=1-\beta_1-\beta_2+\theta> 1-\beta_1-\beta_2\geq0,\\
d_2&=\delta+\eta_0\cdot(1-\eta_1-\eta_2)+\theta'=\delta\cdot(1-\eta_0\cdot\frac{\eta_1+\eta_2}{1+\delta})+\eta_0\cdot(1-\frac{\eta_1+\eta_2}{1+\delta})+\theta'>0.
\end{align*}
Thus, the claim holds.
For each pair of non-empty cross-intersecting families $(\mathcal{A},\mathcal{B})\in 2^{\mathcal{X}}\times2^{\mathcal{Y}}$, $\mathcal{A}\cup \mathcal{B}$ forms a nontrivial independent set of $G(\mathcal{X},\mathcal{Y})$. Therefore, by Claim~\ref{size estamitae}, the inequality~(\ref{eq02}) holds.
To complete the proof, we need to characterize all the nontrivial fragments in $\mathcal{F}(\mathcal{X})$. As a direct consequence of Claim~\ref{size estamitae}, every fragment of $G(\mathcal{X},\mathcal{Y})$ is primitive. Hence, by Theorem~\ref{key00}, when $\prod_{i\in [p]}{n_i \choose t_i}<\prod_{i\in [p]}{n_i\choose s_i}$, $\mathcal{X}$ has only $1$-fragments.
When $\prod_{i\in [p]}{n_i \choose t_i}=\prod_{i\in [p]}{n_i\choose s_i}$, suppose there are nontrivial fragments in $\mathcal{F}(\mathcal{X})$. W.l.o.g., assume that $\mathcal{S}$ is a minimal-sized nontrivial fragment in $\mathcal{X}$. By Theorem~\ref{key00}, $\mathcal{S}$ is semi-imprimitive. Since for any two different elements $A,B\in\mathcal{X}$, $|N(A)\cap N(B)|<\prod_{i\in [p]}{{n_i-t_i}\choose s_i}-1$. Therefore, there are no $2$-fragments in $\mathcal{F(\mathcal{X})}$. By Proposition~\ref{fragment2}, $\mathcal{S}$ is balanced.
Now we are going to prove the non-existence of such $\mathcal{S}$ by analyzing its size and structure, which will yield that $\mathcal{X}$ also has only $1$-fragments when $\prod_{i\in [p]}{n_i \choose t_i}=\prod_{i\in [p]}{n_i\choose s_i}$.
For each $A=\prod_{i\in [p]}A_i\in \mathcal{S}$, let $\Gamma_A=\prod_{i\in [p]}(S_{A_i}\times S_{\bar{A}_i})$, $\Gamma_{\mathcal{S}}=\{\sigma\in \Gamma:~\sigma(\mathcal{S})=\mathcal{S}\}$ and $\Gamma_{A,\mathcal{S}}=\{\sigma\in \Gamma_A:~\sigma(\mathcal{S})=\mathcal{S}\}$. We claim that there exists a subset $C\in \mathcal{S}$ such that $\Gamma_{C}\neq \Gamma_{C,\mathcal{S}}$. Otherwise, for any two different subsets $B,B'\in \mathcal{S}$, we have $\Gamma_{B}=\Gamma_{B,\mathcal{S}}$ and $\Gamma_{B'}= \Gamma_{B',\mathcal{S}}$. Since $\Gamma_{B,\mathcal{S}}$ and $\Gamma_{B',\mathcal{S}}$ are both subgroups of $\Gamma_{\mathcal{S}}$, we have $\langle\Gamma_{B},\Gamma_{B'}\rangle$ is a subgroup of $\Gamma_{\mathcal{S}}$. Let $T\subseteq [p]$ be the factors where $B'_i=B_i~(\text{or}~\bar{B}_i~\text{if}~2t_i=n_i)$, write
$$\Gamma_B=\prod_{i\in T}(S_{B_i}\times S_{\bar{B}_i})\times\prod_{i\in [p]\setminus T}(S_{B_i}\times S_{\bar{B}_i}),$$
then we have,
$$\Gamma_{B'}=\prod_{i\in T}(S_{B_i}\times S_{\bar{B}_i})\times\prod_{i\in [p]\setminus T}(S_{B'_i}\times S_{\bar{B}'_i}).$$
Since $\langle S_{B_i}\times S_{\bar{B}_i}, S_{B'_i}\times S_{\bar{B}_i'}\rangle=S_{n_i}$ for each $B'_i\ne B_i~(\text{and}~B'_i\ne\bar{B}_i~\text{if}~2t_i=n_i)$, we have $$\langle\Gamma_{B},\Gamma_{B'}\rangle=\prod_{i\in T}(S_{B_i}\times S_{\bar{B}_i})\times\prod_{i\in [p]\setminus T}S_{n_i}.$$
Therefore, for some fixed $B\in\mathcal{S}$, $\Gamma_{\mathcal{S}}$ contains $\prod_{i\in T'}(S_{B_i}\times S_{\bar{B}_i})\times\prod_{i\in [p]\setminus T'}S_{n_i}$ as a subgroup, where
$$T'=\{i|i\in[p],\text{ such that }A_i=B_i~(\text{or }\bar{B}_i\text{ if }2t_i=n_i)\text{ for all }A\in\mathcal{S}\}.$$
When $T'=\emptyset$, we have $\Gamma_{\mathcal{S}}=\prod_{i\in [p]}S_{n_i}$, thus $\mathcal{S}=\mathcal{X}$, yielding a contradiction. When $T'\neq \emptyset$, if $|T'|=1$, w.l.o.g., taking $T'=\{1\}$, we have $(S_{B_1}\times S_{\bar{B}_1})\times\prod_{i\in [p]\setminus \{1\}}S_{n_i}\subseteq\Gamma_{\mathcal{S}}$. Therefore, since $\mathcal{S}\neq\mathcal{X}$, from the definition of $T'$ we have
$$\mathcal{S}=\{B_1\}\times\prod_{i\in [p]\setminus \{1\}}{[n_i]\choose t_i}, \text{~or~} S=\{B_1,\bar{B}_1\}\times\prod_{i\in [p]\setminus \{1\}}{[n_i]\choose t_i}\text{~when $2t_1=n_1$}.$$
In both cases, $|\mathcal{S}|<\frac{\alpha(\mathcal{X},\mathcal{Y})}{2}$. If $|T'|\geq 2$, we have
$$\mathcal{S}\subseteq\{B_{i_0}\}\times\prod_{i\in [p]\setminus \{i_0\}}{[n_i]\choose t_i}, \text{~or~} S\subseteq\{B_{i_0},\bar{B}_{i_0}\}\times\prod_{i\in [p]\setminus \{i_0\}}{[n_i]\choose t_i}\text{~when $2t_{i_0}=n_{i_0}$},$$
for some $i_0\in T'$. Therefore, when $T'\neq \emptyset$, we always have $|\mathcal{S}|<\frac{\alpha(\mathcal{X},\mathcal{Y})}{2}$, which contradicts the fact that $\mathcal{S}$ is balanced. Hence, the existence of $C$ is guaranteed.
By Proposition~\ref{fragment2} we have that $[\Gamma_{C}:\Gamma_{C,\mathcal{S}}]$, the index of $\Gamma_{C,\mathcal{S}}$ in $\Gamma_{C}$, equals 2. Now let $\Gamma_{C,\mathcal{S}}[C_i]$ be the projection of $\Gamma_{C,\mathcal{S}}$ onto $S_{C_i}$, $\Gamma_{C,\mathcal{S}}[C_i]$ must be a subgroup of $S_{C_i}$ of index no greater than 2. Thus $\Gamma_{C,\mathcal{S}}[C_i]=S_{C_i}$ or $A_{C_i}$. Since $\Gamma_C=\prod_{i\in [p]}(S_{C_i}\times S_{\bar{C}_i})$, we know that $\Gamma_{C,\mathcal{S}}=\prod_{i\in [p]\setminus\{j\}}(S_{C_i}\times S_{\bar{C}_i})\times(A_{C_{j}}\times S_{\bar{C}_{j}})$ or $\prod_{i\in [p]\setminus\{j\}}(S_{C_i}\times S_{\bar{C}_i})\times(S_{C_{j}}\times A_{\bar{C}_{j}}$), for some $j\in [p]$.
Since for all $i\in [p]$, $t_i=|B_i\cap C_i|+|B_i\cap \bar{C}_i|$ for each pair $B,C\in \mathcal{S}$. If $|B_i\cap C_i|>1$, let $s,t\in B_i\cap C_i$, then the transposition $(s~t)$ fixes both $C_i$ and $B_i$. Taking $i=j$, the semi-imprimitivity of $\mathcal{S}$ implies that $(s~t)\in \Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}$. This yields $\Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}=S_{C_{j}}\times A_{\bar{C}_{j}}$. From this process it follows that, for each $B\in \mathcal{S}$, there exists at most one of $|B_j\cap C_j|$ and $|B_j\cap \bar{C}_j|$ to be greater than $1$. Note that if $B_j\in\bar{C}_j$, then $S_{C_j}$ and $S_{B_j}$ fix both $C_j$ and $B_j$, i.e., $S_{C_j}\times S_{B_j}\subseteq \Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}$. Since $\Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}=A_{C_{j}}\times S_{\bar{C}_{j}}$ or $S_{C_{j}}\times A_{\bar{C}_{j}}$, and neither $A_{C_{j}}\times S_{\bar{C}_{j}}$ nor $S_{C_{j}}\times A_{\bar{C}_{j}}$ contains $S_{C_j}\times S_{B_j}$. Therefore, we obtain that $|B_j\cap C_j|=1$ for each $B\in \mathcal{S}$, or $|B_j\cap C_j|=t_j-1$ for each $B\in \mathcal{S}$.
We claim that for both cases, $\mathcal{S}$ can not be balanced.
Suppose $|B_j\cap C_j|=1$ for each $B\in \mathcal{S}$. W.l.o.g., assume $B_j\cap C_j=\{1\}$ for some $B\in \mathcal{S}$. From the semi-imprimitivity of $\mathcal{S}$, we know that for all $\gamma\in \Gamma,~\gamma(\mathcal{S})\cap\mathcal{S}=\emptyset, ~\mathcal{S}$ or $\{A\}$ for some $A\in \mathcal{S}$. Thus $(\gamma(\mathcal{S})\cap\mathcal{S})|_j=\emptyset,~\mathcal{S}|_j$ or $\{A_j\}$ for some $A_j\in {[n_j]\choose t_j}$. If $t_j>2$, then $|B_j\cap \bar{C}_j|\geq 2$, so $\Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}=A_{C_{j}}\times S_{\bar{C}_{j}}$. On the other hand, we can find distinct $s,t\in C_j$ such that $(1~s~t)(B_j)=B_j\setminus\{1\}\cup\{s\}\in \mathcal{S}|_j$ since $(1~s~t)\in A_{C_j}$. Then $(1~s)(\mathcal{S}|_j)$ has more than one element of $\mathcal{S}|_j$, therefore $(1~s)\in \Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}$. This contradiction proves that $t_j=2$. Thus $\mathcal{S}|_j=\mathcal{C}=\{A_j\in{[n_j]\choose 2}:~1\in A_j\}$. Otherwise, w.l.o.g., assume $C_j=\{1,2\}$ and there exists $B\in \mathcal{S}$ such that $B_j\cap C_j=\{2\}$. Since $\Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}=A_{C_{j}}\times S_{\bar{C}_{j}}$ or $S_{C_{j}}\times A_{\bar{C}_{j}}$, we have $\mathcal{C}\subseteq \mathcal{S}|_j$ and $\mathcal{C'}=\{A_j\in{[n_j]\choose 2}:~2\in A_j\}\subseteq\mathcal{S}|_j$. Thus $\mathcal{S}|_j=\mathcal{C}\cup\mathcal{C'}$. This yields $\Gamma_{C,\mathcal{S}}|_{S_{C_{j}}\times S_{\bar{C}_{j}}}=S_{C_{j}}\times S_{\bar{C}_{j}}$, leading to a contradiction.
Suppose now $|B_j\cap C_j|=t_j-1>1$ for each $B\in \mathcal{S}$. Similarly, we can prove that $n_j-t_j=2$, which contradicts $n_j\geq s_j+t_j+1$ and $2\leq s_j$, $t_j\leq \frac{n}{2}$. Therefore, for each $B\in \mathcal{S}$, $|B_j\cap C_j|=1$
From the analysis above, we know that for each $B\in \mathcal{S}$, $B_j=\{1,b\}$ for some $b\in[n_j]$. Thus, for each $B\in \mathcal{S}$, we have $\Gamma_{B,\mathcal{S}}|_{S_{B_{j}}\times S_{\bar{B}_{j}}}=A_{B_{j}}\times S_{\bar{B}_{j}}$, and $\Gamma_{B,\mathcal{S}}=\prod_{i\in[p]\setminus \{j\}}(S_{B_i}\times S_{\bar{B}_i})\times(A_{B_{j}}\times S_{\bar{B}_{j}})$ since $[\Gamma_{B}:\Gamma_{B,\mathcal{S}}]=2$. Therefore $\Gamma_{\mathcal{S}}$ contains
$$\langle\Gamma_{B,\mathcal{S}},\text{~for~all~}B\in \mathcal{S}\rangle=\prod_{i\in T''}(S_{C_i}\times S_{\bar{C}_i})\times\prod_{i\in [p]\setminus (T''\cup\{j\})}S_{n_i}\times S_{[n_j]\setminus\{1\}}$$
as a subgroup, where $T''=\{i|i\in[p],\text{ such that }B_i=C_i~(\text{or }\bar{C}_i\text{ if }2t_i=n_i)\text{ for all }B\in\mathcal{S}\}$. Similarly, by arguing the structure of $\mathcal{S}$, if $T''\neq\emptyset$, we can prove that $|\mathcal{S}|<\frac{\alpha(\mathcal{X},\mathcal{Y})}{2}$. Thus we have $T''=\emptyset$ and $\mathcal{S}=\prod_{i\in[p]\setminus\{j\}}{[n_i]\choose t_i}\times \mathcal{C}$.
Since $\mathcal{S}$ is balanced, $\prod_{i\in [p]}{n_i \choose t_i}=\prod_{i\in [p]}{n_i\choose s_i}$ and $|\mathcal{S}|=\prod_{i\in[p]\setminus\{j\}}{n_i\choose t_i}\cdot (n_j-1)$, we have
\begin{equation}\label{eq05}
2\prod_{i\in[p]\setminus\{j\}}{n_i\choose t_i}\cdot (n_j-1)=\prod_{i\in[p]\setminus\{j\}}{n_i\choose t_i}\cdot{n_j\choose 2}-\prod_{i\in[p]\setminus\{j\}}{{n_i-s_i}\choose t_i}\cdot{{n_j-s_j}\choose 2}+1,
\end{equation}
which means $n_j$ must be an integral zero of the following function
\begin{align*}
H(x)=(1-a_0)\cdot x^2-(5-a_0\cdot(2s_j+1))\cdot x+(2b_0+4-a_0\cdot (s_j^2+s_j)),
\end{align*}
where $a_0=\prod_{i\in [p]\setminus \{j\}}\frac{{{n_i-s_i}\choose t_i}}{{n_i\choose t_i}}$ and $b_0=\prod_{i\in [p]\setminus \{j\}}{{n_i\choose t_i}^{-1}}$. Since $n_j\geq3+s_j$ and $2\leq s_j\leq \frac{n_j}{2}$, by Vieta's formulas for quadratic polynomials, there is no such $n_j$ satisfying $H(n_j)=0$ when $s_j\geq 3$. Hence $\mathcal{S}=\prod_{i\in[p]\setminus\{j\}}{[n_i]\choose t_i}\times \mathcal{C}$ is a nontrivial balanced fragment of $\mathcal{X}$ if and only if $t_j=s_j=2$ and equation~(\ref{eq05}) holds. Using the fact that $\frac{{{n_i-s_i}\choose t_i}}{{n_i\choose t_i}}\leq (1-\frac{s_i}{n_i})(1-\frac{s_i}{n_i-1})$ and the assumption $n_i\leq \frac{7}{4}n_j$ for distinct $i,j\in[p]$, it can be easily verified that the LHS of equation~(\ref{eq05}) is strictly less than the RHS when $s_j=2$. Therefore, $\mathcal{S}$ can not be balanced.
This completes the proof.
\end{proof}
\section{Concluding remarks}
In this paper we have investigated two multi-part generalizations of the cross-intersecting theorems. Our main contribution is determining the maximal size and the corresponding structures of the families for both trivially and nontrivially (with the non-empty restriction) cross-intersecting cases.
The method we used for the proof was originally introduced by Wang and Zhang in~\cite{WZ11}, which was further generalized to the bipartite case in~\cite{WZ13}. This method can deal with set systems, finite vector spaces and permutations uniformly. It is natural to ask whether we can extend the single-part cross-intersecting theorems for finite vector spaces and permutations to the multi-part case. It is possible for permutations when considering the case without the non-empty restriction, and we believe it is also possible for finite vector spaces. But when it comes to the case where the families are non-empty, as far as we know, there is still no result for finite vector spaces and permutations.
For single-part families $\mathcal{A}$ and $\mathcal{B}$, it is natural to define cross-t-intersecting as $|A\cap B|\geq t$ for each pair of $A\in \mathcal{A}$ and $B\in \mathcal{B}$. But for multi-part families, when defining cross-t-intersecting between two families, the simple extension of the definition for single-part case can be confusing. Therefore, a reasonable definition and related problems for multi-part cross-$t$-intersecting families are also worth considering.
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{sec:introduction}
\textbf{Motivation.} We study the transmission and reflection
properties of meta-materials, i.e., of periodic microstructures of a
composite material with two components. The interest in
meta-materials has immensely grown in the last years as they exhibit
astonishing properties such as band gaps or negative refraction; see
\cite{EP04negphC, PE03lefthanded, CJJP02negrefraction}. The
propagation of electromagnetic waves in such materials is modelled by
time-harmonic Maxwell's equations for the electric field $E$ and the
magnetic field $H$:
\begin{subequations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\curl E &= \phantom{-} i \omega \mu_0\mu H\,, \\
\curl H &= - i \omega \varepsilon_0\varepsilon E\,.
\end{empheq}
\end{subequations}
We use the standard formulation with $\mu_0, \varepsilon_0 > 0$ the
permeability and permittivity of vacuum, $\mu$ and $\varepsilon$ the
corresponding relative parameters, and $\omega>0$ the imposed frequency.
While most materials are non-magnetic, i.e., $\mu=1$, the electric
permittivity $\varepsilon$ covers a wide range. In this paper, we
study meta-materials consisting of air (i.e., $\varepsilon=1$) and a
(metal) microstructure $\Sigma_\eta$. The microstructure is
assumed to be an $\eta$-periodic repetition of scaled copies of some
geometry $\Sigma$. In the present study, we investigate in detail four
different geometries: $\Sigma$ can be a metal cylinder (in two
rotations), a metal plate, or the complement of an air cylinder;
see Fig. \ref{fig:Two-examples-of-perfect-conductors-for-analysis}
and \eqref {eq:Analysis-Definition-of-metal-cylinder}--\eqref
{eq:Analysis-Definition-of-air-cylinder} for a detailed definition.
For the electric permittivity in the microstructure $\Sigma_\eta$, we
consider two different cases: perfect conductors that are formally
obtained by setting $\varepsilon=\infty$, and high-contrast materials
with $\varepsilon=\varepsilon_1\eta^{-2}$, where $\varepsilon_1\in \mathbb{C}$ is
some complex number with $\Im(\varepsilon_1) > 0$. In both cases, our study
is based on the effective equations for the electric and magnetic
field in the limit $\eta \to 0$.
The numerical simulation of electromagnetic wave propagation in such
meta-materials is very challenging because of the rapid variations in
the electric permittivity. Standard methods require the resolution of
the $\eta$-scale, which often becomes infeasible even with today's
computational resources. Instead, we resort to
homogenization and multiscale methods to extract macroscopic features
and the behaviour of the solution. The effective equations obtained
by homogenization can serve as a good motivation and
starting point in this process.
\smallskip \textbf{Literature.} Effective equations for Maxwell's
equations in meta-materials are obtained in several different settings
with various backgrounds in mind: Dielectric bulk inclusions with
high-contrast media \cite{BBF09hommaxwell, BBF15hommaxwell,
CC15hommaxwell} can explain the effect of artificial magnetism and
lead to unusual effective permeabilities $\mu$, while long wires
\cite{BB12homwire} lead to unusual effective permittivities. A
combination of both structures is used to obtain a negative-index
meta-material in \cite{LS15negindex}. Topological changes in the
material in the limit $\eta\to 0$, such as found in split rings
\cite{BS10splitring}, also incite unusual effective behaviour.
Perfect conductors were recently studied as well: split rings in
\cite{LS16pecrings} and different geometries in \cite{SU17hommaxwell}.
Finally, we briefly mention that the Helmholtz equation---as the
two-dimensional reduction of Maxwell's equations---is often studied as
the first example for unusual effective properties: high-contrast
inclusions in \cite{BF04homhelmholtz} or high-contrast layer materials
in \cite{BS13plasmonwaves}, just to name a few. An overview on this
vast topic is provided in \cite{Schw17metamaterial}.
Concerning the numerical treatment, we focus on the Heterogeneous
Multiscale Method (HMM) \cite{EE03hmm, EE05hmm}. For the HMM, first
analytical results concerning the approximation properties for
elliptic problems have been derived in \cite{Abd05hmmanalysis,
EMZ05hmmanalysis, Ohl05HMM} and then extended to other problems,
such as time-harmonic Maxwell's equations \cite{HOV15maxwellHMM} and
the Helmholtz equation and Maxwell's equations with high-contrast
\cite{OV16hmmhelmholtz, Ver17hmmmaxwell}. Another related work is the
multiscale asymptotic expansion for Maxwell's equations
\cite{CZAL10multiscalemaxwell}.
For further recent contributions to HMM approximations for
Maxwell's equations we refer to \cite{CFS17hmmmaxwell,HS16hmmmaxwelltime}.
Sparse tensor product finite elements for multiscale Maxwell-type equations
are analyzed in \cite{CH17femmaxwellmultiscale} and an adaptive generalized
multiscale finite element method is studied in \cite{CL19gmsfem}.
\smallskip \textbf{Main results.} We perform an analytical and a
numerical study of transmission properties of meta-materials that
contain either perfect conductors or high-contrast materials. The
main results are the following:
1.) Using the effective equations of \cite{SU17hommaxwell}, we
calculate the reflection and transmission coefficients for four
microscopic geometries $\Sigma$. Few geometrical parameters are
sufficient to fully describe the effective coefficients. We show that
only certain polarizations can lead to transmission.
2.) For the two geometries that are invariant in the $\operatorname{e}_3$-direction,
we study the limit behaviour of the electromagnetic fields for
high-contrast media. When the electric field is parallel to $\operatorname{e}_3$,
all fields vanish in the limit. Instead, when the magnetic field is
parallel to $\operatorname{e}_3$, transmission cannot be excluded due to resonances.
3.) Extensive numerical experiments for high-contrast media confirm
the analytical results. The numerical experiments underline the
applicability of the Heterogeneous Multiscale Method to these
challenging setting.
Some further remarks on 2.) are in order. The results are related to
homogenization results of \cite{BBF15hommaxwell, CC15hommaxwell}, but
we study more general geometries, since the highly conducting material
can be connected. Furthermore, the results are related to
\cite{BF04homhelmholtz, BS13plasmonwaves}, where connected structures
are investigated, but in a two-dimensional formulation. We treat here
properties of the three-dimensional solutions. We emphasize that the
transmission properties of a high-contrast medium cannot be captured
in the framework of perfect conductors, since the latter excludes
resonances on the scale of the periodicity (except if three different
length-scales are considered as in \cite {LS16pecrings}).
\smallskip \textbf{Organization of the paper.} The paper is organized
as follows: In Section \ref{sec:problem}, we detail the underlying
problem formulations and revisit existing effective equations. In
Section \ref{sec:analysis}, we compute the transmission coefficients
for perfect conductors and derive effective equations for
high-contrast media. In Section \ref{sec:numerics}, we briefly
introduce the Heterogeneous Multiscale Method. Finally, in Section
\ref{subsec:experiment} we present several numerical experiments
concerning the transmission properties of our geometries for
high-contrast materials.
\section{Problem formulation and effective equations}
\label{sec:problem}
This section contains the precise formulation of the problem,
including the description of the four microscopic geometries. We
summarize the relevant known homogenization results and apply them to
the cases of interest.
\subsection{Geometry and material parameters}
\label{sec:geometry}
We study the time-harmonic Maxwell equations with linear material
laws. The geometry is periodic with period $\eta>0$; solutions depend
on this parameter and are therefore indexed with $\eta$. On a domain
$G \subset \mathbb{R}^3$, the problem is to find $E^\eta, H^\eta :G \to \mathbb{C}^3$,
such that
\begin{subequations}\label{eq:time-harmonic-Maxwell-eq}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\curl E^\eta& = \phantom{-} i \omega \mu_0 H^\eta \,,
\label{eq:time-harmonic-Maxwell-eq1}\\
\curl H^\eta&= - i \omega \varepsilon_0\varepsilon_\eta E^\eta \,,
\label{eq:time-harmonic-Maxwell-eq2}
\end{empheq}
\end{subequations}
subject to appropriate boundary conditions. In the following, we will
give details on the geometry $G$ and on the choice of the material
parameter $\varepsilon_\eta$, the relative permittivity. Note that
the system allows to eliminate one unknown. Indeed, if we insert
$H^\eta$ from \eqref{eq:time-harmonic-Maxwell-eq1} into
\eqref{eq:time-harmonic-Maxwell-eq2}, we obtain
\begin{equation}
\curl\curl E^\eta = \omega^2 \mu_0\varepsilon_0\varepsilon_\eta E^\eta\,.
\label{eq:time-harmonic-Maxwell-eq-E}
\end{equation}
Alternatively, substituting $E^\eta$ from \eqref
{eq:time-harmonic-Maxwell-eq2} into \eqref
{eq:time-harmonic-Maxwell-eq1}, we obtain
\begin{equation}
\curl \varepsilon_\eta^{-1}\curl H^\eta = \omega^2 \mu_0\varepsilon_0
H^\eta\,.\label{eq:time-harmonic-Maxwell-eq-H}
\end{equation}
\begin{figure}
\centering
\includegraphics[width=.7\linewidth]{geometry}
\caption{Waveguide domain $G$ with periodic scatterer $\Sigma_{\delta}$
contained in the middle part $Q_M$ and incident wave from the right.
\label{fig:waveguide}}
\end{figure}
\textbf{Geometry.} As sketched in Fig. \ref{fig:waveguide}, with positive numbers $\ell_2, \ell_3 > 0$, the
unbounded macroscopic domain is the waveguide domain
\begin{equation}\label{eq:Analysis-Definition-of-G}
G \coloneqq \set[\big]{x =(x_1, x_2, x_3) \in \mathbb{R}^3 \given x_2 \in (-
\ell_2, \ell_2) \text{ and } x_3 \in (- \ell_3, \ell_3)}\,.
\end{equation}
With another positive number $L > 0$, the domain is divided into three
parts (left, middle, right) as
\begin{equation*}
Q_L \coloneqq \set[\big]{x \in G \given x_1 \leq -L}\, ,\ Q_M
\coloneqq \set[\big]{x \in G \given x_1 \in (-L, 0)}\, , \ \text{
and } \ Q_R \coloneqq \set{x \in G \given x_1 \geq 0}\, .
\end{equation*}
The scatterer $\Sigma_{\delta}$ is contained in the middle part
$Q_M$. For the boundary conditions, we consider an incident wave
from the right that travels along the $x_1$-axis to the left. We
restrict ourselves here to normal incidence. For the analysis, we
impose periodic boundary conditions on the lateral boundaries of the
domain $G$. For the numerics, we will modify the boundary conditions
slightly: we truncate $G$ in $x_1$-direction (to obtain a bounded
domain) and consider impedance boundary conditions (with the incident
wave as data) on the whole boundary of $G$.
\begin{figure}
\centering
\begin{subfigure}[t]{0.3\textwidth}
\begin{tikzpicture}[scale=2.5]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\fill[gray!20!white, opacity=.5] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!20!white, opacity=.5] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!20!white, opacity=.5] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!20!white, opacity=.5] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!75!white] (-.25, -.5) -- (-.25,.5)
arc(180:0:0.25cm and 0.125cm) -- (0.25, -0.5)
(-0.25, -.5) arc(180:360:0.25cm and 0.125cm);
\filldraw[gray!75!white] (-.25,-.5) -- (0.25,-.5);
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (-.25,0.5) arc (180:-180:0.25cm and 0.125cm);
\draw[dashed] (-.25,-.5) arc (180:360:0.25cm and 0.125cm);
\draw[dashed] (-.25,-.5) arc (180:0:0.25cm and 0.125cm);
\draw[dashed] (-0.25, -.5) -- (-0.25, .5);
\draw[dashed] (.25,-.5) -- (.25,.5);
\draw[->] ( -.8, -.7, .5)--( -.4, -.7, .5 );
\draw[->] (-.8, -.7, .5) -- (-.8, -.3, .5);
\draw[->] (-.8, -.7, .5) -- (-.8, -.7, .2);
\node[] at (0,0,0) {$\Sigma_1$};
\node[] at (-.4, -.8, .5){$\operatorname{e}_1$};
\node[] at (-.68, -.3, .5){$\operatorname{e}_3$};
\node[] at (-.8, -.8, .-.1){$\operatorname{e}_2$};
\end{tikzpicture}
\caption{ }
\label{fig:the metal cylinder}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\begin{tikzpicture}[scale=2.5]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (M1) at (-0.5, -0.5, 0.15);
\coordinate (M2) at (-0.5, -0.5, -0.15);
\coordinate (M3) at (-0.5, 0.5, -0.15);
\coordinate (M4) at (-0.5, 0.5, 0.15);
\coordinate (M5) at (0.5, 0.5, 0.15);
\coordinate (M6) at (0.5,0.5,-0.15);
\coordinate (M7) at (0.5, -0.5, -0.15);
\coordinate (M8) at (0.5, -0.5, 0.15);
\fill[gray!20!white, opacity=.5] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!20!white, opacity=.5] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!20!white, opacity=.5] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!20!white, opacity=.5] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!75!white, opacity=.9] (M4) -- (M5) -- (M6) -- (M3) -- cycle;
\fill[gray!75!white, opacity=.9] (M1) -- (M4) -- (M5) -- (M8) -- cycle;
\fill[gray!75!white, opacity=.9] (M2) -- (M3) -- (M6) -- (M7) -- cycle;
\fill[gray!75!white, opacity=.9] (M1) -- (M2) -- (M7) -- (M8) -- cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[dashed] (M1) -- (M4) (M2) -- (M3) ;
\draw[] (M4) -- (M5) -- (M6) -- (M3) -- cycle;
\draw[dashed] (M5) -- (M8) (M7) -- (M6);
\draw[dashed] (M1) -- (M8) (M7) -- (M2);
\draw[->] ( -.8, -.7, .5)--( -.4, -.7, .5 );
\draw[->] (-.8, -.7, .5) -- (-.8, -.3, .5);
\draw[->] (-.8, -.7, .5) -- (-.8, -.7, .2);
\node[] at (0, 0, 0) {$\Sigma_3$};
\node[] at (-.4, -.8, .5){$\operatorname{e}_1$};
\node[] at (-.68, -.3, .5){$\operatorname{e}_3$};
\node[] at (-.8, -.8, .-.1){$\operatorname{e}_2$};
\end{tikzpicture}
\caption{ }
\label{fig:the metal plate}
\end{subfigure}
\hfill\begin{subfigure}[t]{0.3\textwidth}
\begin{tikzpicture}[scale=2.5]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\fill[gray!75!white, opacity=.9] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!75!white, opacity=.9] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!75!white, opacity=.9] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!75!white, opacity=.9] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!20!white] (-0.5, 0.25, 0) -- (.5,.25,0)
arc(90:-90:0.125cm and 0.25cm) --
(0.5, -.25,0) --
(-0.5, -0.25,0) arc(-90:-270:0.125cm and 0.25cm);
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (0.375, 0,0) arc (-180:180:0.125cm and 0.25cm);
\draw[dashed] (-.625,0, 0) arc (180:360:0.125cm and 0.25cm);
\draw[dashed] (-.625,0,0) arc (180:0:0.125cm and 0.25cm);
\draw[dashed] (-0.5, 0.25,0) -- (0.5, 0.25, 0);
\draw[dashed] (-0.5,-.25,0) -- (.5,-.25,0);
\draw[->] ( -.8, -.7, .5)--( -.4, -.7, .5 );
\draw[->] (-.8, -.7, .5) -- (-.8, -.3, .5);
\draw[->] (-.8, -.7, .5) -- (-.8, -.7, .2);
\node[] at (0, -.4, .4) {$\Sigma_4$};
\node[] at (-.4, -.8, .5){$\operatorname{e}_1$};
\node[] at (-.68, -.3, .5){$\operatorname{e}_3$};
\node[] at (-.8, -.8, .-.1){$\operatorname{e}_2$};
\end{tikzpicture}
\caption{ }
\label{fig:the air cylinder}
\end{subfigure}
\caption{The cube shows the periodicity cell $Y$. The microstructures
$\Sigma_1$, $\Sigma_3$, and $\Sigma_4$ are shown in dark grey. (a)
The metal cylinder $\Sigma_1$. (b) The metal plate $\Sigma_3$. (c)
The metal part $\Sigma_4$ is the complement of a cylinder.
\label{fig:Two-examples-of-perfect-conductors-for-analysis}}
\end{figure}
The scatterer $\Sigma_\eta$ is given as an $\eta$-periodic structure.
We use the periodicity cell $Y \coloneqq [-\frac 12 ,\frac 12]^3$ and
introduce the set $I_\eta$ of all vectors such that a scaled and shifted
copy of $Y$ is contained in $Q_M$, $I_\eta \coloneqq \{j\in
\gz^3|\eta(j+Y)\subset Q_M\}$. A set $\Sigma\subset Y$ specifies the
meta-material, which is defined as
\begin{equation}
\Sigma_\eta \coloneqq \bigcup_{j\in
I_\eta}\ \eta\, (j+\Sigma)\,.\label{eq:Sigma-eta}
\end{equation}
For the microscopic structure $\Sigma$ we consider the following four
examples. The metal cylinder (see Fig.~\ref{fig:the metal
cylinder}) is defined for $r \in (0, 1/2)$ as
\begin{equation}\label{eq:Analysis-Definition-of-metal-cylinder}
\Sigma_{1} \coloneqq \set[\big]{y = (y_1, y_2, y_3) \in Y \given y_1^2 + y_2^2 < r^2}\, .
\end{equation}
The set $\Sigma_2$ is obtained by a rotation which aligns the cylinder
with the $\operatorname{e}_1$-axis,
\begin{equation}\label{eq:Analysis-Definition-of-metal-cylinder-rotated}
\Sigma_2 \coloneqq \set[\big]{y = (y_1, y_2, y_3) \in Y \given
y_2^2+y_3^2 < r^2}\, .
\end{equation}
To define the metal plate (see Fig.~\ref{fig:the metal plate}), we
fix $r \in (0, 1/2)$ and set
\begin{equation}\label{eq:Analysis-Definition-of-rotated-metal-plate}
\Sigma_3 \coloneqq \set[\big]{y = (y_1, y_2, y_3)
\in Y \given y_2 \in (- r, r)}\, .
\end{equation}
The fourth geometry is obtained by removing an \enquote{air cylinder} from
the unit cube (see Fig.~\ref{fig:the air cylinder}); for
$r \in (0, 1/2)$ we set
\begin{equation}\label{eq:Analysis-Definition-of-air-cylinder}
\Sigma_4 \coloneqq Y \setminus \set[\big]{y = (y_1, y_2, y_2) \in Y
\given y_2^2 + y_3^2 < r^2}\,.
\end{equation}
\textbf{Material parameters.} We recall that all materials are
non-magnetic, the relative magnetic permeability is $\mu \equiv
1$. Outside the central region, there is no scatterer; we hence set
$\varepsilon_\eta=1$ in $Q_L$ and $Q_R$. The middle part $Q_M$
contains $\Sigma_\eta$. We set $\varepsilon_\eta = 1$ in $Q_M\setminus
\Sigma_\eta$. It remains to specify the electric permittivity
$\varepsilon_\eta$ in $\Sigma_\eta$. We consider two different
settings.
\smallskip (PC) In the case of \emph{perfect conductors}, we set,
loosely speaking, $\varepsilon_\eta = +\infty$ in $\Sigma_\eta$. More
precisely, we require that $E^\eta$ and $H^\eta$ satisfy
\eqref{eq:time-harmonic-Maxwell-eq} in $G\setminus
\overline{\Sigma}_\eta$ and $E^\eta=H^\eta=0$ in
$\Sigma_\eta$. Boundary conditions are induced on $\partial
\Sigma_\eta$: The magnetic field $H^\eta$ has a vanishing normal
component and the electric field $E^\eta$ has vanishing tangential
components on $\partial \Sigma_\eta$.
\smallskip (HC) In the case of \emph{high-contrast media}, we define
the permittivity as
\begin{equation}\label{eq:Analysis-high-contrast-permittivity}
\varepsilon_\eta(x) \coloneqq
\begin{dcases}
\frac{\varepsilon_1}{\eta^2}& \text{ if } x \in \Sigma_\eta\, ,\\
1 & \text{ if } x \in G \setminus \Sigma_\eta\, ,
\end{dcases}
\end{equation}
where $\varepsilon_1 \in \mathbb{C}$ with $\Re(\varepsilon_1) >0$,
$\Im(\varepsilon_1)>0$. Physically speaking, this means that the
scatterer $Q_M$ consists of periodically disposed metal inclusions
$\Sigma_\delta$ embedded in vacuum. The scaling with $\eta^2$ means
that the optical thickness of the inclusions remains constant; see
\cite{BBF15hommaxwell}.
\smallskip In both settings and throughout this paper, we consider
sequences of solutions $(E^\eta, H^\eta)_{\eta}$ to
\eqref{eq:time-harmonic-Maxwell-eq} which are bounded in $L^2(G;
\mathbb{C}^3)$,
\begin{equation}
\label{eq:Analysis-high-contrast-boundedness-assumption-on-fields}
\sup_{\eta > 0} \int_G \abs{E^\eta}^2 + \abs{H^\eta}^2 < \infty\, .
\end{equation}
Let us remark that the specific geometry of the microstructures $\Sigma_1,
\Sigma_2$, and $\Sigma_4$
is not important; the cylinders could as well be cuboids.
\subsection{Effective equations}
\label{sec:homogenization}
Homogenization theory allows to consider the limit $\eta\to 0$. One
identifies limiting fields $\hat{E}$ and $\hat{H}$ (the latter does
not coincide with the weak limit of $H^\eta$) and limiting equations
for these fields. Using the tool of two-scale convergence, such
results have been obtained for perfect conductors as well as for
high-contrast materials. We briefly summarize the main findings here;
analysis and numerics below are built upon these results.
\smallskip \textbf{Perfect conductors (PC).} The homogenization analysis for
this case has been performed in \cite{SU17hommaxwell}. Since the
parameters of vacuum are used outside the scatterer, the original
Maxwell equations describe the limiting fields in $Q_L$ and $Q_R$. In
the meta-material $Q_M$, however, different equations hold. There
holds $E^\eta\wto\hat{E}$ and $H^\eta \wto \hat{\mu}\hat{H}$ in $L^2(G)$
and the fields $\hat{E}$ and $\hat{H}$ solve
\begin{subequations}\label{eq:Analysis-the-general-effective-equations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\curl \hat{E} & = \phantom{-} i \omega \mu_0 \hat{\mu} \hat{H} &&
\text{ in } G\,
, \label{eq:Analysis-the-general-effective-equations-1}\\
\curl \hat{H} & = - i \omega \varepsilon_0 \hat{\varepsilon} \hat{E}
&& \text{ in } G \setminus Q_M \,
, \label{eq:Analysis-the-general-effective-equations-2}\\
(\curl \hat{H})_k & = -i \omega \varepsilon_0 (\hat{\varepsilon }
\hat{E})_k && \text{ in } G\, , \text{ for every } k \in
\mathcal{N}_{\Sigma}\,
, \label{eq:Analysis-the-general-effective-equations-3}\\
\hat{E}_k & = 0 && \text{ in } Q_M \, , \text{ for every } k \in
\mathcal{L}_{\Sigma}\,
, \label{eq:Analysis-the-general-effective-equations-4}\\
\hat{H}_k & = 0 && \text{ in } Q_M \, , \text{ for every } k \in
\mathcal{N}_{Y \setminus \overline{\Sigma}}\, .
\label{eq:Analysis-the-general-effective-equations-5}
\end{empheq}
\end{subequations}
The effective coefficients $\hat{\mu}$ and $\hat{\varepsilon}$ are
determined by cell-problems. For the cell-problems, details on the
index sets, and the derivation of system
\eqref{eq:Analysis-the-general-effective-equations}, we refer to
\cite{SU17hommaxwell}. The index sets $\mathcal{N}_{\Sigma}$,
$\mathcal{L}_{\Sigma}$, and $\mathcal{N}_{Y \setminus
\overline{\Sigma}}$ are subsets of $\set{1,2,3}$ and can be
determined easily from topological properties of $\Sigma$. Loosely
speaking: An index $k$ is in the set $\mathcal{L}_{\Sigma}$, if there
is a curve (loop) that runs in $\Sigma$ and connects opposite faces of
$Y$ in direction $\operatorname{e}_k$. An index $k$ is in $\mathcal{N}_{\Sigma}$, if
there is no loop of that kind. We collect the index sets
$\mathcal{N}_{\Sigma}$, $\mathcal{L}_{\Sigma}$, and $\mathcal{N}_{Y
\setminus \overline{\Sigma}}$ for the geometries $\Sigma_1$ to
$\Sigma_4$ in Table \ref{tab:analysissummary}.
\begin{table}
\caption{Index sets $\mathcal{N}_{\Sigma}$, $\mathcal{L}_{\Sigma}$,
and $\mathcal{N}_{Y \setminus \overline{\Sigma}}$ for
microstructures $\Sigma_1$ to $\Sigma_4$ of
\eqref{eq:Analysis-Definition-of-metal-cylinder}--\eqref
{eq:Analysis-Definition-of-air-cylinder}.}
\label{tab:analysissummary}
\centering
\begin{tabular}{@{}ccccc@{}}
\toprule
geometry&metal cylinder $\Sigma_1$&metal cylinder $\Sigma_2$&metal plate $\Sigma_3$&air cylinder $\Sigma_4$\\
\midrule
$\mathcal{N}_{\Sigma}$& $\{1, 2\}$ & $\{2,3\}$ & $\{2\}$ & $\emptyset$ \\
\midrule
$\mathcal{L}_{\Sigma}$& $\{3\}$ & $\{1\}$ & $\{1,3\}$ & $\{1,2,3\}$ \\
\midrule
$\mathcal{N}_{Y \setminus \overline{\Sigma}}$& $\emptyset$ & $\emptyset$ & $\{2\}$ & $\{2,3\}$ \\
\bottomrule
\end{tabular}
\end{table}
We will specify equations \eqref
{eq:Analysis-the-general-effective-equations-3}--\eqref
{eq:Analysis-the-general-effective-equations-5} for the four chosen
geometries in Section \ref
{sec:Calculation-of-transmission-and-reflection-coefficient}. With
the effective equations for the perfect conductors at hand, one can
ask for the transmission and reflection coefficients of the
meta-material. This is the goal of our analysis in Section
\ref{sec:Calculation-of-transmission-and-reflection-coefficient}.
\smallskip \textbf{high-contrast media (HC).} Homogenization results
for high-contrast media are essentially restricted to the case of
non-connected metal parts, i.e., to geometries that are obtained by
$\Sigma$ which is compactly embedded in $Y$ (it does not touch the
boundary of the cube); see, e.g., \cite {BBF09hommaxwell,
BS10splitring, BBF15hommaxwell}. The few exceptions are mentioned
below.
For such geometries, the limit equations have again the form of
Maxwell's equations,
\begin{subequations}\label{eq:Analysis-high-contrast-effective-equations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\curl \hat{E} & = \phantom{-} i \omega \mu_0 \hat{\mu} \hat{H} &&
\text{ in } G\,
, \label{eq:Analysis-high-contrast-effective-equations-1}\\
\curl \hat{H} & = - i \omega \varepsilon_0 \hat{\varepsilon} \hat{E}
&& \text{ in } G .
\end{empheq}
\end{subequations}
In $Q_L\cup Q_R$, the effective fields coincide with the weak limits
of the original fields, and the effective relative coefficients are
unit tensors. In the meta-material $Q_M$, however, the high-contrast
in the definition of the permittivity $\varepsilon_\eta$ in
\eqref{eq:Analysis-high-contrast-permittivity} leads to non-trivial
limit equations. The effective material parameters $\hat{\varepsilon}$
and $\hat{\mu}$ are obtained via cell problems and they can take
values that are not to be expected from the choice of the material
parameters in the $\eta$-problem.
As discussed in Section \ref{sec:geometry}, time-harmonic Maxwell's
equations can equivalently be written as a single second order PDE for
the $H$-field or the $E$-field. For the $H$-field we obtain
\begin{align}
\label{eq:Analysis-high-contrast-effective-eq-H}
\curl \widehat{\varepsilon^{-1}}\curl \hat{H}=\omega^2\varepsilon_0\mu_0\hat{\mu} \hat{H}
\qquad \text{in} \quad G\,.
\end{align}
Again, the effective material parameters $\widehat{\varepsilon^{-1}}$
and $\hat{\mu}$ are defined via solutions of cell problems and we
refer to \cite{CC15hommaxwell, Ver17hmmmaxwell} for details. We
remark that the equivalence of the two formulations
\eqref{eq:Analysis-high-contrast-effective-equations} and
\eqref{eq:Analysis-high-contrast-effective-eq-H} has been shown in
\cite{Ver17hmmmaxwell}. In particular, the effective permeability
$\hat{\mu}$ agrees between both formulations and we have the relation
$\widehat{\varepsilon^{-1}}=(\hat{\varepsilon})^{-1}$.
The effective equations
\eqref{eq:Analysis-high-contrast-effective-equations} or
\eqref{eq:Analysis-high-contrast-effective-eq-H} mean that, in the
limit $\eta\to 0$, the meta-material $Q_M$ with high-contrast
permittivity $\varepsilon_\eta$ behaves like a homogeneous material
with permittivity $\hat{\varepsilon}$ and permeability $\hat{\mu}$.
The occurrence of a permeability $\hat{\mu}$ in the effective
equations is striking and this effect is known as artificial
magnetism; see \cite{BF04homhelmholtz}. Moreover, $\hat{\mu}$ depends
on the frequency $\omega$ and it can have a negative real part for
certain frequencies. Negative values of the permeability are caused
by (Mie) resonances in the inclusions $\Sigma$ and are studied in
detail in \cite{BBF15hommaxwell, Ver17hmmmaxwell}.
As mentioned, a crucial assumption for the homogenization analysis in
\cite{BBF15hommaxwell, CC15hommaxwell} is that $\Sigma$ is compactly
contained in the unit cube. For the four geometries
$\Sigma_1$ to $\Sigma_4$, this assumption is clearly not met; we
therefore ask whether certain components of the effective fields
$\hat{E}$ and $\hat{H}$ vanish in this case as in the case of perfect
conductors. This motivates our analysis in Section \ref
{sec:high-contrast-media} as well as the numerical experiments in
Section \ref{subsec:experiment}.
Regarding known results on non-compactly contained inclusions we
mention the thin wires in \cite {BF-thinwires} and \cite
{LS15negindex}, and the dimensionally reduced analysis of the metal
plates $\Sigma_3$ in \cite {BS13plasmonwaves}.
\section{Analysis of the microscopic geometries $\Sigma_1$ to
$\Sigma_4$}
\label{sec:analysis}
In Section \ref
{sec:Calculation-of-transmission-and-reflection-coefficient}, we treat
the case of perfect conductors and compute the transmission
coefficients from the effective equations \eqref
{eq:Analysis-the-general-effective-equations}. In Section \ref
{sec:high-contrast-media}, we treat the case of high-contrast media
and discuss the possibility of nontrivial transmission coefficients.
\subsection{Transmission and reflection coefficients for perfect
conductors}
\label{sec:Calculation-of-transmission-and-reflection-coefficient}
We compute the transmission and reflection coefficients for four
different geometries: metal cylinders, metal plate, and air cylinder.
We consider the wave guide $G=Q_L\cup \bar Q_M\cup Q_R$ of Section
\ref{sec:geometry} and impose periodic boundary conditions on the
lateral boundary of $G$. We recall that the four microscopic
structures $\Sigma_1$ to $\Sigma_4$ are defined in
\eqref{eq:Analysis-Definition-of-metal-cylinder}--\eqref
{eq:Analysis-Definition-of-air-cylinder}. Based on the effective
equations \eqref {eq:Analysis-the-general-effective-equations} for the
perfect conductors, we compute the transmission and reflection
coefficients for these geometries.
\smallskip \textbf{Results for perfect conductors.} Before we discuss the
examples in detail, we present an overview of the results. The
propagation of the electromagnetic wave in vacuum is described by the
time-harmonic Maxwell equations
\begin{subequations}\label{eq:Analysis-time-harmonic-Maxwell-eq-in-QR}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\curl \hat{E}& = \phantom{-} i \omega \mu_0 \hat{H}\quad && \text{ in }
Q_L \cup Q_R\, , \label{eq:Analysis-first-time-harmonic-Maxwell-eq-in-QR} \\
\curl \hat{H} &= - i \omega \varepsilon_0 \hat{E} && \text{ in }
Q_L \cup Q_R\, .\label{eq:Analysis-second-time-harmonic-Maxwell-eq-in-QR}
\end{empheq}
\end{subequations}
For the electromagnetic fields, we use the time-convention $\operatorname{e}^{- i
\omega t}$.
From~\eqref{eq:Analysis-time-harmonic-Maxwell-eq-in-QR} we deduce that
both fields are divergence-free in $Q_L \cup Q_R$. We shall assume
that the electric field $\hat{E} \colon G \to \mathbb{C}^3$ in $Q_R$
is the superposition of a normalized incoming wave with normal
incidence and a reflected wave:
\begin{equation}\label{eq:Analysis-general-ansatz-for-electric-field-in-QR}
\hat{E}(x) \coloneqq \big(\operatorname{e}^{-i k_0 x_1} + R \operatorname{e}^{i k_0 x_1}
\big) \operatorname{e}_k \, ,
\end{equation}
for $x=(x_1, x_2, x_3) \in Q_R$ and $k \in \set{2,3}$. Here, $R \in \mathbb{C}$
is the reflection coefficient and $k_0 = \omega \sqrt{\varepsilon_0
\mu_0}$. Note that the electric field $\hat{E}$
in~\eqref{eq:Analysis-general-ansatz-for-electric-field-in-QR} travels
along the $x_1$-axis from right to left.
Due to~\eqref{eq:Analysis-first-time-harmonic-Maxwell-eq-in-QR}, the effective
magnetic field $\hat{H} \colon G \to \mathbb{C}^3$ is given by
\begin{equation}\label{eq:Analysis-general-ansatz-for-magnetic-field-in-QR}
\hat{H}(x) = (-1)^{l}\frac{k_0}{\omega \mu_0} \big(\operatorname{e}^{-i k_0 x_1}
- R \operatorname{e}^{i k_0
x_1} \big) \operatorname{e}_l\, ,
\end{equation}
where $l = 2$ if $k = 3$ and $l = 3$ if
$k=2$ and $x \in Q_R$. Equation~\eqref{eq:Analysis-second-time-harmonic-Maxwell-eq-in-QR}
is satisfied in $Q_R$ by our choice of $k_0$.
On the other hand, for the transmitted electromagnetic wave in the
left domain $Q_L$, we make the ansatz
\begin{equation}\label{eq:Analysis-general-ansatz-for-E-and-H-in-QL}
\hat{E}(x) = T \operatorname{e}^{-i k_0 (x_1+L)}\operatorname{e}_k \quad \text{ and } \quad
\hat{H}(x) = (-1)^l \frac{k_0}{\omega \mu_0} T \operatorname{e}^{-i k_0 (x_1+L)} \operatorname{e}_l \,,
\end{equation}
where $T\in \mathbb{C}$ is the transmission coefficient. We recall that $L >0$
is the width of the meta-material $Q_M$ and $\{ x_1 = -L\}$ is the
interface between left and middle domain. Since the meta-material in
$Q_M$ can lead to reflections, the transmission coefficient $T \in
\mathbb{C}$ does not necessarily satisfy $\abs{T} = 1$; by energy
conservation there always holds $\abs{T} = 1- \abs{R}$.
Our results are collected in Table \ref {table:res-PC}. The table
lists transmission coefficients for the four geometries in the case
that the incoming magnetic field $H$ is parallel to $\operatorname{e}_3$.
\begin{table}[h]
\centering
\begin{tabular}{c c}
\toprule
microstructure $\Sigma$ &
transmission coefficient $T$
\\
\midrule
metal cylinder $\Sigma_1$ &
$T =4 p_1\sqrt{\alpha\gamma}
\Big[(\alpha + \gamma)(1-p_1^2) + 2
\sqrt{\alpha \gamma} (1+ p_1^2)\Big]^{-1}$ \\[0.3cm]
metal cylinder $\Sigma_2$
& $T = 4 p_2 \sqrt{\gamma}
\Big[(1+ \gamma)(1- p_2^2) + 2 \sqrt{\gamma}(1+p_2^2) \Big]^{-1}$ \\[0.3cm]
metal plate $\Sigma_3$ & $ T = 4p_0 \alpha
\Big[(1+\alpha^2)(1-p_0^2) + 2 \alpha (1+
p_0^2)\Big]^{-1}$ \\[0.3cm]
air cylinder $\Sigma_4$ & $T =0$ \\
\bottomrule
\end{tabular}
\caption{Overview of the transmission coefficients $T$ when $H$ is parallel
to $\operatorname{e}_3$. We see, in particular, that $T$ is vanishing for the structure
$\Sigma_4$, but it is nonzero for the other micro-structures. The constant
$\gamma \in \mathbb{C}$ depends on
the microstructure and on solutions to cell problems, and is defined in the
subsequent sections, $\alpha \coloneqq \abs{ Y \setminus \Sigma}$
is the volume fraction of air, $L > 0$ is the width of the meta-material $Q_M$. We use
$k_0 = \omega \sqrt{\varepsilon_0 \mu_0}$ and the numbers $p_0
\coloneqq \operatorname{e}^{i k_0 L}$, $p_1 \coloneqq p_0 \operatorname{e}^{ i \sqrt{\alpha
\gamma} L}$, and $p_2 \coloneqq p_0 \operatorname{e}^{ i \sqrt{\gamma}L}$.
\label{table:res-PC}}
\end{table}
In the remainder of this section we compute the transmission
coefficient $T$ and the reflection coefficient $R$ for the four
microscopic geometries and verify, in particular, the formulas of
Table \ref {table:res-PC}. Moreover, the effective fields in the
meta-material $Q_M$ are determined.
\subsubsection{The metal cylinder $\Sigma_1$}
\label{sec:metal-cylinder}
The metal cylinder $\Sigma_1$ has a high symmetry, which allows to
compute the effective permeability $\hat{\mu}$. To do so, we define
the projection $\pi \colon Y \to \mathbb{R}^2$ onto the first two components,
i.e., $\pi(y_1, y_2, y_3) \coloneqq (y_1, y_2)$. Moreover, we set
$Y^2 \coloneqq \pi(Y)$ and $\Sigma^2_1 \coloneqq \pi(\Sigma_1)$.
Choose $l \in \set{1,2}$ and denote by $H^l \in L^2(Y; \mathbb{C}^3)$
the distributional periodic solution of
\begin{subequations}\label{eq:Analysis-Cell-problem-for-H-field}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\curl H^l & = 0&& \text{ in } Y \setminus \overline{\Sigma}_1\, ,\\
\Div H^l & = 0 && \text{ in } Y\, ,\\
H^l & = 0 && \text{ in } \Sigma_1\, ,
\end{empheq}
\text{with}
\begin{equation}
\oint H^l = \operatorname{e}_l\, .
\end{equation}
\end{subequations}
The normalization of the last equation is defined in \cite
{SU17hommaxwell}; loosely speaking, the left hand collects values of
line integrals of $H^l$, where the lines are curves in $Y\setminus
\Sigma$ and connect opposite faces of $Y$. Problem \eqref
{eq:Analysis-Cell-problem-for-H-field} is uniquely solvable by
\cite[Lemma 3.5]{SU17hommaxwell}. Given the field $H^l = (H^l_1,
H^l_2, H^l_3)$ we define the field $h^l \colon Y^2 \to \mathbb{C}^2$
as
\begin{equation}\label{eq:Anaylsis-Definition-of-2-dimensional-H-field}
h^l(y_1, y_2) \coloneqq \int_0^1 (-H^l_2, H^l_1)(y_1, y_2, y_3) \d y_3\, .
\end{equation}
\begin{lemma}\label{lem:}
Let $H^l \in L^2(Y; \mathbb{C}^3)$ be the solution of
\eqref{eq:Analysis-Cell-problem-for-H-field}. Then $h^l \in L^2(Y^2;
\mathbb{C}^2)$ of \eqref
{eq:Anaylsis-Definition-of-2-dimensional-H-field} is a
distributional periodic solution to the two-dimensional problem
\begin{subequations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\Div h^l & = 0 && \text{ in } Y^2 \setminus \Sigma_1^2\, ,\\
\nabla^{\bot} \cdot h^l & = 0 && \text{ in } Y^2\, , \\
h^l & = 0 && \text{ in } \Sigma_1^2\, .
\end{empheq}
\end{subequations}
Moreover, there exists a potential $\psi \in H_{\sharp}^1(Y^2; \mathbb{C})$
such that $h^l = \nabla \psi - \olddelta_{2l}\operatorname{e}_1 +
\olddelta_{1l}\operatorname{e}_2$.
\end{lemma}
\begin{proof}
The proof consists of a straightforward calculation.
\end{proof}
The decomposition of $h^l$ allows to determine the effective
permeability $\hat{\mu}$, which, by \cite{SU17hommaxwell}, is given as
\begin{equation}\label{eq:Analysis-effective-permeability}
\hat{\mu}(x) \coloneqq \mu_{\eff} \mathds{1}_{Q_M}(x) + \Id \mathds{1}_{G \setminus
\overline{Q_M}}(x)\, ,
\end{equation}
where
\begin{equation}
\label{eq:Analysis-definition-of-the-effective-permeability}
(\mu_{\eff})_{kl} \coloneqq \int_Y H^l \cdot \operatorname{e}_k\, .
\end{equation}
\begin{lemma}[Effective permeability for the metal cylinder]
For the microstructure $\Sigma = \Sigma_1$ the
permeability $\mu_{\eff}$ is given by
\begin{equation}\label{eq:Analysis-metal-cylinder-effective-permeability}
\mu_{\eff} =
\diag \big( 1, 1, \abs{Y \setminus \Sigma_1} \big) \, .
\end{equation}
\end{lemma}
\begin{proof}
To shorten the notation, we write $y^{\prime} \coloneqq (y_1, y_2)
\in Y^2$. Applying Fubini's theorem and using the decomposition of
$h^1$, we find that
\begin{equation*}
(\mu_{\eff})_{11} = \int_Y H^1 \cdot \operatorname{e}_1 = \int_{Y^2} h^1_2(y^{\prime}) \d y^{\prime}
= \int_{Y^2} \partial_2 \psi (y^{\prime}) \d y^{\prime} + \abs{Y^2} = 1\,,
\end{equation*}
where, in the last equality, we exploited that $\psi$ is
$Y^2$-periodic and that $\abs{Y^2} = 1$. A similar computation shows
that ${(\mu_{\eff})}_{22} = 1$.
To compute ${(\mu_{\eff})}_{12}$, we note that $h^1_1(y^{\prime})
= \partial_1 \psi(y^{\prime}) $. Applying Fubini's theorem, we find
\begin{equation*}
{(\mu_{\eff})}_{12} = \int_Y H^1 \cdot \operatorname{e}_2
= - \int_{Y^2} h^1_1(y^{\prime}) \d y^{\prime}
= -\int_{Y^2} \partial_1 \psi(y^{\prime}) \d y^{\prime} = 0\,.
\end{equation*}
As $h^2_2(y^{\prime}) = \partial_2 \psi(y^{\prime})$, we can proceed
as before and find ${(\mu_{\eff})}_{21} = 0$.
One readily checks that $H^3(y ) \coloneqq \mathds{1}_{Y \setminus
\Sigma_1}(y) \operatorname{e}_3$ is the solution of the cell problem
\eqref{eq:Analysis-Cell-problem-for-H-field} with $\oint H^3 =
\operatorname{e}_3$. The missing entries of the effective permeability matrix
$\mu_{\eff}$ can now be computed using the formula for $H^3$ and the
definition of $\mu_{\eff}$ in
\eqref{eq:Analysis-definition-of-the-effective-permeability}.
\end{proof}
Besides $\hat{\mu}$, we also need the effective permittivity
$\hat{\varepsilon}$.
For $l \in \set{1,2,3}$ we denote by $E^l \in L^2(Y ; \mathbb{C}^3)$ the weak
periodic solution to
\begin{subequations}\label{eq:Analysis-cell-problem-for-electric-field}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\label{eq:3}
\curl E^l & = 0 && \text{ in } Y \, ,\\
\Div E^l & = 0 && \text{ in } Y \setminus \overline{\Sigma}_1\, ,\\
E^l &= 0 && \text{ in } \Sigma_1\, ,
\end{empheq}
with
\begin{equation}
\int_Y E^l = \operatorname{e}_l\, .
\end{equation}
\end{subequations}
Problem~\eqref{eq:Analysis-cell-problem-for-electric-field} is
uniquely solvable by~\cite[Lemma 3.1]{SU17hommaxwell}.
Consequently, the solutions
to~\eqref{eq:Analysis-cell-problem-for-electric-field} are real vector
fields. Indeed, for each index $l \in \set{1,2,3}$ the vector field
$\Im (E^l) \colon Y \to \mathbb{R}^3$ is a weak solution
to~\eqref{eq:Analysis-cell-problem-for-electric-field} with $\int_Y
\Im (E^l) = 0$ and hence $\Im(E^l) = 0$ in $Y$.
As in~\cite{SU17hommaxwell} we set
\begin{equation}\label{eq:Analysis-effective-permittivity}
\hat{\varepsilon} (x) \coloneqq \varepsilon_{\eff}
\mathds{1}_{Q_M}(x) + \Id \mathds{1}_{G \setminus
\overline{Q_M}}(x)\, ,
\end{equation}
where
\begin{equation}\label{eq:definition-effective-permittivity}
(\varepsilon_{\eff})_{kl} \coloneqq \int_Y E^k \cdot E^l\, .
\end{equation}
\begin{lemma}[Effective permittivity for the metal cylinder]
For the microstructure $\Sigma = \Sigma_1$, the permittivity
$\varepsilon_{\eff}$ is given by
\begin{equation}\label{eq:Analysis-metal-cylinder-effective-permittivity}
\varepsilon_{\eff} = \diag(\gamma, \gamma, 0)
\, ,
\end{equation}
where $\gamma \coloneqq ({\varepsilon}_{\eff})_{1,1}$.
\end{lemma}
\begin{proof}
As shown in Table~\ref{tab:analysissummary}, we find that
$\mathcal{N}_{\Sigma_1}= \set{1,2}$. From~\cite[Lemma
3.2]{SU17hommaxwell} we hence deduce that
$(\varepsilon_{\eff})_{k,3}$ as well as $(\varepsilon_{\eff})_{3,k}$
vanish for all $k \in \set{1,2,3}$. We claim that the matrix
$\varepsilon_{\eff}$ is symmetric. Because of
$(\varepsilon_{\eff})_{k,3} = (\varepsilon_{\eff})_{3,k} = 0$ we
only have to prove that
$(\varepsilon_{\eff})_{1,2} = (\varepsilon_{\eff})_{2,1}$. As the
solutions $E^1$ and $E^2$ of the cell
problem~\eqref{eq:Analysis-cell-problem-for-electric-field} are real
vector fields, we compute that
\begin{equation*}
(\varepsilon_{\eff})_{1,2} = \int_Y E^1 \cdot E^2 = \int_Y E^2 \cdot
E^1 = (\varepsilon_{\eff})_{2,1}\, .
\end{equation*}
To show that $(\varepsilon_{\eff})_{1,2} =
(\varepsilon_{\eff})_{2,1} = 0$, we consider the map $M \colon Y \to
Y$ that is defined by the diagonal matrix $\diag (-1, 1, 1)$.
Note that $M(\Sigma_1) = \Sigma_1$. To shorten the notation, we set
$E \coloneqq E^1$. Consider the vector field $F \colon Y \to \mathbb{R}^3$,
\begin{equation*}
F(x) \coloneqq ME(Mx) =
\begin{pmatrix}
-E_1 \\
E_2 \\
E_3
\end{pmatrix}
(-x_1, x_2, x_3)\, .
\end{equation*}
One readily checks that $F$ is a solution to the cell
problem~\eqref{eq:Analysis-cell-problem-for-electric-field} with
$
\int_Y F = -\operatorname{e}_1\, .
$
Due to the unique solvability of the cell
problem~\eqref{eq:Analysis-cell-problem-for-electric-field}, we conclude that
that $F=-E $. Similarly, we find that $ME^2\circ M = E^2$. Thus
\begin{equation*}
(\varepsilon_{\eff})_{1,2} = \int_Y E^1 \cdot E^2 = -\int_{Y}
ME^1(My)
\cdot ME^2(My) \d y = - \int_Y E^1 \cdot E^2 = -
(\varepsilon_{\eff})_{1,2}\, .
\end{equation*}
Hence $(\varepsilon_{\eff})_{1,2} = (\varepsilon_{\eff})_{2,1} = 0$.
We are left to prove $(\varepsilon_{\eff})_{2,2}
=(\varepsilon_{\eff})_{1,1}$. To do so, we consider the rotation map $R
\colon Y \to Y$ which is defined by the matrix
\begin{equation*}
\begin{pmatrix}
0 & -1 & 0\\
1 & 0 & 0 \\
0 & 0 & 1
\end{pmatrix}
\, .
\end{equation*}
Then $R(\Sigma_1) =
\Sigma_1$. Moreover, as the cell
problem~\eqref{eq:Analysis-cell-problem-for-electric-field} is uniquely
solvable, we find that $RE^2\circ R = -E^1$. Thus
\begin{equation*}
(\varepsilon_{\eff})_{1,1} = \int_Y E^1 \cdot E^1 = \int_{Y} RE^2(Ry)
\cdot RE^2(Ry) \d y = \int_Y E^2 \cdot E^2 = (\varepsilon_{\eff})_{2,2}\, .
\end{equation*}
This proves the claim.
\end{proof}
By Theorem 4.1 of~\cite{SU17hommaxwell}, the microstructure
$\Sigma_1$ together with the effective
permittivity from~\eqref{eq:Analysis-metal-cylinder-effective-permittivity}
and
permeability from~\eqref{eq:Analysis-metal-cylinder-effective-permeability}
implies that the effective equations are
\begin{subequations}\label{eq:Analysis-effective-equations-metal-cylinder-case-1}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\partial_2 \hat{H}_3 - \partial_3 \hat{H}_2 & = - i \omega
\varepsilon_0
(\hat{\varepsilon}\hat{E})_1 && \text{ in } G\, ,\\
\partial_3 \hat{H}_1 - \partial_1 \hat{H}_3 & = - i \omega
\varepsilon_0 (\hat{\varepsilon }\hat{E})_2 && \text{ in } G \label{eq:Analysis-effective-equations-metal-cylinder-perp-case-curlH=E-for-some-components} \, , \\
\hat{E}_3 & = 0 && \text{ in } Q_M \label{eq:Analysis-effective-equations-metal-cylinder-perp-case-E=0-in-QM}\, .
\end{empheq}
\end{subequations}
The
equations~\eqref{eq:Analysis-effective-equations-metal-cylinder-case-1}
do not repeat~\eqref{eq:Analysis-the-general-effective-equations-1}
and~\eqref{eq:Analysis-the-general-effective-equations-2}.
Due
to~\eqref{eq:Analysis-the-general-effective-equations-1},
the effective electric field is divergence-free. As we assume that $\hat{E}$ travels along
the $x_1$-axis, the first component $\hat{E}_1$ vanishes. Due
to~\eqref{eq:Analysis-effective-equations-metal-cylinder-perp-case-E=0-in-QM}
we expect no transmission if the effective electric field is polarized
in $\operatorname{e}_3$-direction. For nontrivial transmission, we may therefore make the following ansatz for
the effective electric field $\hat{E} \colon G \to \mathbb{C}^3$,
\begin{equation*}
\hat{E} (x) \coloneqq \big( \operatorname{e}^{-ik_0 x_1} + R \operatorname{e}^{i k_0 x_1}\big)
\operatorname{e}_2 \quad \text{ for } x = (x_1, x_2, x_3) \in Q_R\,.
\end{equation*}
Thanks
to~\eqref{eq:Analysis-the-general-effective-equations-2}
the magnetic field $\hat{H}$ is given by
\begin{equation*}
\hat{H} (x) = - \frac{k_0}{\omega \mu_0} \big( \operatorname{e}^{-i k_0 x_1} - R
\operatorname{e}^{ i k_0 x_1} \big)\operatorname{e}_3 \quad \text{ for } x = (x_1, x_2, x_3)
\in Q_R\, .
\end{equation*}
In the meta-material $Q_M$ we write
\begin{equation*}
\hat{E}(x) = \big( T_M \operatorname{e}^{- i k_1 x_1} + R_M \operatorname{e}^{i k_1 x_1}\big)\operatorname{e}_2
\text{ and } \hat{H}(x) = - \frac{k_1}{\omega \mu_0 \alpha}\big( T_M \operatorname{e}^{-i k_1 x_1} - R_M
\operatorname{e}^{i k_1 x_1}) \operatorname{e}_3
\end{equation*}
for $x \in Q_M$,
where we used
equation~\eqref{eq:Analysis-the-general-effective-equations-1}
and~\eqref{eq:Analysis-metal-cylinder-effective-permeability} to
determine the magnetic field with
$\alpha \coloneqq \abs{ Y \setminus \Sigma_1}$. To compute the value
of $k_0$ we use
equation~\eqref{eq:Analysis-the-general-effective-equations-2}
and we find that $k_0 = \omega \sqrt{\varepsilon_0
\mu_0}$. From~\eqref{eq:Analysis-effective-equations-metal-cylinder-perp-case-curlH=E-for-some-components}
we deduce that $k_1 = k_0 \sqrt{\alpha \gamma}$. In $Q_L$ we
choose~\eqref{eq:Analysis-general-ansatz-for-E-and-H-in-QL} as the
ansatz for $\hat{E}$ and $\hat{H}$, where $k = 2$ and $l=3$.
\begin{lemma}[Transmission and reflection coefficients]\label{lem:Analysis-metal-cylinder-case-1-transmission-and-reflection-coefficients}
Given the electric and magnetic fields $\hat{E}$ and $\hat{H}$ as
described above. Set $\alpha \coloneqq \abs{Y \setminus \Sigma_1}$,
$k_1 = \omega \sqrt{\varepsilon_0 \mu_0 \alpha \gamma}$, and
$p_1 \coloneqq \operatorname{e}^{ i k_1 L }$.
The coefficients are then given by
\begin{alignat*}{2}
R &= \frac{( \alpha - \gamma)
(1 - p_1^2)}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)}\, , \quad
T_M & =\frac{2 \sqrt{\alpha} ( \sqrt{\alpha} +
\sqrt{\gamma})}{(\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)}\, ,
\\
R_M &= - \frac{2 \sqrt{\alpha} p_1^2 ( \sqrt{\alpha} -
\sqrt{\gamma})}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)} \, , \quad
T & = \frac{4 \sqrt{ \alpha\gamma}p_1}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)} \, .
\end{alignat*}
\end{lemma}
\begin{proof}
By~\eqref{eq:Analysis-the-general-effective-equations-1}
the tangential trace of $\hat{E}$ has no jump across the surfaces
$\set{ x \in G \given x_1 = 0}$ and $\set{ x \in G \given x _1 = -
L}$. Thus
\begin{equation} \label{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-1}
T_M + R_M = 1+R \quad \text{ and } \quad T = p_1 T_M + \frac 1p_1
R_M\, .
\end{equation}
The effective field $\hat{H}$ is parallel to $\operatorname{e}_3$ and hence,
by~\eqref{eq:Analysis-effective-equations-metal-cylinder-perp-case-curlH=E-for-some-components},
the third component $\hat{H}_3$ does not jump across the surfaces $\set{
x \in G \given x_1 = 0}$ and $\set{ x \in G \given x_1 = -L}$. We
may therefore conclude that
\begin{equation}\label{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-2}
\sqrt{\frac{\gamma}{\alpha}} \big(T_M-R_M \big) = 1- R \quad \text{ and } \quad
T = \sqrt{\frac{\gamma}{\alpha}} \bigg( p_1 T_M - \frac 1p_1 R_M
\bigg)\, .
\end{equation}
Here we used that $k_0 = \omega \sqrt{ \mu_0 \varepsilon_0}$ and $k_1
= k_0 \sqrt{ \alpha \gamma}$.
Solving the equations on the left-hand side
in~\eqref{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-1}
and~\eqref{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-2}
for $R$ and the other two equations for $T$,
we find that
\begin{align}\label{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-3}
T_M + R_M - 1& = R = 1- \sqrt{\frac{\gamma}{\alpha}}(T_M-R_M)
\shortintertext{and }
p_1T_M + \frac{1}{p_1}R_M& = T = \sqrt{\frac{\gamma}{\alpha}}
\bigg(p_1 T_M - \frac{1}{p_1}R_M \bigg)\, .\label{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-5}
\end{align}
Setting $d_+ \coloneqq 1+ \sqrt{\gamma/\alpha}$ and
$d_{-}\coloneqq 1 - \sqrt{\gamma/\alpha}$,
equations~\eqref{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-3}
and~\eqref{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-5}
can be written as
\begin{equation}\label{eq:eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-4}
d_+ T_M = 2 - d_{-} R_M \quad \text{ and } \quad
p_1 d_{-}T_M = - \frac{1}{p_1} d_+ R_M\, .
\end{equation}
Solving each of the two equations
in~\eqref{eq:eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-4}
for $R_M$ and then equating the two
expressions for $R_M$, we obtain
\begin{align*}\label{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-5}
T_M &= \frac{2 d_+}{d_+^2 - d_{-}^2p_1^2}
= \frac{ 2 ( 1 + \sqrt{\gamma / \alpha}) }{
(1 + \sqrt{\gamma / \alpha} )^2
- (1 - \sqrt{\gamma / \alpha}
)^2p_1^2}
= \frac{2 \sqrt{\alpha} ( \sqrt{\alpha} +
\sqrt{\gamma})}{(\sqrt{\alpha} + \sqrt{\gamma})^2 -
(\sqrt{\alpha} - \sqrt{\gamma})^2 p_1^2}\, .
\end{align*}
Note that $(\sqrt{\alpha} + \sqrt{\gamma})^2 - (\sqrt{\alpha} -
\sqrt{\gamma})^2p_1^2 = (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)$, which yields the formula for $T_M$.
From the second equation
in~\eqref{eq:eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-4},
we deduce that
\begin{equation*}
R_M = - p_1^2\frac{d_{-}}{d_+}T_M
= - p_1^2 \frac{\sqrt{\alpha} - \sqrt{\gamma}}{\sqrt{\alpha} +
\sqrt{\gamma}} T_M
= - \frac{2 \sqrt{\alpha} p_1^2 ( \sqrt{\alpha} -
\sqrt{\gamma})}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)}\, .
\end{equation*}
By~\eqref{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-3},
we have that
\begin{align*}
R &= T_M + R_M - 1 = \frac{2 \sqrt{\alpha} (\sqrt{\alpha} +
\sqrt{\gamma} ) - 2 \sqrt{\alpha} p_1^2 (\sqrt{\alpha} - \sqrt{\gamma} )}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)} - 1 \\
&= \frac{( \alpha - \gamma)
(1 - p_1^2)}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)}\, .
\end{align*}
To compute the coefficient $T$ we use
equation~\eqref{eq:Analysis-equation-for-transmission-and-reflection-metal-cylinder-case-1-1}
and find that
\begin{align*}
T &= \frac{2 \sqrt{\alpha} ( \sqrt{\alpha} +
\sqrt{\gamma})p_1 - 2 \sqrt{\alpha} (\sqrt{\alpha} - \sqrt{\gamma})p_1}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)}
= \frac{4 \sqrt{ \alpha} \sqrt{\gamma}p_1}{ (\alpha + \gamma)(1-p_1^2)+ 2 \sqrt{\alpha
\gamma} (1+ p_1^2)}\, .
\end{align*}
This proves the claim.
\end{proof}
\subsubsection{The metal cylinder $\Sigma_2$}
\label{sec:metal-cylinder-case}
Similar to the previous section, we shall determine the transmission
and reflection coefficients for a metal cylinder, considering the
microstructure $\Sigma_2$. We define the effective permeability and
the effective permittivity
$\hat{\mu}, \hat{\varepsilon } \colon G \to \mathbb{C}^3$ as
in~\eqref{eq:Analysis-effective-permeability}
and~\eqref{eq:Analysis-effective-permittivity}.
Following the
reasoning of Section~\ref{sec:metal-cylinder}, we find that the
$\mu_{\eff}$ and
$\varepsilon_{\eff}$ are given by
\begin{equation*}
\mu_{\eff} = \diag \big( \abs{Y \setminus \Sigma_2}, 1, 1 \big)
\quad \text{ and } \quad
\varepsilon_{\eff} = \diag (0, \gamma, \gamma)\, ,
\end{equation*}
where $\gamma \in \mathbb{C}$ is defined as $\gamma \coloneqq \int_Y E^2
\cdot E^2$. The effective equations for the microstructure
$\Sigma_2$ are
\begin{subequations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\partial_3 \hat{H}_1 - \partial_1 \hat{H}_3 & = -i \omega
\varepsilon_0 (\hat{\varepsilon} \hat{E})_2 && \text{ in } G \, , \label{eq:Analysis-effective-equations-metal-cylinder-case-2-curlH=E-for-some-components} \\
\partial_1 \hat{H}_2 - \partial_2 \hat{H}_1 & = - i \omega
\varepsilon_0 (\hat{\varepsilon } \hat{E})_3 && \text{ in } G \, ,
\\
\hat{E}_1& = 0 && \text{ in } Q_M\, .
\end{empheq}
\end{subequations}
We may take a similar ansatz for the effective fields as in
Section~\ref{sec:metal-cylinder} and obtain the following transmission
and reflection coefficients. Note that $k_1$ in
Section~\ref{sec:metal-cylinder} has to be replaced by $k_2 \coloneqq
k_0 \sqrt{\gamma}$.
\begin{lemma}[Transmission and reflection coefficients]
Within the setting of Section~\ref{sec:metal-cylinder}, we set $k_2 = \omega \sqrt{\varepsilon_0
\mu_0 \gamma}$ and $p_2 \coloneqq e^{ i k_2 L}$. The reflection and transmission
coefficients for $\Sigma_2$ are given by
\begin{alignat*}{3}
R &= \frac{(1- \gamma) (1-p_2^2)}{ (1+ \gamma)(1-p_2^2) + 2
\sqrt{\gamma}(1+p_2^2)}\, , & \quad & T_M &=\frac{2 (1 +
\sqrt{\gamma})}{(1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)}\, , \\
R_M &= - \frac{2 p_2^2(1 - \sqrt{\gamma})}{ (1+ \gamma)(1-p_2^2) +
2 \sqrt{\gamma}(1+p_2^2)}\, , &\quad & T &= \frac{4 p_2
\sqrt{\gamma}}{ (1+ \gamma)(1-p_2^2) + 2
\sqrt{\gamma}(1+p_2^2)} \, .
\end{alignat*}
\end{lemma}
Note that in the above transmission and reflection coefficients the
volume fraction of air $\alpha = \abs{ Y \setminus \Sigma_2}$ does not
appear. This is different for the metal cylinder $\Sigma_1$; see Lemma~\ref{lem:Analysis-metal-cylinder-case-1-transmission-and-reflection-coefficients}.
\begin{proof}
Thanks
to~\eqref{eq:Analysis-the-general-effective-equations-1}
we know that the tangential components of $\hat{E}$ do not jump across
the surfaces $\set{ x \in G \given x_1 = 0}$ and $\set{x \in G
\given x_1 = -L}$. Hence
\begin{equation}\label{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-1}
1+R = T_M + R_M \quad \text{ and } T = p_2 T_M + \frac{1}{p_2}R_M \, .
\end{equation}
The effective field $\hat{H}$ is parallel to $\operatorname{e}_3$ and hence, due
to~\eqref{eq:Analysis-effective-equations-metal-cylinder-case-2-curlH=E-for-some-components},
the third component $\hat{H}_3$ does neither jump across $\set{ x \in G
\given x_1 = 0}$ nor across $\set{ x \in G \given x_1 = -L}$. We may
therefore conclude that
\begin{equation}\label{eq:eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-2}
1- R = \sqrt{\gamma} (T_M - R_M) \quad \text{ and } \quad T =
\sqrt{\gamma} \bigg(p_2 T_M - \frac{1}{p_2} R_M \bigg)\, .
\end{equation}
Here we used that $k_2 = k_0 \sqrt{\gamma} = \omega \sqrt{
\varepsilon_0 \mu_0 \gamma}$.
Solving the equations on the left-hand side
in~\eqref{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-1}
and~\eqref{eq:eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-2}
for $R$ and the other two for $T$, we find that
\begin{equation*}
T_M +R_M - 1 =R = 1 - \sqrt{\gamma} (T_M - R_M) \: \text{ and }
\:
p_2 T_M + \frac{1}{p_2} R_M =T = \sqrt{\gamma} \bigg(p_2 T_M - \frac{1}{p_2}R_M \bigg)\, .
\end{equation*}
Setting $c_+ \coloneqq 1 + \sqrt{\gamma}$ and $c_{-} \coloneqq 1 -
\sqrt{\gamma}$, these two equations can be re-written as
\begin{equation}\label{eq:eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-3}
c_+ T_M = 2 - c_{-}R_M \quad \text{ and } \quad c_{-}p_2 T_M = -
\frac{1}{p_2} c_+ R_M\, .
\end{equation}
We can solve for $T_M$ and obtain
\begin{equation*}
T_M = \frac{2 c_+}{c_+^2 - c_{-}^2p_2^2} = \frac{2 (1 +
\sqrt{\gamma})}{(1 + \sqrt{\gamma})^2 - (1 - \sqrt{\gamma})^2
p_2^2}\, .
\end{equation*}
Note that $(1+ \sqrt{\gamma})^2 - (1- \sqrt{\gamma})^2p_2^2 = (1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)$, which proves the formula
for $T_M$.
By~\eqref{eq:eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-3}
we then conclude that
\begin{equation*}
R_M = - p_2^2\frac{c_{-}}{c_+}T_M = - p_2^2 \frac{1 -
\sqrt{\gamma}}{1+\sqrt{\gamma}}T_M = - \frac{2 p_2^2(1 - \sqrt{\gamma})}{ (1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)}\, .
\end{equation*}
To determine the coefficient $R$ we recall from above that $R = T_M +
R_M - 1$ and hence
\begin{equation*}
R = \frac{2 (1+ \sqrt{\gamma}) - 2 p_2^2 (1 -
\sqrt{\gamma}) - (1 + \sqrt{\gamma})^2 + (1 - \sqrt{\gamma})^2p_2^2}{ (1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)}
= \frac{(1- \gamma) (1-p_2^2)}{ (1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)}\, .
\end{equation*}
As $T = p_2 T_M + 1/p_2 R_M$, we find that
\begin{equation*}
T = \frac{2 p_2 (1+ \sqrt{\gamma}) - 2 p_2 (1- \sqrt{\gamma})}{ (1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)} = \frac{4 p_2 \sqrt{\gamma}}{ (1+
\gamma)(1-p_2^2) + 2 \sqrt{\gamma}(1+p_2^2)}\, .
\end{equation*}
This proves the claim.
\end{proof}
We chose the same polarization for the electric and the magnetic field
as in Section~\ref{sec:metal-cylinder}. By symmetry of the
microstructure, we may as well assume that $\hat{E}$ is parallel to $\operatorname{e}_3$ and
$\hat{H}$ is parallel to $\operatorname{e}_2$ and obtain the same reflection and
transmission coefficients.
\subsubsection{The metal plate}
\label{sec:metal-plate}
We consider the microstructure $\Sigma_3$; that is, a metal plate
which is perpendicular to $\operatorname{e}_2$.
Following the reasoning in Section 5.2 in~\cite{SU17hommaxwell}, we
determine the effective equations and obtain:
\begin{subequations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\partial_3 \hat{H}_1 - \partial_1 \hat{H}_3 & = - i \omega
\varepsilon_0
\alpha^{-1} \hat{E}_2 && \text{ in } Q_M\, , \label{eq:Analysis-effective-equations-metal-plate-1-case-curlH=E-for-some-components} \\
\hat{E}_1 &= \hat{E}_3 = 0 && \text{ in } Q_M\, , \label{eq:Analysis-effective-equations-metal-plate-1-case-electric-field-in-meta-material}\\
\hat{H}_2 &= 0 && \text{ in } Q_M\, ,
\end{empheq}
\end{subequations}
where $\alpha \coloneqq \abs{Y \setminus \Sigma_3}$.
The electromagnetic wave is assumed to travel in $\operatorname{e}_1$-direction from right to
left. Moreover, by~\eqref{eq:Analysis-the-general-effective-equations-1},
the electric field is divergence free. Hence, the first component
$\hat{E}_1$ vanishes. Because of~\eqref{eq:Analysis-effective-equations-metal-plate-1-case-electric-field-in-meta-material}
we expect no transmission if the electric field is polarized in
$\operatorname{e}_3$-direction. We may therefore make the following ansatz for the
effective electric field $\hat{E} \colon G \to \mathbb{C}^3$,
\begin{equation*}
\hat{E}(x) \coloneqq \big(\operatorname{e}^{-i k_0 x_1} + R \operatorname{e}^{i k_0
x_1}\big)\operatorname{e}_2\quad \text{ for } x = (x_1, x_2, x_3) \in Q_R\, .
\end{equation*}
Thanks to~\eqref{eq:Analysis-the-general-effective-equations-2},
the magnetic field $\hat{H}$ is given by
\begin{equation*}
\hat{H}(x) =- \frac{k_0}{\omega \mu_0} \big( \operatorname{e}^{-i k_0 x_1} - R
\operatorname{e}^{i k_0 x_1} \big) \operatorname{e}_3 \quad \text{ for } x \in Q_R\, .
\end{equation*}
By
equation~\eqref{eq:Analysis-effective-equations-metal-plate-1-case-electric-field-in-meta-material},
the first and the third component of
the effective electric field are trivial; from this and
equation~\eqref{eq:Analysis-the-general-effective-equations-1},
we deduce that
\begin{equation*}
\hat{E}(x) = \big(T_M \operatorname{e}^{-i k_3 x_1} + R_M\operatorname{e}^{i k_3 x_1}\big) \operatorname{e}_2
\quad \text{ and } \quad \hat{H}(x) = -\frac{k_3}{\omega \mu_0
\alpha} \big(T_M \operatorname{e}^{-ik_3x_1} - R_M \operatorname{e}^{i k_3x_1} \big)\operatorname{e}_3\: \text{ in
} Q_M\, .
\end{equation*}
The value of $k_3$ can be determined
by~\eqref{eq:Analysis-effective-equations-metal-plate-1-case-curlH=E-for-some-components}
and we find that $k_3 = k_0 = \omega \sqrt{\varepsilon_0 \mu_0}$. In
$Q_L$, we choose~\eqref{eq:Analysis-general-ansatz-for-E-and-H-in-QL}
as the ansatz for $\hat{E}$ and $\hat{H}$, where $k = 2$ and $l = 3$.
\begin{lemma}[Transmission and reflection coefficients]
Given the effective fields $\hat{E}$ and $\hat{H}$ as described
above. Set $\alpha \coloneqq \abs{Y \setminus \Sigma_3}$ and $p_0
\coloneqq \operatorname{e}^{ i \omega \sqrt{\varepsilon_0 \mu_0} L}$. The
reflection and transmission coefficients are given by
\begin{alignat}{3}
R &= \frac{(\alpha^2 - 1)(1- p_0^2)}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)}\, , &\qquad & T_M
&= \frac{2 \alpha (\alpha +1)}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)}\,
,
\\
R_M &= -\frac{2 \alpha p_0^2 (\alpha - 1)}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)} \, ,& \qquad & T & = \frac{4 p_0 \alpha}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)} \, .
\end{alignat}
\end{lemma}
\begin{proof}
From~\eqref{eq:Analysis-the-general-effective-equations-1}
we deduce that $\curl \hat{E}$ has no singular part and hence the
tangential trace of $\hat{E}$ along the surfaces $\set{x \in G \given
x_1 = 0}$ and $\set{x \in G \given x_1 = -L}$ does not jump. Thus
\begin{equation}\label{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-1}
T_M + R_M= 1+R \quad \text{ and } \quad T = p_0T_M + \frac{1}{p_0} R_M\,.
\end{equation}
As $\hat{H}$ is parallel to $\operatorname{e}_3$, we deduce
from~\eqref{eq:Analysis-effective-equations-metal-plate-1-case-curlH=E-for-some-components}
that $\hat{H}_3$ does not jump across the surfaces $\set{x \in G
\given x_1 = 0}$ and $\set{x \in G \given x_1 = -L}$. This implies
that
\begin{equation}\label{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-2}
1 - R = \frac{1}{\alpha} \big(T_M - R_M \big) \quad \text{ and }
\quad T = \frac{1}{\alpha}\Big( p_0 T_M - \frac{1}{p_0} R_M\Big) \, .
\end{equation}
Here we used that $k_0 = k_3 = \omega \sqrt{\varepsilon_0 \mu_0}$.
Note that $\alpha >0$ and hence we find $a >0$ such that $\sqrt{a} =
1/ \alpha$. With this new parameter $a$, the equations
in~\eqref{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-2}
read
\begin{equation}\label{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-3}
\sqrt{a}(T_M-R_M) = 1-R \quad \text{ and } \quad
T = \sqrt{a} \bigg(p_0 T_M - \frac{1}{p_0}R_M \bigg) \, .
\end{equation}
Thus the equations
in~\eqref{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-1}
and~\eqref{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-3}
have the same structure as the equations
in~\eqref{eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-1}
and~\eqref{eq:eq:Analysis-equations-for-transmission-and-reflection-coefficients-metal-plate-case-2-2}. We
may therefore use the formulas for $R, T, R_M$, and $T_M$ derived in
Section~\ref{sec:metal-cylinder-case}. Note that
\begin{equation*}
1+ \sqrt{a} = \frac{\alpha +1}{\alpha}\, \quad 1- \sqrt{a} =
\frac{\alpha - 1}{\alpha}\, , \quad \text{ and } \quad 1-a =
\frac{\alpha^2 - 1}{\alpha^2}\, .
\end{equation*}
Thus
\begin{align*}
T_M &= \frac{2 (1+ \sqrt{a})}{(1+ \sqrt{a})^2 - (1- \sqrt{a})^2p_0^2}
= \frac{2 \alpha (\alpha +1)}{(\alpha + 1)^2 -
(\alpha-1)^2p_0^2}
= \frac{2 \alpha (\alpha +1)}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)}\,
, \\
R_M &= - \frac{2 p_0^2 (1- \sqrt{a})}{(1+ \sqrt{a}) - (1-
\sqrt{a})p_0^2}
= - \frac{2 \alpha p_0^2 (\alpha - 1)}{(\alpha + 1)^2 - (\alpha
-1)^2p_0^2}
= -\frac{2 \alpha p_0^2 (\alpha - 1)}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)}\, ,\\
R &= \frac{(1-a)(1-p_0^2)}{(1 + \sqrt{a}) - (1- \sqrt{a})p_0^2}
= \frac{(\alpha^2 - 1)(1-p_0^2)}{(\alpha +1)^2 -
(\alpha-1)^2p_0^2}
= \frac{(\alpha^2 - 1)(1- p_0^2)}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)}\, ,
\shortintertext{and}
T &= \frac{4 p_0\sqrt{\alpha}}{(1+ \sqrt{\alpha})^2 - (1-
\sqrt{\alpha})^2p_0^2}
= \frac{4 p_0 \alpha}{(\alpha+1)^2 - (\alpha-1)^2 p_0^2}
= \frac{4 p_0 \alpha}{(1 + \alpha^2)(1-p_0^2) + 2
\alpha (1+ p_0^2)}\, .
\end{align*}
This proves the claim.
\end{proof}
\subsubsection{The air cylinder}
\label{sec:air-cylinder}
We consider the microstructure $\Sigma_4$; that is, an air cylinder
with symmetry axis parallel to $\operatorname{e}_1$ (see Fig.~\ref{fig:the air
cylinder}).
Combining the effective
equations~\eqref{eq:Analysis-the-general-effective-equations} with the
index sets in Table~\ref{tab:analysissummary}, we obtain the effective
system for this case:
\begin{subequations}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\hat{E} & = 0 && \text{ in } Q_M \,
, \label{eq:Analysis-effective-equations-air-cylinder-E=0}\\
\hat{H}_2 & = \hat{H}_3 = 0 && \text{ in } Q_M \, . \label{eq:Analysis-effective-equations-air-cylinder-H=0-for-some-components}
\end{empheq}
\end{subequations}
As in the previous sections, we choose the following ansatz for the
effective fields $\hat{E}, \hat{H} \colon G \to \mathbb{C}^3$,
\begin{equation*}
\hat{E}(x) \coloneqq \big( \operatorname{e}^{-i k_0 x_1} + R \operatorname{e}^{i k_0 x_1}
\big)\operatorname{e}_2 \quad \text{ and } \quad \hat{H}(x) =- \big( \operatorname{e}^{i k_0 x_1} - R
\operatorname{e}^{-i k_0 x_1}\big) \operatorname{e}_3 \quad
\text{ for } x \in Q_R\, .
\end{equation*}
Equation~\eqref{eq:Analysis-the-general-effective-equations-2}
determines the wave number and we find that $k_0 = \omega \sqrt{\varepsilon_0 \mu_0}$. The effective
electric field $\hat{E}$ vanishes in the meta-material $Q_M$ and hence,
by~\eqref{eq:Analysis-the-general-effective-equations-1}
and~\eqref{eq:Analysis-effective-equations-air-cylinder-H=0-for-some-components},
there is also no effective magnetic field in $Q_M$. So $\hat{E} =
\hat{H} = 0$ in
$Q_M$. Equation~\eqref{eq:Analysis-the-general-effective-equations-1}
implies that the tangential trace of $\hat{E}$ does not jump across the
surface $\set{ x \in G \given x_1 =0}$. Thus
\begin{equation*}
R = -1\, .
\end{equation*}
As no field is transmitted through the meta-material $Q_M$, there is
neither an electric nor an magnetic field in $Q_L$. In other words,
$\hat{E} = \hat{H} = 0$ in $Q_L$. We have thus shown that
\begin{equation*}
R = -1 \quad \text{ and } \quad T = 0\, .
\end{equation*}
\subsection{Vanishing limiting fields in high-contrast media,
2D-analysis}
\label{sec:high-contrast-media}
In this section, we perform an analysis of high-contrast media. Of the
four geometries $\Sigma_1$ to $\Sigma_4$, we study the two
$\operatorname{e}_3$-invariant geometries: the metal cylinder $\Sigma_1$ and the
metal plate $\Sigma_3$, compare Fig.~\ref
{fig:Two-examples-of-perfect-conductors-for-analysis}. We analyze the
time-harmonic Maxwell equations \eqref{eq:time-harmonic-Maxwell-eq}
with the high-contrast permittivity $\varepsilon_\eta$ of \eqref
{eq:Analysis-high-contrast-permittivity}. We recall that the sequence
of solutions $(E^\eta, H^\eta)_\eta$ is assumed to satisfy the
$L^2(G)$-bound \eqref
{eq:Analysis-high-contrast-boundedness-assumption-on-fields}. We are
interested in the limit behaviour of $(E^{\eta}, H^{\eta})_\eta$ as
$\eta \to 0$.
When we consider perfect conductors, the effective equations \eqref
{eq:Analysis-the-general-effective-equations} imply that some
components of $E^\eta$ or $H^\eta$ converge weakly to $0$ in the
meta-material $Q_M$. For media with high-contrast, we do not have such
a result (we recall that homogenization usually considers compactly
contained geometries $\Sigma\subset Y$). In this section we ask for
$\Sigma_1$ and $\Sigma_3$: do the electric fields $(E^\eta)_{\eta}$
converge weakly in $L^2(Q_M; \mathbb{C}^3)$ to $0$ as $\eta \to 0$? Is
this weak convergence in fact a strong convergence? The same questions
are considered for the magnetic fields $(H^\eta)_\eta$.
Let us point out that $E^\eta\cdot \mathds{1}_{\Sigma_\eta} \to 0$ in
$L^2(Q_M; \mathbb{C}^3)$. Indeed, the $L^2$-estimate \eqref
{eq:Analysis-high-contrast-boundedness-assumption-on-fields} can be
improved to
\begin{equation}
\label{eq:Analysis-high-contrast-improved-boundedness-of-fields}
\sup_{\eta >0} \int_G \Big(\abs{\varepsilon_\eta}\, \abs{E^\eta}^2 +
\abs{H^\eta}^2\Big) < \infty\, ,
\end{equation}
as was shown in \cite[Section 3.1]{BS10splitring}. Thus
\begin{equation}
\label{eq:E-eta-in-Sigma-eta}
\frac{\varepsilon_1}{\eta^2} \int_{\Sigma_\eta} \abs{E^\eta}^2 = \int_G
\abs{\varepsilon_\eta} \,
\abs{E^\eta}^2 \mathds{1}_{\Sigma_\eta}
\leq \int_G \Big(\abs{\varepsilon_\eta}\, \abs{E^\eta}^2 +
\abs{H^\eta}^2\Big) \leq C\, .
\end{equation}
So we have that $\norm{L^2(G)}{E^\eta \mathds{1}_{\Sigma_\eta}}^2 \leq
\eta^2 C$ which implies that $E^\eta \mathds{1}_{\Sigma_\eta} \to 0$
in $L^2(Q_M; \mathbb{C}^3)$ as $\eta \to 0$.
We recall that the two geometries of interest are
$x_3$-independent. We therefore consider two different cases: In
Section \ref{sec:parall-electr-field}, we study electric fields
$E^\eta$ that are parallel to $\operatorname{e}_3$. In Section
\ref{sec:parall-magn-field}, we study magnetic fields $H^\eta$ that
are parallel to $\operatorname{e}_3$. By linearity of the equations, superpositions
of these two cases provide the general behaviour of the material.
We will assume that the fields are $x_3$-independent. This is a strong
assumption, which can be justified for $x_3$-independent incoming
fields with a uniqueness property of solutions. In the rest of this section
the fields $E^\eta(x)$ and $H^\eta(x)$ depend only on $(x_1, x_2)$.
\smallskip \textbf{Results for high-contrast media.} In the
$E$-parallel setting, the electric fields $(E^\eta)_{\eta}$ converge
strongly to $0$ in $L^2(Q_M; \mathbb{C}^3)$, the magnetic fields
converge weakly to $0$ in $L^2(Q_M; \mathbb{C}^3)$. On the other hand,
when the magnetic fields $H^\eta$ are parallel to $\operatorname{e}_3$, we can
neither expect the electric fields nor the magnetic fields to converge
weakly to $0$ in $L^2(Q_M; \mathbb{C}^3)$.
\subsubsection{Parallel electric field}
\label{sec:parall-electr-field}
We consider here the case of parallel electric fields, i.e.,
$E^\eta(x) \coloneqq (0, 0, u^\eta(x))$ with $u^\eta = u^\eta(x_1,
x_2)$. By abuse of notation, we will consider $G$ also as a domain in
$\mathbb{R}^2$ and write $(x_1, x_2)\in G$ when $(x_1, x_2,0)\in G$; similarly
for $\Sigma_\eta$. In this setting, the magnetic field $H^\eta$ has no
third component, $H^\eta(x) = (H_1^\eta(x_1, x_2), H_2^\eta(x_1, x_2),
0)$, and Maxwell's equations \eqref{eq:time-harmonic-Maxwell-eq}
reduce to the two-dimensional system
\begin{subequations}
\label{Analysis-high-contrast-two-dimensional-system-of-Maxwell}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
- \nabla^{\perp} u^\eta &
=\phantom{-} i \omega \mu_0 (H_1^\eta, H_2^\eta) &&
\text{ in } G \, , \\
\nabla^{\perp}\cdot (H^\eta_1, H^\eta_2) & =- i \omega
\varepsilon_0 \varepsilon_\eta u^\eta && \text{ in } G \, ,
\end{empheq}
\end{subequations}
where we used the two-dimensional orthogonal gradient,
$\nabla^{\perp}u \coloneqq (- \partial_2 u, \partial_1 u)$, as well as
the two-dimensional curl, $\nabla^{\perp} \cdot (H_1, H_2) \coloneqq
-\partial_2 H_1 +
\partial_1 H_2$. The
system~\eqref{Analysis-high-contrast-two-dimensional-system-of-Maxwell}
can equivalently be written as a scalar
Helmholtz equation
\begin{equation}\label{eq:Analysis-high-contrast-E-parallel-scalar-Helmholtz}
- \Delta u^\eta = \omega ^2 \varepsilon_0 \mu_0 \varepsilon_\eta u^\eta
\quad \text{ in } G \subset \mathbb{R}^2\, .
\end{equation}
A solution of this Helmholtz equation provides the fields in the form
$E^\eta = (0 ,0 ,u^\eta)$ and $H^\eta = i (\omega \mu_0)^{-1}
(\nabla^{\perp}u^\eta ,0)$.
\begin{lemma}[Trivial limits for $E^\eta \parallel \operatorname{e}_3$]
For $\eta > 0$ small, let $\Sigma_\eta \subset G \subset \mathbb{R}^2$ be
a microscopic geometry that is given by $\Sigma_1$ or $\Sigma_3$, and let
the permittivity $\varepsilon_\eta \colon G \to \mathbb{C}$ be
defined by \eqref {eq:Analysis-high-contrast-permittivity}. Let
$E^\eta, H^\eta \colon G \to \mathbb{C}^3$ be solutions with
$E^\eta(x) = (0,0,u^\eta(x_1,x_2))$ that satisfy the estimate \eqref
{eq:Analysis-high-contrast-boundedness-assumption-on-fields}. Then
\begin{equation*}
E^\eta \to 0 \quad \text{ and } \quad H^\eta \rightharpoonup 0
\quad\text{ in } L^2(Q_M) \text{ as } \eta \to 0\, .
\end{equation*}
\end{lemma}
\begin{proof}
The $L^2$-boundedness of $E^\eta$ implies the $L^2$-boundedness of
$u^\eta$, and the $L^2$-boundedness of $H^\eta$ implies the
$L^2$-boundedness of $\nabla u^\eta$. Therefore, the sequence
$(u^\eta)_{\eta}$ is bounded in $H^1(G)$, and we find a limit
function $u \in H^1(G)$ such that $u^\eta \rightharpoonup u$ in $H^1(G)$ and
$u^\eta \to u$ in $L^2(G)$ as $\eta \to 0$.
We write
\begin{equation}\label{eq:strong-convergence-in-microstructure}
u^\eta \mathds{1}_{Q_M}
= u^\eta \mathds{1}_{Q_M\setminus \Sigma_\eta}
+ u^\eta \mathds{1}_{\Sigma_\eta}\,.
\end{equation}
The left hand side converges strongly to $u \mathds{1}_{Q_M}$. The
first term on the right hand side of \eqref
{eq:strong-convergence-in-microstructure} is the product of a
strongly $L^2(Q_M)$-convergent sequence and a weakly
$L^2(Q_M)$-convergent sequence: $\mathds{1}_{Q_M \setminus \Sigma_\eta} \rightharpoonup
\alpha$ in $L^2(Q_M)$, where $\alpha \in (0,1)$ is the volume fraction
of $Y \setminus \Sigma$. We find that the first term on the right hand side
converges in the sense of distributions to $\alpha u$. The estimate
\eqref {eq:E-eta-in-Sigma-eta} provides the strong convergence of
the second term on the right hand side of \eqref
{eq:strong-convergence-in-microstructure} to zero. The
distributional limit of \eqref
{eq:strong-convergence-in-microstructure} provides
\begin{equation}\label{eq:strong-convergence-in-microstructure-limit}
u \mathds{1}_{Q_M} = \alpha u \mathds{1}_{Q_M} + 0\,,
\end{equation}
and hence $u=0$, since $\alpha \neq 0$. We have therefore found
\begin{equation*}
E^\eta = (0, 0, u^\eta) \to 0 \quad \text{ and } \quad H^\eta =
(\nabla^{\perp} u^\eta, 0) \rightharpoonup 0 \quad \text{ in }L^2(Q_M;
\mathbb{C}^3) \text{ as } \eta \to 0\,,
\end{equation*}
which was the claim.
\end{proof}
\subsubsection{Parallel magnetic field}
\label{sec:parall-magn-field}
We now consider a magnetic field that is parallel to $\operatorname{e}_3$,
$H^\eta(x) = (0, 0, u^\eta(x_1, x_2))$, with all quantities being
$x_3$-independent. This $H$-parallel case is the interesting case for
homogenization and it has the potential to generate magnetically
active materials. It was analyzed e.g.\,in \cite {BBF09hommaxwell,
BBF15hommaxwell, BS13plasmonwaves, BF97homfibres}. In this setting,
Maxwell's equations \eqref{eq:time-harmonic-Maxwell-eq} reduce to
\begin{subequations}
\label{eq:Analysis-high-contrast-two-dimensional-Maxwell-system-parallel-H}
\begin{empheq}[left=\empheqlbrace]{alignat=2}
\nabla^{\perp} \cdot (E_1^\eta, E_2^\eta) & = \phantom{-} i \omega
\mu_0 u^\eta &&\quad \text{ in } G \, , \label{eq:Analysis-high-contrast-two-dimensional-Maxwell-system-parallel-H-1}\\
- \nabla^{\perp}u^\eta & = - i \omega \varepsilon_0
\varepsilon_\eta (E_1^\eta, E_2^\eta) &&\quad \text{ in } G\, .
\end{empheq}
\end{subequations}
System \eqref
{eq:Analysis-high-contrast-two-dimensional-Maxwell-system-parallel-H}
can equivalently be written as a scalar Helmholtz equation:
\begin{equation}
\label{eq:Analysis-high-contrast-scalar-Helmholtz-parallel-H}
- \nabla \cdot \bigg( \frac{1}{\varepsilon_\eta} \nabla u^\eta\bigg) =
\omega^2 \varepsilon_0 \mu_0 u^\eta \quad \text{ in } G\, .
\end{equation}
In \eqref {eq:Analysis-high-contrast-E-parallel-scalar-Helmholtz}, the
high-contrast coefficient is outside the differential operator, which
induces a trivial limit behaviour of solutions. Instead, \eqref
{eq:Analysis-high-contrast-scalar-Helmholtz-parallel-H} has the
high-contrast coefficient inside the differential operator, which
leads to a much richer behaviour of solutions.
The case $\Sigma = \Sigma_3$ is the metal plate (see Fig.~%
\ref{fig:the metal plate}) that was studied in
\cite{BS13plasmonwaves}. The result of~\cite{BS13plasmonwaves} is the derivation of a
limit system with nontrivial solutions. In particular, the weak limit
of $(u^\eta)_{\eta}$ can be non-trivial. Similar results are available
for metallic wires $\Sigma = \Sigma_1$ (see Fig.~\ref{fig:the metal
cylinder}); the results of \cite {BBF15hommaxwell} imply that also
in this case the weak limit of $(u^\eta)_{\eta}$ can be non-trivial.
We therefore observe that the $H$-parallel case does not allow to conclude
$H^\eta \rightharpoonup 0$ in $L^2(Q_M; \mathbb{C}^3)$. We note that
$H^\eta =(0, 0, u^\eta)\not\rightharpoonup 0$ implies, by boundedness of the
magnetic field and equation~\eqref{eq:Analysis-high-contrast-two-dimensional-Maxwell-system-parallel-H-1}, also
$E^\eta\not \rightharpoonup 0$.
\section{Finite element based multiscale approximation}
\label{sec:numerics}
In this section, we present numerical multiscale methods that are used
to study Maxwell's equations in high-contrast media from a numerical point of view.
We introduce the necessary notation for finite element discretisations and briefly
discuss the utilized approaches. Based on these methods, numerical
experiments illustrating the transmission properties of the
microstructures are presented in Section \ref{subsec:experiment}.
\subsection{Variational problem for the second order formulation}
We study time-harmonic Maxwell's
equations in their second-order formulation for the magnetic field $H$
\eqref{eq:time-harmonic-Maxwell-eq-H}. The macroscopic domain $\tilde G$ is
assumed to be bounded and we impose impedance boundary conditions on
the Lipschitz boundary $\partial \tilde G$ with the outer unit normal $n$:
\[\curl H\times n-ik_0(n \times H)\times n=g,\]
where $g\in L^2(\partial \tilde G)$ with $g\cdot n=0$ is given and
$k_0=\omega\sqrt{\varepsilon_0\mu_0}$ is the wavenumber. These
boundary conditions can be interpreted as first-order approximation to
the Silver-M\"uller radiation conditions (used for the full space
$\rz^3$); the data $g$ are usually computed from an incident wave.
For the material parameters, we choose $\mu=1$ and $\varepsilon_\eta$
as specified in \ref{eq:Analysis-high-contrast-permittivity}.
Multiplying with a test function and integrating by parts results in
the following variational formulation: Find $H^\eta\in H_{\imp}(\tilde G)$
such that
\begin{align}
\label{eq:numerics-maxwell-hfield-pde}
\int_G\varepsilon_\eta^{-1}\curl H^\eta\cdot \curl \psi
-k_0^2H^\eta\cdot \psi\, dx-ik_0\int_{\partial G}H^\eta_T\cdot
\psi_T\, d\sigma=\int_{\partial G}g\cdot \psi_T\qquad \forall
\psi\in H_{\imp}(\tilde G),
\end{align}
where $H_{\imp}(\tilde G):=\{v\in L^2(\tilde G; \cz^3)|\curl v\in L^2(\tilde G;
\cz^3),\,v_T\in L^2(\partial \tilde G)\}$ and $v_T:=v-(v\cdot n)n$ denotes
the tangential component. Existence and uniqueness of the solution to
this problem for fixed $\eta$ is shown, for instance, in \cite{Monk}.
\subsection{Traditional finite element discretisation} The standard finite
element discretisation of \eqref{eq:numerics-maxwell-hfield-pde} is a
Galerkin procedure with a finite-dimensional approximation space
$V_h\subset H_{\imp}(\tilde G)$ which consists of piecewise polynomial
functions on a (tetrahedral) mesh of $\tilde G$. In detail, we denote by
$\CT_h=\{ T_j|j\in J\}$ a partition of $\tilde G$ into tetrahedra. We assume
that $\CT_H$ is regular (i.e., no hanging nodes or edges occur), shape
regular (i.e., the minimal angle stays bounded under mesh refinement),
and that it resolves the partition into the meta-material $Q_M$ and
air $\tilde G\setminus \overline{Q}_M$. To allow for such a partition, we
implicitly assume $\tilde G$ and $Q_M$ to be Lipschitz polyhedra. Otherwise,
boundary approximations have to used which only makes the following
description more technical. We define the local mesh size
$h_j:=\diam(T_j)$ and the global mesh size $h:=\max_{j\in J}h_j$. As
conforming finite element space for $H_{\imp}(G)$, we use the lowest
order edge elements introduced by N\'ed\'elec, i.e.,
\[V_h:=\{v_h\in H_{\imp}(\tilde G)|v_h|_K(x)=a+b\times x \text{ with }a, b\in
\cz^3, \; \forall K\in \CT_h\}.\] It is well known (see \cite{Monk},
for instance), that the finite element method with this test and trial
space in \eqref{eq:numerics-maxwell-hfield-pde} yields a well-posed
discrete solution $H_h$. Furthermore, the following a priori error
estimate holds
\[\|H^\eta-H_h\|_{H(\curl)}\leq Ch (\|H^\eta\|_{H^1(\tilde G)}+\|\curl
H^\eta\|_{H^1(\tilde G)}).\] For the setting of
\eqref{eq:numerics-maxwell-hfield-pde} as discussed in this paper,
however, two major problems arise. First, due to the discontinuities
of the electric permittivity $\varepsilon^{-1}_\eta$ the necessary
regularity of $H^\eta$ is not available, see
\cite{BGL13regularitymaxwell, CDN99maxwellinterface,
CD00maxwellsingularities}. Second, even in the case of sufficient
regularity the right-hand side of the error estimate experiences a
blow-up with $\|H^\eta\|_{H^1(\tilde G)}+\|\curl H^\eta\|_{H^1(\tilde G)}\to \infty$
for $\eta\to 0$. In other words, a typical solution of
\eqref{eq:numerics-maxwell-hfield-pde} is subject to (strong)
oscillations in $\eta$ such that its derivative does not remain
bounded for the periodicity length tending to zero. As a consequence,
the error estimate has a $\eta$-dependent right-hand side of the type
$h\eta^{-1}$. Therefore, convergence of standard finite element
discretisations is only to be expected in the asymptotic regime when
$h\ll\eta$, i.e., the mesh has to resolve the oscillations in the PDE
coefficients. As discussed in the introduction, this can become
prohibitively expensive.
\subsection{Heterogeneous Multiscale Method} As a remedy to these
limitations of the standard finite element method, we consider a
specific multiscale method. The idea is to extract macroscopic
properties of the solution with $\eta$-independent complexity or
computational effort, respectively. The basic idea directly
comes up from the effective equation
\eqref{eq:Analysis-high-contrast-effective-eq-H}: Since this effective
equation is independent of $\eta$, it can be discretised on a rather
coarse mesh $\CT_h$ without the need to resolve the $\eta$-scale,
i.e., we can have $h>\eta$. This results in an approximation of the
homogenized solution $\hat{H}$, which contains important macroscopic
information of $H^\eta$. However, for the discretisation of the
homogenized equation, the effective material parameters
$\hat{\varepsilon}$ and $\hat{\mu}$ need to be known, at least at the
(macroscopic) quadrature points. This can be achieved by introducing
another, again $\eta$-independent, mesh $\CT_{h_Y}=\{S_l|l\in I\}$ of
the unit cube $Y$ with maximal mesh size $h_Y=\max_{l\in
I}\diam(S_l)$. We assume that $\CT_{h_Y}$ is regular and shape
regular and resolves the partition of $Y$ into $\Sigma$ and
$Y\setminus \overline{\Sigma}$. Furthermore, $\CT_{h_Y}$ has to be
periodic in the sense that it can be wrapped to a regular
triangulation of the torus, i.e., no hanging nodes or edges over the
periodic boundary. Note that $h_Y$ denotes the mesh size of the
triangulation of the unit cube. Thus, it is in no way related to
$\eta$ and can be of the same order as $h$. Based on this mesh, the
cell problems occurring in the definition of $\hat{\varepsilon}$ and
$\hat{\mu}$ can be discretised with standard (Lagrange and
N\'ed\'elec) finite element spaces. For details we refer to
\cite{Ver17hmmmaxwell}. All in all, we can now compute the
homogenized solution $\hat{H}$ of
\eqref{eq:Analysis-high-contrast-effective-eq-H} as follows:
1. Compute discrete solutions of the cell problems (see
\cite{BBF15hommaxwell} or\cite{Ver17hmmmaxwell}) using the mesh
$\CT_{h_Y}$ and the associated standard finite element
spaces. 2. Compute the effective parameters $\hat{\varepsilon}$ (or
$\widehat{\varepsilon^{-1}}$) and $\hat{\mu}$ approximatively with the
discrete cell problems solutions. 3. Compute the discrete homogenized
solution of \eqref{eq:Analysis-high-contrast-effective-eq-H} with the
approximated effective coefficients and using the mesh $\CT_h$ with
the associated finite element space $V_h$ as introduced above.
This (naive) discretisation scheme for the effective equation
\eqref{eq:Analysis-high-contrast-effective-eq-H} in fact can be
interpreted as a specification of the Heterogeneous Mutliscale Method
(HMM) in the perfectly periodic case. The Finite Element
Heterogeneous Multiscale Method, introduced by E and Enguist
\cite{EE03hmm, EE05hmm}, sets up a macroscopic sesquilinear form to
compute the HMM solution $H_h$, which is an approximation of the
homogenized solution $\hat{H}$. The macroscopic sesquilinear form is
very similar to the effective sesquilinear form associated with the
left-hand side of \eqref{eq:Analysis-high-contrast-effective-eq-H},
but the effective material parameters are not computed a priori.
Instead local variants of the cell problems are set up on
$\eta$-scaled cubes $Y_j^\eta=\eta Y+x_j$ around macroscopic
quadrature points $x_j$. We can still use the mesh $\CT_{h_Y}$ of the
unit cube $Y$ and transform it to a partition $\CT_{h_Y}(Y_j^\eta)$ of
the scaled unit cell $Y_j^\eta$. Similarly, also the finite element
spaces associated with $\CT_{h_Y}$ can be transferred to spaces on
$\CT_{h_Y}(Y_j^\eta)$ using a suitable affine mapping. The finescale
computations result in so called local reconstructions, which consist
of the macroscopic basis functions and the corresponding (discrete)
cell problem solutions. Averages (over $Y_j^\eta$) of these local
reconstructions then enter the macroscopic sesquilinear form. A
detailed definition of the HMM for Maxwell's equations in
high-contrast media is presented in \cite{Ver17hmmmaxwell}, where also
the connection to analytical homogenization as well as the possibility
to treat more general than purely periodic problems are discussed. We
only want to emphasize one important feature of the HMM in
\cite{Ver17hmmmaxwell}: Apart from the (macroscopic) approximation
$H_h$, discrete correctors $H_{h_Y, 1}$, $H_{h_Y, 2}$, and $H_{h_Y,
3}$ can be determined from the discrete cell problems (in a second
post-processing step). Via these correctors, we can define the zeroth
order $L^2$-approximation $H^0_{\HMM}:=H_h+\nabla_y H_{h_Y, 2}(\cdot,
\frac{\cdot}{\delta})+H_{h_Y,3}(\cdot, \frac{\cdot}{\delta})$, which
corresponds to the first term of an asymptotic expansion and is used
to approximate the true solution $H^\eta$. We again refer to
\cite{Ver17hmmmaxwell} for details and note that it has been observed
in several numerical examples that these correctors are a vital part
of the HMM-approximation, see \cite{HOV15maxwellHMM, GHV17lodcurl,
OV16hmmhelmholtz, Ver17hmmmaxwell}. We close by remarking that in
Section \ref{subsec:experiment} below, we extend the described HMM of
\cite{Ver17hmmmaxwell} to general microstructures although the
validity of the homogenized models in these cases is not shown so far,
see the discussion in Sections \ref{sec:homogenization} and
\ref{sec:high-contrast-media}.
\section{Numerical study of transmission properties for high-contrast
inclusions}
\label{subsec:experiment}
In this section, we numerically study the transmission properties in
the case of high-contrast for the three micro-geometries: the metal
cylinder, the metal plate, and the air cylinder. Since the aim of
this paper is a better understanding of the different microstructures
and their effect, we focus on the qualitative behaviour rather than
explicit convergence rates. The implementation was done with the
module {\sffamily dune-gdt} \cite{wwwdunegdt} of the DUNE software
framework \cite{BB+08dune1, BB+08dune2}.
\smallskip \textbf{Setting.} We consider Maxwell's equations in the
second-order formulation for the $H$-field
\eqref{eq:numerics-maxwell-hfield-pde} with a high-contrast medium as
defined in \eqref{eq:Analysis-high-contrast-permittivity}. It remains
to specify the macroscopic geometry, the boundary date $g$, the
material parameter $\varepsilon_1$, and the frequency. We use a slab-like
macroscopic geometry similar to Section \ref{sec:geometry}, but we
truncate $G$ also in the $x_1$-direction to have a finite
computational domain $\tilde G$, as described in the previous section.
We choose $\tilde G=(0,1)^3$ with the meta-material located in
$Q_M=\{x\in G|0.25\leq x_1\leq 0.75\}$. Note that $Q_M$ is translated
in $x_1$-direction compared to Section \ref{sec:geometry}, but this
does not influence the qualitative results of the analysis. As in
Section \ref{sec:geometry}, we assume that an incident wave
$H_{\mathrm{inc}}$ from the right travels along the $x_1$-axis to the
left, i.e., $H_{\mathrm{inc}}=\exp(-ikx_1)p$ with a normalized
polarization vector $p\perp \Ve_1$. This incident wave is used to
compute the boundary data $g$ as $g=\curl H_{\mathrm{inc}}\times
n-ik_0 n\times (H_{\mathrm{inc}}\times n)$. We choose the inverse
permittivity as $\varepsilon_1^{-1}=1.0-0.01i$ and note that
$\varepsilon_1$ is only slightly dissipative. In all experiments, we
choose the same wavenumber $k_0=12$ and the periodicity parameter
$\eta=1/8$.
As explained in the previous section, we want to use the Heterogeneous
Multiscale Method to obtain good approximations with reasonable
computational effort. We use the mesh sizes $h_y=h=\sqrt{3}\cdot
1/16$ and compute the macroscopic HMM-approximation $H_h$ as well as
the zeroth order approximation $H_{\HMM}^0$, which also utilizes
information of the discrete correctors. To demonstrate the validity
of the HMM, we use two different reference solutions. First, the
homogenized reference solution $\hat{H}$ is computed as solution to
\eqref{eq:Analysis-high-contrast-effective-eq-H} on a mesh with size
$h=\sqrt{3}\cdot 1/48$, where the effective material parameters are
calculated approximatively using a discretisation of the unit cube
with mesh size $h_Y=\sqrt{3}\cdot 1/24$. Second, the (true) reference
solution $H^\eta$ is computed as direct finite element discretisation
of \eqref{eq:numerics-maxwell-hfield-pde} on a fine grid with
$h_{\mathrm{ref}}=\sqrt{3}\cdot 1/64$.
\begin{table}
\caption{Summary of analytical predictions of the transmission properties
and references to numerical results. The first row provides the geometry.
The second row indicates possible transmission polarizations (of $H$) according
to the theory of perfect conductors of Section
\ref {sec:Calculation-of-transmission-and-reflection-coefficient}.
The third row indicates the possibility of transmission based on
Section \ref {sec:high-contrast-media}: We mention cases in which we
cannot derive weak convergence to $0$. An entry \enquote{-} indicates that no analytical
result can be applied. The last row provides the reference
to the visualization of the numerical calculation for high-contrast media.}
\label{tab:numsummary}
\centering
\begin{tabular}{@{}ccccc@{}}
\toprule
geometry
&metal cylinder $\Sigma_1$&metal cylinder $\Sigma_2$&metal plate $\Sigma_3$
&air cyl. $\Sigma_4$\\
\midrule
transmission (PC)& $\Ve_3$-polarized&$\Ve_2$ and $\Ve_3$-polarized &
$\Ve_3$-polarized&no\\
\midrule
nontriv. limit (HC)& $\Ve_3$-polarized&- &$\Ve_3$-polarized&-\\
\midrule
numerical example& Fig.~\ref{fig:Sigma1-homrefsol}& Fig.~\ref{fig:Sigma2-homrefsol}
& Fig.~\ref{fig:Sigma3-homrefsol} & Fig.~\ref{fig:Sigma4-refsol}\\
\bottomrule
\end{tabular}
\end{table}
\smallskip \textbf{Main results.} Before we discuss the examples in
detail, we present an overview of the results. The qualitative
transmission properties of the meta-material are in good agreement
with the theory of Section
\ref{sec:Calculation-of-transmission-and-reflection-coefficient},
although the numerical examples consider high-contrast media instead
of perfect conductors. The predictions and the corresponding
numerical examples are summarized in Table \ref{tab:numsummary}. In
contrast to perfect conductors, the high-contrast medium leads to
rather high intensities and amplitudes of the $H$-field inside the
inclusions $\Sigma_\eta$. Depending on the chosen wavenumber,
Mie-resonances inside the inclusions can occur for high-contrast
media; see Section \ref{sec:high-contrast-media} and
\cite{BBF15hommaxwell, Ver17hmmmaxwell}. Our numerical experiments
also show that the HMM yields (qualitatively) good approximations,
although the validity of the underlying effective models is not proved
for the studied geometries.
\subsection{Metal cuboids $\tilde \Sigma_1$ and $\tilde \Sigma_2$}
\begin{figure}[t]
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm, clip=true, keepaspectratio=false]{"hom_sol_xpolz_zslice05_new"}%
\hspace{1.9cm}%
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm, clip=true, keepaspectratio=false]{"hom_sol_xpoly_yslice0545_new"}
\begin{tikzpicture}[scale=1.5, overlay, xshift=30ex, yshift=6.5ex]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (A1) at (-0.25, -0.5, 0.25);
\coordinate (B1) at (0.25,-0.5,0.25);
\coordinate (C1) at (0.25,0.5,0.25);
\coordinate (D1) at (-0.25, 0.5, 0.25);
\coordinate (E1) at (-0.25, 0.5, -0.25);
\coordinate (F1) at (0.25, 0.5, -0.25);
\coordinate (G1) at (0.25, -0.5, -0.25);
\coordinate (H1) at (-0.25, -0.5, -0.25);
\fill[gray!20!white, opacity=.5] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!20!white, opacity=.5] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!20!white, opacity=.5] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!20!white, opacity=.5] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!75!white] (A1)--(B1)--(C1)--(D1)--cycle;
\filldraw[gray!75!white] (E1)--(F1)--(G1)--(H1)--cycle;
\filldraw[gray!75!white] (D1)--(E1)--(H1)--(A1)--cycle;
\filldraw[gray!75!white] (B1)--(C1)--(F1)--(G1)--cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (A1)--(B1)--(C1)--(D1)--cycle (E1)--(D1) (F1)--(C1) (G1)--(B1);
\draw[] (E1)--(F1)--(G1);
\draw[densely dashed] (E1)--(H1) (H1)--(G1) (H1)--(A1);
\draw[->] ( -.4, -.5, .5)--( 0, -.5, .5 );
\draw[->] (-.4, -.5, .5) -- (-.4, -.1, .5);
\draw[->] (-.4, -.5, .5) -- (-.4, -.5, .2);
\node[] at (0,0,0) {$\tilde \Sigma_1$};
\node[] at (-0, -.6, .5){$\operatorname{e}_1$};
\node[] at (-.28, -.1, .5){$\operatorname{e}_3$};
\node[] at (-.4, -.6, .-.1){$\operatorname{e}_2$};
\end{tikzpicture}
\caption{Metal cuboid $\tilde \Sigma_1$, the magnitude of $\Re(\hat{H})$ is
plotted. Left: The $H$-field is $\operatorname{e}_3$-polarized and the plot shows
values in the plane $x_3=0.5$. The analysis of both, (PC) and (HC)
yields: transmission is possible. Right: The $H$-field is
$\operatorname{e}_2$-polarized and the plot shows values in the plane $x_2 =
0.545$. Since the $H$-field is not parallel to $\operatorname{e}_3$, the analysis
of (PC) and (HC) predicts that no transmission is possible. Inlet in
the middle: Microstructure in the unit cube.}
\label{fig:Sigma1-homrefsol}
\end{figure}
\begin{figure}
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm, clip=true, keepaspectratio=false]{"ref_sol_xpolz_zslice05_new"}%
\hspace{1.9cm}%
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm, clip=true, keepaspectratio=false]{"correc_xpolz_zslice05_new"}
\begin{tikzpicture}[scale=1.5, overlay, xshift=30ex, yshift=6.5ex]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (A1) at (-0.25, -0.5, 0.25);
\coordinate (B1) at (0.25,-0.5,0.25);
\coordinate (C1) at (0.25,0.5,0.25);
\coordinate (D1) at (-0.25, 0.5, 0.25);
\coordinate (E1) at (-0.25, 0.5, -0.25);
\coordinate (F1) at (0.25, 0.5, -0.25);
\coordinate (G1) at (0.25, -0.5, -0.25);
\coordinate (H1) at (-0.25, -0.5, -0.25);
\fill[gray!20!white, opacity=.5] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!20!white, opacity=.5] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!20!white, opacity=.5] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!20!white, opacity=.5] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!75!white] (A1)--(B1)--(C1)--(D1)--cycle;
\filldraw[gray!75!white] (E1)--(F1)--(G1)--(H1)--cycle;
\filldraw[gray!75!white] (D1)--(E1)--(H1)--(A1)--cycle;
\filldraw[gray!75!white] (B1)--(C1)--(F1)--(G1)--cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (A1)--(B1)--(C1)--(D1)--cycle (E1)--(D1) (F1)--(C1) (G1)--(B1);
\draw[] (E1)--(F1)--(G1);
\draw[densely dashed] (E1)--(H1) (H1)--(G1) (H1)--(A1);
\draw[->] ( -.4, -.5, .5)--( 0, -.5, .5 );
\draw[->] (-.4, -.5, .5) -- (-.4, -.1, .5);
\draw[->] (-.4, -.5, .5) -- (-.4, -.5, .2);
\node[] at (0,-0.35,0) {$\tilde \Sigma_1$};
\node[] at (-0, -.6, .5){$\operatorname{e}_1$};
\node[] at (-.28, -.1, .5){$\operatorname{e}_3$};
\node[] at (-.4, -.6, .-.1){$\operatorname{e}_2$};
\filldraw[red, opacity=.5] (-0.5, 0, -0.5) -- (0.5, 0, -0.5) --(0.5, 0, 0.5) --(-0.5, 0, 0.5)--cycle;
\end{tikzpicture}
\caption{Test of numerical schemes for the metal cuboid
$\tilde \Sigma_1$. We consider an $\operatorname{e}_3$-polarized incoming $H$-field and
plot the solution in the plane $x_3=0.5$; the colors indicate the
magnitude of the reference solution $\Re(H^\eta)$ (left) and the
zeroth order approximation $\Re(H^0_{\HMM})$ (right). Inlet in the
center: Microsctructure in the unit cube with visualization plane in
red.}
\label{fig:Sigma1-HMM}
\end{figure}
\begin{figure}
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm,
clip=true,
keepaspectratio=false]{"hom_sol_cylinderx_yslice0545_new"}%
\hspace{1.9cm}%
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm,
clip=true,
keepaspectratio=false]{"ref_sol_cylinderx_yslice0545_new"}
\begin{tikzpicture}[scale=1.5, overlay, xshift=30ex, yshift=6.5ex]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (A1) at (-0.5, -0.25, 0.25);
\coordinate (B1) at (0.5,-0.25,0.25);
\coordinate (C1) at (0.5,0.25,0.25);
\coordinate (D1) at (-0.5, 0.25, 0.25);
\coordinate (E1) at (-0.5, 0.25, -0.25);
\coordinate (F1) at (0.5, 0.25, -0.25);
\coordinate (G1) at (0.5, -0.25, -0.25);
\coordinate (H1) at (-0.5, -0.25, -0.25);
\fill[gray!20!white, opacity=.5] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!20!white, opacity=.5] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!20!white, opacity=.5] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!20!white, opacity=.5] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!75!white] (A1)--(B1)--(C1)--(D1)--cycle;
\filldraw[gray!75!white] (E1)--(F1)--(G1)--(H1)--cycle;
\filldraw[gray!75!white] (D1)--(E1)--(H1)--(A1)--cycle;
\filldraw[gray!75!white] (B1)--(C1)--(F1)--(G1)--cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (A1)--(B1)--(C1)--(D1)--cycle (E1)--(D1) (F1)--(C1) (G1)--(B1);
\draw[] (E1)--(F1)--(G1);
\draw[densely dashed] (E1)--(H1) (H1)--(G1) (H1)--(A1);
\draw[->] ( -.4, -.5, .5)--( 0, -.5, .5 );
\draw[->] (-.4, -.5, .5) -- (-.4, -.2, .5);
\draw[->] (-.4, -.5, .5) -- (-.4, -.5, .2);
\node[] at (0,-0,0) {$\tilde \Sigma_2$};
\node[] at (-0, -.6, .5){$\operatorname{e}_1$};
\node[] at (-.2, -.2, .5){$\operatorname{e}_3$};
\node[] at (-.4, -.6, .-.1){$\operatorname{e}_2$};
\filldraw[red, opacity=.5] (-0.5, -0.5, -0) -- (0.5, -0.5, -0) --(0.5, 0.5, 0) --(-0.5, 0.5, 0)--cycle;
\end{tikzpicture}
\caption{Metal cuboid $\tilde \Sigma_2$. We study an $\operatorname{e}_3$-polarized
incident $H$-field and plot the magnitude of $\Re(\hat{H})$ (left) and
$\Re(H^\eta)$ (right) in the plane $x_2=0.545$. The analysis (PC)
predicts transmission in this case, the analysis (HC) does not
exclude transmission. Middle: Microstructure in the unit cube with
visualization plane in red.}
\label{fig:Sigma2-homrefsol}
\end{figure}
Instead of metal cylinders with circular base we study metal cuboids
with square base, so that we do not have to deal with boundary
approximations in our numerical method. We choose $\tilde
\Sigma_1=(0.25, 0.75)^2\times (0,1)$ and $\tilde \Sigma_2=(0,1)\times
(0.25, 0.75)^2$. Note that this choice influences the value of
$\gamma$, but not the other results of Sections
\ref{sec:metal-cylinder} and \ref{sec:metal-cylinder-case}. Due to
the symmetry of the microstructure, the effective material parameters
are diagonal matrices with $a_{11}=a_{22}$, but with a different value
$a_{33}$; see the analytical computations in Section
\ref{sec:metal-cylinder}. Up to numerical errors, we obtain the same
structure for the computed approximative effective parameters.
Comparing the homogenized reference solution $\hat{H}$ for $\operatorname{e}_2$- and
$\operatorname{e}_3$-polarized incoming waves for $\tilde \Sigma_1$ in
Fig.~\ref{fig:Sigma1-homrefsol}, we observe that the $\operatorname{e}_3$-polarized
wave is transmitted almost undisturbed through the meta-material. For
the $\operatorname{e}_2$-polarization, however, the field intensity in $Q_L\coloneqq
\set{x \in G \given x\leq 0.25}$ is very low, corresponding to small
transmission factors. This matches the analytical predictions of
Section \ref{sec:metal-cylinder}, which yields transmission only for
$\operatorname{e}_3$-parallel $H$-fields. The same effect is predicted for
high-contrast media by the analysis of Section
\ref{sec:parall-electr-field}.
The HMM can reproduce the behaviour of the homogenized and of the
heterogeneous solution. For the comparison, we only consider the
$\operatorname{e}_3$-polarized incoming wave in Fig.~\ref{fig:Sigma1-HMM} and
compare the zeroth order approximation $H_{\HMM}^0$ (right) to the
(true) reference solution $H^\eta$ (left). Errors are still visible,
but the qualitative agreement is good, even for the coarse mesh size
of $h=h_Y=\sqrt{3}\cdot 1/16$ chosen for the HMM. In particular, the
rather cheaply computable zeroth order approximation $H^0_{\HMM}$ can
capture most of the important features of the \emph{true} solution,
even for inclusions of high-contrast. This clearly underlines the
potential of the HMM. Moreover, Fig.~\ref{fig:Sigma1-HMM} underlines
the specific behaviour of $H^\eta$ in the inclusions $\tilde
\Sigma_\eta$ for high-contrast media. As analyzed in Section
\ref{sec:parall-magn-field}, $H^\eta$ cannot be expected to vanish in
the inclusions in the limit $\eta\to 0$ due to possible resonances;
see \cite{BBF09hommaxwell}. We observe rather high field intensities
in the inclusions; see also \cite{Ver17hmmmaxwell} for a slightly
different inclusion geometry.
\smallskip We also study the rotated metal cuboid $\tilde \Sigma_2$. In
correspondence to the analysis of Section
\ref{sec:metal-cylinder-case}, we observe transmission for an
$\operatorname{e}_3$-polarized incident wave; see Fig.~\ref{fig:Sigma2-homrefsol}.
Note that the homogenized reference solution looks different to
$\tilde \Sigma_1$ because of the rotation of the geometry, which is also
reflected in the different structure and values of the reflection and
transmission coefficients. The (true) reference solution in Fig.~%
\ref{fig:Sigma3-homrefsol} shows the high field intensities in the
metal cuboids induced by the high-contrast permittivity.
\subsection{Metal plate $\Sigma_3$}
\begin{figure}
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm,
clip=true,
keepaspectratio=false]{"hom_sol_yplate_xpolz_zslice05_new"}%
\hspace{1.9cm}%
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm,
clip=true,
keepaspectratio=false]{"hom_sol_yplate_xpoly_zslice05_new"}
\begin{tikzpicture}[scale=1.5, overlay, xshift=30ex, yshift=6.5ex]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (M1) at (-0.5, -0.5, 0.15);
\coordinate (M2) at (-0.5, -0.5, -0.15);
\coordinate (M3) at (-0.5, 0.5, -0.15);
\coordinate (M4) at (-0.5, 0.5, 0.15);
\coordinate (M5) at (0.5, 0.5, 0.15);
\coordinate (M6) at (0.5,0.5,-0.15);
\coordinate (M7) at (0.5, -0.5, -0.15);
\coordinate (M8) at (0.5, -0.5, 0.15);
\fill[gray!20!white, opacity=.5] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!20!white, opacity=.5] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!20!white, opacity=.5] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!20!white, opacity=.5] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!75!white, opacity=.9] (M4) -- (M5) -- (M6) -- (M3) -- cycle;
\fill[gray!75!white, opacity=.9] (M1) -- (M4) -- (M5) -- (M8) -- cycle;
\fill[gray!75!white, opacity=.9] (M2) -- (M3) -- (M6) -- (M7) -- cycle;
\fill[gray!75!white, opacity=.9] (M1) -- (M2) -- (M7) -- (M8) -- cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[dashed] (M1) -- (M4) (M2) -- (M3) ;
\draw[] (M4) -- (M5) -- (M6) -- (M3) -- cycle;
\draw[dashed] (M5) -- (M8) (M7) -- (M6);
\draw[dashed] (M1) -- (M8) (M7) -- (M2);
\draw[->] ( -.4, -.5, .5)--( 0, -.5, .5 );
\draw[->] (-.4, -.5, .5) -- (-.4, -.1, .5);
\draw[->] (-.4, -.5, .5) -- (-.4, -.5, .2);
\node[] at (0,0,0) {$\Sigma_1$};
\node[] at (-0, -.6, .5){$\operatorname{e}_1$};
\node[] at (-.28, -.1, .5){$\operatorname{e}_3$};
\node[] at (-.4, -.6, .-.1){$\operatorname{e}_2$};
\end{tikzpicture}
\caption{Metal plate $\Sigma_3$. The colors indicate the magnitude
of $\Re(\hat{H})$ in the plane $x_3=0.5$. Left: The $H$-field is
$\operatorname{e}_3$-polarized. The analysis (PC) predicts transmission, the
analysis (HC) cannot exclude transmission. Right: The $H$-field is
$\operatorname{e}_2$-polarized. The analysis (PC and HC) predicts that no
transmission is possible.}
\label{fig:Sigma3-homrefsol}
\end{figure}
As in Section \ref{sec:metal-plate}, we choose a metal plate
perpendicular to $\operatorname{e}_2$ of width $0.5$, i.e.\
$\Sigma_3=(0,1)\times (0.25, 0.75)\times (0,1)$. Discretising the
cell problems with mesh size $h_Y=\sqrt{3}\cdot 1/24$, we obtain---up
to numerical errors---the effective material parameters as diagonal
matrices with
\begin{align*}
\widehat{\varepsilon^{-1}}&\approx \diag(10^{-4}, 0.5, 10^{-4})\,,\\
\Re\hat{\mu}&\approx\diag (0.228303, -0.044672, 0.228303)\,.
\end{align*}
Although we consider high-contrast media, this correspond
astonishingly well to the analytical results for perfect conductors of
Section \ref{sec:metal-plate}: The structure of the matrices agrees
and the non-zero value of $\widehat{\varepsilon^{-1}}=|Y\setminus
\overline{\Sigma}|$ is as expected from the theory of perfect
conductors. Due to the contributions of the inclusions, the values of
$\mu_{\hom}$ are different from the case of perfect conductors.
Section \ref{sec:metal-plate} shows that, for perfect conductors, only
an $H$-field polarized in $\operatorname{e}_3$-direction can be transmitted through
the meta-material. Our numerical experiments allow a similar
observation for high-contrast media in Fig.~%
\ref{fig:Sigma3-homrefsol}: The homogenized reference solution only
shows a non-negligible intensity in the domain $Q_L=\{x\in G|x_3\leq
0.25\}$ left of the scatterer if the incident wave is polarized in
$\operatorname{e}_3$ direction. Note that we have some reflections from the
boundary in Fig.~\ref{fig:Sigma3-homrefsol} since we do not use
perfectly matched layers as boundary conditions. The observed
transmission properties for high-contrast media are in accordance with
the theory in Section \ref{sec:high-contrast-media}: For an
$\operatorname{e}_3$-polarized $H$-field as in the left figure, we cannot expect a
(weak) convergence to zero. This corresponds to the observed
non-trivial transmission. By contrast, in the right figure, the
$H$-field is $\operatorname{e}_2$-polarized and no transmission can be
observed. This corresponds to the analysis of Section
\ref{sec:parall-electr-field}, which shows that $H^\eta$ converges to
zero, weakly in $L^2(Q_M)$.
\subsection{Air cuboid $\tilde \Sigma_4$}
\begin{figure}
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm,
clip=true,
keepaspectratio=false]{"ref_sol_cylx_xpolz_zslice0545_new"}
\hspace{1.9cm}%
\includegraphics[width=0.44\textwidth, trim= 75mm 33mm 51mm 32mm,
clip=true,
keepaspectratio=false]{"ref_sol_cylz_xpolz_zslice05_new"}%
\begin{tikzpicture}[scale=1.5, overlay, xshift=-35ex, yshift=4.7ex]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (A1) at (-0.5, -0.25, 0.25);
\coordinate (B1) at (0.5,-0.25,0.25);
\coordinate (C1) at (0.5,0.25,0.25);
\coordinate (D1) at (-0.5, 0.25, 0.25);
\coordinate (E1) at (-0.5, 0.25, -0.25);
\coordinate (F1) at (0.5, 0.25, -0.25);
\coordinate (G1) at (0.5, -0.25, -0.25);
\coordinate (H1) at (-0.5, -0.25, -0.25);
\fill[gray!75!white, opacity=.9] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!75!white, opacity=.9] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!75!white, opacity=.9] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!75!white, opacity=.9] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!25!white] (A1)--(B1)--(C1)--(D1)--cycle;
\filldraw[gray!25!white] (E1)--(F1)--(G1)--(H1)--cycle;
\filldraw[gray!25!white] (D1)--(E1)--(H1)--(A1)--cycle;
\filldraw[gray!25!white] (B1)--(C1)--(F1)--(G1)--cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (A1)--(B1)--(C1)--(D1)--cycle (E1)--(D1) (F1)--(C1) (G1)--(B1);
\draw[] (E1)--(F1)--(G1);
\draw[densely dashed] (E1)--(H1) (H1)--(G1) (H1)--(A1);
\draw[->] ( -.4, -.5, .5)--( 0, -.5, .5 );
\draw[->] (-.4, -.5, .5) -- (-.4, -.1, .5);
\draw[->] (-.4, -.5, .5) -- (-.4, -.5, .2);
\node[] at (0,0,0) {$\tilde \Sigma_4$};
\node[] at (-0, -.6, .5){$\operatorname{e}_1$};
\node[] at (-.28, -.1, .5){$\operatorname{e}_3$};
\node[] at (-.4, -.6, .-.1){$\operatorname{e}_2$};
\filldraw[red, opacity=.5] (-0.5, 0, -0.5) -- (0.5, 0, -0.5) --(0.5, 0, 0.5) --(-0.5, 0, 0.5)--cycle;
\end{tikzpicture}
\begin{tikzpicture}[scale=1.5, overlay, xshift=60ex, yshift=6.7ex]
\coordinate (A) at (-0.5, -0.5, 0.5);
\coordinate (B) at (0.5,-0.5,0.5);
\coordinate (C) at (0.5,0.5,0.5);
\coordinate (D) at (-0.5, 0.5, 0.5);
\coordinate (E) at (-0.5, 0.5, -0.5);
\coordinate (F) at (0.5, 0.5, -0.5);
\coordinate (G) at (0.5, -0.5, -0.5);
\coordinate (H) at (-0.5, -0.5, -0.5);
\coordinate (A1) at (-0.25, -0.5, 0.25);
\coordinate (B1) at (0.25,-0.5,0.25);
\coordinate (C1) at (0.25,0.5,0.25);
\coordinate (D1) at (-0.25, 0.5, 0.25);
\coordinate (E1) at (-0.25, 0.5, -0.25);
\coordinate (F1) at (0.25, 0.5, -0.25);
\coordinate (G1) at (0.25, -0.5, -0.25);
\coordinate (H1) at (-0.25, -0.5, -0.25);
\fill[gray!75!white, opacity=.9] (A) -- (B) -- (C) -- (D) -- cycle;
\fill[gray!75!white, opacity=.9] (E) -- (F) -- (G) --(H) -- cycle;
\fill[gray!75!white, opacity=.9] (D) -- (E) -- (H) -- (A)-- cycle;
\fill[gray!75!white, opacity=.9] (B) -- (C) -- (F) -- (G) -- cycle;
\fill[gray!25!white] (A1)--(B1)--(C1)--(D1)--cycle;
\filldraw[gray!25!white] (E1)--(F1)--(G1)--(H1)--cycle;
\filldraw[gray!25!white] (D1)--(E1)--(H1)--(A1)--cycle;
\filldraw[gray!25!white] (B1)--(C1)--(F1)--(G1)--cycle;
\draw[] (A) -- (B) -- (C) -- (D) --cycle (E) -- (D) (F) -- (C) (G) -- (B);
\draw[] (E) -- (F) -- (G) ;
\draw[densely dashed] (E) -- (H) (H) -- (G) (H) -- (A);
\draw[] (A1)--(B1)--(C1)--(D1)--cycle (E1)--(D1) (F1)--(C1) (G1)--(B1);
\draw[] (E1)--(F1)--(G1);
\draw[densely dashed] (E1)--(H1) (H1)--(G1) (H1)--(A1);
\draw[->] ( -.4, -.5, .5)--( 0, -.5, .5 );
\draw[->] (-.4, -.5, .5) -- (-.4, -.1, .5);
\draw[->] (-.4, -.5, .5) -- (-.4, -.5, .2);
\node[] at (0,-0.35,0) {$\tilde{\Sigma}_4$};
\node[] at (-0, -.6, .5){$\operatorname{e}_1$};
\node[] at (-.28, -.1, .5){$\operatorname{e}_3$};
\node[] at (-.4, -.6, .-.1){$\operatorname{e}_2$};
\filldraw[red, opacity=.5] (-0.5, 0, -0.5) -- (0.5, 0, -0.5) --(0.5, 0, 0.5) --(-0.5, 0, 0.5)--cycle;
\end{tikzpicture}
\caption{Metal block with holes. Left: The structure $\tilde
\Sigma_4$, we plot the magnitude of $\Re(H^\eta)$ in the plane
$x_3=0.545$ for $\operatorname{e}_3$-polarized incoming $H$-field. The analysis
(PC) predicts no transmission, the analysis (HC) cannot exclude
transmission. Right: A geometry in which the cylinders $\tilde
\Sigma_4$ are rotated in $\operatorname{e}_3$-direction. We plot the magnitude of
$\Re(H^\eta)$ in the plane $x_3=0.5$ for $\operatorname{e}_3$-polarized incoming
$H$-field. Small pictures show the microstructures in the unit cube
and the visualization planes in red.}
\label{fig:Sigma4-refsol}
\end{figure}
As with the metal cylinder, we equip the air cylinder of Section
\ref{sec:air-cylinder} with a square base in order to have a
geometry-fitting mesh. To be precise, we define the microstructure
$\tilde \Sigma_4=(0,1)^3\setminus((0,1)\times(0.25, 0.75)^2)$. The
effective permittivity $\widehat{\varepsilon^{-1}}$ vanishes almost
identically for this setting; numerically we obtain only entries of
order $10^{-5}$ for a discretisation of the corresponding cell problem
with mesh size $h_Y=\sqrt{3}\cdot 1/24$. As discussed in Section
\ref{sec:air-cylinder}, no transmission through this meta-material is
expected for the high conductors. We observe the same for
high-contrast media in Fig.~\ref{fig:Sigma4-refsol}: The (true)
reference solution (almost) vanishes in the left part $Q_L$ in all
situations. Here, we only depict $\operatorname{e}_3$-polarized incident waves, once
for $\tilde \Sigma_4$ as described and once for the rotated air cuboid
with main axis in $\operatorname{e}_3$-direction (this is the setting of Section
\ref{sec:metal-cylinder} with interchanged roles of metal and air).
Note that inside the microstructure, high intensities and amplitudes
of the $H^\eta$-field occur due to resonances in the high-contrast
medium.
\section*{Conclusion}
We analyzed the transmission properties of meta-materials consisting
of perfect conductors or high-contrast materials. Depending on the
geometry of the microstructure, certain entries in the effective
material parameters vanish, which induces that also certain components
of the solution vanish. This influences the transmission properties
of the material. Transmission is possible only for certain
polarizations of the incoming wave. For perfect conductors, we
derived closed formulas for the reflection and transmission
coefficients. Using the Heterogeneous Multiscale Method, the
homogenized solution as well as some features of the exact solution
can be approximated on rather coarse meshes and, in particular, with a
cost that is independent of the periodicity length. Our numerical
experiments of three representative geometries with high-contrast
materials confirm the theoretical predictions of their transmission
properties.
\bibliographystyle{abbrv}
|
\section{Algorithms for almost all induced patterns}
In this section, we prove a result that is similar in spirit to
Theorem~{\ref{thm:np-universal-full}} in \cite{NP85} which states that the
time complexity of induced subgraph isomorphism problem for $K_k$
upper bounds that of any $k$-vertex pattern graph. We show that the
circuit complexity of $\ghom{K_k - e}$ upper bounds the time
complexity of the induced subgraph isomorphism problem for all
$k$-vertex pattern graphs $H$ except $K_k$ and $I_k$. The algorithms
obtained from this statement can be obtained from known
results. However, we believe that restating these upper-bounds in
terms of circuits for $K_k - e$ homomorphism polynomials may give new
insights to improve these algorithms.
The key idea is that an efficient construction of homomorphism
polynomial for $K_k - e$ enables efficient construction of
homomorphism polynomials for all smaller graphs. First, we prove the
following technical result.
\begin{proposition}
If $\sub{H} \preceq f$ and $f$ is a graph pattern polynomial family
with uniform $s(n)$-size circuits, then $\ghom H$ has uniform
$O(s(n))$-size circuits.
\end{proposition}
\begin{proof}
We can assume w.l.o.g.\ that $H$ does not have isolated vertices. Let
$H$ have $k$ nodes and let $K_{n}^k$ be the complete $k$-partite graph
with $n$ nodes in each partition. The nodes of $K_{n}^{k}$ are of the
form $(i, \kappa)$, $1 \le i \le n$, $1 \le \kappa \le k$. Let
$\sigma$ be a family of substitutions realizing $\sub H \preceq f$.
Consider $\sub {H,kn}$. We know that $\ml{\sigma_{m}(f_m)} = v_{[q]}
\sub {H,kn}$ for some $m = O(n)$ and $q = O(1)$. Since $H$ does not
contain isolated vertices, there is a function $g$ that maps $V(H)$ to
$E(H)$ such that the image of $f(v)$ for any $v$ is an edge incident
on $v$. Now we define the substitution $\tau$ on the edge and vertex
variables:
\begin{align*}
\tau(x_{\{(i, \kappa),(j, \mu)\}}) &=
\begin{cases}
Y_{i, \kappa, \{\kappa, \mu\}} Y_{j, \mu, \{\kappa, \mu\}} x_{\{i,j\}} &
\text{if $i \not= j$}, \\ 0 & \text{if $i = j$ or
$\{\kappa,\mu\}\not\in E(H)$},
\end{cases}\\
\tau(y_{(i,\kappa)}) &= \hat a_\kappa,
\end{align*}
where the variables $\hat a$ are fresh variables that we need
for book-keeping and we define:
\begin{align*}
Y_{i, \kappa', \{\kappa, \mu\}} &= z_{\kappa, i} y_i &\text{if $g(\kappa') = \{\kappa, \mu\}$}\\
Y_{i, \kappa', \{\kappa, \mu\}} &= 1 &\text{if $g(\kappa') \not= \{\kappa, \mu\}$}
\end{align*}
Every embedding of $H$ into $K_{n}^k$ such that each node of $H$ goes
into another part will contribute a term that is multilinear in the
$\hat a_\kappa$-variables in $\tau(N_{H,kn})$. The substitution also
ensures that the colours of edges correspond to edges in $H$ and
labels of adjacent vertices are different. It is easy to see that
these embeddings correspond to homomorphisms to $K_n$. We have proved
that the part of $\tau(\sub{H, kn})$ multilinear in $\hat a$ variables
is,
\begin{equation*}
\hat a_{V(H)} \sum_{\phi: H \overset{\mathit{hom}}{\mapsto} K_n} \prod_{v\in V(H)} z_{v,
\phi(v)} y_{\phi(v)} \prod_{e\in E(H)} x_{\phi(e)} = \hat a_{[k]}
\ghom {H,n}.
\end{equation*}
Furthermore the part of $\tau(\sigma_{m}(f_m))$ multilinear in $\hat
a$ and $v_i$ variables is $v_{[q]} \hat a_{[k]}\ghom {H,n}$ since
every non-multilinear term stays non-multilinear under $\tau$.
Therefore, we get an exact computation for $\ghom {H,n}$ by
differentiating the circuit with respect to $v_1,\dots,v_q,\hat
a_1,\dots,\hat a_k$ once and then setting all variables $v_i$ for all
$i$ and $\hat a_1,\dots,\hat a_k$ to $0$. Note that each
differentiation will increase the circuit size by a constant factor
and we differentiate a constant number of times. This operation is
linear-time in the size of the circuit.\footnote{Note that unlike in
the Baur-Strassen theorem, we only compute \emph{one} derivative!}
\end{proof}
The above result can be interpreted in two different ways: (1)
Homomorphism polynomials are the best graph pattern polynomials or (2)
Efficient constructions for homomorphism polynomials can be obtained
by obtaining efficient constructions for \emph{any} pattern family $f$
such that $\sub{H}\preceq f$.
\begin{lemma}
Let $k>2$. If $H\neq K_k$ is a $k$-vertex graph, then
$2\sub{H}\preceq \ghom{K_k-e}$.
\end{lemma}
\begin{proof}
The proof of this claim is similar to the proof of
Theorem~\ref{thm:ind_graph_to_clique-full}. Let $M$ be the labelling of
$K_k - e$ using $[k]$ such that vertices $1$ and $k$ are not
adjacent. Let $L$ be a labelling of $H$ using $[k]$ such that $1$
and $k$ are not adjacent. Therefore, the labelled graph $L$ is a
subgraph of the labelled graph $M$. Let $q_1, \dotsc, q_\ell$ be
the edges of $L$ and $q_{\ell+1},\dotsc,q_m$ be the non-edges of
$L$. Let $S$ be the set of all labellings of $H$. For each
labelling $L'$ in $S$, associate a permutation with $L'$ such that
applying it to $L'$ yields $L$. Let $P$ be the set of all such
permutations.
We partition $P$ into $P_1$ and $P_2$ as follows: A permutation
$\phi\in P_1$ if given a sequence of $k$ numbers, we can determine
whether the sequence is consistent with $\phi$, i.e., the
$i^\text{th}$ smallest element in the sequence is at position
$\phi(i)$, without comparing the first and last elements in the
sequence. Otherwise, $\phi\in P_2$. We start with the
$\ghom{K_k - e}$ polynomial over the vertex set
$[n]\times [k]\times P$ and apply the following substitution.
\setcounter{equation}{0}
\begin{align}
\sigma_{H}(y_{(v, p, \phi)}) &= y_v\\
\sigma_{H}(x_{\{(v_1, p_1, \phi), (v_2, p_2, \phi')\}}) &= 0, \text{ if $\phi\neq \phi'$}\\
\sigma_{H}(x_{\{(v_1, p_1, \phi), (v_2, p_2, \phi)\}}) &= 0, \phi^{-1}(p_1) < \phi^{-1}(p_2) \wedge v_1 > v_2\\
\sigma_{H}(x_{\{(v_1, p_1, \phi), (v_2, p_2, \phi)\}}) &= \begin{cases}
x_{\{v_1, v_2\}}, & \{p_1, p_2\}\in E(L)\\
1, &\{p_1, p_2\}\in E(M)\setminus E(L)\\
0, &\text{otherwise}
\end{cases}\\
\sigma_{H}(z_{(1, (v, 1, \phi))}) &= \begin{cases}
u_1, &\phi\in P_2\\
2u_1, &\phi\in P_1
\end{cases}\\
\sigma_{H}(z_{(i, (v, i, \phi))}) &= u_i, i > 1\\
\sigma_{H}(z_{(i, (v, j, \phi))}) &= u_i^2, i \neq j
\end{align}
First, we state some properties satisfied by the surviving
monomials. Rule~1 ensures that all vertices have different
labels. Rule~2 ensures that all variables in a surviving monomial
are indexed by the same permutation. Rules~6 and 7 ensure that all
vertices have different colours. Let $\tau = (1\ k)(2)\dotsm
(k-1)$. Consider an arbitrary surviving monomial indexed by a
permutation $\phi$. If $\phi\in P_1$, then Rule~3 ensures that the
vertices of the monomial are consistent with $\phi$. Assume that
the vertices are $(v_1, 1, \phi),\dotsc, (v_k, k, \phi)$ and they
are not consistent with $\phi$. This is possible only if
$\phi^{-1}(1) < \phi^{-1}(k)$ and $v_1 > v_k$ or $\phi^{-1}(1) >
\phi^{-1}(k)$ and $v_1 < v_k$. Since $\phi\in P_1$, there exists
an $i'$ such that $\phi^{-1}(1) < \phi^{-1}(i') < \phi^{-1}(k)$ or
$\phi^{-1}(1) > \phi^{-1}(i') > \phi^{-1}(k)$. Therefore, we have
that the vertices are inconsistent at either $\{1, i'\}$ or $\{i',
k\}$, a contradiction. If $\phi\in P_2$, then Rule~3 ensures that
the vertices are consistent with $\phi$ or $\tau\circ\phi$. To see
this, observe that, by Rule~3, the inconsistency with $\phi$ can
only occur $\{1, k\}$. This implies that the vertices are
consistent with $\tau\circ\phi$ because ${(\tau\circ\phi)}^{-1}(1)
= \phi^{-1}(k)$ and ${(\tau\circ\phi)}^{-1}(k) = \phi^{-1}(1)$
removing the inconsistency at $\{1, k\}$ and for all other $i$, we
have ${(\tau\circ\phi)}^{-1}(i) = \phi^{-1}(i)$ preserving
consistency at all other points.
Consider a labelled $H$, say $L'$, labelled using
$v_1 < \dotsb < v_k$ with associated permutation $\phi$. Let
$\psi : v_i \mapsto i$. Let $e_1, \dotsc, e_m$ be the edges and
non-edges of $L'$ such that $e_i = \psi^{-1}(\phi^{-1}(q_i))$ for
all $i$. We split the proof into two cases: If $\phi\in P_1$, the
monomial
$z_{(1, (v_{\phi^{-1}(1)}, 1, \phi))}\dotsm z_{(1,
(v_{\phi^{-1}(k)}, k, \phi))}$ (A monomial in $\ghom{K_k - e}$
is completely determined by the homomorphism variables and we will
not specify the other variables for brevity) uniquely generates
the monomial in $\sub{H}$ that corresponds to $L'$. If
$\phi\in P_2$, then there are two cases to consider depending on
whether the permutation $\tau$ is in $Aut(L)$ or not. If
$\tau\not\in Aut(L)$, then the monomials
$z_{(1, (v_{\phi^{-1}(1)}, 1, \phi))}\dotsm z_{(1,
(v_{\phi^{-1}(k)}, k, \phi))}$ and
$z_{(1, (v_{\phi^{-1}(1)}, 1, \tau\circ\phi))}\dotsm z_{(1,
(v_{\phi^{-1}(k)}, k, \tau\circ\phi))}$ are the only two
monomials that yield the required monomial. If $\tau\in Aut(L)$,
then the monomials
$z_{(1, (v_{\phi^{-1}(1)}, 1, \phi))}\dotsm z_{(1,
(v_{\phi^{-1}(k)}, k, \phi))}$ and
$z_{(1, (v_{\phi^{-1}(k)}, 1, \phi))}\dotsm z_{(1,
(v_{\phi^{-1}(1)}, k, \phi))}$ are the only two monomials that
yield the required monomial.
\end{proof}
The above lemma shows that, as expected, the polynomial $\ghom{K_k -
e}$ is strong enough to compute every other graph homomorphism
except that of $K_k$. This allows us to parameterize many existing
results in terms of the size of the arithmetic circuits computing
$\ghom{K_k - e}$.
\begin{theorem}
If there are uniform $O(n^{s(k)})$ size circuits for $\ghom{K_k -
e}$, then the number of subgraph isomorphisms for any $k$-vertex
$H\neq K_k$ can be computed in $O(n^{s(k)})$ time on $n$-vertex
graphs.
\end{theorem}
\begin{proof}
For all $k$-vertex $H\neq K_k$, we have $2\sub{H}\preceq \ghom{K_k -
e}$. For all $H$ on less than $k$ vertices, we have
$\sub{H}\preceq\ind{K_k}\preceq\ghom{K_k - e}$. Therefore, for all
graphs $H\neq K_k$ on at most $k$ vertices, we can construct
$O(n^{s(k)})$ size circuits that compute $2\ghom{H}$. We know that
the number of subgraph isomorphisms for $H$ can be expressed as a
linear combination of the number of homomorphisms for $H$ and the
number of homomorphisms for graphs on less than $k$ vertices
\cite{CDM17}.
\end{proof}
\begin{theorem}
\label{thm:Kk-e-full}
If there are uniform $O(n^{s(k)})$ size circuits for $\ghom{K_k -
e}$, then the induced subgraph isomorphism problem for all
$k$-vertex pattern graphs except $K_k$ and $I_k$ have an
$O(n^{s(k)})$ time algorithm.
\end{theorem}
\begin{proof}
We will show how to decide induced subgraph isomorphism for $H\neq
K_k$ in $O(n^{s(k)})$ time. Now, choose a prime $p$ such that $p$
divides the number of occurences of $H$ in $K_k$. The number of
induced subgraph isomorphisms modulo $p$ for $H$ can be expressed as
a linear combination of the number of subgraph isomorphisms modulo
$p$ of $k$-vertex graphs except $K_k$ and can be computed in
$O(n^{s(k)})$ time. It is known that the induced subgraph
isomorphism problem for $H$ is randomly reducible to this problem
\cite{WWWY15}.
\end{proof}
\begin{theorem}
\label{thm:Kk-e-counting-full}
If there are uniform $O(n^{s(k)})$ size circuits for $\ghom{K_k -
e}$ and if there is an $O(t(n))$ time algorithm for counting the
number of induced subgraph isomorphisms for a $k$-vertex pattern
$H$, then the number of induced subgraph isomorphisms for all
$k$-vertex patterns can be computed in $O(n^{s(k)} + t(n))$ time on
$n$-vertex graphs.
\end{theorem}
\begin{proof}
We know that $i_H = \sum_{H'\sqsupseteq H} a_{H'} n_{H'}$, where all
$a_{H'}\neq 0$, $i_H$ is the number of induced subgraph isomorphisms
from $H$ to $G$ and $n_H$ is the number of subgraph isomorphisms from
$H$ to $G$. Furthermore, we can compute $n_{H'}$ for all $H'\neq K_k$
in $O(n^{s(k)})$ time. Therefore, if we can compute $i_H$ in $t(n)$
time, we can compute $n_{K_k}$ in $O(n^{s(k)} + t(n))$ time.
\end{proof}
The following corollary follows by observing that $\tw{K_k - e} =
k-2$.
\begin{corollary}
All $k$-vertex pattern graphs except $K_k$ and $I_k$ have an
$O(n^{k-1})$ time combinatorial algorithm for deciding induced
subgraph isomorphism on $n$-vertex graphs.
\end{corollary}
\begin{corollary}
For $k\in\{4,5,6,7,8\}$, the induced subgraph isomorphism problem
for any $k$-vertex pattern graph $H$ except $K_k$ and $I_k$ can be
decided faster than currently known best clique algorithms.
\end{corollary}
\begin{proof}
The polynomial family $\ghom{K_k - e}$ can be computed by uniform
arithmetic circuits of size
$O(n^{\omega(\lceil \frac{k-2}{2} \rceil, 1, \lfloor \frac{k-2}{2}
\rfloor)})$ for all $k$. The construction is similar to the other
constructions for homomorphism polynomials using fast matrix
multiplication in this paper.
\end{proof}
\section{Omitted Proofs}
\begin{proof}{(Of Theorem~\ref{thm:test:1-full})}
Suppose that the arithmetic circuit given as input computes a
polynomial on $n$ variables. For each variable $x_i$ select
$a\in \mathbb{Z}_{p}$ and $y\in \{y_1,\dotsc,y_k\}$ uniformly at
random and substitute $ay$ for $x_i$. Evaluate the circuit over
$\mathbb{Z}_{p}[y_1,\dotsc,y_k]/\langle
y_{1}^{2},\dotsc,y_{k}^{2} \rangle$ and accept if and only if the
result is non-zero. The correctness of this algorithm can be proved
by induction on the number of multilinear terms in the polynomial
that are non-zero modulo $p$.
\end{proof}
\begin{proof}{(Of Theorem~\ref{thm:diaz-full})}
We will describe how to construct an arithmetic circuit of size
$O(n^{t+1})$ for $\ghom{H}$ where $t = \tw{H}$. The construction
mirrors the algorithm in Theorem~{3.1} in \cite{DiazST02}. We start
with a nice tree decomposition $D$ of $H$. Each gate in the circuit
will be labelled by some node (say $p$) in $D$ and a partial
homomorphism $\phi : V(H) \mapsto [n]$. The label is $I_p(\phi)$.
Let $p$ be a node in the tree decomposition $D$. Construct the
circuit in a bottom-up fashion as follows:
\begin{description}
\item [$p$ is a start node with $X_p = \{ a \}$] Add $n$ input gates
labelled $I_p(\{(a, v)\})$ with the constant $1$ as value for each
$v\in[n]$.
\item [$p$ is an introduce node] Let $q$ be the child of $p$ and
$X_p - X_q = \{ a \}$. Add gates labelled $I_p(\phi \cup \{(a,
v)\}) = I_q(\phi)$ for each $v\in[n]$. Since there are at most
$O(n^{t+1})$ choices for $\phi \cup \{(a, v)\}$, there are at most
$O(n^{t+1})$ gates.
\item [$p$ is a join node] Let $q_1$ and $q_2$ be the children of
$p$. Add gates labelled $I_p(\phi) = I_{q_1}(\phi)
. I_{q_2}(\phi)$. Since there are at most $O(n^{t+1})$ choices for
$\phi$, there are at most $O(n^{t+1})$ gates.
\item [$p$ is a forget node] Let $q$ be the child of $p$ such that
$X_q - X_p = \{ a \}$. Add gates $I_p(\phi) = \sum_{v\in[n]} z_{a,
v} y_v x_{\{v, u_1\}} \dotsm x_{\{v, u_k\}} I_q(\phi \cup \{(a,
v)\})$ where $\{v, u_i\}, 1\leq i\leq k$ are the images of the
edges incident on $a$ in partial homomorphism $\phi \cup \{(a,
v)\}$. Note that there are $O(n)$ gates corresponding to the tuple
$(p, \phi)$. Since $p$ is a forget node, there are at most
$O(n^t)$ such tuples and therefore at most $O(n^{t+1})$ gates.
\end{description}
\end{proof}
\section{Discussion}
Since the subgraph isomorphism and homomorphism polynomials for
cliques have the same size complexity, there is no advantage to be
gained by using homomorphism polynomials instead of subgraph
isomorphism problem. How hard is it to obtain better circuits for
$\ghom{K_k}$? As the following proposition shows, improving the size
of $\ghom{K_3}$ implies improving matrix multiplication.
\begin{proposition}
If $\sub{K_3}$ (or $\ind{K_3}$ or $\ghom{K_3}$) has $O(n^\tau)$-size
circuits then the exponent of matrix multiplication $\omega \le
\tau$.
\end{proposition}
\begin{proof}
Let $G$ be the complete tripartite graph $T_n$ on $3n$-vertices with
partitions of size $n$. The vertex set of $T_n$ is
$[3] \times [n]$. Instead of substituting a $1$ for every edge in
$T_n$, we substitute the variables $a_{i,j}$ for edges
$\{(1,i),(2,j)\}$, $b_{i,j}$ for edges $\{(2,i),(3,j)\}$, and
$c_{i,j}$ for edges $\{(3,i),(1,j)\}$. The resulting polynomial is:
\begin{equation*}
N' = \sum_{i = 1}^n \sum_{j = 1}^n \sum_{k = 1}^n y_{1,i} y_{2,j} y_{3,k} \cdot a_{i,j} b_{j, k} c_{k, i}
\end{equation*}
We subsitute $1$ for all vertex variables and obtain
\begin{equation*}
N'' = \sum_{i = 1}^n \sum_{j = 1}^n \sum_{k = 1}^n a_{i,j} b_{j, k} c_{k, i}
\end{equation*}
$N''$ has $O(n^\tau)$-size circuits. It is well-known that
$\omega \le \tau$ follows from this, see e.g.~\cite{DBLP:journals/toc/Blaser13}.
\end{proof}
It is interesting to know whether such connections exist for $k > 3$.
\section{Pattern graphs easier than cliques}
\label{sec:easypattern-full}
In this section, we describe a family $H_{3k}$ of pattern graphs such that the
induced subgraph isomorphism problem for $H_{3k}$ has an $O(n^{\omega(k, k-1,
k)})$ time algorithm when $k = 2^\ell, \ell \geq 1$. Note that for the currently
known best algorithms for fast matrix multiplication, we have $\omega(k, k-1, k)
< k\omega$. Therefore, these pattern graphs are strictly easier to detect than
cliques.
The pattern graph $H_{3k}$ is defined on $3k$ vertices and we consider the
canonical labelling of $H_{3k}$ where there is a $(3k-1)$-clique on vertices
$\{1,\dotsc,3k-1\}$ and the vertex $3k$ is adjacent to the vertices
$\{1,\dotsc,2k-1\}$.
\begin{lemma}
\label{lem:IH3kNH3kmod2-full}
$\ind{H_{3k}} = \sub{H_{3k}} \pmod 2$ when $k = 2^\ell, \ell \geq 1$
\end{lemma}
\begin{proof}
We show that the number of times $H_{3k}$ is contained in any of its proper
supergraphs is even if $k$ is a power of $2$. The graph $K_{3k}$ contains
$3k{\binom{3k-1}{2k-1}}$ copies of $H_{3k}$. This number is even when $k$ is
even. The graph $K_{3k} - e$ contains $2{\binom{3k-2}{2k-1}}$ copies of
$H_{3k}$. This number is always even. The remaining proper supergraphs of
$H_{3k}$ are the graphs $K_{3k-1} + (2k+i)e$, i.e., a $(3k-1)$-clique with
$2k+i$ edges to a single vertex, for $0\leq i < k-2$. There are $m_i =
{\binom{2k+i}{2k-1}}$ copies of the graph $H_{3k}$ in these supergraphs. We
observe that the numbers $m_i$ are even when $k = 2^\ell, \ell \geq 1$ by
Lucas' theorem. Lucas' theorem states that ${\binom{p}{q}}$ is even if and
only if in the binary representation of $p$ and $q$, there exists some bit
position $i$ such that $q_i = 1$ and $p_i = 0$. To see why $m_i$ is even,
observe that in the binary representation of $2k-1$, all bits $0$ through
$\ell$ are $1$ and in the binary representation of $2k+i, 0\leq i < k-2$, at
least one of those bits is $0$.
\end{proof}
\begin{lemma}
\label{lem:NH3kHomH3k-full}
$\sub{H_{3k}} \preceq \ghom{H_{3k}}$
\end{lemma}
\begin{proof}
We start with $\ghom{H_{3k}}$ over the vertex set $[n]\times [3k]$ and apply
the following substitution.
\setcounter{equation}{0}
\begin{align}
\sigma(z_{a, (v, a)}) &= z_a\\
\sigma(z_{a, (v, b)}) &= z_a^2, a\neq b\\
\sigma(y_{(v, a)}) &= y_v\\
\sigma(x_{(u, a), (v, b)}) &= 0, \text{if $a, b\in\{1,\dotsc,2k-1\}$ and $a < b$ and $u > v$}\\
\sigma(x_{(u, a), (v, b)}) &= 0, \text{if $a, b\in\{2k,\dotsc,3k-1\}$ and $a < b$ and $u > v$}\\
\sigma(x_{(u, a), (v, b)}) &= x_{\{u, v\}}, \text{otherwise}
\end{align}
Rule 3 ensures that in any surviving monomial, all vertices have distinct
labels. Rule 4 ensures that the vertices coloured $1,\dotsc, 2k-1$ are in
increasing order and Rule 5 ensures that the vertices coloured
$2k,\dotsc,3k-1$ are in increasing order.
Consider an $H_{3k}$ labelled using $[n]$ where the vertices in the
$(3k-1)$-clique are labelled $v_1,\dotsc, v_{3k-1}$ and the
remaining vertex is labelled $v_{3k}$ which is connected to
$v_1 < \dotso < v_{2k-1}$. Also, $v_{2k} < \dotso < v_{3k-1}$. We
claim that the monomial corresponding to this labelled $H_{3k}$ (say
$m$) is uniquely generated by the monomial
$m' = \prod_{1\leq i\leq 3k} z_{i, (v_i, i)} w$ in
$\ghom{H_{3k}}$. Note that the vertices and edges in the image of
the homomorphism is determined by the map $i\mapsto (v_i, i)$. The
monomial $w$ is simply the product of these vertex and edge
variables. It is easy to see that this monomial yields the required
monomial under the above substitution. The uniqueness is proved as
follows: observe that in any monomial $m''$ in $\ghom{H_{3k}}$ that
generates $m$, the vertex coloured $3k$ must be $v_{3k}$. This
implies that the vertices coloured $1,\dotsc,2k-1$ must be the set
$\{v_1,\dotsc,v_{2k-1}\}$. Rule~4 ensures that vertex coloured $i$
must be $v_i$ for $1\leq i \leq 2k-1$. Similarly, the vertices
coloured $2k,\dotsc,3k-1$ must be the set
$\{v_{2k},\dotsc,v_{3k-1}\}$ and Rule~5 ensures that vertex coloured
$i$ must be $v_i$ for $2k\leq i\leq 3k-1$ as well. But then the
monomials $m'$ and $m''$ are the same.
\end{proof}
\begin{lemma}
\label{lem:HomH3kcircuits-full}
$\ghom{H_{3k}}$ can be computed by arithmetic circuits of size
$O(n^{\omega(k, k-1, k)})$ for $k > 1$.
\end{lemma}
\begin{proof}
Consider $H_{3k}$ labelled as before. We define the sets
$S_{1,k,2k,3k-1} = \{1,\dotsc,k,2k\dotsc,3k-1\}$, $S_{k+1, 3k-1} =
\{k+1,\dotsc,3k-1\}$, $S_{k+1, 2k-1} = \{k+1,\dotsc,2k-1\}$, and
$S_{1, 2k-1} = \{1,\dotsc,2k-1\}$. We also define the tuples $V_{1,
k} = (v_1, \dotsc, v_k)$, $V_{2k, 3k-1} = (v_{2k}, \dotsc,
v_{3k-1})$, and $V_{k+1, 2k-1} = (v_{k+1}, \dotsc, v_{2k-1})$ for
any set $v_i$ of $3k-1$ distinct vertex labels. The algorithm also
uses the matrices defined below. The dimensions of each matrix are
specified as the superscript. All other entries of the matrix are
$0$.
\begin{align*}
A^{n^k\times n^k}_{V_{1, k}, V_{2k, 3k-1}} &= \smashoperator{\prod\limits_{i\in S_{1,k,2k,3k-1}}} z_{i, v_i} y_{v_i} \prod\limits_{\substack{i, j\in S_{1,k,2k,3k-1}\\ i\neq j}} x_{\{v_i, v_j\}}
,\text{$v_i$ distinct for $1\leq i \leq 3k-1$}\\
B^{n^k\times n^{k-1}}_{V_{2k, 3k-1}, V_{k+1, 2k-1}} &= \smashoperator{\prod\limits_{i\in S_{k+1,2k-1}}} z_{i, v_i} y_{v_i} \prod\limits_{\substack{i\in S_{k+1,3k-1}\\ j\in S_{k+1,2k-1}\\ i\neq j}} x_{\{v_i, v_j\}}
,\text{$v_i$ distinct for $k+1\leq i\leq 3k-1$}\\
C^{n^{k-1}\times n^k}_{V_{k+1,2k-1}, V_{1,k}} &= x_{\{{(v_i, i)}_{i\in S_{1,2k-1}}\}} \smashoperator{\prod\limits_{\substack{i\in S_{k+1,2k-1}\\ j\in [k]\\ i\neq j}}} x_{\{v_i, v_j\}}
,\text{$v_i$s are distinct for $1\leq i\leq 2k-1$}\\
D^{n^k\times n}_{V_{1,k}, v_{3k}} &= z_{3k, v_{3k}} y_{v_{3k}} \prod\limits_{i\in [k]} x_{\{v_i, v_{3k}\}}
, \text{$v_i$ distinct for $i\in\{1,\dotsc,k,3k\}$}\\
E^{n\times n^{k-1}}_{v_{3k}, V_{k+1,2k-1}} &= \smashoperator{\prod\limits_{i\in S_{k+1,2k-1}}} x_{\{v_i, v_{3k}\}}
, \text{$v_i$ distinct for $i\in\{k+1,\dotsc,2k-1,3k\}$}
\end{align*}
Compute the matrix products $ABC$ and $DE$. Replace the $n^{2k-1}$
variables $x_{\{{(v_i, i)}_{i\in S_3}\}}$ with
${(DE)}_{V_{1,k}, V_{k+1,2k-1}}$. The required
polynomial is then just
\begin{align*}
\ghom{H_{3k}} &= \sum_{(v_1, \dotsc, v_k)} {(ABC)}_{(v_1, \dotsc,
v_k),(v_1, \dotsc, v_k)}
\end{align*}
Consider a homomorphism of $H_{3k}$ defined as $\phi: i\mapsto
u_i$. The monomial corresponding to this homomorphism is uniquely
generated as follows. Let $U_{*}$ be defined similarly to the
tuples $V_{*}$. Set $v_i = u_i$ for $i\in [k]$ in the summation and
consider the monomial generated by the product $A_{U_{1,k},
U_{2k,3k-1}}B_{U_{2k,3k-1}, U_{k+1, 2k-1}}C_{U_{k+1, 2k-1}, U_{1,
k}}$ after replacing the variable $x_{\{{(u_i, i)}_{i\in
S_3}\}}$ by ${(DE)}_{U_{1, k}, U_{k+1,2k-1}}$ taking the
monomial $D_{U_{1,k},u_{3k}}E_{u_{3k}, U_{k+1, 2k-1}}$ from that
entry. It is easy to verify that this generates the required
monomial. For uniqueness, observe that this is the only way to
generate the required product of the homomorphism variables.
Computing $ABC$ can be done using $O(n^{\omega(k, k-1, k)})$ size circuits.
Computing $DE$ can be done using $O(n^{\omega(k, 1, k-1)})$ size circuits. The
top level sum contributes $O(n^k)$ gates. This proves the lemma.
\end{proof}
We conclude this section by stating our main theorem.
\begin{theorem}\label{thm:H3k-full}
The induced subgraph isomorphism problem for $H_{3k}$ has an $O(n^{\omega(k,
k-1, k)})$ time algorithm when $k = 2^\ell, \ell \geq 1$.
\end{theorem}
\section{Graph pattern polynomial families}
We will consider polynomial families $f = (f_n)$ of the following
form: Each $f_n$ will be a polynomial in variables $y_1,\dots,y_n$,
the vertex variables, and variables $x_1,\dots,x_{\binom n 2}$, the
edge variables, and at most linear in $n$ number of additional
variables.The degree of each $f_n$ will usually be constant.
The (not necessarily induced) subgraph isomorphism polynomial family
$\sub H = {(\sub{H,n})}_{n\geq 0}$ for a fixed pattern graph $H$ on
$k$ vertices and $\ell$ edges is a family of multilinear polynomials
of degree $k+\ell$. The $n^\text{th}$ polynomial in the family,
defined over the vertex set $[n]$, is the polynomial on
$n + \binom{n}{2}$ variables given by (\ref{eq:sub-full}):
\begin{equation}
\sub{H,n} = \sum_{\phi : V(H)\overset{\mathit{sub}}{\mapsto} V(K_n)} y_{\phi(V(H))} x_{\phi(E(H))}
\label{eq:sub-full}
\end{equation}
When context is clear, we will often omit the subscript $n$ and simply
write $\sub H$. Given a (host) graph $G$ on $n$ vertices, we can
substitute values for the edge variables of $\sub{H,n}$ depending on
the edges of $G$ ($x_e = 1$ if $e\in E(G)$ and $x_e = 0$ otherwise) to
obtain a polynomial $\sub{H,n}(G)$ on the vertex variables. The
monomials of this polynomial are in one-to-one correspondence with the
$H$-subgraphs of $G$. i.e., a term $ay_{v_1}\dotsm y_{v_k}$, where $a$
is a positive integer, indicates that there are $a$ subgraphs
isomorphic to $H$ in $G$ on the vertices $v_1,\dotsc,v_k$. Therefore,
to detect if there is an $H$-subgraph in $G$, we only have to test
whether $\sub{H,n}(G)$ has a multilinear term.
The induced subgraph isomorphism polynomial family
$\ind{H} = {(\ind{H,n})}_{n\geq 0}$ for a pattern graph $H$ over the
vertex set $[n]$ is defined in (\ref{eq:ind-full}).
\begin{equation}
\ind{H,n} = \sum_{\phi : V(H)\overset{\mathit{ind}}{\mapsto} V(K_n)} y_{\phi(V(H))} x_{\phi(E(H))} \prod_{e\not\in E(H)} (1 - x_{\phi(e)})
\label{eq:ind-full}
\end{equation}
If we substitute the edge variables of $\ind{H,n}$ using a host graph
$G$ on $n$ vertices, then the monomials of the resulting polynomial
$\ind{H,n}(G)$ on the vertex variables are in one-to-one
correspondence with the induced $H$-subgraphs of $G$. In particular,
all monomials have coefficient $0$ or $1$ because there can be at most
one induced copy of $H$ on a set of $k$ vertices. This implies that to
test if there is an induced $H$-subgraph in $G$, we only have to test
whether $\ind{H,n}(G)$ has a multilinear term and we can even do this
modulo $p$ for any prime $p$. Also, note that $\ind{H}$ is simply
$\ind{\overline{H}}$ where all the edge variables $x_e$ are replaced
by $1 - x_e$.
The homomorphism polynomial family $\ghom{H} = {(\ghom{H,n})}_{n\geq
0}$ for pattern graph $H$ over the vertex set $[n]$ is defined in
(\ref{eq:hom-full}).
\begin{equation}
\ghom{H,n} = \sum_{\phi: H \overset{\mathit{hom}}{\mapsto} K_n}
\prod_{v\in V(H)} z_{v, \phi(v)} y_{\phi(v)} \prod_{e\in E(H)} x_{\phi(e)}
\label{eq:hom-full}
\end{equation}
The variables $z_{a, v}$'s are called the \emph{homomorphism
variables}. They keep track how the vertices of $H$ are mapped by
the different homomorphisms in the summation. We note that the size of
the arithmetic circuit computing $\ghom{H,n}$ is independent of the
labelling chosen to define the homomorphism polynomial.
The induced subgraph isomorphism polynomial for any graph $H$ and
subgraph isomorphism polynomials for supergraphs of $H$ are related as
follows:
\begin{equation}\label{eq:indsub-full}
\ind{H,n} = \sum_{H' \sqsupseteq H} {(-1)}^{e(H') - e(H)} \nsub{H}{H'} \sub{H',n}
\end{equation}
Here $e(H)$ is the number of edges in $H$ and $\nsub{H}{H'}$ is the
number of times $H$ appears as a subgraph in $H'$. The sum is taken
over all supergraphs $H'$ of $H$ having the same vertex set as
$H$. Equation~\ref{eq:indsub-full} is used by Curticapean, Dell, and
Marx \cite{CDM17} in the context of counting subgraph isomorphisms.
\begin{example}
Let $P_3$ be the path on $3$ vertices and let $K_3$ be the triangle.
\begin{align*}
\sub{P_3, 3} &= y_1y_2y_3(x_{\{1,2\}}x_{\{2,3\}} +
x_{\{1,3\}}x_{\{2,3\}} +
x_{\{1,2\}}x_{\{1,3\}})\\
\ind{P_3, 3} &= y_1y_2y_3\bigl(x_{\{1,2\}}x_{\{2,3\}}(1 - x_{\{1,3\}})\\
&+ x_{\{1,3\}}x_{\{2,3\}}(1 - x_{\{1,2\}})\\
&+ x_{\{1,2\}}x_{\{1,3\}}(1 - x_{\{2,3\}})\bigr)\\
&= \sub{P_3,3} - 3\sub{K_3,3}
\end{align*}
\end{example}
For any fixed pattern graph $H$, the degree of polynomial families
$\sub{H}$, $\ind{H}$, and $\ghom{H}$ are bounded by a constant
depending only on the size of $H$. Such polynomial families are called
constant-degree polynomial families.
\begin{definition}
A constant-degree polynomial family $f = (f_n)$ is called a
\emph{graph pattern} polynomial family if the $n^{\text{th}}$
polynomial in the family has $n$ vertex variables, $\binom{n}{2}$
edge variables, and at most $cn$ other variables, where $c$ is a
constant, and every non-multilinear term of $f_n$ has at least one
non-edge variable of degree greater than 1.
\end{definition}
It is easy to verify that $\ind{H}$, $\sub{H}$, and $\ghom{H}$ are all
graph pattern polynomial families. For a graph pattern polynomial $f$,
we denote by $f(G)$ the polynomial obtained by substituting $x_e = 0$
if $e\not\in E(G)$ and $x_e = 1$ if $e\in E(G)$ for all edge variables
$x_e$. Note that for any graph pattern polynomial $f$, we have
$\ml{f(G)} = \ml{f}(G)$. This is because any non-multilinear term in
$f$ has to remain non-multilinear or become 0 after this substitution.
\begin{definition}
\begin{enumerate}
\item A constant degree polynomial family $f = (f_n)$ has circuits
of size $s(n)$ if there is a sequence of arithmetic circuits
$(C_n)$ such that $C_n$ computes $f_n$ and has size at most
$s(n)$.
\item $f$ has uniform $s(n)$-size circuits, if on input $n$, we can
construct $C_n$ in time $O(s(n))$ on a Word-RAM.\footnote{Since we
are dealing with fine-grained complexity, we have to be precise
with the encoding of the circuit. We assume an encoding such
that evaluating the circuit is linear time and substituting for
variables with polynomials represented by circuits is
constant-time.}
\end{enumerate}
\end{definition}
We now define a notion of reducibility among graph pattern
polynomials.
\begin{definition} \label{def:subfam-full}
A \emph{substitution family} $\sigma = (\sigma_n)$
is a family of mappings
\[
\sigma_n: \{y_1,\dots,y_n,x_1,\dots,x_{\binom n2},u_1,\dots,u_{m(n)}\}
\to K[y_1,\dots,y_{n'},x_1,\dots,x_{\binom {n'}2},v_1,\dots,v_{r(n)}]
\]
mapping variables to polynomials such that:
\begin{enumerate}
\item $\sigma$ maps vertex variables to constant-degree monomials
containing one or more vertex variables or other variables, and
no edge variables.
\item $\sigma$ maps edge variables to polynomials with
constant-size circuits containing at most one edge variable and
no vertex variables.
\item $\sigma$ maps other variables to constant-degree monomials
containing no vertex or edge variables and at least one
other variable.
\end{enumerate}
$\sigma_n$ naturally extends to $K[y_1,\dots,y_n,$
$x_1,\dots,x_{\binom{n}{2}},$ $u_1,\dots,u_m]$.
\end{definition}
\begin{definition}
A substitution family $\sigma = (\sigma_n)$ is \emph{constant-time
computable} if given $n$ and a variable $z$ in the domain of
$\sigma_n$, we can compute $\sigma_n(z)$ in constant-time on a
Word-RAM. (Note that an encoding of any $z$ fits into on cell of
memory.)
\end{definition}
\begin{definition} \label{def:reduc-full}
Let $f = (f_n)$ and $g = (g_n)$ be graph pattern polynomial
families. Then $f$ is reducible to $g$, denoted $f\preceq g$, via a
constant time computable substitution family $\sigma = (\sigma_n)$
if for all $n$ there is an $m = O(n)$ and $q = O(1)$ such that
\begin{enumerate}
\item $\sigma_m(\ml{g_m})$ is a graph pattern polynomial and
\item $\ml{\sigma_m(g_m)} = v_{[q]}\ml{f_n}$. (Recall that
$v_{[q]} = v_1 \cdots v_q$.)
\end{enumerate}
For any prime $p$, we say that $f\preceq g \pmod p$ if there exists an
$f' = f \pmod p$ such that $f'\preceq g$.
\end{definition}
Property~1 of Definition~\ref{def:reduc-full} and Properties~1 and 3 of
Definition~\ref{def:subfam-full} imply that $\sigma_m(g_m)$ is a graph
pattern polynomial because Properties~1 and 3 of
Definition~\ref{def:subfam-full} ensure that non-multilinear terms remain
so after the substitution. It is easy to see that $\preceq$ is
reflexive via the identity substitution. We can also assume
w.l.o.g.\ that the variables $v_1,\dotsc,v_q$ are fresh variables
introduced by the substitution family $\sigma$.
What is the difference between $\sigma_m(\ml{g_m})$ and
$\ml{\sigma_m(g_m)}$ in the Definition~\ref{def:reduc-full}? Every
monomial in $\ml{\sigma_m(g_m)}$ also appears in $\sigma_m(\ml{g_m})$,
however, the latter may contain further monomials that are not
multilinear.
\begin{proposition}
$\preceq$ is transitive.
\end{proposition}
\begin{proof}
Let $f\preceq g$ via $\sigma$ and $g\preceq h$ via $\tau$. Assume
that $f_n$ is written as a substitution instance of $g_{m(n)}$ by
$\sigma$ and $g_m$ is written as a substitution instance of
$h_{r(m)}$ by $\tau$ for some linearly bounded functions $m$ and
$r$. Let $\sigma_{m(n)}(g_{m(n)})$ and
$\tau_{r(m(n))}(h_{r(m(n)))})$ have $u_1,\dotsc,u_p$ and
$v_1,\dotsc,v_q$, respectively, as other variables that are
multiplied with the multilinear terms. We can assume w.l.o.g.\ that
these two sets of other variables are disjoint.
Define $\sigma'$ as $\sigma$ extended to $v_i$ by $\sigma'_n(v_i) =
v_i$ for all $i$ and $n \in \mathbb{N}$. We claim that $f\preceq h$
via the family $(\sigma'_{m(n)} \circ \tau_{r(m(n))})$. We need to
verify the two properties of Definition \ref{def:reduc-full}.
\emph{Property~1}:
$\sigma'_{m(n)}(\tau_{r(m(n))}(\ml{h_{r(m(n))}})) =
\sigma'_{m(n)}(v_{[q]}\ml{g_{m(n)}} + h')$ where $h'$ is a graph
pattern polynomial containing only non-multilinear terms. Now, we
have $h'' = \sigma'_{m(n)}(v_{[q]}\ml{g_{m(n)}})$
$= v_{[q]}\sigma_{m(n)}(\ml{g_{m(n)}})$ because $\ml{g_{m(n)}}$
cannot contain $v_i$ and $\sigma'_{m(n)}(v_i) = v_i$ for $i\in
[q]$. This implies that $h''$ is a graph pattern polynomial because
$\sigma_{m(n)}(\ml{g_{m(n)}})$ is a graph pattern polynomial. Also,
$\sigma'_{m(n)}(h')$ is a graph pattern polynomial containing only
non-multilinear terms by Properties~1 and 3 of
Definition~\ref{def:subfam-full} proving that
$(\sigma'_{m(n)} \circ \tau_{r(m(n))})(\ml{h_{r(m(n))}})$ is a graph
pattern polynomial.
\emph{Property~2} is proved as follows:
\begin{align*}
\ml{(\sigma'_{m(n)} \circ \tau_{r(m(n))})(h_{r(m(n))})}
&= \ml{\sigma'_{m(n)}(\tau_{r(m(n))}(h_{r(m(n)))}))}\\
&= \ml{\sigma'_{m(n)}(v_{[q]}\ml{g_{m(n)}} + h')}\\
&= \ml{v_{[q]}\sigma_{m(n)}(\ml{g_{m(n)}})}\\
&= v_{[q]}\ml{\sigma_{m(n)}(\ml{g_{m(n)}})}\\
&= v_{[q]}u_{[p]}\ml{f_n }
\end{align*}
Note that the term $h'$ vanishes, since $\sigma_{m(n)}$ does not
introduce new multilinear monomials and also $\ml{.}$ is a linear
operator. The same happens in the second-last line, we did not write
the additional term in the equation, since it vanishes anyway.
We also have $r(m(n)) = O(n)$ and $p + q = O(1)$. It is easy to
verify that $(\sigma'_{m(n)} \circ \tau_{r(m(n))})$ is a
constant-time computable substitution family.
\end{proof}
Efficient algorithms are known for detecting multilinear terms of
\emph{small} degree with non-zero coefficient modulo primes. We state
two such theorems that we use in this paper.
\begin{theorem}
\label{thm:test:1-full}
Let $k$ be any constant and let $p$ be any prime. Given an
arithmetic circuit of size $s$, there is a randomized, one-sided
error $O(s)$-time algorithm to detect whether the polynomial
computed by the circuit has a multilinear term of degree atmost $k$
with non-zero modulo $p$ coefficient.
\end{theorem}
\begin{theorem}
\label{thm:multiparity-full}
Let $k$ be any constant. Given an arithmetic circuit of size $s$
computing a polynomial of degree $k$ on $n$ variables, there is a
deterministic $O(s+n^{\lceil k/2 \rceil})$-time algorithm to compute
the parity of the sum of coefficient of multilinear terms.
\end{theorem}
An important algorithmic consequence of reducibility is stated in
Proposition~\ref{prop:red-alg-full}.
\begin{proposition}
Let $p$ be any prime. Let $f$ and $g$ be graph pattern polynomial
families. Let $s(n)$ be a polynomially-bounded function. If
$f\preceq g$ and $g$ has size uniform $s(n)$-size arithmetic
circuits, then we can test whether $f_n(G)$ has a multilinear term
with non-zero coefficient modulo $p$ in $O(s(n))$ (randomized
one-sided error) time for any $n$-vertex graph $G$.
\label{prop:red-alg-full}
\end{proposition}
\begin{proof}
Assume that $f_n$ is reducible to $g_m$ where $m = O(n)$. Since
$s(n)$ is polynomially bounded, we have $\mathit{size}(g_m) =
O(s(n))$. Apply the substitution $\sigma_m$ to $g_m$ to obtain
$g'$. Let $u_1,\dots,u_r$ be the other variables of $g'$. We claim
that testing whether the polynomial $g'(G)$ has a multilinear term
is equivalent to testing whether $f_n(G)$ has a multilinear term. We
have $u_{[r]}\ml{f_n} = \ml{g'}$. Since both $f_n$ and $g'$ are
graph pattern polynomials, we have
$u_{[r]}\ml{f_n(G)} = u_{[r]}\ml{f_n}(G) = \ml{g'}(G) =
\ml{g'(G)}$. Therefore, testing whether the polynomial $f_n(G)$ has
a multilinear term of degree at most $k$, where $k$ is some
constant, reduces to testing whether $g'(G)$ has a multilinear term
of degree $k + r = O(1)$. Since $g'$ has $O(s(n))$ size circuits,
this can be done in $O(s(n))$ (randomized one-sided error) time.
\end{proof}
On the other hand, if we only have $f\preceq g \pmod p$ for some
specific prime $p$, then it is only possible to test for multilinear
terms in $f$ that have non-zero coefficients modulo $p$ for that prime
$p$.
\begin{corollary}
Let $f\preceq g \pmod p$ and $g$ has $s(n)$ size circuits where
$s(n)$ is polynomially bounded. Then we can test whether $f_n(G)$ has
a multilinear term with non-zero coefficient modulo $p$ in $O(s(n))$
time for any $n$-vertex graph $G$.
\end{corollary}
More relaxed notions of reduction allowing an increase of
$\polylog(n)$ factors in size or allowing multilinear terms to be
multiplied by arbitrary sets of other variables could also be useful
to obtain better algorithms. We do not pursue this because we could
not find any reductions that make use of this freedom.
The following result allows efficient construction of $\ghom{H}$ when
$H$ has small treewidth.
\begin{theorem}{(Implicit in \cite{DiazST02}, Also used in \cite{FominLRSR12} and \cite{DMMRS14})}
$\ghom{H}$ can be computed by $O\bigl(n^{\tw{H}+1}\bigr)$ size
arithmetic circuits for all graphs $H$.
\label{thm:diaz-full}
\end{theorem}
\section{Introduction}
The \emph{induced subgraph isomorphism problem} asks, given simple and
undirected graphs $G$ and $H$, whether there is an induced subgraph of
$G$ that is isomorphic to $H$. The graph $G$ is called the host graph
and the graph $H$ is called the pattern graph. This problem is
$\NP$-complete (See \cite{GareyJohnson90}, problem [GT21]). If the
pattern graph $H$ is fixed, there is a simple $O(n^{|V(H)|})$ time
algorithm to decide the induced subgraph isomorphism problem for
$H$. We study the time complexity of the induced subgraph isomorphism
problem for fixed pattern graphs on the Word-RAM model.
The earliest non-trivial algorithm for this problem was given by Itai
and Rodeh \cite{IR78} who showed that the the number of triangles can
be computed in $O(n^\omega)$ time on $n$-vertex graphs, where $\omega$
is the exponent of matrix multiplication. Later,
Ne\v{s}et\v{r}il and Poljak\ \cite{NP85} generalized this algorithm to count $K_{3k}$
in $O(n^{k\omega})$ time, where $K_{3k}$ is the clique on $3k$
vertices. Eisenbrand and Grandoni \cite{Eisenbrand2004} extended this
algorithm further to count $K_{3k+j}$ for $j\in\{0, 1, 2\}$ using
rectangular matrix multiplication in $O(n^{\omega(k + \lceil j/2
\rceil, k, k + \lfloor j/2 \rfloor)})$ time. This algorithm uses
fast matrix multiplication to achieve the speedup and in fact works
for all pattern graphs on $3k+j$ vertices. Hence we call this
algorithm the universal algorithm. It is reasonable to expect that one
might be able to obtain faster algorithms for specific pattern
graphs. However, algorithms faster than the universal algorithm are
only known for finitely many pattern graphs.
Algorithms that do not use fast matrix multiplication, called
\emph{combinatorial algorithms}, have also been studied. No
combinatorial algorithm that beats the trivial $O(n^k)$ time algorithm
is known for detecting $k$-cliques in $n$ vertex graphs. However,
improvements for certain pattern graphs such as $K_k-e$ has been shown
by Virginia Williams (See \cite{VV08}, p.45). They show a
combinatorial algorithm that decides the induced subgraph isomorphism
problem for $K_k - e$ in time $O(n^{k-1})$. An $O(n^{k-1})$
combinatorial algorithm is also known for deciding induced subgraph
isomorphism problem for $P_k$.
The use of algebraic methods has been particularly useful in finding
fast combinatorial algorithms for detecting pattern graphs. Ryan
Williams \cite{Williams09} gave a linear time algorithm for the (not
necessarily induced) subgraph isomorphism problem for $P_k$. This was
later generalized by Fomin, Lokshtanov, Raman, Saurabh, and Rao
\cite{FominLRSR12} to give $O(n^{\tw{H}+1})$ time algorithms for the
(not necessarily induced) subgraph isomorphism problem for $H$ in $n$
vertex graphs. These results use efficient constructions for
\emph{homomorphism polynomials} (defined later).
The question of whether improving algorithms for detecting a certain
pattern implies faster algorithms for another pattern has also been
studied. In particular, Ne\v{s}et\v{r}il and Poljak\ show that improved algorithms for
detecting $k$-cliques yield improved algorithms for all $k$-vertex
pattern graphs. More precisely:
\begin{theorem}{(\cite{NP85})}
If the induced subgraph isomorphism problem for $K_k$ can be decided
in $O(n^{f(k)})$ time for some $f(k)$, then the induced subgraph
isomorphism problem for $H$ can be decided in time $O(n^{f(k)})$
time, where $H$ is any $k$-vertex pattern graph.
\label{thm:np-universal-full}
\end{theorem}
In this sense, the $k$-clique is a \emph{universal} pattern.
Ne\v{s}et\v{r}il and Poljak's \cite{NP85} algorithm can be easily modified to output the
homomorphism polynomial no host graphs of $n$ vertices for the pattern
$K_{3k}$ in $O(n^{k\omega})$ time given $1^n$ as input. For cliques,
counting (or detecting) homomorphisms\footnote{For host $G$ and
pattern $H$, a function $f:V(H) \mapsto V(G)$ such that
$\{u, v\}\in E(H) \implies \{f(u), f(v)\}\in E(G)$} and counting
(or detecting) induced subgraph isomorphisms have the same
complexity. It is unclear whether computing homomorphism polynomials
efficiently for other pattern graphs help with the induced subgraph
isomorphism problem for those pattern graphs.
\subsection*{Our Results}
In this paper, we show that we can obtain algorithms that are faster
than the universal algorithm for infinitely many pattern graphs.
\begin{reptheorem}{thm:H3k-full}
There exists a family of pattern graphs ${(H_{3k})}_{k\geq 0}$ where
$H_{3k}$ is a $3k$-vertex graph such that the induced subgraph
isomorphism problem for $H_{3k}$ has an $O(n^{\omega(k, k-1, k)})$
time algorithm for infinitely many $k$.
\end{reptheorem}
Here, $\omega(p, q, r)$ is the exponent of $n^p\times n^q$ and
$n^q\times n^r$ matrix multiplication. The exponent of matrix
multiplication is defined as $\omega = \omega(1, 1, 1)$ (See
\cite{DBLP:journals/toc/Blaser13} for a more detailed
introduction). The best known algorithm for $K_{3k}$ takes time
$O(n^{k\omega})$ and the upper-bound on $\omega(k, k-1, k)$ is
strictly smaller than the upper-bound on $k\omega$ for the currently
known fastest matrix multiplication algorithms. If $\omega = 2$, then
we have $\omega(k, k-1, k) = k\omega$ and the above algorithm does not
improve upon the universal algorithm. However, the best upper-bound
that we know is $\omega < 2.373$. It is known that current methods
cannot show $\omega = 2$ (See \cite{AmbainisFG15}, \cite{AlmanW18}).
We develop an algebraic framework to study algorithms for the induced
subgraph isomorphism problems where we consider the size of the
pattern graph a constant. The above algorithm is obtained using this
framework. We show that the existing algorithms for natural pattern
graphs such as $k$-paths and $k$-cycles can be improved by efficiently
computing homomorphism polynomials for pattern graphs that are much
sparser than $k$-cliques.
We obtain, in Theorem~\ref{thm:AlgosForPaths-full} and
Theorem~\ref{thm:AlgosForCycles-full}, the following faster
(randomized, one-sided error) algorithms:
\begin{itemize}
\item Faster algorithms for induced subgraph isomorphism problem for
$P_k$ for $5\leq k \leq 9$.
\item Faster algorithms for induced subgraph isomorphism problem for
$C_k$ for $k\in\{5,7,9\}$.
\item $O(n^{k-2})$ time combinatorial algorithm for induced subgraph
isomorphism problem for $P_k$ and $C_k$.
\item $O(n^{k-2})$ time deterministic combinatorial algorithms for
computing the parity of the number of induced subgraphs isomorphic
to $P_k$ and $C_k$ in $n$-vertex graphs.
\end{itemize}
Unfortunately, we do not know how to compute these homomorphism
polynomials for smaller graphs using circuits of size smaller than
that for homomorphism polynomials for $k$-cliques when $k$ is
arbitrary. Therefore, we do not have an improvement similar to the one
in Theorem~\ref{thm:H3k-full} for paths or cycles.
In light of Theorem~\ref{thm:np-universal-full}, which shows that
$k$-cliques are universal, we show that homomorphism polynomials for
$K_k - e$, the $k$-vertex graph obtained by deleting an edge from
$K_k$, are \emph{almost} universal. We show that the arithmetic
circuit complexity of $\ghom{K_k - e}$ can be used to unify many
existing results. We show that if $\ghom{K_k - e}$ has $O(n^{f(k)})$
size circuits for some function $f(k)$, then:
\begin{enumerate}
\item (Theorem~\ref{thm:Kk-e-full}) The induced subgraph isomorphism
problem for all $k$-vertex pattern graphs other than $K_k$ and $I_k$
can be decided by an $O(n^{f(k)})$ time algorithm, where $k$ is
regarded as a constant and $f(k)$ is any function of
$k$. (\cite{VV08} gives a combinatorial algorithm for $K_k - e$,
\cite{FKLLTCS15} gives an algorithm for $P_k$)
\item (Theorem~\ref{thm:Kk-e-counting-full}) If there is an $O(t(n))$ time
algorithm for counting the number of induced subgraph isomorphisms
for a $k$-vertex pattern $H$, then the number of induced subgraph
isomorphisms for all $k$-vertex patterns can be computed in
$O(n^{f(k)} + t(n))$ time on $n$-vertex graphs. (\cite{KloksKM00}
gives this result for $k = 4$ and \cite{KowalukLL13} gives a weaker
result similar to this one)
\end{enumerate}
Even though these theorems are known for specific values of $f(k)$ as
cited above. We believe that the connection to homomorphism
polynomials for $\ghom{K_k - e}$ is new.
On the lower bounds front, we show in
Theorem~\ref{thm:supergraphs-full}, Theorem~\ref{thm:clique-hard-full}
and Theorem~\ref{thm:inducedHarderThanNonInduced-full} that within the
framework that we develop:
\begin{enumerate}
\item The induced subgraph isomorphism problem for any pattern
containing a $k$-clique or a $k$-independent set is at least as
hard as the isomorphism problem for $k$-clique.
\item For almost all pattern graphs $H$, the induced subgraph
isomorphism problem for $H$ is harder than the subgraph
isomorphism problem for $H$.
\item For almost all pattern graphs $H$, the subgraph isomorphism
problem for $H$ is easier than subgraph isomorphism problems for
all supergraphs of $H$.
\end{enumerate}
We note that only randomized algorithmic reductions are known for
Part~2 of the above theorem and Part~3 is unknown. It is not clear
whether our reductions imply algorithmic hardness for these problems.
\subsection*{Technique}
The \emph{Homomorphism polynomial} for a pattern graph $H$ denoted
$\ghom{H,n}$ is a polynomial such that the monomials of the polynomial
correspond one-to-one with homomorphisms from $H$ to an $n$-vertex
graph. Similarly, we define the graph pattern polynomial families
$\ind{H} = {(\ind{H,n})}_{n\geq 0}$ and $\sub{H} =
{(\sub{H,n})}_{n\geq 0}$ that correspond to the induced subgraph
isomorphism problem for $H$ and the (not necessarily induced) subgraph
isomorphism problem\footnote{Given $(G, H)$, decide whether there
exists an injective $f:V(H) \mapsto V(G)$ such that $\{u, v\}\in
E(H) \implies \{f(u), f(v)\}\in E(G)\}$.} for $H$ respectively. It
can be shown that testing for subgraph isomorphism is equivalent to
testing whether the homomorphism polynomial has multilinear terms
because subgraph isomorphisms are exactly the injective
homomorphisms. In fact, any polynomial family $f$ such that the
multilinear terms of $f$ correspond to multilinear terms of $\sub{H}$
is enough. This naturally leads to a notion of reduction between these
graph pattern polynomial families (denoted by $\preceq$). For example,
we say that $\sub{H}\preceq\ghom{H}$). This notion of reduction allows
us to compare the hardness of different pattern detection problems as
well as construct new algorithms as follows:
\begin{repproposition}{prop:red-alg-full}
Let $f$ and $g$ be graph pattern polynomial families. If $f\preceq g$
and $g$ has $O(n^{s(k)})$ size arithmetic circuits, then we can detect
patterns corresponding to $f$ using an $O(n^{s(k)})$ time algorithm.
\end{repproposition}
This framework naturally raises the question whether one can find
families $f$ such that $\sub{H}\preceq f$ and $f$ has smaller circuits
than $\ghom{H}$. We show that this is not possible by showing that in
this case $\ghom{H}$ has circuits that is as small as circuits for
$f$.
\subsection*{Other related work}
Curticapean, Dell, and Marx \cite{CDM17} showed that algorithms that
count homomorphisms can be used to count subgraph isomorphisms.
Williams, Wang, Williams, and Yu \cite{WWWY15} gave $O(n^\omega)$ time
algorithms for the induced subgraph isomorphism problems for four
vertex pattern graphs, except for $I_4$ and $K_4$. Floderus, Kowaluk,
Lingas, and Lundell \cite{FKLLTCS15} invented a framework that gives
$O(n^{k-1})$ combinatorial algorithms for induced subgraph isomorphism
problems for many pattern graphs on $k$ vertices.
Floderus, Kowaluk, Lingas, and Lundell \cite{FKLLSIAM15} showed
reductions between various induced subgraph isomorphism problems.
They proved that induced subgraph isomorphism problem for $H$ when $H$
contains a $k$-clique (or $k$-independent set) that is vertex-disjoint
from all other $k$-cliques (or $k$-independent sets) is at least as
hard as the induced subgraph isomorphism problem for $K_k$. They also
proved that detecting an induced $C_4$ is at least as hard as
detecting a $K_3$. This is the only example known where a pattern is
harder than another pattern that is not a subgraph. Hardness results
are also known for arithmetic circuits computing homomorphism
polynomials. Austrin, Kaski, and Kubjas \cite{AKK17} proved that
tensor networks (a restricted form of arithmetic circuits) computing
homomorphism polynomials for $k$-cliques require $\Omega(n^{\lceil 2k
/ 3 \rceil})$ time. Durand, Mahajan, Malod, Rugy-Altherre, and
Saurabh \cite{DMMRS14} proved that homomorphism polynomials for
certain pattern families are complete for the class $\VP$, the
algebraic analogue of the class $\P$. This is the only known
polynomial family that is complete for $\VP$ other than the canonical
complete family of universal circuits.
\section{Algorithms for induced paths and cycles}
In this section, we will prove that the time complexity of the induced
subgraph isomorphism problems for paths and cycles are upper bounded
by the circuit complexities of the homomorphism polynomials for
$\overline{P_k}$ and $K_k - P_{k-1}$ respectively. Using this we
derive efficient algorithms for induced subgraph isomorphism problem
for $P_k$ for $k\in \{5,6,7,8,9\}$ and $C_k$ for $k\in \{5,7,9\}$. We
also obtain efficient combinatorial algorithms for the induced
subgraph isomorphism problem for $P_k$ for all $k$ and $C_k$ when $k$
is odd.
The proof has two main steps: First, we show that the induced subgraph
isomorphism polynomials for these patterns are reducible to the
aforementioned homomorphism polynomials
(Lemmas~\ref{lem:path-ind-sub-full}, \ref{lem:path-hom-full},
\ref{lem:cycle-ind-sub-full}, \ref{lem:cycle-hom-full}). Then, we
prove that these homomorphism polynomials can be computed efficiently
(Theorems~\ref{thm:path-algos-full} and \ref{thm:cycle-algos-full}).
\begin{lemma}
$\ind{\overline{P_k}} = \sub{\overline{P_k}} \pmod 2$ for $k\geq 4$.
\label{lem:path-ind-sub-full}
\end{lemma}
\begin{proof}
We will prove that for any proper super-graph $H$ of
$\overline{P_k}$, the number $\nsub{\overline{P_k}}{H}$ is even.
Observe that this number is the same as the number of ways to extend
a proper labelled subgraph of $P_k$ to some labelled $P_k$. Let $H$
be an arbitrary proper subgraph of $P_k$. Let $2\leq \ell \leq k$ be
the number of connected components in $H$ out of which
$0\leq s \leq \ell$ of them consists only of a single vertex. Then
the number of ways to extend $H$ to a $P_k$ is
$\ell!2^{\ell-s}/2$. We can extend $H$ to a $P_k$ by ordering the
connected components from left to right and then connecting the
endpoints from left to right. There are $\ell!$ ways to order $\ell$
components and $2$ ways to place all components with more than one
vertex. Out of these, a configuration and its reverse will lead to
the same labelled $P_k$. Since $\ell \geq 2$, this number is even if
$\ell > s$. Otherwise, this number is $k!/2$ because $\ell = s$
implies that there are $k$ components. This is even when $k\geq
4$. We conclude that
$\ind{\overline{P_k}} = \sub{\overline{P_k}} \pmod 2$.
\end{proof}
\begin{lemma}
$\sub{\overline{P_k}} \preceq \ghom{\overline{P_k}}$
\label{lem:path-hom-full}
\end{lemma}
\begin{proof}
Let $f = \sub{\overline{P_k}}$ and $g = \ghom{\overline{P_k}}$. We
fix the labelling of $\overline{P_k}$ where the vertices of the
complementary $P_k$ are labelled $1,2,\dotsc,k$ with $1$ and $k$ as
the endpoints and for every other vertex $i$, the neighbours are
$i-1$ and $i+1$. Start with $g$ over the vertex set $[n]\times [k]$
and use the following substitution.
\setcounter{equation}{0}
\begin{align}
\sigma(z_{a,(v,a)}) &= z_a\\
\sigma(z_{a,(v,b)}) &= z_a^2 \text{, if $a\neq b$}\\
\sigma(y_{(v, a)}) &= y_v\\
\sigma(x_{\{(u, p), (v, q)\}}) &= 0 \text{, if $\{p,q\}\not\in E(\overline{P_k})$ or if $p=1$ and $q=k$ and $u>v$}\\
\sigma(x_{\{(u, p), (v, q)\}}) &= x_{\{u, v\}} \text{, otherwise}
\end{align}
The resulting polynomial $g'$ satisfies
$\ml{g'} = z_1\ldots z_k\ml{f_n}$ as required. The reduction works
because there is exactly one non-trivial automorphism for
$\overline{P_k}$ and that automorphism maps $1$ to $k$. The monomial
corresponding to one of these automorphisms become $0$ because of
$u>v$ where $u$ has colour 1 and $v$ has colour $k$.
\end{proof}
\begin{theorem}
If $\ghom{\overline{P_k}}$ can be computed by circuits of size
$n^{f(k)}$, then there is an $O(n^{f(k)})$ time algorithm for the
induced subgraph isomorphism problem for $P_k$ on $n$-vertex graphs.
\label{thm:path-main-theorem-full}
\end{theorem}
\begin{theorem}
\label{thm:AlgosForPaths-full}
The following algorithms exist
\begin{enumerate}
\item An $O(n^\omega)$-time algorithm for induced subgraph
isomorphism problem for $P_5$ in $n$-vertex graphs.
\item An $O(n^{\omega(2,1,1)})$-time algorithm for induced subgraph
isomorphism problem for $P_6$ in $n$-vertex graphs.
\item An $O(n^{k-2})$-time combinatorial algorithm for induced subgraph
isomorphism problem for $P_k$ in $n$-vertex graphs.
\item An $O(n^{k-2})$-time deterministic combinatorial algorithm for
computing the parity of the number of induced subgraphs isomorphic
to $P_k$ in $n$-vertex graphs.
\end{enumerate}
\label{thm:path-algos-full}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item We describe how to compute $\ghom{\overline{P_5}}$ using
arithmetic circuits of size $O(n^\omega)$. We start by defining
the following matrices.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node (1) at (0, 0) {1};
\node (2) at (2, 0) {2};
\node (5) at (2, 1) {5};
\node (3) at (1, -1) {3};
\node (4) at (0, 1) {4};
\draw (1) -- (2);
\draw (1) -- (4);
\draw (1) -- (3);
\draw (2) -- (3);
\draw (2) -- (5);
\draw (4) -- (5);
\end{tikzpicture}
\end{center}
\caption{A labelled $\overline{P_5}$}
\label{fig:p5bar-full}
\end{figure}
\begin{align*}
A^{n\times n}_{i, j} &= x_{\{i, j\}}, i\neq j\\
B^{n\times n}_{i, i} &= y_i z_{3, i}\\
C^{n\times n}_{i, i} &= y_i z_{4, i}\\
D^{n\times n}_{i, i} &= y_i z_{5, i}
\end{align*}
Consider the labelled $\overline{P_5}$ in Figure~\ref{fig:p5bar-full}.
Then we can write
\[
\ghom{\overline{P_5}} = \sum_{i, j\in[n], i\neq j} z_{1, i} z_{2,
j} x_{\{i, j\}} y_i y_j {(ABA)}_{i, j} {(ACADA)}_{i,
j}
\]
Clearly, this can be implemented using $O(n^\omega)$ size
circuits. We will now prove that this circuit correctly computes the
polynomial $\ghom{\overline{P_5}}$. Consider a homomorphism
$\phi: j\mapsto i_j$. Consider the monomial generated by
$i = i_1, j = i_2$ in the outer sum, the monomial
$A_{i_1, i_3}B_{i_3, i_3}A_{i_3, i_2}$ in the product
${(ABA)}_{i_1, i_2}$, and the monomial
$A_{i_1, i_4}C_{i_4, i_4}$ $A_{i_4, i_5}D_{i_5,i_5}A_{i_5,i_2}$ in the
product ${(ACADA)}_{i_1, i_2}$. This monomial corresponds to the
homomorphism $\phi$ and one can observe that this is the only way to
generate this monomial. On the other hand, any monomial in the
computed polynomial is generated as described above and therefore
corresponds to a homomorphism.
\item We show how to compute $\ghom{\overline{P_6}}\,$ using
arithmetic circuits of size $O(n^{\omega(2, 1, 1)})$. We define
the following matrices.
\begin{align*}
A^{n\times n^2}_{i, (j, k)} &=
z_{2, i}z_{1, j}z_{6, k} y_i y_j y_k
x_{\{(2, i), ((1, j), (6, k))\}} x_{\{j, k\}} x_{\{k, i\}},
j\neq k, i\neq k\\
B^{n^2\times n}_{(j, k), \ell} &=
z_{5,\ell} y_\ell x_{\{((1, j), (6, k)), (5, \ell)\}}
x_{\{j, \ell\}}, j\neq k, j\neq \ell\\
C^{n\times n}_{\ell, i} &= x_{\{\ell, i\}}, \ell\neq i\\
D^{n^2\times n}_{(j, k), p} &= y_p z_{3, p} x_{\{((1, j), (6, k)), (3, p)\}},
j\neq k, j\neq p, k\neq p\\
E^{n\times n}_{p,\ell} &= x_{\{p, \ell\}}, p\neq \ell\\
F^{n^2\times n}_{(j, k), q} &= y_q z_{4, q} x_{\{((1, j), (6, k)), (4, q)\}},
j\neq k, j\neq q, k\neq q\\
G^{n\times n}_{q,i} &= x_{\{q, i\}}, q\neq i
\end{align*}
Compute the matrix products $ABC$, $DE$, and $FG$. The output of
the circuit is $\sum_i {(ABC)}_{i, i}$ after substituting for the
variables as follows. Replace each $x_{\{((1, j), (6,
k)), (5, \ell)\}}$ with $DE_{(j, k), \ell}$ and each
$x_{\{((1, j), (6, k)), (2, i)\}}$ with $FG_{(j, k),
i}$. Replace each $x_{\{((1, j), (6, k)), (3, p)\}}$
with $x_{\{j, p\}} x_{\{k, p\}}$ and each
$x_{\{((1, j), (6, k)), (4, q)\}}$ with $x_{\{j, q\}} x_{\{k,
q\}}$.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node (1) at (0, 0) {1};
\node (6) at (0, 2) {6};
\node (2) at (1, -0.5) {2};
\node (5) at (1, 2.5) {5};
\node (3) at (2, 0.5) {3};
\node (4) at (2, 1.5) {4};
\draw (1) -- (6);
\draw (1) -- (5);
\draw (1) -- (4);
\draw (1) -- (3);
\draw (2) -- (4);
\draw (2) -- (5);
\draw (2) -- (6);
\draw (3) -- (5);
\draw (3) -- (6);
\draw (4) -- (6);
\end{tikzpicture}
\end{center}
\caption{A labelled $\overline{P_6}$}
\label{fig:p6bar-full}
\end{figure}
Consider the labelling of $\overline{P_6}$ in
Figure~\ref{fig:p6bar-full}. After substituting for all variables as
mentioned above, the monomials of ${(ABC)}_{i, i}$ correspond to
homomorphisms from this labelled $\overline{P_6}$ to $K_n$ that
maps vertex $2$ to $i$. Therefore, the circuit correctly computes
$\ghom{\overline{P_6}}$.
\item We observe that $\tw{\overline{P_k}} = k-3$ and therefore
using Theorem~\ref{thm:diaz-full}, we can compute
$\ghom{\overline{P_k}}$ using $O(n^{k-2})$ size circuits.
\item Consider the substitution in the proof of
Lemma~\ref{lem:path-hom-full} and replace rules (1) and (2) by the
following rules.
\begin{align*}
\sigma(z_{a, (v, a)}) &= 1 \tag{1'}\\
\sigma(z_{a, (v, b)}) &= 0 \tag{2'}
\end{align*}
The multilinear part of the resulting polynomial $f$ is the same
as $\sub{\overline{P_k}}$ and hence has degree-$k$. Therefore, we
only have to compute the parity of the sum of coefficients of the
multilinear terms of $f(G)$. By Theorem~\ref{thm:multiparity-full},
this can be done in $O(n^{k-2})$ time.
\end{enumerate}
\end{proof}
We remark that by computing homomorphism polynomials for
$\overline{P_{k}}$ for $k = 7, 8, 9$ using small-size circuits, we can
obtain the following algorithms for the induced subgraph isomorphism
problem for paths: An $O(n^{2\omega})$ time algorithm for $P_7$, an
$O(n^{\omega(3, 2, 2)})$ time algorithm for $P_8$, and an
$O(n^{\omega(3,3,2)})$ time algorithm for $P_9$. All these algorithms
are faster than the corresponding algorithms for $k$-cliques.
\begin{lemma}
$\ind{\overline{C_k}} = \sub{\overline{C_k}} + \sub{\overline{P_k}}
+ \sub{K_k - P_{k-1}} \pmod 2$ for $k\geq 5$.
\label{lem:cycle-ind-sub-full}
\end{lemma}
\begin{proof}
We claim that the only proper supergraphs of $\overline{C_k}$
containing it an odd number of times are $\overline{P_k}$ and
$K_k - P_{k-1}$. There is exactly one way to extend a $P_k$ or a
$P_{k-1} + v$ to a $C_k$. Let $H$ be a proper subgraph of $C_k$
other than these two graphs. Assume that $H$ has
$2 \leq \ell \leq k$ connected components out of which
$0\leq s\leq \ell$ are single vertices. Then there are
$m = \ell! 2^{\ell-s}/2\ell$ ways to extend $H$ to $C_k$. If
$\ell > s$, then $m$ is even because $(\ell-1)!$ is even when
$\ell\geq 3$ and when $\ell = 2$ the number $s$ is $0$ and $m =
2$. If $\ell = 2$ and $s = 1$, then $H = P_{k-1} + v$. If
$\ell = s$, then $m = \ell!/2\ell = (\ell-1)!/2$. But $\ell = s$
implies that $\ell = k$ and therefore $m = (k-1)!/2$ which is even
when $k\geq 5$.
\end{proof}
\begin{lemma}
\begin{enumerate}
\item $\sub{\overline{C_k}} \preceq \ghom{K_k - P_{k-1}} \pmod 2$
for odd $k\geq 5$.
\item $\sub{\overline{P_k}} \preceq \ghom{K_k - P_{k-1}}$ for
$k\geq 5$.
\item $\sub{K_k - P_{k-1}} \preceq \ghom{K_k - P_{k-1}}$ for
$k\geq 5$.
\item $\ind{\overline{C_k}} \preceq \ghom{K_k - P_{k-1}} \pmod 2$ for
odd $k\geq 5$.
\end{enumerate}
\label{lem:cycle-hom-full}
\end{lemma}
\begin{proof}
We start with $\ghom{K_k - P_{k-1}}$ over the vertex set
$[n]\times [k]$ in all cases and apply the following substitutions.
\begin{enumerate}
\item Fix the labelling of $\overline{C_k}$ where the complementary
$C_k$ is labelled $1,\dotsc,k$ such that the vertex $1$ has
neighbours $2$ and $k$ and $k$ has neighbours $1$ and $k-1$ and
every other vertex $i$ has $i+1$ and $i-1$ as its neighours. The
crucial observation is that $\overline{C_k}$ has $2k$
automorphisms and if we only select automorphisms where the label
of the vertex coloured $1$ is strictly less than the label of the
vertex coloured $3$, then we select exactly $k$
automorphisms. This allows us to compute a polynomial family $h$
such that $k.\sub{\overline{C_k}} \preceq h$ and
$k.\sub{\overline{C_k}} = \sub{\overline{C_k}} \pmod 2$.
\setcounter{equation}{0}
\begin{align}
\sigma_1(z_{a,(v,a)}) &= z_a\\
\sigma_1(z_{a,(v,b)}) &= z_a^2 \text{, if $a\neq b$}\\
\sigma_1(y_{(v, a)}) &= y_v\\
\sigma_1(x_{\{(u, p), (v, q)\}}) &= 0 \text{, if $p=1$ and $q=3$ and $u>v$}\\
\sigma_1(x_{\{(u, p), (v, q)\}}) &= 1 \text{, if $p=1$ and $q=2$ or $p=1$ and $q=k$}\\
\sigma_1(x_{\{(u, p), (v, q)\}}) &= x_{\{u, v\}} \text{, otherwise}
\end{align}
\item
Fix the labelling of $\overline{P_k}$ where the complementary
$P_k$ is
\tikz[baseline=-\the\dimexpr\fontdimen22\textfont2\relax,node distance=0.5cm]{\node
(1) {1};\node[right of=1] (2) {2};\node[right of=2] (dots) {$\dotsm$};\node[right of=dots] (k) {$k$};
\draw (1) -- (2) -- (dots) -- (k);}.
\setcounter{equation}{0}
\begin{align}
\sigma_2(z_{a,(v,a)}) &= z_a\\
\sigma_2(z_{a,(v,b)}) &= z_a^2 \text{, if $a\neq b$}\\
\sigma_2(y_{(v, a)}) &= y_v\\
\sigma_2(x_{\{(u, p), (v, q)\}}) &= 0 \text{, if $p=1$ and $q=k$ and $u>v$}\\
\sigma_2(x_{\{(u, p), (v, q)\}}) &= 1 \text{, if $p=1$ and $q=2$}\\
\sigma_2(x_{\{(u, p), (v, q)\}}) &= x_{\{u, v\}} \text{, otherwise}
\end{align}
\item Fix the labelling of $K_k - P_{k-1}$ where the complementary
$P_{k-1} + v$ is
\tikz[baseline=-\the\dimexpr\fontdimen22\textfont2\relax,node
distance=0.5cm]{\node (1) {1};\node[right of=1] (2) {2};\node[right
of=2] (3) {3};\node[right of=3] (dots) {$\dotsm$};\node[right
of=dots] (k) {$k$}; \draw (2) -- (3) -- (dots) -- (k);}.
\setcounter{equation}{0}
\begin{align}
\sigma_3(z_{a,(v,a)}) &= z_a\\
\sigma_3(z_{a,(v,b)}) &= z_a^2 \text{, if $a\neq b$}\\
\sigma_3(y_{(v, a)}) &= y_v\\
\sigma_3(x_{\{(u, p), (v, q)\}}) &= 0 \text{, if $p=2$ and $q=k$ and $u>v$}\\
\sigma_3(x_{\{(u, p), (v, q)\}}) &= x_{\{u, v\}} \text{, otherwise}
\end{align}
\item We prove that
$k\sub{\overline{C_k}} + \sub{\overline{P_k}} + \sub{K_k -
P_{k-1}} \preceq \ghom{K_k - P_{k-1}}$. Start with
$\ghom{K_k - P_{k-1}}$ over the vertex set
$[n]\times [k] \times [3]$ and apply the following substitution.
\setcounter{equation}{0}
\begin{align}
\sigma(z_{a,(v,b,i)}) &= \sigma_i(z_{a, (v, b)}))\\
\sigma(y_{(v, a, i)}) &= \sigma_i(y_{(v, a)})\\
\sigma(x_{\{(u, p, i), (v, q, j)\}}) &= 0 \text{, if $i\neq j$}\\
\sigma(x_{\{(u, p, i), (v, q, i)\}}) &= \sigma_i(x_{\{(u, p), (v, q)\}}) \text{, otherwise}
\end{align}
Rule~3 ensures that only the monomials where every vertex is
indexed by the same element in $[3]$ survive. The other rules
ensure that any monomial $m$ indexed by $i\in[3]$ are mapped to
$\sigma_i(m')$, where $m'$ is the same as $m$ but with $i$
removed.
\end{enumerate}
The proof of correctness of these reductions is the same as the
argument in Theorem~\ref{thm:supergraphs-full}. In addition, the
condition that $u > v$ when $u$ is coloured $1$ and $v$ is coloured
$k$ rules out one out of two automorphisms for $\overline{P_k}$ in
part~2 and the condition that $u > v$ when $u$ is coloured $2$ and
$v$ is coloured $k$ rules out one out of two automorphisms for
$K_k - P_{k-1}$ in part~3.
\end{proof}
\begin{theorem}
If $\ghom{K_k - P_{k-1}}$ can be computed by circuits of size
$n^{f(k)}$, then there is an $O(n^{f(k)})$ time algorithm for
induced subgraph isomorphism problem for $C_k$ on $n$-vertex graphs
for odd $k\geq 5$.
\label{thm:cycle-main-theorem-full}
\end{theorem}
\begin{theorem}
\label{thm:AlgosForCycles-full}
The following algorithms exist
\begin{enumerate}
\item An $O(n^\omega)$-time algorithm for induced subgraph
isomorphism problem for $C_5$ in $n$-vertex graphs.
\item An $O(n^{k-2})$-time combinatorial algorithm for induced
subgraph isomorphism problem for $C_k$ in $n$-vertex graphs, where
$k\geq 5$ is odd.
\item An $O(n^{k-2})$-time deterministic combinatorial algorithm for
computing the parity of the number of induced subgraphs isomorphic
to $C_k$ in $n$-vertex graphs, where $k\geq 5$ is odd.
\end{enumerate}
\label{thm:cycle-algos-full}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item We describe how to compute $\ghom{K_5 - P_4}$ using
arithmetic circuits of size $O(n^\omega)$. We start by defining the
following matrices.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node (1) at (0, 0) {1};
\node (2) at (2, 0) {2};
\node (5) at (3, -1) {5};
\node (3) at (1, -2) {3};
\node (4) at (-1, -1) {4};
\draw (1) -- (2);
\draw (1) -- (3);
\draw (2) -- (3);
\draw (1) -- (4);
\draw (3) -- (4);
\draw (2) -- (5);
\draw (3) -- (5);
\end{tikzpicture}
\end{center}
\caption{A labelled $K_5 - P_4$}
\label{fig:k5-p4-full}
\end{figure}
\begin{align*}
A^{n\times n}_{i, j} &= x_{\{(i, 1), (j, 3)\}}, i\neq j\\
E^{n\times n}_{i, j} &= x_{\{(i, 3), (j, 2)\}}, i\neq j\\
F^{n\times n}_{i, j} &= x_{\{i, j\}}, i\neq j\\
B^{n\times n}_{i, i} &= y_i z_{3, i}\\
C^{n\times n}_{i, i} &= y_i z_{4, i}\\
D^{n\times n}_{i, i} &= y_i z_{5, i}
\end{align*}
Consider the labelled $K_5 - P_4$ in
Figure~\ref{fig:k5-p4-full}. Compute the matrix products $FCF$, $FDF$,
and $ABE$. Compute the polynomial
$\sum_{i, j\in [n], i\neq j} z_{1, i} z_{2, j} y_i y_j x_{\{i,
j\}} {(ABE)}_{i, j}$ and replace $x_{\{(i, 1), (j, 3)\}}$ with
${(FCF)}_{i, j}$ and replace $x_{\{(i, 3), (j, 2)\}}$ with
${(FDF)}_{i, j}$. It is easy to see that the resulting polynomial
is $\ghom{K_5 - P_4}$ for this labelled $K_5 - P_4$ and the
circuit has size $O(n^\omega)$.
\item $\tw{K_k - P_{k-1}} = k-3$.
\item The proof is similar to the proof of Part~4 of
Theorem~\ref{thm:path-algos-full}.
\end{enumerate}
\end{proof}
We remark that by computing homomorphism polynomials for $K_k -
P_{k-1}$ for $k = 7, 9$ using small-size circuits, we can obtain an
$O(n^{2\omega})$ time algorithm for induced subgraph isomorphism for
$C_7$ and an $O(n^{\omega(3,3,2)})$ time algorithm for induced
subgraph isomorphism for $C_9$. These algorithms are faster than the
corresponding algorithms for $k$-cliques.
\section{Preliminaries}
For a polynomial $f$, we use $\pdeg{f}$ to denote the degree of $f$. A
monomial is called multilinear, if every variable in it has degree at
most one. We use $\ml{f}$ to denote the multilinear part of $f$, that
is, the sum of all multilinear monomials in $f$. An arithmetic circuit
computing a polynomial $P\in K[x_1,\dotsc,x_n]$ is a circuit with $+$,
$\times$ gates where the input gates are labelled by variables or
constants from the underlying field and one gate is designated as the
output gate. The size of an arithmetic circuit is the number of wires
in the circuit. For indeterminates $x_1,\dots,x_n$ and a set
$S = \{s_1, \dotsc, s_p\} \subseteq \{1,\dots,n\}$ of indices, we
write $x_S$ to denote the product $x_{s_1}\dotsm x_{s_p}$.
An induced subgraph isomorphism from $H$ to $G$ is an injective
function $\phi : V(H) \overset{\mathit{ind}}{\mapsto} V(G)$ such that
$\{u, v\}\in E(H) \iff \{\phi(u), \phi(v)\}\in E(G)$. Any function
from $V(H)$ to $V(G)$ can be extended to unordered pairs of vertices
of $H$ as $\phi(\{u, v\}) = \{\phi(u), \phi(v)\}$. A subgraph
isomorphism from $H$ to $G$ is an injective function
$\phi : V(H) \overset{\mathit{sub}}{\mapsto} V(G)$ such that
$\{u, v\}\in E(H) \implies \{\phi(u), \phi(v)\}\in E(G)$. Two subgraph
isomorphisms or induced subgraph isomorphisms are considered different
only if the set of edges in the image are different. A graph
homomorphism from $H$ to $G$ is a function $\phi: V(H) \overset{\mathit{hom}}{\mapsto} V(G)$
such that $\{u, v\}\in E(H) \implies \{\phi(u), \phi(v)\}\in
E(G)$. Unlike isomorphisms, we consider two distinct functions that
yield the same set of edges in the image as distinct graph
homomorphisms. We define $\phi(S) = \{\phi(s) : s\in S\}$.
We write $H\sqsubseteq H'$ ($H\sqsupseteq H'$) to specify that $H$ is a subgraph
(supergraph) of $H'$. The number $\tw{H}$ stands for the treewidth of
$H$. The graph $K_n$ is the complete graph on $n$ vertices labelled
using $[n]$. We use the fact that $\naut{H} = 1$ for almost all graphs
in many of our results. In this paper, we will frequently consider
graphs where vertices are labelled by tuples. A vertex $(i, p)$ is
said to have \emph{label} $i$ and \emph{colour} $p$. An edge $\{(i_1,
p_1), (i_2, p_2)\}$ has \emph{label} $\{i_1, i_2\}$ and \emph{colour}
$\{p_1, p_2\}$. We will sometimes write this edge as $(\{i_1, i_2\},
\{p_1, p_2\})$. Note that both $\{(i_1, p_1), (i_2, p_2)\}$ and
$\{(i_2, p_1), (i_1, p_2)\}$ are written as $(\{i_1, i_2\}, \{p_1,
p_2\})$. But the context should make it clear which edge is being
rewritten.
\section{Reductions between patterns}
The following proposition is analogous to the obvious fact that the
complexity of the induced subgraph isomorphism problem is the same for
any pattern and its complement.
\begin{proposition}
$\ind{H}\preceq \ind{\overline{H}}$ for all graphs $H$.
\end{proposition}
\begin{proof}
Use the substitution that maps $x_e$ to $1-x_e$ for any edge
variable $x_e$ and maps any vertex variable to itself.
\end{proof}
It is known that $\naut{H} = 1$ for almost all graphs $H$. Therefore,
the following proposition can be interpreted as stating that the
homomorphism polynomial is harder than the subgraph isomorphism
polynomial for almost all pattern graphs $H$. This is used in
\cite{FominLRSR12} to obtain algorithms for subgraph isomorphism
problems.
\begin{proposition}
$\naut{H}\sub{H} \preceq \ghom{H}$ for all graphs $H$.
\end{proposition}
\begin{proof}
Let $H$ be a $k$ vertex graph labelled using $[k]$. Use the
substitution $\sigma(z_{a, v}) = u_a$ for all $a\in V(H), v\in V(G)$
and $\sigma(w) = w$ for all the other variables $w$ in $\ghom{H}$
over the vertex set $[n]$. We have $\naut{H}.u_{[k]}.\ml{\sub{H}} =
\ml{\sigma(\ghom{H})} = \sigma(\ml{\ghom{H}})$. Consider an
arbitrary automorphism $\phi$ of $H$. For every monomial $m =
y_{v_1}\dotso y_{v_k}x_{e_1}\dotso x_{e_\ell}$ in $\sub{H}$, there
are exactly $\naut{H}$ monomials $m_{\phi} = z_{(\phi(1),
v_1)}\dotso z_{(\phi(k), v_k)}y_{v_1}\dotso y_{v_k}x_{e_1}\dotso
x_{e_\ell}$ in $\ghom{H}$ that satisfy $\sigma(m_\phi) =
u_{[k]}m$. This proves Properties~1 and 2 of the reduction. It is
easy to see that the reduction satisfies the other properties too.
\end{proof}
Intuitively, the subgraph isomorphism problem should become harder
when the pattern graph becomes larger. However, it is not known
whether this is the case. Nevertheless, we can show this hardness
result holds for subgraph isomorphism polynomials for almost all
pattern graphs.
\begin{theorem}
If $H\sqsubseteq H'$, then $\naut{H}\sub{H} \preceq \sub{H'}$.
\label{thm:supergraphs-full}
\end{theorem}
\begin{proof}
Let $|V(H)| = k$ and $|V(H')| = k + \ell$ for some $\ell \ge
0$. Choose a labelling $L$ of the vertices of $H'$ such that the
vertices of an $H$ in $H'$ are labelled $1,\dotsc, k$. Consider the
polynomial $\sub{H'}$ over the vertex set $([n]\times [k]) \cup
\{(n+i, k+i) : 1\le i\le \ell\}$. Substitute for the variables as
follows:
\setcounter{equation}{0}
\begin{align}
\sigma(y_{(i, p)}) &= \begin{cases}
y_i u_p, &\text{ for all } i\in[n], p\in[k]\\
u_p, &\text{ otherwise}
\end{cases}\\
\sigma(x_{\{(i_1, p_1), (i_2, p_2)\}})
&=
\begin{cases}
x_{\{i_1, i_2\}} & \quad \text{if } \{p_1, p_2\}\in E(H)\\
1 & \quad \text{if } \{p_1, p_2\}\in E(H')\setminus E(H)\\
0 & \quad \text{otherwise}
\end{cases}
\end{align}
We say that a monomial in $\sub{H'}$ \emph{survives} if the monomial
does not become non-multilinear or 0 after the substitution. First,
we will prove that all surviving monomials correspond to
$H'$-subgraphs where the labels and colours of vertices are
different and the colours of edges are the same as in the labelling
$L$. Rule~1 ensures that the colours and labels of all vertices in
the surviving monomials are different. Rule~2 ensures that there is
a one-to-one correspondence between the edges $\{p_1, p_2\}$ in the
labelling $L$ and the edge variables
$x_{\{(i_1, p_1), (i_2, p_2)\}}$. To see this, observe that each
monomial in $\sub{H'}$ has $|E(H')|$ edge variables. Since all
vertices in a surviving monomial have different colours, all edges
in the monomial must have different colours. Since any edge variable
that has a colour not in the labelling $L$ is set to $0$, the
colours of edges must be in one-to-one correspondence with the edges
in the labelling $L$. This proves the all surviving monomials are of
the form
$y_{(u_1, 1)}\dotsm y_{(u_k, k)} (\prod_i y_{(n+i, k+i)}) x_{(e_1,
q_1)} \dotsm x_{(e_m, q_m)} w$ for $u_1, \dotsc ,u_k\in [n]$,
where $w$ is the product of edge variables with colour $\{p, q\}$
such that $\{p, q\}$ is an edge in $H'$ but not in $H$ in the
labelling $L$, $u_1, \dotsc, u_k$ are all different, and
$q_1, \dotsc ,q_m$ are edges in $H$ in the labelling $L$. Note that
the product $w$ is determined uniquely by $u_1,\dotsc,u_k$.
We claim that for each monomial $y_{S} x_{T}$ in $\sub{H}$ over the
vertex set $[n]$ there are $\naut{H}$ monomials $y_{S} x_{T}
u_{[k]}$ in $\sigma(\sub{H'})$. Consider an arbitrary monomial $y_S
x_T = y_{v_1}\dotsm y_{v_k} x_{e_1}\dotsm x_{e_m}$ in $\sub{H}$
where $m = |E(H)|$. The monomials in $\sub{H'}$ that yield $y_S x_T
u_{[k+\ell]}$ after the substitution are exactly the monomials $y_{(w_1,
1)}\dotsm y_{(w_k, k)} (\prod_i y_{(n+i, k+i)}) x_{(e_1', q_1)}
\dotsm x_{(e_m', q_m)} w$ where $w$ is the product of edge variables
with colour $\{p, q\}$ such that $\{p, q\}$ is an edge in $H'$ but
not in $H$ in the labelling $L$, $\{w_1, \dotsc, w_k\} =
\{v_1,\dotsc, v_k\}$, and $\{e_1, \dotsc ,e_m\} = \{e_1',\dotsc,
e_m'\}$. But this monomial corresponds to the automorphism $\phi :
v_i\mapsto w_i$. Since $w$ is uniquely determined given
$w_1,\dotsc,w_k$, the number of such monomials is $\naut{H}$. Also,
each surviving monomial yields a monomial in $\sub{H}$.
Additionally, each non-multilinear term in the polynomial obtained
after the substitution contains at least one vertex or other
variable with degree more than one. This proves the theorem.
\end{proof}
The following theorem states that the induced subgraph isomorphism
polynomial is harder than the subgraph isomorphism polynomial for
almost all graphs.
\begin{theorem}
\label{thm:inducedHarderThanNonInduced-full}
$\naut{H}\sub{H} \preceq \ind{H}$ for all graphs $H$.
\end{theorem}
\begin{proof}
Observe that $\ind{H} = \sub{H} + \sum_{H' \sqsupset H}
a_{H'}\sub{H'}$. Let $k$ be the number of vertices in $H$ and fix
some labelling of $H$ using $[k]$. Now consider the polynomial
$\ind{H}$ over the vertex set $[n] \times [k]$ and apply the
following substitution.
\setcounter{equation}{0}
\begin{align}
\sigma(y_{(i, p)}) &= y_iu_p\\
\sigma(x_{\{(i_1, p_1), (i_2, p_2)\}}) &=
\begin{cases}
x_{\{i_1, i_2\}} & \text{ if $\{p_1, p_2\}\in E(H)$}\\
0 & \text{ otherwise}
\end{cases}
\end{align}
Now observe that any monomial in $\sub{H'}$ for $H' \sqsupset H$ must
vanish because it will have at least one more edge than $H$. By the
same argument as in the proof of Theorem~\ref{thm:supergraphs-full}, we
conclude that there are exactly $\naut{H}$ monomials in $\sub{H}$
over $[n] \times [k]$ that yield the monomial $y_Sx_Tu_{[k]}$ after
the substitution for any monomial $y_Sx_T$ in $\sub{H}$ over $[n]$.
\end{proof}
We now prove the analogue of Theorem~{\ref{thm:np-universal-full}} in
\cite{NP85} which states that $k$-clique is harder than any other
$k$-vertex pattern graph.
\begin{theorem}
For any $k$-vertex graph $H$, $\ind{H}\preceq \ind{K_k}$.
\label{thm:ind_graph_to_clique-full}
\end{theorem}
\begin{proof}
Fix a canonical labelling $L$ of the graph $H$ using $[k]$. Let
$q_1, \dotsc ,q_\ell$ be the edges in the canonical labelling $L$
and let $q_{\ell+1}, \dotsc , q_m$ be the non-edges in $L$ where
$\ell$ is the number of edges in $H$ and $m = {\binom{k}{2}}$. Let
$S$ be the set of distinct labellings of $H$ using $[k]$. Associate
all labellings $L'\in S$ with a permutation $\phi$ such that
applying $\phi$ to an $H$ labelled $L'$ yields an $H$ labelled
$L$. Let $P$ be the set of all such permutations. For example, there
are three distinct labellings for $P_3$: $L =$ 1 -- 2 -- 3, 1 -- 3
-- 2, and 2 -- 1 -- 3 with associated permutations $(1)(2)(3)$,
$(1)(2 3)$, and $(1 2)(3)$ (Note that the these permutations are not
unique if the graph has non-trivial automorphisms). Apply the
following substitution to $\ind{K_k}$ over the vertex set
$[n]\times [k]\times P$:
\setcounter{equation}{0}
\begin{align}
\sigma(y_{(v, p, \phi)}) &= y_vu_p\\
\sigma(x_{\{(v_1, p_1,\phi), (v_2, p_2,\phi')\}}) &= 0 \text{ if $\phi\neq \phi'$ or $p_1 = p_2$ or $v_1 = v_2$}\\
\sigma(x_{\{(v_1, p_1,\phi), (v_2, p_2,\phi)\}}) &= 0
\text{ if $\phi^{-1}(p_1)< \phi^{-1}(p_2)$ and $v_1 > v_2$}\\
\sigma(x_{\{(v_1, p_1,\phi), (v_2, p_2,\phi)\}}) &=
\begin{cases}
x_{\{v_1, v_2\}} & \text{ if $\{p_1, p_2\}\in E(L)$}\\
1 - x_{\{v_1, v_2\}} & \text{ if $\{p_1, p_2\}\not\in E(L)$}\\
\end{cases}
\end{align}
The first two rules ensure that in any surviving monomial, the
labels and colours of all vertices are different and all vertices
are indexed by the same permutation.
We can extend the correspondence between labellings of $H$ and
permutations to arbitrary labellings (as opposed to labellings using
$[k]$). Given a labelling of $H$ using $v_1 < \dotsb < v_k$, we can
obtain a labelling $L'$ of $H$ using $[k]$ by replacing each $v_i$
by $i$ for all $i$. The permutation associated with the labelling
$M$ is the same as the permutation associated with labelling $L'$.
Consider an arbitrary labelling $M$ of $H$ using
$v_1 < \dotsb < v_k$ where each $v_i\in [n]$. Let $L'\in S$ be the
labelling corresponding to the labelling $M$ such that
$\psi : v_i \mapsto i$ is the permutation that maps $M$ to $L'$. Let
$\phi\in P$ be the permutation associated with $L'$. For
convenience, we denote the edges and non-edges of $M$ by
$e_1,\dotsc,e_m$ such that $e_i = \psi^{-1}(\phi^{-1}(q_i))$ for all
$i$. We will prove that for the term
$t = y_{v_1}\dotsm y_{v_k}x_{e_1}\dotsm x_{e_\ell}(1 -
x_{e_{\ell+1}})\dotsm (1 - x_{e_m})$ in $I(H)$ that encodes $M$,
there is a unique monomial $s$ in $\ind{K_k}$ such that
$\sigma(s) = u_{[k]}t$. The monomial
$s = y_{(v_1, \phi(1), \phi)}\dotsm y_{(v_k, \phi(k), \phi)}x_{(e_1,
q_1, \phi)}\dotsm x_{(e_m, q_m, \phi)}$. First of all, we have to
prove that given that $v_i$ has colour $\phi(i)$, the edges are
coloured such that $e_i$ gets colour $q_i$. Start with an arbitrary
$q_i = (j, k)$. Then,
$e_i = \psi^{-1}((\phi^{-1}(j), \phi^{-1}(k))) = (v_{\phi^{-1}(j)},
v_{\phi^{-1}(k)})$ which has colour $(j, k)$ as required. Also, we
have $\sigma(s)\neq 0$ because if
$\phi^{-1}(\phi(i)) = i < j = \phi^{-1}(\phi(j))$, then $v_i <
v_j$. Given that $\sigma(s)\neq 0$, it is easy to see that
$\sigma(s) = u_{[k]}t$ by applying rules 1 and 4.
Given an arbitrary surviving monomial $r = y_{(v_1, 1, \phi)}\dotsm
y_{(v_k, k, \phi)}\allowbreak x_{(e_1, q_1, \phi)}\dotsm x_{(e_m,
q_m, \phi)}$ in $\ind{K_k}$ such that $\sigma(r) = u_{[k]}w$ for
some $w$, we claim that $w$ encodes a labelling $M$ of $H$ where the
permutation associated with $M$ is $\phi$. It is easy to see that
$w$ encodes some labelling of $H$. Observe that for $r$ to survive,
the vertices $(v_i, i, \phi)$ for all $i$ has to be consistent with
$\phi$, i.e., the vertex coloured $\phi(i)$ must be the
$i^{\text{th}}$ smallest among all $v_j$s by Rule~3. By the
definition of $\phi$, we have $\{i, j\}\in E(L')$ if and only if
$\{\phi(i), \phi(j)\}\in E(L)$. By Rule~4, we also have if
$\{\phi(i), \phi(j)\}\in E(L)$ then $x_{\{v_{\phi(i)},
v_{\phi(j)}\}}$ appears in the term $w$ and otherwise $(1 -
x_{\{v_{\phi(i)}, v_{\phi(j)}\}})$ appears in $w$. In other words,
in the graph encoded by $w$, the $i^{\text{th}}$ smallest and
$j^{\text{th}}$ smallest vertices are connected if and only if the
$i^{\text{th}}$ smallest and $j^{\text{th}}$ smallest vertices are
connected in $L'$. Therefore, the associated permutation is $\phi$
as claimed. We can now prove that $u_{[k]}t$ is uniquely generated
from $s$. Suppose for contradiction that the monomial $s' =
y_{(v_1', 1, \phi')}\dotsm y_{(v_k', k, \phi')} x_{(e_1', q_1,
\phi')}\dotsm x_{(e_m', q_m, \phi')}$ also satisfies $\sigma(s') =
u_{[k]}t$. Then, it must be that $\{v_1',\dotsc, v_k'\} =
\{v_1,\dotsc ,v_k\}$, $\{e_1,\dotsc, e_\ell\} = \{e_1',\dotsc
,e_\ell'\}$, and $\{e_{\ell+1},\dotsc, e_m\} = \{e_{\ell+1}',\dotsc
,e_m'\}$. We know that $\phi = \phi'$ because the permutation in the
monomial must correspond to the labelling encoded by $t$. But, $\phi
= \phi'$ implies $v_i' = v_i$ for all $i$ (Otherwise, the third rule
ensures that at least one edge variable in $s'$ becomes $0$ under
$\sigma$). But, if $v_i' = v_i$ for all $i$, then $e_j = e_j'$ for
all $j$ contradicting $s\neq s'$.
We have proved that $\ml{\sigma(\ind{K_k})} =
u_{[k]}\ind{H}$. Observe that the polynomial obtained after the
substitution cannot contain edge variables of degree more than one
because of Rule~2. It is easy to see that the substitution satisfies
the other properties.
\end{proof}
The theorem below shows that the induced subgraph isomorphism
polynomial for any graph containing a $k$-clique or $k$-independent
set is harder than the $k$-clique polynomial. An analogous hardness
result is known for algorithms, only when the pattern $H$ contains a
$k$-clique (or $k$-independent set) that is disjoint from all other
$k$-cliques (or $k$-independent sets) \cite{FKLLTCS15}.
\begin{theorem}
If $H$ contains a $k$-clique or a $k$-independent set, then
$\ind{K_k} \preceq \ind{H}$.
\label{thm:clique-hard-full}
\end{theorem}
\begin{proof}
We will prove the statement when $H$ contains a $k$-clique. The
other part follows because if $H$ contains a $k$-independent set,
then the graph $\overline{H}$ contains a $k$-clique and $\ind{K_k}
\preceq \ind{\overline{H}} \preceq \ind{H}$.
Fix a labelling of $H$ where the vertices of a $k$-clique are
labelled using $[k]$ and the remaining vertices are labelled
$k+1,\dotsc,k+\ell$. Consider the polynomial $\ind{H}$ over the
vertex set $([n] \times [k]) \cup \{(n+i, k+i) : 1\leq i \leq \ell\}$
and apply the following substitution.
\setcounter{equation}{0}
\begin{align}
\sigma(y_{(i, p)}) &=
\begin{cases}
y_iu_p & \text{ if $i\in [n]$ and $p\in[k]$}\\
u_p & \text{ otherwise}
\end{cases}\\
\sigma(x_{\{(i_1, p_1), (i_2, p_2)\}}) &=
\begin{cases}
x_{\{i_1, i_2\}} & \text{ if $\{p_1, p_2\}\in E(K_k)$ and $p_1 < p_2$ and $i_1 < i_2$}\\
1 & \text{ if $\{p_1, p_2\} \in E(H) \setminus E(K_k)$}\\
0 & \text{ otherwise}
\end{cases}
\end{align}
Consider a $k$-clique on the vertices $i_1,\dotsc,i_k\in[n]$ on an
$n$-vertex graph where $i_1 < \dotsb < i_k$. The monomial in
$\ind{K_k}$ corresponding to this clique is generated uniquely from
the monomial $y_{(i_1, 1)} \dotso y_{(i_k, k)}$$\prod_i y_{(n+i,
k+i)}x_{\{(i_1, 1), (i_2, 2)\}}$$\dotso x_{\{(i_{k-1}, k-1), (i_k,
k)\}} w$ in $\ind{H}$, where $w$ is the product of all edge
variables corresponding to edges in $H$ but not in $K_k$. Note that
Rules~1 and 2 ensure that in any surviving monomial, the labels and
colours of all vertices are distinct and the colours of the edges
must be the same as $E(H)$. The product $w$ is determined by $i_1,
\dotsc, i_k$. This proves that $\ml{\sigma(\ind{H})} =
u_{[k+\ell]}\ml{\ind{K_k}}$. It is easy to verify that the
substitution satisfies the other properties.
\end{proof}
Theorem~\ref{thm:clique-hard-full} is true with $\sub{H}$ or
$\ghom{H}$ instead of $\ind{H}$. In fact, the same proof works for
$\sub{H}$. For $\ghom{H}$, use the substitution in the proof of
Theorem~\ref{thm:clique-hard-full} along with $z_{a, (v, a)} = u_a$
and $z_{a, (v, b)} = u_a^2$ when $a\ne b$ for all homomorphism
variables.
\section{A Motivating Example: Induced-$P_4$ Isomorphism}
In this section, we sketch a one-sided error, randomized $O(n^2)$ time
algorithm for the induced subgraph isomorphism problem for $P_4$ to
illustrate the techniques used to derive algorithms in this paper.
We start by giving an algorithm for the subgraph isomorphism problem
for $P_4$. Consider the following polynomial:
\begin{equation*}
\sub{P_4,n} = \sum_{(p, q, r, s) : p < s} y_p y_q y_r y_s x_{\{p, q\}} x_{\{q, r\}} x_{\{r, s\}}
\end{equation*}
where the summation is over all quadruples over $[n]$ where all four
elements are distinct. Each monomial in the above polynomial
corresponds naturally to a $P_4$ in an $n$-vertex graph. The condition
$p < s$ ensures that each path has exactly one monomial corresponding
to it.
Given an $n$-vertex host graph $G$ and an arithmetic circuit for
$\sub{P_4, n}$, we can construct an arithmetic circuit for the
polynomial $\sub{P_4, n}(G)$ on the $y$ variables obtained by
substituting $x_e = 0$ when $e\not\in E(G)$ and $x_e = 1$ when
$e\in E(G)$. The polynomial $\sub{P_4, n}(G)$ can be written as
$\sum_{X} a_{X} y_{X}$ where the summation is over all four vertex
subsets $X$ of $V(G)$ and $a_X$ is the number of $P_4$s in the induced
subgraph $G[X]$. Therefore, we can decide whether $G$ has a subgraph
isomorphic to $P_4$ by testing whether $\sub{P_4, n}(G)$ is
identically $0$. Since the degree of this polynomial is a constant
$k$, this can be done in time linear in the size of the arithmetic
circuit computing $\sub{P_4, n}$.
However, we do not know how to construct a $O(n^2)$ size arithmetic
circuit for $\sub{P_4, n}$. Instead, we construct a $O(n^2)$ size
arithmetic circuit for the following polynomial called the walk
polynomial:
\begin{equation*}
\ghom{P_4,n} = \sum_{\phi: P_4 \overset{\mathit{hom}}{\mapsto} K_n} \prod_{v\in V(P_4)}
z_{v, \phi(v)} y_{\phi(v)} \prod_{e\in E(P_4)} x_{\phi(e)}
\end{equation*}
This polynomial is also called the homomorphism polynomial for $P_4$
because its terms are in one-to-one correspondence with graph
homomorphisms from $P_4$ to $K_n$. As before, we consider the
polynomial $\ghom{P_4, n}(G)$ obtained by substituting for the $x$
variables appropriately. The crucial observation is that
$\ghom{P_4, n}(G)$ contains a multilinear term if and only if
$\sub{P_4, n}(G)$ is not identically zero. This is because the
multilinear terms of $\ghom{P_4, n}$ correspond to injective
homomorphisms from $P_4$ which in turn correspond to subgraph
isomorphisms from $P_4$. More specifically, each $P_4$ corresponds to
two injective homomorphisms from $P_4$ since $P_4$ has two
automorphisms. Therefore, we can test whether $G$ has a subgraph
isomorphic to $P_4$ by testing whether $\ghom{P_4, n}(G)$ has a
multilinear term. We can construct a $O(n^2)$ size arithmetic circuit
for the polynomial $p_4 = \ghom{P_4, n}$ inductively as follows:
\begin{align*}
p_{1, v} &= y_v, v\in [n]\\
p_{i+1, v} &= \sum_{u\in [n]} p_{i, u} y_v x_{\{u, v\}}, v\in [n], i\geq 1\\
p_4 &= \sum_{v\in [n]} p_{4, v}
\end{align*}
The above construction works for any $k$ and not just $k = 4$. This
method is used by Ryan Williams \cite{Williams09} to obtain an
$O(2^{k}(n+m))$ time algorithm for the subgraph isomorphism problem
for $P_k$.
In fact, the above method works for any pattern graph $H$. Extend the
definitions above to define $\sub{H, n}$ and $\ghom{H, n}$ in the
natural fashion. Then, we can test whether an $n$-vertex graph $G$ has
a subgraph isomorphic to $H$ by testing whether $\sub{H, n}(G)$ is
identically zero which in turn can be done by testing whether
$\ghom{H, n}(G)$ has a multilinear term. Therefore, the complexity of
subgraph isomorphism problem for any pattern $H$ is as easy as
constructing the homomophism polynomial for $H$. This method is used
by Fomin et. al. \cite{FominLRSR12} to obtain efficient algorithms for
subgraph isomorphism problems.
We now turn our attention to the induced subgraph isomorphism problem
for $P_4$. We note that the induced subgraph isomorphism problem for
$P_k$ is much harder than the subgraph isomorphism problem for
$P_k$. The subgraph isomorphism problem for $P_k$ has a linear time
algorithm as seen above but the induced subgraph isomorphism problem
for $P_k$ cannot have $n^{o(k)}$ time algorithms unless
$\class{FPT} = \fptw1$. We start by considering the polynomial:
\begin{equation*}
\ind{P_4,n} = \sum_{(p, q, r, s) : p < s} y_p y_q y_r y_s x_{\{p, q\}} x_{\{q, r\}} x_{\{r, s\}} (1 - x_{\{p, r\}})(1 - x_{\{p, s\}})(1 - x_{\{q, s\}})
\end{equation*}
The polynomial $\ind{P_4, n}(G)$ can be written as $\sum_X y_X$ where
the summation is over all four vertex subsets of $V(G)$ that induces a
$P_4$. Notice that all coefficents are $1$ because there can be at
most $1$ induced-$P_4$ on any four vertex subset. By expanding terms
of the form $1-x_*$ in the above polynomial, we observe that we can
rewrite $\ind{P_4, n}$ as follows:
\begin{equation*}
\ind{P_4, n} = \sub{P_4, n} - 4\sub{C_4, n} - 2\sub{K_3+e, n} +
6\sub{K_4-e, n} + 12\sub{K_4, n}
\end{equation*}
Since the coefficients in $\ind{P_4, n}(G)$ are all $0$ or $1$, it is
sufficient to check whether $\ind{P_4, n}(G) \pmod 2$ is non-zero to
test whether $\ind{P_4, n}(G)$ is non-zero. From the above equation,
we can see that $\ind{P_4, n} = \sub{P_4, n} \pmod 2$. Therefore,
instead of working with $\ind{P_4, n} \pmod 2$, we can work with
$\sub{P_4, n} \pmod 2$. We have already seen that we can use
$\ghom{P_4, n}(G)$ to test whether $\sub{P_4, n}(G)$ is
non-zero. However, this is not sufficient to solve induced subgraph
isomorphism. We want to detect whether $\sub{P_4, n}(G)$ is non-zero
modulo $2$. Therefore, the multilinear terms of $\ghom{P_4, n}(G)$ has
to be in one-to-one correspondence with the terms of
$\sub{P_4, n}(G)$. We have to divide the polynomial $\ghom{P_4, n}(G)$
by $2$ before testing for the existence of multilinear terms modulo
$2$. However, since we are working over a field of characteristic $2$,
this division is not possible. We work around this problem by starting
with $\ghom{P_4, n'}$ for $n'$ slightly larger than $n$ and we show
that this enables the ``division'' by $2$.
The reader may have observed that instead of the homomorphism
polynomial, we could have taken any polynomial $f$ for which the
multilinear terms of $f(G)$ are in one-to-one correspondence with
$\sub{P_4, n}(G)$. This observation leads to the definition of a
notion of reduction between polynomials. Informally, $f\preceq g$ if
detecting multilinear terms in $f(G)$ is as easy as detecting
multilinear terms in $g(G)$. Additionally, for the evaluation $f(G)$
to be well-defined, the polynomial $f$ must have some special
structure. We call such polynomials graph pattern polynomials.
On first glance, it appears hard to generalize this algorithm for
$P_4$ to sparse pattern graphs on an arbitrary number of vertices (For
example, $P_k$) because we have to argue about the coefficients of
many $\sub{*}$ polynomials in the expansion. On the other hand, if we
consider the pattern graph $K_k$, we have $\ind{K_k} = \ghom{K_k}$. In
this paper, we show that for many graph patterns sparser than $K_k$,
the induced subgraph isomorphism problem is as easy as constructing
arithmetic circuits for homomorphism polynomials for those patterns
(or patterns that are only slightly denser).
|
\section*{Acknowledgement}
Shuaa Alharbi is supported by the Saudi Arabian Ministry of Higher Education Doctoral Scholarship and Qassim University in Saudi Arabia. \c{C}i\u{g}dem~Sazak is funded by the Turkey Ministry of National Education. Carl J. Nelson is funded by EPSRC UK (EP/N509668/1).
\subsection{Background}
\subsubsection{Mathematical Morphology and the Top-Hat Transform}
Mathematical morphology operations are a set of non-linear filtering methods, and almost all of them formed through a combination of two basic operators: dilation and erosion.
If $I(\mathbf{p})$ is a grey-scale image and $B(\mathbf{p})$ is structuring element where $\mathbf{p}$ denotes the pixel position $[x,y]^T$ in the 2D images and $[x,y,z]^T$ in the 3D images. Dilation, ($\oplus$) can be defined as the maximum of the points in a weighted neighbourhood determined by the structuring element, and mathematically:
\begin{equation}
(I \oplus B)(\mathbf{p}) = \sup_{\mathbf{x} \in E}[I(\mathbf{x})+B(\mathbf{p}-\mathbf{x})],
\end{equation}
where `$\sup$' is the supremum and $\mathbf{x} \in E$ denotes all points in Euclidean space within the image. Likewise, we mathematically represent erosion ($\ominus$), as the minimum of the points in the neighbourhood determined by the structuring element:
\begin{equation}
(I \ominus B)(\mathbf{p}) = \inf_{\mathbf{x} \in E}[I(\mathbf{x})+B(\mathbf{p}-\mathbf{x})],
\end{equation}
where `$\inf$' is the infimum. The behaviour of dilation is expanding bright areas and reducing dark areas, while erosion is expanding dark regions reducing bright areas~\cite{haralick1987image}.
From these two operators we can define two further commonly used morphological filters:
\begin{align}
\text{opening}: & \quad (I \circ B)(\mathbf{p}) = ((I \ominus B) \oplus B)(\mathbf{p}) \\
\text{closing}: & \quad (I \bullet B)(\mathbf{p}) = ((I \oplus B) \ominus B)(\mathbf{p})
\end{align}
where an opening ($\circ$) will preserve dark features and patterns, suppressing bright features, and a closing ($\bullet$) will preserve bright features whilst suppressing dark patterns.
By comparing the original image and the result of opening or closing, two region extraction operations, which are called top-hat (TH) and bottom-hat (BH) transform, and defined as follows;
\begin{align}
\text{TH}= I(\mathbf{p}) - (I \circ B)(\mathbf{p}) \\
\text{BT}= (I \bullet B)(\mathbf{p})-I(\mathbf{p})
\end{align}
The TH is usually used to extract bright structures, while BT is used to extract dark structures.
\subsubsection{Vesselness and Neuriteness Measurements}
\paragraph{2D Vesselness}\label{sec:2DVes}
One of the most popular Hessian-based approaches that used the eigenvalues of the Hessian to compute the likeliness of an image region to contain vessels or other image ridges~\cite{frangi1998multiscale}. It is computed based on the ratio of eigenvalues of the Hessian matrix as follows:
\begin{equation}
\resizebox{.8\linewidth}{!}{$
V_\sigma =\begin{cases}
0, & \lambda_2>0 \\
\text{exp}\Bigg(-\dfrac{R_\beta^2}{2\beta^2}\Bigg)
\Bigg( 1-\text{exp}\bigg(-\dfrac{S^2}{2c^2}\bigg) \Bigg), & \text{otherwise}
\end{cases}$}
\label{2DVesselness}
\end{equation}
where:
\begin{align}
\resizebox{.5\linewidth}{!}{$
R_\beta=\lambda _{1}/\lambda_{2}, \quad
S = \sqrt{\lambda_1^2 + \lambda_2^2}.$}
\nonumber
\end{align}
The $\lambda_{1}$, $\lambda_{2}$ are eigenvalues of the Hessian matrix, and $\lambda_{1} \geq \lambda_{2}$. Where $\beta$ and $c$ are positive real user-defined parameters.
If the magnitude of both eigenvalues is small, i.e. the local image structure is likely to be part of the background, then the vesselness measure will be small. If one eigenvalue is small and the other large then the local structure is likely to be curvilinear and the vesselness measure is large. In case both of the eigenvalues magnitudes are large, then the structure is likely to be a blob and the vesselness measure will again small.
\paragraph{3D Vesselness}\label{sec:3DVess}
A 3D vesselness measure~\cite{frangi1998multiscale} is extended on the basis of all eigenvalues of the 3D Hessian matrix. Then, the vesselness for the 3D images is computed as follows:
\begin{equation}
\resizebox{.9\hsize}{!}{$
V_\sigma = \begin{cases}
0, & \lambda_2, \lambda_3>0 \\
\text{exp}\Bigg(-\dfrac{R_{\beta}^2}{2\beta^2}\Bigg)
\Bigg(1-\text{exp}\bigg(-\dfrac{R_\alpha^2}{2\alpha^2}\bigg)\Bigg)
\Bigg(1-\text{exp}\bigg(-\dfrac{S^2}{2c^2}\bigg)\Bigg), & \text{otherwise} \end{cases}
$}
\label{3Dvesselness}
\end{equation}
and where;
\begin{align}
\resizebox{.8\hsize}{!}{$
S=\sqrt {\lambda_{1}^{ 2 }+\lambda _{ 2 }^{ 2 }+\lambda _{ 3 }^{ 2 } },\quad
R_{\beta}=\lambda_{1}/{\sqrt{\lambda_2\lambda_3}},\quad
R_{\alpha}=\lambda _{ 2 } / \lambda _{ 3 } .$}
\nonumber
\end{align}
Similar to vesselness measure in 2D, the $\alpha $, $\beta$ and $c$ are real-valued positive user-defined parameters.
\paragraph{2D Neuriteness} \label{sec:2DNeurite}
This method introduced by~\cite{meijering2004design} and designed to enhance low contrast and highly inhomogeneous neurites in the biomedical images. They changed the Hessian matrix by including a tuning parameter, alpha and derive two tuned eigenvalues $\lambda_1^{'}$ and $\lambda_2^{'}$ as follows:
\begin{align}
\lambda_1^{'} =\lambda _ 1 +\alpha \lambda _{ 2 }, \nonumber\\
\lambda_2^{'}=\lambda _{ 2 } +\alpha \lambda _{ 1 }.
\end{align}
Then, they consider the maximum and minimum eigenvalues across the whole image as describe below, and define a new neurite-enhancing metric $N_{\sigma}$.
\begin{align}
\lambda_{max} =\max(|{\lambda_{1}}'|,|{\lambda_{2}}'|), \nonumber
\\ \lambda_{min} = \min(\lambda_{max}).
\label{lamdamax}
\end{align}
The neuritenees measurement define as:
\begin{equation}
N_\sigma=\begin{cases}
\dfrac{\lambda _{ max }}{\lambda _{ min }} & $if $ \lambda_{ max } <0 \\
0 & $if $ \lambda _{ max }\ge 0 \end{cases},
\newcounter{mytempeqncnt}
\label{2DNeuriteness}
\end{equation}
where $ { \lambda }_{ i }^{ ' }$ are symbolised the normalized eigenvalues of modify Hessian matrix. The $\lambda_\text{min}$ denotes the smallest eigenvalue while $\lambda_\text{max}$ represents the largest one of eigenvalues. Additionally, line like structures which is dark $\left( \lambda _{ max } \ge 0 \right)$ are ignored by the detector.
\paragraph{3D Neuriteness} \label{sec:3DNeurite}
The neuritenees measurement for the 3D image~\cite{meijering2004design} can define using a 3D modified Hessian matrix. Then, the 3D neuriteness measurement can define as:
\begin{equation}
N_\sigma=\begin{cases}
\dfrac{\lambda _{ max } }{\lambda _{ min }} & \text{if } \lambda _{ max }<0 \\
0 & \text{if } \lambda _{ max } \geq 0
\end{cases},
\label{3DNeuriteness}
\end{equation}
and where;
\begin{align}
& \lambda_1^{'} =\lambda _ 1 +\alpha \lambda _{ 2 }+\alpha \lambda _{ 3 },\nonumber\\
& \lambda_2^{'}=\lambda _{ 2 } +\alpha \lambda _{ 1 }+\alpha \lambda _{ 3 },\nonumber\\
& \lambda_3^{'}=\lambda _{ 3 } +\alpha \lambda _{ 1 }+\alpha \lambda _{ 2 },\nonumber\\
& \lambda_{max} =\max(|{\lambda_{1}}'|,|{\lambda_{2}}'|,|{\lambda_{3}}'),\nonumber\\
& \lambda_{min} = \min(\lambda_{max}).\nonumber \notag
\end{align}
\section{Conclusion }\label{sec:conclusion}
The enhancement of curvilinear structures is important for many image processing applications.
In this research, we have proposed a novel approach that combines the advantages of a morphological multiscale top-hat transform and a local tensor to enhance the curvilinear structures in a wide range of 2D and 3D biological and medical images.
The proposed MTHT approach is evaluated qualitatively and quantitively using different 2D and 3D images.
The experimental results show that the approach is comparable with the Hessian-based vesselness and neuriteness approaches, as well with the Zana's top-hat, PCT, SCIRD-TS and RORPO approach.
In general, the MTHT proposed approach showed better enhancement results compared with the state-of-art approaches.
Although the proposed approach achieves good enhancement results in all tested biomedical images, there is room for improvement.
In particular, the top-hat transform using different structural elements for an improved enhancement of the image background, as well as better handling of junctions should be explored further.
\section{Implementation}
The software was implemented and written in MATLAB 2017a on Windows 8.1 pro 64-bit PC running an Intel Core i7-4790 CPU (3.60 GHz) with 16GB RAM. The software is made available at: \url{https://github.com/ShuaaAlharbi/MTHT}.
\section{Introduction}\label{sec:intro}
The enhancement and detection of curvilinear structures are important and essential tasks in biomedical image processing.
There is a wide range of curvilinear structure in biomedical imaging data, such as blood vessels, neurons, leaf veins, and fungal networks.
Curvilinear structure enhancement is an important step, especially where the subjective quality of images of curvilinear structures is necessary for human interpretation.
A wide range of curvilinear structure enhancement approaches have used mathematical morphology operations to enhance curvilinear structures in 2D and 3D images.
The top-hat transform~\cite{haralick1987image} is a popular approach, which extracts bright features from a dark background that match the shape and orientation of a specified structuring element~\cite{zana2001segmentation}.
This approach has been used to extract curvilinear structures in retinal~\cite{liao2014retinal} and fingerprint~\cite{bibiloni2017general} images.
A local tensor representation~\cite{Knutsson1989} of an image measures how image structures change across dominant directions, and the eigenvalues and eigenvectors of the tensor can provide information that can be used to enhance, extract and analyse curvilinear structures.
In this paper, we combine these two approaches by representing curvilinear structures filtered by morphological operations in a local tensor representation of the image.
We apply a multiscale top-hat with a line structuring element at different scales and orientations.
Then, we produce a stack of top-hat images and combine them into a local tensor, find the eigenvalues to calculate vesselness and neuriteness to enhance the curvilinear structure in the biomedical images. This approach works with 2D and 3D images.
Compared with other existing approaches, the gathered results prove that our proposed approach achieves high-quality curvilinear structure enhancement in the synthetic examples and in a wide range of real 2D and 3D biomedical image types.
\section{Method}\label{sec:method}
In this section, we introduce the proposed approach that consolidates the advantages of mathematical morphology and local tensor representation to enhance curvilinear structures in 2D/3D images.
Before explaining the proposed approach in detail, it is useful to provide some more background of the concepts that are applied in this paper.
\input{tex/background}
\subsection{Proposed Method Framework}
Since curvilinear structures can appear at different scales and directions in images, a top-hat transform using multiscale and multi-directional structuring elements should be applied to detect them.
The image is processed by using line structuring elements of different sizes (scale) and directions (orientations) and is then represented as a tensor, the Multiscale Top-Hat Tensor (MTHT), which intrinsically contains information on scale and orientation.
Then, through the use of its eigenvalues and eigenvectors, vesselness and neuriteness are calculated to enhance curvilinear structures. The details of the proposed approach are given below.
\subsubsection{Multiscale Top-Hat Transform}
For a given 2D/3D grayscale image $I(\mathbf{p})$, where $\mathbf{p}$ donates the pixel position, a stack of 2D/3D line structuring elements $B_{{\sigma}_i,\mathbf{u}_j}$, for $m$ different scales $\sigma_i$ and $n$ different orientations $\mathbf{u}_j$, is defined.
In 2D, the $\mathbf{u}_j$ orientation of line structuring element is defined as follows:
\begin{equation}
\label{eq:unitvector}
\mathbf{u}_j = \left[cos({\theta}_j),sin({\theta}_j)\right]^{T},
\end{equation}
where ${\theta}_j \in [0; 180)$.
In 3D, as proposed in~\cite{hacihaliloglu20062a,sazak2017contrast}, a point distribution on the sphere of unit radius is used to define the orientation $\mathbf{u}_j$ of the 3D line structuring element as follows:
\begin{equation}
\mathbf{u}_j = [sin(\theta_j)cos(\phi_j), sin(\theta_j)sin(\phi_j), cos(\theta_j)]^T,
\end{equation}
where ${\theta}_j \in [0; 180]$ and $\phi_j \in [0; 360)$.
Then, we produced a top-hat image using a line structuring element defined by scale $\sigma_i$ and orientation $\mathbf{u}_j$ as follows:
\begin{equation}
TH(\mathbf{p})_{{\sigma}_i,\mathbf{u}_j}=I(\mathbf{p}) - (I \circ B_{\sigma_i,\mathbf{u}_j})(\mathbf{p}).
\label{TH}
\end{equation}
\subsubsection{Tensor Representation}
In general, the tensor representation of an image can provide information about how much the image differs along and across the dominant orientations within a particular region~\cite{Knutsson1989}.
In our case, the local tensor $T(\mathbf{p})_{\sigma_i}$ representation of an image $I(\mathbf{p})$ is generated by combining the bank of top-hat images from Equation~\ref{TH} as follows:
\begin{equation}
T(\mathbf{p})_{\sigma_i}=\sum_{j=1}^{n}\|TH(\mathbf{p})_{\sigma_i,\mathbf{u}_j}\|(\mathbf{u}_j\mathbf{u}_j^T).
\label{eq:tensor}
\end{equation}
\subsubsection{MTHT Vesselness}
As described in Section~\ref{RelatedWork}, piecewise curvilinear segments can be detected by analysing the relations between eigenvalues and eigenvectors of the locally calculated Hessian~\cite{frangi1998multiscale}.
In a similar way~\cite{obara2012contrast}, the vesselness of the proposed approach is defined where the eigenvalues of the Hessian matrix are substituted with those of the MTHT.
Finally, multiscale vesselness, for a given set of $m$ scales can be calculated as follows:
\begin{align}
V = \underset{{i}} \max \left(V_{\sigma_i}\right).
\end{align}
\subsubsection{MTHT Neuriteness}
When combining the neuriteness with our approach, it is necessary to modify the neuriteness measurement introduced by ~\cite{meijering2004design} for 2D and 3D images respectively.
In~\cite{meijering2004design}, they normalised eigenvalues correspondingly to the smaller absolute eigenvalue which is a negative value.
Whereas, in our approach, we used a morphological line structuring element instead of the second order derivative of the Gaussian function used by~\cite{meijering2004design}, so the smaller absolute eigenvalue will be equal to $0$.
The modify neuriteness equation is:
\begin{equation}
N_{\sigma_i}=\begin{cases}
\dfrac{\lambda}{\lambda _{ max }} & $if $ \lambda>0 \\
0 & $if $ \lambda = 0 \end{cases},
\label{2DNeuritenessModify}
\end{equation}
where $\lambda$ is the larger in the magnitude of the two eigenvalues $\lambda_{1}$ and $\lambda_{2}$ for 2D images or the larger in the magnitude of the three eigenvalues $\lambda_{1}$, $\lambda_{2}$ and $\lambda_{3}$ for 3D images. $\lambda_{max}$ denotes the largest $\lambda$ over all pixels in the image. Similar to vesselness, a multiscale neuriteness can be calculated as:
\begin{align}
N = \underset{{i}} \max \left(N_{\sigma_i}\right).
\end{align}
\section{Related Work}
\label{RelatedWork}
Many curvilinear structure enhancement approaches for 2D and 3D images for a wide range of applications have been proposed in the literature to date.
In this section, we list a small selection of the most relevant approaches divided into several subclasses according to the underlying concepts.
\subsection{Hessian Matrix-based Approaches}
A Hessian matrix-based image representation is constructed using responses of an image convolution with a set of matching filters, defined by second-order derivatives of the Gaussian at multiple scales~\cite{frangi1998multiscale,meijering2004design}.
This concept is used to enhance and detect curve / tubular, sheet-like, and blob-like structures in the 2D and 3D images by exploring the relationships between eigenvectors and eigenvalues of the Hessian matrix.
The three most common measurements proposed to date are: vesselness, neuriteness, and regularised volume ratio.
\subsubsection{Vesselness}
The vesselness measure~\cite{frangi1998multiscale} is calculated by computing the ratio of the eigenvalues of the Hessian matrix. The vesselness reaches its maximum when the scale and orientation of the filter matches the size and orientation of the local curvilinear structure. However, vesselness fails at junctions of curvilinear structures / networks due to the low filters responses.
\subsubsection{Neuriteness}
On the other hand, the neuriteness measure~\cite{meijering2004design} is based on a slightly modified Hessian matrix by adding a new tuning parameter.
Neuriteness, in the same way as vesselness, fails at junctions of curvilinear structures / networks due to the low filters responses.s
\subsubsection{Volume Ratio-based Approach}
Hessian-based approaches rely on the eigenvalues and this leads to several problems: (1) eigenvalues are non-uniform throughout an elongated or rounded structure that has uniform intensity; (2) eigenvalues vary with image intensity; and (3) enhancement is not uniform across scales.
A recent volume ratio-based approach~\cite{jerman2016enhancement} aims to solve such problems by computing the ratio of Hessian eigenvalues to handle the low magnitudes of eigenvalues and uniform responses across different structures.
This approach intends to intimate vascular elongated structures in 2D and 3D angiography images.
However, it has drawbacks; despite enhancing the curvilinear structures, it also enhances the noise.
\subsection{Mathematical Morphology-based Approaches}
Morphological operations probe an image with a structuring element placed at all possible locations in the image and match it with the corresponding neighbourhood of pixels.
This structuring element applied to an input image uses a set of operators (intersection, union, inclusion, complement).
Morphological operations are easy to implement and are suitable for many shape-oriented problems. A great number of approaches have been proposed to enhance and detect the curvilinear structures based on different mathematical morphological transforms such as~\cite{soille2001directional,sazak2017multiscale2d,sazak2018multiscale3d}.
\subsubsection{Top-Hat Transform}
The top-hat transform has been widely used to enhance and detect curvilinear structures in retinal~\cite{zana2001segmentation} and aerial~\cite{roman2017top} images.
Zana and Klein~\cite{zana2001segmentation} enhance the curvilinear structures using the top-hat transform with line structuring elements at different directions and with a fixed scale.
Then, they computed the sum of the top-hat along each direction, followed by a curvature measure that is calculated using a Laplacian of Gaussian.
Thus, any small bright noise will be reduced and the contrast of curvilinear structures will be improved.
\subsubsection{Path Operators Transform}
A mathematical morphology-based path opening and closing operation to detect the curvilinear structures in retinal images was introduced by~\cite{sigurdhsson2014automatic}.
Recently, a new path operator called Ranking the Orientation Responses of Path Operators (RORPO) has been proposed to distinguish curvilinear objects from blob-like and planar structures in images~\cite{merveille20172d,merveille2018curvilinear}.
The main disadvantage of the RORPO approach is its high computation cost when applied to large volume image datasets.
Furthermore, this approach required an isotropic image resolution.
\subsection{Phase Congruency Tensor-based Approaches}
Phase congruency (PC) was first introduced in~\cite{kovesi2003phase} and later combined with a local tensor to enhance curvilinear structures in 2D~\cite{obara2012contrast} and 3D~\cite{sazak2017contrast} images.
The majority of Hessian-based approaches rely on image intensity, which leads to poor enhancement or suppression of finer and lower intensity vessels, where Phase Congruency Tensor-based approaches are image contrast-independent.
Moreover, the local tensor has a better representation of directions and the main advantage of using the local tensor is its ability to detect structures oriented in any direction.
However, a major drawback of the PC-based concept is the complexity of its parameter space.
\subsection{Histogram-based Approaches}
Histogram-based approaches are the most popular technique for improving image contrast.
Contrast Limited Adaptive Histogram Equalisation (CLAHE)~\cite{pisano1998contrast} is a widely used pre-processing stage in order to improve the local details of an image.
A major drawback of this method is its sensitivity to noise.
An improvement proposed by~\cite{zhao2014retinal} employs the anisotropic diffusion filter to reduce the noise and smooth the image, especially near the boundary.
\subsection{Wavelet Transform-based Approaches}
The wavelet transform has been widely used for curvilinear structure enhancement in biomedical images.
In~\cite{sihalath2010fingerprint}, the authors propose a new approach to enhance the curvilinear structures in fingerprint images by involving the second derivative of a Gaussian filter with a directional wavelet transform.
Another approach combines the Discrete Wavelet Transform and morphological filter (opening and closing) to enhance curvilinear structures in MRI images~\cite{srivastava2011combination}. In addition, two different wavelets in parallel were applied in~\cite{demirel2011image} to achieve an enhanced high-resolution image.
In~\cite{bankhead2012fast}, the authors proposed an approach exploring the isotropic undecimated wavelet transform.
However, similar to Hessian-based approaches, wavelet transform-based approaches fail to enhance low-intensity and fine curvilinear structures.
\subsection{Learning-based Approaches}
Recent learning-based methods are more suitable to deal with the scene complexity problem in natural images~\cite{lam2010general,azzopardi2015trainable,zhang2016robust}. In particular~\cite{annunziata2015scale}, proposed a new regression architecture based on the basis of filter banks learned by sparse convolutional coding to speed-up the training process. They are carefully designed hand-crafted filters (SCIRD-TS) which are modelling appearance properties of curvilinear structures.
\section{Results} \label{sec:result}
In this section, we present quantitative and qualitative validations for the proposed approach against both synthetic and real-world 2D and 3D imaging data. We then compare the results with state-of-the-art approaches.
In order to validate the approach quantitatively in 2D and 3D images, we calculate the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC), further details can be found in~\cite{fawcett2006introduction}.
\input{./tex/figure-retina-2d}
\input{./tex/ROC-2d}
\input{./tex/table-retina-2d}
\subsection{Application to 2D Retinal Images}
Although a visual inspection can provide some information regarding the effectiveness of the curvilinear structure enhancement approaches, a more rigorous form of quantitative validation is required.
As in~\cite{jerman2016enhancement}, we chose to use the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC) metrics to compare the curvilinear structure enhancement approaches. We derive the ROC curve and then calculate the AUC value. Each enhanced image is segmented at different threshold levels and compared with the corresponding ground truth segmentation of curvilinear structures in the image.
We measure the quality of the approach by using publicly available retinal image datasets: DRIVE~\cite{niemeijer2004comparative}, STARE~\cite{hoover2000locating} and HRF~\cite{odstrcilik2013retinal}.
These datasets have been chosen because of their availability and their ground truth data. We have used these ground truth segmentations to quantitatively compare the proposed approach with the other curvilinear structure enhancement approaches.
In particular, we evaluate our approach, alongside the state-of-the-art methods, calculating the Receiver Operating Characteristic (ROC) curve and the mean of Area Under the Curve (AUC) between the enhanced images and the ground truth.
The results are displayed accordingly in Figure~\ref{fig:rtina}, Figure~\ref{fig:retinaROC} and Table~\ref{tab:auc2}. A higher AUC value indicates a better enhancement of curvilinear structures, with a value of $1$ indicating that the enhanced image is identical to the ground truth image.
Our experimental results clearly show that our proposed approach works better than the state-of-the-art approaches for the STARE dataset.
Furthermore, the proposed approach achieved a high score overall on the HRF healthy and unhealthy images, as illustrated in Table~\ref{tab:auc2}.
The average computation time for the proposed method is $13.7$ seconds for DRIVE image and $16.4$ seconds for STARE image. Please make a note that the proposed method has been implemented and tested in Matlab, however, C++ implementation could be much faster.
\input{./tex/figure-Vascular-3d}
\input{./tex/ROC-3d}
\input{./tex/table-Vascular-3d}
\subsection{3D Vascular Network Complexity}
In order to validate our approach in 3D, we used synthetic vascular networks produced by the free software package called VascuSynth ~\cite{hamarneh2010vascusynth}.
The tree generation is performed by iteratively growing a vascular structure based on an oxygen demand map. Each generated image is associated with it's ground truth.
In this experiment, we generated 9 volumetric images with an increasing complexity and their corresponding ground truth.
In addition, in order to make the image more realistic, we added a small amount of Gaussian noise of level $\sigma^2=10$ and applied a Gaussian smoothing kernel with a standard deviation of 1.
The results, in terms of AUC, are presented in Table~\ref{tab:auc3} and a sample of the results are shown in Figure~\ref{fig:vascu:input}.
We also demonstrate the mean ROC curve over the 9 enhanced images, as shown in Figure~\ref{fig:roc:vascu}.
Our proposed approach is compared with vesselness~\cite{frangi1998multiscale}, neuriteness~\cite{meijering2004design}, PCT (vesselness and neuriteness)~\cite{sazak2017contrast} and with the latest 3D enhancement approach~\cite{merveille2018curvilinear}.
Our proposed approach clearly has the highest mean AUC value (0.995) with a standard deviation equal to (0.006) for the proposed MTHT-vesselness.
On the other hand, we obtained an AUC value (0.978) with a standard deviation equal to (0.014) for the proposed MTHT-neuriteness compared to the state-of-art approaches.
\subsection{2D and 3D Qualitative Validation}
Additionally, as displayed in Figures~\ref{fig:2d} and~\ref{fig:3d}, we have demonstrated the robustness of the proposed approach when applied to a wide range of 2D and 3D real-world images.
It is clear that our approach has the best performance compared with the state-of-the-art approaches.
In particular, our proposed approach can handle complex curvilinear networks as shown in Figure~\ref{fig:3d}(1) and (2).
\input{./tex/figure-real-1-2d}
\input{./tex/figure-real-3d}
|
\section{Introduction}
Studying strong interactions between hadrons is one of the most
important issues in hadron physics. The best known strong interaction
is the nuclear force between nucleons ($N$s), which generates a large
number of atomic nuclei composed of protons and neutrons. In addition
to the nuclear force, recent interest in the strong interactions
between hadrons is to explore bound states of mesons and baryons
governed by strong interactions between them, which are so-called
hadronic molecules. A classic example of hadronic molecule candidates
is the $\Lambda (1405)$ resonance, which may be an $S$-wave quasibound
state of antikaon ($\bar{K}$) and $N$~\cite{Dalitz:1960du}. Recently
an analysis of the lattice QCD energy levels with an
effective-field-theory model showed that the $\Lambda (1405)$ is
dominated by the bound $\bar{K} N$ component with isospin $I =
0$~\cite{Hall:2014uca}. The $\bar{K} N$ molecular picture for the
$\Lambda (1405)$ was supported also in Refs.~\cite{Sekihara:2014kya,
Kamiya:2015aea} in terms of the compositeness, which is defined as
the norm of a two-body wave function for hadronic
resonances~\cite{Hyodo:2011qc, Aceti:2012dd, Sekihara:2016xnq}.
Furthermore, recent experiments in high-energy colliders such as
Belle, BaBar, BESIII, and LHCb have revealed fruitful physics in the
charm- and bottom-quark sectors. Besides the $X$, $Y$, and $Z$
resonances as exotic candidates, the $D_{s 0}(2317)^{-}$
resonance~\cite{Aubert:2003fg, Abe:2003jk} is of interest from the
viewpoint of hadronic molecules.
Because its mass is located just below the $\bar{K} \bar{D}$
threshold, it is natural to think that the $D_{s 0}(2317)^{-}$ is an
$S$-wave $\bar{K} \bar{D}$ bound state with isospin $I = 0$. A
dominant $\bar{K} \bar{D}$ component for the $D_{s 0}(2317)^{-}$ was
implied by theoretical calculations~\cite{Gamermann:2006nm,
Navarra:2015iea}, and was supported also by theoretical analyses of
lattice QCD data~\cite{Mohler:2013rwa, Torres:2014vna,
Albaladejo:2018mhb} and of experimental
data~\cite{Albaladejo:2016hae}.
We can extend discussions on hadronic molecules of two-body systems to
those of three-body systems. In this respect, three-body systems are
not only a key to understand interactions between hadrons but also a
good ground to investigate three-body dynamics. For example,
properties of two-body bound states may disappear in three-body bound
states, or, conversely, some properties may emerge uniquely in the
three-body bound states. Three-body forces may become significant,
and in general its form may differ from the three-nucleon force.
Furthermore, in case that a two-body interaction depends on the energy
owing to implicit channels which do not appear as explicit degrees of
freedom, it is not trivial how to treat the energy dependence of the
two-body interaction in the three-body calculations. We can discuss
these three-body dynamics from properties of three-body hadronic
molecules by applying and extending approaches to solve few-body
problems developed for usual atomic nuclei.
An important progress for three-body hadronic molecules takes place
recently on the $\bar{K} N N$ quasibound state. The $\bar{K} N N$
quasibound state was predicted based on the strong attraction in the
$\bar{K} N$ system in Ref.~\cite{Akaishi:2002bg}, which was followed
by more sophisticated theoretical
calculations~\cite{Shevchenko:2006xy, Shevchenko:2007ke, Ikeda:2007nz,
Ikeda:2008ub, Ikeda:2010tk, Dote:2008in, Dote:2008hw, Wycech:2008wf,
Bayar:2011qj, Barnea:2012qa, Dote:2014via, Dote:2017veg,
Dote:2017wkk, Ohnishi:2017uni}. Eventually, the J-PARC E15
experiment very recently observed a peak structure which can be a
signal of the $\bar{K} N N$ quasibound state~\cite{Sada:2016nkb,
Sekihara:2016vyd, Ajimura:2018iyx}. The study of the $\bar{K} N N$
quasibound state also triggered theoretical studies of similar
three-body hadronic molecules: for instance, $\bar{K} \bar{K}
N$~\cite{Kanada-Enyo:2008wsu, Shevchenko:2015oea}, $K \bar{K}
N$~\cite{Jido:2008kp, MartinezTorres:2008kh, MartinezTorres:2009cw,
MartinezTorres:2010zv, Xie:2010ig}, $K K
\bar{K}$~\cite{Torres:2011jt}, and $K D N$~\cite{Xiao:2011rc}.
In this study we propose a new candidate of three-body hadronic
molecules, the $\bar{K} \bar{D} N$ three-body system with spin/parity
$J^{P} = 1/2^{+}$ and isospin $I = 1/2$. This system has two kinds of
attraction which could be essential to make the $\bar{K} \bar{D} N$
bound state. One is the $\bar{K} N ( I = 0 )$ interaction and the
other is the $\bar{K} \bar{D} ( I = 0 )$ interaction, which
dynamically generate the $\Lambda (1405)$ and $D_{s 0} (2317)$,
respectively. On the other hand, the $\bar{D} N$ interaction is
moderate, but some models implied that the coupling to the
$\bar{D}^{\ast} N$ channel brings attraction to the $\bar{D} N$
interaction and generates an $S$-wave $\bar{D} N ( I = 0 )$ bound
state with binding energy $\sim 1 \text{ MeV}$~\cite{Yasui:2009bz,
Gamermann:2010zz}. To clarify whether the $\bar{K} \bar{D} N$
three-body system is bound or not, in this manuscript we will solve a
nonrelativistic three-body potential model for the $\bar{K} \bar{D} N$
system and search for the bound state. Throughout this study, we
assume isospin symmetry for the hadron masses and interactions, and
concentrate on the zero-charge $\bar{K} \bar{D} N$ system, which
exists in the $K^{-} \bar{D}^{0} p$-$\bar{K}^{0} D^{-} p$-$\bar{K}^{0}
\bar{D}^{0} n$ coupled channels.
We here mention that the $\bar{K} \bar{D} N$ three-body system with $I
= 1/2$ has the minimal quark configuration of $u u d s \bar{c}$ or $u
d d s \bar{c}$, so this bound state, if exists, is explicitly a
``pentaquark'' state. This is in contrast to the
charmonium-pentaquark $P_{c} (4450)$~\cite{Aaij:2015tga}, in which the
charm and anticharm quarks are hidden inside the $P_{c} (4450)$ and
the minimal quark configuration is $u u d$.
This paper is organized as follows. In Sec.~\ref{sec:2} we construct
the two-body local potentials between $\bar{K} N$, $\bar{K} \bar{D}$,
and $\bar{D} N$ for the subsystems of $\bar{K} \bar{D} N$. By using
this two-body local potentials, we formulate the $\bar{K} \bar{D} N$
three-body problem in Sec.~\ref{sec:3}. In Sec.~\ref{sec:4} we show
our numerical results and discuss properties of the $\bar{K} \bar{D}
N$ system. Section~\ref{sec:5} is devoted to the summary and
concluding remarks of this study.
\section{\boldmath Two-body systems}
\label{sec:2}
\subsection{How to construct two-body local potentials}
\label{sec:2A}
First of all, we explain how to construct the two-body interactions
for the $\bar{K} N$, $\bar{K} \bar{D}$, and $\bar{D} N$ systems.
Because we are interested in the two-body interactions in $S$ wave, we
extract the $S$-wave projected interaction and then construct the
local and orbital-angular-momentum independent potentials for the
two-body systems which reproduce the two-body phenomena in $S$ wave.
In general, the $\bar{K} N$, $\bar{K} \bar{D}$, and $\bar{D} N$
channels couple to inelastic channels, but in this study we integrate
out them so that only the $\bar{K} N$, $\bar{K} \bar{D}$, and $\bar{D}
N$ channels are explicit degrees of freedom, according to the method
in Ref.~\cite{Hyodo:2007jq}. For instance, in the $\bar{K}
\bar{D}$-$\pi \bar{D}_{s}$-$\eta \bar{D}_{s}$ coupled channels for the
$\bar{K} \bar{D}$ interaction, the $\pi \bar{D}_{s}$ and $\eta
\bar{D}_{s}$ channels are taken as inelastic and are integrated out.
We refer to the two-body interactions in which inelastic channels are
integrated out as effective interactions.
We start with a full coupled-channels interaction in isospin basis
$V_{j k}$ with the channel indices $j$ and $k$, which is calculated in
a certain model. We project the interaction to the $S$ wave and take
the so-called on-shell factorization~\cite{Oset:1997it}, so $V_{j k}$
depends only on the two-body center-of-mass energy $\epsilon$. This
interaction generates the full coupled-channels scattering amplitude
$T_{j k} ( \epsilon )$ as
\begin{align}
T_{j k} ( \epsilon )
= & V_{j k} ( \epsilon )
+ \sum _{l} V_{j l} ( \epsilon ) G_{l} ( \epsilon ) T_{l k} ( \epsilon )
\notag \\
= & \left \{ \left [ 1 - V ( \epsilon ) G ( \epsilon ) \right ]^{-1}
V ( \epsilon ) \right \} _{j k}.
\end{align}
Here, $G_{j}$ is the hadron--hadron loop function
\begin{equation}
G_{j} ( \epsilon ) = i \int \frac{d^{4} k}{( 2 \pi )^{4}}
\frac{1}{k^{2} - m_{j}^{2}} \frac{1}{( P - k )^{2} - M_{j}^{2}} ,
\end{equation}
where $m_{j}$ and $M_{j}$ are masses of particles in channel $j$ and
$P^{\mu} = ( \epsilon , \, \bm{0} )$. We calculate the loop function with
the dimensional regularization, which brings a subtraction constant
corresponding to the cutoff for the loop.
Now suppose that we explicitly treat only channel $j = 1$ and
integrate out inelastic channels $j > 1$. In this condition, we can
calculate the effective interaction $V^{\rm eff} $ as
\begin{equation}
V^{\rm eff} ( \epsilon )
= \left \{ \left [ 1 - V ( \epsilon ) \tilde{G} ( \epsilon ) \right ]^{-1}
V ( \epsilon ) \right \} _{j=1, k=1} ,
\end{equation}
with
\begin{equation}
\tilde{G}_{j} ( \epsilon ) =
\begin{cases}
0 & ( j = 1 ) , \\
G_{j} ( \epsilon ) & ( j > 1 ) .
\end{cases}
\end{equation}
Physically, $V^{\rm eff}$ is the sum of the bare interaction $V_{1 1}$
and terms which include resummation of loop contributions from the
inelastic channels $j > 1$ to all orders~\cite{Hyodo:2007jq}. Then,
the effective interaction $V^{\rm eff}$ is translated into the local
two-body potential $U ( r )$ with the relative distance $r$ in the
nonrelativistic reduction:
\begin{equation}
U ( r ; \, \epsilon )
= \frac{g ( r )}{4 \omega _{1} ( \epsilon ) \Omega _{1} ( \epsilon )}
V^{\rm eff} ( \epsilon ) .
\end{equation}
Here, $g ( r )$ is a form factor defined as
\begin{equation}
g ( r ) = \frac{1}{\pi ^{3/2} b^{3}}
e^{- r^{2} / b^{2}} ,
\end{equation}
and
\begin{equation}
\omega _{1} ( \epsilon ) = \frac{\epsilon ^{2} + m_{1}^{2} - M_{1}^{2}}
{2 \epsilon} ,
\quad
\Omega _{1} ( \epsilon ) = \frac{\epsilon ^{2} + M_{1}^{2} - m_{1}^{2}}
{2 \epsilon} .
\end{equation}
The range parameter $b$ can be fixed independently in three systems:
$\bar{K} N$, $\bar{K} \bar{D}$, and $\bar{D} N$. Note that the local
potential $U$ depends on the energy of the two-body system $\epsilon$
according to the integration of the implicit channels as well as
intrinsic energy dependence of the full interaction $V_{j k}$.
The above potential $U ( r ; \, \epsilon )$ is described in isospin
basis. The translation into the potential in particle basis is
straightforward.
\subsection{\boldmath $\bar{K} N$ system}
In the $\bar{K} N$ subsystem in $\bar{K} \bar{D} N$, we consider three
channels: $K^{-} p$, $\bar{K}^{0} n$, and $\bar{K}^{0} p$. The former
two channels couple to each other.
We employ the Kyoto $\bar{K} N$ effective potential developed in the
above manner in Ref.~\cite{Miyahara:2015bya}, which reproduces
experimental results on the $K^{-} p$ scattering phenomena based on
chiral SU(3) coupled-channels dynamics~\cite{Ikeda:2011pi,
Ikeda:2012au}. The range parameter is $b = 0.38 \text{ fm}$. The Kyoto
$\bar{K} N$ potential in its original form is written in isospin basis
as $U_{\bar{K} N ( I = 0 )} ( r; \, \epsilon )$ and $U_{\bar{K} N ( I
= 1 )} ( r; \, \epsilon )$. The expression in particle basis is
\begin{align}
& U_{K^{-} p \to K^{-} p}
= U_{\bar{K}^{0} n \to \bar{K}^{0} n}
= \frac{U_{\bar{K} N ( I = 0 )} + U_{\bar{K} N ( I = 1 )}}{2} ,
\end{align}
\begin{align}
& U_{K^{-} p \to \bar{K}^{0} n}
= U_{\bar{K}^{0} n \to K^{-} p}
= \frac{U_{\bar{K} N ( I = 0 )} - U_{\bar{K} N ( I = 1 )}}{2} ,
\end{align}
\begin{align}
& U_{\bar{K}^{0} p \to \bar{K}^{0} p}
= U_{\bar{K} N ( I = 1 )} ,
\end{align}
where we omitted parameters $(r ; \, \epsilon)$ for $U$.
The $\bar{K} N ( I = 0 )$ effective potential $U_{\bar{K} N ( I = 0 )}
( r ; \, \epsilon )$ generates two $\Lambda (1405)$ poles at $\epsilon
_{\rm pole} = 1424 - 26 i \text{ MeV}$ and $1381 - 81 i
\text{ MeV}$~\cite{Miyahara:2015bya} as $S$-wave bound-state solutions of the
Schr\"{o}dinger equation
\begin{align}
& \left [ m_{K} + m_{N} - \frac{\nabla ^{2}}{2 \mu _{K N}}
+ U_{\bar{K} N ( I = 0 )} ( r ; \, \epsilon _{\rm pole} ) \right ]
\psi
( r )
\notag \\
& = \epsilon _{\rm pole} \psi
( r ) ,
\end{align}
where $m_{K}$ and $m_{N}$ are kaon and nucleon masses, respectively,
and $\mu _{K N} \equiv m_{K} m_{N} / (m_{K} + m_{N})$ is the reduced
mass of the $\bar{K} N$ system. Among the two $\Lambda (1405)$ poles,
the higher pole at $\epsilon _{\rm pole} = 1424 - 26 i \text{ MeV}$
corresponds to the $\bar{K} N$ quasibound state in chiral
dynamics~\cite{Sekihara:2014kya, Kamiya:2015aea}. Properties of the
$\bar{K} N$ quasibound state in Kyoto $\bar{K} N$ potential was
discussed in Ref.~\cite{Miyahara:2015bya}; we here quote that the
average of the $\bar{K} N$ distance is $\sqrt{\langle r^{2} \rangle} =
1.06 - 0.57 i \text{ fm}$ with the Gamow-vector normalization
method~\cite{Miyahara:2015bya}.
\subsection{\boldmath $\bar{K} \bar{D}$ system}
In the $\bar{K} \bar{D}$ subsystem in $\bar{K} \bar{D} N$, we consider
three channels: $K^{-} \bar{D}^{0}$, $\bar{K}^{0} D^{-}$ and
$\bar{K}^{0} \bar{D}^{0}$. The former two channels couple to each
other.
We employ a phenomenological Lagrangian constructed in
Ref.~\cite{Gamermann:2006nm}. An important point is that in this
model the $D_{s 0} (2317)$ state is dynamically generated by the
strong attraction of the elastic $\bar{K} \bar{D} (I = 0)$
interaction. In isospin $I = 0$, we have $\bar{K} \bar{D}$-$\eta
\bar{D}_{s}$ coupled channels with the interaction
\begin{align}
& V_{\bar{K} \bar{D} \to \bar{K} \bar{D} ( I = 0 )} ( \epsilon )
\notag \\
& = - \frac{1}{3 f_{\pi} f_{D}}
\left [ \gamma ( \bar{t} - \bar{u} ) + \epsilon ^{2} - \bar{u}
+ m_{D}^{2} + m_{K}^{2} \right ] ,
\label{eq:V_KDKD0}
\end{align}
\begin{align}
V_{\bar{K} \bar{D} \to \eta \bar{D}_{s}} ( \epsilon )
& = V_{\eta \bar{D}_{s} \to \bar{K} \bar{D}} ( \epsilon )
\notag \\
& = - \frac{1}{6 \sqrt{3} f_{\pi} f_{D}}
\left [ \gamma ( \bar{u} - \bar{t} ) - ( 3 + \gamma )
( \epsilon ^{2} - \bar{u} ) \right .
\notag \\
& \left . \phantom{\bar{u} - \bar{t}}
- m_{D}^{2} - 3 m_{K}^{2}
+ 2 m_{\pi}^{2} \right ] ,
\end{align}
\begin{align}
& V_{\eta \bar{D}_{s} \to \eta \bar{D}_{s}} ( \epsilon )
\notag \\
& = - \frac{1}{9 f_{\pi} f_{D}}
\left [ \gamma ( - \epsilon ^{2} + 2 \bar{t} - \bar{u} )
+ 2 m_{D}^{2} + 6 m_{K}^{2} - 4 m_{\pi}^{2} \right ] ,
\end{align}
where $f_{\pi}$ and $m_{\pi}$ ($f_{D}$ and $m_{D}$) are decay constant
and mass of pion ($D$ meson), respectively, and $\gamma$ is the
squared ratio of the masses of the light to heavy vector mesons. In
the above expressions, $\bar{t}$ and $\bar{u}$ are the Mandelstam
variables projected to the $S$ wave in the on-shell factorization.
Owing to the $S$-wave projection, $\bar{t}$ and $\bar{u}$ are
functions only of $\epsilon$:
\begin{align}
\bar{t}
= - \frac{1}{2 \epsilon ^{2}} & \left [ \epsilon ^{4} - \epsilon ^{2}
( m_{a}^{2} + m_{b}^{2} + m_{c}^{2} + m_{d}^{2} ) \right .
\notag \\
& \left . + ( m_{a}^{2} - m_{b}^{2} ) ( m_{c}^{2} - m_{d}^{2} ) \right ] ,
\end{align}
\begin{align}
\bar{u}
= - \frac{1}{2 \epsilon ^{2}} & \left [ \epsilon ^{4} - \epsilon ^{2}
( m_{a}^{2} + m_{b}^{2} + m_{c}^{2} + m_{d}^{2} ) \right .
\notag \\
& \left . - ( m_{a}^{2} - m_{b}^{2} ) ( m_{c}^{2} - m_{d}^{2} ) \right ] ,
\end{align}
where $m_{a, b, c, d}$ are masses of particles in the $a b \to c d$
reaction. Similarly, in isospin $I = 1$, we have $\bar{K}
\bar{D}$-$\pi \bar{D}_{s}$ coupled channels with the interaction
\begin{align}
& V_{\bar{K} \bar{D} \to \bar{K} \bar{D} ( I = 1 )} ( \epsilon ) = 0,
\label{eq:V_KDKD1}
\end{align}
\begin{align}
& V_{\bar{K} \bar{D} \to \pi \bar{D}_{s}} ( \epsilon )
= V_{\pi \bar{D}_{s} \to \bar{K} \bar{D}} ( \epsilon )
\notag \\
& = - \frac{1}{6 f_{\pi} f_{D}}
\left [ \gamma ( \epsilon ^{2} - \bar{t} ) + \epsilon ^{2} - \bar{u}
- m_{D}^{2} - m_{K}^{2} \right ] ,
\end{align}
\begin{align}
& V_{\pi \bar{D}_{s} \to \pi \bar{D}_{s}} ( \epsilon ) = 0 .
\end{align}
Note that in our model the $\bar{K} \bar{D} (I = 1)$ interacts only
through the $\pi \bar{D}_{s}$ intermediate channel.
Then, the inelastic channels $\pi \bar{D}_{s}$ and $\eta \bar{D}_{s}$
are integrated out and $\bar{K} \bar{D}$ effective potentials in
isospin basis, $U_{\bar{K} \bar{D} ( I = 0 )} (r ; \, \epsilon)$ and
$U_{\bar{K} \bar{D} ( I = 1 )} (r ; \, \epsilon)$, are calculated in
the method in Sec.~\ref{sec:2A}. These potentials are translated into
those in particle basis as:
\begin{align}
U_{K^{-} \bar{D}^{0} \to K^{-} \bar{D}^{0}}
= & U_{\bar{K}^{0} D^{-} \to \bar{K}^{0} D^{-}}
\notag \\
= & \frac{U_{\bar{K} \bar{D} ( I = 0 )} + U_{\bar{K} \bar{D} ( I = 1 )}}{2} ,
\end{align}
\begin{align}
U_{K^{-} \bar{D}^{0} \to \bar{K}^{0} D^{-}}
= & U_{\bar{K}^{0} D^{-} \to K^{-} \bar{D}^{0}}
\notag \\
= & \frac{U_{\bar{K} \bar{D} ( I = 0 )} - U_{\bar{K} \bar{D} ( I = 1 )}}{2} ,
\end{align}
\begin{align}
U_{\bar{K}^{0} \bar{D}^{0} \to \bar{K}^{0} \bar{D}^{0}}
= U_{\bar{K} \bar{D} ( I = 1 )} .
\end{align}
In this study, we use parameters fixed in Ref.~\cite{Navarra:2015iea}:
$\gamma = ( 800 \text{ MeV} / 2050 \text{ MeV} )^{2} \approx 0.152$, $f_{\pi} = 93
\text{ MeV}$, and $f_{D} = 165 \text{ MeV}$. Furthermore, we fix the range of the
effective local potential as $b = 0.36 \text{ fm}$. With these parameters,
we can generate the scalar meson $D_{s 0}(2317)^{-}$ with its mass
$\epsilon _{\rm pole} = 2317 \text{ MeV}$ in the $\bar{K} \bar{D} ( I = 0 )$
channel as an $S$-wave bound-state solution of the Schr\"{o}dinger equation with
the effective potential $U_{\bar{K} \bar{D} ( I = 0 )} ( r ; \,
\epsilon )$
\begin{align}
& \left [ m_{K} + m_{D} - \frac{\nabla ^{2}}{2 \mu _{K D}}
+ U_{\bar{K} \bar{D} ( I = 0 )} ( r ; \, \epsilon _{\rm pole} ) \right ] \psi
( r )
\notag \\
&
= \epsilon _{\rm pole} \psi
( r ) ,
\end{align}
with the $\bar{K} \bar{D}$ reduced mass $\mu _{K D} \equiv m_{K} m_{D}
/ (m_{K} + m_{D})$. In Fig.~\ref{fig:KD-DN} (solid line) we plot the
density distribution for the $\bar{K} \bar{D}$ system in the
$D_{s 0}(2317)$
\begin{equation}
\rho _{\bar{K} \bar{D}} ( r )
\equiv r^{2} \left [ \psi ( r ) \right ] ^{2} ,
\label{eq:rho_KD}
\end{equation}
calculated from the wave function of the $\bar{K} \bar{D}$ bound state
$\psi ( r )$ with the normalization that integral of $\rho _{\bar{K}
\bar{D}}$ with respect to $r$ in the range $[ 0, \, \infty )$ is
unity. The average of the $\bar{K} \bar{D}$ distance is calculated
as $\sqrt{ \langle r^{2} \rangle _{\bar{K} \bar{D}}} = 0.93 \text{ fm}$,
where $\langle r^{2} \rangle _{\bar{K} \bar{D}}$ is defined as
\begin{equation}
\langle r^{2} \rangle _{\bar{K} \bar{D}}
\equiv \int _{0}^{\infty} d r \, r^{2} \rho _{\bar{K} \bar{D}} ( r ) .
\end{equation}
\begin{figure}[!t]
\centering
\PsfigA{8.6cm}{Fig1_WF.ps}
\caption{Density distributions for the $\bar{K} \bar{D}$ system in
the $D_{s 0}(2317)$ (solid line) and for the $\bar{D} N$ system in
the $\bar{D} N$ bound state (dashed line).}
\label{fig:KD-DN}
\end{figure}
\begin{figure}[!t]
\centering
\Psfig{8.6cm}{Fig2.eps}
\caption{Real parts of the effective local potentials with the
energies at the respective thresholds. The $\bar{K} \bar{D} ( I =
1 )$ potential is very tiny, as $U_{\bar{K} \bar{D} ( I = 1 )} ( r
= 0 ) \approx - 2 \text{ MeV}$.}
\label{fig:Ueff}
\end{figure}
In Fig.~\ref{fig:Ueff}, to compare the strength of the potentials, we
show the real parts of the effective local potentials for the $\bar{K}
\bar{D} ( I = 0 )$ and the $\bar{K} \bar{D} ( I = 1 )$ systems as
thick and thin dashed-dotted lines, respectively. The energy for the
potentials is fixed as the threshold, $\epsilon = m_{K} + m_{D}$. As
one can see, the $\bar{K} \bar{D} ( I = 0 )$ potential has very strong
attraction, while the $\bar{K} \bar{D} ( I = 1 )$ potential is very
tiny and negligible.
We here mention that in terms of heavy quark symmetry we may have to
introduce the $\bar{D}^{\ast}$ and $\bar{D}_{s}^{\ast}$ vector mesons,
which exist $\sim 140 \text{ MeV}$ above the ground $\bar{D}$ and
$\bar{D}_{s}$ mesons, respectively. However, in this study we do not
take into account them in the $\bar{K} \bar{D}$ system because the
contributions from the $\bar{K} \bar{D}^{\ast}$, $\pi
\bar{D}_{s}^{\ast}$, and $\eta \bar{D}_{s}^{\ast}$ channels are
expected to be negligible compared to the $\bar{K} \bar{D}$ dynamics
around its threshold.
\subsection{\boldmath $\bar{D} N$ system}
In the $\bar{D} N$ subsystem in $\bar{K} \bar{D} N$, we consider three
channels: $D^{-} p$, $\bar{D}^{0} n$, and $\bar{D}^{0} p$. The former
two channels couple to each other.
For the $\bar{D} N$ interaction, we take the approach discussed in
Ref.~\cite{Gamermann:2010zz}. We introduce the $S$-wave channels
$\bar{D} N$ (specified by the channel $j = 1$) and $\bar{D}^{\ast} N$
($j = 2$) both in isospin $I = 0$ and $1$.\footnote{We neglect the
$\bar{D}^{\ast} \Delta$ channel in $I = 1$, which was included in
Ref.~\cite{Gamermann:2010zz} but was not important.} We calculate
the interaction with a Lagrangian invariant under $\SUN{8}$ rotations
which treats heavy pseudoscalar and vector mesons on an equal footing
as required by heavy quark symmetry. The $S$-wave interaction can be
expressed as~\cite{Gamermann:2010zz}
\begin{equation}
V_{j k} ( \epsilon ) = \frac{\xi _{j k}}{2 f_{D}^{2}} ( \epsilon - m_{N} )
\sqrt{[ \Omega _{j} ( \epsilon ) + m_{N} ][ \Omega _{k} ( \epsilon ) + m_{N} ]}
\end{equation}
with the on-shell nucleon energy in $j$th channel $\Omega _{j} (
\epsilon )$. The coefficient $\xi _{j k}$ comes from the $\SUN{8}$
group structure of the couplings, whose expression is
\begin{equation}
\xi _{(I = 0)} =
\left (
\begin{array}{@{\,}cc@{\,}}
0 & - \sqrt{12} \\
- \sqrt{12} & 4 \\
\end{array}
\right ) ,
\quad
\xi _{(I = 1)} =
\left (
\begin{array}{@{\,}cc@{\,}}
2 & 4 / \sqrt{3} \\
4 / \sqrt{3} & -2/3 \\
\end{array}
\right ) ,
\end{equation}
Parameters are taken from Ref.~\cite{Gamermann:2010zz}: $f_{D} = 157.4
\text{ MeV}$. An interesting feature is that, although the elastic $\bar{D}
N$ interaction is zero in $I = 0$, dynamics with the $\bar{D}^{\ast}
N$ coupled channel generates a $\bar{D} N$ bound state in $I = 0$ with
binding energy $\sim 1
\text{ MeV}$~\cite{Gamermann:2010zz}.\footnote{Importance of the
$\bar{D}^{\ast} N$ channel in the $\bar{D} N$ dynamics was pointed
out also in Ref.~\cite{Yasui:2009bz}, where a $\bar{D} N$ bound
state with binding energy $\sim 1 \text{ MeV}$ was predicted as well.}
With such an attractive $\bar{D} N$ interaction, it is possible to
study the formation of $D$ mesic nuclei in, e.g.,
Ref.~\cite{GarciaRecio:2010vt, GarciaRecio:2011xt,
Yamagata-Sekihara:2015ebw}.
Then we integrate out the $\bar{D}^{\ast} N$ channel and obtain
$\bar{D} N$ effective potentials in isospin basis, $U_{\bar{D} N ( I =
0 )} (r ; \, \epsilon)$ and $U_{\bar{D} N ( I = 1 )} (r ; \,
\epsilon)$, with which we calculate the potentials in particle basis
as
\begin{align}
& U_{D^{-} p \to D^{-} p} = U_{\bar{D}^{0} n \to \bar{D}^{0} n}
= \frac{U_{\bar{D} N ( I = 0 )} + U_{\bar{D} N ( I = 1 )}}{2} ,
\end{align}
\begin{align}
& U_{D^{-} p \to \bar{D}^{0} n} = U_{\bar{D}^{0} n \to D^{-} p}
= - \frac{U_{\bar{D} N ( I = 0 )} - U_{\bar{D} N ( I = 1 )}}{2} ,
\end{align}
\begin{align}
& U_{\bar{D}^{0} p \to \bar{D}^{0} p} = U_{\bar{D} N ( I = 1 )} .
\end{align}
As for the range parameter $b$, we fix it so as to reproduce an
$S$-wave $\bar{D} N ( I = 0 )$ bound state with $1 \text{ MeV}$ binding
($\epsilon _{\rm pole} = 2805 \text{ MeV}$) as a solution of the Schr\"{o}dinger
equation with the effective potential $U_{\bar{D} N ( I = 0 )} ( r ;
\, \epsilon )$
\begin{align}
& \left [ m_{D} + m_{N} - \frac{\nabla ^{2}}{2 \mu _{D N}}
+ U_{\bar{D} N ( I = 0 )} ( r ; \, \epsilon _{\rm pole} ) \right ] \psi
( r )
\notag \\
&
= \epsilon _{\rm pole} \psi
( r ) ,
\end{align}
with the $\bar{D} N$ reduced mass $\mu _{D N} \equiv m_{D} m_{N} /
(m_{D} + m_{N})$. The result of the range parameter is $b = 0.26
\text{ fm}$. In Fig.~\ref{fig:KD-DN} (dashed line) we plot the density
distribution for the $\bar{D} N$ system in the $\bar{D} N$ bound state
calculated in the same manner as in the $\bar{K} \bar{D}$ case. The
calculated $\bar{D} N$ distance $\sqrt{ \langle r^{2} \rangle
_{\bar{D} N}} = 3.66 \text{ fm}$ is larger than $\sqrt{ \langle r^{2}
\rangle _{\bar{K} \bar{D}}} = 0.93 \text{ fm}$ due to the loosely bound
nature of $\bar{D} N$.
In Fig.~\ref{fig:Ueff}, we show the real parts of the effective local
potentials for the $\bar{D} N ( I = 0 )$ and the $\bar{D} N ( I = 1 )$
systems as thick and thin dotted lines, respectively. The energy for
the potentials is fixed as the threshold, $\epsilon = m_{D} + m_{N}$.
As one can see, the $\bar{D} N ( I = 0 )$ potential is the smallest
among the $\bar{K} N$, $\bar{K} \bar{D}$, and $\bar{D} N$ potentials
with $I = 0$. The $\bar{D} N ( I = 1 )$ potential is much smaller
than the $\bar{D} N ( I = 0 )$ potential.
\section{\boldmath $\bar{K} \bar{D} N$ three-body problem}
\label{sec:3}
\begin{figure}[!t]
\centering
\PsfigII{0.169}{Jacobi1.eps}
\PsfigII{0.169}{Jacobi2.eps}
\PsfigII{0.169}{Jacobi3.eps}
\caption{Three types of Jacobi coordinates.}
\label{fig:Jacobi}
\end{figure}
Next let us formulate the $\bar{K} \bar{D} N$ three-body problem. For
this purpose, we set the coordinates of the $\bar{K}$, $\bar{D}$, and
$N$ as $\bm{r}_{1}$, $\bm{r}_{2}$, and $\bm{r}_{3}$, respectively, and
introduce the Jacobi coordinates as in Fig.~\ref{fig:Jacobi}:
\begin{align}
& \bm{\lambda}_{1} = \bm{r}_{1}
- \frac{m_{D} \bm{r}_{2} + m_{N} \bm{r}_{3}}{m_{D} + m_{N}} ,
\quad
\bm{\rho}_{1} = \bm{r}_{2} - \bm{r}_{3} ,
\\
& \bm{\lambda}_{2} = \bm{r}_{2}
- \frac{m_{N} \bm{r}_{3} + m_{K} \bm{r}_{1}}{m_{N} + m_{K}} ,
\quad
\bm{\rho}_{2} = \bm{r}_{3} - \bm{r}_{1} ,
\\
& \bm{\lambda}_{3} = \bm{r}_{3}
- \frac{m_{K} \bm{r}_{1} + m_{D} \bm{r}_{2}}{m_{K} + m_{D}} ,
\quad
\bm{\rho}_{3} = \bm{r}_{1} - \bm{r}_{2} .
\end{align}
We consider the $\bar{K} \bar{D} N$ system in its center-of-mass rest
frame
\begin{equation}
\bm{R}
= \frac{m_{K} \bm{r}_{1} + m_{D} \bm{r}_{2} + m_{N} \bm{r}_{3}}{M_{K D N}}
= \bm{0} ,
\end{equation}
\begin{equation}
M_{K D N} \equiv m_{K} + m_{D} + m_{N} ,
\end{equation}
and do not treat the center-of-mass motion of the three-body system.
We employ particle basis and describe the zero-charge $\bar{K} \bar{D}
N$ system. For the three-body system, we use index $j$ to specify the
channel in particle basis: $j = 1$ for $K^{-} \bar{D}^{0} p$, $2$ for
$\bar{K}^{0} D^{-} p$, and $3$ for $\bar{K}^{0} \bar{D}^{0} n$. We do
not explicitly take into account other three-body channels such as
$\pi \bar{D} \Sigma$ and $\pi \bar{D}_{s} N$ but they are implemented
in the effective potentials. We neglect the transitions to two-hadron
channels such as $\bar{D} \Lambda$ and limit our model space to the
$\bar{K} \bar{D} N$ states.
By using the Jacobi coordinates and channel index, we can express the
three-body wave function in coordinate space as
\begin{equation}
\Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} )
= \braket{ j ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ) | \Psi } .
\end{equation}
Here and below, the Jacobi coordinate $( \bm{\lambda}_{1} , \,
\bm{\rho}_{1} )$ is used just as a representative to specify the set
of the coordinates $( \bm{r}_{1} , \, \bm{r}_{2} , \, \bm{r}_{3} )$.
The three-body wave function $\ket{\Psi}$ satisfies the Schr\"{o}dinger equation
\begin{equation}
\hat{H} \ket{\Psi} = E \ket{\Psi} ,
\end{equation}
with the Hamiltonian $\hat{H}$ and an eigenvalue $E$ of the three-body
system, respectively. Multiplying bra vector $\bra{j (
\bm{\lambda}_{1} , \, \bm{\rho}_{1} )}$ from the left and inserting
the complete set
\begin{equation}
1 = \sum _{k = 1}^{3} \int d^{3} \lambda _{1} \, d^{3} \rho _{1} \,
\ket{k ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} )}
\bra{k ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} )}
\end{equation}
between $\hat{H}$ and $\ket{\Psi}$ in the left-hand side, we obtain
\begin{align}
& \sum _{k = 1}^{3} \left [ \delta _{j k} \hat{H}_{0}
+ V_{j k} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ; \, E ) \right ]
\Psi _{k} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} )
\notag \\ &
= E \Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ) .
\label{eq:Schr}
\end{align}
The kinetic term of the three-body Hamiltonian $\hat{H}_{0}$ is
\begin{align}
\hat{H}_{0}
= & M_{K D N} - \frac{1}{2 \mu _{1}^{\prime}}
\left ( \frac{\partial}{\partial \bm{\lambda}_{1}} \right ) ^{2}
- \frac{1}{2 \mu _{1}}
\left ( \frac{\partial}{\partial \bm{\rho}_{1}} \right ) ^{2}
\end{align}
where $\mu _{1}^{\prime}$ and $\mu _{1}$ are the reduced masses
\begin{equation}
\mu _{1}^{\prime} = \frac{( m_{D} + m_{N} ) m_{K}}{M_{K D N}} ,
\quad
\mu _{1} = \frac{m_{D} m_{N}}{m_{D} + m_{N}} .
\end{equation}
As for the potential term $V_{j k}$, we employ the
orbital-angular-momentum independent potentials developed in the
previous section. The diagonal parts $V_{j j}$ consist of all the
three combinations of two particles among $\bar{K}$, $\bar{D}$, and
$N$ in each channel:
\begin{align}
V_{1 1} = &
U_{\bar{D}^{0} p \to \bar{D}^{0} p}
( \rho _{1} ; \, \epsilon _{\bar{D} N} )
+ U_{K^{-} p \to K^{-} p} ( \rho _{2} ; \, \epsilon _{\bar{K} N} )
\notag \\
& + U_{K^{-} \bar{D}^{0} \to K^{-} \bar{D}^{0}}
( \rho _{3} ; \, \epsilon _{\bar{K} \bar{D}} ) ,
\end{align}
\begin{align}
V_{2 2} = &
U_{D^{-} p \to D^{-} p}
( \rho _{1} ; \, \epsilon _{\bar{D} N} )
+ U_{\bar{K}^{0} p \to \bar{K}^{0} p} ( \rho _{2} ; \, \epsilon _{\bar{K} N} )
\notag \\
& + U_{\bar{K}^{0} D^{-} \to \bar{K}^{0} D^{-}}
( \rho _{3} ; \, \epsilon _{\bar{K} \bar{D}} ) ,
\end{align}
\begin{align}
V_{3 3} = &
U_{\bar{D}^{0} n \to \bar{D}^{0} n}
( \rho _{1} ; \, \epsilon _{\bar{D} N} )
+ U_{\bar{K}^{0} n \to \bar{K}^{0} n} ( \rho _{2} ; \, \epsilon _{\bar{K} N} )
\notag \\
& + U_{\bar{K}^{0} \bar{D}^{0} \to \bar{K}^{0} \bar{D}^{0}}
( \rho _{3} ; \, \epsilon _{\bar{K} \bar{D}} ) ,
\end{align}
where $\epsilon _{\bar{D} N}$, $\epsilon _{\bar{K} N}$, and $\epsilon
_{\bar{K} \bar{D}}$ are energies of the subsystems $\bar{K} N$,
$\bar{D} N$, and $\bar{K} \bar{D}$, respectively, fixed later. The
nondiagonal components of the potential consist of the charge
transition of two particles among $\bar{K}$, $\bar{D}$, and $N$:
\begin{align}
V_{1 2} = V_{2 1} = U_{K^{-} \bar{D}^{0} \to \bar{K}^{0} D^{-}}
( \rho _{3} ; \, \epsilon _{\bar{K} \bar{D}} ) ,
\end{align}
\begin{align}
V_{1 3} = V_{3 1} = U_{K^{-} p \to \bar{K}^{0} n}
( \rho _{2} ; \, \epsilon _{\bar{K} N} ) ,
\end{align}
\begin{align}
V_{2 3} = V_{3 2} = U_{D^{-} p \to \bar{D}^{0} n}
( \rho _{1} ; \, \epsilon _{\bar{D} N} ) .
\end{align}
Because the potentials have imaginary parts according to the
implementation of the open channels, the Hamiltonian $\hat{H}$ is not
Hermitian. Therefore, the Hamiltonian can have an eigenstate with a
complex eigenvalue, called a quasibound state. We do not consider
three-body forces in this study.
As we have constructed in Sec.~\ref{sec:2}, the two-body potential
in the subsystem depends on its energy. There is ambiguity to fix
energy of a two-body subsystem in a three system, but we here simply
divide the total energy $E$ among three particles according to
the ratio of masses, i.e.,
\begin{align}
& \epsilon _{\bar{D} N} = \frac{m_{D} + m_{N}}{M_{K D N}} E ,
\label{eq:eps_DN}
\\
& \epsilon _{\bar{K} N} = \frac{m_{K} + m_{N}}{M_{K D N}} E ,
\label{eq:eps_KN}
\\
& \epsilon _{\bar{K} \bar{D}} = \frac{m_{K} + m_{D}}{M_{K D N}} E .
\label{eq:eps_KD}
\end{align}
Note that the subsystem energy is complex when the total energy $E$ is
complex.
In this study we concentrate on the ground state of the $\bar{K}
\bar{D} N$ system with spin/parity $J^{P} = 1/2^{+}$, so we limit the
basis function for the three-body wave function in the channel $c$ to
having zero orbital angular momenta both for the $\lambda _{c}$ and
$\rho _{c}$ modes: $l_{\lambda _{c}} = l_{\rho _{c}} = 0$. We then
employ the Gaussian expansion method~\cite{Hiyama:2003cu} and take the
sum of all the three rearrangements of the Jacobi coordinates, which
results in
\begin{equation}
\Psi _{j}
= \sum _{c = 1}^{3} \sum _{n , n^{\prime} = 1}^{N} \mathcal{C}_{j, n n^{\prime}}^{c}
\exp \left ( - \frac{\lambda _{c}^{2}}{r_{n}^{2}}
- \frac{\rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right ) ,
\label{eq:WFj}
\end{equation}
with number of the expansion $N$, coefficients $\mathcal{C}_{j, n
n^{\prime}}^{c}$, and different ranges $r_{n}$ in a geometric
progression
\begin{equation}
r_{n} = r_{\rm min} \times
\left ( \frac{r_{\rm max}}{r_{\rm min}} \right )^{(n - 1) / (N - 1)} .
\end{equation}
The minimal and maximal ranges, $r_{\rm min}$ and $r_{\rm max}$,
respectively, are fixed according to the physical condition of
interactions. We comment that, although each $c$ channel in
Eq.~\eqref{eq:WFj} have zero orbital angular momentum, $l_{\lambda
_{c}} = l_{\rho _{c}} = 0$, the sum of all the three rearrangements
allows us to take into account components with nonzero orbital angular
momenta of two-body subsystems.
By using the wave function~\eqref{eq:WFj}, the Schr\"{o}dinger
equation~\eqref{eq:Schr} becomes
\begin{align}
& \sum _{c = 1}^{3} \sum _{n , n^{\prime} = 1}^{N} \sum _{k = 1}^{3}
\left \{ \delta _{j k} \left [ \frac{1}{\mu _{c}^{\prime} r_{n}^{2}}
\left ( 3 - \frac{2 \lambda _{c}^{2}}{r_{n}^{2}} \right )
\right . \right .
\notag \\
&
\left . \left . ~
+ \frac{1}{\mu _{c} r_{n^{\prime}}^{2}}
\left ( 3 - \frac{2 \rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right )
+ M_{K D N} - E \right ] + V_{j k} ( E )
\right \}
\notag \\
& \times
\mathcal{C}_{k, n n^{\prime}}^{c}
\exp \left ( - \frac{\lambda _{c}^{2}}{r_{n}^{2}}
- \frac{\rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right ) = 0 ,
\label{eq:GEM}
\end{align}
where we introduced the reduced masses
\begin{equation}
\mu _{2}^{\prime} = \frac{( m_{K} + m_{N} ) m_{D}}{M_{K D N}} ,
\quad
\mu _{2} = \frac{m_{K} m_{N}}{m_{K} + m_{N}} ,
\end{equation}
\begin{equation}
\mu _{3}^{\prime} = \frac{( m_{K} + m_{D} ) m_{N}}{M_{K D N}} ,
\quad
\mu _{3} = \frac{m_{K} m_{D}}{m_{K} + m_{D}} .
\end{equation}
Then we multiply $\exp ( - \lambda _{a}^{2} / r_{m}^{2} - \rho
_{a}^{2} / r_{m^{\prime}}^{2} )$ to the Schr\"{o}dinger equation~\eqref{eq:GEM}
and integrate it with respect to $\bm{\lambda}_{1}$ and
$\bm{\rho}_{1}$, which results in
\begin{align}
\sum _{\beta} & \left [ \mathcal{T}_{\alpha \beta}
+ \mathcal{V}_{\alpha \beta} ( E )
+ ( M_{\bar{K} \bar{D} N} - E ) \mathcal{N}_{\alpha \beta} \right ]
\mathcal{C}_{k, n n^{\prime}}^{c} = 0 ,
\label{eq:three}
\end{align}
where we introduced sets of indices $\alpha = \{ j, \, m, \,
m^{\prime}, \, a \}$ and $\beta = \{ k, \, n, \, n^{\prime}, \, c \}$,
and define $\mathcal{T}_{\alpha \beta}$, $\mathcal{V}_{\alpha \beta}$,
and $\mathcal{N}_{\alpha \beta}$ as
\begin{align}
\mathcal{T}_{\alpha \beta} \equiv & \delta _{j k}
\int d^{3} \lambda _{1} \, d^{3} \rho _{1} \,
\exp \left ( - \frac{\lambda _{a}^{2}}{r_{m}^{2}}
- \frac{\rho _{a}^{2}}{r_{m^{\prime}}^{2}}
- \frac{\lambda _{c}^{2}}{r_{n}^{2}}
- \frac{\rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right )
\notag \\ &
\times \left [ \frac{1}{\mu _{c}^{\prime} r_{n}^{2}}
\left ( 3 - \frac{2 \lambda _{c}^{2}}{r_{n}^{2}} \right )
+ \frac{1}{\mu _{c} r_{n^{\prime}}^{2}}
\left ( 3 - \frac{2 \rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right ) \right ] ,
\end{align}
\begin{align}
\mathcal{V}_{\alpha \beta} ( E )
\equiv & \int d^{3} \lambda _{1} \, d^{3} \rho _{1} \,
\exp \left ( - \frac{\lambda _{a}^{2}}{r_{m}^{2}}
- \frac{\rho _{a}^{2}}{r_{m^{\prime}}^{2}}
- \frac{\lambda _{c}^{2}}{r_{n}^{2}}
- \frac{\rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right )
\notag \\ & \times
V_{j k} ( E ) ,
\end{align}
\begin{align}
\mathcal{N}_{\alpha \beta} \equiv \delta _{j k}
\int d^{3} \lambda _{1} \, d^{3} \rho _{1} \,
\exp \left ( - \frac{\lambda _{a}^{2}}{r_{m}^{2}}
- \frac{\rho _{a}^{2}}{r_{m^{\prime}}^{2}}
- \frac{\lambda _{c}^{2}}{r_{n}^{2}}
- \frac{\rho _{c}^{2}}{r_{n^{\prime}}^{2}} \right ) ,
\end{align}
respectively. We can regard Eq.~\eqref{eq:three} as a generalized
eigenvalue problem of linear algebra. We numerically solve this to
evaluate the eigenvalue $E = E_{\rm pole}$ and eigenvector
$\mathcal{C}_{j, n n^{\prime}}^{c}$.
\section{\boldmath $\bar{K} \bar{D} N$ molecular state}
\label{sec:4}
\subsection{Eigenenergy}
Now we solve the Schr\"{o}dinger equation~\eqref{eq:three} in the $K^{-}
\bar{D}^{0} p$-$\bar{K}^{0} D^{-} p$-$\bar{K}^{0} \bar{D}^{0} n$
coupled channels and search for the $\bar{K} \bar{D} N$ bound state.
For the study of the $\bar{K} \bar{D} N$ system, we fix $r_{\rm min} =
0.1 \text{ fm}$ and $r_{\rm max} = 20.0 \text{ fm}$. Taking the number of the
expansion $N = 10$, we find a solution of Eq.~\eqref{eq:three} with
its eigenenergy $E_{\rm pole} = 3244 - 17 i \text{ MeV}$. The convergence of
the expansion can be checked by the trace of the eigenenergy $E_{\rm
pole}$ from $N = 4$ to $10$, which is plotted in the complex energy
plane of Fig.~\ref{fig:trace}. As one can see, we achieve the
convergence of the expansion with $N \ge 8$.
\begin{figure}[!t]
\centering
\PsfigA{8.6cm}{Fig4_pole.ps}
\caption{Eigenenergy $E_{\rm pole}$ as a function of the number of
the Gaussian expansion $N$ (open circles) from $N = 4$ to $N =
10$. We also plot the threshold points for the $\Lambda (1405) +
\bar{D}$, $D_{s 0} (2317) + N$, and $\bar{K} + \bar{D} + N$ as
filled symbols.}
\label{fig:trace}
\end{figure}
The real part of the eigenenergy $E_{\rm pole}$ is below the $D_{s 0}
(2317) + N$ threshold ($3256 \text{ MeV}$) as well as below the $\Lambda
(1405) + \bar{D}$ and $\bar{K} \bar{D} N$ thresholds ($3291 - 26 i
\text{ MeV}$ and $3302 \text{ MeV}$, respectively). This means that this state is
indeed a $\bar{K} \bar{D} N$ quasibound state which cannot decay into
$D_{s 0}(2317) + N$, $\Lambda (1405) + \bar{D}$, nor $\bar{K} \bar{D}
N$. The binding energy of the quasibound state is $58 \text{ MeV}$ measured
from the $\bar{K} \bar{D} N$ threshold, $48 \text{ MeV}$ from the $\Lambda
(1405) \bar{D}$ threshold, and $13 \text{ MeV}$ from the $D_{s 0} (2317) N$
threshold.
The imaginary part of the eigenenergy indicates decay of the
quasibound state, as we introduced complex-valued potentials
reflecting implicit decay channels such as $\pi \Sigma$ in $\bar{K}
N$. We emphasize that the imaginary part of the eigenenergy is
obtained in a full calculation rather than in a perturbative one.
From the eigenenergy, we find that the decay width of the quasibound
state is $- 2 \times \text{Im} E_{\rm pole} = 34 \text{ MeV}$. The decay of
the quasibound state will be discussed in the next subsection.
\subsection{Decay}
Because $\Lambda (1405)$ decays into $\pi \Sigma$ and $D_{s
0}(2317)^{-}$ into $\pi \bar{D}_{s}$, we expect that the main decay
channels of the $\bar{K} \bar{D} N$ quasibound state may be $\pi
\Sigma + \bar{D}$ and $\pi \bar{D}_{s} + N$. To check this, we
perform the same three-body calculations but without the imaginary
parts of the two-body potentials.
When we neglect the imaginary part of the $\bar{K} N$ potential, we
obtain the bound-state eigenenergy at $3240 - 0 i \text{ MeV}$. On the other
hand, when we neglect the imaginary part of the $\bar{K} \bar{D}$
($\bar{D} N$) potential, we obtain the bound-state eigenenergy at
$3244 - 18 i \text{ MeV}$ ($3245 - 20 i \text{ MeV}$). These results indicate that
the decay width of the $\bar{K} \bar{D} N$ quasibound state originates
from the $\bar{K} N$ interaction.
Hence, one could conclude that the decay is dominated by the $\pi
\Sigma + \bar{D}$ channel, but it is not the all of the main decay
modes. As we will see in the next subsection, the $\bar{K} \bar{D} N$
quasibound state has a significant $\bar{K} N ( I = 1 )$ component.
Furthermore, the $\bar{K} N$ effective potential has similar values
of the imaginary parts for the $I = 0$ and $I = 1$ channels (see
Figs.~7 and 8 of Ref.~\cite{Miyahara:2015bya}). These indicate that
the bound-state decay originating from the $\bar{K} N$ interaction
includes not only the $\pi \Sigma ( I = 0 ) + \bar{D}$ mode but also
the $\pi \Lambda + \bar{D}$ and $\pi \Sigma ( I = 1 ) + \bar{D}$
modes. As a consequence, the main decay modes of the $\bar{K} \bar{D}
N$ quasibound state are the $\pi \Lambda + \bar{D}$ and $\pi \Sigma +
\bar{D}$ channels.
We here note that the decay width of the $\bar{K} \bar{D} N$
quasibound state, $34 \text{ MeV}$, is smaller than that of the $\Lambda
(1405)$ as the $\bar{K} N$ quasibound state, $\sim 50 \text{ MeV}$. If the
$\bar{K} N$ subsystem in the $\bar{K} \bar{D} N$ quasibound state
behaved like the $\Lambda (1405)$, the decay width of the $\bar{K}
\bar{D} N$ quasibound state would be similar to the width of the
$\Lambda (1405)$. Indeed, this reduction of the decay width is caused
by the three-body dynamics, in particular the slight extension of the
$\bar{K} N$ distance in the $\bar{K} \bar{D} N$ quasibound state
compared to that in the $\Lambda (1405)$, as we will see in the next
subsection. From this result we can say that, in general, the decay
width of a three-body quasibound state $A B C$ is not the sum of the
decay widths of the two-body quasibound states $A B$, $B C$, and $C
A$ but it depends on the internal structure of the three-body
quasibound state.
We may consider the two-hadron decay modes, $\bar{K} \bar{D} N \to
\bar{D} \Lambda$, $\bar{D} \Sigma$, and $\bar{D}_{s} N$, as well,
which are not included in our formulation. For these two-hadron decay
modes, we can use the same argument as in Ref.~\cite{Jido:2008kp} (see
also Fig.~4 therein). First, the transition to two-hadron states via
a contact interaction is strongly suppressed for a three-body
quasibound state, because the three constituents should meet at a
point for the contact interaction to take place. Second, the
transition to two-hadron states via virtual meson exchanges is also
suppressed due to the dilute nature of the three-body quasibound
state. Such a virtual meson exchange process is expected to take
place in the nonmesonic decay of $\bar{K}$-nucleus
systems~\cite{Sekihara:2009yk, Sekihara:2012wj}. Therefore, in
analogy to the nonmesonic decay of $\bar{K}$-nucleus systems, we can
estimate that the branching ratio of the two-hadron decays of the
$\bar{K} \bar{D} N$ quasibound state will be $\sim 20 \%$, about as
large as the empirical values of the branching ratio of the nonmesonic
decay of $\bar{K}$-nucleus systems.
\subsection{Structure}
Then, we investigate the internal structure of the $\bar{K} \bar{D} N$
quasibound state by using the wave function $\Psi _{j} (
\bm{\lambda}_{1} , \, \bm{\rho}_{1} )$ which is normalized as
\begin{equation}
\sum _{j = 1}^{3} \int d^{3} \lambda _{1} \, d^{3} \rho _{1} \,
\left [ \Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} )
\right ] ^{2} = 1 .
\label{eq:norm}
\end{equation}
We emphasize that, because the $\bar{K} \bar{D} N$ quasibound state is
a resonance, we calculate the complex value squared of the wave
function rather than the absolute value squared to normalize the
resonance wave function $\Psi _{j}$ as a Gamow vector.
\begin{table}[!t]
\caption{Isospin components $X$ and averages of the distances $d$
for two hadrons in the $\bar{K} \bar{D} N$ quasibound state.}
\label{tab:comp}
\begin{ruledtabular}
\begin{tabular}{lc}
$X_{\bar{K} N ( I = 0 )}$ & $0.24 + 0.02 i$
\\
$X_{\bar{K} N ( I = 1 )}$ & $0.76 - 0.02 i$
\\
$d_{\bar{K} N}$ & $1.13 - 0.39 i \text{ fm}$
\\
\hline
$X_{\bar{K} \bar{D} ( I = 0 )}$ & $0.98 - 0.01 i$
\\
$X_{\bar{K} \bar{D} ( I = 1 )}$ & $0.02 + 0.01 i$
\\
$d_{\bar{K} \bar{D}}$ & $0.79 - 0.05 i \text{ fm}$
\\
\hline
$X_{\bar{D} N ( I = 0 )}$ & $0.27 - 0.02 i$
\\
$X_{\bar{D} N ( I = 1 )}$ & $0.73 + 0.02 i$
\\
$d_{\bar{D} N}$ & $1.05 - 0.35 i \text{ fm}$
\end{tabular}
\end{ruledtabular}
\end{table}
We first perform the isospin decomposition. To this end, we construct
the projection operator to the $\bar{K} N ( I = 0 )$ state as
\begin{equation}
\mathcal{P}_{\bar{K} N ( I = 0 )} =
\frac{1}{2} \ket{ K^{-} p + \bar{K}^{0} n }
\bra{ K^{-} p + \bar{K}^{0} n } .
\end{equation}
By using this projection operator, we can calculate the fraction of
the $\bar{K} N ( I = 0 )$ component in the $\bar{K} \bar{D} N$
quasibound state as
\begin{align}
X_{\bar{K} N ( I = 0 )} \equiv &
\braket{\Psi | \mathcal{P}_{\bar{K} N ( I = 0 )} | \Psi }
= \frac{\left \langle \Psi _{1}^{2} + 2 \Psi _{1} \Psi _{3}
+ \Psi _{3}^{2} \right \rangle}
{2} ,
\end{align}
where
\begin{align}
\left \langle \Psi _{j} \Psi _{k} \right \rangle
= \int d^{3} \lambda _{1} \, d^{3} \rho _{1} \,
\Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} )
\Psi _{k} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ) ,
\end{align}
while the $\bar{K} N ( I = 1 )$ component is
\begin{equation}
X_{\bar{K} N ( I = 1 )} = 1 - X_{\bar{K} N ( I = 0 )} .
\end{equation}
Similarly, we can express the projection operators and fractions of
the components for other states as
\begin{equation}
\mathcal{P}_{\bar{K} \bar{D} ( I = 0 )} =
\frac{1}{2} \ket{ K^{-} \bar{D}^{0} + \bar{K}^{0} D^{-} }
\bra{ K^{-} \bar{D}^{0} + \bar{K}^{0} D^{-} } ,
\end{equation}
\begin{align}
X_{\bar{K} \bar{D} ( I = 0 )} \equiv &
\braket{\Psi | \mathcal{P}_{\bar{K} \bar{D} ( I = 0 )} | \Psi }
= \frac{\left \langle \Psi _{1}^{2} + 2 \Psi _{1} \Psi _{2}
+ \Psi _{2}^{2} \right \rangle}{2} ,
\end{align}
\begin{equation}
X_{\bar{K} \bar{D} ( I = 1 )} = 1 - X_{\bar{K} \bar{D} ( I = 0 )} ,
\end{equation}
\begin{equation}
\mathcal{P}_{\bar{D} N ( I = 0 )} =
\frac{1}{2} \ket{ \bar{D}^{0} n - D^{-} p }
\bra{ \bar{D}^{0} n - D^{-} p } ,
\end{equation}
\begin{align}
X_{\bar{D} N ( I = 0 )} \equiv &
\braket{\Psi | \mathcal{P}_{\bar{D} N ( I = 0 )} | \Psi }
= \frac{\left \langle \Psi _{2}^{2} - 2 \Psi _{2} \Psi _{3}
+ \Psi _{3}^{2} \right \rangle}{2} ,
\end{align}
\begin{equation}
X_{\bar{D} N ( I = 1 )} = 1 - X_{\bar{D} N ( I = 0 )} ,
\end{equation}
respectively. The results of the fractions $X$ are listed in
Table~\ref{tab:comp}. All the fractions are complex because the
quasibound state is a resonance. Nevertheless, the $\bar{K} \bar{D}$
component in isospin $I = 0$ is very close to unity with small
imaginary part, which implies the dominant $\bar{K} \bar{D} ( I = 0 )$
component inside the $\bar{K} \bar{D} N$ quasibound state. This is a
consequence of the three-body dynamics to maximize the attraction
among three constituents. Namely, as one can see from
Fig.~\ref{fig:Ueff}, the $\bar{K} \bar{D} ( I = 0 )$ interaction is
most attractive among the pairs of the two constituents in the present
formulation, and $\bar{K} N ( I = 0 )$ comes next. Besides, in
contrast to the moderate $\bar{K} N ( I = 1 )$ attraction, the
$\bar{K} \bar{D} ( I = 1 )$ interaction is negligible, as the $\bar{K}
\bar{D} (I = 1)$ interacts only through the $\pi \bar{D}_{s}$
intermediate channel [see Eq.~\eqref{eq:V_KDKD1}]. Therefore, the
three-body dynamics increase the $\bar{K} \bar{D} ( I = 0 )$ fraction
as much as possible so as to maximize the attraction in the $\bar{K}
\bar{D} N$ quasibound state. In this sense, the $\bar{K} \bar{D}$
interaction, both the $I = 0$ and $1$ components, is most essential
for the internal structure of the $\bar{K} \bar{D} N$ quasibound
state. Furthermore, we have checked that the following relations
hold:
\begin{align}
\left \langle \Psi _{1}^{2} \right \rangle \approx
\left \langle \Psi _{1} \Psi _{2} \right \rangle \approx
\left \langle \Psi _{2}^{2} \right \rangle \approx \frac{1}{2} ,
\end{align}
\begin{align}
\left \langle \Psi _{1} \Psi _{3} \right \rangle \approx
\left \langle \Psi _{2} \Psi _{3} \right \rangle \approx
\left \langle \Psi _{3}^{2} \right \rangle \approx 0 .
\end{align}
These relations indicate that the quasibound state is indeed described
by the $[ \bar{K} \bar{D} (I = 0) ] p$ configuration. The negligible
contribution from the $\bar{K}^{0} \bar{D}^{0} n$ channel also
explains the results that the $\bar{K} N ( I = 0 )$ and $\bar{D} N ( I
= 0 )$ components are close to $1/4$.
\begin{figure}[!t]
\centering
\PsfigA{8.6cm}{Fig5_pd.ps}
\caption{Density distributions for the $\bar{K} N$, $\bar{K}
\bar{D}$, and $\bar{D} N$ subsystems in the $\bar{K} \bar{D} N$
quasibound state.}
\label{fig:density}
\end{figure}
Next we investigate how the two-hadron subsystems behave in the
$\bar{K} \bar{D} N$ quasibound state by calculating the density
distribution for each pair of two constituents, which is defined as
\begin{align}
& \text{P}_{\bar{K} N} ( \rho _{2} )
\equiv \rho _{2}^{2} \int d \Omega _{\bm{\rho} _{2}} \, d^{3} \lambda _{2}
\, \sum _{j = 1}^{3}
\left [ \Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ) \right ] ^{2} ,
\\
& \text{P}_{\bar{K} \bar{D}} ( \rho _{3} )
\equiv \rho _{3}^{2} \int d \Omega _{\bm{\rho} _{3}} \, d^{3} \lambda _{3}
\, \sum _{j = 1}^{3}
\left [ \Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ) \right ] ^{2} ,
\\
& \text{P}_{\bar{D} N} ( \rho _{1} )
\equiv \rho _{1}^{2} \int d \Omega _{\bm{\rho} _{1}} \, d^{3} \lambda _{1}
\, \sum _{j = 1}^{3}
\left [ \Psi _{j} ( \bm{\lambda}_{1} , \, \bm{\rho}_{1} ) \right ] ^{2} ,
\end{align}
where $\Omega _{\bm{\rho}}$ is the solid angle of the vector
$\bm{\rho}$. The integration of $\text{P}_{\bar{K} N}$,
$\text{P}_{\bar{K} \bar{D}}$, and $\text{P}_{\bar{D} N}$, with respect
to $\rho _{2}$, $\rho _{3}$, and $\rho _{1}$, respectively, in the
range $[ 0, \, \infty )$ is unity according to the
normalization~\eqref{eq:norm}.\footnote{The measures of the Jacobi
coordinates satisfy $d^{3} \lambda _{1} \, d^{3} \rho _{1} = d^{3}
\lambda _{2} \, d^{3} \rho _{2} = d^{3} \lambda _{3} \, d^{3} \rho
_{3}$.} The resulting density distributions for the pairs of two
constituents are plotted in Fig.~\ref{fig:density}. From the
figure, the $\bar{K} N$ and $\bar{D} N$ distributions are similar to
each other and extend typical hadronic scale $1 \text{ fm}$. However, the
$\bar{K} \bar{D}$ distribution in the $\bar{K} \bar{D} N$ quasibound
state is significant only below $1 \text{ fm}$, which is similar to the
$\bar{K} \bar{D}$ distribution in the $D_{s 0}(2317)$ in
Fig.~\ref{fig:KD-DN}. This result supports the $[ \bar{K} \bar{D}
(I = 0) ] p$ configuration for the $\bar{K} \bar{D} N$ quasibound
state. Furthermore, because the $\bar{K} \bar{D} ( I = 0 )$
subsystem is compact, the distributions of the $\bar{K} N$ and
$\bar{D} N$ in the $\bar{K} \bar{D} N$ quasibound state are similar
to each other.
From the density distributions, we can calculate the averages of the
distances between two constituents:
\begin{align}
& d_{\bar{K} N} \equiv
\sqrt{\int _{0}^{\infty} d \rho _{2} \, \rho _{2}^{2}
\text{P}_{\bar{K} N} ( \rho _{2} ) } ,
\\ &
d_{\bar{K} \bar{D}} \equiv
\sqrt{\int _{0}^{\infty} d \rho _{3} \, \rho _{3}^{2}
\text{P}_{\bar{K} \bar{D}} ( \rho _{3} )} ,
\\
& d_{\bar{D} N} \equiv
\sqrt{\int _{0}^{\infty} d \rho _{1} \, \rho _{1}^{2}
\text{P}_{\bar{D} N} ( \rho _{1} )} .
\end{align}
The results are listed in Table~\ref{tab:comp}. As one can see, the
averages of the distances have small imaginary parts due to the
resonance nature but their real parts are dominant. Therefore, below
we focus on the real parts of the distances. Among the three
distances, the distance between $\bar{K} \bar{D}$ is the smallest,
which indicates the compact $\bar{K} \bar{D} ( I = 0 )$ subsystem.
The distance between $\bar{K} \bar{D}$ in the $\bar{K} \bar{D} N$
quasibound state is smaller than that in the $D_{s 0}(2317)$ as the
$\bar{K} \bar{D}$ two-body bound state, $0.93 \text{ fm}$. This is because
the $N$ assists the $\bar{K} \bar{D}$ attraction in the $\bar{K}
\bar{D} N$ quasibound state via the $\bar{K} N$ and $\bar{D} N$
interactions. The distance between $\bar{D} N$ in the $\bar{K}
\bar{D} N$ quasibound state becomes much smaller than that of the
$\bar{D} N$ two-body bound state, $3.66 \text{ fm}$, owing to the compact
$\bar{K} \bar{D}$ subsystem and the strong attraction between $\bar{K}
N$. We also note that, although the distance between $\bar{K} N$ in
the $\bar{K} \bar{D} N$ quasibound state is similar to that in the
$\Lambda (1405)$, the $\Lambda (1405)$ is not effective degrees of
freedom in the $\bar{K} \bar{D} N$ quasibound state because the
isospin component $X_{\bar{K} N ( I = 0 )}$ is only about $1/4$.
In terms of the structure, we can understand the decrease of the decay
width of the $\bar{K} \bar{D} N$ quasibound state compared to the
$\Lambda (1405)$. This is caused by the fact that the $\bar{K} N$
distance in the $\bar{K} \bar{D} N$ quasibound state is slightly
larger than that in the $\Lambda (1405)$, which is crucial to the
decay width of the $\bar{K} \bar{D} N$ quasibound state originating
from the $\bar{K} N$ interaction. Actually, because the $\bar{K} N$
effective potential has a finite range, which is $b = 0.38 \text{ fm}$ in our
study, the increase of the $\bar{K} N$ distance directly reduces the
probability of overlapping $\bar{K} N$ for the decay. As a
consequence, we obtain the smaller decay width of the $\bar{K} \bar{D}
N$ quasibound state than that of the $\Lambda (1405)$. We note that
the isospin structure of the $\bar{K} \bar{D} N$ quasibound state,
i.e., dominant $\bar{K} N ( I = 1 )$ component, is irrelevant to the
decrease of the decay width, because the $\bar{K} N$ effective
potential takes similar values of the imaginary parts for the $I = 0$
and $I = 1$ channels.
\subsection{Theoretical ambiguities}
Finally, we discuss theoretical ambiguities in our scenario of the
generation of the $\bar{K} \bar{D} N$ quasibound state.
As we have seen, the $\bar{K} \bar{D} N$ quasibound state is generated
by the two kinds of strong attraction, the $\bar{K} N ( I = 0 )$
interaction and $\bar{K} \bar{D} ( I = 0 )$ interaction, and moderate
$\bar{D} N$ attraction. Among them, the $\bar{D} N$ interaction is
not well determined due to poor experimental data. To check the
influence of the $\bar{D} N$ interaction, we switch off the $\bar{D}
N$ interaction and perform the three-body calculations. As a result,
we obtain the $\bar{K} \bar{D} N$ quasibound state with eigenenergy
$3248 - 21 i \text{ MeV}$, whose value is similar to the full-calculation
value $3244 - 17 i \text{ MeV}$. Therefore, we can say that ambiguity of
the $\bar{D} N$ interaction is irrelevant.
Besides, although the $\bar{K} \bar{D}$ interaction is fixed to
reproduce the $D_{s 0}(2317)$ as the $\bar{K} \bar{D}$ bound state, it
is not clear how much the $D_{s 0}(2317)$ contains a ``bare'' $s
\bar{c}$ component rather than the $\bar{K} \bar{D}$ molecular
component. Such an $s \bar{c}$ component will weaken the attraction
of the $\bar{K} \bar{D}$ effective potential and may affect our
scenario. However, the $D_{s 0}(2317)$ generated in the present
formulation already contains some missing-channel contribution rather
than the $\bar{K} \bar{D}$-$\eta \bar{D}_{s}$ channels via the
intrinsic energy dependence of the interaction~\eqref{eq:V_KDKD0}.
Actually, the intrinsic energy dependence of the $\bar{K}
\bar{D}$-$\eta \bar{D}_{s}$ interaction introduces missing-channel
fraction $= 22 \%$ to the $D_{s 0}(2317)$, as seen in
Ref.~\cite{Navarra:2015iea}, which can be interpreted as a bare $s
\bar{c}$ component. Therefore, our scenario allows the $D_{s
0}(2317)$ to have $\sim 20 \%$ fraction of the bare $s \bar{c}$
component.
Our scenario would be affected by the treatment of the energy
dependence of the two-body interaction in the three-body dynamics [see
Eqs.~\eqref{eq:eps_DN}, \eqref{eq:eps_KN}, and \eqref{eq:eps_KD}].
To check this, we firstly fix two-body energy of only one of the three
pairs in the three-body system to its threshold energy while we keep
the energy dependence for other two pairs. When we keep the energy
dependence for the $\bar{K} \bar{D}$ and $\bar{D} N$ potentials but
fix the $\bar{K} N$ energy as $\epsilon _{\bar{K} N} = m_{K} + m_{N}$,
we obtain the eigenenergy $3242 - 21 i \text{ MeV}$. Similarly, when we fix
the $\bar{K} \bar{D}$ ($\bar{D} N$) energy as $\epsilon _{\bar{K}
\bar{D}} = m_{K} + m_{D}$ ($\epsilon _{\bar{D} N} = m_{D} + m_{N}$),
we obtain the eigenenergy $3238 - 18 i \text{ MeV}$ ($ 3231 - 25 i \text{ MeV}$).
The shifts of the eigenenergy in these cases are not significant
because the energy dependence of the two-body effective potentials is
not essential in the energy region of interest, i.e., around their
thresholds. Secondly, if we fix all the two-body energies to the
two-body threshold energies, $\epsilon _{\bar{D} N} = m_{D} + m_{N}$,
$\epsilon _{\bar{K} N} = m_{K} + m_{N}$, and $\epsilon _{\bar{K}
\bar{D}} = m_{K} + m_{D}$, the eigenenergy becomes $3219 - 34 i
\text{ MeV}$. Thirdly, when we fix the two-body energies to the pole
positions of the two-body bound states, i.e., $\epsilon _{\bar{K} N} =
1424 - 26 i \text{ MeV}$, $\epsilon _{\bar{K} \bar{D}} = 2317 \text{ MeV}$, and
$\epsilon _{\bar{D} N} = 2805 \text{ MeV}$, the eigenenergy becomes $3226 -
28 i \text{ MeV}$. These treatments bring more biding energy and width to
the $\bar{K} \bar{D} N$ state, but the $\bar{K} \bar{D} N$ quasibound
state exists in any case.
From the above discussions, we conclude that the $\bar{K} \bar{D} N$
quasibound state will exist even if we take into account theoretical
ambiguities.
\section{Summary and concluding remarks}
\label{sec:5}
In this study we investigated the $\bar{K} \bar{D} N$ quasibound state
with spin/parity $J^{P} = 1/2^{+}$ and isospin $I = 1/2$ in a
nonrelativistic three-body potential model. Following the approach in
Refs.~\cite{Hyodo:2007jq, Miyahara:2015bya} for the $\bar{K} N$
effective local potential, we constructed the $\bar{K} \bar{D}$ and
$\bar{D} N$ effective local potentials based on phenomenological
models, with which we obtained the $D_{s 0}(2317)$ as a $\bar{K}
\bar{D}$ bound state and $\bar{D} N$ bound state, respectively. These
two-body effective potentials implicitly contain inelastic channels.
In particular, the inclusion of the open channels is essential to
describe the $\bar{K} \bar{D} N$ system as a decaying state.
By solving the three-body Schr\"{o}dinger equation in the $K^{-} \bar{D}^{0}
p$-$\bar{K}^{0} D^{-} p$-$\bar{K}^{0} \bar{D}^{0} n$ coupled channels
with the constructed two-body effective local potentials, we obtained
the $\bar{K} \bar{D} N$ quasibound state with eigenenergy $3244 - 17 i
\text{ MeV}$. The real part of the eigenenergy is below the $D_{s 0} (2317)
+ N$ and $\Lambda (1405) + \bar{D}$ thresholds as well as the $\bar{K}
\bar{D} N$ threshold, so the $\bar{K} \bar{D} N$ quasibound state
cannot decay into $D_{s 0}(2317) + N$, $\Lambda (1405) + \bar{D}$, nor
$\bar{K} \bar{D} N$. From the imaginary part of the eigenenergy, we
calculated the decay width of the $\bar{K} \bar{D} N$ quasibound state
into $\pi \bar{D} \Lambda$, $\pi \bar{D} \Sigma$ to be $34 \text{ MeV}$. In
addition to the three-hadron decay modes, the $\bar{K} \bar{D} N$
quasibound state will have two-hadron decay modes $\bar{K} \bar{D} N
\to \bar{D} \Lambda$, $\bar{D} \Sigma$, and $\bar{D}_{s} N$, and the
branching ratio of the two-hadron decays was estimated to be $\sim 20
\%$.
As for the internal structure, the $\bar{K} \bar{D} N$ quasibound
state takes $[ \bar{K} \bar{D} (I = 0) ] p$ configuration with compact
$\bar{K} \bar{D}$ subsystem because this configuration can maximize
the attraction among three constituents by utilizing the strong
$\bar{K} \bar{D} ( I = 0 )$ attraction fully. We found that the
three-body dynamics may increase distance between two constituents and
hence may reduce decay width of the three-body quasibound state
compared to that of the two-body quasibound state of the constituents,
as is the relation between the $\bar{K} \bar{D} N$ quasibound state
and $\bar{K} N$ quasibound state [$\Lambda (1405)$]. We also
discussed theoretical ambiguities, and conclude that the $\bar{K}
\bar{D} N$ quasibound state will exist even if we take into account
theoretical ambiguities.
Finally, we remark the possibility of the experimental search for the
$\bar{K} \bar{D} N$ quasibound state. Because the $\bar{K} \bar{D} N$
quasibound state has both strangeness $S = -1$ and charm $C = -1$,
practical candidate is the production in relativistic heavy-ion
collisions~\cite{Cho:2010db, Cho:2011ew, Cho:2017dcy}. One can search
for the $\bar{K} \bar{D} N$ quasibound state in, e.g., the $\pi
\bar{D} \Lambda$ and/or $\bar{D} \Lambda$ invariant-mass spectra of
relativistic heavy-ion collisions. The $B$ meson decays in $B$
factories are feasible as well. In this case, for instance, the
$\bar{B}_{s} ( s \bar{b} ) \to \pi \bar{D} \Lambda + \bar{p}$,
$\bar{D} \Lambda + \bar{p}$ processes are suitable.
\begin{acknowledgments}
The authors acknowledge E.~Hiyama for fruitful discussions on the
few-body calculations. The authors are grateful to D.~Jido for
helpful discussions on dynamics which emerges uniquely in three-body
systems.
This work was partly supported by JSPS KAKENHI Grants No.~JP15K17649
and No.~JP18K13545.
\end{acknowledgments}
|
\section{\clearpage\stdsection}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{conjecture}[theorem]{Conjecture}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{question}[theorem]{Question}
\newtheorem{example}[theorem]{Example}
\newtheorem{convention}[theorem]{Convention}
\numberwithin{equation}{section}
\oddsidemargin 0in
\textwidth 6.5in
\textheight 9in
\topmargin 0in
\headheight 0in
\headsep 0in
\newcommand{\mathrm{crit}}{\mathrm{crit}}
\newcommand{\mathrm{subcrit}}{\mathrm{subcrit}}
\renewcommand{\H}{\mathcal H}
\newcommand{\mathbf 1}{\mathbf 1}
\newcommand{\mathcal J}{\mathcal J}
\newcommand{\mathcal A}{\mathcal A}
\newcommand{\mathcal B}{\mathcal B}
\newcommand{\mathcal C}{\mathcal C}
\newcommand{\mathcal E}{\mathcal E}
\newcommand{A_\infty}{A_\infty}
\newcommand{\mathcal S}{\mathcal S}
\newcommand{\mathbb L}{\mathbb L}
\newcommand{\mathcal L}{\mathcal L}
\newcommand{\mathbb Q}{\mathbb Q}
\newcommand{\mathbb Z}{\mathbb Z}
\newcommand{\mathbb C}{\mathbb C}
\newcommand{\mathbf C}{\mathbf C}
\newcommand{\mathbf D}{\mathbf D}
\newcommand{\mathbb R}{\mathbb R}
\newcommand{\mathbb P}{\mathbb P}
\newcommand{\mathbf R}{\mathbf R}
\newcommand{\mathcal Q}{\mathcal Q}
\newcommand{\mathcal M}{\mathcal M}
\newcommand{\mathcal N}{\mathcal N}
\newcommand{\mathcal U}{\mathcal U}
\newcommand{\mathcal T}{\mathcal T}
\newcommand{\mathcal G}{\mathcal G}
\newcommand{\mathcal X}{\mathcal X}
\newcommand{\mathfrak f}{\mathfrak f}
\newcommand{\mathfrak g}{\mathfrak g}
\newcommand{\mathfrak c}{\mathfrak c}
\newcommand{\overline{\mathcal M}}{\overline{\mathcal M}}
\newcommand{\overline{\mathcal C}}{\overline{\mathcal C}}
\newcommand{\overline{\mathcal R}}{\overline{\mathcal R}}
\newcommand{\overline{\mathcal S}}{\overline{\mathcal S}}
\newcommand{\mathcal O}{\mathcal O}
\newcommand{\mathcal D}{\mathcal D}
\newcommand{\mathcal W}{\mathcal W}
\newcommand{\mathrm{std}}{\mathrm{std}}
\newcommand{\operatorname{im}}{\operatorname{im}}
\newcommand{\operatorname{Op}}{\operatorname{Op}}
\newcommand{\operatorname{id}}{\operatorname{id}}
\newcommand{\operatorname{Tw}}{\operatorname{Tw}}
\newcommand{\mathcal F}{\mathcal F}
\newcommand{\operatorname{End}}{\operatorname{End}}
\newcommand{\operatorname{cone}}{\operatorname{cone}}
\newcommand{\operatorname{\mathsf{cones}}}{\operatorname{\mathsf{cones}}}
\newcommand{\operatorname{coker}}{\operatorname{coker}}
\newcommand{[\operatorname{fiber}]}{[\operatorname{fiber}]}
\newcommand{[\operatorname{cocore}]}{[\operatorname{cocore}]}
\newcommand{\mathrm{op}}{\mathrm{op}}
\newcommand{\mathrm{pt}}{\mathrm{pt}}
\newcommand{\operatorname{Hom}}{\operatorname{Hom}}
\newcommand{\operatorname{Fun}}{\operatorname{Fun}}
\newcommand{\operatorname{Spin}}{\operatorname{Spin}}
\newcommand{\operatorname{Pin}}{\operatorname{Pin}}
\newcommand{\operatornamewithlimits{colim}}{\operatornamewithlimits{colim}}
\newcommand{\operatornamewithlimits{hocolim}}{\operatornamewithlimits{hocolim}}
\newcommand{\mathrm{sm}}{\mathrm{sm}}
\newcommand{\operatorname{Perf}}{\operatorname{Perf}}
\newcommand{\operatorname{Ch}}{\operatorname{Ch}}
\newcommand{\operatorname{Coh}}{\operatorname{Coh}}
\newcommand{\operatorname{Mod}}{\operatorname{Mod}}
\newcommand{\operatorname{Prop}}{\operatorname{Prop}}
\newcommand{\operatorname{\mathcal LGr}}{\operatorname{\mathcal LGr}}
\newcommand{\operatorname{\widetilde{\mathcal LGr}}}{\operatorname{\widetilde{\mathcal LGr}}}
\newcommand{\operatorname{Sh}}{\operatorname{Sh}}
\newcommand{\operatorname{Wh}}{\operatorname{Wh}}
\renewcommand{\star}{\operatorname{star}}
\renewcommand{\ss}{\operatorname{ss}}
\newcommand{\operatorname{\mathsf{Set}}}{\operatorname{\mathsf{Set}}}
\begin{document}
\title{Microlocal Morse theory of wrapped Fukaya categories}
\author{
Sheel Ganatra,
John Pardon,
and
Vivek Shende
}
\date{24 September 2018}
\maketitle
\begin{abstract}
Consider on the one hand the partially wrapped Fukaya category of a cotangent bundle stopped at an appropriately stratifiable singular isotropic.
Consider on the other hand the derived category of sheaves with microsupport along the singular isotropic.
We show here that the former is equivalent to the compact objects of the latter.
Applications include a sheaf-theoretic description of the wrapped Fukaya categories of plumbings and of codimension one Weinstein hypersurfaces
of cosphere bundles.
While inspired by the Nadler--Zaslow correspondence, our results do not depend on it.
\end{abstract}
\section{Introduction}
Let $X$ be a Liouville manifold (or sector) and $\Lambda \subseteq \partial_\infty X$ a closed subset at infinity. We write $\mathcal W(X, \Lambda)$
for the Fukaya category in which objects are Lagrangians disjoint at infinity from $\Lambda$ and in which morphisms are defined by wrapping
in the complement of $\Lambda$. This category is defined and studied in \cite{abouzaidseidel,sylvanthesis, gpssectorsoc, gpsstructural}, and one has particularly good control over it when $X$ is Weinstein and $\Lambda$ is a (possibly singular) isotropic.
Our interest here is in the case of cotangent bundles, $X = T^*M$. In this case, we can also study
the sheaves on $M$, microsupported inside $\Lambda$.
We write $\operatorname{Sh}(M)$ for the (dg-)category of unbounded complexes of sheaves of $\mathbb Z$-modules on $M$, localized
along the stalkwise quasi-isomorphisms. The data of an object $\mathcal F \in \operatorname{Sh}(M)$ is the assignment of some complex of $\mathbb Z$-modules $\mathcal F(U)$ (called the sections of $\mathcal F$ over $U$) to every open set $U \subseteq M$, along with restriction maps $\mathcal F(U) \to \mathcal F(V)$ for inclusions $V \subseteq U$.
The microsupport of a sheaf $\mathcal F$ is a closed conical locus $\ss(F) \subseteq T^*M$. Its role is to encode
when restriction maps are quasi-isomorphisms. The basic idea is illustrated by the fact \cite[Corollary 5.4.19]{kashiwara-schapira}
that for a smooth function $\phi: M \to \mathbb R$ and real numbers $a < b$, one has
$\mathcal F(\phi^{-1}(-\infty, b)) \xrightarrow{\sim} \mathcal F(\phi^{-1}(-\infty, a))$ whenever the graph of $d \phi$ is
disjoint from $\ss(F)$ over $\phi^{-1}([a,b))$. For example, armed with the further knowledge that the microsupport of the constant sheaf
is the zero section, one may conclude from this that the cohomology of sublevel sets of $\phi$ changes only at levels where $d \phi = 0$.
We prove here the following result:
\begin{theorem}\label{sheaffukayaequivalence}
Let $M$ be a real analytic manifold, and let $\Lambda \subseteq T^*M$ be a subanalytic closed isotropic subset.
Then $\operatorname{Perf}\mathcal W(T^*M,\Lambda)^\mathrm{op}$ is equivalent to the category of compact objects in $\operatorname{Sh}_\Lambda(M)$.
Moreover, this equivalence carries the linking disk at any smooth Legendrian point $p\in\Lambda$ to a co-representative of the microstalk functor at $p\in\Lambda$.
\end{theorem}
Rather than writing $\mathcal W(T^*M,\Lambda)^\mathrm{op}$, one could equivalently negate the Liouville form and write $\mathcal W(-T^*M,\Lambda)$,
or pull back by the antipodal map and write $\mathcal W(T^*M, -\Lambda)$. We also remind the reader that the notation $\operatorname{Perf}$ refers to the idempotent-completed pre-triangulated closure of an $A_\infty$ or dg category (see \S \ref{perfpropsec}).
The reader is cautioned that the compact objects of $\operatorname{Sh}_\Lambda(M)$ \emph{do not} necessarily have perfect stalks or bounded homological degree.
That is, they need not be constructible sheaves in the usual sense. The necessity of considering such objects on the sheaf side was pointed
out in \cite{nadler-wrapped}, where the above result was implicitly conjectured by the title.
The main idea of the proof of Theorem \ref{sheaffukayaequivalence} is to first establish the result by direct calculation when $\Lambda=N^*_\infty\mathcal S$ is the union of conormals of strata of a triangulation $\mathcal S$ of $M$, and then to show that both sides transform in the same way when $\Lambda$ becomes smaller.
In Section \ref{sec:axioms}, we formulate these two steps as axioms for a system of categories $\Lambda\mapsto\mathcal C(\Lambda)$ parameterized by closed subanalytic singular isotropics $\Lambda\subseteq S^*M$.
The remainder of the paper consists in establishing these axioms on both sides.
On the sheaf side, these amount to the (standard) fact that the category of sheaves constructible with respect to an appropriate stratification is the representation category of the poset of the stratification, plus stratified Morse theory \cite{goresky-macpherson}
as reformulated microlocally with sheaf coefficients in \cite{kashiwara-schapira}. We describe this
in detail in Section \ref{sec:sheaves}. On the Fukaya
side, the analogue of this microlocal Morse theory is the wrapping exact triangle of \cite{gpsstructural}.
The structural results of \cite{gpsstructural} are also fundamentally useful in the computation of the Fukaya category of a cotangent bundle, with wrapping stopped by the union of conormals to a triangulation.
This is carried out in Section \ref{sec:fukaya}.
That there should be a relationship between constructible sheaves on $M$ and the Fukaya category of $T^*M$ was first suggested by Nadler and Zaslow
in \cite{nadlerzaslow,nadlermicrolocal}. The appeal of this statement was that the category of constructible sheaves
is essentially combinatorial, leading readily to computations.
By contrast with their work, our theorem concerns \emph{wrapped}, rather than infinitesimal, Fukaya categories.
Wrapped Floer cohomology
and the wrapped Fukaya category are more complicated, global, typically infinite rank, invariants than their infinitesimal
counterparts studied in \cite{nadlerzaslow,nadlermicrolocal}. From this point of view, it may seem at first surprising that sheaves
can model both infinitesimal and wrapped categories. This possibility was suggested by \cite{nadler-wrapped}; as noted there
one can expect to recover the infinitesimal category from the wrapped category, but not vice versa.
A discussion
from our point of view, recovering a version of the original Nadler--Zaslow correspondence from Theorem \ref{sheaffukayaequivalence},
appears in Section \ref{NZ}.
Because we treat the wrapped category, our result has broader implications than
the original \cite{nadlerzaslow, nadlermicrolocal}. In particular, it is
necessary to use the wrapped category on the $A$-model
side of mirror symmetry to match categories of coherent sheaves
when the $B$-model side is not both smooth and compact.
Additionally, the connection between sheaves and Legendrian contact homology is
best understood from the
wrapped perspective. We detail in
Section \ref{sec:corollaries} the extent to which various sheaf calculations
\cite{FLTZ-Morelli, shende-treumann-zaslow, STWZ, shende-treumann-williams, kuwagaki, nadler-mirrorLG,
nadler-wrapped, gammage-shende} can now be understood as computations of
wrapped Fukaya categories. We also give a new computation of the wrapped Fukaya categories of cotangent bundles
(in particular generalizing
\cite{abouzaidcotangent, abouzaidtwisted} to the case where $M$ may be non-compact or have boundary), and we
compute the wrapped Fukaya categories of plumbings.
In the proof of Theorem \ref{sheaffukayaequivalence}, we \emph{do not} rely on the results \cite{nadlerzaslow,nadlermicrolocal} of Nadler and Zaslow.\footnote{We emphasize this because some symplectic
geometers do not view \cite{nadlerzaslow} as fully rigorous; the worry involves the construction of the infinitesimal Fukaya category given in \cite{nadlerzaslow}. We do not
wish to express any opinion on this matter, beyond noting that we are aware of it and that it does not affect the present results.}
In fact it would not have helped to do so:
beyond the foundations we set up in \cite{gpssectorsoc, gpsstructural},
the main new ideas and difficulties in the proof of Theorem
\ref{sheaffukayaequivalence} concern the geometry of wrapping (as underlies the results of \cite{gpsstructural}), as opposed to the calculation of Floer cohomology.
In fact, the only Floer cohomology calculations which need to be made in this entire article are between Lagrangians which intersect in at most one point.
In particular, we expect the proofs of the results in this paper, in particular Theorem \ref{sheaffukayaequivalence}, would apply in the case of more general (e.g.\ sphere spectrum) coefficients, provided one has access to the definitions of the sheaf and Fukaya categories, respectively, in these settings.
\vspace{2mm}
{\bf{Convention.}} Throughout this document, we work in the setting of dg- and, equivalently, $A_\infty$-categories over $\mathbb Z$ (or more generally any commutative ring). We only ever
consider ``derived'' functors, we only ever mean ``homotopy'' limits or colimits,
and we systematically omit the word ``quasi''. By modules,
we mean dg- or $A_\infty$-modules, e.g.\ by $\mathbb Z$-modules we mean the category of chain complexes of abelian $\mathbb Z$-modules, localized at quasi-isomorphisms,
except when, as in this sentence, we qualify it with the word `abelian'. In \S \ref{categoricalsection} we detail our assumptions about these categories and collect relevant categorical notions which will appear throughout the paper.
\subsection{Acknowledgements}
We thank Mohammed Abouzaid, Roger Casals, Tobias Ekholm, Benjamin Gammage, David Nadler, Amnon Neeman, and Lenhard Ng
for helpful discussions, some of which took place during visits to the American Institute of Mathematics.
S.G.\ would like to thank the Mathematical Sciences Research Institute for its
hospitality during a visit in Spring 2018 (supported by NSF grant DMS--1440140)
during which some of this work was completed.
This research was conducted during the period J.P.\ served as a Clay Research Fellow and was partially supported by a Packard Fellowship and by the National Science Foundation under the Alan T.\ Waterman Award, Grant No.\ 1747553.
V.S.\ was supported by the NSF CAREER grant DMS--1654545.
\section{Stratifications}
Let $X$ be a topological space. By a stratification $\mathcal S$ of $X$, we mean a locally finite decomposition into disjoint locally closed subsets $\{X_\alpha\}_{\alpha\in\mathcal S}$, called
strata, such that each boundary $\overline{X_\alpha}\setminus X_\alpha$ is a union of other strata $X_\beta$.
The collection of strata $\mathcal S$ is naturally a poset, in which there is a map $\beta \to \alpha$ iff $X_\alpha \subseteq \overline{X_\beta}$.
When $X$ is a manifold, we implicitly require the strata to be as well.
\begin{remark}\label{stratvsposet}
The poset $\mathcal S$ does not generally reflect the topology of the space $X$.
Conditions under which it does (contractibility of various strata/stars) are well known and recalled below.
\end{remark}
We will say a subset $Y \subseteq X$ is $\mathcal S$-constructible when it is a union of strata of the stratification $\mathcal S$ of $X$.
We say that a stratification $\mathcal T$ refines a stratification $\mathcal S$ when the strata of $\mathcal S$ are $\mathcal T$-constructible.
We recall that an abstract simplicial complex on a vertex set $\mathcal{V}$ is a collection $\Sigma$
of nonempty finite subsets of $\mathcal{V}$, containing all singletons and all subsets of elements of $\Sigma$. By a simplicial complex,
we mean the geometric realization $\left|\Sigma\right|$ of an abstract simplicial complex $\Sigma$; it comes with a stratification by the `open simplices' (which, of course, are locally closed, not necessarily open, subsets of $\left|\Sigma\right|$).
We say a stratification $\mathcal S$ on $X$ is a triangulation when there is a homeomorphism $\left|\Sigma\right|\xrightarrow\sim X$ identifying stratifications.\footnote{Strictly speaking, it would be better to define a triangulation as a simplicial complex $(\mathcal V,\Sigma)$ together with a map $\left|\Sigma\right|\xrightarrow\sim X$ identifying stratifications. This distinction, however, will not concern us.}
Note the following are \emph{not} triangulations:
a stratification of a circle into single point and its complement, or into two points and their complement.
The stratification into three points and their complement is a triangulation.
The open star of a stratum is the union of strata whose closures contain it. Taking stars reverses the inclusion: we have
$X_\alpha \subseteq \overline{X_\beta} \iff \star(X_\beta) \subseteq \star(X_\alpha)$.
Note that $\star(X_\alpha) \cap \star(X_\beta) = \bigcup_{\alpha, \beta \to \gamma} \star(X_\gamma)$.
For triangulations, we can do better: $\star(X_\alpha) \cap \star(X_\beta) = \star(X_\gamma)$ where $\gamma$ is the simplex spanned by the vertices of $\alpha$ union the vertices of $\beta$ (if this simplex is present), and otherwise $\star(X_\alpha) \cap \star(X_\beta) =\emptyset$.
Given a $C^1$ stratification $\mathcal S$ of a $C^1$ manifold $M$, we write $N^* \mathcal S \subseteq T^*M$ for the union of conormals to the strata.
The stratification is said to be Whitney (a) iff $N^* \mathcal S$ is closed; or, as it is usually formulated, if limits of tangent spaces to strata contain
the tangent spaces of their boundary strata. In any setting where we consider cotangent bundles, we will only ever consider
Whitney (a) stratifications.
Eventually we will restrict to the setting of analytic manifolds and subanalytic stratifications.
We recall that a set is defined to be subanalytic when locally (i.e.\ in a neighborhood of every point of its closure) it is the analytic image of a relatively compact semianalytic set.
The canonical modern reference is \cite{bierstone-milman}.\footnote{Wherever we have written `subanalytic', one could substitute any fixed analytic-geometric category, so long as one is willing (we are) to have strata that are $C^p$ for arbitrarily large $p$ but not necessarily $C^\infty$ \cite{vandendries-miller}.}
By a subanalytic stratification, we mean a stratification in which all strata are subanalytic.
By a subanalytic triangulation, we mean a triangulation $f:\left|\Sigma\right|\xrightarrow\sim X$, where the restriction of $f$ to every closed simplex is analytic and the restriction of $f$ to every open simplex is an immersion.
It is a fundamental result that for any locally finite subanalytic partition $M = \coprod M_\alpha$, there is a subanalytic
triangulation in which all $M_\alpha$ are constructible.
See \cite{bierstone-milman} for proofs of the above results.
In \cite{kashiwara-schapira} there is a notion of $\mu$-stratification, this being a certain strengthening
of the Whitney conditions; it is shown in \cite[Thm.\ 8.3.20]{kashiwara-schapira} that any subanalytic stratification can
be refined to a $\mu$-stratification.
\section{Microlocal Morse categories} \label{sec:axioms}
\subsection{Strata poset categories and refinement functors}
Let $\mathcal S$ be a stratification.
We fix the following notation for the Yoneda embedding:
\begin{align}
\mathcal S &\to \operatorname{Fun}(\mathcal S^\mathrm{op},\operatorname{\mathsf{Set}}),\\
\alpha&\mapsto\operatorname{Hom}(\cdot,\alpha)=:1_{\star(\alpha)}.
\end{align}
Note that
\begin{equation}
\operatorname{Hom}(1_{\star(\alpha)},1_{\star(\beta)}) = 1_{\star(\beta)}(\alpha) = \operatorname{Hom}(\alpha, \beta) = \begin{cases} \{1\} & \star(\alpha) \subseteq \star(\beta) \\ \emptyset & \mathrm{otherwise}. \end{cases}
\end{equation}
For any $\mathcal S$-constructible open set $U$, we introduce the functor $1_U\in\operatorname{Fun}(\mathcal S^\mathrm{op},\operatorname{\mathsf{Set}})$ defined by the analogous formula
\begin{equation}\label{indicatorUdef}
\operatorname{Hom}(1_{\star(\alpha)}, 1_{U}) = 1_{U}(\alpha) := \begin{cases} \{1\} & \star(\alpha) \subseteq U \\ \emptyset & \mathrm{otherwise}. \end{cases}
\end{equation}
Note that $\star(\alpha) \subseteq U$ iff $\alpha\subseteq U$.
Now let $\mathcal S'$ be a stratification refining $\mathcal S$. There is a natural map $r:\mathcal S'\to\mathcal S$, sending a stratum in $\mathcal S'$ to the
unique stratum in $\mathcal S$ containing it.
We write
$$r^*:\operatorname{Fun}(\mathcal S^\mathrm{op},\operatorname{\mathsf{Set}})\to\operatorname{Fun}(\mathcal S^{\prime\mathrm{op}},\operatorname{\mathsf{Set}})$$
for the pullback of functors along this map $r$.
For $\tau' \in \mathcal S'$ and an $\mathcal S$-constructible open set $U$, we have
\begin{equation}
\operatorname{Hom}(1_{\star(\tau')}, r^*1_U) = (r^*1_U)(\tau') = 1_U(r(\tau')) =
\begin{cases} \{1\} & \star(r(\tau')) \subseteq U \\ \emptyset & \mathrm{otherwise}. \end{cases}
\end{equation}
Since $U$ is open and $\mathcal S$-constructible, we have $\star(r(\tau')) \subseteq U$ iff $\star(\tau')\subseteq U$, so we conclude that $r^*1_U=1_U$.
We now linearize. We write $\mathbb Z[\mathcal S]$ for the linearization of a poset $\mathcal S$.
We write $\operatorname{Mod}\mathcal S$ for the category of modules $\operatorname{Fun}(\mathcal S^\mathrm{op},\operatorname{Mod}\mathbb Z) = \operatorname{Fun}(\mathbb Z[\mathcal S]^\mathrm{op}, \operatorname{Mod} \mathbb Z)$, and we use $r^*:\operatorname{Mod}\mathcal S\to\operatorname{Mod}\mathcal S'$ for pullback of modules as above.
As with any pullback of modules, this functor has a left adjoint given by extension of scalars,
which by abuse of notation we write as $r:\operatorname{Mod}\mathcal S'\to\operatorname{Mod}\mathcal S$ due to the commuting diagram
\begin{equation}
\begin{tikzcd}[column sep = large]
\mathcal S'\ar{r}{s\mapsto 1_{\star(s)}}\ar{d}[swap]{r}& \operatorname{Mod}\mathcal S'\ar{d}{r}\\
\mathcal S\ar{r}{s\mapsto 1_{\star(s)}}& \operatorname{Mod}\mathcal S.
\end{tikzcd}
\end{equation}
Restriction of scalars $r^*$ is co-continuous, so its left adjoint $r$ extension of scalars preserves compact objects, giving a map $r:\operatorname{Perf}\mathcal S'\to\operatorname{Perf}\mathcal S$ (which can also be viewed as the canonical extension of $r:\mathcal S'\to\mathcal S$ to the idempotent completed pre-triangulated hulls).
\subsection{A category for any \texorpdfstring{$\Lambda$}{Lambda}}
We now wish to define a \emph{microlocal Morse category} $\mathcal C(\Lambda)$ for any subanalytic singular isotropic $\Lambda\subseteq S^\ast M$, together with functors $\mathcal C(\Lambda')\to\mathcal C(\Lambda)$ for inclusions $\Lambda'\supseteq\Lambda$.
We define this system of categories $\Lambda\mapsto\mathcal C(\Lambda)$, the \emph{microlocal Morse theatre}, by formulating axioms which characterize it uniquely.
The previous subsection defined categories $\operatorname{Perf}\mathcal S$ together with functors $r:\operatorname{Perf}\mathcal S'\to\operatorname{Perf}\mathcal S$ whenever $\mathcal S'$ is a refinement of $\mathcal S$.
For our current purpose, these categories do not have the correct significance for general stratifications $\mathcal S$ (compare Remark \ref{stratvsposet}).
As such, we will consider these categories only for triangulations $\mathcal S$.\footnote{In fact, there are weaker conditions on a stratification $\mathcal S$ (which are satisfied if $\mathcal S$ is a triangulation) implying that $\operatorname{Perf}\mathcal S$ is the correct category to associate to $\mathcal S$.}
The microlocal Morse theatre is an extension of this functor $\mathcal S\mapsto\operatorname{Perf}\mathcal S$ on triangulations in the following sense:
\begin{definition}\label{categorysystemextension}
A \emph{microlocal Morse pre-theatre} $\Lambda\mapsto\mathcal C(\Lambda)$ is a functor from the category of subanalytic singular isotropics inside $S^\ast M$ to the category of dg-categories over $\mathbb Z$, together with an isomorphism of functors $(\mathcal S\mapsto\mathcal C(N^*_\infty\mathcal S))=(\mathcal S\mapsto\operatorname{Perf}\mathcal S)$ on $\mu$-triangulations $\mathcal S$.
\end{definition}
\begin{remark}
Any isomorphism of functors $(\mathcal S\mapsto H^*\mathcal C(N^*_\infty\mathcal S))=(\mathcal S\mapsto H^*\operatorname{Perf}\mathcal S)$ automatically lifts to an isomorphism $(\mathcal S\mapsto\mathcal C(N^*_\infty\mathcal S))=(\mathcal S\mapsto\operatorname{Perf}\mathcal S)$ by Lemma \ref{ffhenceainf}.
This will be crucial when discussing Fukaya categories.
\end{remark}
We will characterize the microlocal Morse theatre in terms of microlocal Morse theory.\footnote{More conventionally \cite{goresky-macpherson},
this is called stratified Morse theory. We find the term `microlocal' more descriptive, and also the word stratified would otherwise
take on too many meanings in this article.}
Let $f:M\to\mathbb R$ be a function and $\mathcal S$ a stratification.
An intersection of $\Gamma_{df}$ with $N^*\mathcal S$ is called an $\mathcal S$-critical point, which is said to be Morse if it is a transverse intersection at a smooth point of $N^*\mathcal S$.
The function $f$ is said to be $\mathcal S$-Morse when all its $\mathcal S$-critical points are Morse.
When $\mathcal S$ is subanalytic, such functions are plentiful, and can be chosen analytic.
(See \cite[Thm.\ 2.2.1]{goresky-macpherson} for this assertion, which is collected there from various results in the literature.)
More generally, for any singular isotropic $\Lambda\subseteq S^*M$, a $\Lambda$-critical point of $f$ is an intersection of $\Gamma_{df}$ with the union of the zero section and the cone over $\Lambda$, it is said to be Morse if transverse and at a smooth point, and any $f$ whose $\Lambda$-critical points are all Morse is called $\Lambda$-Morse.
\begin{definition} \label{morsecharacter}
In any microlocal Morse pre-theatre $\Lambda\mapsto\mathcal C(\Lambda)$, the \emph{Morse characters} $\mathcal X_{\Lambda,p}(f,\epsilon,\mathcal S)\in\mathcal C(\Lambda)$ are defined as follows for smooth Legendrian points $p\in\Lambda$.
Let $f:M\to\mathbb R$ be an analytic function with a Morse $\Lambda$-critical point at $p$ with critical value $0$, no other $\Lambda$-critical points with critical values in the interval $[-\epsilon,\epsilon]$, and with relatively compact sublevel set $f^{-1}(-\infty,\epsilon)$.
Let $\mathcal S$ be a triangulation for which $\Lambda\subseteq N^*_\infty\mathcal S$ and for which both $f^{-1}(-\infty, -\epsilon)$ and $f^{-1}(-\infty, \epsilon)$ are $\mathcal S$-constructible.
The Morse character $\mathcal X_{\Lambda,p}(f,\epsilon,\mathcal S)$ is then defined as the image of
\begin{equation}
\operatorname{cone}(1_{f^{-1}(-\infty, -\epsilon)} \to 1_{f^{-1}(-\infty, \epsilon)}) \in \operatorname{Perf}\mathcal S=\mathcal C(N^*_\infty\mathcal S).
\end{equation}
under the map $\mathcal C(N^*_\infty\mathcal S)\to\mathcal C(\Lambda)$.
\end{definition}
The Morse character $\mathcal X_{\Lambda,p}(f,\epsilon,\mathcal S)\in\mathcal C(\Lambda)$ depends \emph{a priori} on the `casting directors' $(f,\epsilon,\mathcal S)$.
Casting directors $(f,\epsilon)$ exist at any smooth Legendrian point $p\in\Lambda$ by general position, and $\mathcal S$ exists by the following result:
\begin{proposition}[{\cite[8.3.10]{kashiwara-schapira}}]\label{putlambdainconormal}
For any closed subanalytic singular isotropic $\Lambda\subseteq S^*M$, there exists a subanalytic stratification $\mathcal S$ of $M$ such that $\Lambda\subseteq N^*_\infty\mathcal S$.\qed
\end{proposition}
\begin{definition}\label{microlocalmorsetheatre}
A \emph{microlocal Morse theatre} is a microlocal Morse pre-theatre $\Lambda\mapsto\mathcal C(\Lambda)$ such that for any inclusion $\Lambda\subseteq\Lambda'$ and any collection of Morse characters $\mathcal X_{\Lambda',p}(f,\epsilon,\mathcal S)\in\mathcal C(\Lambda')$ at smooth Legendrian points $p\in\Lambda'\setminus\Lambda$ with at least one in every component of the smooth Legendrian locus of $\Lambda'\setminus\Lambda$, the functor $\mathcal C(\Lambda')\to\mathcal C(\Lambda)$ is (the idempotent completion of) the quotient by these Morse characters.
\end{definition}
The definition of a microlocal Morse theatre allows one to readily compute any particular microlocal Morse category $\mathcal C(\Lambda)$: embed $\Lambda$ into some $N^*_\infty\mathcal S$ using Proposition \ref{putlambdainconormal}, cast Morse characters in $\mathcal C(N^*_\infty\mathcal S)=\operatorname{Perf}\mathcal S$ for all Legendrian components of $N^*_\infty\mathcal S\setminus\Lambda$, and take the quotient of $\operatorname{Perf}\mathcal S$ by these characters and idempotent complete.
It follows that:
\begin{proposition}
Any two microlocal Morse theatres $\Lambda\mapsto\mathcal C(\Lambda)$ are uniquely isomorphic.\qed
\end{proposition}
A \emph{dramatic realization} is a particular construction of the microlocal Morse theatre $\Lambda\mapsto\mathcal C(\Lambda)$.
We give two dramatic realizations, namely via sheaves and via Lagrangians in Sections \ref{sec:sheaves} and \ref{sec:fukaya}, respectively.
Both these dramatic realizations \emph{cast} the Morse characters as certain familiar objects.
They moreover show that the Morse characters in fact depend only on $p$ and are independent of the casting directors.
\begin{theorem}
The microlocal Morse theatre $\Lambda\mapsto\mathcal C(\Lambda)$ exists, and the Morse characters $\mathcal X_{\Lambda,p}\in\mathcal C(\Lambda)$ are independent of the casting directors and form a local system over the smooth Legendrian locus of $\Lambda$.
\end{theorem}
\begin{proof}
This follows from either Theorem \ref{sheafrealization} or Theorem \ref{fukcasting}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{sheaffukayaequivalence}]
Combine Theorems \ref{sheafrealization} and Theorem \ref{fukcasting}.
\end{proof}
In fact, both dramatic realizations show that $\mathcal C(\Lambda)$ is invariant under contact isotopy of $S^*M$, something which is not apparent from the present combinatorial prescription.
This is immediate on the Fukaya side, and on the sheaf side it is `sheaf quantization' \cite{guillermou-kashiwara-schapira}.
In fact, there are even stronger invariance statements: it is shown in \cite{gpsstructural} that in fact $\mathcal C(\Lambda)$ is invariant under isotopy of $S^*M \setminus \Lambda$ inside $S^*M$;
meanwhile, it is shown in \cite{nadler-shende} that $\mathcal C(\Lambda)$ is invariant under ``non-characteristic deformations'' of $\Lambda$.
\begin{remark}
The construction of this subsection makes sense in any stable setting, e.g.\ over the sphere spectrum.
To show existence of the microlocal Morse theatre in such a more general setting, one could set up either microlocal sheaf theory or the Fukaya category over the sphere spectrum.
In principle, one could also show existence directly from the stratified Morse theory of \cite{goresky-macpherson}, as it already establishes results about homotopy types of spaces (not just their cohomologies).
A more interesting question is whether any symplectically invariant statement can be made beyond the stable setting.
\end{remark}
\section{Sheaf categories} \label{sec:sheaves}
We recall the general formalism of sheaves, and properties of stratifications. We then recall from \cite{kashiwara-schapira}
the notion of microsupport, and the category $\operatorname{Sh}_\Lambda(M)$ of sheaves on $M$ whose microsupport at infinity is contained in $\Lambda$.
We show that the assignment $\Lambda \mapsto \operatorname{Sh}_\Lambda(M)$ is a microlocal Morse theatre in the sense of Definition \ref{microlocalmorsetheatre}.
\subsection{Categories of sheaves and functors between them}
Here we give a brief review of the general formalism of sheaves. Our presentation is somewhat modern in that we \emph{never} discuss
sheaves of abelian groups, rather we work at the dg level and with unbounded complexes from the beginning, but it is essentially the same as
any standard account such as \cite{iversen, kashiwara-schapira, schapira-notes}, complemented by \cite{spaltenstein} in order to work
with unbounded complexes, and in particular for the proper base change theorem in this setting.
Given a topological space $T$, we write $\operatorname{Op}(T)$ for the category whose objects are open sets and morphisms are inclusions.
A ($\mathbb Z$-module valued) presheaf on $T$ is by definition a functor $\operatorname{Op}(T)^\mathrm{op} \to \operatorname{Mod}\mathbb Z$.
In particular, a presheaf $\mathcal F$ takes a value $\mathcal F(U) \in \operatorname{Mod}\mathbb Z$ on an open set $U \subseteq T$, termed its sections; given open sets $U \subseteq V$ it gives
a morphism $\mathcal F(V) \to \mathcal F(U)$, termed the restriction, etcetera. Given any subset $X\subseteq T$, we write $\mathcal F(X) = \varinjlim_{X \subseteq U} \mathcal F(U)$;
when $X$ is a point, this is termed the stalk and is written $\mathcal F_x$.
The category of sheaves is the full subcategory of presheaves on objects $\mathcal F$ taking covers to limits:
$$\mathcal F\left(\bigcup_{i \in I} U_i\right) \xrightarrow{\sim} \varprojlim_{\emptyset\ne J \subseteq I} \mathcal F\left( \bigcap_{j \in J} U_j \right)$$
It is also the localization of the category of presheaves along the stalkwise quasi-isomorphisms. The composition of these
two operations is termed ``sheafification'',
giving, for any presheaf $\mathcal F^{\mathrm{pre}}$, a sheaf $\mathcal F$ such that any map from
$\mathcal F^{\mathrm{pre}}$ to a sheaf factors uniquely through $\mathcal F$.
We write $\operatorname{Sh}(T)$ for the (dg) category of sheaves of (dg) $\mathbb Z$-modules on $T$. It is complete and cocomplete.
It has very few compact
objects, but is well generated
\cite{neeman-sheaves, neeman-book}.\footnote{Well generation is a significantly weaker version of compact generation, which is
nevertheless sufficient to appeal to Brown representability, i.e. the assertion that co-continuous functors are representable; that
consequently co-continuous functors are left adjoints and continuous functors are right adjoints, etc.}
Its homotopy category is what was classically called the unbounded derived category of sheaves on $T$.
For any continuous map $f: S \to T$, there is an adjoint pair
$f^*: Sh(T) \leftrightarrow Sh(S): f_*$. The pushforward $f_*$ is given by the formula
$(f_* \mathcal F)(U) = \mathcal F(f^{-1}(U))$, while the pullback $f^*$ is the sheafification of the presheaf given by
$(f^*\mathcal G)(V) = \mathcal G(f(V))$.
\begin{example}
Consider $f: S \to \mathrm{point}$, and the constant sheaf $\mathbb Z_S := f^* \mathbb Z$. Note that in
our conventions, $\mathbb Z_S(U)$ is a chain complex computing the cohomology of $U$. This should illustrate
where, in this account of sheaf theory, is hiding the usual homological algebra of resolutions: it is in the sheafification.
\end{example}
Being a left adjoint, $f^*$ is cocontinuous
(preserves colimits, in particular, sums). When $j: U \to T$ is the inclusion of an open set, $j^*$ is given by
the simpler formula $(j^*\mathcal F)(V) = \mathcal F(V)$, no sheafification required, and hence preserves limits
as well. In particular it must also be a right adjoint. The corresponding left adjoint is easy to describe:
$$j_!\mathcal F(V) = \begin{cases} \mathcal F(V) & V \subseteq U \\ 0 & \mathrm{otherwise} \end{cases}$$
The sheaf $j_! \mathcal F$ is termed the extension by zero, since its stalks in $U$ are isomorphic to the corresponding stalks
of $\mathcal F$, and its stalks outside of $U$ are zero.
For a sheaf $\mathcal F$ on $T$, we write $\mathcal F_U := j_! j^* \mathcal F$. By adjunction there is a canonical morphism
$\mathcal F_U \to \mathcal F$. The object $\mathbb Z_U$ co-represents the functor of sections over $U$, i.e. $\operatorname{Hom}(\mathbb Z_U,\mathcal F)=\mathcal F(U)$.
Being a right adjoint, $f_*$ is continuous. When $f$ is proper, it is in addition co-continuous. More generally,
for a morphism of locally compact spaces $f: S \to T$, one defines\footnote{This particular way of defining the
$!$ pushforward is taken from \cite{schapira-notes}. It has the virtue of making the co-continuity of $f_!$ obvious.}
\begin{eqnarray*}
f_!: \operatorname{Sh}(S) & \to & \operatorname{Sh}(T) \\
\mathcal F & \mapsto & \varinjlim_{U \subset \subset T} f_* \mathcal F_U
\end{eqnarray*}
Here the notation $U \subset \subset T$ means that the closure of $U$ is compact. When $S$ is an open subset,
this recovers the original definition. When $f$ is proper, then $f_! = f_*$. When $f$ is the map to a point, then
$f_! f^* \mathbb Z$ is the compactly supported cohomology.
As $f_!$ is built from colimits, left adjoints, and pushforwards from compact sets, it is co-continuous.
As such it has a right adjoint, denoted $f^!$. When $f$ is the inclusion
of an open subset, we already had the right adjoint $f^*$, so in this case $f^* = f^!$.
For any locally closed subset $v: V \subseteq T$, we extend the notation $\mathcal F_V := v_! v^* \mathcal F$. This sheaf
has the same stalks as $\mathcal F$ at points in $V$, and has vanishing stalks outside.
For an open-closed decomposition $U \xhookrightarrow{j} T \xhookleftarrow{i} V$ ($j$ open, $i$ closed), the functors $j_{*}, j_!$ and $i_*, i_!$ are fully
faithful, and there are exact triangles
\begin{equation}\label{triangles}
i_! i^! \to \operatorname{id} \to j_* j^* \xrightarrow{[1]} \qquad \qquad \qquad j_! j^! \to \operatorname{id} \to i_* i^* \xrightarrow{[1]}
\end{equation}
Denoting by $\operatorname{Op}(M)$ the poset of open sets, there are functors
\begin{align}\label{generatingsheaves}
\operatorname{Op}(M)&\xrightarrow{!}\operatorname{Sh}(M)&\operatorname{Op}(M)^\mathrm{op}&\xrightarrow{*}\operatorname{Sh}(M)\cr
U&\mapsto u_!\mathbb Z&U&\mapsto u_\ast\mathbb Z
\end{align}
We have the following criterion for when (pullbacks of) these functors are fully faithful:
\begin{lemma} \label{lem:posetsheaf}
Let $\Pi$ be a poset with a map to $\operatorname{Op}(M)$, and let $\mathbb Z[\Pi]$ denote its dg linearization.
The following are equivalent:
\begin{itemize}
\item $H^*(U) \cong \mathbb Z$ for all $U \in \Pi$ and $H^*(U) \xrightarrow{\sim} H^*(U \setminus V)$ whenever $U \nsubseteq V$.
\item The composition $\mathbb Z[\Pi]\to\operatorname{Op}(M)\xrightarrow{!}\operatorname{Sh}(M)$ is fully faithful.
\item The composition $\mathbb Z[\Pi]^\mathrm{op}\to\operatorname{Op}(M)^\mathrm{op}\xrightarrow{*}\operatorname{Sh}(M)$ is fully faithful.
\end{itemize}
\end{lemma}
\begin{proof}
Let us show that the first condition is equivalent to the second.
We have
\begin{align}
\operatorname{Hom}_M(\mathbb Z_U, \mathbb Z_V) &= \operatorname{Hom}_M(u_! \mathbb Z, v_! \mathbb Z) = \operatorname{Hom}_U (\mathbb Z, u^! v_! \mathbb Z) = \operatorname{Hom}_U(\mathbb Z, u^* v_! \mathbb Z)\cr
&=H^*(U, \mathbb Z_{V \cap U})=\operatorname{cone}(H^*(U) \to H^*(U \setminus V)),
\end{align}
where we have used the second triangle in \eqref{triangles}.
The second condition asks that this be $\mathbb Z$ when $U \subseteq V$ and zero otherwise, which is exactly what is asserted in the first condition.
A similar calculation, or taking Verdier duals and noting $\operatorname{Hom}(\mathcal F,\mathcal G) = \operatorname{Hom}(D\mathcal G, D\mathcal F)$, shows the first is equivalent to the third.
\end{proof}
\begin{lemma} \label{lem:sheafrefine}
Let $\Pi$ be a poset with a map to $\operatorname{Op}(M)$ satisfying the equivalent conditions in Lemma \ref{lem:posetsheaf},
and suppose that $W\subseteq M$ is an open set such that
$H^*(U) \xrightarrow{\sim} H^*(U \setminus W)$ is an isomorphism whenever $U\nsubseteq W$.
Then the pullback of the module $\operatorname{Hom}(-, \mathbb Z_W)$ along $\mathbb Z[\Pi] \xrightarrow{!} \operatorname{Sh}(M)$ is the indicator functor
\begin{equation}
1_{W}: U \mapsto \begin{cases} \mathbb Z & U \subseteq W, \cr 0 & \mathrm{otherwise.} \end{cases}
\end{equation}
\end{lemma}
\begin{proof}
This is true by the same calculation as above.
\end{proof}
\subsection{Constructible sheaves}
Let $T$ be a topological space and $\mathcal S: T = \coprod T_\alpha$ a stratification. Write $T_\mathcal S$ for the topological
space with underlying set $T$ and base given by the stars of strata in $\mathcal S$ (note that the intersection of any two stars is expressible a union of stars).
Note the continuous map $T \to T_\mathcal S$.
\begin{remark}\label{weakremark}
Let $\pi: T \to T'$ be any map weakening a topology. For any open set $U$ of $T'$, and any sheaf $\mathcal F$ on $T$, one
has by definition $\pi_*\mathcal F(U) = \mathcal F(U)$. It follows that $\pi^*\mathbb Z_U = \mathbb Z_U$, as this sheaf co-represents
the functor of sections over $U$.
\end{remark}
\begin{lemma} \label{posetmeaning}
Pulling back sheaves under $\mathcal S\xrightarrow\star\operatorname{Op}(T_\mathcal S)$ defines an equivalence
\begin{align}
\operatorname{Sh}(T_\mathcal S)&\xrightarrow{\sim}\operatorname{Fun}(\mathcal S^\mathrm{op},\operatorname{Mod}\mathbb Z)=\operatorname{Mod}\mathcal S\\
\mathcal F&\mapsto(s\mapsto\mathcal F(\star(s)))=\operatorname{Hom}_{T_\mathcal S}(\mathbb Z_{\star(-)},\mathcal F)
\end{align}
which sends $\mathbb Z_{\star(s)}$ to $\operatorname{Hom}_{\mathcal S}(\cdot,s)=1_{\star(s)}$.
\end{lemma}
\begin{proof}
The functor in question is simply restricting a sheaf on $T_\mathcal S$ to the base consisting of stars of strata.
This functor is fully faithful because a map of sheaves is determined by its restriction to a base for the topology.
It is essentially surjective because
there are no nontrivial covers of stars of strata by stars of strata. The behavior on objects is as asserted because $\mathbb Z_{\star(s)}$ and $s$
are the co-representatives of the functors of sections over $s$ and the value of the module at $s$, respectively.
\end{proof}
\begin{lemma}\label{posetmeaningpushforward}
If $\mathcal S'$ refines $\mathcal S$, then the following diagram commutes:
\begin{equation}
\begin{tikzcd}
\operatorname{Sh}(T_{\mathcal S'})\ar{r}&\operatorname{Mod}\mathcal S'\\
\operatorname{Sh}(T_\mathcal S)\ar{r}\ar{u}{\pi^*}&\operatorname{Mod}\mathcal S\ar{u}[swap]{r^*}
\end{tikzcd}
\end{equation}
where $\pi^\ast$ denotes pullback of sheaves under the continuous map $\pi:T_{\mathcal S'}\to T_\mathcal S$ and $r^*:\operatorname{Mod}\mathcal S\to\operatorname{Mod}\mathcal S'$ denotes the pullback along the natural map $r:\mathcal S'\to\mathcal S$.
\end{lemma}
\begin{proof}
By Remark \ref{weakremark} and the characterization of the horizontal functors as $\mathbb Z_U\mapsto 1_U$.
\end{proof}
A sheaf is said to be constant when it is isomorphic to the star pullback of a sheaf on a point,
and locally constant when this is true after restriction to an open cover.
For a stratification $\mathcal S$ of $M$, we say a sheaf is $\mathcal S$-constructible\footnote{Some sources, such as \cite{kashiwara-schapira}, also
ask that the word constructible should mean that sheaves should have perfect stalks and bounded cohomological degree. We do not.}
if it is locally constant when star restricted to each stratum of $\mathcal S$.
Note that the image of the pullback map $\operatorname{Sh}(T_{\mathcal S}) \to \operatorname{Sh}(T)$ is contained in $\operatorname{Sh}_\mathcal S(T)$ (i.e.\ consists of $\mathcal S$-constructible sheaves).
\begin{lemma} \label{sheafposetequivalence}
For a triangulation $\mathcal S$, the map $\operatorname{Sh}(T_{\mathcal S}) \to \operatorname{Sh}_{\mathcal S}(T)$ is an equivalence.
\end{lemma}
\begin{proof}
To show full faithfulness, in view of the equivalence of Lemma \ref{posetmeaning} it is enough to check that $\operatorname{Hom}_{T_\mathcal S}(\mathbb Z_{\star(s)},\mathbb Z_{\star(t)})=\operatorname{Hom}_T(\mathbb Z_{\star(s)},\mathbb Z_{\star(t)})$.
The former is the indicator of $\star(s)\subseteq\star(t)$ again by Lemma \ref{posetmeaning}.
To show that $\operatorname{Hom}_T(\mathbb Z_{\star(s)},\mathbb Z_{\star(t)})$ is as well, by Lemma \ref{lem:posetsheaf} it is enough to show that $H^*(\star(s))\to H^*(\star(s)\setminus\star(t))$ is an isomorphism for $\star(s)\nsubseteq\star(t)$.
If $\star(s)\nsubseteq\star(t)$, then $\star(s)\setminus\star(t)$ is the join of something with $s$, and is hence contractible.
Regarding essential surjectivity, note that the rightmost exact triangle of \eqref{triangles} serves to decompose
any sheaf into an iterated extension of (extensions by zero of) sheaves on the strata; hence any constructible sheaf into
(extensions by zero of) locally constant sheaves on the strata. Since the strata are all contractible, these sheaves are in fact constant.
This shows that the $\mathbb Z_s$ generate. To conclude that the $\mathbb Z_{\star(s)}$ generate,
use the exact triangle $\mathbb Z_{\star(s) \setminus s} \to \mathbb Z_{\star(s)} \to \mathbb Z_s \xrightarrow{[1]}$ and induction on dimension of strata (noting that the first term is in the span of $\mathbb Z_t$ for $\dim(t) < \dim(s)$).
\end{proof}
\subsection{Microsupport}\label{subsec:microsupport}
The notion of microsupport is developed in \cite{kashiwara-schapira}.\footnote{In \cite{kashiwara-schapira}, the authors
work in the bounded derived category. As noted in \cite{robalo-schapira}, the only real dependence on this
was in the proof of one lemma, which is extended to the unbounded setting in that reference.} We recall some basic facts here.
For what follows, let $M$ denote an analytic manifold.
Given a sheaf $\mathcal F$ and
a smooth function $\phi: M \to \mathbb R$, consider a point $m$ in a level set $\phi^{-1}(t)$. We say that $m \in M$ is a cohomological $\mathcal F$-critical
point of $\phi$ if, for inclusion of the superlevelset $i: \phi^{-1}(\mathbb R_{\ge t}) \hookrightarrow M$, one has $(i^! \mathcal F)_m \ne 0$.
The \emph{microsupport} $\ss(F) \subseteq T^*M$ is by definition the closure of the locus of differentials of functions at their cohomological $\mathcal F$-critical points \cite{kashiwara-schapira}.
It is conical.
If $\mathcal F$ is locally constant, then a cohomological $\mathcal F$-critical point can only occur where the function in question has zero derivative. Thus
the microsupport of a locally constant sheaf is contained in the zero section (and is equal to it where the sheaf is not locally zero).
If $U \subseteq M$ is an open set and $m$ is a point in the smooth locus of $\partial U$, then over $m$, the locus $\ss(\mathbb Z_U)=\ss(u_!\mathbb Z)$ is the half-line of outward
conormals to $\partial U$. The locus $\ss(u_\ast\mathbb Z)$ is the inward conormal.
For a subset $X \subseteq T^*M$, we write $\operatorname{Sh}_X(M)$ for the full subcategory of $\operatorname{Sh}(M)$ spanned by objects with microsupport contained in $X$.
Similarly, for $X \subseteq S^*M$, we write $\operatorname{Sh}_X(M)$ for the full subcategory of $\operatorname{Sh}(M)$ with microsupport at infinity contained in $X$.
Evidently if $0_M \subset X$, then $\operatorname{Sh}_X(M) = \operatorname{Sh}_{\partial_{\infty}X}(M)$.
If $\mathcal F$ is a sheaf and $\phi$ is a smooth function with $d\phi_x = \xi$, then for $i: \phi^{-1}(\mathbb R_{\ge t}) \to M$,
if $(i^! F)_x \ne 0$, we have $(x, \xi) \in \ss(\mathcal F)$.
Given this, one wants to assign the complex $(i^! F)_x$ itself as an invariant of $\mathcal F$ at $(x, \xi)$. This is not generally
possible, but it can be done when $\xi$ is a point in the smooth Lagrangian locus of $\ss(\mathcal F)$ \cite[Proposition 7.5.3]{kashiwara-schapira}.
Namely, at any smooth Lagrangian point $(x,\xi)\in X\subseteq S^\ast M$, there is a `microstalk' functor
\begin{equation}
\mu_{(x, \xi)}:\operatorname{Sh}_X(M) \to \operatorname{Sh}(\mathrm{pt}).
\end{equation}
It is given by a shift of $\mathcal F \mapsto (i^! \mathcal F)_x$ for any $\phi$ with $d_x \phi = \xi$ with the graph of $d \phi$ transverse to $X$.
The shift can be fixed using the index of the three transverse Lagrangians $(\ss(\mathcal F), T^*_x M, \Gamma_{d \phi})$.
When $\xi = 0$, the microstalk functor is simply the stalk functor.
\begin{theorem}[{\cite[5.4.19, 5.4.20, 7.5.3]{kashiwara-schapira} or \cite{goresky-macpherson, schmid-vilonen}}]\label{thm:sheafwex}
Let $X \subseteq T^*M$ be a closed conical subset, let
$\phi: M \to \mathbb R$ be a proper function, and assume that over $\phi^{-1}([a, b))$, one has $\Gamma_{d \phi} \cap \ss(\mathcal F) = (x, \xi)$,
where $(x, \xi)$ is a smooth Lagrangian point of $X$.
Let $A: \phi^{-1}((-\infty, a)) \to M$, $A': \phi^{-1}((a, \infty)) \to M$, $B: \phi^{-1}((-\infty, b)) \to M$, and $B': \phi^{-1}((b, \infty)) \to M$ be the inclusions.
Then (up to a shift), the following functors $\operatorname{Sh}_X(M) \to \operatorname{Sh}(\mathrm{pt})$ are isomorphic:
\begin{itemize}
\item The microstalk functor $\mu_{(x,\xi)}$.
\item $\operatorname{Hom}(\operatorname{cone}(A_! \mathbb Z \to B_! \mathbb Z), -)$.
\item $\operatorname{Hom}(\operatorname{cone}(A'_* \mathbb Z \to B'_* \mathbb Z), -)$.
\end{itemize}
Here the maps are the canonical ones coming from restriction of sections.
\end{theorem}
We do not say that $\operatorname{cone}(A_! \mathbb Z \to B_! \mathbb Z)$ co-represents the microstalk because it is not an element of $\operatorname{Sh}_X(M)$.
As observed in \cite{nadler-wrapped}, such co-representatives do exist, for categorical reasons.
Indeed, the microsupport of a sum or product is contained in the union of the microsupports. It follows that the subcategory
$\operatorname{Sh}_X(M) \subseteq \operatorname{Sh}(M)$ is closed under sums and products. In particular, $\operatorname{Sh}_X(M)$ is complete and co-complete,
and the inclusion $\operatorname{Sh}_X(M) \to \operatorname{Sh}(M)$ is continuous and co-continuous. More generally, if $X \subseteq X'$, then
the inclusion $\iota: \operatorname{Sh}_X(M) \to \operatorname{Sh}_{X'}(M)$ is continuous and co-continuous.
Thus it has both adjoints: $(\iota^*, \iota, \iota^!)$.
For example, if $V \subseteq M$
is a closed subset, then taking $X = T^*M|_V$ and $X' = T^* M$ recovers the adjoint triple for the pushforward along $V \to M$,
because $\operatorname{Sh}_{T^*M|_V}(M)=\operatorname{Sh}(V)$.
Using the left adjoint and Theorem \ref{thm:sheafwex}, we can obtain a co-representative for the microstalk as follows. Take any $X' \supseteq \ss(\operatorname{cone}(A_! \mathbb Z \to B_! \mathbb Z))$, e.g.\ $X' = T^*M$. Then $\iota^*\operatorname{cone}(A_! \mathbb Z \to B_! \mathbb Z) \in \operatorname{Sh}_X(M)$ co-represents the microstalk.
We do not generally have a good understanding of $(\iota^*, \iota, \iota^!)$, but when $X' \setminus X$ is isotropic we have the following:
\begin{theorem} \label{thm:sheafstopremoval}
Let $X \subseteq T^*M$ be closed and conical, and let $\Lambda\subseteq T^*M \setminus X$ be a closed conical subanalytic isotropic.
Then $\operatorname{Sh}_X(M) \subseteq \operatorname{Sh}_{X \cup \Lambda}(M)$ is the kernel of all microstalks at Lagrangian points of $\Lambda$.
Thus the left adjoint $\iota^*$ to this inclusion $\iota$ realizes the quotient
\begin{equation}\label{sheafstopremovalqe}
\operatorname{Sh}_{X \cup \Lambda}(M)/\mathcal D\xrightarrow\sim \operatorname{Sh}_X(M),
\end{equation}
where $\mathcal D$ denotes co-representing objects for the microstalks at Lagrangian points of $\Lambda$.
\end{theorem}
\begin{proof}
If $\ss(\mathcal F)\subseteq X$, then the microstalks of $\mathcal F$ at Lagrangian points of $\Lambda$ vanish by definition of microsupport.
To prove the converse, suppose that $\ss(\mathcal F)\subseteq X\cup\Lambda$ and that the microstalks of $\mathcal F$ vanish at all Lagrangian points of $\Lambda$, and let us show that $\ss(\mathcal F)\subseteq X$.
By the fundamental result \cite[6.5.4]{kashiwara-schapira} that the microsupport is co-isotropic, it is enough to show that $p\notin\ss(\mathcal F)$ for every Lagrangian point $p\in\Lambda$.
It is not quite immediate from the definitions that vanishing of the microstalk implies there is no microsupport,
since the microsupport is defined in terms of arbitrary test functions, whereas microstalks are defined in terms of microlocally transverse test functions.
To see it is true, and that moreover the microstalk is locally
constant along $\Lambda$, one can apply a contact transformation so that $\Lambda$ becomes locally the conormal
to a smooth hypersurface; for details see \cite[Chap. 7]{kashiwara-schapira}.
\end{proof}
\subsection{Compact objects}
Here we elaborate upon some assertions of \cite{nadler-wrapped}.
We write $\operatorname{Sh}_X(M)^c$ for the compact objects in the category $\operatorname{Sh}_X(M)$. Be warned:
\begin{proposition}[\cite{neeman-sheaves}]
When $M$ is non-compact, $\operatorname{Sh}(M)^c = 0$.\qed
\end{proposition}
There are not many more compact objects in the compact case.
However, for constructible sheaves with respect to a \emph{fixed} triangulation, the situation is different:
\begin{lemma} \label{constructibleimpliescompact}
For $\mathcal S$ a triangulation, the category $\operatorname{Sh}_\mathcal S(M)$ is compactly generated, and
the objects of $\operatorname{Sh}_\mathcal S(M)^c$ are the sheaves with perfect stalks and compact support.
\end{lemma}
\begin{proof}
Under the identification (Lemma \ref{sheafposetequivalence}) $\operatorname{Sh}_\mathcal S(M) = \operatorname{Mod} \mathcal S$, the $\mathbb Z_{\star(s)}$ go to compact generators.
The devissage in the proof of the same Lemma shows that $\mathbb Z_s$ also generate, and can be expressed using finitely many
$\mathbb Z_{\star(s)}$, hence are compact. The $\mathbb Z_s$ evidently generate the sheaves with perfect stalks and compact support.
\end{proof}
\begin{remark}\label{noncompacttriangulation}
Note that while a non-compact manifold does not admit a finite triangulation, it can sometimes be a relatively compact constructible subset of a larger manifold.
\end{remark}
Let us carry what we can of this to the categories $\operatorname{Sh}_\Lambda(M)$ when $\Lambda$ is subanalytic.
\begin{proposition} \label{compactmicrostalks}
Let $\mathcal S$ be a subanalytic stratification.
The microstalk at a smooth point of $N^* \mathcal S$ is
co-representable by a compact object of $\operatorname{Sh}_{\mathcal S}(M)$.
The same holds for the stalk at any point in the zero section.
\end{proposition}
\begin{proof}
Consider the microstalk at some point $(x, \xi)$. It is possible to choose
real analytic $\phi$ as in Theorem \ref{thm:sheafwex}, see \cite[Thm.\ 2.2.1]{goresky-macpherson} or
\cite[Proposition 8.3.12]{kashiwara-schapira}. We keep the notation of Theorem \ref{thm:sheafwex}.
Refine the stratification to some $\mathcal S'$ so the $A_!\mathbb Z$ and $B_!\mathbb Z$ are constructible.
By Lemma \ref{constructibleimpliescompact}, $\operatorname{cone}(A_! \mathbb Z \to B_! \mathbb Z)$ is a compact object in $\operatorname{Sh}_{\mathcal S'}(M)$.
The functor $\iota$ is co-continuous,
so its left adjoint $\iota^*$ preserves compact objects.
Thus the co-representing object $\iota^*\operatorname{cone}(A_! \mathbb Z \to B_! \mathbb Z) \in \operatorname{Sh}_\mathcal S(M)$ is compact.
Regarding stalks, note that for any $x \in M$, the functor of taking stalks at $x$, which is by definition $\mathcal F_x := \varinjlim \mathcal F(B_\epsilon(x))$,
is in fact computed by some fixed $\mathcal F_x = \mathcal F(B_{\epsilon_x}(x))$. Indeed, further shrinking of the ball will be non-characteristic with respect
to $N^*\mathcal S$, as follows from Whitney's condition B.
Now we argue as above, choosing any analytic function with sublevelset $B_{\epsilon_x}(x)$.
\end{proof}
\begin{corollary} \label{compactsheafstopremoval}
Let $X\subseteq T^*M$ and $\Lambda\subseteq T^*M\setminus X$ be closed conical subanalytic isotropics. Then
\begin{equation}
(\operatorname{Sh}_{X \cup \Lambda}(M)^c/\mathcal D)^\pi\xrightarrow\sim \operatorname{Sh}_X(M)^c,
\end{equation}
where $\mathcal D$ denotes co-representing objects for the microstalks at Lagrangian points of $\Lambda$.
\end{corollary}
\begin{proof}
The microstalks are compact by Proposition \ref{compactmicrostalks} and the fact (observed in the proof of that result) that the left adjoint to the inclusion $\operatorname{Sh}_{X \cup \Lambda}(M)\hookrightarrow\operatorname{Sh}_\mathcal S(M)$ preserves compact objects (for $\mathcal S$ a stratification with $X\cup\Lambda\subseteq N^*\mathcal S$).
Now apply Lemma \ref{compactinquotient} to Theorem \ref{thm:sheafstopremoval}.
\end{proof}
\begin{corollary}\label{microstalksgenerate}
For any subanalytic isotropic $\Lambda$, the category $\operatorname{Sh}_\Lambda(M)$ is compactly generated by co-representatives of the microstalks at smooth points of $\Lambda$ (including the stalks at points of the zero section $M$ whose cotangent sphere is disjoint from $\Lambda$).
\end{corollary}
\begin{proof}
By Theorem \ref{thm:sheafstopremoval}, the quotient by this set of objects is the zero
category.
\end{proof}
The following result was shown in \cite{nadler-wrapped} using arborealization; here is a direct argument.
\begin{corollary} \label{propersheaves}
The Yoneda embedding induces an equivalence between the full subcategory of $\operatorname{Sh}_\Lambda(M)$ of objects with perfect stalks
and the category $\operatorname{Prop} \operatorname{Sh}_\Lambda(M)^c$.
\end{corollary}
\begin{proof}
From the argument in Proposition \ref{compactmicrostalks}, we see that the microstalks are calculated by comparing sections over
precompact sets; it follows that a sheaf microsupported in $\Lambda$ (thus constructible) with perfect stalks has perfect microstalks.
Together the stalk and microstalk functors split-generate $\operatorname{Sh}_\Lambda(M)^c$ by Corollary \ref{microstalksgenerate}, so we see that a sheaf with perfect stalks defines a proper module.
To see the converse, recall from the argument in Proposition \ref{compactmicrostalks} that the stalk functors can be expressed
in terms of sections over open sets constructible with respect to some $\mathcal S$ satisfying $N^*\mathcal S\supseteq\Lambda$.
The left adjoint to $\operatorname{Sh}_\Lambda(M)\hookrightarrow\operatorname{Sh}_\mathcal S(M)$ preserves compact objects as observed previously, hence proper over $\operatorname{Sh}_\Lambda(M)^c$ implies perfect stalks.
\end{proof}
\begin{remark}
Note that if we restricted from the beginning to sheaves constructible with
respect to some fixed finite subanalytic stratification, and for intermediate constructions allowed ourselves only fixed
finite refinements, then we could in principle argue without appeal to the ``well-generated'' version of Brown
representability.
The results of this article, restricted to compact manifolds, can all be deduced in this setting.
\end{remark}
For compact $M$, we establish smoothness and/or properness for some of these categories.
\begin{proposition} \label{sheafcompactsmoothproper}
If $M$ is compact and $\mathcal S$ is a triangulation, then $\operatorname{Sh}_\mathcal S(M)^c$ is smooth and proper.
\end{proposition}
\begin{proof}
The $\mathbb Z_{\star(s)}$ give a finite generating exceptional collection which is proper, and this implies smoothness by Lemma \ref{exceptionalpropersmooth}.
\end{proof}
More generally,
\begin{corollary}
If $M$ is compact and $\Lambda$ is subanalytic isotropic, then the category $\operatorname{Sh}_\Lambda(M)^c$ is smooth,
$\operatorname{Prop} \operatorname{Sh}_\Lambda(M)^c \subseteq \operatorname{Perf} \operatorname{Sh}_\Lambda(M)^c$, and the category $\operatorname{Prop} \operatorname{Sh}_\Lambda(M)^c $ is proper.
\end{corollary}
\begin{proof}
By Proposition \ref{sheafcompactsmoothproper} and Corollary \ref{compactsheafstopremoval}, the category
$\operatorname{Sh}_\Lambda(M)^c$ is a quotient of a smooth category, hence smooth (Lemma \ref{smoothquotient}).
Smoothness implies proper modules are perfect (Lemma \ref{propperf}) and that the category of proper modules is proper.
\end{proof}
\begin{remark}\label{noncompactsmoothness}
When $(M,\Lambda)$ are non-compact but finite-type in a suitable sense, the
same result is true. One can prove it by embedding into a compact manifold as
in Remark \ref{noncompacttriangulation}.
\end{remark}
\subsection{In conclusion}
Collecting the results of this section, we have shown:
\begin{theorem}\label{sheafrealization}
The functor $\Lambda \mapsto \operatorname{Sh}_\Lambda(M)^c$ is a microlocal Morse theatre in the sense of Definition \ref{microlocalmorsetheatre}, which casts the co-representatives of the microstalk functors at smooth points of $\Lambda$ as the Morse characters.
\end{theorem}
\begin{proof}
The most obvious functor $\Lambda \to Sh_\Lambda(M)$ is the one which carries \emph{inclusions} $\Lambda \subseteq \Lambda'$ to
\emph{inclusions} $\operatorname{Sh}_\Lambda(M)\hookrightarrow\operatorname{Sh}_{\Lambda'}(M)$; note that this is in fact a strict diagram of categories (as all are simply full subcategories of $\operatorname{Sh}(M)$) and takes values
in the category whose objects are large dg categories and whose morphisms are continuous and co-continuous.
Passing to left adjoints and taking compact objects (a left adjoint of a co-continuous functor preserves compact objects), we obtain a functor $\Lambda\mapsto\operatorname{Sh}_\Lambda(M)^c$.
For triangulations $\mathcal S$, the functors
\begin{equation}
\mathcal S\xrightarrow{s\mapsto\mathbb Z_{\star(s)}}\operatorname{Sh}_\mathcal S(M)\xrightarrow{\mathcal F\mapsto\operatorname{Hom}(\mathbb Z_{\star(-)},\mathcal F)}\operatorname{Mod}\mathcal S
\end{equation}
define an equivalence $\operatorname{Perf}\mathcal S=\operatorname{Sh}_\mathcal S(M)^c$ by Lemmas \ref{posetmeaning} and \ref{sheafposetequivalence}.
When $\mathcal S$ is a $\mu$-stratification, we have $\operatorname{Sh}_\mathcal S(M)^c=\operatorname{Sh}_{N^*_\infty\mathcal S}(M)^c$ by \cite[Prop.\ 8.4.1]{kashiwara-schapira}.
Taking the commutative diagram in Lemma \ref{posetmeaningpushforward} and passing to the left adjoints of the vertical maps
shows that this equivalence respects refinement of triangulations.
This shows that $\Lambda\mapsto\operatorname{Sh}_\Lambda(M)^c$ is a microlocal Morse pre-theatre.
By Theorem \ref{thm:sheafwex}, the Morse characters in $\operatorname{Perf}\mathcal S$ correspond, under this isomorphism, to co-representatives of the microstalks.
According to Corollary \ref{compactsheafstopremoval}, the functor $\operatorname{Sh}_{\Lambda'}(M)^c\to\operatorname{Sh}_\Lambda(M)^c$ is the quotient by co-representatives of the microstalks.
\end{proof}
\section{Wrapped Fukaya categories} \label{sec:fukaya}
\subsection{Notation}
Here we quickly fix notation and review basic facts (see, e.g., \cite[\S 3.3]{gpssectorsoc} for more details).
Fix a Liouville manifold or sector $X$.
For exact Lagrangians $L, K \subseteq X$, conical at infinity,
we write $HF^\ast(L, K)$ for the Floer cohomology. We write $HF^\ast(L, L)$ to mean $HF^\ast(L^+, L)$, where $L^+$ denotes an (unspecified) small positive pushoff of $L$.
This group $HF^\ast(L, L)=HF^\ast(L^+, L)$ is a unital algebra, and its unit is termed the \emph{continuation element}.
Composition of continuation elements defines
a continuation element in $HF^\ast(L^{++}, L)$ for $L^{++}$ any (not necessarily small) positive wrapping of $L$.
If the positive isotopy $L\leadsto L^{++}$ takes place in the complement of $\partial_\infty K$,
then composition with the continuation map gives an isomorphism $HF^\ast(L, K) \xrightarrow{\sim} HF^\ast(L^{++}, K)$ (and similarly in the reverse). More generally, if $L \leadsto L'$ is any isotopy taking place in the complement of $\partial_\infty K$, then there is an induced identification $HF^\ast(L,K) = HF^\ast(L',K)$ (see \cite[Lem.\ 3.21]{gpssectorsoc}).
The wrapped Floer cohomology $HW^\ast(L, K)_X$ is equivalently calculated by
$$\varinjlim_{L \leadsto L^{++}} HF^\ast(L^{++}, K) = \varinjlim_{\substack{L \leadsto L^{++} \\ K^{--}\leadsto K}} HF^\ast(L^{++}, K^{--}) = \varinjlim_{K^{--}\leadsto K} HF^\ast(L, K^{--}).$$
Here, the direct limits are taken using the continuation maps over positive-at-infinity isotopies of $L$ and negative-at-infinity isotopies of $K$.
The freedom to wrap in only one factor is extremely useful in practice.
Given any closed subset $\mathfrak f \subset \partial_\infty X$, and $L, K$ disjoint at infinity from $\mathfrak f$,
we similarly define $HW^\ast(L, K)_{(X,\mathfrak f)}$ by restricting wrappings to take place in the complement of $\mathfrak f$.
One main point of \cite{gpssectorsoc} was the construction of a covariant functor $\mathcal W(X) \to \mathcal W(Y)$ for an inclusion of Liouville sectors $X \subseteq Y$.
In \cite{gpsstructural} we remarked that the same construction gives a functor $\mathcal W(X, \mathfrak f\cap(\partial_\infty X)^\circ) \to \mathcal W(Y, \mathfrak f)$.
This covariance is a nontrivial result having to do with the fact that holomorphic disks do not cross the boundary of a Liouville sector. By
contrast, it is immediate from the definition that if $\mathfrak g \subseteq \mathfrak f$ then there is a natural map $\mathcal W(X, \mathfrak f) \to \mathcal W(X, \mathfrak g)$: just wrap more. Both
covariance statements allow one to calculate in a potentially simpler geometry, and push forward the result.
The following Lemma allows one to explicitly describe cofinal wrapping sequences. Its typical use is the following. To compute $HW^\ast(L, K)$,
one finds a cofinal sequence $L_t$ as in the Lemma, such that the induced maps $HF^\ast(L_t, K) \to HF^\ast(L_{t+1}, K) $ are eventually
all isomorphisms. Then $HW^\ast(L, K) = HF^\ast(L_t, K)$ for any $L_t$ in this stable range.
\begin{lemma}[{\cite[Lemma 2.1]{gpsstructural}}]\label{cofinalitycriterion}
Let $Y$ be a contact manifold and $\Lambda_t$ a positive isotopy of Legendrians.
If $\Lambda_t$ escapes to infinity as $t\to\infty$ (i.e.\ is eventually disjoint from any given compact subset of $Y$), then it is a cofinal wrapping of $\Lambda_0$.
\qed
\end{lemma}
\subsection{Foundations}
In \cite{gpssectorsoc,gpsstructural}, for any Liouville sector $X$ and any closed subset $\mathfrak f\subseteq(\partial_\infty X)^\circ$, we constructed $A_\infty$ categories $\mathcal W(X, \mathfrak f)$ whose objects are exact Lagrangians in $X \setminus \mathfrak f$,
conical at infinity. The cohomology category is simply wrapped Floer cohomology $H^*\mathcal W(L, K)=HW^\ast(L, K)_{(X,\mathfrak f)}$.
We wish to consider here categories $\mathcal W(T^*M,\Lambda)$ for manifolds $M$ and closed subsets $\Lambda\subseteq S^*M=\partial_\infty T^*M$.
When $M$ is non-compact, this does not strictly fit into the framework of \cite{gpssectorsoc,gpsstructural}, so we describe here the construction (which is only a minor variation on \cite[Sec.\ 3]{gpssectorsoc} and \cite[Sec.\ 2]{gpsstructural}, to which we refer the reader for more details).
We do not discuss the case when $M$ has boundary, since it is not difficult to check that the category one would define for such $M$ is equivalent to the category we define here for its interior $M^\circ$.
To define $\mathcal W(T^*M,\Lambda)$, we choose the following data:
\begin{enumerate}
\item A countable poset $\mathcal O$ of exact conical at infinity Lagrangians inside $T^*M$ (equipped with grading/orientation data) disjoint from $\Lambda$ and with relatively compact image in $M$ (and $\mathcal O$ must contain at least one in every isotopy class of such Lagrangians). We require $\mathcal O$ to be cofinite, namely $\{K\in\mathcal O\,|\,K\leq L\}$ is finite for all $L\in\mathcal O$. We require that every totally ordered subset $L_0>\cdots>L_k\in\mathcal O$ must be mutually transverse. Finally, we require that for every $L\in\mathcal O$, there exist a cofinal sequence $L=L_0<L_1<\cdots\in\mathcal O$ along with positive isotopies $L_0=L_0\leadsto L_1\leadsto\cdots$ which are cofinal inside the positive wrapping category of $L$ inside $T^*M$ away from $\Lambda$.
\item A collection $C$ of elements of $HF^*(L,K)$ for pairs $L>K\in\mathcal O$ consisting only of continuation elements for various positive isotopies $K\leadsto L$ disjoint from $\Lambda$. This collection $C$ must be such that for every $L\in\mathcal O$, there exists a sequence $L=L_0<L_1<\cdots\in\mathcal O$ cofinal in $\mathcal O$ and positive isotopies $L=L_0\leadsto L_1\leadsto\cdots$ cofinal in the wrapping category of $L$ (away from $\Lambda$) such that every associated continuation element in $HF^*(L_{i+1},L_i)$ is in $C$.
\item For every $L\in\mathcal O$, a choice of compact codimension zero submanifold-with-boundary $M_L\subseteq M$ containing the image of $L$, such that $M_L\subseteq(M_K)^\circ$ for $L<K$.
\item Floer data for $\mathcal O$, consisting of choices of strip-like coordinates $\underline\xi$ as in \cite[(3.54)--(3.55)]{gpssectorsoc} and almost complex structures
\begin{equation}
J_{L_0,\ldots,L_k}:\overline{\mathcal S}_{k,1}\to\mathcal J(T^*M_{L_0})
\end{equation}
for $L_0>\cdots>L_k\in\mathcal O$.
These almost complex structures are required to make a fixed choice of projection from near $\partial T^*M_{L_0}$ to $\mathbb C_{\Re\geq 0}$ holomorphic.
\end{enumerate}
\begin{remark}\label{indliouvillesectors}
The essential difference between the current situation and the setup of \cite{gpssectorsoc,gpsstructural} is that $T^*M$, rather than being a Liouville sector itself, is only a filtered ascending union of Liouville sectors $T^*M_\alpha\subseteq T^*M$ (over compact codimension zero submanifolds-with-boundary $M_\alpha\subseteq M$).
The present discussion, while phrased in terms of cotangent bundles, would apply without change to any such `ind-Liouville sector', i.e.\ any filtered ascending union of Liouville sectors.
\end{remark}
By counting holomorphic disks with respect to given data $(\mathcal O,C,\underline M,\underline\xi,\underline J)$, we define a directed (by the poset) strictly unital $A_\infty$ category $\mathcal O$ with $\hom_{\mathcal O}(K,L) = 0$ unless $K\geq L$.
The definition of $\mathcal W(T^*M,\Lambda)$ is as the localization $\mathcal W:=\mathcal O[C^{-1}]$.
This has the correct cohomology category, calculated by wrapping as above (see \cite[Lemma 3.37]{gpssectorsoc}).
Given $(\mathcal O,C)$, it is straightforward to construct $(\underline M,\underline\xi,\underline J)$ by induction to achieve transversality.
The construction of $(\mathcal O,C)$ proceeds by applying the following Lemma to $(\mathcal O_0, C_0 = \emptyset)$ where $\mathcal O_0$ is any countable set containing at least one Lagrangian in every isotopy class, thought of as a poset with no relations:
\begin{lemma}\label{findO}
Let $\mathcal O_0$ be any cofinite countable poset of Lagrangians inside $(T^*M,\Lambda)$ (with every totally ordered subset mutually transverse), and let $C_0$ be a collection of continuation elements.
There exists another such pair $(\mathcal O_1,C_1)$ together with a downward closed embedding $\mathcal O_0\hookrightarrow\mathcal O_1$ with $C_1|_{\mathcal O_0}=C_0$, such that $(\mathcal O_1,C_1)$ in addition satisfies: for every $L\in\mathcal O_1$, there exists a cofinal sequence $L=L_0<L_1<\cdots\in\mathcal O_1$ and positive isotopies $L=L_0\leadsto L_1\leadsto\cdots$ cofinal in the wrapping category of $L$ (away from $\Lambda$) such that the associated continuation elements are in $C_1$.
\end{lemma}
\begin{proof}
Exhaust $\mathcal O_0$ by downward closed finite subsets $Z_0\subseteq Z_1\subseteq\cdots$.
The additional Lagrangians $\mathcal O_1\setminus\mathcal O_0$ are indexed by integers $i\geq 0$ (ordered accordingly), and the $i$th such additional Lagrangian lies above precisely the Lagrangians $Z_i$ inside $\mathcal O_0$.
For each Lagrangian $L\in\mathcal O_0$, we choose (generically) cofinal wrappings $L\leadsto L_1\leadsto L_2\leadsto\cdots$, and we add the $L_1<L_2<\cdots$ to the list of Lagrangians we want to include in $\mathcal O_1\setminus\mathcal O_0$.
There are countably many such sequences $(L_1,L_2,\cdots)$, so we can include all of them into $\mathbb Z_{\geq 0}$ at the same time, ensuring moreover that $L<L_1$.
Namely, we enumerate $\mathcal O_0=\{K_0,K_1,\ldots\}$, and we process these $K_i$ in order as follows: given $L=K_i$, embed $L_1<L_2<\cdots$ into $\mathbb Z_{\geq 0}$ (in the complement of everything else previously embedded there) such that $L_1$ is put high enough to ensure $L<L_1$, and such that there are still infinitely many `slots' in $\mathbb Z_{\geq 0}$ remaining (to be used in the countably many subsequent steps).
At the end of the process, there may be `unfilled' slots in $\mathbb Z_{\geq 0}$, however this is of no importance.
After the completion of the process of embedding all of the countably many
sequences $(L_1,L_2,\cdots)$ into $\mathbb Z_{\geq 0}$, we further inductively
perturb these $L_i$ to $\tilde L_i$ (in a manner preserving positivity and cofinality of the isotopies $L\leadsto\tilde L_1\leadsto \tilde L_2\leadsto$) to ensure mutual transversality
of every totally ordered subset of the thusly defined poset $\mathcal O_1$.
The continuation elements $C_1$ are simply $C_0$ union
those associated to the positive isotopies $L\leadsto \tilde L_1\leadsto
\tilde L_2\leadsto\cdots$.
\end{proof}
Applying Lemma \ref{findO}, we obtain a pair $(\mathcal O,C)$ for any $\Lambda\subseteq S^*M$, thus giving rise to a category $\mathcal W(T^*M,\Lambda)$.
This defines each of the categories $\mathcal W(T^*M,\Lambda)$ individually, however we also want pushforward functors $\mathcal W(T^*M,\Lambda')\to\mathcal W(T^*M,\Lambda)$ for inclusions $\Lambda'\supseteq\Lambda$.
That is, denoting by $\mathbf\Lambda$ the poset of all closed subsets $\Lambda\subseteq S^*M$ (ordered by reverse inclusion), we want a functor
\begin{align}
\mathcal W:\mathbf\Lambda&\to\textsf{A$_\infty$-cat}\\
\Lambda&\mapsto\mathcal W(T^*M,\Lambda)
\end{align}
(in the sense that $\mathcal W(T^*M,\Lambda'')\to\mathcal W(T^*M,\Lambda')\to\mathcal W(T^*M,\Lambda)$ coincides with $\mathcal W(T^*M,\Lambda'')\to\mathcal W(T^*M,\Lambda)$ on the nose).
To construct this functor $\mathcal W$, it suffices to construct it for finite subsets of $\mathbf\Lambda$, in the following sense.
Namely, suppose that for every finite subset $F\subseteq\mathbf\Lambda$, we have a functor
\begin{align}
\mathcal W_F:F&\to\textsf{A$_\infty$-cat}\\
\Lambda&\mapsto\mathcal W_F(T^*M,\Lambda)
\end{align}
along with quasi-equivalences $\mathcal W_F\to\mathcal W_{F'}|_F$ (again, strictly compatible for triples $F\subseteq F'\subseteq F''$) which are all \emph{naive inclusions}, i.e.\ are injective on objects and with all higher ($k\geq 2$) functor operations vanishing.
Then we may define
\begin{equation}
\mathcal W(\Lambda):=\varinjlim_{\Lambda\in F\subseteq\mathbf\Lambda}\mathcal W_F(\Lambda)
\end{equation}
(which makes sense since the transition functors are naive inclusions).
Now for $\Lambda \supseteq \Lambda'$, and for any $F \ni \Lambda$,
there is a (naive) inclusion $\mathcal W_F(\Lambda) \to \mathcal W_{F \cup \{\Lambda'\}}(\Lambda)
\to \mathcal W_{F \cup \{\Lambda'\}}(\Lambda')$ living over the map of directed systems
$\{\Lambda \in F \subseteq \mathbf\Lambda\}\to \{\Lambda' \in F' \subseteq
\mathbf\Lambda\}$ which sends $\Lambda \in F$ to $\Lambda' \in F \cup
\{\Lambda'\}$, compatibly with maps in the system. This defines $\mathcal W(\Lambda)
\to \mathcal W(\Lambda')$ as desired, and one can check that these maps indeed compose as desired.
It thus suffices to construct the compatible systems of categories $\mathcal W_F$ for finite subsets $F\subseteq\mathbf\Lambda$.
To construct these $\mathcal W_F$, it is enough to define the corresponding $(\mathcal O_F,C_F)$ (the subsequent inductive construction of Floer data is straightforward, and hence will not be discussed further).
We would thus like to define, for every finite $F\subseteq\mathbf\Lambda$, a functor $\Lambda\mapsto(\mathcal O_F(\Lambda),C_F(\Lambda))$ such that each map $(\mathcal O_F(\Lambda'),C_F(\Lambda'))\hookrightarrow(\mathcal O_F(\Lambda),C_F(\Lambda))$ is as in the conclusion of Lemma \ref{findO}, namely downward closed and satisfying $C_F(\Lambda)|_{\mathcal O_F(\Lambda')}=C_F(\Lambda')$.
Furthermore, we would like to have the same sort of maps $(\mathcal O_F,C_F)\hookrightarrow(\mathcal O_{F'},C_{F'})|_F$ for $F\subseteq F'$.
The construction of $(\mathcal O_F,C_F)$ is by induction on finite subsets $F\subseteq\mathbf\Lambda$ (ordered by inclusion).
For a fixed $F$, the construction of $(\mathcal O_F(\Lambda),C_F(\Lambda))$ is by induction on $\Lambda\in F$ (ordered by reverse inclusion).
To define $(\mathcal O_F(\Lambda),C_F(\Lambda))$, we begin with $(\mathcal O_0, C_0)$ equal to the colimit of everything which must map into $(\mathcal O_F(\Lambda),C_F(\Lambda))$, namely $(\mathcal O_0, C_0)$ is the colimit of $(\mathcal O_{F'}(\Lambda'),C_{F'}(\Lambda'))$ over pairs $(\Lambda', F')$ with $\Lambda\subseteq\Lambda'\in F'\subseteq F$ (with its natural poset structure) with at least one of the inclusions strict. If necessary, add on to $\mathcal O_0$ additional Lagrangians in order to represent all isotopy classes (with no relations to the rest of $\mathcal O_0$, and no additions to $C_0$).
Applying Lemma \ref{findO} to the result defines $(\mathcal O_F(\Lambda),C_F(\Lambda))$.
\subsection{Gradings and orientations}\label{gradorsec}
We briefly review the setup for defining gradings and orientations in Floer theory; for more details see Seidel \cite{seidelgraded,seidelbook}.
The Grassmannian of Lagrangian subspaces of a given symplectic vector space of dimension $2n$ is homotopy equivalent to $U(n)/O(n)$, and its fundamental group has a preferred isomorphism with $\mathbb Z$ given by $\det^2:U(n)/O(n)\to U(1)$.
A $\mathbb Z$-grading on a symplectic manifold $X$ is a choice of fiberwise universal cover $\operatorname{\widetilde{\mathcal LGr}} X\to\operatorname{\mathcal LGr} X$.
A $\mathbb Z$-grading on a Lagrangian $L\subseteq X$ with respect to a given $\mathbb Z$-grading on $X$ is a lift of the tautological section $L\to(\operatorname{\mathcal LGr} X)|_L$ to $(\operatorname{\widetilde{\mathcal LGr}} X)|_L$.
The Lagrangian Grassmannian of the symplectic manifold $T^*M$ has a canonical section (over the zero section $M$) namely $[x\mapsto T^*_xM]$.
Using this section as the fiberwise basepoint, we obtain a canonical $\mathbb Z$-grading $\operatorname{\widetilde{\mathcal LGr}} T^\ast M\to\operatorname{\mathcal LGr} T^\ast M$.
We consider exclusively this $\mathbb Z$-grading on $T^*M$ for defining the Fukaya category of a cotangent bundle $T^\ast M$.
The Lagrangian fibers $T_x^\ast M\subseteq T^\ast M$ have tautological $\mathbb Z$-gradings (in particular, varying continuously in $x$) relative to this canonical $\mathbb Z$-grading on $T^*M$.
A $\operatorname{Pin}$-structure on a Lagrangian $L\subseteq X$ relative to a $K(\mathbb Z/2,1)$-bundle $b$ over $X$ is an isomorphism of $K(\mathbb Z/2,1)$-bundles over $L$ between $b|_L$ and $\varinjlim\mathbb P(TL\oplus\mathbb R^n)$.
We consider exclusively the $K(\mathbb Z/2,1)$-bundle $b:=\varinjlim\mathbb P(T^\ast M\oplus\mathbb R^n)$ over $M$ (pulled back to $T^\ast M$) for defining the Fukaya category of a cotangent bundle $T^\ast M$.
The Lagrangian fibers $T_x^\ast M\subseteq T^\ast M$ are tautologically $\operatorname{Pin}$ relative to $b$ (varying continuously in $x$).
\subsection{Wrapping exact triangle, stop removal, generation}
The fundamental ingredients underlying our work in this section are the wrapping exact triangle and its consequence stop removal, both proved in \cite{gpsstructural}.
The wrapping exact triangle can be thought of as quantifying the price of wrapping through a stop.
It should be compared with Theorem \ref{thm:sheafwex}.
\begin{theorem}[{Wrapping exact triangle \cite[Theorem 1.7]{gpsstructural}}]\label{wrapcone}
Let $(X,\Lambda)$ be a stopped Liouville sector, and let $p\in\Lambda$ be a point near which $\Lambda$ is a Legendrian submanifold.
If $L\subseteq X$ is an exact Lagrangian submanifold and $L^w\subseteq X$ is obtained from $L$ by passing $\partial_\infty L$ through $\Lambda$ transversally at $p$ in the positive direction, then there is an exact triangle
\begin{equation}
L^w\to L\to D_p\xrightarrow{[1]}
\end{equation}
in $\mathcal W(X,\Lambda)$, where $D_p\subseteq X$ denotes the small Lagrangian disk linking $\Lambda$ at $p$ and the map $L^w\to L$ is the continuation map.
\end{theorem}
The following result about wrapped Fukaya categories is a consequence of the wrapping exact triangle, and can be compared with Theorem \ref{thm:sheafstopremoval}.
\begin{theorem}[{Stop removal \cite[Theorem 1.13]{gpsstructural}}]\label{stopremoval}
Let $(X,\mathfrak f)$ be a stopped Liouville manifold (or sector), and let $\Lambda\subseteq(\partial_\infty X)^\circ\setminus\mathfrak f$ be an isotropic submanifold.
Then pushforward induces a quasi-equivalence
\begin{equation}\label{stopremovalqe}
\mathcal W(X,\mathfrak f\cup\Lambda)/\mathcal D\xrightarrow\sim\mathcal W(X,\mathfrak f),
\end{equation}
where $\mathcal D$ denotes the collection of small Lagrangian disks linking (Legendrian points of) $\Lambda$.
\end{theorem}
We will also need to know that fibers generate:
\begin{theorem}\label{generation}
The cotangent fibers split-generate $\mathcal W(T^*M)$.
\end{theorem}
\begin{proof}
When $M$ is compact (including the case with boundary), this
is \cite[Theorem 1.9 and Example 1.10]{gpsstructural}. For a
general possibly non-compact $M$, we observe that any Lagrangian $L \in \mathcal W(T^*M)$
is in the essential image of the pushforward functor $\mathcal W(T^*M_L) \to
\mathcal W(T^*M)$, for some compact codimension zero submanifold-with-boundary $M_L
\subseteq M$. Now push foward the fact that $L$ is split-generated by a fiber in $\mathcal W(T^*M_L)$.
\end{proof}
\begin{remark}
In fact, \cite[Theorem 1.9]{gpsstructural} and the argument above shows the fibers generate $\mathcal W(T^*M)$, however we only need split-generation.
\end{remark}
Another ingredient which proves useful in our computations is the K\"unneth theorem for Floer cohomology and wrapped Fukaya categories, also proved in \cite{gpsstructural}.
\subsection{Conormals}
For a relatively compact open set $U \subseteq M$ with smooth boundary,
we write $L_U \subseteq T^*M$ for (a smoothing of) the union of $U \subseteq T_M^* M$
with its \emph{outward} conormal.
We write $-$ to mean the antipodal map on $T^*M$, hence $-L_U$ is a smoothing of the union of
$U$ with its \emph{inward} conormal.
For a Lagrangian $L$, we write $L^+$ for an unspecified small positive Reeb pushoff of $L$, and $L^-$ for a negative pushoff.
If $U$ is an relatively compact open set with smooth boundary, we write $U^{\epsilon}$ for an $\epsilon$ neighborhood of $U$ in some metric,
and $U^{-\epsilon}$ for the open set such that $(U^{-\epsilon})^{\epsilon} = U$. When $\epsilon$ is unimportant and unchanging, we write these as
$U^+$ and $U^-$. Our conormal conventions are chosen to ensure that the conormal to $U^+$ is a \emph{positive} pushoff of the conormal to $U$; in other words
$$L_{U^+} = L_U^+ \qquad \qquad L_{U^-} = L_U^-$$
That is, positive Reeb flow pushes outward conormals out.
More generally, if $\overline{U}$ is a manifold with corners, then we write we write $L_U$ to mean $L_{\tilde U}$ where $\tilde U$ is obtained from $U$ by smoothing out the boundary.
In all of the above cases, we could also equivalently say that $L_U$ is a rounding of $\ss(\mathbb Z_U)$ (compare Section \ref{subsec:microsupport}).
Each $L_U$ is exact and possesses a canonical relative Pin
structure and grading: the codimension zero inclusion $\overline{U^-} \subset L_U$ is a
homotopy equivalence, and $\overline{U^-}$ is a codimension 0 submanifold of, and
thereby inherits all of this data from, the zero section.
\subsection{Floer cohomology between conormals of balls and stable balls}
By a \emph{ball}, we mean an open set with smooth boundary whose closure is diffeomorphic to the unit ball.
\begin{lemma} \label{continuationballs}
Let $U,V\subseteq M$ be balls with $\overline U\subseteq V$.
Then $HF^\ast(L_V, L_U) = \mathbb Z$, and is canonically generated by the continuation element.
\end{lemma}
\begin{proof}
In this case there is a positive isotopy from $L_U^+$ to $L_V$ in the complement of $\partial_\infty L_U$.
\end{proof}
\begin{lemma} \label{deformtozero}
Let $V$ be an open set with smooth boundary, and let $U$ be a $\epsilon$-ball centered a point on $\partial V$.
Then $HF^\ast(L_U, L_V) = 0 = HF^\ast(L_V, L_U)$.
\end{lemma}
\begin{proof}
During the obvious isotopy of $U$ outward to become disjoint from $V$, their conormals never intersect at infinity.
\end{proof}
By a \emph{stable ball}, we mean a contractible open set with smooth boundary.
The reason we study stable balls is that we do not know how to prove that for a subanalytic triangulation, the inward perturbation (in the sense of Section \ref{constructibleconormals}) of an open star is a ball; it is, however, obviously a stable ball.
To compute Floer cohomology between conormals of stable balls, we reduce to the case of conormals to balls by stabilizing (i.e.\ taking their product with conormals to standard balls in $\mathbb R^k$) and appealing to the K\"unneth theorem for Floer cohomology.
We begin by showing that the stabilization of a stable ball is indeed a ball (thus justifying the name).
This uses the following famous corollary of the $h$-cobordism theorem:
\begin{theorem}
A stable ball of dimension $\ge 6$ with simply connected boundary is a ball.\qed
\end{theorem}
\begin{corollary}\label{stableballisstableball}
Let $M$ be a stable ball. Then $M\times I^k$ is a ball provided $\dim M+k\ge 6$ and $k\geq 1$.\qed
\end{corollary}
\begin{proof}
We just need to check that the boundary of $M\times I^k$ is simply connected.
It suffices to show that for any stable ball $N$ of dimension $\ge 2$, the boundary of $N\times I$ is simply connected.
The boundary of $N\times I$ is (up to homotopy) two copies of $N$ glued along their common boundary.
Since $N$ is contractible, the fundamental group of this gluing vanishes provided $\partial N$ is connected.
If $\partial N$ were disconnected, then by Poincar\'e duality, the cohomology group $H^{\dim N-1}(N)$ would be nonzero, which contradicts contractibility as $\dim N\ge 2$.
\end{proof}
\begin{proposition} \label{fakecontinuation}
Let $U,V\subseteq M$ be stable balls with $\overline U\subseteq V$.
Then $HF^\ast(L_V, L_U) = \mathbb Z$, and it is equipped with a canonical generator $1_{VU}$ (which is just the continuation map when $U$ and $V$ are balls).
These generators behave well under composition: for any triple of stable balls $U,V,W\subseteq M$ with $\overline U\subseteq V$ and $\overline V\subseteq W$, we have $1_{WV} 1_{VU} = 1_{WU}$.
\end{proposition}
\begin{proof}
We multiply by $L_U,L_V$ by $L_{B_{1}(0)},L_{B_{2}(0)} \subseteq T^*\mathbb R^k$ where $k$ is sufficiently large to guarantee that $U\times B_1(0)$ and $V\times B_2(0)$ are balls by Corollary \ref{stableballisstableball}.
By the K\"unneth formula for Floer cohomology (see e.g.\ \cite[Lemma 6.3]{gpsstructural}), we have
$$HF^*(L_V \times L_{B_{2}(0)}, L_U \times L_{B_{1}(0)}) = HF^*(L_V, L_U) \otimes HF^*(L_{B_{2}(0)}, L_{B_{1}(0)}) = HF^*(L_V, L_U).$$
On the other hand, by the result for balls Lemma \ref{continuationballs}, we have
$$HF^*(L_V \times L_{B_{2}(0)}, L_U \times L_{B_{1}(0)})=HF^*(L_{V \times B_{2}(0)}, L_{U \times B_{1}(0)})=\mathbb Z.$$
After arguing that the above identification is compatible with rounding of corners, this defines the canonical generator $1_{VU}\in HF^*(L_V, L_U)$.
The proof that $1_{WV} 1_{VU} = 1_{WU}$ is the same: stabilize to reduce to the corresponding fact for honest continuation maps.
\end{proof}
\begin{corollary} \label{squish}
Let $U \subseteq M$ be any stable ball. Then the map $L_U\to T^*_pM$ from Proposition \ref{fakecontinuation} is an isomorphism in $\mathcal W(T^*M)$ for any point $p\in U$.
\end{corollary}
\begin{proof}
By pushing forward, it suffices to treat the case $M=U^+$.
It further suffices to show the result after applying the K\"unneth embedding $\mathcal W(T^*U^+)\hookrightarrow\mathcal W(T^*(U^+\times I^k))$ (see \cite[Theorem 1.5]{gpsstructural}).
The canonical map $L_U\to[\operatorname{fiber}]$ from Proposition \ref{fakecontinuation} is, by definition, sent by the K\"unneth functor to the continuation map $L_{U\times I^k}\to[\operatorname{fiber}]$ (which is defined since the stabilized stable ball $U^+\times I^k$ is a ball).
The continuation map is an isomorphism in the wrapped Fukaya category of the ball, so we are done by full faithfulness of K\"unneth.
\end{proof}
There is similarly an improved version of Lemma \ref{deformtozero}:
\begin{lemma} \label{fakedeformtozero}
Let $V$ be an open set with smooth boundary, and let $U$ be stable ball such that
$U \cap \partial V$ is also a stable ball.
Then $HF^\ast(L_U, L_V) = 0 = HF^\ast(L_V, L_U)$.
\end{lemma}
\begin{proof}
Stabilization reduces to Lemma \ref{deformtozero} (note that $U \cap \partial V$ necessarily divides $U$ into two stable balls).
\end{proof}
A more subtle result about stable balls is the following, which will be important later:
\begin{proposition} \label{relsquish}
Let $X^m \subseteq Y^n$ be an inclusion of stable balls, with $\partial X \subseteq \partial Y$.
Assume there exists another stable ball (with corners) $Z^{m +1} \subseteq Y^n$ such that $\partial Z$ is the union of $X$ with a smooth submanifold
of $\partial Y$.
Then the map $L_Y \to L_{B_\epsilon(x)}$ from Proposition \ref{fakecontinuation} is an isomorphism in $\mathcal W(T^*Y, N^*_\infty X)$ for any $x \in X$.
\end{proposition}
\begin{proof}
By stabilization, we reduce to the case that $X$, $Y$, and $Z$ are all balls.
This implies that, up to diffeomorphism, everything is standard: $Y$ is the unit ball, $X$ is the intersection of $Y$ with a linear subspace, and $Z$ is the intersection of $Y$ with a linear halfspace.
Indeed, since $X$ and $Z$ are balls, we can use $Z$ to push $X$ to $Z\cap\partial Y$, thus showing that $X$ is simply a slight inward pushoff of the ball $Z\cap\partial Y\subseteq\partial Y$.
By definition, the map from Proposition \ref{fakecontinuation} becomes the continuation map under stabilization.
Once everything is standard, it is obvious that the continuation map $L_Y \to L_{B_\epsilon(x)}$ is an isomorphism, since the positive isotopy $L_{B_\epsilon(x)}\to L_Y$ is disjoint from $N^*_\infty X$ at infinity.
\end{proof}
\begin{remark}
We will apply Proposition \ref{relsquish} when (before rounding) $X$ is a simplex in a triangulation, $Y$ is its star, and $Z$ is any simplex containing $X$ of dimension one larger.
\end{remark}
\subsection{Fukaya categories of conormals to stars}\label{constructibleconormals}
Let $\mathcal S$ be a stratification of $M$ by locally closed smooth submanifolds.
Whitney's conditions on $\mathcal S$ are:
\begin{itemize}
\item[(a)]For strata $X\subseteq\overline Y$ and points $x\in X$ and $y\in Y$, as $y\to x$ the tangent spaces $T_yY$ become arbitrarily close to containing $T_xX$ (uniformly over compact subsets of $X$).
\item[(b)]For strata $X\subseteq\overline Y$ and points $x\in X$ and $y\in Y$, as $y\to x$ the secant lines between $x$ and $y$ become arbitrarily close to being contained in $T_yY$ (uniformly over compact subsets of $X$).
\end{itemize}
We assume $\mathcal S$ satisfies Whitney's conditions (a) and (b).
Consider an $\mathcal S$-constructible relatively compact open set $U\subseteq M$. Fixing a Riemannian metric on $M$, let
$N_\varepsilon$ denote the $\varepsilon$-neighborhood.
For $\underline{\varepsilon} = ( \varepsilon_0,\ldots,\varepsilon_{\dim M})$ some positive real numbers, we define
\begin{equation}
U^{-\underline{\varepsilon}}:=U\Bigm\backslash\bigcup_{\mathrm{strata}\,X\subseteq M\setminus U}N_{\varepsilon_{\dim X}}X
\end{equation}
When $\varepsilon_i$ sufficiently small in terms of $\varepsilon_0,\ldots,\varepsilon_{i-1}$ (a condition we indicate simply by ``$\underline{\varepsilon}$ sufficiently small'', and other authors \cite{goresky-macpherson, nadlerzaslow} express in terms of ``fringed sets''), $U^{-\underline \varepsilon}$ is (the interior of) a manifold with corners; taking $\underline\varepsilon\to 0$ we see that $U^{-\underline \varepsilon}$ is diffeomorphic to $U$ (rel any fixed compact subset of $U$).
When the choice of sufficiently small $\underline \varepsilon$ is unimportant and not varying, we write $U^-$ for $U^{-\underline \varepsilon}$.
\begin{definition}\label{inwardperturbation}
For a constructible open subset $U$ of $S$, we refer to the $U^{-\underline \varepsilon}$ above as
the {\em inward perturbation} of $U$ with respect to $\mathcal S$.
\end{definition}
\begin{remark}
More generally, one can define, for every stratum $X_\alpha$ of $\mathcal S$, the set $X_\alpha':=N_{\varepsilon_{\dim X_\alpha}}X_\alpha\setminus\bigcup_{\beta\leq\alpha}N_{\varepsilon_{\dim X_\beta}}X_\beta$, and to any (not necessarily open) $\mathcal S$-constructible subset $\bigcup_{\alpha\in A}X_\alpha\subseteq M$, we can associate the manifold-with-corners $\bigcup_{\alpha\in A}X_\alpha'$.
\end{remark}
\begin{proposition}\label{conormalconvergence}
For sufficiently small $\underline\varepsilon>0$, the conormals $\pm L_{U^{-\underline{\varepsilon}}}$ are disjoint at infinity from $N^*_\infty\mathcal S$, and as $\underline\varepsilon\to 0$, they converge to (i.e.\ become contained in arbitrarily small neighborhoods of) $N^*\mathcal S$.
\end{proposition}
\begin{proof}
Let $x\in\partial U^-$, and let $\mathcal S_x\subseteq\mathcal S$ denote the collection of strata $X$ for which $x\in\partial N_{\varepsilon_{\dim X}}X$.
Note that $\mathcal S_x$ has a unique smallest stratum $X_x$ for sufficiently small $\underline\varepsilon$ (proof: construct $\varepsilon_i$ by induction).
Let $\xi$ be a conormal direction to $\partial U^-$ at $x$, that is $\xi$ is a linear combination of conormal directions to $\partial N_{\varepsilon_{\dim X}}X$ at $x$ for various strata $X\in\mathcal S_x$.
A conormal direction to $\partial N_{\varepsilon_{\dim X}}X$ at $x$ is, for sufficiently small $\underline\varepsilon$, arbitrarily close to the conormal sphere of $X$ near $x$.
By Whitney's condition (a), the conormal sphere of $X$ near $x$ is, in turn, for sufficiently small $\underline\varepsilon$, arbitrarily close to the conormal sphere of the minimal stratum $X_x$ near $x$.
It follows that the conormal direction $\xi$ is arbitrarily close to the conormal of the minimal stratum $X_x$ near $x$.
This shows that the conormals $\pm L_{U^{-\underline{\varepsilon}}}$ converge to $N^*\mathcal S$ as $\underline\varepsilon\to 0$.
To show that the conormals to $U^{-\underline{\varepsilon}}$ are disjoint at infinity from $N^*_\infty\mathcal S$, we use Whitney's condition (b).
The only strata whose conormals could possibly be hit are those strata inside $U$.
Given a point $x\in\partial U^{-\underline\varepsilon}$, taking $\underline\varepsilon>0$ sufficiently small ensures that the secant lines between $x$ and the nearby strata on the boundary of $U$ have span transverse to $\partial U^{-\underline\varepsilon}$.
Such secant lines are approximately tangent to the stratum containing $x$, so we conclude that this stratum is also transverse to $\partial U^{-\underline\varepsilon}$.
\end{proof}
\begin{remark}
Any subanalytic family of Legendrians inside $S^*M$ whose projections converge to $\partial U$ converges to the conormals of \emph{some refinement of $\mathcal S$}.
In contrast, the above proposition does not require refining the stratification.
\end{remark}
\begin{corollary} \label{corneringstopped}
Let $L$ be any Lagrangian, disjoint at infinity from $N^* \mathcal S$. Then for all $\underline \varepsilon$ sufficiently small,
$CF^*(L_{U^{- \underline \varepsilon}}, L) \xrightarrow{\sim} CW^*(L_{U^{- \underline \varepsilon}}, L)_{N^*_\infty\mathcal S}$.
\end{corollary}
\begin{proof}
Proposition \ref{conormalconvergence} and Lemma \ref{cofinalitycriterion} imply that taking $\underline\varepsilon\to 0$ constitutes a cofinal wrapping of $L_{U^{- \underline \varepsilon}}$ in the complement of $N^\ast_\infty\mathcal S$.
\end{proof}
Let $\mathcal S$ be a triangulation of $M$.
To each stratum $s\in\mathcal S$, we associate its open star $\star(s)$ and the conormal $L_{\star(s)}$ to its inward perturbation $\star(s)^{-\underline\varepsilon}$.
By Proposition \ref{conormalconvergence}, for $\underline\varepsilon$ sufficiently small, the conormal to $\star(s)^{-\underline\varepsilon}$ is disjoint at infinity from $N^*_\infty\mathcal S$, so it defines an object of $\mathcal W(T^*M,N^*_\infty\mathcal S)$.
Since $\star(s)$ is contractible, $L_{\star(s)}$ is the conormal to a stable ball, and hence the results of the previous subsection apply, allowing us to deduce the following:
\begin{proposition}\label{homsbetweenstars}
We have
\begin{equation}
HW^*(L_{\star(s)},L_{\star(t)})=\begin{cases}\mathbb Z&t\to s\cr0&\mathrm{otherwise}\end{cases}
\end{equation}
generated in the former case by the map from Proposition \ref{fakecontinuation}.
\end{proposition}
\begin{proof}
Fix a small $\underline\varepsilon>0$ and let $\underline\delta\to 0$.
By Proposition \ref{conormalconvergence}, the wrapped Floer cohomology $HW^*(L_{\star(s)},L_{\star(t)})$ is calculated by $HF^*(L_{\star(s)^{-\underline\delta}},L_{\star(t)^{-\underline\varepsilon}})$.
Now if $t\to s$, then $\star(t)^{-\underline\varepsilon}\subseteq \star(s)^{-\underline\delta}$ is an inclusion of stable balls, so Proposition \ref{fakecontinuation} produces a canonical generator of $HF^*(L_{\star(s)^{-\underline\delta}},L_{\star(t)^{-\underline\varepsilon}})=\mathbb Z$.
Now suppose that $t\nrightarrow s$.
If $\star(s)\cap\star(t)=\emptyset$, then the desired vanishing is trivial.
Otherwise, we have $\star(s)\cap\star(t)=\star(r)$ where $r$ is the simplex spanned by the union of the vertices of $s$ and $t$.
To show the desired vanishing, it suffices by Proposition \ref{fakedeformtozero} to show that $\star(t)^{-\underline\varepsilon}\cap\partial\star(s)^{-\underline\delta}$ is a stable ball.
This space $\star(t)^{-\underline\varepsilon}\cap\partial\star(s)^{-\underline\delta}$ is homotopy equivalent to the star of $t$ inside the link of $s$, and so is contractible.
\end{proof}
It will be convenient to have another perspective on the objects $L_{\star(s)}$.
Let $L_s$ denote the conormal to a small ball centered at any point on the stratum $s$ (this conormal is disjoint from $N^*_\infty\mathcal S$ at infinity by Whitney's condition (b)).
One reason the $L_s$ are nice to consider is the following calculation:
\begin{lemma}\label{Lshoms}
For any $\mathcal S$-constructible open set $U$, we have
\begin{equation}
HW^*(L_U,L_s)_{N^*_\infty\mathcal S}=\begin{cases}\mathbb Z&\star(s)\subseteq U\cr 0&\mathrm{otherwise}\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
We calculate using Corollary \ref{corneringstopped}.
If $s$ is a stratum in the interior of $U$, then the ball centered at $s$ is contained in $U$, so there is a single intersection point.
If $s$ is a stratum not contained in the closure of $U$, then the morphism space obviously vanishes since the two Lagrangians are disjoint.
Finally, we claim that if $s$ is a stratum on the boundary of $U$, the morphism space still vanishes.
To see this, start with a small $\underline\varepsilon>0$, and choose $L_s$ to be the conormal of a small ball disjoint from $U^{-\underline\varepsilon}$.
Now we take $\underline\varepsilon\to 0$ with this $L_s$ fixed, and we claim that the outward conormal to $U^{-\underline\varepsilon}$ never passes through the outward conormal of this small ball.
Indeed, the portion of $\partial U^{-\underline\varepsilon}$ coming from $\partial N_\varepsilon(s)$ will be tangent to the small ball, however with the opposite coorientation.
The remaining nearby parts of $\partial U^{-\underline\varepsilon}$, namely coming from $\partial N_\varepsilon(t)$ for strata $t$ whose boundaries contain $s$, will be transverse to the boundary of the small ball by Whitney's condition (b): any secant line from the center of the ball to a point on its boundary intersected with $t$ is, by Whitney's condition (b) approximately tangent to $t$ (hence to $\partial N_\varepsilon(t)$).
\end{proof}
Another reason that the $L_s$ are nice to consider is that we can show using the wrapping exact triangle and stop removal that they (split-)generate:
\begin{proposition}\label{smallstarsgen}
For any stratification $\mathcal S$, the objects $L_s$ for strata $s$ split-generate $\mathcal W(T^*M,N^*_\infty\mathcal S)$.
\end{proposition}
\begin{proof}
Denote by $N^*_\infty\mathcal S_{\leq k}$ the stratification where we keep all strata of dimension $\leq k$ and combine all other strata into a single top stratum.
We consider the sequence of categories
\begin{multline}
\mathcal W(T^*M,N^*_\infty\mathcal S)=\mathcal W(T^*M,N^*_\infty\mathcal S_{\leq n-1})\to\mathcal W(T^*M,N^*_\infty\mathcal S_{\leq n-2})\to\cdots\cr\cdots\to\mathcal W(T^*M,N^*_\infty\mathcal S_{\leq 0})\to\mathcal W(T^*M).
\end{multline}
Each of these functors removes a locally closed Legendrian submanifold $N^*_\infty\mathcal S_{\leq k}\setminus N^*_\infty\mathcal S_{\leq k-1}$, and thus by stop removal Theorem \ref{stopremoval}, is the quotient by the corresponding linking disks.
The linking disk at a point on $N^*_\infty\mathcal S_{\leq k}\setminus N^*_\infty\mathcal S_{\leq k-1}$ can be described as follows.
A point on $N^*_\infty\mathcal S_{\leq k}\setminus N^*_\infty\mathcal S_{\leq k-1}$ is simply a point $x$ on a $k$-dimensional stratum together with a covector $\xi$ at $x$ conormal to the stratum.
Consider a small ball $B_a$ centered at $x$, and consider a smaller ball $B_b\subseteq B_a$ disjoint from the stratum containing $x$.
There is a family of balls starting at $B_a$ and shrinking down to $B_b$ whose boundaries are tangent to the stratum containing $x$ only at $(x,\xi)$.
It follows from the wrapping exact triangle Theorem \ref{wrapcone} that the cone on the resulting continuation map $L_{B_a}\to L_{B_b}$ is precisely the linking disk at $(x,\xi)$.
We have thus shown that the linking disks to each locally closed Legendrian $N^*_\infty\mathcal S_{\leq k}\setminus N^*_\infty\mathcal S_{\leq k-1}$ are generated by the objects $L_s$.
By Theorem \ref{generation} above, these $L_s$ also split-generate the final category $\mathcal W(T^*M)$.
We conclude that the $L_s$ split-generate $\mathcal W(T^*M,N^*_\infty\mathcal S)$, as the quotient by all of them vanishes.
\end{proof}
\begin{remark}
A small variation on the above proof and an appeal to \cite[Theorem 1.9]{gpsstructural} shows that the objects $L_s$ in fact generate $\mathcal W(T^*M,N^*_\infty\mathcal S)$.
We give the weaker argument above to minimize the results we need to appeal to.
\end{remark}
\begin{proposition}\label{bigsmallstars}
The map $L_{\star(s)}\to L_s$ from Proposition \ref{fakecontinuation} is an isomorphism in $\mathcal W(T^*M,N^*_\infty\mathcal S)$.
\end{proposition}
\begin{proof}
We proceed by induction on the codimension of $s$.
When $s$ has codimension zero, the desired statement follows from Corollary \ref{squish}.
Now suppose that $s$ has positive codimension.
For any $t$ of strictly smaller codimension than $s$, we have $\operatorname{Hom}(L_{\star(t)},L_{\star(s)})=0$ by Proposition \ref{homsbetweenstars} and $\operatorname{Hom}(L_{\star(t)},L_s)=0$ by Lemma \ref{Lshoms}.
Now by the discussion in the proof of Proposition \ref{smallstarsgen}, the functor
\begin{equation}
\mathcal W(T^*M,N^*_\infty\mathcal S)\to\mathcal W(T^*M,N^*_\infty\mathcal S_{\leq\dim s})
\end{equation}
quotients by cones of $L_t$ for $t$ of strictly smaller codimension than $s$.
By the induction hypothesis and the calculations of the previous paragraph, such cones are left-orthogonal to $L_s$ and $L_{\star(s)}$.
Hence it suffices to check that $L_{\star(s)}\to L_s$ is an isomorphism in $\mathcal W(T^*M,N^*_\infty\mathcal S_{\leq\dim s})$.
Finally, we observe that $L_{\star(s)}\to L_s$ is an isomorphism in $\mathcal W(T^*M,N^*_\infty\mathcal S_{\leq\dim s})$ by Proposition \ref{relsquish}.
Namely, we take $Y=\star(s)^-$, $X=s\cap\star(s)^-$, and $Z=t\cap\star(s)^-$ for any simplex $t$ containing $s$ and of one higher dimension.
\end{proof}
\begin{remark}
For a smooth triangulation $\mathcal S$, there is an obvious positive isotopy from $L_s$ to $L_{\star(s)}$ disjoint from $N^*_\infty\mathcal S$ (thus proving Proposition \ref{bigsmallstars} in this case), obtained by expanding a small ball centered at a point on $s$ to $\star(s)$, keeping the boundary transverse to the strata of $\mathcal S$.
We do not know whether this proof can be generalized from smooth triangulations to subanalytic triangulations.
\end{remark}
\subsection{Functors from poset categories to Fukaya categories}
\begin{definition}\label{posettocohofukaya}
Let $M$ be a manifold with stratification $\mathcal S$, and let
$U: \Pi \to \operatorname{Op}_\mathcal S(M)$ be a map from a poset $\Pi$ to the poset of $\mathcal S$-constructible open subsets of $M$.
Suppose further that each $U(\pi)^-$ (from Section \ref{constructibleconormals}) is a stable ball.
Define a functor on cohomology categories
\begin{equation}
H^*F_U:\mathbb Z[\Pi] \to H^*\mathcal W(T^\ast M, N^*_\infty\mathcal S)^\mathrm{op}
\end{equation}
by $H^*F_U(\pi):=L_{U(\pi)}$ and $H^*F_U(1_{\pi,\pi'})=1_{U(\pi'),U(\pi)}\in HW^*(L_{U(\pi')}, L_{U(\pi)})$ is the canonical generator from Proposition \ref{fakecontinuation}.
\end{definition}
\begin{remark}
Note that this definition of $H^*F_U$ depends on having chosen the correct $K(\mathbb Z/2,1)$-bundle over $T^\ast M$ to twist by (compare \S\ref{gradorsec}).
Having chosen the wrong such bundle would show up in the functor respecting composition only up to a sign.
The resulting $2$-cocycle, or rather its class in $H^2(N\Pi,\mathbb Z/2)$, would represent (the pullback to $N\Pi$ from $M$ of) the obstruction to choosing continuously varying relative $\operatorname{Pin}$-structures on all cotangent fibers.
\end{remark}
\begin{proposition}\label{ffhenceainf}
Let $H^*F:\mathbb Z[\Pi]\to H^*\mathcal C$ be any functor on cohomology categories such that $H^*\mathcal C(F(x),F(y))$ is free and concentrated in degree zero for every pair $x\leq y\in\Pi$.
Then there exists an $A_\infty$ functor $F$ lifting $H^*F$, and moreover the space of natural isomorphisms between any two such lifts is contractible.
\end{proposition}
\begin{proof}
We show existence of a lift $F$ by induction.
Take $F^1$ to be any map with the correct action on cohomology.
Having chosen $F^1,\ldots,F^{k-1}$, the obstruction to the existence of an $F^k$ satisfying the $A_\infty$ functor equations of order $k$ is a degree $2-k$ cohomology class in (the cohomology of)
\begin{equation}
\prod_{\pi_0\leq\cdots\leq\pi_k\in\Pi}\operatorname{Hom}(\mathcal C(F(\pi_0),F(\pi_1))\otimes\cdots\otimes\mathcal C(F(\pi_{k-1}),F(\pi_k)),\mathcal C(F(\pi_0),F(\pi_k))).
\end{equation}
Appealing to cofibrancy of $\mathcal C(-,-)$ and the fact that $H^*\mathcal C(F(x),F(y))$ is free and concentrated in degree zero, we conclude that the obstruction class must vanish when $k\geq 3$ for degree reasons.
When $k=2$, the obstruction class measures the failure of $H^*F$ to respect composition, so by hypothesis this obstruction also vanishes. Hence in either case, there exists an $F^k$ compatible with the previously chosen $F^1,\ldots,F^{k-1}$.
(Compare \cite[Lemma 1.9]{seidelbook}, where this obstruction theory argument is explained in more detail.)
To analyze the space of natural quasi-isomorphisms, we again argue inductively to show that all obstructions vanish (again using the $\mathbb Z$-grading).
\end{proof}
\begin{corollary}\label{posettofukaya}
There is a unique up to contractible choice $A_\infty$ functor
\begin{equation}
F_U:\mathbb Z[\Pi] \to \mathcal W(T^\ast M, N^*_\infty\mathcal S)^\mathrm{op}
\end{equation}
lifting the functor on cohomology categories from Defintion \ref{posettocohofukaya}.
\end{corollary}
\begin{proof}
By Corollary \ref{corneringstopped}, the wrapped Floer cohomology group $HW^*(L_{U(\pi)}, L_{U(\pi')})$ is simply the Floer cohomology of two nested stable balls, which is $\mathbb Z$ by Proposition \ref{fakecontinuation}.
Thus Proposition \ref{ffhenceainf} is applicable.
\end{proof}
\begin{remark}
To extend Corollary \ref{posettofukaya} to the Fukaya category with a $\mathbb Z/N$-grading, we would need to add to the requirement that $F$ (and natural transformations $F_1\to F_2$) must lift to $\mathbb Z$-graded categories locally (the $\mathbb Z$-grading is only defined locally, over any contractible open subset of $M$).
\end{remark}
\begin{definition}\label{Fdefn}
For a triangulation $\mathcal S$, let
\begin{equation}
F_\mathcal S:\mathbb Z[\mathcal S]\to\mathcal W(T^*M,N^*_\infty \mathcal S)^\mathrm{op}
\end{equation}
denote the functor induced from Definition \ref{posettocohofukaya} and Corollary \ref{posettofukaya} by the map associating to each simplex of $\mathcal S$ its open star.
\end{definition}
\begin{theorem}\label{Fequivalence}
The functor $F_\mathcal S$ is a Morita equivalence.
\end{theorem}
\begin{proof}
Proposition \ref{homsbetweenstars} shows is full faithfulness of $F_\mathcal S$, and Propositions \ref{smallstarsgen} and \ref{bigsmallstars} together show essential surjectivity of $F_\mathcal S$ (after passing to $\operatorname{Perf}$).
\end{proof}
We now show that $F_\mathcal S$ is compatible with refinement (compare Lemma \ref{posetmeaningpushforward}):
\begin{theorem}\label{Frefinement}
For any refinement of triangulations $\mathcal S'$ refining $\mathcal S$, the following diagram commutes:
\begin{equation}
\begin{tikzcd}
\mathbb Z[\mathcal S']\ar{r}{F_{\mathcal S'}} \ar{d}[swap]{r} & \mathcal W(T^\ast M,N^\ast_\infty \mathcal S')^\mathrm{op}\ar{d}{\rho}\\
\mathbb Z[\mathcal S]\ar{r}{F_\mathcal S} & \mathcal W(T^\ast M,N^\ast_\infty \mathcal S)^\mathrm{op}
\end{tikzcd}
\end{equation}
up to contractible choice.
\end{theorem}
\begin{proof}
There are two functors $\rho\circ F_{\mathcal S'}$ and $F_{\mathcal S}\circ r$ from $\mathbb Z[\mathcal S']$ to $\mathcal W(T^\ast M,N^\ast_\infty \mathcal S)$.
By Corollary \ref{ffhenceainf}, it suffices to define a canonical natural isomorphism between the induced functors on cohomology categories.
It is most natural to define this canonical natural isomorphism in the direction $F_{\mathcal S}\circ r\Longrightarrow\rho\circ F_{\mathcal S'}$.
To a stratum $s$ of $\mathcal S'$, the composition $F_{\mathcal S}\circ r$ associates the conormal of $\star_\mathcal S(r(s))$, and the composition $\rho\circ F_{\mathcal S'}$ associates the conormal of $\star_{\mathcal S'}(s)$.
Both are stable balls, and there is an inclusion $\star_\mathcal S(r(s))\supseteq\star_{\mathcal S'}(s)$, so by Proposition \ref{fakecontinuation} there is a canonical map from one to the other.
Using the composition property of Proposition \ref{fakecontinuation}, it is easy to check that this defines a natural transformation $H^*(F_{\mathcal S}\circ r)\Longrightarrow H^*(\rho\circ F_{\mathcal S'})$.
This natural transformation is in fact a natural isomorphism since the natural maps from both $L_{\star_\mathcal S(r(s))}$ and $L_{\star_{\mathcal S'}(s)}$ to $L_s=L_{r(s)}$ are isomorphisms by Proposition \ref{bigsmallstars}.
\end{proof}
\subsection{In conclusion}
\begin{theorem}\label{fukcasting}
The functor $\Lambda\mapsto\mathcal W(T^*M,\Lambda)^\mathrm{op}$ is a microlocal Morse theater in the sense of Definition \ref{microlocalmorsetheatre}, which casts the linking disks at smooth points of $\Lambda$ as the Morse characters.
\end{theorem}
\begin{proof}
Definition \ref{Fdefn} and Theorems \ref{Fequivalence} and \ref{Frefinement} give the identification between $\mathcal S\mapsto\operatorname{Perf}\mathcal S$ and $\mathcal S\mapsto\mathcal W(T^*M,N^*_\infty\mathcal S)^\mathrm{op}$.
Stop removal Theorem \ref{stopremoval} says that $\mathcal W(T^*M,\Lambda')\to\mathcal W(T^*M,\Lambda)$ is the quotient by the linking disks at the smooth points of $\Lambda'\setminus\Lambda$.
It therefore suffices to show that the Morse characters are precisely these linking disks.
Recall from Definition \ref{morsecharacter} that a Morse character at a smooth point $p\in\Lambda$ is defined as follows.
We choose a function $f:M\to\mathbb R$ and an $\epsilon>0$ such that $f$ has no critical values in $[-\epsilon,\epsilon]$ and $df$ is transverse to $\Lambda$ over $f^{-1}[-\epsilon,\epsilon]$, intersecting it only at $p$ (where $f$ vanishes).
We also choose a triangulation $\mathcal S$ such that $\Lambda\subseteq N^*_\infty\mathcal S$ and $f^{-1}(-\infty,-\epsilon)$ and $f^{-1}(-\infty,\epsilon)$ are constructible.
The Morse character associated to these choices is then defined as the image in $\mathcal W(T^*M,\Lambda)$ of
$$\operatorname{cone}(1_{f^{-1}(-\infty, -\epsilon)} \to 1_{f^{-1}(-\infty, \epsilon)})\in\operatorname{Perf}\mathcal S=\mathcal W(T^*M,N^*_\infty\mathcal S)^\mathrm{op}.$$
To show that this cone is indeed the linking disk at $p$ in $\mathcal W(T^*M,\Lambda)$, we appeal to the wrapping exact triangle Theorem \ref{wrapcone}.
It thus suffices to show that $1_{f^{-1}(-\infty, \pm\epsilon)}\in\mathcal W(T^*M,N^*_\infty\mathcal S)$ is identified with the conormal of $f^{-1}(-\infty, \pm\epsilon)$, and the map $1_{f^{-1}(-\infty, -\epsilon)} \to 1_{f^{-1}(-\infty, \epsilon)}$ is identified with the continuation map associated to the obvious positive isotopy from the conormal of $f^{-1}(-\infty, -\epsilon)$ to the conormal of $f^{-1}(-\infty, \epsilon)$ coming from the fact that $f$ has no critical values in the interval $[-\epsilon,\epsilon]$.
We first note that because the level sets $f^{-1}(\pm \epsilon)$ are smooth, it is straightforward to produce an isotopy between the
above inward conormals and the \emph{the inward perturbation with respect to $\mathcal S$} of $f^{-1}(-\infty, \pm\epsilon)$, in the sense of
Definition \ref{inwardperturbation}. Henceforth we work with the latter, and denote them as $L_{f^{-1}(-\infty, \pm\epsilon)}$.
It follows from Lemma \ref{Lshoms} that
\begin{equation}
HW^*(L_{f^{-1}(-\infty, \pm\epsilon)},L_s)_{N^*_\infty\mathcal S}=\begin{cases}\mathbb Z&\star(s)\subseteq f^{-1}(-\infty, \pm\epsilon)\cr 0&\mathrm{otherwise}\end{cases}
\end{equation}
As $L_{\star(s)} = L_s$,
this identifies the pullback under $F_\mathcal S$ (Definition \ref{Fdefn}) of the Yoneda module of $L_{f^{-1}(-\infty, \pm\epsilon)}$ with
the indicator $1_{f^{-1}(-\infty, \pm\epsilon)}$.
As $\partial_\infty L_{f^{-1}(-\infty, \epsilon)}$ falls immediately into the stop $N^*_\infty\mathcal S$, we have
$$HW^*(L_{f^{-1}(-\infty, \epsilon)},L_{f^{-1}(-\infty,-\epsilon)})_{N^*_\infty\mathcal S} = HF^*(L_{f^{-1}(-\infty, \epsilon)},L_{f^{-1}(-\infty,-\epsilon)}).$$
By Yoneda, the map $1_{f^{-1}(-\infty, -\epsilon)} \to 1_{f^{-1}(-\infty, \epsilon)}$ corresponds to an element of the above group.
It remains to show that this element is the continuation map.
We test both against the generators $L_s$ for $\mathcal W(M, N^*_\infty \mathcal S)$; for $1_{f^{-1}(-\infty, -\epsilon)} \to 1_{f^{-1}(-\infty, \epsilon)}$
the answer is determined by Theorem \ref{Fequivalence}.
Both tests give zero unless the stratum $s$ is contained in ${f^{-1}(-\infty,-\epsilon)}$; in this case we may arrange that
$L_s$ is the conormal to a subset of ${f^{-1}(-\infty,-\epsilon)}$ as well.
For these $L_s$, the evident isotopy
$L_{f^{-1}(-\infty,-\epsilon)} \rightsquigarrow L_{f^{-1}(-\infty, \epsilon)}$ is disjoint at infinity from $L_s$, so
the desired assertion follows from
\cite[Lemma 3.26]{gpssectorsoc}.
\end{proof}
\section{Examples} \label{sec:corollaries}
\subsection{Cotangent bundles}
\label{cotangent}
Let $M$ be a smooth manifold.
The cotangent fibers $F_q\in\mathcal W(T^*M)$ generate by Abouzaid \cite{abouzaidcriterion,abouzaidcotangent} when $M$ is closed and by \cite[Theorem 1.9]{gpsstructural} in general.
When $M$ is closed, Abbondandolo--Schwarz \cite{abbondandoloschwarz} and Abouzaid \cite{abouzaidtwisted} calculated the endomorphism algebra of the fiber as $CW^\ast(F_q,F_q)=C_{-\ast}(\Omega_q M)$ (using relative $\operatorname{Pin}$ structures as in Section \ref{gradorsec}).
The present Theorem \ref{sheaffukayaequivalence} (which does not depend on any of \cite{abbondandoloschwarz, abouzaidtwisted, abouzaidcotangent,abouzaidcriterion}) gives a proof of this fact for all (not necessarily closed) $M$:
\begin{corollary}
There is a quasi-isomorphism $CW^\ast(F_q, F_q) = C_{-\ast}(\Omega_q M)$.
Moreover if $M \subseteq N$ is a codimension zero inclusion, there is a commutative diagram
\begin{equation}
\begin{tikzcd}
CW^*(F_q, F_q)_{T^*M} \ar{r} \ar{d} & C_{-\ast}(\Omega_q M) \ar{d} \\
CW^*(F_q, F_q)_{T^*N} \ar{r} & C_{-\ast}(\Omega_q N)
\end{tikzcd}
\end{equation}
where the left hand vertical arrow is covariant inclusion and the right hand vertical arrow is induced by pushforward of loops.
\end{corollary}
\begin{proof}
Note that there exists a real analytic structure on $M$ whose induced smooth structure agrees with the given one.
Taking $\Lambda = \emptyset$ in Theorem \ref{sheaffukayaequivalence} gives $\operatorname{Perf} \mathcal W(T^*M) = \operatorname{Sh}_\emptyset(M)^c$.
It is well known the latter is the category $\operatorname{Perf} C_{-\bullet}(\Omega_qM)$, e.g.\ because both are the global sections of the constant cosheaf of linear categories with costalk $\operatorname{Mod}\mathbb Z$.
We may derive the more precise assertion that $C_{-\bullet}(\Omega_qM)$ is endomorphisms of the cotangent fiber by following a fiber through the equivalence, e.g.\ by considering the inclusion of the cotangent bundle of a disk, or equivalently by introducing a stop along the conormal of the boundary of a disk and then removing it.
\end{proof}
\subsection{Plumbings}
Many authors have studied Fukaya categories of plumbings \cite{abouzaidplumbing,abouzaidsmith, etgu-lekili} and their sheaf counterparts \cite{bezrukavnikov-kapranov}.
Here we compute the wrapped category of a plumbing.
Let $\Pi_{2n}$ be the Liouville pair $(\mathbb C^n, \partial_\infty (\mathbb R^n \cup i\mathbb R^n))$; we term it the plumbing sector.
Plumbings are formed by taking a manifold $M$ (usually disconnected) with spherical boundary $\partial M = \coprod S^{n-1}$,
and gluing the Liouville pair $(T^*M, \partial M)$ to some number of plumbing sectors along the spheres.
One can model the wrapped Fukaya category of the plumbing sector directly in sheaf theory: we can view it as the pair
$(T^*\mathbb R^n, N^*_\infty\{0\})$, and the category $\operatorname{Sh}_{N^*_\infty\{0\}}(\mathbb R^n)$ has a well-known description in terms of the Fourier
transform as described in \cite{bezrukavnikov-kapranov}.
This category is equivalent to $\mathcal W(\Pi_{2n})$ by Theorem \ref{sheaffukayaequivalence}.
To apply the gluing results of \cite{gpsstructural}, however, we need to know how the wrapped Fukaya categories of the two boundary sectors include, which is slightly more than what Theorem \ref{sheaffukayaequivalence} tells us.
Hence we give a direct computation of the wrapped Fukaya category of the plumbing sector.
Take a positive Reeb pushoff of the boundary of a cotangent fiber in $T^*\mathbb R^n$, so it is now the outward conormal of a small ball.
Deleting the original cotangent fiber, we obtain the Liouville sector $T^* S^{n-1} \times \mathcal A_2$ where $\mathcal A_2$ denotes the Liouville sector $(\mathbb C,\{e^{2\pi ik/3}\}_{k=0,1,2})$.
We can get back to the plumbing sector $\Pi_{2n}$ by adding back the missing fiber, which amounts to attaching a Weinstein handle along one of the boundary sectors $T^*(S^{n-1} \times I)$.
We may thus deduce from \cite[Thm.\ 1.20, Thm.\ 1.5, and Cor.\ 1.11]{gpsstructural} that:
\begin{lemma}
$$\operatorname{Perf}\mathcal W(\Pi_{2n}) = \operatorname{Perf}(\operatornamewithlimits{colim}(\operatorname{Perf}(\bullet) \leftarrow \operatorname{Perf} C_*(\Omega S^{n-1}) \to \operatorname{Perf}(\bullet \to \bullet) \otimes \operatorname{Perf} C_*(\Omega S^{n-1})))$$
\end{lemma}
Gluing in the remaining manifolds, we conclude:
\begin{corollary}
The wrapped Fukaya category of a plumbing is calculated by ($\operatorname{Perf}$ applied to) the colimit of the diagram
$$
\begin{tikzcd}
\coprod \operatorname{Perf}(\bullet) & & \\
\coprod \operatorname{Perf} C_*(\Omega S^{n-1}) \ar{u} \ar{r} & \coprod \operatorname{Perf}(\bullet \to \bullet) \otimes \operatorname{Perf} C_*(\Omega S^{n-1}) \\
& \coprod \operatorname{Perf} C_*(\Omega S^{n-1}) \ar{u} \ar{r} & \coprod \operatorname{Perf} C_*(\Omega M_i)
\end{tikzcd}
$$
where $M_i$ are the components of $M$.
\end{corollary}
\subsection{Proper modules and infinitesimal Fukaya categories} \label{NZ}
Recall that for a dg or $A_\infty$ category $\mathcal C$, we write $\operatorname{Prop} \mathcal C := \operatorname{Fun}(\mathcal C, \operatorname{Perf}\mathbb Z)$ for the category
of proper (aka pseudo-perfect) modules. It is immediate from our main result that
$\operatorname{Prop} \operatorname{Sh}_\Lambda(M)^c = \operatorname{Prop} \mathcal W(T^*M, \Lambda)^\mathrm{op}$.
Recall from Corollary \ref{propersheaves} that any proper $\operatorname{Sh}_\Lambda(M)^c$-module is representable by an object of $\operatorname{Sh}_\Lambda(M)$ with perfect stalks, i.e.\ a constructible sheaf in the classical sense.
Let us describe some objects in the Fukaya category $\mathcal W(T^*M, \Lambda)$ which necessarily give rise to proper modules (and thus to sheaves on $M$ with perfect stalks, microsupported inside $\Lambda$).
\begin{definition}
For any stopped Liouville manifold $(X, \mathfrak f)$, we define the \emph{forward stopped subcategory} $\mathcal W^{\epsilon}(X, \mathfrak f)$ to be
the full subcategory of $\mathcal W(X,\mathfrak f)$ generated by Lagrangians which admit a positive wrapping into $\mathfrak f$, meaning $\partial_\infty L$ becomes contained in arbitrarily small neighborhoods of $\mathfrak f$.
By Lemma \ref{cofinalitycriterion}, such a wrapping is necessarily cofinal.
\end{definition}
\begin{example}
If $\mathfrak f$ admits a ribbon $F$ (or, alternatively, is itself equal to a Liouville hypersurface $F$) then $\mathcal W^{\epsilon}(X, \mathfrak f)$ contains all Lagrangians whose boundary at infinity is contained in a neighborhood of a small negative Reeb pushoff of $\mathfrak f$ (or $F$).
\end{example}
\begin{example}
All compact (exact) Lagrangians are contained in $\mathcal W^{\epsilon}(X, \mathfrak f)$, as their boundary at infinity $\emptyset$ is wrapped into $\mathfrak f$ by the trivial wrapping.
\end{example}
\begin{proposition}
All objects of $\mathcal W^{\epsilon}(X, \mathfrak f)$ co-represent proper modules over $\mathcal W(X, \mathfrak f)$; that is, the restriction of the Yoneda embedding $\mathcal W(X,\mathfrak f)\hookrightarrow\operatorname{Mod}\mathcal W(X, \mathfrak f)^\mathrm{op}$ to $\mathcal W^\epsilon(X,\mathfrak f)$ has image contained in $\operatorname{Prop}\mathcal W(X, \mathfrak f)^\mathrm{op}$.
\end{proposition}
\begin{proof}
Morphisms in the wrapped category can be computed by cofinally positively wrapping the first factor.
Any $L \in \mathcal W^{\epsilon}(X, \mathfrak f)$ admits such a wrapping $\{L_t\}_{t\geq 0}$ which converges at infinity to $\mathfrak f$.
It follows that after some time $t$, its boundary at infinity stays disjoint at infinity from $K$, and hence $CW^*(L,K)=CF^*(L_t,K)$ for sufficiently large $t$.
\end{proof}
\begin{corollary}\label{epsilonproper}
The equivalence $\operatorname{Perf}\mathcal W(T^*M, \Lambda)^\mathrm{op}=\operatorname{Sh}_\Lambda(M)^c$ sends $\mathcal W^{\epsilon}(T^*M, \Lambda)$ fully faithfully into $\operatorname{Prop}\operatorname{Sh}_\Lambda(M)^c$.
\end{corollary}
Recall that for a triangulation $\mathcal S$, the category $\mathcal W(T^*M, N^*_\infty \mathcal S)^\mathrm{op}$ is Morita equivalent to $\mathbb Z[\mathcal S]$, hence smooth and proper.
The generators $L_{\star(s)}$ of $\mathcal W(T^*M, N^*_\infty \mathcal S)$ used to prove this equivalence were shown in that proof to lie in $\mathcal W^\epsilon(T^*M, N^*_\infty \mathcal S)$, so we have:
\begin{proposition}\label{wrappedisinf}
For a triangulation $\mathcal S$, the inclusion $\mathcal W^{\epsilon}(T^*M, N^*_\infty \mathcal S) \subseteq \mathcal W(T^*M, N^*_\infty \mathcal S)$ is a Morita equivalence.\qed
\end{proposition}
\begin{remark}
Corollary \ref{epsilonproper} is very similar to the original Nadler--Zaslow correspondence \cite{nadlerzaslow}, restricted to Lagrangians with fixed asymptotics.
To be more precise, recall that Nadler--Zaslow wish to consider an infinitesmially wrapped Fukaya category $\mathcal W^{\inf}_\mathfrak f(T^*M)$ of Lagrangians `asymptotic at infinity to $\mathfrak f$' and then show it is equivalent to a category of sheaves on $M$ with microsupport inside $\mathfrak f$.
If $\mathfrak f$ is a smooth Legendrian and $\mathcal W^{\inf}_\mathfrak f(T^*M)$ is defined to consist of Lagrangians which are conical at infinity, ending inside $\mathfrak f$, then there is a fully faithful embedding $\mathcal W^{\inf}_\mathfrak f(T^*M)\hookrightarrow\mathcal W^\epsilon(T^*M,\mathfrak f)$, sending a Lagrangian ending inside $\mathfrak f$ to its small negative pushoff (which then tautologically wraps positively back into $\mathfrak f$).
Hence Corollary \ref{epsilonproper} recovers a version of \cite{nadlerzaslow} when $\mathfrak f$ is a smooth Legendrian.
One can certainly imagine constructing such an embedding $\mathcal W^{\inf}_\mathfrak f(T^*M)\hookrightarrow\mathcal W^\epsilon(T^*M,\mathfrak f)$ for more general (e.g.\ subanalytic isotropic) $\mathfrak f$ (for some particular definition of $\mathcal W^{\inf}_\mathfrak f(T^*M)$).
\end{remark}
\begin{remark}\label{infinitesimalproper}
We \emph{do not} know when $\mathcal W^{\epsilon}(T^*M, \Lambda)^\mathrm{op} \hookrightarrow \operatorname{Prop} \operatorname{Sh}_\Lambda(M)^c$ is a Morita equivalence.
Note that the assertion of such an equivalence (for $\mathcal W^{\inf}_\Lambda(T^*M)^\mathrm{op}$) \emph{is not} made in \cite{nadlermicrolocal}, although that work is occasionally misquoted to suggest that it is.
What is actually said is that one can get all objects of $\operatorname{Prop} \operatorname{Sh}_\Lambda(M)^c$ from twisted complexes of objects of $\mathcal W^{\inf}_{\Lambda'}(T^*M)^\mathrm{op}$ for a \emph{possibly larger} $\Lambda'$ which, as twisted complexes, pair trivially
with all Lagrangians contained in a neighborhood of $\Lambda'\setminus\Lambda$ (and thus could be said to be ``Floer-theoretically supported away from $\Lambda'\setminus\Lambda$'').
\end{remark}
To make a precise statement along the lines of Remark \ref{infinitesimalproper},
realizing a version of the Nadler--Zaslow equivalence, we have:
\begin{proposition}
If $\mathcal S$ is any triangulation of compact $M$ with $\Lambda\subseteq N^*_\infty\mathcal S$, and $\mathcal D$ denotes the collection of linking disks
to smooth points of $N^*_\infty \mathcal S \backslash \Lambda$, then
\begin{equation}
\operatorname{Prop}\operatorname{Sh}_\Lambda(M)^c = \operatorname{Prop} \mathcal W(T^*M, \Lambda)^\mathrm{op} = (\operatorname{Tw}\mathcal W^{\epsilon}(T^*M, N^*_\infty \mathcal S)^\mathrm{op})_{\operatorname{Ann}(\mathcal D)}
\end{equation}
where $\operatorname{Tw}$ denotes twisted complexes (i.e.\ any model for the the
pre-triangulated, non idempotent-completed, hull), and the subscript
$\operatorname{Ann}(\mathcal D)$ indicates taking the full subcategory of objects annihilated by $CW^*(-,D)=0$ for all $D\in\mathcal D$.
\end{proposition}
\begin{proof}
For such an $\mathcal S$, the functor $j: \mathcal W(T^*M, N^*_\infty \mathcal S) \to \mathcal W(T^*M,
\Lambda)$ is the quotient by $\mathcal D$ by Theorem \ref{stopremoval}.
Pullback of modules under any localization is a fully faithful embedding, identifying the category of modules over the localized category with the full subcategory of modules over the original category which annihilate the objects quotiented by (see \S\ref{quotients} and \cite[Lemmas 3.12 and 3.13]{gpssectorsoc}).
Properness of a module is also clearly equivalent to properness of its pullback.
We thus conclude that
$$j^*: \operatorname{Prop} \mathcal W(T^*M, \Lambda)^\mathrm{op} \hookrightarrow \operatorname{Prop} \mathcal W(T^*M, N^*_\infty \mathcal S)^\mathrm{op}$$
embeds the former as the full subcategory of the latter annihilating $\mathcal D$.
Now $\mathcal W(T^*M, N^*_\infty \mathcal S)$ (Morita equivalent to $\operatorname{Perf}\mathcal S^\mathrm{op}$ by Proposition \ref{Fequivalence}) is
smooth and proper by Lemma \ref{exceptionalpropersmooth} (since $M$ is compact and thus there are finitely many simplices).
Hence $\operatorname{Prop} \mathcal W(T^*M, N^*_\infty \mathcal S)^\mathrm{op} = \operatorname{Perf} \mathcal W(T^*M,
N^*_\infty \mathcal S)^\mathrm{op} = \operatorname{Perf} \mathcal W^{\epsilon}(T^*M, N^*_\infty \mathcal S)^\mathrm{op}$ (by Proposition \ref{wrappedisinf}).
Finally, we observe that idempotent completion is
unecessary by Lemma \ref{exceptionalidempotentcomplete}, as $\operatorname{Perf}\mathcal S$ has a generating exceptional collection.
\end{proof}
\begin{remark}
For non-compact $M$, the same proof implies that
$$\operatorname{Prop}\operatorname{Sh}_\Lambda(M)^c = (\operatorname{Prop} \mathcal W^{\epsilon}(T^*M, N^*_\infty
\mathcal S)^\mathrm{op})_{\operatorname{Ann}(\mathcal D)} \supseteq (\operatorname{Perf}\mathcal W^{\epsilon}(T^*M, N^*_\infty
\mathcal S)^\mathrm{op})_{\operatorname{Ann}(\mathcal D)}$$
but the inclusion is not generally an equality.
\end{remark}
\begin{example}
Let us explain how our `stopped' setup can be used to make ordinary (not wrapped) Floer cohomology calculations using sheaves.
Suppose given two Lagrangians $L, K \subseteq T^*M$ for which $\Lambda:=\partial_\infty L\cup\partial_\infty K$ is subanalytic.
We are interested in computing $HF^\ast(L^+, K)$. Thus consider the wrapped category $\mathcal W(T^*M, \Lambda)$ and small negative pushoffs $L^-,K^-\in\mathcal W(T^*M, \partial L \cup \partial K)$, and observe that
$$HF^*(L^+,K)=HW^*(L^-,K^-)_\Lambda.$$
By our main result, the right hand side can be computed as $\operatorname{Hom}(\mathcal F_K,\mathcal F_L)$ in the sheaf category $\operatorname{Sh}_\Lambda(M)$, provided we can determine the
sheaves $\mathcal F_L$ and $\mathcal F_K$ to which $L^-$ and $K^-$ are sent by our Theorem \ref{sheaffukayaequivalence}.
Here we make only a few observations regarding how to determine these sheaves.
Because linking disks go to microstalks and $L^-,K^-$ are forward stopped, we can see immediately
that $\mathcal F_L, \mathcal F_K$ have microstalk $\mathbb Z$ along the respective loci $\partial_\infty L,\partial_\infty K\subseteq\Lambda$.
For the same reason, for $p$ away from the front projection of $\Lambda=\partial_\infty L\cup\partial_\infty K$, we have
$$\mathcal F_L|_p \cong CF^*(L, T_p^* M) \qquad \qquad \mathcal F_K|_p \cong CF^*(K, T_p^* M).$$
In some cases, e.g.\ in case that $L$ intersects every cotangent fiber either once or not at all, this data already suffices to determine
$\mathcal F_L$. In particular, this situation occurs in \cite{STWZ}, where
sheaf calculations are made exhibiting cluster transformations arising from comparing different fillings
of Legendrian knots. The present discussion suffices to translate those calculations into calculations in Lagrangian Floer theory.
\end{example}
\subsection{Legendrians and constructible sheaves}
\label{aas}
\begin{corollary}
Let $\Lambda \subseteq J^1 \mathbb R^n \subseteq S^*\mathbb R^{n+1}$ be a smooth compact Legendrian. Let $D$ be its linking disk, and consider the algebra
$$\mathcal A_\Lambda := CW^*(D, D)_{T^* \mathbb R^{n+1}, \Lambda}.$$
Then $\operatorname{Mod}\mathcal A_\Lambda^\mathrm{op}$ is equivalent to the category $\operatorname{Sh}_\Lambda(T^* \mathbb R^{n+1})_0$ of sheaves microsupported inside $\Lambda$ and with vanishing stalk at infinity.
This equivalence identifies the forgetful functor $\operatorname{Mod}\mathcal A_\Lambda^\mathrm{op}\to\operatorname{Mod}\mathbb Z$ with the microstalk along $\Lambda$.
Hence $\operatorname{Prop}\mathcal A_\Lambda^\mathrm{op}$ is equivalent to the subcategory of $\operatorname{Sh}_\Lambda(T^*\mathbb R^{n+1})_0$ of objects with perfect microstalk along $\Lambda$ (or equivalently, with perfect stalks).
\end{corollary}
\begin{proof}
Our generation results \cite[Theorem 1.9]{gpsstructural} imply that $\mathcal W(T^*\mathbb R^{n+1}, \Lambda)$ is generated by $D$ and a cotangent fiber near infinity.
Because we assume that $\Lambda \subseteq J^1 \mathbb R^n$, the cotangent fiber at negative (in the last coordinate) infinity
can be cofinally positively wrapped without intersecting $\Lambda$, and likewise the (isomorphic) cotangent
fiber at positive infinity can be cofinally negatively wrapped without intersecting $\Lambda$.
These large wrappings are conormals to large disks in $\mathbb R^{n+1}$ containing the projection of $\Lambda$;
they thus have vanishing wrapped Floer cohomology (in both directions) with the linking disk $D$ to $\Lambda$.
Thus $\mathcal W(T^* \mathbb R^{n+1}, \Lambda)$ is generated by the two orthogonal objects $D$ and the fiber at infinity.
Denote by $\mu,\sigma\in\operatorname{Sh}_\Lambda(T^*\mathbb R^{n+1})^c$ the objects corresponding to $D$ and the fiber at infinity, respectively.
They are orthogonal, and have endomorphism algebras $\mathcal A_\Lambda^\mathrm{op}$ and $\mathbb Z$, respectively.
We have $\operatorname{Sh}_\Lambda(T^*\mathbb R^{n+1})=\operatorname{Mod}\operatorname{Sh}_\Lambda(T^*\mathbb R^{n+1})^c=\operatorname{Mod}\mathcal W(T^* \mathbb R^{n+1}, \Lambda)^\mathrm{op}=\operatorname{Mod}\mathcal A_\Lambda^\mathrm{op}\oplus\operatorname{Mod}\mathbb Z$, and this equivalence is given concretely by $\mathcal F\mapsto\operatorname{Hom}(\mu,\mathcal F)\oplus\operatorname{Hom}(\sigma,\mathcal F)$.
By Theorem \ref{sheaffukayaequivalence}, $\operatorname{Hom}(\mu,\mathcal F)$ is the microstalk along $\Lambda$ and $\operatorname{Hom}(\sigma,\mathcal F)$ is the stalk at infinity.
To see that perfect stalks is equivalent to perfect microstalks along $\Lambda$ for objects of $\operatorname{Sh}_\Lambda(T^*\mathbb R^{n+1})_0$, argue as follows.
Suppose microstalks are perfect.
Stalks are computed by $\operatorname{Hom}(\mathbb Z_{B_\epsilon(x)},\mathcal F)$ for some sufficiently small $\epsilon>0$ (in terms of $\Lambda$), since changing $\epsilon$ is non-characteristic by Whitney's condition B for the image of $\Lambda$.
Now moving $B_\epsilon(x)$ generically to infinity picks up some number of microstalks when its conormal passes through $\Lambda$ (transversally), and eventually gives zero since the stalk of $\mathcal F$ near infinity vanishes.
Thus perfect microstalks implies perfect stalks.
To see that perfect stalks implies perfect microstalks, recall that $\operatorname{Sh}_\Lambda(T^*\mathbb R^{n+1})$ is a full subcategory of $\operatorname{Sh}_\mathcal S(T^*\mathbb R^{n+1})$ for a triangulation $\mathcal S$ whose conormal contains $\Lambda$.
Perfect stalk and compact support objects of $\operatorname{Sh}_\mathcal S(T^*\mathbb R^{n+1})$ are generated by $\mathbb Z_{\star(s)}$, which all have perfect microstalks.
\end{proof}
Let us comment on the relation of the above result to the `augmentations are sheaves' statement in \cite{shende-treumann-zaslow, nrssz}.
There is an evident similarity: both relate augmentations of an algebra associated to a Legendrian to categories of sheaves microsupported
in that Legendrian. But they are not exactly the same: the algebra $\mathcal A_\Lambda$ is not by definition the Chekanov--Eliashberg dga, and
moreover in \cite{nrssz} the category of augmentations is defined by a somewhat complicated procedure, not just as proper modules over a dga.
Also in \cite{nrssz}, the authors restrict attention to augmentations, i.e.\ $1$-dimensional representations of the dga, whereas the above
result concerns the entire representation category (the underlying $\mathbb Z$-module of the representation being the microstalk), specializing to a comparison of rank $k$ representations with rank $k$ microstalk sheaves for every $k$.
In fact, $\mathcal A_\Lambda$ was conjectured by Sylvan to be a version of the Chekanov--Eliashberg dga with enhanced $C_*(\Omega \Lambda)$ coefficients.
A precise statement comparing $\mathcal A_{\Lambda}$ to such a generalized ``loop space
dga'' can be found in \cite[Conj.\ 3]{ekholm-lekili}, where it is explained that
the comparison should follow from a slight variant of the surgery techniques of
\cite{bourgeoisekholmeliashberg} (the complete proof of which has yet to
appear).
The relation between the multiple copy
construction of \cite{nrssz} and the loop space dga can also be extracted from \cite{ekholm-lekili}.
Finally we note that a version of the above discussion serves to translate between the arguments of \cite{shende-conormal} and \cite{ekholmngshende}.
\subsection{Fukaya-Seidel categories of cotangent bundles}
Let $W: T^*M \to \mathbb C$ be an exact symplectic fibration with singularities. The associated
Fukaya--Seidel category is by definition $\mathcal W(T^*M, W^{-1}(-\infty))$.
According to \cite{gpsstructural}, retracting the stop to its core does not affect the category:
$\mathcal W(T^*M, W^{-1}(-\infty)) = \mathcal W(T^*M, \mathfrak c_{W^{-1}(-\infty)})$.
Thus if the fiber is Weinstein, then we may calculate the corresponding Fukaya--Seidel category using Theorem \ref{sheaffukayaequivalence} (provided the core can be made subanalytic).
In particular, the sheaf-theoretic work on mirror symmetry for toric varieties may now be translated into assertions
regarding the wrapped Fukaya category. Recall that
\cite{FLTZ-Morelli} introduced for any $n$-dimensional toric variety
$\mathbf{T}$ a certain
Lagrangian $\Lambda_\mathbf{T} \subset T^*(S^1)^n$. They conjectured,\footnote{Strictly speaking, they conjectured the proper module version of
this statement.} and \cite{kuwagaki} proved, that
$\operatorname{Sh}_{\partial_\infty\Lambda_\mathbf{T}}((S^1)^n)^c = \operatorname{Coh}(\mathbf{T})$, where we use $\operatorname{Coh}$ to denote the dg category of coherent complexes. By Theorem \ref{sheaffukayaequivalence}, we may conclude:
\begin{corollary}\label{wrappedcoh}
$\operatorname{Perf}\mathcal W(T^*(S^1)^n, \partial_\infty\Lambda_\mathbf{T})^\mathrm{op} = \operatorname{Coh}(\mathbf{T})$
\end{corollary}
We note that this is a much stronger statement than \cite{abouzaid-toric}, in that it includes
the case of non-Fano, non-compact, singular, and stacky $\mathbf{T}$.
When $\mathbf{T}$ is smooth and Fano, it was expected that the $\operatorname{Coh}(\mathbf{T})$ should be
equivalent to the Fukaya--Seidel category of the mirror
Hori--Vafa superpotential \cite{hori-vafa}.
To compare this expectation with Corollary \ref{wrappedcoh}, it
suffices to show that $\partial_\infty \Lambda_\mathbf{T}$ is in fact the core of the fiber of said superpotential in the Fano case. This is
shown under certain hypotheses in \cite{gammage-shende} and in general in \cite{pengzhou}.
For more general $\mathbf{T}$ this equivalence is known to be false, though $\operatorname{Coh}(\mathbf{T})$
conjecturally still embeds into the Fukaya--Seidel category of the Hori--Vafa mirror restricted to an open subset in
$\mathbb C$ containing some of the critical values of the superpotential, see \cite[\S 5]{aurouxkatzarkovorlovprojective} or \cite{abouzaid-toric}.
It may be interesting to explore this conjecture using the present methods.
\subsection{Weinstein hypersurfaces in cosphere bundles}
Let $X \subset S^*M$ be a Weinstein hypersurface (i.e.\ codimension one), and let $X^\epsilon$ be its positive pushoff.
We showed in \cite{gpsstructural} that the covariant pushforward $\mathcal W(X) \to \mathcal W(T^*M, X \sqcup X^\epsilon)$ is
fully faithful. Retracting the stops and appealing to generation by cocores \cite[Theorem 1.9]{gpsstructural}, we see that $\mathcal W(X)$ is equivalent to the full subcategory of
$\mathcal W(T^*M, \mathfrak c_X \sqcup \mathfrak c_X^\epsilon)$ generated by the linking disks to $\mathfrak c_X$.
By Theorem \ref{sheaffukayaequivalence}, we may conclude that $\mathcal W(X)$ is Morita equivalent to
the full subcategory
of $\operatorname{Sh}_{\mathfrak c_X \sqcup \mathfrak c_X^\epsilon}(M)$ generated by the co-representatives of the microstalks at smooth points of $\mathfrak c_X$.
This provides a ``sheaf theoretic prescription'' for $\mathcal W(X)$.
\begin{remark}
We do not know how to tell whether a given Weinstein manifold can be embedded in some cosphere bundle as a hypersurface.
\end{remark}
The above sheaf theoretic prescription for $\mathcal W(X)$ does not yet match the category typically associated by sheaf theorists to $\mathfrak c_X$.
This latter category is defined as follows. One forms the ``Kashiwara--Schapira stack'' by sheafifying the presheaf \emph{of categories}
on $T^*M$ given by the formula
$\mu\operatorname{Sh}^{\mathrm{pre}}(\Omega) := \operatorname{Sh}(M) / \operatorname{Sh}_{T^*M \setminus \Omega}(M)$.
The presheaf $\mu\operatorname{Sh}^{\mathrm{pre}}$ is already discussed in \cite{kashiwara-schapira}; working with its sheafification is a more modern
phenomenon, see e.g.\ \cite{guillermou, nadler-wrapped}.
The notion of microsupport makes sense for a section of this sheaf, and we write $\mu\operatorname{Sh}_\Lambda$ for the subsheaf of full
subcategories of objects with microsupport inside $\Lambda$.
The category typically associated to $\mathfrak c_X$ in the sheaf theory literature is $\mu\operatorname{Sh}_{\mathfrak c_X}(\mathfrak c_X)$.
The remainder of the present discussion depends on the following ``anti-microlocalization lemma'', proved in \cite{nadler-shende}:
\begin{lemma}[\cite{nadler-shende}]\label{antimicrolocalization}
For any compact singular Legendrian $\Lambda\subseteq S^*M$, the natural map $\mu\operatorname{Sh}_\Lambda(\Lambda)^c \to \operatorname{Sh}_{\Lambda\sqcup\Lambda^\epsilon}(X)^c$ is fully faithful, and its image is generated by co-representatives of the microstalks.
\end{lemma}
Sheaf theorists may be more accustomed to the map in the other direction
$\operatorname{Sh}_{\Lambda \sqcup \Lambda^\epsilon}(X) \to \mu \operatorname{Sh}_{\Lambda}(\Lambda)$; the map above is its left adjoint restricted to compact
objects as in \cite{nadler-wrapped}. The fact that the image is generated by microstalks is the usual fact that
sheaves have no microsupport along a smooth Lagrangian iff the microstalk is zero. The content of Lemma \ref{antimicrolocalization} is
in the full faithfulness. Similar results can be found in \cite{guillermou}.
The anti-microlocalization embedding from Lemma \ref{antimicrolocalization} parallels the embedding $\mathcal W(X) \hookrightarrow \mathcal W(T^*M, \mathfrak c_X \sqcup \mathfrak c_X^\epsilon)$ discussed just above.
As the equivalence $\operatorname{Sh}_{\mathfrak c_X \sqcup \mathfrak c_X^\epsilon}(X)^c=\operatorname{Perf}\mathcal W(T^*M, \mathfrak c_X \sqcup \mathfrak c_X^\epsilon)^\mathrm{op}$ from Theorem \ref{sheaffukayaequivalence} identifies microstalks and linking disks, and these generate the images of $\mu\operatorname{Sh}_{\mathfrak c_X}(\mathfrak c_X)^c$ and $\mathcal W(X)$ respectively, we conclude that by restriction Theorem \ref{sheaffukayaequivalence} defines an equivalence $\operatorname{Perf} \mathcal W(X)^\mathrm{op}=\mu\operatorname{Sh}_{\mathfrak c_X}(\mathfrak c_X)^c$.
\begin{example}[Attachment to cotangent bundles]
It is known that if $\Lambda \subseteq S^*M$ is smooth, then $\mu\operatorname{Sh}_\Lambda$ is locally the stack of local systems; hence
$\mu\operatorname{Sh}_\Lambda(\Lambda)$ is a category of twisted local systems, the twist coming from Maslov obstructions
(see e.g.\ \cite{guillermou}).
Of course, we learn the same thing from the identification
$\mu\operatorname{Sh}_\Lambda(\Lambda) = \operatorname{Perf} \mathcal W(T^*\Lambda)^\mathrm{op}$ above. Note that we have to define $\operatorname{Perf} \mathcal W(T^* \Lambda)$ using not the twisting data intrinsic to $T^*\Lambda$ as discussed in \S\ref{gradorsec}, but rather from the data restricted from $T^*M$, hence the appearence of
the Maslov obstructions.
Now suppose we have some other manifold $N$ with some component of $\partial N$ diffeomorphic to $\Lambda$
(and that we choose Maslov data consistent across this identification).
Then we can study the category of $T^*M \#_{T^*\Lambda} T^*N$.
We know that its wrapped Fukaya category is ($\operatorname{Perf}$ of) the pushout of $\mathcal W(T^*N)\leftarrow\mathcal W(T^*\Lambda)\to\mathcal W(T^*M,\Lambda)$ by \cite[Theorem 1.20]{gpsstructural}, and the above discussion allows us to describe these functors on the sheaf side.
Thus we may compute $\mathcal W(T^*M \#_{T^*\Lambda} T^*N)$ using sheaves. We can also allow stops in
$T^*M$ and $T^*N$ disjoint from $\Lambda$, and we can also treat the case where $\Lambda$ is embedded into $S^*N$ instead of being identified with one of its boundary components.
The simplest case is when $N$ is just a disjoint union of disks, so we are attaching handles. For an example of sheaf calculations
in this setting, see \cite{shende-treumann-williams}.
\end{example}
\begin{example}[Mirror symmetry for very affine hypersurfaces]\label{veryaffine}
As in Corollary \ref{wrappedcoh}, let $\Lambda_\mathbf{T}$ be the Lagrangian of \cite{FLTZ-Morelli}.
Consider $\operatorname{Sh}_{\partial_\infty \Lambda_\mathbf{T}}(T^* (S^1)^n) \to \mu\operatorname{Sh}_{\partial_\infty \Lambda_\mathbf{T}}(\partial_\infty \Lambda_\mathbf{T})$. As mentioned above, in
\cite{kuwagaki} the source category was computed, and shown to be equivalent to the category of coherent sheaves on an appropriate
toric variety. In \cite{gammage-shende} the map and the target were computed, and identified with the restriction of coherent sheaves to the toric boundary.
As we have already mentioned, $\partial_\infty \Lambda_\mathbf{T}$ was also shown there (under a hypothesis later removed in \cite{pengzhou})
to be the core of the fiber of the Hori--Vafa mirror $W_\mathbf{T}: (\mathbb C^*)^n \to \mathbb C$.
(\cite{nadler-wrapped} had previously treated the case of pants, i.e.\ when $W$ is the sum of the coordinates.)
Thus we may translate the sheaf theoretic results of \cite{gammage-shende} to the following:
\begin{equation}
\begin{tikzcd}
\operatorname{Perf} \mathcal W(W_\mathbf{T}^{-1}(-\infty)) \ar{r}{\sim} \ar{d} & \operatorname{Coh}(\partial \mathbf{T}) \ar{d}\\
\operatorname{Perf} \mathcal W((\mathbb C^*)^n, W_\mathbf{T}^{-1}(-\infty)) \ar{r}{\sim} & \operatorname{Coh}(\mathbf{T})
\end{tikzcd}
\end{equation}
\end{example}
|
\section{Introduction}
High efficiency video coding (HEVC) \cite{sullivan2012overview} is the state-of-the-art video coding standards developed by the joint collaborative team on video coding (JCT-VC).
HEVC supports four depth levels of coding units (CUs) at maximum, from $64\times64$ to $8\times8$, also known as CU quad-tree partition.
By contrast with the fixed-size marcoblock supported in H.264 \cite{wiegand2003overview}, this flexible partition structure leads that at the meantime of performance enhancement, CU partition decision becomes one of the most time-consuming modules in HEVC encoder.
Currently, CU partition decision relies on a rate-distortion optimization (RDO) quad-tree search algorithm.
Specifically HEVC encoder recursively in three depths, makes decision whether a CU split or not by comparing if the rate-distortion cost of the CU as a whole is higher than the sum of rate-distortion costs of four splited sub-CUs.
In the light of this condition, so much encoding time is spent on CU partition decision that a faster approach of it is urgently required.
In the last few years, deep learning has successfully demonstrated its impressive performance on many varieties of fields.
Convolutional neural networks (CNNs) specifically are able to extract features from images and then classify them to several types.
Accordingly, an idea comes up that we can take advantages of CNNs to assist many decision modules in HEVC encoder.
CNNs are able to boost the decision modules in HEVC encoding, such as CU partition, intra and inter prediction mode.
For instance, CNNs can assist HEVC encoder to decide CU partition much faster than traditional method.
At the meantime, since CNNs are able to achieve considerable accuracy of CU partition decision, RD performance of encoded video will hardly be aggravated.
Several CNN-based fast CU partition decision algorithms are emerging.
Liu \textit{et al.} \cite{liu2016cnn} proposed a simple CNN structure to predict CU mode.
Liu's work mainly focuses on pipe-line to adapt hardware implementation, which is tough to be ported to another hardware or platform.
Moreover, Li \textit{et al.} \cite{li2017deep} presented a deep CNN to decide three level CU partition, and in their CNN there are three branches with different sizes and quantities of convolution layers.
On the other hand, none of feasible measure is taken during training process against the imbalance training samples issue.
Xu \textit{et al.} \cite{xu2017reducing} raised a novel CNN structure to decide CU quad-tree partition over three depths at one time.
Within a $64\times64$ CU, Xu's proposed CNNs utilize redundant neighbor sub-CUs to sub-CU partition decision in second and third depth.
In this paper, we propose a novel two-stage approach in order to accelerate CU partition decision.
The first stage of the proposed approach is a rough early-termination algorithm based on the range of CU luma samples that directly related to CU partition mode.
This early-termination algorithm will be applied before CNN-based stage to filter some evident non-split CUs out of the subsequent CNN in order to tightly restrain computational complexity.
The second stage is CNNs designed and trained to precisely estimate the partition mode of ambiguous CUs that the first stage conveys.
Due to the depth and width of our neural networks, our CNNs are able to accomplish much lower error rate.
In a nutshell, the first stage, or the early-termination algorithm, is a rough but fast method to filter some evident homogeneous CUs, which plays an important part in complexity reduction.
Then in the second stage the CNN-based CU partition algorithm is a final method with higher accuracy, which mainly concentrates on RD performance retention.
As a result, the proposed two-stage approach gains the balance of time saving and RD loss.
\section{Two-Stage CU Partition Approach}
CU partition mode is decided over three depths totally and CUs are determined to split or to non-split at each depth.
In other word, CU partition decision at each depth is actually a binary classification problem, so we design a CU non-split/split classifier.
The proposed approach measures range of CU luma samples firstly to filter homogeneous CUs to early terminate.
Then CUs passed early-termination filter will be decided by three CNNs whether to split or not in each depth.
\subsection{Early-Termination Algorithm}
Aside from CNNs, we craft an early-termination algorithm to improve efficiency of whole proposed approach further.
Apparently, homogeneous CUs have tendency to split into four sub-CUs and vice versa.
Considering that, we do a statistical analysis on every CU luma samples in our training sequence.
Given feasibility and computational complexity, we choose two most possible CU classification indicators, range and standard deviation of CU luma samples.
Every range and standard deviation samples of our training dataset are separated into two groups i.e. ones of non-split CUs and ones of split CUs.
Then, both selected indicators are assessed by probability density.
\begin{figure}[!h]
\centering
\subfigure[Range of $64\times64$ CUs]{
\begin{minipage}{40mm}
\includegraphics[width=42mm]{cu64_range.pdf}
\label{fig:indicators:cu64_range}
\end{minipage}
}
\subfigure[Std. dev. of $64\times64$ CUs]{
\begin{minipage}{40mm}
\includegraphics[width=42mm]{cu64_stddev.pdf}
\label{fig:indicators:cu64_stddev}
\end{minipage}
}
\subfigure[Range of $32\times32$ CUs]{
\begin{minipage}{40mm}
\includegraphics[width=42mm]{cu32_range.pdf}
\label{fig:indicators:cu32_range}
\end{minipage}
}
\subfigure[Std. dev. of $32\times32$ CUs]{
\begin{minipage}{40mm}
\includegraphics[width=42mm]{cu32_stddev.pdf}
\label{fig:indicators:cu32_stddev}
\end{minipage}
}
\subfigure[Range of $16\times16$ CUs]{
\begin{minipage}{40mm}
\includegraphics[width=42mm]{cu16_range.pdf}
\label{fig:indicators:cu16_range}
\end{minipage}
}
\subfigure[Std. dev. of $16\times16$ CUs]{
\begin{minipage}{40mm}
\includegraphics[width=42mm]{cu16_stddev.pdf}
\label{fig:indicators:cu16_stddev}
\end{minipage}
}
\caption{Probability density of both selected indicators}
\label{fig:indicators}
\end{figure}
As Fig. \ref{fig:indicators} shows, the possibility densities of two selected indicators, range and standard deviation of CU luma samples, are shown respectively in each graph within non-split and split groups of three CU sizes.
As we can see, split CUs basically have larger range and standard deviation than non-split CUs, which results in an truth that CUs can be classified into non-split and split by their range and standard deviation.
However, the differential of range is obviously larger than standard deviation, as shown in Fig. \ref{fig:indicators}, which manifests that the range is more effective to distinguish non-split/split CUs.
Another advantage of range is that it is much easier and faster to calculate range than standard deviation.
Consequently, we regard the range of CU luma samples as the indicator to classify CU partition.
More clearly, the threshold are configured as 20 in $64\times64$ CUs, 25 in $32\times32$ CUs and 30 in $16\times16$ CUs.
CUs whose range of luma samples below the threshold will be early terminated in this stage as shown in grey areas in Fig. \ref{fig:indicators}.
\subsection{CNN-based CU Partition Algorithm}
In pursuit of higher accuracy than current CNN-based CU partition approaches, a deeper and wider CNN structure is introduced and crafted for CU partition decision in HEVC.
Nevertheless, upon the trade-off of accuracy and computation complexity, the computational complexity of CNNs has to be controlled in an acceptable limitation.
Hence we cannot only seek the highest accuracy by sacrifice of computational performance and the proposed CNNs are illustrated below.
$64\times64$, $32\times32$ and $16\times16$ CU will be determined to non-split or split during encoding process.
Thus, we design three large-scale CNNs to decide CU partition mode at each depth from 0 ($64\times64$) to 3 ($16\times16$).
In each CNN, the input layer is two-dimensional normalized luma samples of the CU, and the output layer is two softmax activation units representing the probabilities of non-split and split.
Our large-scale CNN structures are enlightened by GoogLeNet \cite{szegedy2015going} comprising several Inception layers.
\begin{figure*}[ht]
\centering
\subfigure[CNN for $64\times64$ CUs]{
\includegraphics[width=175mm]{cnn64.pdf}
\label{fig:cnn:cu64}
}
\subfigure[CNN for $32\times32$ CUs]{
\includegraphics[width=175mm]{cnn32.pdf}
\label{fig:cnn:cu32}
}
\subfigure[CNN for $16\times16$ CUs]{
\includegraphics[width=175mm]{cnn16.pdf}
\label{fig:cnn:cu16}
}
\caption{Proposed CNN structures \protect\footnotemark[1]}
\label{fig:cnn}
\end{figure*}
Particularly, frameworks and parameters of the proposed CNNs for each size of CU are elaborated in Fig. \ref{fig:cnn}.
At very beginning of each CNN, there is a $3\times3$ convolutional layer to increase input channel from 1 to 32.
Then, core framework of each CNN consists of four Inception layers for features extraction and many max pooling layer for dimension reduction.
In core framework, only sequence arrangement of four Inception layers and pooling layers is different from each other CNN, and four Inception layers in each CNN have same numbers of channels.
Specifically, max poling layers are followed every Inception layer for $64\times64$ CUs, the first and third Inception layer for $32\times32$ CUs, and the second Inception layer for $16\times16$ CUs.
For each CNN, the output of core framework is a $8\times8\times256$ feature map, and there is a $8\times8$ valid average pooling layer to flat that into a one-dimensional array.
Final layer is a fully connected layer and a softmax activation.
\subsection{Self-Adaptive Weighted Loss Function}
Our training dataset is extracted from original HEVC encoder, consisting of the luma samples (as feature) and the split flag (as label) of every CU in the training sequence.
To prevent overlap the test sequences, the training sequence is concatenation of 1,000 raw images that are arbitrarily selected from RAISE database \cite{dang2015raise}.
It should be noticed that as a matter of fact larger CUs tend to split, so our training dataset actually is a classification-imbalance dataset in which the ratio of positive and negative samples is far larger or smaller than 1 as shown in Table \ref{tab:dataset}.
It undoubtedly results in a severe problem in the training process.
That is under any circumstances CNNs, without significantly increasing loss, will always choose the partition mode of the majority of training samples.
\footnotetext[1]{Each block indicated a layer of CNN in the form of ``$Layer Type\ [@Channel] \ KernelWidth \times KernelHeight+Stride(Padding(S/V))$"}
\renewcommand\arraystretch{1.2}
\begin{table}[!h]
\centering
\caption{Partition Distribution of CUs in Training Dataset}
\label{tab:dataset}
\begin{small}
\begin{tabular}{|p{16mm}<{\centering}
|p{15mm}<{\centering}|p{15mm}<{\centering}|p{15mm}<{\centering}|}
\hline
{\bfseries CU Size} & $64\times64$ & $32\times32$ & $16\times16$ \\
\hline
{\bfseries Non-Split} & 508,105 & 3,828,038 & 29,467,909 \\
\hline
{\bfseries Split} & 1,667,895 & 4,875,962 & 5,348,091 \\
\hline
{\bfseries NS/S Ratio} & 0.30 & 0.78 & 5.51 \\
\hline
{\bfseries Total} & 2,176,000 & 8,704,000 & 34,816,000 \\
\hline
\end{tabular}
\end{small}
\end{table}
Against this classification-imbalance problem, we propose the self-adaptive weighted loss function.
Based on the traditional cross-entropy loss function, our self-adaptive weighted loss function is given by
\begin{equation}
\label{eq:loss}
{L = - \frac{1}{N}\sum\limits_i {\alpha_1 y_i\ln {{\hat y}_i} + \alpha_0 (1 - y_i)} \ln (1 - {\hat y_i})}
\end{equation}
where $\alpha_0$ and $\alpha_1$ is weight of 0 (non-split) samples' loss and 1 (split) samples' loss respectively and $N$ is batch size.
Parameter $\alpha_0$ and $\alpha_1$ will be adjusted during training process to adapt to the training results.
In detail, for every $M$ steps, the next $\alpha_1 / \alpha_0$ will be set to equal the weighted sum of current $\alpha_1 / \alpha_0$ of proportion of amounts of 0 samples and 1 samples in test results of present CNN.
When updating $\alpha_0$ and $\alpha_1$, keep $\alpha_0 + \alpha_1 = 2$ in order to control loss in a reasonable range.
Hence the update method of loss weights is given by
\begin{equation}
\label{eq:loss_update}
{(\frac {\alpha_1} {\alpha_0})_{t + 1} = (1 - \eta )(\frac {\alpha_1} {\alpha_0})_t + \eta \frac{N_0}{N_1}}
\end{equation}
where $N_0$, $N_1$ is the amounts of 0 samples and 1 samples in test results, and $\eta$ is a hyperparameter representing the weight of the proportion of two samples to update loss weights.
Since loss of minority of predictions will be multiplied a relatively large penalty item, the estimations of CNN will not remain the majority.
Note that $\alpha_0$ and $\alpha_1$ are a self-adaptive auxiliary parameters instead of hyperparameters.
\section{Experimental Results}
In our experiment, CPU is Intel Xeon CPU E5-2620 v4, GPU is NVIDIA GeForce GTX 1080 and OS is Ubuntu 16.04.
Our CNNs are implemented with TensorFlow \cite{abadi2016tensorflow} deep learning library
The proposed approach is integrated with HEVC Test Model (HM) 16.18, and we use the test sequences with QP = \{22, 27, 32, 37\} under all intra (AI) configuration.
\renewcommand\arraystretch{1.2}
\begin{table*}[ht]
\centering
\caption{BD-Bitrate(\%), BD-PSNR(dB) and Time Saving(\%) of the Proposed and Relative Approach}
\label{tab:results}
\begin{small}
\begin{tabular}{|p{24mm}<{\centering}|p{18mm}<{\centering}
|p{10mm}<{\centering}|p{10mm}<{\centering}|p{10mm}<{\centering}
|p{10mm}<{\centering}|p{10mm}<{\centering}|p{10mm}<{\centering}
|p{10mm}<{\centering}|p{10mm}<{\centering}|p{10mm}<{\centering}|}
\hline
\multirow{3}*{\bfseries Sequences} &
\multirow{3}*{\bfseries Resolution} &
\multicolumn{3}{c|}{\bfseries Liu \textit{et al.} \cite{liu2016cnn}} &
\multicolumn{3}{c|}{\bfseries Cen \textit{et al.} \cite{cen2015fast}} &
\multicolumn{3}{c|}{\bfseries Proposed} \\
\cline{3-11}
& & {\bfseries BD-} & {\bfseries BD-} & \multirow{2}*{\bfseries TS} &
{\bfseries BD-} & {\bfseries BD-} & \multirow{2}*{\bfseries TS} &
{\bfseries BD-} & {\bfseries BD-} & \multirow{2}*{\bfseries TS} \\
& & {\bfseries BR} & {\bfseries PSNR} & &
{\bfseries BR} & {\bfseries PSNR} & &
{\bfseries BR} & {\bfseries PSNR} & \\
\hline
BQTerrace & $1920\times1080$ & 4.21 & -0.10 & 36.71 &
{\bfseries 2.26} & {\bfseries -0.07} & 13.14 &
2.42 & -0.14 & {\bfseries 43.30} \\
\hline
BasketballDrill & $832\times480$ & 5.67 & -0.21 & 34.75 &
7.33 & {\bfseries -0.10} & 18.35 &
{\bfseries 2.36} & {\bfseries -0.10} & {\bfseries 37.71} \\
\hline
PartyScene & $832\times480$ & 4.34 & -0.12 & {\bfseries 34.83} &
4.00 & -0.18 & 11.45 &
{\bfseries 1.39} & {\bfseries -0.10} & 31.69 \\
\hline
RaceHorses & $416\times240$ & 5.23 & -0.23 & {\bfseries 32.70} &
{\bfseries 0.01} & -0.14 & 12.75 &
1.99 & {\bfseries -0.13} & 27.03 \\
\hline
BasketballDrillText & $832\times480$ & 5.96 & -0.26 & {\bfseries 37.66} &
{\bfseries 0.09} & -1.82 & 18.90 &
4.42 & {\bfseries -0.22} & 35.52 \\
\hline
ChinaSpeed & $1024\times768$ & 4.66 & -0.39 & 34.15 &
{\bfseries 1.14} & {\bfseries -0.11} & 20.02 &
4.11 & -0.38 & {\bfseries 43.05} \\
\hline
SlideEditing & $1280\times720$ & 3.52 & -0.50 & 31.69 &
2.42 & {\bfseries -0.13} & 13.70 &
{\bfseries 2.18} & -0.34 & {\bfseries 38.84} \\
\hline
\multicolumn{2}{|c|}{\bfseries Average} & 4.79 & -0.25 & 34.64 &
2.85 & -0.36 & 15.47 &
{\bfseries 2.69} & {\bfseries -0.22} & {\bfseries 36.74} \\
\hline
\end{tabular}
\end{small}
\end{table*}
We compare our proposed approach with other relative approaches by Liu \textit{et al.} \cite{liu2016cnn} and Cen \textit{et al.} \cite{cen2015fast} on our aforementioned experimental environment.
Cen's approach initially calculates the rarest depth (0 or 3) which will be skipped in subsequent algorithm, and then CU depth is determined from the remaining depths by the comparison algorithm with adjacent CUs.
The experimental results are shown in Table \ref{tab:results}, including the Bjontegaard delta bitrate(BD-BR), Bjontegaard delta peak signal-to-noise(BD-PSNR) \cite{bjontegaard2001calculation} and the encoding time saving (TS) of three approaches with original HM encoder as anchor.
And the time saving is given by
\begin{equation}
\label{eq:time_saving}
{TS = \frac {T_{orig} - T_{prop}} {T_{orig}} \times 100\%}
\end{equation}
where $T_{orig}$ and $T_{prop}$ is the encoding time of original and proposed encoder respectively.
Firstly, the proposed approach reduce 36.74\% computational complexity on average, and specifically at maximum 43.40\% in ``\textit{BQTerrace}".
As the brute-force RDO quad-tree search in original HM is skipped, the proposed approach is able to reduce such computational complexity.
Because the calculation of range of CU luma samples is fast and easy, our novel early-termination algorithm also assist to mitigate compuational complexity further.
Compared to the approach by Cen \textit{et al.} \cite{cen2015fast}, their approach reduces only 15.47\% encoding time and the proposed approach saves much more encoding time.
In addition, the proposed approach reach a better RD performance with 2.69\% BD-BR increment and 0.22dB BD-PSNR decrease.
As our two-stage approach has considerable accuracy on CU partition mode prediction, so that RD performance can not be aggravated.
The approach by Cen \textit{et al.} \cite{cen2015fast} has 2.85\% BD-BR increment which is higher than our approach.
Moreover, the shallow and narrow CNN-based approach by Liu \textit{et al.} \cite{liu2016cnn} on dedicated hardware incurs 4.79\% BD-BR rise.
However, our proposed approach achieves lower 2.69\% BD-BR rise.
Thus our large-scale CNN-based approach has stronger ability to extract and generalize CU features, but with relatively more computation burden on general hardware.
And the simple early-termination algorithm based on the range of CU luma samples mitigates this issue without accuracy loss.
Higher accuracy of CU partition estimation of the proposed two-stage approach makes great contribution to lower BD-BR rise.
Therefore, experimental results indicate that the proposed approach on CU partition acceleration outperforms conventional method in HM 16.18 with 36.74\% encoding time saving and negligible BD-BR increment.
In terms of both RD performance and encoding time saving the proposed approach is better than other approaches in \cite{liu2016cnn} and \cite{cen2015fast}.
\section{Conclusion}
In this paper, we proposed a novel CU partition decision approach in HEVC, which has been proved to have ability to reduce encoding time without significantly impairing RD performance.
As presented above, the proposed approach has two stages in which range of CU luma samples and three CNNs targeting each CU size are utilized to decide a CU partition mode.
Relieving computational complexity of CU partition decision, the proposed approach reduces nearly 37\% encoding time with ignorable 2.69\% BD-BR rise in comparison with the original HM 16.18 encoder.
\enlargethispage{0mm}
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{s1}
It is well-known that massive quiescent galaxies are more predominant in the centres
of galaxy clusters relative to the general fields in the present day. This trend is
often characterised by colours
\citep{Vaucouleurs:1961,Visvanathan:1977,Butcher:1984,Bower:1992,Bower:1998,Terlevich:2001,Tanaka:2005,Kodama:2007,Mei:2009,Bamford:2009,Peng:2010,Muzzin:2012,Wetzel:2012,Darvish:2016}
and morphological types
\citep{Dressler:1980,Dressler:1997,Couch:1998,Goto:2003,Kauffmann:2004,Wel:2008,Cappellari:2011,Houghton:2013,Fogarty:2014,Brough:2017,Lopes:2017}.
Over ten billion years ago, the most massive structures in the Universe -- galaxy
protoclusters -- played a prominent role in the star formation and mass assembly
of massive galaxies \citep{Chiang:2017}. Massive protoclusters\footnotemark[1]
\citep{Albada:1961,Peebles:1970a,Sunyaev:1972} at redshift $z\sim$ 2--3 are ideal
test-beds to probe this rapid transition, and thus develop our understanding of which
physical phenomena have driven such early and/or fast growth in centres of distant
galaxy clusters
\citep{Steidel:2005,Doherty:2010,Tanaka:2010b,Hatch:2011,Gobat:2011,Koyama:2013b,Tanaka:2013,Kubo:2013,Alexander:2016,Kubo:2017,Shimakawa:2018}.
\footnotetext[1]{Various survey bias and restrictions result in vague and inconsistent
definitions of the {\it protocluster} in any work. This series of papers refers to
overdense fields on the scale of $\gtrsim10$, $\sim$1--10, and $\lesssim1$ comoving
Mpc as large-scale structures, protoclusters, and dense cores (groups) for the target,
respectively.}
The rapid change of star formation rate (SFR) density in the centres of clusters seems
to follow $(1+z)^{6\sim8}$ from $z\sim2$ to now
\citep{Kodama:2001,Clements:2014,Smail:2014,Shimakawa:2014,Kato:2016}.
Such a drastic variation is not only due to the increase in the number of quenched
galaxies in clusters at lower redshifts
\citep{Blanton:2009,Muzzin:2012,Wetzel:2012,Burg:2013,Darvish:2016,Paulino:2018}
but also due to very active star formation in high-$z$ protoclusters
\citep{Dannerbauer:2014,Umehata:2015,Tadaki:2015,Wang:2016,Oteo:2017}.
These populations are complicated to reproduce with the classical semi-analytic models
\citep{Romeo:2015}. Moreover, past studies have reported protoclusters which host
large numbers of active galactic nuclei (AGNs) (e.g.,
\citealt{Lehmer:2009,Lehmer:2013,Hennawi:2015,Cai:2017b,Krishnan:2017}, but see
\citealt{Macuga:2018}) including radio-loud sources
\citep{Pentericci:2002,Rottgering:2003,Venemans:2007,Hatch:2014}. A few studies have
investigated the energy injection from central AGNs into the ambient gas surrounding
high-$z$ (proto-) clusters \citep{Nesvadba:2006,Valentino:2016}. There is no good
understanding of how large an impact AGNs have on the proto-intercluster medium of
protocluster members. This uncertainty makes it even more difficult to understand the
mechanism behind the difference in star formation histories in and outside cluster centres.
It is, therefore, important to characterise massive galaxies in protoclusters. Our
MAHALO-Subaru (Mapping H-Alpha and Lines of Oxygen with Subaru; \citealt{Kodama:2013})
surveys have extensively studied star formation in high-$z$ clusters and
protoclusters. High-density sampling of line emitters at limited redshift ranges
($\pm2000$ km~s$^{-1}$) with narrow-band filters have found the inside-out propagation
of star formation and mapped bottom-up structure growth based on the spatial
distributions of emission line galaxies and their physical properties. We have
identified that the regions dominated by bright line emitters are shifted from the
densest cluster cores to lower-density outskirts and filamentary outer structures, on
timescales from $z\sim3$ to present (e.g.,
\citealt{Hayashi:2010,Koyama:2010,Koyama:2011,Tadaki:2012,Hayashi:2012,Koyama:2013a}).
Recent deep follow-up {\rm\,H$\alpha$}\ imaging towards a young protocluster, USS~1558$-$003 at
$z=2.53$, finds enhanced star formation and concentration of massive {\rm\,H$\alpha$}\ emitters
(HAEs) in fragmented group cores \citep{Shimakawa:2018}.
Furthermore, a follow-up sub-mm/radio campaign with ALMA has shown gas-depleted
massive galaxies in the very centre of an X-ray cluster, XMMXCS~J2215.9$-$1738 at
$z=1.46$ \citep{Hayashi:2017}. Their typical gas fraction is no more than 10 per cent
as opposed to gas-rich sources in the outer regions with gas fractions of $\gtrsim50$
percent (\citealt{Hayashi:2018}, see also \citealt{Noble:2017}). Such a sharp contrast
in time and radial distribution would require e.g., a strong quenching mechanism like
AGN feedback
\citep{Springel:2005b,Sijacki:2007,Fabjan:2010,McCarthy:2010,Barnes:2017}, and/or
rapid gas consumption via starbursts
\citep{Hopkins:2009,Hayward:2011,Hopkins:2013,Narayanan:2015}.
Here, in the second part of our MAHALO-Deep cluster survey (MDCS), we investigate the
properties of massive galaxies in a protocluster associated with a radio galaxy,
PKS~1138$-$262 at $z=2.16$. This protocluster is known to have an apparent red
sequence \citep{Kurk:2004a,Kodama:2007,Tanaka:2013}; at the same time, there is a
strong excess of red {\rm\,H$\alpha$}-emitting galaxies \citep{Koyama:2013a}. \citet{Koyama:2013a}
also found that higher fractions of redder and more massive HAEs in higher-density
regions than in under dense regions in the protocluster, implying that the build-up of
stellar mass has mostly completed for massive galaxies in the densest parts of the
protocluster at this time (see also \citealt{Doherty:2010,Hatch:2011,Tanaka:2013}).
These unique trends suggest that the protocluster is in a critical transition phase
from young, fragmented, protoclusters, to the classical X-ray clusters at
$z\lesssim2$. The primary goal of this paper is to determine the stellar mass function
of protocluster members and then quantify passive fraction and AGN fraction as a
function of stellar mass. Also, based on multi-wavelength datasets from literature, we
investigate local number densities, rest-frame colours and SFRs for individual HAEs
and check if properties of HAEs are different from the field. These will enable us to
investigate how galaxies in the protocluster stop forming stars.
We assume the cosmological parameters of $\Omega_M=0.3$, $\Omega_\Lambda=0.7$ and
$h=0.7$ and adopt a \citet{Chabrier:2003} stellar initial mass function. The AB
magnitude system \citep{Oke:1983} is employed throughout the Paper.
\section{Target and dataset}\label{s2}
\subsection{PKS~1138$-$262}\label{s2.1}
This paper focuses on a dense protocluster associated with a radio galaxy,
PKS~1138$-$262 (or MRC~1138$-$262, $\alpha_\mathrm{J2000}=$
11$^\mathrm{h}$40$^\mathrm{m}$48$^\mathrm{s}$, $\delta_\mathrm{J2000}=$
$-$26$^\mathrm{d}$29$^\mathrm{m}$09$^\mathrm{s}$,
\citealt{Bolton:1979,Roettgering:1994,Roettgering:1997,Carilli:1997}) at $z=2.156$
known as the Spiderweb galaxy \citep{Pentericci:1998,Miley:2006}. The PKS~1138
protocluster (hereafter PKS~1138) was first explored by \citet{Pentericci:1997},
\citet{Kurk:2000}, and \citet{Pentericci:2000}. PKS~1138, together with the SSA22
protocluster at $z=3.09$ \citep{Steidel:1998,Steidel:2000}, has been extensively
studied over a long period. The following is a short summary of previous findings on
PKS~1138 over the past two decades.
Since the X-ray gas density of nearby galaxy clusters is correlated with the large
rotation measures \citep{Taylor:1994}, \citet{Carilli:1997} and
\citet{Pentericci:1997} have suggested that the Spiderweb galaxy resides in a dense
cluster environment given its observed very high rotation measure of the polarized
radio emission (6200 rad~m$^{-2}$; see also \citealt{Athreya:1998}).
\citet{Kurk:2000} and their series of papers
\citep{Pentericci:2000,Kurk:2003,Kurk:2004a,Kurk:2004b,Croft:2005} identified a
significant overdensity in this field based on imaging and follow-up spectroscopic
searches towards {\rm\,Ly$\alpha$}\ and {\rm\,H$\alpha$}\ emitters (LAEs and HAEs, respectively) and distant
red galaxies (DRGs). Their narrow-band imaging surveys succeeded in selecting 50 LAE
and 40 HAE candidates, and then spectroscopically confirmed 14 and 9 objects
respectively. They also found high concentrations of HAEs and DRGs within 0.5 Mpc
of the Spiderweb galaxy, which are 4--5 times greater than those outside
the central region. Such massive overdensities have been subsequently confirmed on
higher dynamic-scales \citep{Koyama:2013a,Shimakawa:2014} and by comparing with other
radio galaxy environments \citep{Venemans:2007,Mayo:2012,Galametz:2012}. Spectroscopic
observations tentatively suggested that the protocluster centre of PKS~1138 may have
halo mass $\sim10^{14}$ {\rm M}$_{\odot}$\ and virial radius of 0.5 Mpc
\citep{Pentericci:2000,Kuiper:2011,Shimakawa:2014} assuming that the system is
collapsed (but see \citealt{Kuiper:2011}). Such a massive overdensity has the
potential to grow into a massive, Coma-like, galaxy cluster by the present-day
\citep{Chiang:2013,Lovell:2018}.
\subsection{Data}\label{s2.2}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig1.pdf}
\caption{System throughputs of $NB_{2071}$ and $K_s$ filters with MOIRCS, represented
by the black solid and dotted lines, respectively. The red histogram shows spec-$z$
distribution of 29 HAEs, 23 of which have been confirmed by \citet{Shimakawa:2014};
spectroscopic redshifts for the remainder are taken from the literature
\citep{Pentericci:2002,Kurk:2004b,Croft:2005,Doherty:2010,Tanaka:2013}.
One should note that a strong dip at $z\sim2.16$ is caused by the strong OH lines at
$\lambda=2.0729$ $\mu$m preventing us from spectroscopically identifying the {\rm\,H$\alpha$}\ line
of HAEs at this redshift \citep{Shimakawa:2014}.}
\label{fig1}
\end{figure}
We employ the multi broad-band and narrow-band dataset from MDCS and the literature.
The data consist of $B$, $F475W$, $F814W$, $z'$, $Y$, $J$, $H$,
$K_{s,\mathrm{MOIRCS}}$ (hereafter $K_s$), $K_{s,\mathrm{HAWKI}}$, and $NB_{2071}$.
Table~\ref{tab1} summarises the seeing FWHM and limiting magnitudes for these images.
The $z'$, $J$, $K_s$, and $NB_{2071}$ images are based on the past MAHALO-Subaru
campaign (S10B-028I, Kodama et al.; \citealt{Koyama:2013a}) and MDCS (S15A-047, Kodama
et al.). The reduced $B$-band image is provided by Koyama et al. (in preparation),
and was recently obtained with Suprime-Cam on the Subaru Telescope between May and
June 2017. The narrow-band filter, $NB_{2071}$ has a central wavelength of 2.071
$\mu$m with the full width at half-maximum (FWHM) of 270 \AA, which covers the
{\rm\,H$\alpha$}-redshift $2.15\pm0.02$ (fig.~\ref{fig1}).
In addition, we use the reduced Hubble Space Telescope (HST) ACS/WFC data ($F475W$,
$F814W$), obtained from the Hubble Legacy Archive (HLA), and reduced near-infrared
(NIR) images ($Y$, $H$, $K_{s,\mathrm{HAWKI}}$) taken with HAWK-I on Very Large
Telescope (VLT). These original data have been reported in detail by
\citet{Miley:2006} and \citet{Dannerbauer:2017}, respectively. Moreover, this work
employs 3.6 and 4.5 $\mu$m IRAC bands \citep{Seymour:2007}. We use the Post-BCD (PBCD)
products from the Spitzer data archive library. Each IRAC band covers 89 percent of
the entire narrow-band emitters. The IRAC images are shallow (21.4--21.6 in $3\sigma$
limiting magnitude), and we confirmed that has a negligible effect on the measurement
of physical properties with the SED fitting (\S\ref{s2.4.1}). However, we solely use
the photometry in these bands to impose restrictions on the rest-frame NIR spectra of
the targeting HAEs at $z=2.2$; this is crucial when constraining the rest-frame $J$
band magnitudes (\S\ref{s3.2}).
We also introduce here $NB_{2071}$ data, taken as part of the MDCS with the
Multi-Object Infrared Camera and Spectrograph (MOIRCS;
\citealt{Ichikawa:2006,Suzuki:2008}) on the Subaru Telescope (the same instrument that
was used in the past MAHALO-Subaru survey; \citealt{Koyama:2013a}). The observations
were executed between April 30 and May 6, 2015, under photometric conditions with
seeing FWHM $\sim0.6$ arcsec. The integration time is 125 min which was split into 180
sec individual exposures. After combining with the existing $NB_{2071}$ data (186 min
integration), we reconstructed all the data using the reduction pipeline
{\sc mcsred}\footnotemark[2] \citep{Tanaka:2011}, which is written as
{\sc iraf}\footnotemark[3] scripts \citep{Tody:1993}. As described in
\citet{Shimakawa:2018}, we executed flat fielding, masking objects from the combined
data in the first run (thus the whole reduction process was conducted twice to remake
secure object masks), sky subtraction (by median sky and then the polynomially-fitted
plane for residual sky subtraction), distortion correction, cross-matching, and image
mosaicing with this pipeline. The reconstructed $NB_{2071}$ image reaches 23.95 mag in
$3\sigma$ limiting magnitude using a 1.4 arcsec diameter aperture, and its seeing FWHM
is 0.63 arcsec. The image depth becomes deeper by 0.5 mag than the previous data
\citep{Koyama:2013a}. The world coordinate system (WCS,
\citealt{Calabretta:2002,Greisen:2002}) of the narrow-band image is carefully matched
by the {\sc iraf} scripts ({\sc ccmap} and {\sc ccsetwcs}) to that of the F814W image,
based on 67 point sources. F814W has one of the best spatial resolutions amongst our
dataset. The standard deviation of point source separations between the NB$_{2071}$
and F814 images suggests that the relative WCS uncertainty would be around 0.04 arcsec
in the survey area. One should note, however, that the absolute astrometry would have
0.3 arcsec errors in right ascension and declination based on comparison with the
Guide Star Catalogue 2 \citep{Lasker:2008}.
\footnotetext[2]{\url{http://www.naoj.org/staff/ichi/MCSRED/mcsred.html}}
\footnotetext[3]{\url{http://iraf.noao.edu}}
\begin{table}
\centering
\caption{Data summary. The first to fourth columns indicate filter name,
instrument/telescope, seeing FWHM, $3\sigma$ limiting magnitude in 1.4 arcsec diameter
aperture including galactic extinction correction, respectively. The fifth column
shows the galactic extinction based on the NASA Extragalactic Database extinction law
calculator \citep{Schlegel:1998,Fitzpatrick:1999}$^a$. We employ recalibrated
estimates from \citet{Schlafly:2011}.
}
\begin{tabular}{lcccc}
\hline
Filter & Instrument & FWHM & 3$\sigma$ & A$\mathrm{_\lambda}$ \\
& /Telescope & (arcsec) & (AB) & (mag) \\
\hline
$NB_{2071}$ & MOIRCS/Subaru & 0.63 & 23.95 & 0.01 \\
$K_s$ & MOIRCS/Subaru & 0.63 & 23.99 & 0.01 \\
\hline
$B$ & S-Cam/Subaru & 1.15 & 26.56$^c$ & 0.14 \\
$F814W$ & ACS/HST & 0.11 & 26.33$^b$ & 0.06 \\
$F475W$ & ACS/HST & 0.11 & 27.02$^b$ & 0.13 \\
$z'$ & S-Cam/Subaru & 0.70 & 26.35 & 0.05 \\
$Y$ & HAWK-I/VLT & 0.37 & 26.08$^b$ & 0.04 \\
$J$ & MOIRCS/Subaru & 0.69 & 24.33 & 0.03 \\
$H$ & HAWK-I/VLT & 0.49 & 25.11$^b$ & 0.02 \\
$K_s$ & HAWK-I/VLT & 0.38 & 24.75$^b$ & 0.01 \\
3.6 $\mu$m & IRAC/Spitzer & 1.8 & 21.42$^d$ & 0.00 \\
4.5 $\mu$m & IRAC/Spitzer & 1.8 & 21.57$^d$ & 0.00 \\
\hline
\multicolumn{5}{l}{$^a$ \url{http://irsa.ipac.caltech.edu/applications/DUST/}}\\
\multicolumn{5}{l}{$^b$ limiting magnitudes after PSF matching with $NB_{2071}$}\\
\multicolumn{5}{l}{$^c$ limiting magnitude in 2.5 arcsec aperture diameter}\\
\multicolumn{5}{l}{$^d$ limiting magnitude in 8.0 arcsec aperture diameter}\\
\end{tabular}
\label{tab1}
\end{table}
\subsection{Sample selection}\label{s2.3}
\subsubsection{Narrow-band selection}\label{s2.3.1}
We selected the sample of HAEs by the combined technique of narrow-band selection
\citep{Bunker:1995} and $Bz'K_s$ colour selection \citep{Daddi:2004}.
The former selection is defined by the following criteria,
\begin{eqnarray}
K_s-NB &>& -2.5\log(1-\Sigma\delta10^{-0.4(ZP-NB)}) +\zeta \label{eq1} \\
K_s-NB &>& 0.253 \label{eq2}
\end{eqnarray}
where $\Sigma$ is the confidence level (in sigma) of the colour-excess and $\delta$ is
defined by the combined $1\sigma$ background noise at $NB$ ($\equiv NB_{2071}$) and
$K_s$ bands, ($\delta=\sqrt{\sigma_{NB}(S)^2+\sigma_{Ks}(S)^2}$ where $S$ is the
photometric aperture area). $ZP$ is the zero point magnitude of the $NB_{2071}$ image.
$\zeta$ is a correction factor of the colour term. We use $\zeta=-0.04$ which
corresponds to the median value of the colour terms in the entire HAEs
(Appendix~\ref{a1}). The former equation reflects the narrow-band flux limit
($>3\times10^{-17}$ erg~s$^{-1}$cm$^{-2}$) and the latter colour threshold
(eq.~\ref{eq2}) corresponds to the equivalent width limit of narrow-band flux
(EW$_\mathrm{NB}=30$ \AA\ in the rest frame for $z=2.15$). The EW$_\mathrm{NB}$ limit
is chosen so as not to accidentally pick up contaminant non-emitters
(Appendix~\ref{a1}).
We note that the measurement of background noise ($\sigma(S)$) in this work and recent
other narrow-band studies (e.g., \citealt{Hayashi:2010,Matthee:2017,Hayashi:2018b}) is
different from the original calculation by \citet{Bunker:1995}. \citet{Bunker:1995}
define the noise by $\sigma(S)=\sqrt{\pi r^2\sigma_0^2}$ where $r$ is an aperture
radius and $\sigma_0$ is $1\sigma$ background noise in pixel. This definition assumes
that photometric error is proportional to the aperture radius, however, the real
science images have pixel-to-pixel correlations (see e.g., \citealt{Skelton:2014}),
which lead to underestimated background noise especially in the larger aperture area.
We indeed obtained the power law functions $N=$ 1.412 and 1.345 ($N$ is defined by
$\sigma(S)\propto r^N$) at $NB_{2071}$ and $K_s$ images based on randomly-positioned
empty apertures with different radii across the image. This work thus employs the
fixed background noise at each band ($NB_{2071,1\sigma}=25.14$ mag and
$K_{s,1\sigma}=25.18$ mag), derived by placing random empty apertures with the same
diameter (1.4 arcsec) in the selection process. We here ignore the local sky variance
that is estimated to be $\le0.1$ mag across each image.
We performed source detection in the reduced narrow-band image, using SExtractor (ver.
2.19.5, \citealt{Bertin:1996}). We set detection parameters of {\sc detect$\_$minarea}
$=9$, {\sc detect$\_$thresh} $=1.2$, {\sc analysis$\_$thresh} $=1.2$, and
{\sc deblen$\_$mincont} $=1\times10^{-4}$. $K_s$-band photometry was conducted by the
double-image mode of the SExtractor with the $NB_{2071}$ image for the source
detection. Input parameters for source photometry are set with {\sc back$\_$size}
$=64$, {\sc back$\_$filtersize} $=5$, {\sc backphoto$\_$type} = {\sc local}, and
{\sc backphoto$\_$thick} $=32$ (the same applies all source photometric processes
hereafter). According to the Monte Carlo simulation with randomly-positioned PSF models
embedded in the narrow-band image (see Appendix~\ref{a2} for details), the magnitude
limit of 95 percent completeness in the source detection is $\sim22.8$ mag.
Figure~\ref{fig2} shows the colour--magnitude diagram ($NB_{2071}$ versus
$K_s-NB_{2071}$) for $NB_{2071}$ detected sources in the PKS~1138 region. One should
note here that their $NB_{2071}$ magnitudes and colours are based on the fixed
aperture photometry of 1.4 arcsec diameter. We assume two sigma limiting magnitude for
non-detections at $K_s$ band. We then select the objects with $\Sigma>3$
colour-excesses as our narrow-band emitter (NBE) sample. The $\Sigma=3$ limit in this
work is more conservative when compared to $\Sigma=2$ in \citet{Kurk:2004a} and
$\Sigma=2.5$ in \citet{Koyama:2013a} for the same field. Nevertheless, thanks to the
deeper observing depth than the previous work, $\gtrsim1.5$ times more emitters (97
samples) meet our colour criteria in the same area. Our Monte Carlo simulation claims
that this narrow-band selection has 68 and 95 percent completeness at
$NB_{2071}=$ 22.80 and 22.45 mag, respectively (Appendix~\ref{a2}).
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig2.pdf}
\caption{Colour--magnitude diagram, $NB_{2071}$ versus $K_s-NB_{2071}$.
The black dots are all NB-detected sources. The red squares indicate narrow-band
emitters showing their narrow-band flux excesses greater than three sigma levels and
EW$_\mathrm{NB}$ higher than 94.5 \AA\ (30 \AA\ in the rest frame at $z=2.15$).
Blue solid, dash-dotted, and dotted lines are 3, 5, and 10 $\Sigma$ excess,
respectively. The blue horizontal line shows the EW limit.}
\label{fig2}
\end{figure}
\subsubsection{Colour selection}\label{s2.3.2}
Combined with the past spectroscopic observations
\citep{Pentericci:2002,Kurk:2004b,Croft:2005,Doherty:2010,Tanaka:2013,Shimakawa:2014}
and narrow-band {\rm\,Ly$\alpha$}\ imaging \citep{Kurk:2000}, we already have 36 secure HAE sources
with spectroscopic confirmation or narrow-band excess in two filters in {\rm\,H$\alpha$}\ and {\rm\,Ly$\alpha$}\
lines. For the remaining NBEs, even though the survey field is known to be a massive
overdense region, it is important to carry out further selection to select HAEs more
likely to be associated with PKS~1138 at $z=2.2$, and exclude other line contaminants
e.g., background {\rm\,[O{\sc iii}]}, {\rm\,H$\beta$}\ line emitters at $z>3$ and foreground {\rm\,Pa$\alpha$}\ emitter at
$z=0.1$.
Colour--colour selection has been widely used for further selection to remove other
line contaminants \citep{Koyama:2013a,Tadaki:2013}. Although it would be better to also
integrate with photometric redshifts as demonstrated by the High-redshift(Z) Emission
Line Survey (HiZELS; \citealt{Geach:2008,Sobral:2013}), this work does not employ
photometric redshifts since available photometric bands are not many as used in
such large panoramic surveys.
For $z\sim2.2$ sources, we employ the well known $BzK$ ($\equiv(z-K_s)-(B-z)$)
selection \citep{Daddi:2004,Daddi:2005} which is accessible given our imaging dataset
($Bz'K_s$) as employed in the previous work \citep{Koyama:2013a}. The $BzK$ colour
criteria enable culling of star-forming and passive galaxies at $z\sim$ 1.4--2.5
without extinction correction. We plot NB detections with $>2\sigma$ detection at $K_s$
band on the $Bz'K_s$ colour--colour diagram (fig.~\ref{fig3}). In addition, we also
show spec-$z$ sources in the COSMOS--CANDELS field
\citep{Scoville:2007,Capak:2007,Grogin:2011,Koekemoer:2011} from the MOSFIRE Deep
Evolution Field survey (MOSDEF; \citealt{Kriek:2015}) as a reference sample. $Bz'K_s$
colours of these spec-$z$ sources are derived from the 3D-HST database
\citep{Brammer:2012,Skelton:2014}, which were originally taken by the large legacy
surveys with the Subaru Telescope and Vista \citep{Taniguchi:2007,McCracken:2012}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig3.pdf}
\caption{$Bz'K_s$ colour--colour diagram for the PKS~1138 region. Red squares and blue
symbols indicate the narrow-band emitters and the spec-$z$ samples at $z=$ 2.0--2.3
(circles) and 3.0--3.3 (crosses) from the MOSDEF survey \citep{Kriek:2015},
respectively. Spectroscopically-confirmed HAEs are highlighted by open black squares.
Grey dots are NB-detected sources. The figure only shows objects with $>2\sigma$
detection at $K_s$-band. Two sigma limiting magnitudes replace band photometry for
faint sources at $B$ or $z'$-band. The black solid line is our colour threshold defined
to remove those foreground or background contaminants. The horizontal dashed line is
the colour criterion of DRG. The black cross on the upper right shows the typical
$1\sigma$ photometric error of the narrow-band emitters.}
\label{fig3}
\end{figure}
We derive $Bz'K_s$ colours of NB-detections based on the output from the SExtractor
({\sc mag$\_$auto}) with double image mode. All images were tailored to the size of
the narrow-band image with a scale of 0.117 arcsec per pixel. We employ the outputs of
{\sc mag$\_$auto} in each band and we set {\sc kron$\_$fact} $=2.5$. One should note
that source photometry for the $B$-band image with a much larger seeing FWHM
(table~\ref{tab1}) was also executed independently and then we chose brighter $B$-band
flux densities from single or double image mode for individual sources. Based on
colours of confirmed members and spec-$z$ sources, we set the colour thresholds of
$Bz'K_s>0$ or ($z'-K_s$) $>2.5$, and then select an additional 32 HAE members as well
as remove 16 narrow-band emitters as other line emitters (table~\ref{tab2}). Here, we
adopt a $Bz'K_s$ colour criterion that is different from the \citet{Daddi:2004}
prescription ($BzK_s\geq-0.2$). We assume two sigma limiting magnitudes for
non-detections, and then evaluate them if those upper limits or lower limits can meet
our selection criteria. The $Bz'K_s$ selection cannot perfectly guarantee that the
selected NBEs are our targeting HAEs at $z=2.15\pm0.02$ though (fig.~\ref{fig3}). We
indeed find that two confirmed protocluster members drop below our selection limit. On
the other hand, some reference spec-$z$ sources at $z=$ 3.0--3.3 break into the realm
of $Bz'K_s$-selected galaxies. Also, colour-selected HAEs and rejected line emitters
near boundaries of the colour criteria cannot be securely classified once we take
account of those photometric errors. While we count these colour-selected emitters as
HAEs throughout this paper at this time, we definitely require follow-up spectroscopy
for the robust identification of these sources in the future. The online catalogue
(appendix~\ref{a3}) summarises the identification status for individual HAE samples in
detail.
When taken together, a total number of 68 HAEs have been selected as the protocluster
members in this work. 51 out of them are already discovered by our previous survey
\citep{Koyama:2013a}, meaning that our deeper data increase the number of the HAE
sample by 33 percent. Besides these, we have 13 NBEs which cannot be removed by the
$Bz'K_s$ colour due to insufficient photometric data. These unknown emitter samples
are defined as HAE candidates. Given the ten times higher density in the protocluster
region (\S\ref{s3.4}), most of these faint emitters should be HAEs. The contamination
rate in the HAE candidates may be around $\sim20$ percent considering 16
colour-rejected NBEs amongst the 84 ($36+32+16$) emitters (table~\ref{tab2}).
\begin{table}
\caption{Classification of NBEs and DRG$_\mathrm{nIR}$ samples in PKS~1138 region. This
work employs confirmed and colour-selected emitters as the HAE sample. We also use HAE
candidates when we derive distribution functions (\S\ref{s3.4}). See \S\ref{s2.3} for
details.}
\begin{center}
\begin{tabular}{lrl}
\hline
Class & N & Description \\
\hline
NBEs & 97 & narrow-band emitters ($\Sigma>3$) \\
HAEs (confirmed) & 36 & confirmed by spec-$z$ or {\rm\,Ly$\alpha$}\ line \\
HAEs (by colours) & 32 & selected by $Bz'K_s$ colour \\
other line emitters & 16 & rejected by $Bz'K_s$ colour \\
\hline
HAEs & 68 & confirmed $+$ colour-selected HAEs \\
HAE candidates & 13 & cannot be rejected by colours \\
\hline
DRG$_\mathrm{nIR}$ & 34 & $z'-K_s>2.5$ w/o 24 $\mu$m detection \\
\hline
\end{tabular}
\end{center}
\label{tab2}
\end{table}%
\subsubsection{Distant red galaxies (DRGs)}\label{s2.3.3}
We also establish a reference sample of distant red galaxies (DRGs) that do not show
signs of active star formation. These objects allow us to infer the selection bias of
our narrow-band technique at the massive end, and also provide the upper limit to the
quiescent population in the derivation of the stellar mass function (\S\ref{s3.1}). We
first chose objects with significant $K_s$-band detection, $K_s<23.4$ ($5\sigma$ limit
mag), corresponding to the 95 percent completeness limit for massive galaxies
(M$_\star>10^{10.5}$ {\rm M}$_{\odot}$) according to the photo-$z$ source catalogue in the COSMOS
field \citep{Laigle:2016}. We then select passive $BzK$ (pBzK) galaxies that satisfy
($z'-K_s$) $>2.5$ (fig.~\ref{fig3}) and do not overlap with NBEs nor MIPS/Spitzer 24
$\mu$m sources reported by \citet{Koyama:2013a}. One should note that this colour
threshold is different from the classic definition of DRGs ($J-K_s>2.3$ in vega) by
\citet{Dokkum:2004} and \citet{Franx:2003}. The cross-checking with MIPS 24$\mu$m
sources allows us to remove significant dusty starburst populations. The detection
limit at the MIPS 24$\mu$m image roughly corresponds to the infrared luminosity of
$L_\mathrm{IR}\sim10^{12}$ $L_\odot$ and SFR $\sim100$ {\rm M}$_{\odot}$~yr$^{-1}$ at $z=2.15$.
This selection results in 34 DRG$_\mathrm{nIR}$ candidates without bright-IR emission
($L_\mathrm{IR}\gtrsim10^{12}$ $L_\odot$), which are described as DRG$_\mathrm{nIR}$
hereafter. Three of these are known to be protocluster members confirmed with
spectrophotometric analysis \citep{Tanaka:2013}. According to a photometric redshift
code, {\sc eazy} \citep{Brammer:2008,Brammer:2011}, measured photometric redshifts
fall within $z=2.1\pm0.2$ in 17 sources.
\subsubsection{X-ray sources}\label{s2.3.4}
We checked the presence of X-ray emission from our HAE samples using an image from the
Chandra X-ray Observatory. Our survey field is covered by the S3 chip with the ACIS-S
detector. The data quality and source catalogue were published in \citet{Carilli:2002}
and \citet{Pentericci:2002}. However, we double-checked the data independently based
on the Chandra Source Catalogue (CSC v1.1, \citealt{Evans:2010}) and also by analysing
the original data with the Chandra Interactive Analysis of Observations (CIAO v4.7.6)
to obtain more detailed coordinates.
Based on X-ray detections selected by the CIAO code {\sc wavdetect} for an
exposure-weighted reduced image with {\sc mkexpmap}, we found that six HAEs
(\#40,46,58,68,73,95) have X-ray detections within 0.4 arcsec separation angle at
higher than four sigma levels. The faintest X-ray source has $4\times10^{-15}$
erg~s$^{-1}$cm$^{-2}$ and $1.4\times10^{44}$ erg~s$^{-1}$ in unabsorbed flux and
luminosity (assuming the redshift of $z=2.15$) at the broadband (0.5--7.0 keV)
according to the CSC, respectively. Given such a shallow detection limit, these X-ray
sources are expected to originate from active galactic nuclei. All of these X-ray
sources have been identified as \#3,5,6,7,16 in \citet{Pentericci:2002}, whereas \#7
contains two HAE sources defined in this work: one is the Spiderweb radio galaxy (\#73
in this work), and the other is HAE-058. Corresponding identification numbers to each
HAE are fully described in our catalogue (appendix~\ref{a3}).
\subsubsection{Other resources}\label{s2.3.5}
The Spiderweb protocluster is a well-surveyed region, with numerous studies in
addition to those already mentioned, e.g., MIPS 24 $\mu$m imaging with
the Spitzer Space Telescope \citep{Mayo:2012,Koyama:2013a,Koyama:2013b}, LABOCA 870
$\mu$m imaging with the APEX telescope \citep{Dannerbauer:2014}, CO($1-0$) observation
with ATCA \citep{Emonts:2016,Dannerbauer:2017,Emonts:2018}, and CO($3-2$) observation
with ALMA (Tadaki et al. in preparation).
Because of the restricted field coverage relative to our survey area, or serious
blending issue due to poor spatial resolutions, we do not use these other resources,
mostly in the mid-IR to radio regime, unless otherwise mentioned. On the other hand,
these past studies are useful to characterise some specific HAEs, and thus, such
information is referenced where appropriate throughout the paper.
\subsection{Derivation of physical properties}\label{s2.4}
This section explains how we derive line flux, stellar mass, and amount of dust
reddening. The measuring methods are similar to those in the first paper of the
MDCS series \citep{Shimakawa:2018}.
\subsubsection{SED fitting}\label{s2.4.1}
We use SED-fitting to derive stellar masses and dust extinctions of our samples based
on the SED-fitting code ({\sc fast}) distributed by \citet{Kriek:2009}. We use the
\citet{Bruzual:2003} stellar population model, the \citet{Calzetti:2000} extinction
law, and the \citet{Chabrier:2003} IMF. We then run the code with a fixed redshift of
$z=2.15$ independently of spectroscopic confirmation and low metal abundance of
$Z=0.004$ ($0.2 Z_\odot$), and assume delayed exponentially declining star formation
history (SFR $\propto~t\cdot exp(-t/\tau)$). $\tau$ value and age are allowed to be
$10^9$--$10^{11}$ yr and $10^{7.6}$--$10^{9.4}$ yr, respectively. We allow the amount
of stellar extinction ($A_V$) to be between 0 and 3 mag. The outcome of the choice of
these parameter sets does not significantly affect the stellar mass estimations.
However, the obtained dust extinction systematically depends on input parameters. If
we employ solar metal abundance instead of $Z=0.2 Z_\odot$, for instance, derived
$A_V$ values become systematically lower by 0.2--0.3 mag than those with $Z=0.004$.
Indeed, dust correction is the major issue for narrow-band studies due to the lack of
{\rm\,H$\beta$}\ line information. Considering this model dependency, we pay special attention to
physical properties requiring dust correction, such as SFR, throughout the paper.
We then carried out the SED fitting based on multi-band photometry derived in the same
way as for the $Bz'K_s$ colour estimation in the previous subsection. First, we
performed PSF-matching for the $F475W$, $F814W$, $Y$, $H$, $K_{s,\mathrm{HAWKI}}$
images to the seeing size of $NB_{2071}$. Source photometry at IRAC 3.6 and 4.5 $\mu$m
bands were conducted independently, and we cross-matched those to NBEs within 1 arcsec
distance. Whilst we executed the SED-fitting with IRAC photometry if available, we
confirm that IRAC data have negligible effects on stellar mass and $A_v$ estimation in
our samples. This lack of systematic discrepancy regardless of the availability of
IRAC photometry is consistent with past work \citep{Elsner:2008,Muzzin:2009}.
Derived stellar masses and dust reddening are summarised in appendix~\ref{a3}. We
employ $1\sigma$ errors of obtained parameters from 100 Monte Carlo simulations
carried out with the {\sc fast} code.
Typical fitted SED spectra of massive HAEs (M$_\star>10^{10.5}$ M$_\odot$) are
presented in fig.~\ref{fig4}, which are divided into the spectra of HAEs with or
without X-ray counterparts. We should note that, in spite of the importance of the SED
decomposition into the stellar light and the nuclear component
\citep{Merloni:2010,Santini:2012}, this work ignores this procedure due to the lack of
photometric bands at rest-frame IR bands. Although we have Spitzer/MIPS 24 $\mu$m
data, the serious blending issue does not provide us with reliable mid-IR photometry.
At the least, the stellar mass measurement of the Spiderweb radio galaxy, the most
luminous X-ray source in our sample, is highly uncertain. Also, we see a clear excess
at IRAC bands from the extrapolation of model-inferred SED in one of X-ray HAEs (\#58)
and whose derived stellar mass should be overestimated as well. For other AGN host
HAEs, since their photometry can be fitted only by stellar components, it remains
unclear how reliable our stellar mass estimates are. High resolution deep mid-IR data,
by e.g., JWST/MIRI \citep{Rieke:2015}, are needed to decompose those SEDs and obtain
pure stellar components.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig4.pdf}
\caption{
Median stellar spectra of massive (M$_\star>10^{10.5}$ M$_\odot$) HAEs with
and without X-ray emission, which are represented by purple thick and black thin
lines, respectively. These are derived from the median values of fitted SED spectra
normalised at $\lambda_\mathrm{rest}=5500$ \AA\ for individuals. Their median values
and $1\sigma$ scatters of observed flux densities at 11 photometric bands are shown by
purple squares and black diamonds, respectively (the data points are slightly shifted
in a transverse direction for better visibility). Minimum--median--maximum values of
derived log stellar mass in each group are 10.81--11.01--11.07 in the HAEs with X-ray
sources and 10.55--10.77--11.33 in the HAEs without X-ray counterparts.}
\label{fig4}
\end{figure}
\subsubsection{Narrow-band flux}\label{s2.4.2}
We obtained narrow-band line flux ($F_\mathrm{NB}$), emission subtracted flux
density at $K_s$-band ($f_c$), and rest-frame equivalent width of narrow-band flux
(EW$_\mathrm{NB}$) by the following formula,
\begin{eqnarray}
F_\mathrm{NB} &=& \Delta_\mathrm{NB} \frac{f_\mathrm{NB}-f_\mathrm{Ks'}}{1-\Delta_\mathrm{NB}/\Delta_\mathrm{Ks}} \label{eq3} \\
f_{c} &=& \frac{f_\mathrm{Ks'}-f_\mathrm{NB}\cdot\Delta_\mathrm{NB}/\Delta_\mathrm{Ks}}{1-\Delta_\mathrm{NB}/\Delta_\mathrm{Ks}} \label{eq4} \\
\mathrm{EW_{NB}} &=& \frac{F_\mathrm{NB}}{f_c\cdot(1-z)} \label{eq5}
\end{eqnarray}
where $\Delta_\mathrm{NB}$ and $\Delta_\mathrm{Ks}$ are full-width-half-maximum (FWHM)
of NB$_{2071}$ (270 \AA) and $K_s$ band (3100 \AA) filters, respectively.
$f_\mathrm{Ks'}$ is $K_s$-band flux density including the colour term correction
(\S\ref{s2.3.1}). We employed $-0.04$ mag for the colour term correction based on the
median value of model-inferred SED spectra of the entire HAEs (fig.~\ref{fig1a}). Given
the similar centre wavelength between the two filters, uncertainty from the colour
correction is negligibly small relative to the total flux errors.
On the other hand, the shape of the filter throughput including atmospheric
transmission on Maunakea (fig.~\ref{fig1}) may cause an additional $\sim12$ percent
error in the flux estimation according to the standard deviation of the response curve
at wavelengths within the filter FWHM. We incorporate this error budget into the
narrow-band flux errors of HAEs individually. One should note that we likely
overestimate this error value given that HAEs tend to gather towards the protocluster
system ($z\sim2.156$) along the line of sight \citep{Shimakawa:2014}.
We then obtained observed {\rm\,H$\alpha$}\ luminosities of HAEs as follows. We assume a fixed
redshift of $z=2.15$ that corresponds to the {\rm\,H$\alpha$}\ redshift captured by the centre of
the $NB_{2071}$ filter. Flux contribution from the {\rm\,[N{\sc ii}]}$\lambda\lambda6550,6585$
doublet is corrected based on their stellar masses derived by the SED fitting
(\S\ref{s2.4.1}). Our past spectroscopic observation has derived typical
{\rm\,[N{\sc ii}]}$\lambda6585$/{\rm\,H$\alpha$}\ flux ratios (N2; \citealt{Pettini:2004}) for HAEs in PKS~1138
at different stellar mass bins \citep{Shimakawa:2015}, which can be characterised by
the following relation,
\begin{equation}
\mathrm{N2} = -0.71 +0.33 \times [\log(\mathrm{M_\star/M_\odot}) - 10].
\label{eq6}
\end{equation}
This prescription enables relatively self-consistent N2 correction for the narrow-band
flux in our sample. We assume {\rm\,[N{\sc ii}]}$\lambda6550$:{\rm\,[N{\sc ii}]}$\lambda6585$ = 1:3 to remove
{\rm\,[N{\sc ii}]}$\lambda6550$ flux as well \citep{Osterbrock:1974}. In addition, we incorporate
the uncertainty of the N2 correction into the derived {\rm\,H$\alpha$}\ luminosities based on the
typical observational scatter $\Delta\log\mathrm{(O/H)_{N2}}\sim0.1$ dex (i.e.,
$\Delta$N2 $=0.18$ dex) of the N2-inferred mass--metallicity relation
\citep{Tremonti:2004,Mannucci:2010,Steidel:2014}. The calculated {\rm\,H$\alpha$}\ luminosities and
those total error budgets are summarised in the online catalogue (appendix~\ref{a3}).
\section{Results}\label{s3}
The goal of this paper is to investigate physical properties of massive HAEs in the
Spiderweb protocluster (PKS~1138) at $z=2.2$. Our previous paper
\citep{Shimakawa:2018} reported the vigorous formation of more massive galaxies in
fragmented dense groups alongside intergroup regions within the USS~1558 protocluster
at $z=2.53$. Compared to USS~1558, PKS~1138 is apparently a more advanced and reddened
protocluster system \citep{Kodama:2007}. More specifically, \citet{Galametz:2012}
have reported three times higher number density of old populations selected by IRAC
colour in this field compared to the typical radio galaxy environments at high
redshifts including USS~1558. Given the fact that massive galaxies in the protocluster
are destined to grow into bright red sequence galaxies in the local Universe,
identifying these massive HAEs will help us to infer the evolutionary steps cluster
galaxies would have experienced in their maturing phases at $z\sim2$.
\subsection{Stellar mass functions}\label{s3.1}
We first derive the stellar mass function of HAEs in PKS~1138 whose stellar masses are
individually derived from the SED fitting.
Since the mass estimations do not include the SED decomposition to remove AGN
contamination, all results obtained in this section must be taken with caution.
Analysing the Spiderweb protocluster region especially suffers from this issue due to
the number excess of luminous X-ray sources \citep{Pentericci:2002}.
We use the same measuring method as in our previous paper \citep{Shimakawa:2018} for
HAEs in USS~1558 protocluster region at $z=2.53$.
The most important part of the derivation of the stellar mass function for narrow-band
selected emitters is the completeness correction. Following \citet{Shimakawa:2018}, we
evaluate both detection completeness and selection completeness with the Monte Carlo
simulation. The detection completeness is defined as the fraction of missing samples
in the source detection process, which is highly dependent on the initial parameters
of the SExtractor code (\S\ref{s2.3.1}). The selection completeness is a specific
problem of the narrow-band selection, which is firstly noted by \citet{Sobral:2009}
and then developed by their following analyses
\citep{Sobral:2012,Sobral:2013,Sobral:2014}. The selection completeness indicates the
completeness in the process of the narrow-band colour selection (\S\ref{s2.3.1}).
Evaluating the selection completeness is especially crucial since the narrow-band
selection is not only based on the depth of the narrow-band image, but is also
dependent on the colour between narrow-band and broad-band photometry (see
\citealt{Sobral:2009}). Indeed, our Monte Carlo analysis indicates that the
narrow-band selection requires an additional 20--70 percent completeness correction at
the faint end relative to the completeness correction only for the detection. The
detailed procedure of our completeness correction is examined in Appendix~\ref{a2}.
Figure~\ref{fig5} shows the number densities as a function of stellar mass for HAEs in
PKS~1138. We evaluate the number densities of HAEs by
$\phi(\log{L})=\Sigma_i(V_\mathrm{max}\cdot C(NB)\cdot\Delta(\log{L}))^{-1}$, where
$L=\mathrm{M}_\star/\mathrm{M}_\odot$ and $V_\mathrm{max}$ ($=3676$ co-Mpc$^3$) is the
volume size, respectively. The latter is obtained from the filter FWHM of the
narrow-band filter (55 co-Mpc) and the survey area (66 co-Mpc$^2$). Since the redshift
distribution of HAEs is more concentrated around the protocluster centre
\citep{Shimakawa:2014}, we tend to overestimate the volume size. Open squares in
fig.~\ref{fig5} include the completeness corrections that also incorporate 13
unclassified HAE candidates (table~\ref{tab2}) to compensate loss from the $Bz'K_s$
colour selection. We count these candidates with the additional correction of 20
percent possible contamination (\S\ref{s2.3.2}). One should note that while this
further correction could affect the resultant fitting parameters, this does not change
our conclusion since all HAE candidates are less-massive galaxies at stellar masses
lower than $10^{10.5}$ {\rm M}$_{\odot}$, most of which are outside of the scope of this work.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{fig5.pdf}
\caption{Stellar mass functions in various fields at $z\sim2$. Open and filled boxes
show the number densities with and without completeness correction at each stellar
mass bin. The blue curve represents the stellar mass function in the USS~1558
protocluster at $z=2.5$. Purple and yellow curves are those in the group and
intergroup regions therein, respectively. Thin grey and pink lines are the stellar
mass function in the general field at $z\sim2$. The former is based on star-forming
galaxies selected from their rest-$UVJ$ colours and photo-$z$ \citep{Davidzon:2017}.
The latter is narrow-band selected HAEs including additional sorting with photo-$z$
and colours \citep{Sobral:2014}. Dotted lines indicate the extrapolated lines from
the available data range in each function.
}
\label{fig5}
\end{figure}
We fit the stellar mass distribution using the Schechter function
\citep{Schechter:1976}, which is given by the following equations,
\begin{eqnarray}
\phi(L)dL &=& \phi^\ast (\frac{L}{L^\ast})^\alpha \exp(-\frac{L}{L^\ast})\frac{dL}{L^\ast} ~~~~~~~~~~~~~~~~ or \\
\phi(L)dL &=& \phi^\ast (\frac{L}{L^\ast})^{\alpha+1} \exp(-\frac{L}{L^\ast})\ln{10}~d(\log{L}),
\end{eqnarray}
where $L^\ast$ is the characteristic stellar mass at which the power law slope cuts
off. We then fit the stellar mass distribution with the Schecter function based on the
{\sc mpfit} code \citep{Markwardt:2009}\footnotemark[4]. We do not use the stellar
mass bins of $10^{8.5-9.7}$ {\rm M}$_{\odot}$\ in the fitting since we cannot probe typical star-forming
galaxies at these bins due to the flux limit (\S\ref{s3.3}). Also, we remove the radio
galaxy (M$_\star=10^{12.4}$ {\rm M}$_{\odot}$). The thick frames in fig.~\ref{fig5} highlight the
sample bins used in the curve fitting.
\footnotetext[4]{\url{http://purl.com/net/mpfit}}
\begin{table}
\centering
\caption{Results of Schechter function fitting for the stellar mass distribution. The
third and fourth columns indicate the normalisation factors between PKS~1138 and the
general fields at the similar redshift range reported by \citet{Davidzon:2017} and
\citet{Sobral:2014} at the stellar mass of $10^{9.7}$ {\rm M}$_{\odot}$, respectively.}
\begin{tabular}{cccc}
\hline
log(M$_\star^\ast$/M$_\odot$) & log($\Phi_{\mathrm{M}_\star}^\ast$/Mpc$^{-3}$) & $\Phi_\mathrm{9.7}$/$\Phi_\mathrm{D17}$ & $\Phi_\mathrm{9.7}$/$\Phi_\mathrm{S14}$ \\
\hline
$11.726\pm0.756$ & $-3.097\pm0.432$ & 9.12 & 13.29 \\
\hline
\end{tabular}
\label{tab3}
\end{table}
The derived parameters of the curve fitting are given in table~\ref{tab3}. Due to the
small sample size, we fix the power law slope to $\alpha=-1.5$, as used in
\citet{Shimakawa:2018}, to minimise the fitting errors. We should note that our
restricted sample sizes and large binning sizes would not be sufficient to determine
the function parameters and even errors robustly. Indeed, it is known that there are
non-negligible variations of derived M$_\star^\ast$ and $\Phi^\ast$ even if one
employs much larger datasets, perhaps due to the cosmic variance and selection effects
\citep{Ilbert:2013,Muzzin:2013,Sobral:2014,Davidzon:2017,Hayashi:2018b}. Although
systematic comparisons are unfair because of such issues, the differences of the
number densities between PKS~1138 and the general fields
\citep{Sobral:2014,Davidzon:2017} at the stellar mass $\sim10^{10}$ {\rm M}$_{\odot}$\
(fig.~\ref{fig5}) suggest that the PKS~1138 protocluster is an approximately ten times
higher density region than the general field at a similar redshift.
The figure also shows the stellar mass function of HAEs in the USS~1558 protocluster
at $z=2.53$, and its group regions and intergroup regions \citep{Shimakawa:2018}.
These functions are derived by the same procedure as this work, which enables a
relatively fair comparison between two protoclusters at different redshifts. We find
that the cut-off stellar mass of HAEs in PKS~1138 at $z=2.2$.
Whilst the derived cut-off values would be overestimated since we may
overestimate stellar masses of AGN host galaxies, this is consistent with that in the
fragmented group regions (M$_\star^\ast\sim10^{11.5}$ {\rm M}$_{\odot}$) in USS~1558 at $z=2.5$,
within the error margins. Both have significantly higher characteristic stellar masses
than those in the intergroup regions (M$_\star^\ast\sim10^{10.6}$ {\rm M}$_{\odot}$) in USS~1558,
meaning that PKS~1138 is associated with the larger number of more massive HAEs than
lower density regions in USS~1558 at $z=2.53$. Also, this could suggest that PKS~1138
is at a point where fragmented cores are about to consolidate into a massive cluster
with a single core if we assume these two protoclusters are on the similar
evolutionary track to massive clusters \citep{Shimakawa:2014}. Comparing these two
protoclusters in the stellar mass function, the PKS~1138 region has a 1.5 times lower
number density of HAEs relative to USS~1558.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{fig6.pdf}
\caption{{\it Lower panel:} Stellar mass function in the PKS~1138 protocluster region
(same as in fig.~\ref{fig5}). This figure also plots the $1\sigma$ upper limits of the
stellar mass function for passive galaxies at M$_\star>10^{10.5}$ {\rm M}$_{\odot}$\ based on the
DRG$_\mathrm{nIR}$ samples. We select DRG$_\mathrm{nIR}$ whose rest-frame $UVJ$ colours
agree with passive galaxy populations within $1\sigma$ errors. The lower limits are
constrained by three quiescent objects with spectroscopic (or spectrophotometric)
confirmation by \citet{Tanaka:2013}. Grey dash-dotted and dashed curves are the stellar
mass function of star-forming and passive galaxies in the general field at $z=$ 2--2.5
\citep{Davidzon:2017}.
{\it Upper panel:} Fraction of X-ray selected AGNs among HAEs (orange crosses) and
passive galaxies (red hatched region) in PKS~1138 as a function of stellar mass. The
latter is calculated based on the $1\sigma$ upper and lower limits shown in the lower
panel, and error-bars include Poisson noise. Blue triangles indicate the X-ray AGN
fraction among HAEs in USS~1558 at $z=2.53$ \citep{Macuga:2018}. The grey dotted line
indicates the passive fraction in the general field at $z=$ 2--2.5 reported by
\citet{Davidzon:2017}.}
\label{fig6}
\end{figure}
We also estimate the fraction of bright X-ray sources among HAEs in each stellar
mass bin. Studying AGN activities across different stellar mass ranges is essential
since AGNs are thought to have mass dependence and especially play critical roles in
massive-end systems \citep{Ferrarese:2000,Kauffmann:2003,Matteo:2005}. We thus merely
check the X-ray AGN fraction in our HAE samples and then find that more than 60 percent
(4/6) of very massive HAEs host AGNs at M$_\star=10^{11-12.5}$ {\rm M}$_{\odot}$\ (fig.~\ref{fig6}).
Such a high AGN fraction may be even more enhanced once we get a better AGN
identification tool (see discussion \S\ref{s4}). The figure also shows AGN fraction in
the other protocluster area, USS~1558 at $z=2.53$ \citep{Macuga:2018}. This detection
limit of $L_X=3\times10^{43}$ erg~s$^{-1}$ is deeper than that in PKS~1138. In
USS~1558, despite that, there is no bright X-ray source in a higher density region
within the protocluster except the radio galaxy that is the only X-ray source (1/3) in
the highest stellar mass bin (M$_\star=10^{11}$--$10^{11.5}$ {\rm M}$_{\odot}$). We summarise the
AGN fraction in each stellar mass bin in table~\ref{tab4}.
\begin{table}
\centering
\caption{X-ray AGN fractions amongst HAEs in PKS~1138 ($z=2.15$) and USS~1558
($z=2.53$) at different stellar mass bins. One should note that these do not include
the HAE candidates and the completeness correction.}
\begin{tabular}{lll}
\hline
log(M$_\star^\ast$/M$_\odot$) & PKS~1138 & USS~1558 \\
\hline
8.0--8.5 & --- & 0/5 (0\%) \\
8.5--9.0 & 0/3 (0\%) & 1/18 (6\%) \\
9.0--9.4 & 0/14 (0\%) & 0/21 (0\%) \\
9.4--9.7 & 0/10 (0\%) & 0/20 (0\%) \\
9.7--10.0 & 0/11 (0\%) & 0/17 (0\%) \\
10.0--10.5 & 1/17 (6\%) & 1/18 (6\%) \\
10.5--11.0 & 1/7 (14\%) & 0/5 (0\%) \\
11.0--11.5 & 3/5 (60\%) & 1$^\ast$/3 (33\%) \\
11.5--12.5 & 1$^\ast$/1 (100\%) & --- \\
\hline
\multicolumn{3}{l}{$^\ast$ Radio galaxies}\\
\end{tabular}
\label{tab4}
\end{table}
In addition, we tentatively constrain the quenching fraction in PKS~1138 by combining
HAEs with DRG$_\mathrm{nIR}$ samples (\S\ref{s2.3.3}). Within the target area, there
are 34 DRG$_\mathrm{nIR}$ sources that do not have flux excesses at narrow-band nor
bright dust emission at MIPS/24$\mu$m band, which provide the upper limit of the
distribution functions of passive galaxies in PKS~1138. We here employ only
DRG$_\mathrm{nIR}$ that can be classified as passive populations on the rest-frame
$UVJ$ plane within margins of $1\sigma$ errors. Also, three of them are
spectroscopically (or spectrophotometrically) identified by \citet{Tanaka:2013}. They
allow us to constrain the lower limit of number densities of passive objects. We derive
their stellar mass from the SED-fitting by assuming the fixed redshift of 2.15 and then
employ only the 23 DRG$_\mathrm{nIR}$ with the stellar mass greater than $10^{10.5}$ {\rm M}$_{\odot}$\
in this analysis.
Roughly expected number densities of passive galaxies and the passive fraction as a
function of the stellar mass can be seen in fig.~\ref{fig6}. The fraction of passive
galaxies in PKS~1138 is estimated to be $\sim36$ percent. One should note that,
however, such a constraint becomes almost irrelevant once we include errors. When we
compare the number density of our DRG$_\mathrm{nIR}$ samples at $K_s=$ 21--23 mag with
that of DRG$_\mathrm{nIR}$ selected in the same way (\S2.3.3) in the COSMOS field
\citep{Laigle:2016}, the excess factor of DRG$_\mathrm{nIR}$ in PKS~1138 is estimated
to be $\sim1.7$ (which is consistent with the estimation by \citealt{Kodama:2007}). The
upper limits of the number densities thus should be overestimated due to the foreground and
background contaminants. A future deep NIR survey with instruments such as Keck/MOSFIRE
and JWST/NIRSpec is required to obtain the passive fraction more reliably.
Regarding these trends, we caution about a potential issue in our narrow-band
selection. Our HAE samples are limited either by the narrow-band flux
($>3\times10^{-17}$ erg~s$^{-1}$cm$^{-2}$) or EW$_\mathrm{NB}$ ($>30$ \AA\ in the rest
frame) depending on their narrow-band magnitude (fig.~\ref{fig2}). Figure~\ref{fig7}
roughly explains how this selection bias may affect the different stellar mass ranges.
Since the narrow-band magnitude correlates with the stellar mass, the flux limit is the
primary bias in the narrow-band selection at the lower stellar mass regime,
$\lesssim10^{10.4}$ {\rm M}$_{\odot}$\ (i.e., eq.~\ref{eq1}). On the other hand, towards the
massive end (M$_\star\gtrsim10^{10.4}$ {\rm M}$_{\odot}$), the EW$_\mathrm{NB}$ limit ($>30$ \AA)
as defined by eq.~\ref{eq2} drives the selection bias. Thus, we should note that our
HAE selection is not fully equal to the selection of star-forming populations at the
massive end, in the meaning that the selection is restricted by EW$_\mathrm{NB}$ that
roughly corresponds to specific SFR instead of the narrow-band flux (i.e., SFR). Such
an additional bias would require special attention when we regard our HAE samples as
star-forming galaxies. For example, we may underestimate the number of star-forming
galaxies at the massive end. Also, the AGN fraction could be overestimated since
EW$_\mathrm{NB}$ could be enhanced by the flux contribution in both {\rm\,H$\alpha$}\ and {\rm\,[N{\sc ii}]}\
lines from AGNs. For reference, \citet{Sobral:2016b} have found that AGNs contribute
$\sim15$ percent of the total {\rm\,H$\alpha$}\ luminosity density at any redshift up to $z=2.23$.
We return to the discussion of the AGN fraction in \S4.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig7.pdf}
\caption{EW of narrow-band flux (EW$_\mathrm{NB}$) against stellar mass. Blue large
and small squares show HAEs and HAE candidates respectively (table~\ref{tab2}). The
horizontal dash-dotted line indicates our EW selection limit that corresponds to the
rest-frame EW$_\mathrm{NB}=30$ \AA\ at $z=2.15$. Purple crosses have X-ray
counterparts. error-bars show $1\sigma$ uncertainties. We also plot upper limits of
EW$_\mathrm{NB}$ for the DRG$_\mathrm{nIR}$ samples by red circles for a reference.}
\label{fig7}
\end{figure}
\citet{Shimakawa:2018} have derived the dust-corrected {\rm\,H$\alpha$}\ luminosity function in the
USS~1558 field, with tentative extinction correction for individual HAEs. However, we
decide not to discuss the {\rm\,H$\alpha$}\ luminosity function towards the PKS~1138 region since
this protocluster is known to be associated with a large number of dusty starbursts
\citep{Stevens:2003,Mayo:2012,Koyama:2013a,Valtchanov:2013,Rigby:2014,Dannerbauer:2014}
and it is thus very challenging to properly estimate dust-corrected {\rm\,H$\alpha$}\ luminosities
for all HAEs on an equitable basis. If we apply dust correction in the same way as for
HAEs in the USS~1558 field \citep{Shimakawa:2018}, we can derive
$\log(L^\ast_\mathrm{H\alpha}$/erg~s$^{-1})=43.61\pm0.20$ and
$\log(\Phi_{L_\mathrm{H\alpha}}^\ast$/Mpc$^{-3})=-2.38\pm0.18$, respectively.
\subsection{Spatial distribution and rest-frame colours}\label{s3.2}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{fig8.pdf}
\caption{Spatial distribution with respect to the Spiderweb galaxy as shown by the star
symbol in the centre. Large filled and small open squares are HAEs and HAE candidates,
respectively. Symbol colours of HAEs indicate their rest-frame $U-V$ colours. Red
circles and black crosses indicate DRG$_\mathrm{nIR}$ and X-ray sources, respectively.
The colour map in the background shows the excess of surface number densities based on the
5th neighbour analysis.}
\label{fig8}
\end{figure*}
We show the positions of our updated HAE samples over the survey area in
fig.~\ref{fig8}. The underlying colour in the figure indicates the excess of the
number density ($\log(1+\delta)$), which is defined by the following equation,
\begin{equation}
\delta = \frac{\Sigma_\mathrm{5th}-C\cdot\overline{\Sigma}_\mathrm{5th}}{C\cdot\overline{\Sigma}_\mathrm{5th}}
\label{eq7}
\end{equation}
where $\Sigma_\mathrm{Nth}$ is the surface density in physical (ph-) Mpc$^{-2}$
including $N$ HAEs therein. We here adopt $N=5$. $C$ is the scaling factor which
normalises the mean number density over the entire protocluster field
($\overline{\Sigma}_\mathrm{5th}=13.5$ ph-Mpc$^{-2}$) to the typical density in
the general field. We tentatively apply $C=0.1$ inferred from the ten times higher
number density of the stellar mass function at M$_\star\sim10^{10}$ {\rm M}$_{\odot}$\ relative to
those in the general fields (\S\ref{s3.1}).
The PKS~1138 protocluster is associated with a very dense region of HAEs around the
Spiderweb galaxy as firstly reported by \citet{Kurk:2004a}: it contains more than 100
times as many galaxies as seen in the random field in the local scale (see also
\citealt{Kuiper:2011}). We find $>10$ new {\rm\,H$\alpha$}\ detections associated with this massive
structure. Also, four out of six HAEs with X-ray detections \citep{Pentericci:2002}
are positioned within or close to the central system. PKS~1138 is also known to have
filamentary structures on the east side \citep{Croft:2005,Koyama:2013a} which are
aligned along the line of sight as well \citep{Shimakawa:2014}. The most compact group
in this region can be seen at four co-Mpc away eastward from the radio galaxy. This
compact group involves four HAEs (\#25,26,27,29) within only 60 ph-kpc distance and
has a $3.6\sigma$ source detection ($5.0\pm1.4$ mJy) at LABOCA 870 $\mu$m (\#DKB12 in
\citealt{Dannerbauer:2014}). The peak density of passive galaxy candidates selected as
DRGs is slightly shifted towards the east direction, though more spectroscopic
identifications are needed to confirm this sub-structure.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig9.pdf}
\caption{Rest-frame $U-V$ colours as a function of surface number densities of HAEs
(from the left, $\Sigma_\mathrm{5th}$ ph-Mpc$^{-2}$ and distance from the Spiderweb
radio galaxy, $d_\mathrm{RG}$ ph-Mpc). The symbols are the same as in fig.~\ref{fig7}.
Solid and dashed lines indicate median values and 68th percentiles distributions of
rest-frame $U-V$ colour of HAEs with respect to each axis.}
\label{fig9}
\end{figure}
We then estimate the rest-frame $UVJ$ colours, and associated errors, of HAEs using
the EAZY code \citep{Brammer:2008,Brammer:2011}, and then investigate colour
dependence on the local scale. The rest $U$, $V$, $J$ photometries roughly correspond
to $Y$, $H$, and 3.6 $\mu$m bands, respectively.
Figure~\ref{fig9} shows the surface number densities vs. the rest-frame $U-V$ colours
of HAEs. We explore the colour dependence on local environments with different density
measurements, i.e., the surface densities including 5th neighbours, and distance from
the radio galaxy ($d_\mathrm{RG}$). We find no clear correlation between $U-V$ colours
and local densities of HAEs within the protocluster, which is unchanged when we use
$\Sigma_\mathrm{3th}$ or $\Sigma_\mathrm{10th}$ for the density measurement. These
results are inconsistent with the concentration of redder HAEs towards the
protocluster centre as reported by \citet{Koyama:2013a}. However, \citet{Koyama:2013a}
investigated the colour dependence of HAEs in an area twice as large as the survey
field of MDCS, and their results are enhanced by the absence of red HAEs in these
outer regions. We, therefore, conclude that the inconsistency between this work and
\citet{Koyama:2013a} is due to the insufficiently large survey area in this work to
confirm the finding of \citet{Koyama:2013a}.
We then characterise HAEs on the rest-frame $UVJ$ plane (fig.~\ref{fig10}). The
shallow depths of the IRAC bands ($m_{3\sigma}\sim21.5$ AB), mean that only 32 percent
of the entire HAE sample are detected at IRAC bands at a more than two sigma
confidence level. These IRAC detected HAEs are shown by the filled symbols in
fig.~\ref{fig10}. Typical errors are $\Delta(U-V)=0.31$ dex and $\Delta(V-J)=0.29$
dex, respectively. The remainders are indicated by open symbols and have rest $J$-band
magnitudes estimated from the extrapolated SED spectra. Uncertainties of rest $V-J$
colours in these non-IRAC detections would be $\sim0.6$ dex according to the EAZY
code.
As a result, we find that rest-frame $UVJ$ colours of HAEs agree with those of the
star-forming population \citep{Williams:2009,Moresco:2013,Straatman:2016,Fang:2018}
within the margin of error. Despite the significant uncertainties of individual
colours, HAEs hosting bright X-ray AGNs tend to have redder rest-frame $U-V$ colours,
which agree with the findings by \citet{Krishnan:2017}. More interestingly, we see that
HAEs with X-ray emissions (\#40,58,68,73,95) are preferentially located near the edge
of the quiescent population. The outlier lying at the bottom on the $UVJ$ plane (\#46)
is known to be an AGN (\#6 in \citealt{Pentericci:2002} and \#215 in
\citealt{Kurk:2004b}), with very broad {\rm\,H$\alpha$}\ line emission identified by near-infrared
spectroscopy with VLT/ISAAC \citep{Kurk:2004b}. The colour trend suggests that HAEs
hosting X-ray AGNs could be in the transition stage from dusty star-forming galaxies
to passive populations, i.e., in the post-starburst phase.
Another causal factor is the effect of the nuclear emission. AGN
contribution would redden more rest-frame $V-J$ colours than $U-V$ colours, and thus
this bias may rather weaken the colour discrepancy between HAEs and X-ray HAEs,
though a more detailed analysis is needed.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig10.pdf}
\caption{Rest-frame $UVJ$ colours of our HAE and DRG$_\mathrm{nIR}$ samples in the
protocluster region. Blue squares and red circles are HAEs and DRG$_\mathrm{nIR}$,
respectively. Filled symbols indicate a $>2\sigma$ detection in either 3.6 or 4.5
$\mu$m, or both. We do not have sufficient IRAC photometry for open symbols, and those
$VJ$ colours are not reliable. Blue bold and red thin crosses on the lower right
indicate typical errors of IRAC-detected HAEs and DRG$_\mathrm{nIR}$, respectively.
According to the EAZY code, the typical uncertainty of the objects without IRAC
detections ($>2\sigma$) is $\sim0.6$ dex. The black arrow corresponds to the colour
shift by dust reddening ($\Delta A_V=1$) along the \citet{Calzetti:2000} extinction
law. The black solid line indicates the quiescent/star-forming classification line
from \citet{Williams:2009}.}
\label{fig10}
\end{figure}
\subsection{Star-forming main sequence}\label{s3.3}
We estimate the SFR$_\mathrm{H\alpha}$ of HAEs from the observed {\rm\,H$\alpha$}\ luminosities
(\S\ref{s2.4.2}) using the same method as \citet{Shimakawa:2018}. We use the
\citet{Kennicutt:1998} prescription on the assumption of \citet{Chabrier:2003} IMF (a
factor of 1.7 reduces SFR from the standard \citealt{Kennicutt:1998} calibration). We
then correct for dust extinction of the {\rm\,H$\alpha$}\ line based on the SED-inferred stellar
extinction ($A_V$) from the {\sc fast} code (\S\ref{s2.4.1}) and the
\citet{Calzetti:2000} extinction law. Throughout the analysis, we used the additional
assumption of E($B-V$)$_\mathrm{stellar}$ = E($B-V$)$_\mathrm{nebular}$. Such a
hypothesis is relatively reasonable for typical star-forming galaxies at $z\sim2$
\citep{Reddy:2015}. However, \citet{Reddy:2015} and \citet{Price:2014} also note that
the ratio of nebular extinction to stellar extinction depends on galaxy properties,
especially SFR; our assumption, therefore, would lead to underestimating the SFRs of
active and dusty star-forming objects. Previous infrared studies also suggest that the
dust extinction law for individual galaxies varies depending on their IR luminosities
and dust geometries \citep{Reddy:2006,Reddy:2010,Casey:2014,Narayanan:2018}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{fig11.pdf}
\caption{Stellar mass versus dust-corrected SFR$_\mathrm{H\alpha}$ of HAEs. Black
filled and open squares indicate HAEs and HAE candidates respectively
(table~\ref{tab2}). The error-bars correspond to $1\sigma$ errors. Grey circles show
HAEs in the USS~1558 protocluster at $z=2.53$ reported by \citet{Shimakawa:2018} which
estimate SFR$_\mathrm{H\alpha}$ in the same way. Purple crosses represent HAEs
overlapping with the Chandra X-ray sources whose observed {\rm\,H$\alpha$}\ fluxes may be
overestimated since we do not consider the flux contribution from AGNs. Grey
dot-dashed and dashed lines are the star-forming main sequence of HAEs at $z=2.2$ in
the general field \citep{Oteo:2015} and star-forming galaxies at $z\sim2$ reported by
\citet{Daddi:2007a}, respectively. Pink diamonds show SFRs derived from UV and IR
luminosities for MIPS/24$\mu$m detected HAEs (see text).}
\label{fig11}
\end{figure}
Indeed, we confirm that SFR$_\mathrm{UV+IR}$ of dusty HAE sources with Spitzer/MIPS
24 $\mu$m detections is significantly higher than those dust-corrected
SFR$_\mathrm{H\alpha}$ with our tentative extinction correction (fig.~\ref{fig11}).
Their IR luminosities are derived from the flux densities at 24 $\mu$m by following
the \citet{Wuyts:2008} conversion prescription and then obtaining their SFRs via
SFR$_\mathrm{UV+IR}$ ($=1.09\times10^{-10}(L_\mathrm{IR}+3.3\nu L_{\nu,2800}$),
\citealt{Bell:2005}). We only measure IR luminosities of 10 HAEs without X-ray
detections (\#5,9,13,14,22,27,61,71,80,93). Two of them (\#27,93 but note \#27 is
blended with \#26,29 in the MIPS/24$\mu$m image) are associated with 870 $\mu$m LABOCA
sources (DKB12 and DKB15 in \citealt{Dannerbauer:2014}). According to the difference
between dust-corrected SFR$_\mathrm{H\alpha}$ and SFR$_\mathrm{UV+IR}$, we expect to
underestimate SFR$_\mathrm{H\alpha}$ by a factor of four on average for these dusty
HAEs. Such a large mismatch is the reason why we do not analyse the {\rm\,H$\alpha$}\ luminosity function in this paper, as discussed in the previous subsection.
The new deeper {\rm\,H$\alpha$}\ data succeed in finding relatively fainter HAEs than those found
by the previous MAHALO-Subaru survey \citep{Koyama:2013a}. As shown in fig.~\ref{fig11}, when
we apply the same method for extinction correction of the {\rm\,H$\alpha$}\ line in HAEs in PKS~1138
as used for those in USS~1558 \citep{Shimakawa:2018}, HAEs in PKS~1138 and USS~1558
seem to follow the same stellar mass -- SFR$_\mathrm{H\alpha}$ relation. We find no
clear trend in star formation with environment, which is in agreement with past studies
in the same field \citep{Koyama:2013a,Koyama:2013b}. However, finding no environmental
dependence on star formation in PKS~1138 is in contrast to USS~1558 where the HAEs in
the group regions are more actively star-forming than those in the intergroup regions
\citep{Shimakawa:2018}. Moreover, we find that three of the X-ray detected HAEs tend to
be positioned below the main sequence, although these SFR$_\mathrm{H\alpha}$ would be
contaminated by additional {\rm\,H$\alpha$}\ emission from AGNs. This is consistent with their
post-starburst like rest-$UVJ$ colours (\S\ref{s3.2}), though we should keep in mind
that our dust corrections mass estimates of X-ray hosts have
substantial uncertainties. One source showing the excess of star formation among HAEs
with X-ray emission is HAE-\#46 (M$_\star=10^{10.4}$ {\rm M}$_{\odot}$) that has a blue $U-V$
colour and \citet{Kurk:2004b} have identified that this HAE (\#215 in their literature)
has a broad {\rm\,H$\alpha$}\ line.
Four HAEs (\#42,48,54,56) are confirmed by ALMA CO($3-2$) observations (Tadaki et al.
in preparation). We cannot use these data to derive their dust-corrected SFRs more
robustly as they are clearly blended with other sources in the MIPS/24 $\mu$m band.
However, \#48,54 are more like dusty starbursts according to their rest-frame $UVJ$
colours (appendix~\ref{a3}).
Our deeper {\rm\,H$\alpha$}\ imaging increases the number of HAE sources by 40--50 percent compared
to our previous narrow-band imaging \citep{Koyama:2013a}. The mean narrow-band flux of
newly identified HAEs is 2.5 times fainter than the mean flux of the entire sample. We
discover eight HAEs at a stellar mass lower than $10^{9}$ {\rm M}$_{\odot}$\ similar to those
found in USS~1558 at $z=2.5$ \citep{Hayashi:2016}. The larger number of lower-mass
HAEs in USS~1558 (fig.~\ref{fig11}) should be due to its 0.5 mag deeper $K_s$ data than
that in PKS~1138. We defer the detailed analyses of these samples to future work since
it remains unclear whether they are truly HAEs associated with the PKS~1138
protocluster, due to a lack of firm detections at multi photometric bands. Future
follow-up deep spectroscopy can provide us with both robust confirmations of these
faint sources, and improved completeness when we investigate gas-phase metallicities
of low-mass HAEs as compared to the previous spectroscopic analyses by
\citet{Shimakawa:2015}.
\subsection{The Spiderweb nebula}\label{s3.4}
The previous {\rm\,H$\alpha$}\ line observations reported that the Spiderweb radio galaxy is
associated with an extended {\rm\,H$\alpha$}\ nebula over a few ten ph-kpc in radius
\citep{Kurk:2002,Nesvadba:2006,Kuiper:2011,Koyama:2013a}. \citet{Kurk:2002} have
derived {\rm\,Ly$\alpha$}/{\rm\,H$\alpha$}\ line ratios in different regions within the {\rm\,H$\alpha$}\ ({\rm\,Ly$\alpha$}) nebula.
These vary between 0.015 (at the nucleus) and 7.6 (at the position blending with the
radio jet). \citet{Nesvadba:2006} have studied the spatial distribution and kinematics
of the optical emission lines in detail by using the near-infrared integral field
spectrograph (VLT/SPIFFI). Their emission line analyses suggest that the Spiderweb
nebula has electron density $n_e\sim388_{-148}^{+182}$ cm$^{-3}$, broad-line emission
with FWHM$>2000$ km~s$^{-1}$, typical {\rm\,[N{\sc ii}]}/{\rm\,H$\alpha$}\ line flux ratio of $\sim1$, and the
total observed {\rm\,H$\alpha$}\ luminosity of $L_\mathrm{H\alpha}=(14.8\pm1.2)\times10^{43}$
ergs~s$^{-1}$. They roughly estimate the least total ionised gas mass of
M$_\mathrm{H\textsc{ii}}=$ (2.3--6.5)$\times10^9$ {\rm M}$_{\odot}$. Such a relatively high {\rm\,H$\alpha$}\
emission brightness cannot be explained by shock heating alone \citep{Dopita:1996}, so
\citet{Nesvadba:2006} conclude that photoionisation from the radio galaxy plays a
dominant role in the emission line properties of this large {\rm\,H$\alpha$}\ nebula (see also
\citealt{Martin:2003,Nesvadba:2008,Nesvadba:2017}).
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{fig12.pdf}
\caption{Multi-faces of the Spiderweb galaxy (20 arcsec on a side in each image).
(a) The left panel shows NB$_{2071}$ image (grey scale) and line-subtracted $K_s$
image (orange contour). The first contour is 1.5 sigma background rms.
(b) The middle panel is an RGB image from F475W and F814W photometry with HST. White
contours are based on the NB$_{2071}$ image, and the first contour corresponds to 1.5
rms in the background.
(c) The right panel represents the surface brightness of narrow-band flux
(SB$_\mathrm{NB}$) where the continuum is subtracted (eq.~\ref{eq3}). 1 sigma error is
$3.2\times10^{-17}$ erg~s$^{-1}$cm$^{-2}$arcsec$^{-2}$ for a given pixel$^2$ area.}
\label{fig12}
\end{figure*}
Our deep {\rm\,H$\alpha$}\ narrow-band imaging can trace extended {\rm\,H$\alpha$}\ (+{\rm\,[N{\sc ii}]}) emission over the
wider FoV, down to $4.8\times10^{-17}$ erg~s$^{-1}$cm$^{-2}$arcsec$^{-2}$ above the
1.5 sigma level in pixel scale. As a result, we succeed in identifying the large {\rm\,H$\alpha$}\
nebula structure that has a maximum projected extent of $\sim100$ ph-kpc
(fig.~\ref{fig12}). Such an enormous {\rm\,H$\alpha$}\ structure has never been discovered before
at $z>2$. The {\rm\,H$\alpha$}\ nebula extends in the NE--SW direction, which is broadly consistent
with the inclination of structures in X-ray and CO($1-0$) emission found by
\citet{Carilli:2002} and \citet{Emonts:2016}, respectively.
When we adopt the conservative assumption of {\rm\,[N{\sc ii}]}/{\rm\,H$\alpha$}=1 in the same manner as
\citet{Nesvadba:2006}, we obtain the total observed {\rm\,H$\alpha$}\ luminosity
$L_\mathrm{H\alpha}=(1.35\pm0.15)\times10^{44}$ ergs~s$^{-1}$ over 50 ph-kpc in
radius. This value is consistent with the past result by \citet{Nesvadba:2006}. We
should note that the obtained {\rm\,H$\alpha$}\ luminosity is a lower limit since dust correction is
not implemented and the bandwidth of the narrow-band filter does not fully cover the
broad-line emission as identified by \citet{Nesvadba:2006}. These complicated issues
prevent us from exploring this unique nebula in more detail. We experimentally carry
out some further analyses with these caveats by comparing with HST/ACS images, which
are shown in appendix~\ref{a4}.
Finally, we inspected the other HAEs in PKS~1138 searching for significantly extended
{\rm\,H$\alpha$}\ structures based on the median stacking with the {\sc iraf} scripts. Our stacking
procedure reaches down to $6.4\times10^{-18}$ erg~s$^{-1}$cm$^{-2}$arcsec$^{-2}$ with
a two sigma confidence level. However, we do not identify such diffuse nebula beyond a
radii of 10 ph-kpc in the other HAEs, even after stacking all HAE samples.
\section{Discussion}\label{s4}
High-redshift protoclusters provide us with a great opportunity to unveil the
formation histories of massive galaxy clusters and their member galaxies in the local
Universe. As claimed by past studies
\citep{Kurk:2004a,Kodama:2007,Doherty:2010,Hatch:2011,Galametz:2012,Koyama:2013a,Tanaka:2013},
PKS~1138 at $z=2.2$ is considered to be one of the most massive protoclusters at
$z\sim2$, with a significant number excess ($\delta\gtrsim10$) of massive red
star-forming and passive galaxies over the MOIRCS FoV relative to the general field at
a similar redshift. This suggests that the massive galaxies in PKS~1138 are in the
maturing phase, and such red galaxies would provide us with direct insights into the
quenching processes of bright red sequence objects seen in the present-day clusters of
galaxies.
With such expectations, the original motivation of this work was to characterise these
massive HAEs. We especially focus on X-ray fraction and the rest-frame $UVJ$ colours of
HAEs based on the Chandra X-ray data and multi-band images, both of which mostly cover
the survey area. We first derive the stellar mass function and confirm the number
excess of massive HAEs seen in the past studies
\citep{Hatch:2011,Koyama:2013a,Koyama:2013b}. The characteristic stellar mass is
comparable to that in dense group cores seen in the USS~1558 protocluster
\citep{Shimakawa:2018}. However, PKS~1138 is associated with a larger number of massive
red HAEs than USS~1558 \citep{Kodama:2007,Galametz:2012}, suggesting that PKS~1138 is a
more developed system. In addition, considering the fact that the PKS~1138 protocluster
hosts at least three passive galaxies in such a small field
(\citealt{Doherty:2010,Tanaka:2013}, see also \citealt{Zirm:2008}), PKS~1138 is ideally
placed to see the rapid transition from massive dusty star-forming galaxies to
quiescent galaxies.
We find that four out of the six very massive HAEs (M$_\star=10^{11-12.5}$ {\rm M}$_{\odot}$),
including the Spiderweb galaxy, host bright X-ray sources ($L_X\gtrsim10^{44}$
erg~s$^{-1}$, see also \citealt{Pentericci:2002}). In general fields at the similar
redshift, \citet{Sobral:2016b} and \citet{Matthee:2017} have shown that the AGN
fraction significantly increases with increase in {\rm\,H$\alpha$}\ luminosities of HAEs, and
reaches $\sim100$ percent at the bright end ($>10^{43.5}$
erg~s$^{-1}$ without dust correction). Similarly, we confirm that the top two HAEs
with high observed narrow-band luminosities ($=10^{43-43.5}$ erg~s$^{-1}$ without
dust correction) and the Spiderweb radio galaxy ($=10^{44}$ erg~s$^{-1}$) host X-ray
counterparts. The fraction is about two times higher than those in general fields
reported by them, although our sample size is quite small. Such a relatively fair
comparison seems to strength our finding of the enhancement of AGN fraction in the
protocluster region.
Intriguingly, they tend to have unique rest-frame $UVJ$ colours corresponding to the
post-starburst sequence between star-forming galaxies and passive populations
\citep{Williams:2009,Whitaker:2011}.
Thus, AGNs may play an important role in suppressing the star formation and the
build-up of massive quiescent galaxies since these massive HAEs are more likely to grow
into the bright-end red sequence systems seen in the massive galaxy clusters in the
local Universe. AGN feedback is indeed considered to be an important mechanism to
suppress the star formation of massive systems as pointed out by previous work (e.g.,
\citealt{Springel:2005,Croton:2006,Bower:2006,Somerville:2008,Voort:2011b,Page:2012,Kormendy:2013,Cicone:2014,Genzel:2014b}).
If a significant number of massive forming galaxies are subject to the influence of
such strong feedback, we could explain the rapidly declining star formation around the
distant cluster centres since $z\sim2$
\citep{Smail:2014,Shimakawa:2014,Clements:2014,Kato:2016}.
We should note that the derived AGN fraction among HAEs is the minimum fraction
inferred from the Chandra X-ray data. Our previous spectroscopic analyses
\citep{Shimakawa:2015} indicate that other three sources (\#14,30,54) might also be
AGNs because of their high {\rm\,[N{\sc ii}]}/{\rm\,H$\alpha$}\ line ratios ($>0.5$), although the spectral data
are obtained with low spectral resolution (R=513) and the derived line ratios have
substantial uncertainties ($\sim0.2$ dex). If we assume that these HAEs also host AGNs,
the AGN fractions increase to 43 and 80 percent at the stellar mass bins of
$10^{10.5-11}$ and $10^{11-11.5}$ {\rm M}$_{\odot}$\ respectively.
Such a large number of AGN host galaxies may affect the derivation of the stellar mass
function, since their stellar mass estimates may be overestimated due to the flux
contamination from the nuclear emission. In the meanwhile, the surprisingly high AGN fraction
implies that massive star-forming galaxies in PKS~1138 may mostly host AGNs and could
have a considerable impact on the current galaxy evolution paradigm.
The most energetic source present is the Spiderweb radio galaxy. We identified the
enormous {\rm\,H$\alpha$}\ structure extending over 100 ph-kpc associated with this massive
system. The spatial extent broadly agrees with previous findings which have reported
an extended component in the {\rm\,Ly$\alpha$}\ line \citep{Gopal:2000,Kurk:2002}, X-ray
\citep{Carilli:2002}, UV \citep{Pentericci:1998,Hatch:2008}, and the CO($1-0$) line
\citep{Emonts:2016}. The Spiderweb galaxy is also known to be a composite of active
star formation and AGN \citep{Ogle:2012,Drouart:2014}. The starburst component has
SFR of $\sim1400$ {\rm M}$_{\odot}$~yr$^{-1}$ \citep{Seymour:2012,Rawlings:2013,Dannerbauer:2014},
supported by a rich gas reservoir of $\sim6\times10^{10}$ {\rm M}$_{\odot}$\
\citep{Emonts:2013,Emonts:2016,Gullberg:2016}. This monster galaxy is expected to grow
into a brightest cluster galaxy, as seen in the local Universe \citep{Hatch:2009}. The
large {\rm\,H$\alpha$}\ nebula could show the occurrence of a pre-heating event in the
proto-intercluster medium \citep{Babul:2002,Dubois:2011,Dubois:2012,Valentino:2016},
though there is still much debate about heating mechanisms of the intercluster medium
at high redshifts \citep{McNamara:2007,Kravtsov:2012}.
Lastly, the very high X-ray fraction in massive HAEs cautions the possibility that the
narrow-band HAE selection may overestimate an AGN fraction, especially at the massive
end. For example, \citet{Sobral:2016b} have shown that AGNs typically contribute 15
percent of the total {\rm\,H$\alpha$}\ luminosity. Moreover, AGNs are expected to enhance {\rm\,[N{\sc ii}]}\ line
flux \citep{Baldwin:1981,Veilleux:1987} and this can additionally increase the
narrow-band flux and EW$_\mathrm{NB}$. For instance, for a given {\rm\,H$\alpha$}\ luminosity, an
increase of log {\rm\,[N{\sc ii}]}/{\rm\,H$\alpha$}\ ratio from $-0.5$ to $0$ raises EW$_\mathrm{NB}$ by 0.18 dex.
These contaminations may cause the serious selection bias in the sense that the
narrow-band selection preferentially detect AGN host HAEs. In particular, our current
analyses cannot rule out the possibility that we are missing massive dusty starburst
populations which would have faint observed {\rm\,H$\alpha$}\ luminosities and low EW$_\mathrm{NB}$
due to heavy dust obscuration. Even if we miss a few dust-obscured objects, this factor
can substantially decrease the bright-end AGN fraction. A wide-field spectroscopic
search at the IR to radio regime with e.g., ALMA and JWST is highly desirable to
clarify this selection issue.
\section{Summary}\label{s5}
In this second instalment of the MDCS campaign, we investigate HAEs associated with
one of the most studied protoclusters, PKS~1138 at $z=2.2$. Using the advanced dataset
we construct samples of HAEs associated with the Spiderweb protocluster; these consist
of 68 HAEs (36 confirmed with spec-$z$ and 32 selected by $Bz'K_s$ colour) and 13 HAE
candidates. 17 and 9 objects amongst them are newly-discovered HAEs and HAE candidates
by this work, respectively. The online catalogue (appendix~\ref{a3}) lists coordinates,
confirmation status, physical properties, and rest-frame $UVJ$ colours of the HAE
samples. The major findings are summarised as follows.
\begin{description}
\item[---]
We investigate the stellar mass distribution function of HAEs in PKS~1138, including
the completeness correction derived using a Monte Carlo simulation. We then identify
the high cut-off stellar mass of log(M$_\star$/{\rm M}$_{\odot}$) $=11.73\pm0.76$ and find that
the number density is about ten times higher over the survey area than that in the
blank field. On the other hand, we find that at least four out of six very massive HAEs
have X-ray counterparts, and those stellar masses may be significantly overestimated.
This uncertainty would especially affect the derivation of the characteristic stellar mass.
The Spiderweb nebula is unique and shows an extensive structure over 100 ph-kpc in
the {\rm\,H$\alpha$}\ emission line, as well as in other tracers such as {\rm\,Ly$\alpha$}, X-ray, UV, and CO($1-0$)
lines. We also tentatively obtain a passive fraction of 36 percent by combining the HAE
sample with DRG$_\mathrm{nIR}$, although the measurement has a substantial error.
\item[---]
We investigate the rest-frame $UVJ$ colours of HAEs and their environmental dependence
on the local scale. Given the limited survey field, we do not see any clear correlation
between colours and local overdensities. On the other hand, HAEs with luminous X-ray
emission tend to have post-starburst like colours, implying that bright AGNs may play
an important role in quenching active star formation in these sources. Because the
Spiderweb protocluster is also associated with a large number of dusty starbursts and
passive galaxies, many massive galaxies in PKS~1138 seem to be undergoing a rapid
transition from the active star formation phase into the quiescent mode.
\end{description}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{fig13.pdf}
\caption{A speculative schematic of formation and evolution histories of galaxy
clusters and galaxies therein, based on our findings through the MDCS. We tentatively
assume that PKS~1138 at $z=2.2$ is a maturing protocluster (middle) and USS~1558 at
$z=2.5$ is a growing protocluster (right), respectively, judging form substructures and
characteristics of member galaxies. The left one represents the classical massive
cluster of galaxies associated with a diffuse X-ray source. Galaxy colours are trying
to express that the star formation declining and red sequence formation in the centre
of galaxy clusters towards the local Universe. Yellow star symbols show energetic AGN
activities. A table at the bottom highlights phenomena that are expected to occur in
galaxies and (proto-)inter-cluster medium (see text).}
\label{fig13}
\end{figure*}
Combined with our previous paper \citep{Shimakawa:2018}, we here summarise the results
that we obtained through the MDCS campaign. We have explored two protoclusters,
USS~1558 ($z=2.53$) and PKS~1138 ($z=2.15$) with the deep narrow-band imaging on the
Subaru Telescope ($NB_{3\sigma}\sim24$ mag).
The former target, USS~1558 is thought to be a young protocluster composed of
fragmented dense groups as seen at higher redshifts
\citep{Umehata:2015,Oteo:2017,Miller:2018}. \citet{Shimakawa:2018} have found that
SFRs of HAEs in dense group cores are statistically higher than those in inter-group
regions. Also, local overdensities show the higher cut-off stellar mass in the stellar
mass function than the inter-group. These trends suggest that galaxy formation is more
enhanced and accelerated in the fragmented groups in a growing phase of hierarchical
growth of galaxy clusters (see also \citealt{Wang:2016,Oteo:2017}). Such an active
star formation could be supported by a rich cold gas reservoir (Tadaki et al. in
preparation) fed by vigorous cold accretion (e.g.,
\citealt{Dekel:2006,Dekel:2009,Dekel:2009b,Keres:2009,Faucher:2010,Voort:2011}).
Another companion paper \citep{Shimakawa:2017b} showing the {\rm\,Ly$\alpha$}\ photon depletion in
the dense group regions could support this scenario.
PKS~1138 is thought to be a more advanced system because of the significant excess of
bright red galaxies \citep{Kurk:2004a,Kodama:2007,Doherty:2010,Galametz:2012}. This
work finds an enormous {\rm\,H$\alpha$}\ structure associated with the Spiderweb radio galaxy and a
high AGN fraction in massive HAEs. Such high energy sources could contribute
pre-heating in the proto-inter-cluster medium \citep{Valentino:2016}, though heating
mechanisms of inter-cluster medium in an early phase are highly uncertain
observationally. Rest-frame $UVJ$ colours of X-ray HAEs imply that they would be in a
post-starburst phase, suggesting the importance of the role of AGN as a driver of star
formation quenching. These results, together with previous work
\citep{Doherty:2010,Tanaka:2013,Koyama:2013b}, all point to PKS~1138 possibly being the
sweet spot where maturing galaxies are undergoing a rapid transition from dusty
starbursts to quenching populations.
At lower redshifts ($z\lesssim2$), member galaxies in the hot inter-cluster medium
enriched by superheated plasma would no longer retain their star formation because of
insufficient gas accretion
\citep{Keres:2005,Sijacki:2007,Schaye:2010,McCarthy:2011,Haines:2013,Hughes:2013,Jaffe:2015,Jaffe:2016,Bianconi:2016,Hayashi:2017}.
Additionally, less massive galaxies falling into the clusters may experience
environmental quenching as has previously been confirmed at $z<2$ (e.g.,
\citealt{Bamford:2009,Peng:2010,Smith:2012,Raichoor:2012,Muzzin:2012,Gobat:2015,Brown:2017}).
Figure~\ref{fig13} briefly summarises these speculations based on the MDCS campaign. We
strongly caution that the arguments discussed above are based on only two protoclusters
and must therefore be viewed with caution. Future work is necessary to properly
comprehend the systematic biases and appropriately understand the diversity of high-$z$
protoclusters, as mentioned in the following section.
\section{Future prospects: Diverse protoclusters}\label{s6}
Our MDCS program enables a high sampling density, down to a certain flux limit, with
the deep narrow-band imaging: a unique advantage over the other common approaches such
as photo-$z$ and the Balmer/Lyman Break technique. On the other hand, we targeted only
two known protoclusters at $z>2$ which were initially discovered by surveys for radio
galaxy environments \citep{Kurk:2000,Kajisawa:2006}. We thus should note that this
work has not only a small sample size of protoclusters but also an unknown bias
towards the environments associated with radio-loud galaxies. In particular, we do not
yet know how the sampling effects affect the properties of member galaxies.
In recent years, many intensive protocluster-search survey projects have been
launched. These utilise continuous strong absorption in the {\rm\,Ly$\alpha$}\ forest
\citep{Cai:2016}, Lyman Break features \citep{Toshikawa:2017}, radio-loud galaxies
\citep{Noirot:2018} and dust emission \citep{Planck:2015,Greenslade:2018} as
signposts. Whilst these surveys will each construct statistical protocluster
catalogues, the sampling effects remain to be resolved. Future work will be needed to
understand such selection biases and categorise individual protoclusters with, e.g.,
X-ray properties \citep{Gobat:2011,Valentino:2016,Wang:2016} and {\rm\,H{\sc i}}\ tomography
\citep{Lee:2016}. The upcoming eROSITA \citep{Merloni:2012}, a successor to the
ASTRO-H (Hitomi, \citealt{Takahashi:2016}), and the Prime Focus Spectrograph on the
Subaru Telescope \citep{Tamura:2016} will possess the exceptional capability needed to
reveal these unique objects.
\section*{Acknowledgements}
RS thanks Emma Rigby for proofreading the manuscript.
We thank the anonymous referee for useful comments.
The data are collected at the Subaru Telescope, which is operated by the National
Astronomical Observatory of Japan, and also based on observations made with the
NASA/ESA Hubble Space Telescope, and obtained from the Hubble Legacy Archive, which
is a collaboration between the Space Telescope Science Institute (STScI/NASA), the
Space Telescope European Coordinating Facility (ST-ECF/ESA) and the Canadian
Astronomy Data Centre (CADC/NRC/CSA). We would like to thank the Subaru staff for
their support throughout all observing and analysing processes.
A part of analyses is conducted with the assistance of the Tool for OPerations on
Catalogues And Tables ({\sc topcat}; \citealt{Taylor:2015}). This work gains the
benefit from the 3D-HST Treasury Program (GO 12177 and 12328) with NASA/ESA HST,
which is operated by the Association of Universities for Research in Astronomy,
Inc., under NASA contract NAS5-26555. This research has made use of data obtained
from the Chandra Source Catalogue, provided by the Chandra X-ray Centre (CXC) as
part of the Chandra Data Archive, and also use of the NASA/ IPAC Infrared Science
Archive, which is operated by the Jet Propulsion Laboratory, California Institute of
Technology, under contract with the National Aeronautics and Space Administration.
R.S. acknowledges the support from the Japan Society for the Promotion of Science
(JSPS) through JSPS overseas research fellowships. T.K. acknowledges KAKENHI No.
21340045. HD acknowledges financial support from the Spanish Ministry of Economy and
Competitiveness (MINECO) under the 2014 Ram\'on y Cajal program MINECO RYC-2014-15686.
We wish to recognize and acknowledge the very significant cultural role and reverence
that the summit of Maunakea has always had within the indigenous Hawaiian community.
We are most fortunate to have the opportunity to conduct observations from this
mountain.
\bibliographystyle{mnras}
|
\section{Motivation}
One of the fundamental questions in medical practice is how diseases progress in individual patients. Accurate continuous monitoring of a patient's condition could considerably improve prevention and treatment. Many medical tests, such as x-ray, MRI, motion capture gait analysis, biopsy, or blood tests are costly, harmful or inconvenient to perform frequently. Since in many situations increase in sampling is not feasible due to inconvenience and costs, practitioners need to reach out for statistical tools to analyze the dynamics of these measurements.
While in many situations multiple data points from patients' histories are available, these data are often underutilized. For the sake of simplicity and convenience, many prognostic models applied in practice only use the last few observations from a series or summary statistics such as the mean over time. However, this simplification ignores important information about the progression, including its dynamics or individual patient's variability. Moreover, the noise inherent to measurements further hinders the inference of changes in patient's health. For making use of these data, practitioners need statistical models and software. To enable appropriate usage and broad adoption these tools should be simple to use and understand.
To illustrate the potential benefits of temporal models in these scenarios, in Figure \ref{fig:motivation} we present measurements of gait pathology progression. The thick solid line represents the estimated mean and indicates a clear trend of growth during puberty. However, by looking at individual processes and by modeling between-subject similarities, we may model the progression more accurately. This approach could result in personalized predictions, new clustering techniques patients based on progression patterns, or extraction of latent representation of progression patterns, which then could be used as predictors in other models.
\begin{figure}[h]
\includegraphics[width=0.49\linewidth]{points}
\includegraphics[width=0.49\linewidth]{grouped}
\caption{We observe a Gait Deviation Index (GDI), a holistic metric of motor impairment, at every visit (left plot) and we derive the population progression (the thick solid curve) using a locally weighted scatterplot smoothing method. The analysis of individual patients (connected dots in the right plot, also differing by color) can reveal patterns in individual trends.}
\label{fig:motivation}
\end{figure}
This kind of data has been commonly analyzed using linear mixed models, where we treat time as a random effect and nonlinearity of this effect is imposed by choice of the functional basis \citep{zeger1988models, verbeke1997linear, mcculloch2001generalized}. When data is very sparse, additional constraints on the covariance structure of trajectories are derived using cross-validation or information criteria \citep{rice2001nonparametric,bigelow2009bayesian}. To further reduce the dimension, practitioners model the covariance as a low-rank matrix \citep{james2000principal,berkey1983longitudinal, yan2017dynamic, hall2006properties, besse1986principal, yao2006penalized, greven2011longitudinal}. Multiple models were developed for incorporating additional covariates \citep{song2002semiparametric, liu2009joint, rizopoulos2014combining}. While statisticians use these methods in practice, they need to fine-tune them each time since the nature of processes differs in each clinical setting. Moreover, the probabilistic formulation of the model and dependence on the underlying distributions might hinder applicability or adoption to other practical problems.
In this work, we propose a flexible and straightforward statistical framework, exploiting matrix factorization techniques.
We focus on the simplicity of the formulation, and we implement software easy to use and extend.
\section{Background and related work}\label{s:background}
In many clinical settings, researchers and practitioners model the patient's history as a multivariate process of clinical measurements. Examples include variables such as the size of a tumor, blood markers, height and weight of a patient. The multivariate measurements are noisy, and they are observed on different time-points. This situation arises when, for example, clinicians perform a blood test at every visit but a biopsy sporadically.
Even though multiple variables are measured, the ones directly related to the disease, such as the size of a tumor, are usually of the special interest. Therefore, to focus our attention and aid further discussion, we start by introducing notation and methodology for the univariate case. Let $N$ denote the number of subjects. For each individual $i \in \{ 1,2,...,N \}$, we measure the process at $n_i$ irregularly sampled time-points $\mathbf{t}_i = [t_{i,1},t_{i,2},...,t_{i,n_i}]'$. We assume that for each $i$ we have $t_{min} < t_{i,1} < t_{i,2} < ... < t_{i,n_i} < t_{max}$, for some $t_{min},t_{max}$.
Let $\mathbf{y}_i = [y_{i,1,},...,y_{i,n_i}]'$ denote observations corresponding to $\mathbf{t}_i$ for each individual $i$. To model individual trajectories given pairs $(\mathbf{t}_i,\mathbf{y}_i)$ practitioners map observations into a low-dimensional space which represents progression patterns. Ideally, a small distance between individuals in the latent space reflects similar progression patterns.
In this section, we discuss state-of-the-art approaches to estimating this low-dimensional latent embedding. We classify them into three categories: the direct approach, mixed-effect models, and low-rank approximations.
\subsection{Direct approach}\label{ss:direct}
If the set of observed values for each individual is dense, elementary interpolation using a continuous basis can be sufficient for approximating the entire trajectory. Let $\{b_i: i \in \mathbb{N} \}$ be a basis of $L_2([t_{\min},t_{\max}])$. In practice, we truncate the basis to a set of the first $K \in \mathbb{N}$ basis elements. Let $\mathbf{b}(t) = [b_1(t),b_2(t),...,b_K(t)]'$ be a vector of $K$ basis elements evaluated at a timepoint $t \in (t_{\min},t_{\max})$. Throughout this article we use the word \emph{basis} to refer to some truncated basis of $K$ elements.
To find an individual trajectory, for each subject $i \in \{ 1,...,N \}$, we may use least squares method and estimate a set of coefficients $\mathbf{w}_i \in \mathbb{R}^K$, minimizing squared euclidean distance to the observed point
\begin{align}\label{eq:direct-individual}
\argmin_{\mathbf{w}_i}\sum_{j=1}^{n_i}\left|y_{i,j} - \mathbf{w}_i'\mathbf{b}(t_{i,j})\right|^2.
\end{align}
Finally, to reduce overfitting, it is often convenient to compute principal functional components across all individuals and represent curves in the space spanned by the first few of them. Such representation, referred to as a {\it Karhunen-Lo\`eve} expansion \citep{watanabe1965karhunen,kosambi2016statistics}, has became a foundation of many functional data analysis workflows \citep{ramsay1991some,yao2005linear,cnaan1997tutorial,laird1988missing,horvath2012inference,besse1997simultaneous}.
This approach to modeling covariance has two main drawbacks. First, if the number of observations $n_i$ for an individual $i$ is smaller or equal to the size of the basis $K$, we can fit a curve with no error leading to overfitting and unreliable estimator of the variance. Second, this approach ignores similarities between the curves, which could potentially improve the fit.
A basic idea to remedy these issues is to estimate both the basis coefficients and the variance structure simultaneously. Linear mixed-effect models provide a convenient solution to this problem.
\subsection{Linear mixed-effect models}\label{ss:lmm}
A common approach to modeling longitudinal observation $(\mathbf{t}_i, \mathbf{y}_i)$ is to assume that data come from a linear mixed-effect model (LMM) \citep{verbeke1997linear, zeger1988models}. We operate in a functional space with a basis $\mathbf{b}(t)$ of $K$ elements and we assume there exists a \emph{fixed effect} $\mu(t) = m' \mathbf{b}(t)$, where $m = [m_1,...,m_K]'$ for $m_i \in \mathbb{R}$ for $1 \leq i \leq K$. We model the individual \emph{random effect} as a vector of basis coefficients. In the simplest form, we assume
\begin{align}\label{eq:latent-probabilistic}
\mathbf{w}_i \sim \mathcal{N}(0, \Sigma),
\end{align}
where $\Sigma$ is a $K \times K$ covariance matrix. We model individual observations as
\begin{align}\label{eq:probabilistic}
\mathbf{y}_i|\mathbf{w}_i \sim \mathcal{N}(\mu_i + B_i\mathbf{w}_i, \sigma^2I_{n_i}),
\end{align}
where $\mu_i = [\mu(t_{i,1}),\mu(t_{i,2}),...,\mu(t_{i,n_i})]'$, $\sigma$ is the standard deviation of observation error and $B_i = [\mathbf{b}(t_{i,1}),...,\mathbf{b}(t_{i,n_i})]'$ is the basis evaluated at timepoints defined in $\mathbf{t}_i$. Estimation of the model parameters is typically accomplished by the expectation-maximization (EM) algorithm \citep{laird1982random}. For estimating coefficients $\mathbf{w}_i$ one can use the best unbiased linear predictor (BLUP) \citep{henderson1950estimation,robinson1991blup}.
Since the LMM estimates the covariance structure and the individual fit simultaneously, it reduces the problem of uncertain estimates of $\mathbf{w}_i$, present in the direct approach. However, this model is only applicable if we observe a relatively large number of observations per subject since we attempt to estimate $K$ coefficients for every subject.
To model trajectories from a small number of observations, practitioners further constrain the covariance structure. If we knew the functions which contribute the most to the random effect, we could fit a LMM in a smaller space spanned by these functions. We explore possibilities to learn the basis from the data using low-rank approximations of the covariance matrix.
\subsection{Low-rank approximations}\label{ss:reduced-rank}
There are multiple ways to constrain the covariance structure. We can use cross-validation or information criteria to choose the best basis, the number of elements or positions of spline knots \citep{rice2001nonparametric,bigelow2009bayesian}. Alternatively, we can place a prior on the covariance matrix \citep{maclehose2009nonparametric}.
Another solution is to restrict the latent space to $q < K$ dimensions and learn from the data the mapping $A \in \mathbb{R}^{K \times q}$ between the latent space and the basis. In the simplest scenario with Gaussian observation noise, observations can be modeled as
\begin{align}\label{eq:james-model}
\mathbf{y}_i | \mathbf{w}_i \sim \mathcal{N}(\mu_i + B_i A \mathbf{w}_i, \sigma^2I_{n_i}),
\end{align}
following the notation from \eqref{eq:probabilistic}.
\citet{james2000principal} propose an EM algorithm for finding model parameters and latent variables $\mathbf{w}_i \in \mathbb{R}^q$ in \eqref{eq:james-model}. In the expectation stage, they marginalize $\mathbf{w}_i$, while
in the maximization stage, with $\mathbf{w}_i$ assumed observed, they maximize the likelihood with respect to $\{\mu,A,\sigma\}$. The maximum likelihood, given $\mathbf{w}_i$, takes form
\begin{align*}
\prod_{i=1}^N \frac{1}{(2\pi)^{n_i/2} \sigma^{n_i} |\Sigma|^{1/2}} \exp\{ &-(\mathbf{y}_i - \mu_i - B_i A \mathbf{w}_i)'(\mathbf{y}_i - \mu_i - B_i A \mathbf{w}_i) / 2\sigma^2 \nonumber\\
&- \frac{1}{2}\mathbf{w}_i' \Sigma^{-1} \mathbf{w} \}.\label{eq:likelihood}
\end{align*}
Another approach to estimating parameters of \eqref{eq:james-model} is to optimize over $\mathbf{w}_i$ and marginalize $A$ \citep{lawrence2004gaussian}. This approach allows modifying the distance measure in the latent space, using the \emph{kernel trick} \citep{schulam2016disease}.
Estimation of the covariance structure of processes is central to the estimation of individual trajectories. \citet{descary2016functional} propose a method where the estimate of the covariance matrix is obtained through matrix completion.
Methods based on low-rank approximations are widely adopted and applied in practice \citep{berkey1983longitudinal, yan2017dynamic, hall2006properties, besse1986principal, yao2006penalized, greven2011longitudinal}. However, due to their probabilistic formulation and reliance on the distribution assumptions, these models usually need to be carefully fine-tuned for specific situations. This shortcoming motivates us to develop an elementary optimization framework.
\section{Modeling sparse longitudinal processes using matrix completion}\label{s:context}
The mixed-effect model, like any other probabilistic model, can be heavily biased when data comes from a distribution considerably different than assumed. In the medical context, since biomarkers can differ in every clinical setting, fine-tuning the models may require an extensive amount of time and expertise. In this work, we develop a more flexible approach based solely on the $\ell_2$ approximation rather than the underlying probabilistic distributions.
We pose the problem of trajectory prediction as a matrix completion problem, and we solve it using sparse matrix factorization techniques \citep{rennie2005fast, candes2009exact}. In the classical matrix completion problem, the objective is to predict elements of a sparsely observed matrix using its known elements while minimizing a specific criterion, often chosen to be the Mean Squared Error (MSE). The motivating example is the ``Netflix Prize'' competition \citep{bennett2007netflix}, where participants were tasked to predict unknown movie ratings using other observed ratings. We can represent these data as a matrix of $N$ users and $M$ movies, with a subset of known elements, measured on a fixed scale, e.g., $1-5$.
To solve the matrix completion problem, we usually assume that the true matrix can be approximated by a low-rank matrix \citep{srebro2005generalization}. In the low-rank representation $WA'$ columns of $A$ spanning the space of movies can be interpreted as ``genre'' components, and each user is represented as a weighted sum of their preferred genres, i.e., a row in the matrix of latent variables $W$ (see Figure \ref{fig:idea}).
We can use the same idea to predict sparsely sampled curves, as long as we introduce an additional smoothing step. The low-dimensional latent structure now corresponds to progression patterns, and a trajectory of each individual can be represented as a weighted sum of these ``principal'' patterns. In Figure \ref{fig:idea}, the patterns are given by $A'B'$, while the individual weights are encoded in $W$.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\linewidth]{intro}
\caption{The crutial observation motivating this paper is the fact that sparse longitudinal trajectory prediction problem can be mapped to the matrix completion problem. Matrix completion can be approached with matrix factorization where we look for $W$ and $A$ of low rank, approximating observed values in $Y$ (circled green rectangles in the matrix $Y$). In the sparse longitudinal setting, we impose continuity by fixing the basis $B$ (e.g., splines, here with three elements), and again we find low-rank $W$ and $A$ approximating observed values in $Y$.}
\label{fig:idea}
\end{figure}
We first introduce a methodology for univariate sparsely-sampled processes. The direct method, mixed-effect models and low-rank approximations described in Section \ref{s:background} have their analogy in the matrix completion setting. We discuss these analogies in sections \ref{ss:direct-matrix} and \ref{ss:low-rank-matrix}. Next, we show that the simple representation of the problem allows for extension to multivariate sparsely-sampled processes and a regression setting.
\subsection{Notation}
For each individual $i \in \{1,2,...,N\}$ we observe $n_i$ measurements $\{\tilde y_{i,1},...,\tilde y_{i,n_i}\}$ at time-points $\{t_{i,1},t_{i,2},...,t_{i,n_i}\}$. Unlike in the prior work introduced in Section \ref{s:background}, here we discretize the time grid to $T$ time-points $G = \left[\tau_1, \tau_2, ..., \tau_T\right],$ where $t_{\min} = \tau_1$, $t_{\max} = \tau_T$ and $T$ is arbitrarily large. Each individual $i$ is expressed as a partially observed vector $r_i \in \mathbb{R}^T$. For each time-point $t_{i,j}$ for $1 \leq j \leq n_i$ we find a corresponding grid-point $g_i(j) = \argmin_{1 \leq k \leq T} |\tau_k - t_{i,j}|$. We assign $y_{i,g_i(j)} = \tilde y_{i,j}$. Let $O_i = \{g_i(j): 1 \leq j \leq n_i \}$ be a set of grid indices corresponding to observed grid-points for an individual $i$. All elements outside $O_i$, i.e. $\{y_{i,j} : j \notin O_i\}$ are considered missing.
For $T$ sufficiently large, our observations can be uniquely represented as a $N \times T$ matrix $Y$ with missing values. We denote the set of all observed elements by pairs of indices as $\Omega = \{ (i,j) : i\in \{1,2,...,N\}, j \in O_i \}$. Let $P_\Omega(Y)$ be the projection onto observed indices, i.e. $P_\Omega(Y) = W$, such that $W_{i,j} = Y_{i,j}$ for $(i,j) \in \Omega$ and $W_{i,j} = 0$ otherwise. We define $P^\perp_\Omega(Y)$ as the projection on the complement of $\Omega$, i.e. $P^\perp_\Omega(Y) = Y - P_\Omega(Y)$. We use $\|\cdot\|_F$ to denote the Frobenius norm, i.e. the square root of the sum of matrix elements, and $\|\cdot\|_*$ to denote the nuclear norm, i.e. the sum of singular values.
As in Section \ref{s:background} we impose smoothness by using a continuous basis $\mathbf{b}(t) = [b_1(t),b_2(t),...,b_K(t)]'$. However, now, the basis is evaluated on the grid $G$ and we define a $T \times K$ matrix $B = [\mathbf{b}(\tau_1),\mathbf{b}(\tau_2),...,\mathbf{b}(\tau_T)]'$.
In our algorithms we use a diagonal-thresholding operators defined as follows. Let $D = \diag(d_1,...,d_p)$ be a diagonal matrix. We define {\em soft-thresholding} as
\begin{equation*}
D_\lambda = \diag((d_1 - \lambda)_+,(d_2 - \lambda)_+,...,(d_p - \lambda)_+),\label{eq:thresholding}
\end{equation*}
where $(x)_+ = \max(x, 0)$, and {\em hard-thresholding} as
\begin{equation*}
D_\lambda^H = \diag(d_1,d_2,...,d_q,0,...,0),\label{eq:hard-thresholding}
\end{equation*}
where $q = \argmin_k(d_k < \lambda)$.
\subsection{Direct approach}\label{ss:direct-matrix}
The optimization problem \eqref{eq:direct-individual} of the direct approach described in Section \ref{ss:direct} can be approximated in the matrix completion setting. First, note that the bias introduced by the grid is negligible if the grid is sufficiently large. We have
\begin{align}
\left|\tilde y_{i,j} - \mathbf{w}_i'\mathbf{b}(t_{i,j})\right| &= \left|y_{i,g(j)} - \mathbf{w}_i'\mathbf{b}(t_{i,j})\right|\nonumber\\
&= \left| y_{i,g(j)} - \mathbf{w}_i' (\mathbf{b} (\tau_{g_i(j)}) + \mathbf{b} (t_{i,j}) - \mathbf{b} (\tau_{g_i(j)}))\right|\nonumber\\
&\leq \left| y_{i,g(j)} - \mathbf{w}_i' \mathbf{b}(\tau_{g_i(j)})\right| + \left|\mathbf{w}_i' (\mathbf{b}(t_{i,j}) - \mathbf{b}(\tau_{g_i(j)}))\right|\label{eq:grid-snap}
\end{align}
and by the continuity of the basis on a closed interval $[t_{\min},t_{\max}]$ the second element in \eqref{eq:grid-snap} can be arbitrarily small if $T \rightarrow \infty$. For the simplicity of notation in the reminder of this work we assume that all the points are observed on the grid $G$ of $T$ equidistributed points.
Now, we rewrite the optimization problem \eqref{eq:direct-individual} as a matrix completion problem
\begin{align}
\argmin_{\{\mathbf{w}_i\}}\sum_{i=1}^N \sum_{j=1}^{n_i}\left|y_{i,g_i(j)} - \mathbf{w}_i' \mathbf{b}(\tau_{g_i(j)}))\right|^2 &= \argmin_{\{\mathbf{w}_i\}}\sum_{(i,k) \in \Omega}\left|y_{i,k} - \mathbf{w}_i' \mathbf{b}(\tau_{k}))\right|^2\nonumber\\
&= \argmin_W \| P_\Omega(Y - WB') \|_F^2,\label{eq:direct-matrix}
\end{align}
where by optimization over $\{\mathbf{w}_i\}$ we mean optimization over all $\{\mathbf{w}_i : i \in \{ 1,2,...,N \}\}$ and $W$ is an $N \times K$ matrix composed of vectors $\{\mathbf{w}_i'\}$ stacked vertically.
The matrix formulation in equation \eqref{eq:direct-matrix} and the classic approach in Section \ref{ss:direct} share multiple characteristics. In both cases, if data is dense, we may find an accurate representation of the underlying process simply by fitting least-squares estimates of $W$ or $\{\mathbf{w}_i\}$. Conversely, if the data is too sparse, the problem becomes ill-posed, and the least-squares estimates can overfit.
However, representations \eqref{eq:direct-individual} and \eqref{eq:direct-matrix} differ algebraically and this difference constitutes the foundation for the method introduced in the sequel of this paper. The matrix representation enables us to use the matrix completion framework and, in particular, in Section \ref{ss:matrix-factorization} we introduce convex optimization algorithms for solving \eqref{eq:direct-matrix}.
Note that some low-rank constraints on the random effect from the mixed-effect model introduced in Section \ref{ss:lmm} can be expressed in terms of constraints on $W$ and they can be potentially solved without imposing probability distributions. In particular in Section \ref{ss:low-rank-matrix} we show that the linear mixed-effect model can be expressed by constraining $\rank(W)$.
\subsection{Low-rank matrix approximation}\label{ss:low-rank-matrix}
In the low-rank approach described in Section \ref{ss:reduced-rank} we assume that individual trajectories can be represented in a low-dimensional space, by constraining the rank of $W$.
We might attempt taking the same route for solving \eqref{eq:direct-matrix}. One difficulty comes from the fact that optimization with a rank constraint on $W$ turns the original least squares problem into a non-convex problem. Motivated by matrix completion literature, we relax the rank constraint in \eqref{eq:direct-matrix} to a nuclear norm constraint and we attempt to solve
\begin{align}
\argmin_W \| P_\Omega(Y - WB') \|_F^2 + \lambda\|W\|_*,
\label{eq:rank-restricted}
\end{align}
for some parameter $\lambda > 0$.
\citet{cai2010singular} propose a first-order singular value thresholding algorithm (SVT), for solving a general class of problems involving a nuclear norm penalty and a linear form of $Y$, which includes \eqref{eq:rank-restricted}. Their algorithm can handle large datasets and has strong guarantees on convergence, but it requires tuning the step size parameter, which can greatly influence performance. This limitation was addressed by \citet{ma2011fixed,mazumder2010spectral,hastie2015matrix} who introduced iterative procedures which do not depend on such tuning. Moreover, \citet{hardt2014fast,chen2015fast} propose methods for the non-convex problem and \citet{ge2016matrix} argue that the non-convex problem has no spurious local minima.
In the last decade, the machine learning and statistical learning communities have introduced multiple algorithms for matrix completion. Many of them are suitable for solving \eqref{eq:rank-restricted}. However, in this article we focus on analyzing the benefits of framing the trajectory prediction problem \eqref{eq:direct-individual} in the matrix completion framework, rather than on benchmarking existing solutions.
We argue that the matrix formulation \eqref{eq:rank-restricted} has two key advantages over the probabilistic approach (Section \ref{ss:reduced-rank}). First, the problem \eqref{eq:rank-restricted} does not impose any assumption on the distribution of $W$ or the distribution of the noise. This property is particularly important whenever the data does not follow the normal distribution. We further describe links between the matrix factorization and low-rank mixed-effect models in Section~\ref{s:the-link}. Second, the simple formulation of \eqref{eq:rank-restricted} allows us to generalize this model to the multivariate or regression setting as we discuss in Section \ref{s:multivariate}.
\subsection{Low-rank approximation with singular value thresholding}\label{ss:matrix-factorization}
The low-rank probabilistic approach, introduced in Section \ref{ss:reduced-rank}, is based on the observation that the underlying processes for each subject can be approximated in a low-dimensional space.
Here, we exploit the same characteristic using a matrix-factorization techniques for solving the optimization problem \eqref{eq:rank-restricted}.
For the sake of clarity and simplicity, we choose to solve the problem \eqref{eq:rank-restricted} with an extended version of the \textsc{Soft-Impute} algorithm designed by \citet{hastie2015matrix,mazumder2010spectral}. As discussed in Section \ref{ss:low-rank-matrix}, many other convex optimization algorithms can be applied.
The key component to the solution is the following property linking problem \eqref{eq:rank-restricted} with the singular value decomposition (SVD). Consider the fully observed case of \eqref{eq:rank-restricted}. Using Theorem 2.1 in \citet{cai2010singular}, one can show that the optimization problem
\begin{align}\label{eq:optsvd}
\argmin_{W} \frac{1}{2} \| Y - WB' \|_F^2 + \lambda\|W\|_*
\end{align}
has a unique solution $W = \mathcal{S}_\lambda (YB)$, where $\mathcal{S}_\lambda(X) = UD_\lambda V'$ and $X = UDV'$ is the SVD of $X$. We refer to $\mathcal{S}_\lambda(X)$ as the singular value thresholding (SVT) of $X$.
In order to solve \eqref{eq:optsvd} with a sparsely observed $Y$, we modify the \textsc{Soft-Impute} algorithm to account for the basis $B$. Our Algorithm \ref{alg:soft-impute} iteratively constructs better approximations of the global solution for each $\lambda$ in some predefined set $\{\lambda_1, \lambda_2, ..., \lambda_k\}$ for $k \in \mathbb{N}$. For a given approximation of the solution $W^{old}$, we use $W^{old}B'$ to impute unknown elements of $Y$ obtaining $\tilde{Y}$. Then, we construct the next approximation $W^{new}$ by computing SVT of $\tilde{Y}$.
As a stopping criterion, we compute the change between subsequent solution, relative to the magnitude of the solution, following the methodology in \cite{cai2010singular}. We set a fixed threshold $\varepsilon > 0$ for this criterion, depending on the desired accuracy.
\begin{algorithm}
\caption{\textsc{Soft-Longitudinal-Impute}\label{alg:soft-impute}}
\begin{enumerate}
\item Initialize $W^{old}$ with zeros
\item Do for $\lambda_1 > \lambda_2 > ... > \lambda_k$:
\begin{enumerate}
\item Repeat:
\begin{enumerate}
\item Compute $W^{new} \leftarrow S_{\lambda_i}( (P_\Omega(Y) + P_\Omega^\perp(W^{old}B'))B )$
\item If $\frac{\|W^{new} - W^{old}\|_F^2}{\|W^{old}\|_F^2} < \varepsilon$ exit
\item Assign $W^{old} \leftarrow W^{new}$
\end{enumerate}
\item Assign $\hat{W}_{\lambda_i} \leftarrow W^{new}$
\end{enumerate}
\item Output $\hat{W}_{\lambda_1}, \hat{W}_{\lambda_2}, ... , \hat{W}_{\lambda_k}$
\end{enumerate}
\end{algorithm}
A common bottleneck of algorithms introduced by \citet{cai2010singular,mazumder2010spectral,ma2011fixed} as well as other SVT-based approaches is the computation of the SVD of large matrices. This problem is particularly severe in standard matrix completion settings, such as the Netflix problem, where the matrix size is over $400{,}000 \times 20{,}000$. However, in our problem,
\begin{align}\label{eq:small-rank}
\rank(WB') \leq \rank(B) = K \ll N,
\end{align}
with $K \sim 10$ in our motivating data example. While algorithms for large matrices are applicable here, the property \eqref{eq:small-rank} makes the computation of SVD feasible in practice with generic packages such as PROPACK \citep{larsen2004propack}.
For making predictions on new data, in practice, we are interested in two scenarios: (1) we collect new measurements of an existing patient $i \in \{1,2,...,N\}$; (2) we include a new patient $(N+1)$ with at least as many observations as the rank of $W$. In both cases, we use the current fit for each parameter $\lambda$ and update all models with newly observed data. This approach not only estimates new predictions but also marginally improves the existing model.
\subsection{$\ell_0$-norm regularization}
While the nuclear norm relaxation \eqref{eq:rank-restricted} is motivated by making the problem convex, \citet{mazumder2010spectral} argue that in many cases it can also give better results than the rank constraint. They draw an analogy to the relation between best-subset regression ($\ell_0$ regularization) and LASSO ($\ell_1$ regularization as in \citet{tibshirani1996regression, friedman2001elements}). In LASSO by shrinking model parameters, we allow more parameters to be included, what can potentially improve the prediction if the true subset is larger than the one derived through $\ell_0$ regularization. In the case of \eqref{eq:rank-restricted}, shrinking the nuclear norm allows us to include more dimensions of $W$ again potentially improving the prediction if the true dimension is high.
Conversely, the same phenomenon can also lead to problems if the underlying dimension is small. In such case, shrinking may allow including unnecessary dimensions emerging from noise. To remedy this issue, following the analogy with penalized linear regression, we may consider another classes of penalties. In particular, we may consider coming back to the rank constraint by modifying the nuclear norm penalty \eqref{eq:rank-restricted} to $\ell_0$. We define $\|W\|_0 = \rank(W)$. The problem
\begin{align}\label{eq:matrixproblem-final-l0}
\min_{W} \frac{1}{2} \|P_\Omega(Y - WB')\|_F^2 + \lambda\|W\|_0,
\end{align}
has a solution $W = \mathcal{S}_\lambda^H (YB)$, where $\mathcal{S}_\lambda^H(X) = UD_\lambda^H V'$ and $X = UDV'$ is the SVD of $X$. We refer to $\mathcal{S}_\lambda^H(X)$ as the hard singular value thresholding (hard SVT) of $X$. We use Algorithm \ref{alg:hard-impute} to find the hard SVT of $YB$.
\begin{algorithm}
\caption{\textsc{Hard-Longitudinal-Impute}\label{alg:hard-impute}}
\begin{enumerate}
\item Initialize $W^{old}_{\lambda_i}$ with solutions $\tilde{W}_{\lambda_i}$ from \textsc{Soft-Longitudinal-Impute}
\item Do for $\lambda_1 > \lambda_2 > ... > \lambda_k$:
\begin{enumerate}
\item Repeat:
\begin{enumerate}
\item Compute $W^{new} \leftarrow S_{\lambda_i}^H( (P_\Omega(Y) + P_\Omega^\perp(W^{old}B'))B )$
\item If $\frac{\|W^{new} - W^{old}\|_F^2}{\|W^{old}\|_F^2} < \varepsilon$ exit
\item Assign $W^{old} \leftarrow W^{new}$
\end{enumerate}
\item Assign $\hat{W}_{\lambda_i} \leftarrow W^{new}$
\end{enumerate}
\item Output $\hat{W}_{\lambda_1}, \hat{W}_{\lambda_2}, ... , \hat{W}_{\lambda_k}$
\end{enumerate}
\end{algorithm}
The problem \eqref{eq:matrixproblem-final-l0} is non-convex, however, by starting from the solutions of Algorithm \ref{alg:soft-impute} we explore the space around the global minimum of the $\ell_1$ version of the problem. \citet{ge2016matrix} show that this strategy is successful empirically.
\subsection{The link between reduced-rank model and \textsc{Soft-Longitudinal-Impute}}\label{s:the-link}
Intuitively we might expect similarity between the principal directions derived using the probabilistic approach \eqref{eq:probabilistic} and their counterparts derived from the SVT-based approach. We investigate this relation by analyzing behavior of SVT for matrices sampled from the probabilistic model given by \eqref{eq:probabilistic}.
For simplicity, let us assume that $\mu = 0$ and the data is fully observed on a grid $G$ of $T$ time-points. Assume that observations $i \in \{1,2,...,N\}$ come from the mixed-effect model
\begin{align}
\mathbf{y}_i | \mathbf{w}_i &\sim \mathcal{N}( B \mathbf{w}_i, \sigma^2 I_T), \label{eq:model-y}\\
\mathbf{w}_i &\sim \mathcal{N}(0 , \Sigma), \nonumbe
\end{align}
where $\Sigma$ is an unknown covariance matrix of rank $q < K$ and variables $\{\mathbf{w}_i, \mathbf{y}_i\}$ are independent. By the spectral decomposition theorem we decompose $\Sigma = V\Lambda V'$, where $V$ is a $K\times K$ orthogonal and $\Lambda$ is a diagonal $K\times K$ matrix with $q$ positive coefficients in decreasing order.
Since $\mathbf{y}_i$ and $\mathbf{w}_i$ are independent, the distribution \eqref{eq:model-y} can be rewritten as
\begin{align}
\mathbf{y}_i &\sim \mathcal{N}(0, B V \Lambda V' B' + \sigma^2 I_T). \label{eq:model}
\end{align}
Note that, the model \eqref{eq:model} is a factor model with $q$ factors---the first $q$ columns of $BV$.
The following theorem constitutes a link between the mixed-effect model and SVT. It is adapted from Theorem 9.4.1 in \citet{mardia1980multivariate}, derived from \citet{joreskog1967some},
\begin{theorem}\label{thm:maxlike}
Let $Y = [\mathbf{y}_1,...,\mathbf{y}_N]'$ be the observed matrix and let $S_{\sigma^2}(YB) = UD_{\sigma^2}Q'$. Then, $(D_{\sigma^2},Q)$ is the maximum likelihood estimator of $(\Lambda,V)$.
\end{theorem}
Factor analysis methods give not only estimates of $\Lambda$ and $V$ but also estimates of the individual latent variables $W = [\mathbf{w}_1,...,\mathbf{w}_N]'$.
In the multivariate analysis literature, there are multiple ways to estimate factor scores, i.e., a matrix $A$ such that $X \sim AD_{\sigma^2}V'$, most notably Spearman's scores, Bartlett's scores, and Thompson's scores \citep{kim1978factor}.
Simply taking $W = U$ as the estimate of the scores corresponds to the solution of \eqref{eq:optsvd} as long as $\lambda = \sigma^2$.
Note that in Theorem \ref{thm:maxlike} we assume that $\sigma^2$ is known, which is infeasible in practice. However, the likelihood of $(V,\sigma)$ can be parametrized by $\sigma$, and we can find the optimal solution analytically. This corresponds to minimizing \eqref{eq:optsvd} for different $\lambda$.
This intuition is confirmed in our simulation study in Section \ref{s:simulation} (see Figure \ref{fig:principal-components}). Note that a similar analogy is drawn between the classical principal component analysis and probabilistic principal component analysis by \citet{tipping1999probabilistic} and \citet{james2000principal}.
\section{Multivariate longitudinal data}\label{s:multivariate}
In practice, we are often interested in prediction of a univariate process in the context of other longitudinal measurements and covariates constant over time. Examples include prediction of disease progression given patient's demographics, data from clinical visits at which multiple blood tests or physical exams are taken, or measurements which are intrinsically multivariate such as gene expression or x-rays collected over time. The growing interest in this setting stimulated research in latent models \citep{sammel1996latent} and multivariate longitudinal regression \citep{gray1998estimating,gray2000multidimensional}. \citet{diggle2002analysis} present an example case study in which longitudinal multivariate modeling enables estimation of joint confidence region of multiple parameters changing over time, shrinking the individual confidence intervals.
In this section, we present an extension of our univariate framework to multivariate measurements sparse in time. We explore two cases: (1) dimensionality reduction, where we project sparsely sampled multivariate processes to a small latent space, and (2) linear regression, where we use a multivariate process and covariates constant over time for prediction of a univariate process of interest. To motivate our approach, we start with a regression involving two variables observed over time.
\subsection{Motivation: Univariate regression}
Suppose, as previously, that the true processes are in a low-dimensional space of some continuous functions (e.g., splines) and that we observe them with noise. More precisely, let
\begin{align}\label{eq:definitions-xy}
\mathbf{x}_i = B\mathbf{w}_i + \mathbf{e}_{x,i} \text{\ \ and\ \ } \mathbf{y}_i = B\mathbf{u}_i + \mathbf{e}_{y,i},
\end{align}
for $1 \leq i \leq N$, where $\mathbf{x}_i,\mathbf{y}_i,\mathbf{e}_{x,i},\mathbf{e}_{y,i}$ are $T \times 1$ vectors, $\mathbf{w}_i, \mathbf{u}_i$ are $K \times 1$ vectors and $B$ is a $T \times K$ matrix of $K$ splines evaluated on a grid of $T$ points. We assume zero-mean independent errors $\mathbf{e}_{x,i},\mathbf{e}_{y,i}$ with fixed covariance matrices $\Sigma_X,\Sigma_Y$ respectively, and that the true processes underlying the observed $\mathbf{x}_i$ and $\mathbf{y}_i$ follow a linear relation in the spline space, i.e.
\begin{align}\label{eq:linear-xy}
\mathbf{u}_i = A'\mathbf{w}_i,
\end{align}
where $A$ is an unknown $K \times K$ matrix.
Let $X,Y,U,W$ be matrices formed from $\mathbf{x}_i',\mathbf{y}_i',\mathbf{w}_i',\mathbf{u}_i'$ stacked vertically. From \eqref{eq:linear-xy} we have $U = WA$, while \eqref{eq:definitions-xy} implies
\begin{align}\label{eq:almost-errors-in-variables}
X = WB' + E_X \text{ \ and \ } Y = WA B' + E_{Y},
\end{align}
where $E_X,E_Y$ are matrices of observation noise. Without loss of generality we assume that $B$ is orthonormal. We have freedom to choose the basis $B$ and any basis can be orthogonalized using, for example, singular value decomposition.
Note that due to orthogonality of $B$ and after multiplying both expressions in \eqref{eq:almost-errors-in-variables} by $B$ we can map the problem to the classical multivariate {\em errors-in-variables models}. Let
\begin{align}\label{eq:errors-in-variables}
\tilde{X} = XB = W + \tilde{E}_X \text{ and } \tilde{Y} = YB = WA + \tilde{E}_Y,
\end{align}
where $\tilde{E}_X = E_XB$ and $\tilde{E}_Y = E_YB$. In errors-in-variables models it is assumed that the parameters $W$ and $A$ are unknown, and are to be estimated. Both regressors and responses are measured with noise (here $\tilde{E}_X$ and $\tilde{E}_Y$). The parameter $W$ can be interpreted as a latent representation of both $\tilde{X}$ and $\tilde{Y}$.
The problem of estimating parameters in \eqref{eq:errors-in-variables} has been extensively studied in literature dating back to \citet{adcock1878problem}. Two main methods for estimating parameters in \eqref{eq:errors-in-variables} are {\em maximum likelihood estimates (MLE)} and {\em generalized least squares estimators (GLSE)}. The estimators in MLE are derived under the assumption of certain distributions of the errors. In GLSE, the only assumption about errors is that they are independent, zero-mean, and they have a common covariance matrix. Then, $\tilde{X} - W$ and $\tilde{Y} - WA$ are zero-mean and estimates for $W$ and $B$ can be found by optimizing some norm of these expressions. \citet{gleser1973estimation} show that in the no-intercept model for $\tilde{X}$ and $\tilde{Y}$ of the same size (as in our case) and under the assumption of Gaussian errors, MLE and GLSE give the same estimates of $A$, if GLSE are derived for the Frobenius norm.
In this work, we focus on the GLSE approach as it can be directly solved in our matrix factorization framework and we find it easier to deploy and extend in practice.
To account for different magnitudes of the noise in $X$ and $Y$, we consider the optimization problem with weights
\begin{align}\label{eq:lag:bivariate}
\minimize_{A,W} & \frac{1}{\sigma_X^2}\|XB - W \|_F^2 + \frac{1}{\sigma_Y^2} \| YB - WA \|_F^2,
\end{align}
where $\sigma_X,\sigma_Y > 0$. Let $\gamma = \sigma_X^2 / \sigma_Y^2$. Then \eqref{eq:lag:bivariate} can be transformed to
\begin{align}\label{eq:lag:bivariate3}
\minimize_{A,W} & \| (XB : \gamma YB) - W(I: \gamma A) \|_F^2
\end{align}
where $(\cdot : \cdot)$ operator stacks vertically matrices with the same number of rows. To solve \eqref{eq:lag:bivariate3}, we show that the SVD of $(XB : \gamma YB)$ truncated to the first $K$ dimensions, can be decomposed to $W(I: \gamma A)$. Let $USV'$ be the SVD of $(XB : \gamma YB)$, with
\[
U = \begin{bmatrix}
U_1 & U_2
\end{bmatrix},
S = \begin{bmatrix}
S_{11} & S_{12}\\
S_{21} & S_{22}
\end{bmatrix}
\text{ and }
V = \begin{bmatrix}
V_{11} & V_{12}\\
V_{21} & V_{22}
\end{bmatrix},
\]
where each $S_{ij}$ and $V_{ij}$ is a $K\times K$ matrix for $1 \leq i,j \leq 2$ and each $U_i$ is a $N \times K$ matrix for $1 \leq i \leq 2$. By Lemma 2.1 and Lemma 2.2 in \citep{gleser1981estimation} matrix $V_{11}$ is almost surely nonsingular. Therefore, $V_{11}^{-1}$ almost surely exists and we can transform the decomposition such that $(I : \gamma A) = (V_{11}')^{-1}\begin{bmatrix}V_{11}' & V_{21}'\end{bmatrix}$ and $W = U_1 S_{11} V_{11}'$, solving \eqref{eq:lag:bivariate3}.
For partially observed data, if they are very sparse, it might be essential to constrain the rank of the solution. The partially-observed and rank-constrained version of the problem \eqref{eq:lag:bivariate3} takes the form
\begin{align*
\minimize_{A,W} & \| P_{\tilde\Omega}((X:\gamma Y) - W(B':\gamma A B')) \|_F^2,\nonumber\\
\subjectto & \rank(W(B':\gamma A B')) = k,
\end{align*}
where $k$ is the desired rank of the solution and $P_{\tilde\Omega}$ is a projection on \[\tilde\Omega = \{(q,r): (q,r) \in \Omega \text{ or } (q,r-T) \in \Omega\}. \] As previously, for an unknown $k$ we can introduce a rank penalty using the nuclear norm
\begin{align}\label{eq:lag:bivariate2-partial}
\minimize_{A,W} & \| P_{\tilde\Omega}((X:\gamma Y) - W(B':\gamma A B')) \|_F^2 + \lambda\|W(B':\gamma A B')\|_*.
\end{align}
The algorithm in the general case of multiple processes is derived in Section \ref{ss:dim-red}.
Although we motivate the problem as a regression of $Y$ on $X$, note that $X$ and $Y$ are symmetric in \eqref{eq:lag:bivariate3}. The low-rank matrix $W$ is, therefore, a joint low-rank representation of matrices $X$ and $Y$ and thus our method can be seen as a dimensionality reduction technique or as a latent space model. In Section \ref{ss:dim-red} we extended this idea to a larger number of variables. In Section \ref{ss:regression} we discuss how this approach can be used for regression.
The linear structure of \eqref{eq:almost-errors-in-variables} allows us to draw analogy not only to the errors-in-variables models but also to vast literature on {\em canonical correlation analysis (CCA)}, {\em partial least squares (PLS)}, {\em factor analysis (FA)}, and {\em linear functional equation (LFE) models}. \cite{borga1997unified} show that solutions of CCA and PLS can also be derived from the SVD of stacked matrices, as we did with $(XB:\gamma YB)$ in \eqref{eq:lag:bivariate3}. \cite{gleser1981estimation} thoroughly discusses the relation between errors-in-variables, FA and LFE.
Finally, note that the method of using SVD for stacked matrices has also been directly applied in the recommender systems context. \citet{condli1999bayesian} showed that for improving prediction of unknown entries in some partially observed matrix $Q$ one might consider a low-rank decomposition of $(Q:R)$, where $R$ are additional covariates for each row, e.g., demographic features of individuals.
\subsection{Dimensionality reduction}\label{ss:dim-red}
Suppose that for every individual we observe multiple variables varying in time (e.g. results of multiple medical tests at different times in a clinic) and we want to find a projection on $\mathbb{R}^d$ maximizing the variance explained for some $d \in \mathbb{N}$. This projection would correspond to characterizing patients by their progression trends of multiple metrics simultaneously.
We extend the equation \eqref{eq:lag:bivariate3} to account for a larger number of covariates and we do not impose decomposition of the solution yet. We formulate the following optimization problem
\begin{align*
\argmin_{W} &\| (X_1B:X_2B:...:X_pB) - W \|_F^2 + \lambda\|W\|_*,
\end{align*}
where $X_i$ are some $N \times T$ matrices corresponding to the processes measured on a grid of $T$ points, $B$ is a basis evaluated on the same grid and orthogonalized (a $T \times K$ matrix), and $W$ is a $N \times pK$ matrix.
Let $ \mathbf{B} = I_p \otimes B $ be the Kronecker product of $p \times p$ identity matrix and $B$, i.e. a $pT \times pK$ matrix with $B$ stacked $p$ times on the diagonal, and let $\mathbf{X} = (X_{1}: X_{2}:...: X_{p})$. Note that $\mathbf{B}$ is orthogonal and therefore results developed in Section \ref{ss:reduced-rank} apply here. In particular, singular value thresholding solves
\begin{align*
\argmin_{W} &\| \mathbf{X} - W \mathbf{B}' \|_F^2 + \lambda\|W\|_*
\end{align*}
and we can use Algorithm \ref{alg:soft-impute} for solving
\begin{align}\label{eq:multivar-partially}
\argmin_{W} &\| P_{\Omega}\left(\mathbf{X} - W\mathbf{B}\right) \|_F^2 + \lambda\|W\|_*,
\end{align}
where $P_{\Omega}$ is the projection on observed indices $\Omega$.
Note that \eqref{eq:multivar-partially} can be further extended. First, we can allow for different basis for each process. Let $B_{i}$ be a basis selected for approximation of processes $X_{i}$ for $1 \leq i \leq p$, where each basis $B_i$ is $d_i$ dimensional and it is evaluated on some time-grid $T_i$. In particular, we allow $B_i = [1]$, which corresponds to a covariate $X_i$ constant in time for a given individual (e.g. gender). Second, if measurements are on different scales, we may normalize them using scaling factors $\gamma_1, \gamma_2,..., \gamma_p > 0$. Then, the problem \eqref{eq:multivar-partially} takes form
\begin{align*}
\argmin_{Z} &\| P_{\Omega}\left((\gamma_1 X_{1}B_1:\gamma_2 X_{2}B_2:...:\gamma_p X_{p}B_p) - W\right) \|_F^2 + \lambda\|W\|_*,
\end{align*}
which we solve with the same techniques as \eqref{eq:multivar-partially}. Third, the observed indices $\Omega$ may vary in each process allowing for, for example, dense measurements of one process and sparse measurements of another one.
\subsection{Regression}\label{ss:regression}
In practice, we are interested in estimating progression of an individual parameter (e.g., cancer growth) given some individual features constant over time (e.g., demographics) or given progressions of other covariates (e.g., blood tests, vitals, biopsy results).
We start with a regression problem with fixed covariates and sparsely observed response trajectories. Let $X$ be a $N \times d$ matrix of observed covariates, $Y$ be a sparsely observed $N \times T$ matrix of trajectories, and $B$ be a $T \times K$ matrix representing a basis of $K$ splines evaluated on a grid of $T$ points. We consider the optimization problem
\begin{align}\label{regression}
\argmin_A \| P_\Omega(Y - XAB')\|^2,
\end{align}
where $A$ is a $d \times K$ matrix and $P_\Omega$ is a projection on the observed indices $\Omega$. To solve \eqref{regression} we propose an iterative Algorithm \ref{alg:longitudinal-regression}.
\begin{algorithm}
\caption{\textsc{Sparse-Regression}\label{alg:longitudinal-regression}}
\vspace{3pt}
\begin{enumerate}
\item Initialize $A$ zeros
\item Repeat till convergence:
\begin{enumerate}
\item Impute regressed values $\hat{Y} = P_\Omega(Y) + P_\Omega^\perp(XAB')$
\item Compute $A^{new} \leftarrow (X'X)^{-1} X'\hat{Y}B$
\item If $\frac{\|A^{new} - A\|_F^2}{\|A\|_F^2} < \varepsilon$ exit
\item Assign $A \leftarrow A^{new}$
\end{enumerate}
\item Return $A$
\end{enumerate}
\end{algorithm}
Suppose we want to incorporate other variables varying in time for prediction of the response process. We can directly apply the method proposed in Section \ref{ss:dim-red} and model the response and regressors together. However, it might be suboptimal for prediction, as it optimizes for least-squares distance in all variables rather than only the response. This difference is analogical to the difference between regression line of some univariate $y$ on independent variables $x_1,...,x_p$ and the first principal component analysis of $(y,x_1,...,x_p)$, used for prediction of $y$. While the first one minimizes the distance to $y$, the latter minimizes the distance to $(y,x_1,...,x_p)$, which is usually less efficient for predicting $y$.
Alternatively, we can use methodology from Section \ref{ss:dim-red} only for covariates. The solution of \eqref{eq:multivar-partially} can be decomposed into $W = USV'$, and we regress $Y$ on $U$ as in \eqref{regression}.
Let $U$ be an $N \times d_2$ orthogonal matrix derived from $(X_1,...,X_p)$, where $d_2$ is the numbers of latent components. We search for a $d_2 \times K$ matrix $A$ solving
\begin{align}\label{eq:pcr}
\minimize_{A} \|P_\Omega(Y - UAB')\|_F^2 + \lambda\|A\|_*,
\end{align}
where $P_\Omega$ is a projection on a set of observed coefficients.
We propose a two-step iterative procedure (Algorithm~\ref{alg:sparse-regression}).
\begin{algorithm}
\caption{\textsc{Sparse-Longitudinal-Regression}\label{alg:sparse-regression}}
\vspace{3pt}
\begin{flushleft}
\textbf{Step 1: Latent representation}
\end{flushleft}
\begin{enumerate}
\item For sparsely observed $\mathbf{X} = (X_1,X_2,...,X_p)$ find latent scores $U$ (Algorithm \ref{alg:soft-impute})
\end{enumerate}
\begin{flushleft}
\textbf{Step 2: Regression}
\end{flushleft}
\begin{enumerate}
\item For each $\lambda_1,\lambda_2,...,\lambda_k$
\begin{itemize}
\item Get $A_{\lambda_i}$ by solving the regression problem \eqref{eq:pcr} with $\textbf{Y},U,\lambda_i$ (Algorithm \ref{alg:longitudinal-regression})
\end{itemize}
\item Return $A_{\lambda_1}, A_{\lambda_2}, ..., A_{\lambda_k}$
\end{enumerate}
\end{algorithm}
\section{Simulations}\label{s:simulation}
We illustrate properties of the multivariate longitudinal fitting in a simulation study. First, we generate curves with quickly decaying eigenvalues of covariance matrices. Then, we compare the performance of the methods in terms of the variance explained by the predictions.
\subsection{Data generation} Let $G$ be a grid of $T = 31$ equidistributed points and let $B$ be a basis of $K = 7$ spline functions evaluated on the grid $G$.
We simulate three $N \times K$ matrices using the same procedure $\mathcal{G}(r_1, r_2, K, N)$, where $r_1, r_2 \in \mathbb{R}_+^K$:
\begin{enumerate}
\item Define the procedure $\mathcal{M}(r)$ generating symmetric matrices $K \times K$ for a given vector $r \in \mathbb{R}_+^K$:
\begin{enumerate}
\item[(i)] Simulate $K \times K$ matrix $R$
\item[(ii)] Use SVD to decompose $S$ to $UDV'$
\item[(iii)] Return $Q = V \diag[r] V' $, where $\diag[r]$ is a diagonal matrix with $r$ on the diagonal
\end{enumerate}
\item Let $\Sigma_1 = \mathcal{M}(r_1)$, $\Sigma_2 = \mathcal{M}(r_2)$ and $\mu = \mathcal{MN}(0, I_K)$, where $\mathcal{MN}$ denotes the multivariate Gaussian distribution
\item Draw $N$ vectors $v_i$ according to the distribution
\begin{align*}
v_i \sim
\begin{cases}
\mathcal{MN}(2\mu, \Sigma_1) & \text{ if } 1 \leq i \leq N/3\\
\mathcal{MN}(-\mu, \Sigma_2) & \text{ if } N/3 < i \leq N
\end{cases}
\end{align*}
\item Return a matrix with rows $[v_i]_{1 \leq i \leq N}$.
\end{enumerate}
Define \[ r_1 = [1, 0.4, 0.005, 0.1 \exp(-3), ..., 0.1 \exp(-K+1)] \] and \[ r_2 = [1.3, 0.2, 0.005, 0.1 \exp(-3), ..., 0.1 \exp(-K+1)],\]
and let $X_1,X_2,Z$ be generated following the procedure $\mathcal{G}(r_1,r_2,K,N)$ and let $Y = Z + X_1 + X_2$. We consider $X_1,X_2$ and $Y$ as coefficients in a spline space $B$. We derive corresponding functions by multiplying these matrices by $B'$, i.e. $X_{f,1} = X_1B'$, $X_{f,2} = X_2B'$ and $Y_f = YB'$. We set $N=100$.
We uniformly sample $10\%$ indices $\Omega \subset \{1,...,N\} \times \{1,...,T\}$, i.e. around $3$ points per curve on average. Each observed element of each matrix $X_{f,1}, X_{f,2}$ and $Y$ is drawn with Gaussian noise with mean $0$ and standard deviation $0.25$. The task is to recover $Y$ from sparse observed elements $\{Y_{i,j} : (i,j) \in \Omega\}$.
\subsection{Methods}
We compare \textsc{Soft-Longitudinal-Impute} (SLI) defined in Algorithm \ref{alg:soft-impute} with the fPCA procedure \citep{james2000principal}, implemented in \citet{peng2009geometric}. Both SLI and fPCA require a set of basis functions. In both cases, we use the same basis $B$ as in the data generation process. In SLI we also need to specify the tuning parameter $\lambda$, while in fPCA we need to choose the rank $R$. We use cross-validation to choose $\lambda$ and $R$ by optimizing for the prediction error on held-out (validation) observations.
We divide the observed coefficients into training ($81\%$), validation ($9\%$) and test ($10\%$) sets. We choose the best parameters of the three models on the validation set and then retrain on the entire training and validation sets combined.
We compute the error of entire curves by taking mean squared Frobenius distance between $Y$ and estimated $\hat{Y}$, i.e.
\begin{align}\label{eq:err}
MSE(\hat{Y}) = \frac{1}{T|S|} \sum_{i\in S} \|Y_i - \hat{Y}_i \|_F^2
\end{align}
on the test set $S$.
For illustration, we also present results of \textsc{Sparse-Longitudinal-Regression} (SLR), regressing $Y$ on $X_1$ and $X_2$. Note, however, that the SLR, unlike two other methods, uses additional information about $Y$ contained in $X_1$ and $X_2$. This comparison is only meant to validate our approach.
We train the algorithms with all combinations of parameters:
regularization parameter for SLI and SLR procedures $\lambda \in \{10, 15, 20, ..., 50\}$ and the rank for fPCA procedure $d \in \{2,3,4\}$. We define the grid of $T = 51$ points. We compare the three methods fPCA, SLI and SLR, to the baseline {\it null model} which we define as the population mean across all visits.
\begin{figure}[ht!]
\includegraphics[width=0.49\linewidth]{components-fpca}
\includegraphics[width=0.49\linewidth]{components-fimp}
\caption{The first three principal components derived using sparse functional PCA (left) and Soft-Longitudinal-Impute (right). The components are ordered as follows: 1st red solid curve, 2nd green dashed curve, 3rd blue dotted line. As expected, estimates of components are very similar.}
\label{fig:principal-components}
\end{figure}
\subsection{Results}
\begin{figure}[ht!]
\includegraphics[width=0.49\linewidth]{pred-mean}
\includegraphics[width=0.49\linewidth]{pred-fpca}
\includegraphics[width=0.49\linewidth]{pred-fimp}
\includegraphics[width=0.49\linewidth]{pred-freg}
\caption{Two example curves (black and red) for each of the four methods. Solid lines are the true curves and dashed lines are the predictions. }
\label{fig:example-predictions}
\end{figure}
The SLI achieves performance similar to \citep{james2000principal}, as presented in Table~\ref{tbl:simulations}. The SLR, having access to additional information about $Y$, clearly outperforms other methods validating its correctness.
\begin{table}[ht]
\centering
\begin{tabular}{rrr}
\hline
& MSE & \% std \\
\hline
fPCA & 0.124 & 0.03 \\
SLI & 0.121 & 0.03 \\
SLR & 0.064 & 0.03 \\
\hline
\end{tabular}
\caption{Average (across 10 trials) variance explained by (1) subject's mean of observed points (mean), (2) functional principal components (fPCA), (3) Sparse-Longitudinal-Impute (SLI) and (4) Sparse-Longitudinal-Regression (SLR) with extra data available for this procedure.}
\label{tbl:simulations}
\end{table}
In Figure \ref{fig:principal-components} we present the first components derived from both sparse functional PCA and SLI. In Figure \ref{fig:example-predictions} we present example predictions from all four methods. In Figure \ref{fig:estimated-rank}, we present the estimated rank and cross-validation error of one of the simulation runs.
\begin{figure}[ht!]
\includegraphics[width=0.49\linewidth]{error-of-lambda}
\includegraphics[width=0.49\linewidth]{K-of-lambda}
\caption{Estimated error of the solution (left) and the estimated rank of the solution (right) depending on the parameter $\lambda$.}
\label{fig:estimated-rank}
\end{figure}
\section{Data study}
The target class of problems motivating this study is a prediction of trajectories of disease progression of individual patients from sparse observations. In this section, we present how our methods are applied for understanding the progression of neurological disorders leading to motor impairments and gait pathologies. First, we discuss how practitioners collect the data and use them to guide the decision process. Next, we describe our dataset and present how our methodology can improve current workflows. Then, we present and discuss results.
In clinical gait analysis, at each visit movement of a child is recorded using optical motion capture. Optical motion capture allows estimating 3D positions of body parts using a set of cameras tracking markers positions on the subject's body. A set of at least three markers is placed at each analyzed body segment usually associated with bones (e.g., tibia, humerus) so that its 3D position and orientation can be identified uniquely. These data are then used to determine relative positions of body segments by computing the angle between the corresponding planes. Typically it is done using a biomechanical model for enforcing biomechanical constraints and improving accuracy.
In gait analysis practitioners are usually concerned about movement pattern of nine joints in lower limbs: ankle, knee, hip in each leg and pelvis (Figure \ref{fig:joint-angles}). Each joint angle is measured in time. For making the curves comparable between the patients, usually, the time dimension is normalized to the percentage of the gait cycle, defined as the time between two foot strikes (Figure \ref{fig:joint-angles-in-time}).
While trajectories of joint angles are a piece of data commonly used by practitioners for taking decisions regarding treatment, their high-dimensional nature hinders their use as a quantitative metric of gait pathology or treatment outcome. This motivates development of univariate summary metrics of gait impairment, such as questionnaire-based metrics Gillette Functional Assessment Walking Scale (FAQ) \citep{gorton2011gillette}, Gross Motor Function Classification System (GMFCS) \citep{palisano2008content} and Functional Mobility Scale (FMS) \citep{graham2004functional}, or observational video analysis scores such as Edinburgh Gait Score \citep{read2003edinburgh}.
One of the most widely adopted quantitative measurements of gait impairments in pediatrics is Gait Deviation Index (GDI) \citep{schwartz2008gait}. GDI is derived from joint angle trajectories and measures deviation of the first ten singular values from the population average of the typically developing population. GDI is normalized in such a way that $100$ corresponds to the mean value of typically developing children, with the standard deviation equal $10$. It is proven to be highly correlated with questionnaire-based methods. Thanks to its deterministic derivation from the motion capture measurements this method is considered more objective than questionnaires.
\begin{figure}[p]
\centering
\includegraphics[width=0.83\linewidth]{bodies-chart-paper-layout.png}
\caption{Four joints measured in clinical gait analysis: pelvis, hip, knee, and ankle. Each joint can be measured in three planes: sagittal plane (top row), frontal plate (middle row), and transverse plane (bottom row).}
\label{fig:joint-angles}
\includegraphics[width=0.83\linewidth]{kinematics.pdf}
\caption{Recordings of joint angles during the gait cycle (fraction of the gait cycle on each X axis)}
\label{fig:joint-angles-in-time}
\end{figure}
In medical practice, GDI has been adapted as a metric for diagnosing the severity of impairment, and it constitutes an integral part of the clinical decision making process and evaluation of treatment outcomes. However, in order to correctly identify the surgery outcome, it is crucial to understand the natural progression of GDI. In particular, a positive outcome of a surgery might be negligible when compared to natural improvement during puberty. Similarly, a surgery maintaining the same level of GDI might be incorrectly classified as a failure, if the decline in patient's function over time is not accounted for.
Methods introduced in this article can be used to approximate individual progressions of GDI. First, we present how a prediction can be made solely based on the patient's GDI history and histories of other patients. Next, using our regression procedure, we predict GDI trajectories using other sparsely observed covariates, namely O2 expenditure and walking speed.
\subsection{Materials and methods}
We analyze a dataset of Gillette Children's Hospital patients visiting the clinic between 1994 and 2014, age ranging between 4 and 19 years, mostly diagnosed with Cerebral Palsy. The dataset contains $84$ visits of $36$ patients without gait disorders and $6066$ visits of $2898$ patients with gait pathologies.
Motion capture data was collected at 120Hz and joint angles in time were extracted. These joint angles were then normalized in time to the gait cycle, resulting in curves as in Figure \ref{fig:joint-angles-in-time}. Points from these curves were then subsampled (51 equidistributed points). Given the data in this form, we computed GDI metric from each visit and each leg.
In the dataset which we received from the hospital, for each patient we observe their birthday and disease subtype. From each visit, we observe the following variables: patient ID, time of the visit, GDI of the left leg, GDI of the right leg, walking speed, and O2 expenditure. Other clinical variables that we received were not included in this study. Note that walking speed is related to information we lose during normalization of the gait cycle in time. O2 expenditure is a measure of a subject's energy expenditure during walking. Pathological gait is often energy inefficient and reduction of O2 expenditure is one of the objectives of treatments. Finally, GDI is computed for two legs while in many cases the neurological disorder affects only one limb. To simplify the analysis, we focus on the more impaired limb by analyzing the minimum of the left and the right GDI.
Our objective is to model individual progression curves. We test three methods: functional principal components (fPCA), \textsc{Soft-Longitudinal-Impute} (SLI) and \textsc{Sparse-Longitudinal-Regression} (SLR). We compare the results to the \emph{null model} -- the population mean across all visits (\verb|mean|). In SLR, we approximate GDI using latent variables of sparsely observed covariates \textit{O2 expenditure} and \textit{walking speed}, following the methodology from Section \ref{ss:regression}.
In our evaluation procedure, for the test set, we randomly select $5\%$ of observations of patients who visited the clinic at least $4$ times. Then, we split the remaining $95\%$ of observations into a training and validation sets in $90:10$ proportion. We train the algorithms with the following combinations of parameters: the regularization parameter for SLI and SLR procedures $\lambda \in \{0, 0.1, 0.2, ..., 2.0\}$ and the rank for fPCA procedure $d \in \{2,3,4,...,K\}$. We define the grid of $T = 51$ points. We repeat the entire evaluation procedure $20$ times.
Let us denote the test set as $\Omega \subset \{1,2,...,N\} \times \{1,2,...,T\}$. We validate each model $M$ on held-out indices by computing the mean squared error as defined in \eqref{eq:err}. We select the parameters of each of the three methods using cross-validation, using the same validation set.
For reproducibility and to simplify adoption we created an \verb|R| package \verb|fcomplete| available at \verb|https://github.com/kidzik/fcomplete|. The package contains implementations of algorithms \ref{alg:soft-impute}, \ref{alg:hard-impute}, \ref{alg:longitudinal-regression}, and \ref{alg:sparse-regression}, and helper functions for transforming the data, sampling training and test datasets, and plotting functions. For convenience, we also provided an interface for using the \verb|fpca| package implementing Sparse Functional Principal Components algorithms \citep{james2000principal,peng2009geometric}. The analysis was perform on a desktop PC with 32 GB RAM memmory and an Intel\textsuperscript{\textregistered} Core\textsuperscript{\texttrademark} i7-6700K CPU @ 4.00GHz, operating on a Ubuntu 18.04 system with \verb|R| version 3.4.4.
\begin{table}[ht]
\centering
\begin{tabular}{rrr}
\hline
& mean & sd \\
\hline
fPCA & 0.73 & 0.16\\
SLI & 0.70 & 0.10\\
SLR & 0.68 & 0.08 \\
\hline
\end{tabular}
\caption{Distribution of cross-validated MSE of the three procedures: functional principal components (fPCA), Soft-Longitudinal-Impute (SLI), Soft-Longitudinal-Regression with two additional predictors: O2 expenditure and walking speed (SLR).}
\label{tbl:data-res}
\end{table}
\subsection{Results}\label{ss:results}
Compared to the null model, all three methods explain around $\sim 30\%$ of the variance. We present detailed results in Table \ref{tbl:data-res}. \textsc{Sparse-Longitudinal-Regression} had smaller mean and variance than two other methods indicating that two other variables. We conclude that O2 expenditure and walking speed provide additional information for prediction of GDI progression.
Both fPCA and \textsc{Sparse-Longitudinal-Impute} provide latent representations of patients' progression curves which can potentially be interpreted. To this end we first analyze the singular value vectors from our SVD solution which we refer to as principal components.
In the left plot in Figure \ref{fig:data-components} we show the first two estimated principal components. We found that the first component estimates the change between GDI before and after age of 20. The second component models changes around age of 10 and around age of 18. In the right plof in Figure \ref{fig:data-components}, by adding a principle component to the population mean curve, we illustrate how differences in the first component are reflected in the patients trajectory. By visual investigation of curves returned by our \textsc{Sparse-Longitudinal-Impute} and by fPCA we found similar trends in the first two components.
Since our SVD decomposition defines a low-rank representation of the progression trends, we can also use it to gain insights on progression in different groups of patients. In Cerebral Palsy we divide paralysis into subtypes depending on which limbs are affected: monolegia (one leg), diplegia (two legs), hemiplegia (one side of the body), triplegia (three limbs), quadriplegia (four limbs). Hemiplegia is the most prevalent in our population and it might be divided depending on severity, from type I (weak muscles, drop foot) to type IV (severe spasticity). We find differences between trends of progression for different subtypes of paralysis of patients ($F_{6,541} = 17.17, p < 10^{-15}$). We illustrate these distributions in Figure \ref{fig:subtypes}.
\begin{figure}[ht!]
\includegraphics[width=0.49\linewidth]{data-components.pdf}
\includegraphics[width=0.49\linewidth]{data-components-added.pdf}
\caption{Left: Trends of variability (principal components). Right: Effect of principal components on individual trajectories.}
\label{fig:data-components}
\includegraphics[width=0.9\linewidth]{paralysis-subtypes}
\caption{Progression trends is different subsets of diseases. Negative values of the score, such as most of the quadriplegic group, correspond to individual trends where the first component (the red curve Figure \ref{fig:data-components} left) is subtracted from the mean (the green curve in Figure \ref{fig:data-components} right). Positive values of the score, such as most of the hemiplegic group, correspond to individual trends where the first component is added (the red curve in Figure \ref{fig:data-components} right).}
\label{fig:subtypes}
\end{figure}
\section{Discussion}
Results presented in Section \ref{ss:results} imply that our \textsc{Sparse-Longitudinal-Impute} and \textsc{Sparse-Longitudinal-Regression} methods can be successfully applied to understands trends of variability of disease progression. We show how to incorporate progressions of O2 expenditure and walking speed in the prediction of the progression of GDI. We present how low-rank representation can be leveraged to gain insights about subtypes of impairment.
While a large portion of variance remains unexplained, it is important to note that in practice the individual progression is not accounted for explicitly in the current decision-making process. Instead, practitioners only use the population-level characteristics of the dependence between age and impairment severity. Our model can greatly improve this practice.
Despite successful application, we identify limitations that could be potentially addressed in the extensions of our model. First, the method is meant to capture natural continuous progression of GDI, while in practice there are many discrete events, such as surgeries that break continuity assumption and render the mean trajectories less interpretable. Second, our methodology does not address the ``cold start problem'', i.e. we do not provide tools for predictions with only one or zero observations. Third, we do not provide explicit equations for confidence bounds of predicted parameters.
While these and other limitations can constrain applicability of the method in the current form, they can be addressed using existing techniques of matrix completion. The focus of this paper is to introduce a computational framework rather than build a full solution for all cases. Elementary formulation of the optimization problem as well as the fully-functional \verb|R| implementation can foster development of new tools using matrix completion for longitudinal analysis and for mixed-effect models.
\renewcommand*{\bibfont}{\small}
|
\section{Introduction}
Let $M$ be a compact complex manifold of complex dimension $n$. Suppose that $M$ admits a K\"ahler metric $\omega_0$.
In \cite{LT}, La Nave-Tian (see also the work of Rubinstein \cite{R}) consider a family of K\"ahler metrics
$\omega = \omega(s)$ satisfying the \emph{continuity equation}
\begin{equation}
\label{eqn: ce}
\omega = \omega_0 - s \operatorname{Ric} (\omega), \quad \textrm{for } s \ge 0.
\end{equation}
Here
$\operatorname{Ric}(\omega) = -\sqrt{-1} \partial \overline{\partial} \log \det g$
is the Ricci curvature $(1,1)$ form of $\omega = \sqrt{-1} g_{i\ov{j}}dz^i \wedge d\ov{z}^j$. This equation was introduced as an alternative to the K\"ahler-Ricci flow in carrying out the Song-Tian analytic minimal model program \cite{ST, ST2}.
The continuity equation has the feature that the Ricci curvature along the path is automatically bounded from below and this has led to several developments \cite{FGS, LTZ, Li, ZZ, ZZ2}.
In this paper we study a natural analogue of (\ref{eqn: ce}) for non-K\"ahler Hermitian metrics. If $\omega$ is any Hermitian metric we still define
\begin{equation}
\label{eqn: ricci curv}
\operatorname{Ric} (\omega) = -\sqrt{-1} \partial \overline{\partial} \log \det g ,
\end{equation}
which we refer to as the \emph{Chern-Ricci form} of $\omega$. Unlike the K\"ahler case, in general this form need not
relate to the full Riemann curvature tensor in any simple fashion. We now consider the continuity equation (\ref{eqn: ce}) for general Hermitian metrics with the definition (\ref{eqn: ricci curv}).
Our first result establishes the maximal existence interval for the continuity equation.
\begin{theorem}
\label{thm: main}
Let $M$ be a compact manifold with a Hermitian metric
$\omega_0$.
Then there exists a unique family of Hermitian metrics
$\omega = \omega(s)$ satisfying
\begin{equation}
\label{eqn: flow}
\omega = \omega_0 - s \operatorname{Ric}(\omega),\quad \omega > 0,
\quad s \in [0, T),
\end{equation}
where $T$ is defined by
\begin{equation}
\label{eqn: maximal time}
T := \sup \{ s >0 \mid \exists \psi \in C^{\infty}(M) \emph{ with } \omega_0 - s \operatorname{Ric}(\omega_0) + \sqrt{-1} \partial \overline{\partial} \psi > 0 \} .
\end{equation}
\end{theorem}
We make some remarks about this result.
\medskip
\noindent
1) Theorem \ref{thm: main} extends the result of La Nave-Tian \cite{LT} who showed that when $\omega_0$ is K\"ahler, there exists a solution to (\ref{eqn: ce}) up to $T = \sup \{ s >0 \mid [\omega_0] -s c_1(M)>0 \}$, where we are writing $c_1(M) = [\operatorname{Ric}(\omega_0)] \in H^{1,1}(M;\mathbb{R})$ for the first Chern class of $M$ (modulo a factor of $2\pi$). This $T$ coincides with the maximal existence time for the K\"ahler-Ricci flow \cite{C, TZ, Ts1, Ts2}.
\medskip
\noindent
2) The continuity equation (\ref{eqn: flow}) for Hermitian metrics can be regarded as an elliptic version of the \emph{Chern-Ricci flow}
$$\frac{\partial}{\partial t} \omega = - \operatorname{Ric}(\omega),$$
first introduced by Gill \cite{G}. Indeed the value $T$ of Theorem \ref{thm: main} coincides with the maximal existence time for the Chern-Ricci flow \cite{TW2}. In particular, if $n=2$ and $\omega_0$ satisfies the Gauduchon condition $\partial \ov{\partial} \omega=0$ then $T$ can be readily computed for many examples (see \cite{TW2, TW3}).
\medskip
\noindent
3) The value $T$ is independent of the choice of $\omega_0$ in the following sense: if we replace $\omega_0$ by $\omega_0 + \sqrt{-1} \partial \overline{\partial} f>0$ for a smooth function $f$ then the value $T$ does not change.
\medskip
\noindent
4) We reduce the proof of Theorem \ref{thm: main} to an existence result of Cherrier \cite{Ch} (see Theorem \ref{thmch} below).
\medskip
Our second theorem gives an example of the continuity equation (\ref{eqn: flow}) in the setting of elliptic surfaces. In particular, it will illustrate the close connection to the Chern-Ricci flow.
Let
$\pi: M \rightarrow S$ be an elliptic bundle
over a Riemann surface $S$ of genus at least $2$. In particular, each point $y$ in $S$ has a neighborhood $U$ so that $\pi^{-1}(U)$ is biholomorphic to $U \times T^2$ for a complex 1-dimensional torus $T^2$. There exist such bundles which are \emph{non-K\"ahler} elliptic surfaces, meaning that they do not admit \emph{any} K\"ahler metric (see the exposition in \cite[Section 8]{TW3}). In fact, by the Kodaira classification, \emph{every} minimal non-K\"ahler surface of Kodaira dimension 1 is such an elliptic surface, or admits a finite cover by one (see \cite[Lemmas 1, 2]{B} or \cite[Theorem 7.4]{Wa}).
Denote by $\omega_S$ the unique K\"ahler-Einstein metric on $S$ satisfying $\operatorname{Ric}(\omega_S)=-\omega_S$. Then the pull-back $[\pi^*\omega_S]$ lies in $c_1(M)$ and it follows from Theorem \ref{thm: main} that the continuity equation (\ref{eqn: flow}) with any initial $\omega_0$ has $T=\infty$ (see Lemma \ref{lemmamax} below).
Take $\omega_0$ to be a Gauduchon ($\partial \overline{\partial} \omega_0=0$) metric on $M$. Note that every Hermitian metric is conformal to a Gauduchon one \cite{Ga}.
There exists a family of Gauduchon metrics
$\omega'(s)$ satisfying the continuity equation
\[
\omega'(s) = \omega_0 - s \textrm{Ric}(\omega'(s)),
\]
for $s \in [0,\infty)$.
It is convenient to make a scaling change (cf. \cite{ZZ})
and consider $\omega(s) = \omega'(s)/(s+1)$ so that the equation becomes
\begin{equation}
\label{eqn: eb ce}
(1+s) \omega(s)
= \omega_0 - s \textrm{Ric}(\omega(s)), \quad s \in [0,\infty).
\end{equation}
We call this the \emph{normalized continuity equation}. Our result describes the behavior of $\omega(s)$ as $s \rightarrow \infty$.
\pagebreak[3]
\begin{theorem}
\label{thm: eb main} Let $\pi: M \rightarrow S$ be an elliptic bundle as above, and let $\omega_0$ be a Gauduchon metric on $M$. Let $\omega(s)$ solve the normalized continuity equation (\ref{eqn: eb ce}).
As $s \rightarrow \infty$,
\begin{enumerate}
\item[(i)] $\displaystyle{\omega(s) \rightarrow \pi^*\omega_S}$
in the $C^0(M, \omega_0)$ topology.
\item[(ii)] $(M, \omega(s))$ converges to $(S, \omega_S)$ in the Gromov-Hausdorff topology.
\item[(iii)] The Chern-Ricci curvature of $\omega(s)$ remains uniformly bounded.
\end{enumerate}
\end{theorem}
The behavior of $\omega(s)$ mirrors the behavior of the Chern-Ricci flow on such elliptic surfaces, which was studied by Tosatti, Yang and the second-named author \cite{TWY}. Indeed (i) and (ii) hold for both equations, and the proof of Theorem \ref{thm: eb main} makes heavy use of the results and techniques of \cite{TWY}. A crucial difference is that the Chern-Ricci curvature bound was not obtained in \cite{TWY},
suggesting a possible advantage of the continuity equation in this setting. We also find some simplifications compared to \cite{TWY}.
Here are some further remarks about Theorem \ref{thm: eb main}.
\medskip
\noindent
1) We note that the Gauduchon assumption is only used to obtain the identity (\ref{TWY}) below (see \cite[Lemma 3.2]{TWY}) which is used for the bound on the potential $\varphi$ (Lemma \ref{lmm: phi bound} below).
\medskip
\noindent
2) For (i) the precise convergence we obtain is $| \omega(s) - \pi^*\omega_S|_{\omega_0} \le Cs^{-\alpha}$ for any $\alpha \in (0,1/8)$, which corresponds to the exponential convergence for the Chern-Ricci flow in \cite{TWY}.
\medskip
\noindent
3) The paper \cite{TWY} considers the metrics restricted to the fibers along the Chern-Ricci flow and obtains convergence (after rescaling) to flat metrics, making use of arguments from \cite{FZ, G2, GTZ, ShW, SW, To}. The analogous result holds for the continuity equation. Moreover, the argument of \cite[Theorem 8.2]{TW3} or \cite[Corollary 1.2]{TWY} gives an extension of Theorem \ref{thm: eb main} to all minimal non-K\"ahler elliptic surfaces, by taking a finite cover. We omit the details for the sake of brevity and to avoid repetition.
\medskip
\noindent
3) It is not even known if the Chern \emph{scalar} curvature $R$ is uniformly bounded for the Chern-Ricci flow on elliptic bundles. The bounds for $R$ proved
in \cite{TWY} were $-C \le R \le Ce^{t/2}$, where $t$ is the time parameter along the flow.
\medskip
\noindent
4) Zhang-Zhang \cite{ZZ} investigated the K\"ahler version of the continuity equation on minimal elliptic K\"ahler surfaces, including those which have non-bundle fibration structures and singular fibers and established the analogue of (ii) (cf. \cite{ST, TZ2}).
\medskip
The outline of the paper is as follows. In Section \ref{section: pre} we establish notation and state a technical but important lemma for later use. In Sections \ref{section: pf} and \ref{section: ell} we prove Theorems \ref{thm: main} and \ref{thm: eb main} respectively.
\section{Preliminaries} \label{section: pre}
Given a Hermitian metric $g= (g_{i\ov{j}})$ with associated $(1,1)$ form $\omega = \sqrt{-1} g_{i\ov{j}} dz^i \wedge d\ov{z}^j$ we write $\nabla$ for its Chern connection. The Christoffel symbols of $\nabla$ are given by $\Gamma^k_{ij} = g^{\ov{q}k} \partial_i g_{j\ov{q}}$, its torsion is given by $T^k_{ij} = \Gamma^k_{ij} -\Gamma^{k}_{ji}$ and the Chern curvature is $R_{k\ov{\ell}i}^{\ \ \ \, p} = - \partial_{\ov{\ell}}\Gamma^p_{ki}$. We will sometimes raise and lower indices in the usual way using the metric $g$.
The Chern-Ricci curvature of $g$ is the tensor $R_{k\ov{\ell}} = g^{\ov{j}i} R_{k\ov{\ell}i\ov{j}} = - \partial_k \partial_{\ov{\ell}} \log \det g$, and the associated Chern-Ricci form is
$$\operatorname{Ric}(\omega) = \sqrt{-1} R_{k\ov{\ell}} dz^k \wedge d\ov{z}^{\ell},$$
a closed real $(1,1)$ form. The Chern scalar curvature is the trace $R = g^{\ov{\ell} k} R_{k\ov{\ell}}$.
We use $\Delta$ to denote the complex Laplacian of $g$ which acts on a function $f$ by the formula $\Delta f = g^{\ov{j}i} \partial_i \partial_{\ov{j}} f$. Given another Hermitian metric $g'$ with associated $(1,1)$ form $\omega'$, we write $\tr{g}{g'} = \tr{\omega}{\omega'} = g^{\ov{j}i} g'_{i\ov{j}}$.
We note here a technical result which will be useful for later sections.
\begin{proposition} \label{propform}
Let $g=(g_{i\ov{j}})$ and $g'= (g'_{i\ov{j}})$ be Hermitian metrics with $g'_{i\ov{j}} = g_{i\ov{j}} + \partial_i \partial_{\ov{j}} \varphi$, for a smooth function $\varphi$, and define
$$f = \log \frac{\det g'}{\det g}.$$
Then
\[
\begin{split}
\Delta' \log \tr{g}{g'} = {} & \frac{1}{\tr{g}{g'}} \left\{ \frac{2}{\tr{g}{g'}} \emph{Re} \left( g'^{\ov{q}k} T^i_{ik} \nabla_{\ov{q}} \tr{g}{g'} \right) +K + \Delta f - R \right. \\
{} & + g'^{\ov{j}i} \nabla_i \ov{T^{\ell}_{j\ell}} + g'^{\ov{j}i} g^{\ov{\ell}k} g_{p\ov{j}} \nabla_{\ov{\ell}} T^p_{ik} - g'^{\ov{j}i} g^{\ov{\ell}k} g'_{k\ov{q}} (\nabla_i \ov{T^q_{j\ell}} - R_{i\ov{\ell} p \ov{j}} g^{\ov{q}p} ) \\ {} & - \left. g'^{\ov{j}i} g^{\ov{\ell}k} T^p_{ik} \ov{T^q_{j\ell}} g_{p\ov{q}} \right\},
\end{split}
\]
for $K= g^{\ov{\ell}i} g'^{\ov{j}p} g'^{\ov{q}k} B_{i\ov{j}k} \ov{B_{\ell \ov{p}q}} \ge 0$
where
$$B_{i\ov{j}k} = \nabla_i g'_{k\ov{j}} - g'_{i\ov{j}} \frac{\nabla_k \tr{g}{g'}}{\tr{g}{g'}} + T^p_{ik} g'_{p\ov{j}},$$
and $\Delta'$ is the complex Laplacian of $g'$.
\end{proposition}
\begin{proof} This identity is due to Cherrier \cite{Ch}. In this precise form it can be found in \cite[Section 9]{TW2}.
\end{proof}
\section{Proof of Theorem \ref{thm: main}} \label{section: pf}
In order to prove Theorem \ref{thm: main} we reduce
the equation (\ref{eqn: flow})
to a complex Monge-Amp\`ere equation on $M$.
Let $\tilde T \in (0,T)$. By definition of $T$ there is
a smooth function $\psi$ such that
$$\omega_0 - \tilde{T} \operatorname{Ric}(\omega_0) + \sqrt{-1} \partial \overline{\partial} \psi>0.$$
Let $\Omega$ be the volume form given by $\Omega = \omega_0^n e^{\psi/\tilde{T}}$, so that
$$\omega_0 + \tilde{T} \sqrt{-1} \partial \overline{\partial} \log \Omega = \omega_0 - \tilde{T} \operatorname{Ric}(\omega_0) + \sqrt{-1} \partial \overline{\partial} \psi>0.$$
By convexity of the space of Hermitian metrics we also have
$\omega_0 + s \partial \overline{\partial} \log \Omega > 0$ for each $s \in [0, \tilde T]$.
\begin{proposition}
\label{lmm: ma} Fix $s \in [0, \tilde{T}]$. Then there exists a metric
$\omega$ satisfying
$\omega = \omega_0 - s \operatorname{Ric}(\omega)$
if and only if
there exists a smooth function $u: M \to \mathbb{R}$ satisfying
\begin{equation} \label{ma}
\begin{split}
& \log \frac{ (\omega_0 + s \sqrt{-1} \partial \overline{\partial} \log \Omega + s \sqrt{-1} \partial \overline{\partial} u)^n }{\Omega} - u = 0, \\
& \quad \omega_0 + s \sqrt{-1} \partial \overline{\partial} \log \Omega + s \sqrt{-1} \partial \overline{\partial} u > 0 .
\end{split}
\end{equation}
\end{proposition}
\begin{proof}
Suppose first that the metric $\omega = \omega(s)$ satisfies
$\omega = \omega_0 - s \operatorname{Ric}(\omega)$. Define $u$ by $u = \log (\omega^n/\Omega)$. Then $\operatorname{Ric}(\omega) = - \sqrt{-1} \partial \overline{\partial} \log \Omega - \sqrt{-1} \partial \overline{\partial} u$ and so
$$\omega = \omega_0 + s \sqrt{-1} \partial \overline{\partial} \log \Omega + s \sqrt{-1} \partial \overline{\partial} u>0,$$
as required.
Conversely, if $u$ satisfies (\ref{ma})
then
it is straightforward to check that
$\omega := \omega_0 + s \sqrt{-1} \partial \overline{\partial} \log \Omega + s \sqrt{-1} \partial \overline{\partial} u$
satisfies $\omega = \omega_0 - s \operatorname{Ric}(\omega)$.
\end{proof}
An immediate consequence of the above proposition is the uniqueness of solutions to the continuity equation.
\begin{corollary}
If $\omega'$ and $\omega$ are two metrics solving the continuity equation (\ref{eqn: flow}) for the same $s$ in $[0,T)$ then $\omega'=\omega$.
\end{corollary}
\begin{proof} For $s=0$ there is nothing to prove. For $s \in (0,T)$, the result follows from uniqueness of solutions $u$ of the equation (\ref{ma}), a consequence of the maximum principle.
\end{proof}
We now proceed to the proof of Theorem \ref{thm: main}. First note that (\ref{ma}) is trivially solved when
$s=0$ by taking $u = \log \omega_0^n/\Omega$. Fix $s \in (0,\tilde{T}]$. Define a new function $\varphi = su$, a Hermitian metric $\hat{\omega}$ by
$$\hat{\omega} = \omega_0 + s \sqrt{-1} \partial \overline{\partial} \log \Omega,$$
and a function $F= \log (\Omega/\hat{\omega}^n)$. Then the equation (\ref{ma}) becomes
\begin{equation*} \label{ma2}
\log \frac{(\hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi)^n}{\hat{\omega}^n} = \frac{1}{s} \varphi +F, \quad \hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi>0.
\end{equation*}
Recall that $s$ here is fixed. Then Theorem \ref{thm: main} follows from the following result.
\begin{theorem}[Cherrier \cite{Ch}] \label{thmch}
Let $(M, \hat{\omega})$ be a compact Hermitian manifold, $F$ a smooth function on $M$ and $\lambda>0$ a constant. Then there exists a unique solution $\varphi$ to the equation
\begin{equation} \label{ma3}
\log \frac{(\hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi)^n}{\hat{\omega}^n} = \lambda \varphi +F, \quad \hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi>0.
\end{equation}
\end{theorem}
\begin{proof}
The complex Monge-Amp\`ere equation (\ref{ma3}) is a well-known one in the special case when $\hat{\omega}$ is K\"ahler, and was solved by Aubin \cite{A} and Yau \cite{Y}. In the Hermitian case, its solution is due to Cherrier \cite{Ch} (note that here $\lambda$ is strictly positive: for $\lambda =0$ see \cite{Y} and \cite{Ch, TW15}). For the sake of completeness, we include here a brief sketch of the proof.
We introduce a parameter $t \in [0,1]$ and consider the family of equations
\begin{equation} \label{ma4}
\log \frac{(\hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi)^n}{\hat{\omega}^n} = \lambda \varphi +tF, \quad \hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi>0.
\end{equation}
for $\varphi =\varphi(t)$. Let $E$ denote the set of those $t \in [0,1]$ for which (\ref{ma4}) has a solution. Note that $0 \in E$ since $\varphi=0$ is trivially a solution. It suffices to show that $E$ is both open and closed.
For the openness of $E$, fix $\alpha \in (0,1)$ and consider the map
$$\Psi : [0,1] \times C^{2,\alpha}(M) \rightarrow C^{\alpha}(M), \quad \Psi(t, \varphi) = \log \frac{(\hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi)^n}{\hat{\omega}^n} - \lambda \varphi -tF.$$
Assume $t_0 \in E$, and that (\ref{ma4}) has a corresponding solution $\varphi_0$. Write $\omega_0 = \hat{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi_0$ and $g_0$ for the corresponding Hermitian metric. The derivative of $\Psi$ in the second variable at $(t_0, \varphi_0)$ is the linear operator
$L: C^{2,\alpha}(M) \rightarrow C^{\alpha}(M)$ given by
$$L f = \Delta_0 f - \lambda f,$$
for $\Delta_0$ the Laplacian of $\omega_0$.
The maximum principle implies that $L$ is injective. By the Implicit Function Theorem, the surjectivity of $L$ is sufficient to show the openness of $E$.
Following an argument similar to that of \cite{TW1}, we compute the $L^2$ adjoint of this operator with respect to a specific volume form on $M$, making use of a theorem of Gauduchon \cite{Ga}. Let $\sigma$ be a smooth function such that $\omega_G := e^{\sigma} \omega_0$ is a Gauduchon metric, namely that $\partial \overline{\partial} \omega_G^{n-1}=0$. Then compute for a smooth function $h$,
$$\int_M (\Delta_0 f) h e^{(n-1)\sigma} \omega_0^n = \int_M f \left( \Delta_0 h + 2n \textrm{Re} \left( \frac{\sqrt{-1} \partial h \wedge \ov{\partial} \omega_G^{n-1}}{e^{(n-1)\sigma} \omega_0^n} \right) \right) e^{(n-1)\sigma}\omega_0^n.$$
Hence the adjoint of $L$ with respect to $e^{(n-1)\sigma}\omega_0^n$ is given by
$$L^* h = \Delta_0 h + 2n \textrm{Re} \left( \frac{\sqrt{-1} \partial h \wedge \ov{\partial} \omega_G^{n-1}}{e^{(n-1)\sigma} \omega_0^n} \right) - \lambda h,$$
and the maximum principle implies that $L^*$ is injective. By the Fredholm alternative, $L$ is surjective.
For the closedness of $E$ we need \emph{a priori} estimates on $\varphi$ solving (\ref{ma4}), independent of $t$. A uniform bound $|\varphi| \le C$ follows immediately from the maximum principle. Here and henceforth, $C$ will denote a uniform constant that may change from line to line.
Write $\omega' = \hat{\omega}+ \sqrt{-1} \partial \overline{\partial} \varphi$, and let $g'$ be the associated Hermitian metric. We will bound $\tr{\hat{g}}{g'}$ from above. By the bound on $|\varphi|$, the equation (\ref{ma4}) and the arithmetic-geometric means inequality this will imply the uniform equivalence of the metrics $\hat{g}$ and $g'$. We follow the argument of \cite[Section 9]{TW2} which uses a trick of Phong-Sturm \cite{PS} and consider the quantity
$$Q = \log \tr{\hat{g}}{g'} - A \varphi + \frac{1}{\varphi - \inf_M \varphi +1},$$
for $A$ a constant to be determined.
As in (9.4) of \cite{TW2},
\begin{equation} \label{DQ}
\Delta' Q \ge \Delta' \log \tr{\hat{g}}{g'} + A \tr{g'}{\hat{g}} + \frac{2|\partial \varphi|^2_{g'}}{(\varphi - \inf_M \varphi+1)^3} - An -n.
\end{equation}
Next we apply Proposition \ref{propform} with $f = \lambda \varphi + tF$ to obtain
\begin{equation} \label{Dp}
\Delta' \log \tr{\hat{g}}{g'} \ge \frac{2}{(\tr{\hat{g}}{g'})^2} \textrm{Re} (g'^{\ov{q}k} \hat{T}^i_{ik} \hat{\nabla}_{\ov{q}} \tr{\hat{g}}{g'}) -C \tr{g'}{\hat{g}} -C,
\end{equation}
noting that $\hat{\Delta} f \ge -C$ since $\hat{\Delta} \varphi > -n$.
We compute at a point $x_0 \in M$ at which $Q$ achieves its maximum. At $x_0$ we have $\partial Q=0$ and, assuming without loss of generality that $\tr{\hat{g}}{g'}$ is large compared to $A$, we obtain
\begin{equation} \label{useful}
\left| \frac{2}{(\tr{\hat{g}}{g'})^2} \textrm{Re} (g'^{\ov{q}k} \hat{T}^i_{ik} \hat{\nabla}_{\ov{q}} \tr{\hat{g}}{g'}) \right| \le \frac{ |\partial \varphi|^2_{g'}}{(\varphi - \inf_M \varphi+1)^3} + C \tr{g'}{\hat{g}},
\end{equation}
recalling that $| \varphi| \le C$. Combining (\ref{DQ}), (\ref{Dp}) and (\ref{useful}) and choosing $A$ sufficiently uniformly large, we obtain that $\tr{g'}{\hat{g}}\le C$ at $x_0$, and an upper bound for $Q$ follows. This implies an upper bound for $\tr{\hat{g}}{g'}$ on $M$ and hence the uniform ellipticity of the equation (\ref{ma4}). Then $C^{2,\alpha}$ estimates for $\varphi$ follow from the Evans-Krylov theory \cite{E,K, Tr} or \cite{TWWY}, and higher order estimates for $\varphi$ follow from a standard bootstrap procedure.
\end{proof}
\section{Elliptic bundles} \label{section: ell}
In this section we give a proof of Theorem \ref{thm: eb main}.
As in the introduction, let
$\pi: M \rightarrow S$ be an elliptic bundle
over a Riemann surface $S$ of genus at least $2$,
and let $\omega_0$ be a Gauduchon metric on $M$.
We follow the notation used in \cite{TWY},
and use several important facts established there.
For convenience we restate the relevant facts here and
refer the reader to the paper for further details.
Note that when comparing our notation here with that
in \cite{TWY}, our quantity $s$ relates to the quantity $t$
in that paper by the equation $1+s = e^t$.
Given $y \in S$ we denote by $E_y := \pi^{-1}(y)$
the fiber over $y$, which by assumption is
isomorphic to a torus.
There is a smooth function $\rho : M \to \mathbb{R}$ such that
the form
\[
\omega_{\textrm{flat}} := \omega_0 + \sqrt{-1} \partial \overline{\partial} \rho
\]
has the property that its restriction to each fiber
$E_y$ is the unique flat metric on $E_y$ in the cohomology class
of $\omega_0 |_{E_y}$.
We refer to $\omega_{\textrm{flat}}$ as the \emph{semi-flat form}.
It is not necessarily a metric since it may not be
positive definite on $M$.
We denote by $\omega_S$ the pullback $\pi^* \omega_S$ of the unique
K\"ahler-Einstein metric on $S$.
This form lies in $-c_1(M)$;
fix the volume form $\Omega$ which satisfies
$\sqrt{-1} \partial \overline{\partial} \log \Omega = \omega_S$ and
$\int_M \Omega = 2 \int_M \omega_0 \wedge \omega_S$. By \cite[Lemma 3.2]{TWY} we have
\begin{equation} \label{TWY}
\Omega = 2 \omega_{\textrm{flat}} \wedge \omega_S.
\end{equation}
From the equation $\sqrt{-1} \partial \overline{\partial} \log \Omega = \omega_S$, every Hermitian metric has Chern-Ricci form equal to $-\omega_S + \sqrt{-1} \partial \overline{\partial} \psi$ for some function $\psi$. Since $\omega_S \ge 0$, Theorem \ref{thm: main} immediately implies the following (cf. \cite[Theorem 1.5]{TW2}).
\begin{lemma} \label{lemmamax}
The maximal existence interval for the continuity equation with any initial metric on $M$ is $[0,\infty)$.
\end{lemma}
We use the \emph{normalized continuity equation} (\ref{eqn: eb ce}), namely
$\omega=\omega(s)$ is a family of Gauduchon metrics solving:
\begin{equation}
\label{ce2}
(1+s) \omega
= \omega_0 - s \textrm{Ric}(\omega), \quad s \in [0,\infty).
\end{equation}
We set
\begin{equation}
\label{eqn: omega tilde defn}
\tilde{\omega} = \tilde{\omega}(s) = \frac{1}{1+s} \omega_{\textrm{flat}}
+ \frac{s}{1+s} \omega_S.
\end{equation}
Note that $\tilde{\omega}$ may not be positive definite for every $s > 0$,
but that it will be for every $s$ sufficiently large.
We will use $\tilde{\omega}(s)$ as a path of reference metrics to
reduce (\ref{ce2}) to a complex Monge-Amp\`ere equation. We claim that (\ref{ce2}) is equivalent to
\begin{equation} \label{mae}
\log \frac{(1+s)(\tilde{\omega}+\sqrt{-1} \partial \overline{\partial} \varphi)^2}{\Omega} = \frac{1+s}{s} \varphi + \frac{1}{s} \rho, \quad \omega = \tilde{\omega}+\sqrt{-1} \partial \overline{\partial} \varphi>0
\end{equation}
Indeed, if $\omega$ solves (\ref{ce2}) then define $\varphi$ by
$$(1+s)\varphi = s \log \left(\frac{ (1+s)\omega^2}{\Omega} \right) -\rho,$$
and then applying $\sqrt{-1} \partial \overline{\partial}$ to both sides and rearranging we obtain
$$(1+s) \left( \tilde{\omega} + \sqrt{-1} \partial \overline{\partial} \varphi \right) = \omega_0 - s \operatorname{Ric}(\omega),$$
from which it follows that $\omega = \tilde{\omega}+\sqrt{-1} \partial \overline{\partial} \varphi$. Likewise, if $\varphi$ solves (\ref{mae}) then $\omega:= \tilde{\omega}+\sqrt{-1} \partial \overline{\partial} \varphi$ solves (\ref{ce2}).
We now turn to the proof of Theorem \ref{thm: eb main}, by establishing uniform estimates for $\omega=\omega(s)$. Note that we may assume without loss of generality that $s$ is sufficiently large so that $\tilde{\omega}(s)$ is positive definite.
We begin with:
\begin{lemma}
\label{lmm: phi bound}
There is a uniform constant $C$ such that
\[
|\varphi| \le \frac{C}{1+s} .
\]
\end{lemma}
\begin{proof} We follow \cite[Lemma 6.7]{SW} and \cite[Lemma 3.4]{TWY}.
Since
$\Omega = 2 \omega_{\textrm{flat}} \wedge \omega_S$
we have:
\begin{equation} \label{ue}
\begin{split}
\frac{(1+s)\tilde{\omega}^2}{\Omega}
= {} & \frac{\omega_{\textrm{flat}}^2 + 2s \omega_{\textrm{flat}} \wedge \omega_S}{(1+s)\Omega} \\
= {} & 1 + \frac1{1+s}\left(\frac{\omega_{\textrm{flat}}^2}{\Omega}-1\right) \\ = {} & 1+O(1/s)
\end{split}
\end{equation}
Since $s \log (1+O(1/s))$ is bounded as $s \rightarrow \infty$ we obtain
\begin{equation} \label{elog}
\left| s \log \frac{(1+s)\tilde{\omega}^2}{\Omega} \right| \le C.
\end{equation}
Now we apply the maximum principle.
Suppose $\varphi$ achieves its maximum at a point $x_0$.
Then at $x_0$ we have $\sqrt{-1} \partial \overline{\partial} \varphi \le 0$ and hence $\omega \le \tilde{\omega}$ and $\omega^2 \le \tilde{\omega}^2$. Then by (\ref{mae}), at $x_0$,
$$\varphi \le - \frac{\rho}{1+s} + \frac{s}{1+s} \log \frac{(1+s)\tilde{\omega}^2}{\Omega} \le \frac{C}{1+s},$$
by (\ref{elog}), giving the upper bound for $\varphi$. The lower bound is similar.
\end{proof}
Next we show that the volume forms of $\omega$ and $\tilde{\omega}$ are uniformly equivalent, and in fact approach each other as $s \rightarrow \infty$.
\begin{lemma}
\label{lemma: det omega bound}
There is a uniform constant $C > 0$ such that for $s$ sufficiently large,
\[
\tilde{\omega}^2\left( 1- \frac{C}{s} \right) \le \omega^2 \le \tilde{\omega}^2 \left(1+\frac{C}{s} \right).
\]
\end{lemma}
\begin{proof}
From (\ref{ue})
we see that for $s$ sufficiently large
$$\frac{\omega^2}{\tilde{\omega}^2} = \frac{(1+s)\omega^2}{\Omega} \frac{\Omega}{(1+s)\tilde{\omega}^2} = \frac{(1+s)\omega^2}{\Omega} (1+O(1/s)).$$
But from (\ref{mae}) and Lemma \ref{lmm: phi bound}
$$\frac{(1+s)\omega^2}{\Omega} = \exp\left( \frac{1+s}{s} \varphi + \frac{\rho}{s} \right) = 1+O(1/s),$$
and the result follows.
\end{proof}
We now turn to proving that $\tr{\tilde{g}}{g}$ is uniformly bounded, where $g, \ti{g}$ are the Hermitian metrics associated to $\omega, \ti{\omega}$.
This, together with the previous lemma will show that $\omega$ and $\tilde{\omega}$ are
uniformly equivalent.
We denote by $\tilde{\nabla}$ the Chern connection of $\tilde{g}$.
Similarly we will write $\tilde{T}_{ik}^p$ and $\tilde{R}_{i {\mybar{j}} k \ov{\ell}}$
for the torsion and curvature tensors
of $\tilde{g}$, and $|\tilde{T}|_{\tilde{g}}$, $|\widetilde{\textrm{Rm}}|_{\tilde{g}}$ for their norms with respect to $\tilde{g}$. We begin with a technical lemma from \cite{TWY}.
\begin{lemma} \label{lemmaTR}
For $s$ sufficiently large,
$$
| \tilde{T} |_{\tilde{g}} \le C,
\quad | \overline{\tilde{\nabla}} \tilde{T} |_{\tilde{g}}
+ | \tilde{\nabla} \tilde{T} |_{\tilde{g}}
+ | \widetilde{\mathrm{Rm}} |_{\tilde{g}} \le C \sqrt{s},
$$
for a uniform constant $C$.
\end{lemma}
\begin{proof} See \cite[Lemma 4.1]{TWY}.
\end{proof}
In order to apply the maximum principle to $\tr{\tilde{g}}{g}$ we will use the following lemma (cf. \cite[Lemma 5.2]{TWY}).
\begin{lemma}
\label{lmm: laplacian log trace omega}
For $s$ sufficiently large,
\[
\Delta \log \tr{\tilde{g}}{g}
\ge
\frac{2}{(\tr{\tilde{g}}{g})^2}
\operatorname{Re} \left( g^{{\mybar{q}} k} \tilde{T}_{ik}^i
\tilde{\nabla}_{{\mybar{q}}} \tr{\tilde{g}}{g} \right)
- C \sqrt{s} \, \tr{g}{\tilde{g}}.
\]
\end{lemma}
\begin{proof} Write (\ref{mae}) as
$$\log\frac{(\tilde{\omega}+\sqrt{-1} \partial \overline{\partial} \varphi)^2}{\tilde{\omega}^2} = f,$$
for
$$f = \frac{1+s}{s} \varphi + \frac{1}{s} \rho + \log \frac{\Omega}{(1+s)\tilde{\omega}^2}.$$
Compute
\[
\begin{split}
\tilde{\Delta} f = {} & \frac{1+s}{s} \tilde{\Delta} \varphi + \frac{1}{s} \tilde{\Delta} \rho + \tr{\tilde{\omega}}\omega_S + \tilde{R} \\
= {} & \frac{1+s}{s} ( \tr{\tilde{\omega}}{\omega} - 2) + \frac{2(1+s)}{s} - \frac{1}{s} \tr{\tilde{\omega}}{\omega_0} + \tilde{R},
\end{split}
\]
where we used $\sqrt{-1} \partial \overline{\partial} \rho = (1+s) \tilde{\omega} - s \omega_S - \omega_0$.
Applying Proposition \ref{propform} we obtain
\[
\begin{split}
\lefteqn{\Delta \log \tr{\ti{g}}{g}} \\ \ge {} & \frac{1}{\tr{\tilde{g}}{g}} \left\{ \frac{2}{\tr{\tilde{g}}{g}} \textrm{Re} \left( g^{\ov{q}k} \ti{T}^i_{ik} \ti{\nabla}_{\ov{q}} \tr{\tilde{g}}{g} \right) -C - \frac{1}{s} \tr{\tilde{\omega}}{\omega_0}
+ g^{\ov{j}i} \ti{\nabla}_i \ov{\ti{T}^{\ell}_{j\ell}} \right. \\ {} & + \left. g^{\ov{j}i} \ti{g}^{\ov{\ell}k} \ti{g}_{p\ov{j}} \ti{\nabla}_{\ov{\ell}} \ti{T}^p_{ik} - g^{\ov{j}i} \ti{g}^{\ov{\ell}k} g_{k\ov{q}} (\ti{\nabla}_i \ov{\ti{T}^q_{j\ell}} - \ti{R}_{i\ov{\ell} p \ov{j}} \ti{g}^{\ov{q}p} ) - g^{\ov{j}i} \ti{g}^{\ov{\ell}k} \ti{T}^p_{ik} \ov{\ti{T}^q_{j\ell}} \ti{g}_{p\ov{q}} \right\},
\end{split}
\]
for a uniform constant $C$. From the definition of $\tilde{\omega}$ we see that for $s$ sufficiently large, $\omega_0 \le C s \tilde{\omega}$ and hence
$\frac{1}{s} \tr{\tilde{\omega}}{\omega_0} \le C$. Note that by Lemma \ref{lemma: det omega bound}, $\tr{g}{\tilde{g}} \ge c$ for a uniform $c>0$ and $\tr{g}{\tilde{g}}$ is uniformly equivalent to $\tr{\tilde{g}}{g}$.
Applying Lemma \ref{lemmaTR} completes the proof.
\end{proof}
We can now obtain the bound on
$\tr{\tilde{g}}{g}$.
\begin{lemma}
\label{lem: eb tr gtilde g bd}
We have
\[
\tr{\tilde{g}}{g} \le C,
\]
and hence $\omega$ and $\tilde{\omega}$ are uniformly equivalent.
\end{lemma}
\begin{proof}
As in \cite[Theorem 5.1]{TWY}, define
\[
Q = \log \tr{\tilde{g}}{g}
- A \sqrt{s}\, \varphi
+ \frac{1}{\tilde{C}+ \sqrt{s} \, \varphi}
\]
where $\tilde{C}$ is chosen so that
$\tilde{C}+ \sqrt{s}\, \varphi \ge 1$
(see Lemma \ref{lmm: phi bound}).
Now $\Delta \varphi = 2 -\tr{g}{\tilde{g}}$ and so
\[
\begin{split}
\Delta \left( - A \sqrt{s}\, \varphi
+ \frac{1}{\tilde{C}+ \sqrt{s} \, \varphi} \right)
= {} & \left( -A - \frac{1}{(\tilde{C} + \sqrt{s} \, \varphi)^2} \right)
\Delta (\sqrt{s} \, \varphi)
+ \frac{2 | \partial (\sqrt{s}\, \varphi ) |^2_g}%
{(\tilde{C} + \sqrt{s} \, \varphi )^3}
\\
{} & \ge - C A \sqrt{s} + A \sqrt{s} \, \tr{g}{\tilde{g}}
+ \frac{2 | \partial (\sqrt{s}\, \varphi ) |^2_g}%
{(\tilde{C} + \sqrt{s} \, \varphi )^3}.
\end{split}
\]
At the point $x_0$ where $Q$ achieves a maximum we have
$\partial_{\ov{q}}Q=0$ so that
\begin{align*}
\frac{2}{(\tr{\tilde{g}}{g})^2} \operatorname{Re} \left(
g^{{\mybar{q}} k} \tilde{T}_{ik}^i \tilde{\nabla}_{{\mybar{q}}} \tr{\tilde{g}}{g} \right)
={} & \frac{2}{\tr{\tilde{g}}{g}} \operatorname{Re} \left(
g^{{\mybar{q}} k} \tilde{T}_{ik}^i (A + \frac1{(\tilde{C} + \sqrt{s} \varphi)^2})
\sqrt{s} \partial_{{\mybar{q}}} \varphi \right) \\
\ge {} & - \frac{4A}{\tr{\tilde{g}}{g}}
\left| g^{{\mybar{q}} k} \tilde{T}_{ik}^i \right|_g
\left| \partial_{{\mybar{q}}} (\sqrt{s} \varphi) \right|_g \\
\ge {} & - \frac{CA^2}{(\tr{\tilde{g}}{g})^2}
\left| g^{{\mybar{q}} k} \tilde{T}_{ik}^i \right|_g^2
(\tilde{C} + \sqrt{s} \varphi)^3
- \frac{\left| \partial (\sqrt{s} \varphi) \right|_g^2}%
{(\tilde{C} + \sqrt{s} \varphi)^3} \\
\ge {} & - \frac{C' A^2}{\tr{\tilde{g}}{g}}
- \frac{\left| \partial (\sqrt{s} \varphi) \right|_g^2}%
{(\tilde{C} + \sqrt{s} \varphi)^3},
\end{align*}
where we have used Lemmas \ref{lmm: phi bound} and \ref{lemmaTR}.
Then, at $x_0$, from Lemma \ref{lmm: laplacian log trace omega},
\begin{align*}
0 &\ge \Delta Q
\ge - C A^2 + (A-C) \sqrt{s} \, \tr{g}{\tilde{g}} - CA \sqrt{s}
\ge - CA^2 + \sqrt{s} \, \tr{g}{\tilde{g}} - CA \sqrt{s}
\end{align*}
if we choose $A \ge C+1$.
So at this point $\tr{g}{\tilde{g}}$, and hence $\tr{\tilde{g}}{g}$ is bounded from above and the result follows.
\end{proof}
Next we show that $g$ and $\ti{g}$ approach each other as $s \rightarrow \infty$.
\begin{lemma} \label{lemmaconv}
For every $\alpha$ with $0<\alpha<1/4$ there is a constant $C$ such that for $s$ sufficiently large,
\begin{enumerate}
\item[(a)] $\displaystyle{\tr{\tilde{g}}{g}-2\le Cs^{-\alpha}}.$
\item[(b)] $\displaystyle{\tr{g}{\tilde{g}} -2 \le C s^{-\alpha}}.$
\item[(c)] $\displaystyle{(1- Cs^{-\alpha/2})\tilde{g} \le g \le (1+Cs^{-\alpha/2}) \ti{g}.}$
\end{enumerate}
\end{lemma}
\begin{proof} We use the idea from \cite[Proposition 7.3]{TWY}, but in our case the argument is slightly easier.
Now that $g$ and $\ti{g}$ are uniformly equivalent, it follows from Lemma \ref{lmm: laplacian log trace omega}
that
$$\Delta \tr{\tilde{g}}{g} = \tr{\tilde{g}}{g} \Delta \log \tr{\tilde{g}}{g} + \frac{ |\nabla \tr{\tilde{g}}{g}|^2_g}{\tr{\tilde{g}}{g}} \ge - C \sqrt{s},$$ where we always assume $s$ is sufficiently large. Define $\beta = 1/2 + 2\alpha <1$ and $Q = s^{\alpha} (\tr{\tilde{g}}{g} -2) - s^{\beta}\varphi$. Compute
$$\Delta Q \ge - Cs^{\alpha+1/2} + s^{\beta}(\tr{g}{\ti{g}}-2) \ge -C s^{\alpha+1/2} + s^{\beta}(\tr{\ti{g}}{g} -2) - C' s^{\beta-1},$$
where for the last inequality we used
\begin{equation} \label{trdet}
\tr{g}{\ti{g}} = \tr{\ti{g}}{g} + \left( \frac{\det \ti{g}}{\det g}-1 \right) \tr{\ti{g}}{g} = \tr{\ti{g}}{g} + O(1/s),
\end{equation}
which follows from Lemma \ref{lemma: det omega bound}.
Hence at the point where $Q$ achieves a maximum,
$$s^{\alpha}(\tr{\ti{g}}{g} -2) \le Cs^{2\alpha +1/2-\beta} +Cs^{\alpha-1} \le 2C.$$
But from Lemma \ref{lmm: phi bound}, $s^{\beta} |\varphi|$ is bounded, and hence $Q$ is bounded, giving (a).
Part (b) follows from (a) and (\ref{trdet}). Part (c) is an elementary consequence of parts (a) and (b) (see \cite[Lemma 7.4]{TWY}).
\end{proof}
Now part (i) of Theorem \ref{thm: eb main} follows from part (c) of this lemma and the definition of $\tilde{\omega}$. Part (ii) is a consequence of (i) (see \cite[Lemma 9.1]{TWY}). The next result completes the proof of Theorem \ref{thm: eb main}.
\begin{lemma}
\label{thm: chern scalar curv bd}
For $\operatorname{Ric}(\omega)$ the Chern-Ricci curvature of $\omega=\omega(s)$, we have
\[
- C\omega \le \operatorname{Ric}(\omega) \le C\omega
\]
for a uniform constant $C$.
\end{lemma}
\begin{proof}
From the continuity equation
(\ref{ce2})
we have
\[
\operatorname{Ric}(\omega) = \frac1s \omega_0 - \frac{(1+s)}{s} \omega.
\]
Hence $\operatorname{Ric}(\omega) \ge - \frac{(1+s)}{s} \omega$
giving immediately the lower bound of $\operatorname{Ric}(\omega)$.
For the upper bound we have for $s$ sufficiently large,
\[
\operatorname{Ric}(\omega)
\le \frac{1}{s} \omega_0 \le C \omega,
\]
since $\omega_0 \le Cs \tilde{\omega}$ and
$\tilde{\omega}$ and $\omega$ are equivalent.
\end{proof}
|
\section{Introduction}
About $40\%$ of planets discovered by the \textit{Kepler}\ spacecraft are in multi-planet systems \footnote{https://exoplanetarchive.ipac.caltech.edu}, some of which have small orbital period ratios between neighboring planets. The observed period ratios between adjacent pairs (Figure \ref{observation}) show that most of the period ratios are smaller than three, and there is a pile-up of period ratios around the 3:2 and 2:1 mean motion resonances (MMRs). The existence of planet pairs near first-order MMRs is often ascribed to disk migration \citep{snellgrove2001,lee2002,lee2009,wang2014}. However, we might expect the overabundance to be larger if disk migration is common, though \citet{Pan:2017} suggested that resonance capture is more difficult for smaller planets in a disk. Additionally, more planet pairs are observed on the far side of MMRs rather than being symmetrically distributed around them \citep{lissauer2011,fabrycky2014}. Numerous mechanisms have been proposed to explain the asymmetrical period ratio distribution around MMRs including dissipative resonant repulsion \citep{lithwick2012,batygin2013}, stochastic and smooth migration \citep{rein2012a}, interactions between the planets and the planetesimal disk \citep{ford2015}, in-situ growth of planets \citep{petrovich2013}, and planet-planet interactions \citep{pu2015}.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{observation.pdf}
\caption{Period ratio distribution of the \textit{Kepler}\ adjacent planet pairs. The samples are obtained from Q1-Q17 DR25 of NASA Exoplanet Archive and only confirmed planets are shown here. The yellow dots are planets in the solar system.}
\label{observation}
\end{figure}
Most of the \textit{Kepler}\ planetary systems are perceived as being quite compact, often containing multiple planets with orbital periods shorter than Mercury. However, since the dynamics of the systems are generally scale invariant (dictated primarily by orbital period ratios rather than the orbital periods themselves \citep{Rice2018}) the term ``compact'' is ambiguous. For example, compared with the physical size of the orbits of \textit{Kepler}\ planets, the planets in our solar system are relatively far apart. However, they have similar period ratios---the quantity that is more fundamental---to those observed in \textit{Kepler}\ planet pairs (shown in Figure \ref{observation}). Thus, either the \textit{Kepler}\ planetary systems are less compact or the solar system is more compact than commonly envisioned.
The relationship between the spacing of the planets in a system and the stability of that system has been studied extensively. For two planet systems, there are several stability criteria including Hill stability \citep{marchal1982,gladman1993,veras2004}, Lagrange stability \citep{barnes2006,zhou2003,barnes2007}, and the resonance overlap criterion \citep{wisdom1980,duncan1989,mardling2008,deck2013,ramos2015,Hadden:2018}. For planetary systems that include more than two planets, the dynamics becomes more complex. \citet{quillen2011} studied three-body resonance overlap in closely-spaced multi-planet systems. Most other results, however, are based on numerical simulations \citep{chambers1996,zhou2007,Smith2009,funk2010,morrison2016,Obertas2017} and typically use Equal Mutual Separation (hereafter EMS) as the system architectures, where the semi-major axis of adjacent planet pairs is determined by
\begin{equation}
a_{i+1}-a_i=K\Delta_H.
\end{equation}
and
\begin{equation}
\Delta_H=\frac{(a_{i}+a_{i+1})}{2}(\frac{m_{i}+m_{i+1}}{3M_\star})^{1/3}.
\end{equation}
where $\Delta_H$ is the mutual Hill radius, $a_i$ is the semi-major axis of the $i_{\rm th}$ planet, $m$ is the planetary mass, and $K$ is a numerical spacing parameter. In this way, the period ratio between adjacent planets is:
\begin{equation}
\frac{P_{i+1}}{P_i} \approx 1+\frac{3}{2}K(\frac{m_{i}+m_{i+1}}{3M_\star})^{1/3}.
\end{equation}
For EMS planetary systems, $K$ is constant within one system. This quantity is a key factor in determining the stability timescale $\tau$ of EMS planetary systems, where $\rm log\tau$ $\propto$ $K$ (see \citet{pu2015} for a review). Another measure of the compactness of a system is the orbital period ratios between the planets. Even with differences in planetary masses, it is clear that planet pairs with smaller period ratios are more compact and can be more strongly perturbed throughout their dynamical history than those with larger period ratios. We see in Figure \ref{observation} that there is an obvious decrease of planet pairs towards small period ratios ($<1.5$), which may be caused, at least in part, by dynamical instability---a conjecture we investigate here.
\citet{Izidoro:2017} studied the influence of dynamical instability on the period ratio distribution of multi-planet systems starting from compact resonant chains. But in this paper we conduct numerical simulations on non-EMS planetary systems with uniformly-distributed initial period ratios which we then evolve to determine the role that instability plays in shaping the final period ratio distribution. By comparing the final distribution to the observed distribution, we should gain insight not only into the effects of dynamical instability, but also into the planet formation process generally. That is, at least a portion of the difference between our simulations and the observations must be a consequence of the formation process itself, independent of the system's subsequent dynamical evolution.
We describe our simulation techniques and the initial conditions in section \ref{sec:set-up}. In section \ref{sec:stability}, we analyze the factors that influence the dynamical stability of multi-planet systems. The consequences that instability has on the period ratio distribution and a comparison of the probability density function between our simulations and the de-biased \textit{Kepler}\ observations are presented in section \ref{sec:periodratio}. Finally, our conclusions are outlined in section \ref{sec:conclusion}.
\section{Simulation setup}\label{sec:set-up}
We consider four kinds of planetary systems containing $N$ planets orbiting a one solar mass star, where $N$ ranges from two to five. We begin with samples of 1000 realizations for each kind of system. The period ratio for each adjacent pair is assigned such that the period ratios in each suite are strictly uniform between one and three. For example, for the five-planet systems, there are 4000 period ratios. We therefore generate an array with 4000 equally-spaced elements between one and three and randomly choose the period ratios in each system from that array (without replacement) until the sample of 1000 systems is complete (as opposed to drawing period ratios from a uniform distribution, which would be subject to unwanted statistical variation). Thus, our planetary systems are non-EMS and the distribution of period ratios for all adjacent planet pairs are uniform. For each system, the innermost planet has an orbital period of 10 days, consistent with the typical orbital period of planets observed by the \textit{Kepler}\ mission \citep{Thompson:2018}. The eccentricity and inclination (in radians) for each planet are drawn from a Rayleigh distribution with $\sigma=10^{-3}$:
\begin{equation}
P(x)=\frac{x}{\sigma^2} e^{-\frac{x^2}{2\sigma^2}}.
\end{equation}
The planetary masses are also Rayleigh distributed with $\sigma_m=$ 6 $m_{\oplus}$, based on the TTV mass of the \textit{Kepler}\ observations\citep{Hadden2017}. The minimal planetary mass is limited to be 1 $m_{\oplus}$. Other orbital elements are randomly distributed between 0 and $360^{\circ}$.
We integrate each system up to $10^6$ years using the $ias15$ integration scheme of the REBOUND package \citep{Rein2012,Rein2015}. This integration time is about $3.65\times10^7$ orbits of the inner-most planet, $t_0$. We include collisions in our integrations. Once the distance between two planets is smaller than the sum of their planetary radius, they merge with momentum and mass conserved. The planetary radius is calculated as $R=(m/(3m_{\oplus}))R_{\oplus}$ \citep{wu2013}.
\section{The stability criteria in multi-planet systems}\label{sec:stability}
The stability of planetary systems containing more than two planets is more challenging and less well understood than two planet systems. Here, we study the stability criteria for both two-planet systems and systems with more than two planets. The relationship between period ratio and planetary mass ($m_1+m_2$) of planet pairs in all four kinds of planetary systems after $3.65\times10^7$ $t_0$ is shown in Figure \ref{pr_mass}. We find that the stable planet pairs have period ratios either smaller than 1.05 or larger than 1.1. The two groups of planet pairs remain stable via different mechanisms, which are discussed in the following sections.
For planet pairs with period ratios larger than 1.1, two different criteria are often invoked to determine their stability, either the resonance overlap criteria \citep{wisdom1980,deck2013} or the Hill stability criteria \citep{gladman1993}. We find that both criteria are reasonable approximations to the stability cutoff, but that the resonance overlap criteria performs better (it is strictly obeyed in our simulations for period ratios larger than 1.1). For period ratios smaller than 1.05, the systems are stable if they are in the 1:1 MMR.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{pr_mass.pdf}
\caption{The total planetary mass and period ratios of all adjacent planet pairs (shown as gray dots) in our simulation after $3.65\times10^7$ $t_0$. The red curve is the first-order MMR overlap criteria of the initially circular case from \citet{deck2013}, the green curve is the Hill stability criteria from \citet{gladman1993}. The blue dot represents the Kepler-36 system \citet{carter2012}.}
\label{pr_mass}
\end{figure}
\subsection{Planet pairs in the 1:1 MMR}\label{sec:11mmr}
After an integration time of $3.65\times10^7 \ t_0$, some planet pairs with period ratios near 1 remain because they are protected by the 1:1 MMR. Co-orbital configurations have been studied extensively, especially in planet-satellite systems \citep{dermott1981,dermott19812,yoder1983,tabachnik2000,christou2011}. These insights are also applied to the problem where a terrestrial planet co-orbits with a gas giant \citep{dvorak2004,erdi2005,beauge2007}. More general problems such as two comparable planets in 1:1 resonance have also been studied \citep{nauenberg2002,laughlin2002}.
Of the stable, co-orbital planetary systems, planet pairs with initial differences of mean longitude far from $180^{\circ}$ and period ratios very close to 1 evolve in tadpole orbits (shown in the left panel of Figure \ref{orbit}), while planet pairs with period ratios slightly farther from 1 have horseshoe orbits (shown in the right panel of Figure \ref{orbit}). The fraction of tadpole orbits among all co-orbital configurations is about $25\%$. (Recall that all of these co-orbital systems were generated randomly from our distributions of initial parameter values).
Planet pairs in systems with more than two planets account for $87\%$ of all co-orbital configurations. Hence, co-orbital planets are also likely to be stable in multi-planet ($N>2$) systems. We check and find that planet pairs survived 1:1 MMR generally have period ratios $<1.03$. For planetary systems with more than two planets, the period ratio between the co-orbital pair and their closest companion should be larger than $1.33$ to ensure the stability of the co-orbital pair. The resonant angle $\phi=\lambda_2-\lambda_1$ of planet pairs in tadpole orbits (where $1$ and $2$ represent the two planets in the resonance) oscillates within a small range and one planet never crosses the $L_{3}$ Lagrange point of the other. For planet pairs in horseshoe orbits, the resonant angles oscillate over a large range $>180^{\circ}$ of values---where one planet crosses the $L_3$, $L_4$ and $L_5$ Lagrange points of the other planet.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{orbit.pdf}
\caption{ Position of one planet (shown as light blue) in the reference frame co-rotating with the other planet (shown as green dot). The host star is shown as the red dot. The left panel is the tadpole orbit and the right panel is the horseshoe orbit.}
\label{orbit}
\end{figure}
The co-orbital configuration of the two-planet case can be stable for as long as $3.65\times10^9$ $t_0$ (possibly longer), for both the tadpole and the horseshoe orbits. \citet{tabachnik2000} showed that the Earth tadpole can be stable for as long as $10^9$ years, while horseshoe orbits are generally considered less stable than tadpole orbits \citep{dermott19812}. \citet{laughlin2002} suggested that the horseshoe configuration can be stable for a long time if $(m_1+m_2)/m_{\star}\leq2\times10^{-4}$, which is the case for our simulations. Although co-orbital planets were not found by \textit{Kepler}\ \citep{janson2013}, \citet{ford2006} and \citet{leleu2017} proposed a method to detect them by combining transit and radial velocity measurements. This method may have different detection sensitivities that may enable their discoveries in the future. Nevertheless, if such planet pairs were common, they would likely have been detected by \textit{Kepler}---especially in high Signal-to-Noise cases. There are a few planet candidate systems that appear to have small period ratios such as KOI-284, KOI-521 and KOI-2248. However, these systems show signs of being false positives, or (as in the case of KOI-284) false multis---where the signal is actually from two separate planetary systems in a stellar binary \citep{lissauer2014}. Thus, we find it unlikely that co-orbital planet pairs are a common byproduct of planet formation.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{k_dis.pdf}
\caption{The distribution of $K$ between each planet pair in the two-planet system (left panel) and systems with more than two planets (right panel). The initial distributions are shown in gray and the final distributions are shown in blue. Note that there are more two-planet systems after $t=3.65\times10^7$ $t_0$ because some of them are produced by planetary systems with more than two planets where collision or ejection occurs.}
\label{kdis}
\end{figure}
\subsection{Stability of planet pairs with period ratio $>$ 1.1}
We now turn from planets in the 1:1 MMR to pairs in multi-planet systems that have larger period ratios. Previous works \citep{chambers1996,zhou2007,Smith2009,funk2010,pu2015,morrison2016,Obertas2017} have shown that the mutual separation in units of mutual hill radius, $K$, is one indicator of the instability timescale of EMS planetary systems. \citet{gladman1993} showed that for two planet systems, the minimal $K$ required to remain stable is $\sim 3.5$. Figure \ref{kdis} shows the initial and final $K$ distributions for systems with two or more planets. The initial values of $K$ are distributed between 0 and 40. After $3.65\times10^7 \ t_0$, however, $K$ of the remaining pairs are either very close to 0 or larger than the predicted stability cutoff of 3.5. For EMS planetary systems, numerical simulation results in \citet{Obertas2017} show that five planet systems can survive at least $10^9$ $t_0$ for $K\geq8.5$. The criterion $K>3.5$ between each planet pair alone can not ensure the stability of the multi-planet (N$>$2) systems. For non-EMS planetary systems, we investigate whether $K$ between each planet pair, or some other statistic derived from $K$, best characterizes the stability of the multi-planet systems in the following sections.
\subsubsection{Factors that determine the stability in multiple planet systems}
We consider three statistics derived from $K$: the minimum $K$ in a system ($K_{\rm min}$), the harmonic mean value of $K$ ($K_{\rm hmn}$), and the arithmetic mean value of $K$ ($K_{\rm avg}$).
Generally, the minimum mutual separation ($K_{\rm min}$) represents the local compactness of the planetary system, with the other two means gradually transitioning between local compactness and global compactness (the harmonic mean is the smallest of the three Pythagorean means and the arithmetic mean is the largest). The stable rates of planetary systems at different $K_{\rm min}$, $K_{\rm hmn}$, and $K_{\rm avg}$ are shown in Figure \ref{stable_rates}. Here, our measure of the stability of a planetary system is whether or not the planetary orbits remain near their initial values throughout the integration. That is, $\left|P_{f}-P_{i}\right|<0.01P_{i}$, where $P_{i}$ and $P_{f}$ represent the initial and final orbital period, respectively.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{stable_rates.pdf}
\caption{The upper, middle and lower panels show stable rates as a function of $K_{\rm min}$, $K_{\rm hmn}$ and $K_{\rm avg}$, respectively. Different colors represent different kind of planetary systems. Two planet systems are shown in light red, three planet systems are shown in yellow, four planet systems are shown in light blue and five planet systems are shown in dark red. Each kind of planet pair is divided into 100 groups based on $K_{\rm min}$, $K_{\rm hmn}$ and $K_{\rm avg}$, respectively, and the stable rates are calculated in these small groups. }
\label{stable_rates}
\end{figure}
We see that the stable rates increase with all three statistics $K_{\rm min}$, $K_{\rm hmn}$, and $K_{\rm avg}$. Once $K_{\rm min}$, $K_{\rm hmn}$ or $K_{\rm avg}$ exceeds a particular critical value (noted as $K_{\rm min, crit}$, $K_{\rm hmn, crit}$ and $K_{\rm avg, crit}$, respectively), the stability rates are $100\%$, meaning that the planetary system is stable for at least $3.65\times10^7 \ t_0$. The critical values for the three statistics of $K$ are shown in Table \ref{tab1}. For the two-planet systems, the critical values of the $K$'s are all near 4 with uncertainties of 0.4---slightly larger than the traditional 3.5. Part of the reason for this larger cutoff may be that we have very few samples of planetary systems around 3.5, and our stability criteria is quite restrictive. Increasing the number of planets within one system increases the critical values of $K_{\rm min}$, $K_{\rm hmn}$, and $K_{\rm avg}$.
Among all stable planetary systems, the fraction of planetary systems with $K_{\rm min}>K_{\rm min, crit}$, $K_{\rm hmn}>K_{\rm hmn, crit}$ and $K_{\rm avg}>K_{\rm avg, crit}$ are shown in Table \ref{tab1}, respectively. A large fraction of stable systems above the critical value indicates a good stability criterion since it places a better constraint on the stable spacings of planets. We see from Table \ref{tab1} that $K_{\rm avg}$ is not a good statistic to determine the stability of multi-planet systems, especially for systems containing four or more planets---nearly 85\% of stable systems have separations smaller than the threshold where all systems are seen to be stable.
For $K_{\rm min}$ and $K_{\rm hmn}$, only about 25\% of the stable systems are below the threshold. To better determine which of these statistics best constrains the dynamics of the system, we move on to compare the instability timescales determined by $K_{\rm min}$ and $K_{\rm hmn}$ between our samples and the EMS systems.
\begin{table}
\begin{center}
\caption{Critical values of different statistics of $K$ for planetary systems containing different number of planets (the second, third and fourth row) and fraction of planetary systems meeting the criteria among all stable planetary systems (the fifth, sixth and seventh row).\label{tab1}}
\resizebox{\linewidth}{!}{%
\begin{tabular}{ccccc}
\hline
\hline
&N=2&N=3&N=4&N=5\\\hline
$K_{\rm min, crit}$&$4.0\pm0.4$&$6.6\pm0.3$&$7.1\pm0.3$&$7.1\pm0.3$\\\hline
$K_{\rm hmn, crit}$&$4.0\pm0.4$&$6.8\pm0.4$&$11.6\pm0.3$&$13.6\pm0.3$\\\hline
$K_{\rm avg, crit}$&$4.0\pm0.4$&$16.2\pm0.4$&$20.2\pm0.3$&$22.9\pm0.3$\\\hline
$K_{\rm min}>K_{\rm min, crit}$&$99.2\%\pm0.5\%$&$88.2\%\pm1.0\%$&$80.7\%\pm2.0\%$&$75.8\%\pm2.7\%$\\\hline
$K_{\rm hmn}>K_{\rm hmn,crit}$&$99.2\%\pm0.5\%$&$97.9\%\pm1.3\%$&$82.9\%\pm2.0\%$&$72.5\%\pm2.4\%$\\\hline
$K_{\rm avg}>K_{\rm avg,crit}$&$99.2\%\pm0.5\%$&$69.1\%\pm2.5\%$&$38.9\%\pm2.4\%$&$15.9\%\pm1.5\%$\\\hline
\end{tabular}}
\end{center}
\end{table}
\subsubsection{Lower limit of instability timescale determined by $K_{\rm min}$}
In this section, we calculate the instability timescale when a first close encounter occurs in our simulations. Figure \ref{tk1} compares our results to the results from EMS systems in \citet{chambers1996}, \citet{Obertas2017} and \citet{Rice2018}. We can see that the instability timescales for these systems have a large scatter, even at the same $K_{\rm min}$, $K_{\rm hmn}$, or $K_{\rm avg}$. However, the lower limit of the instability timescale at different $K_{\rm min}$ is consistent with the value calculated in EMS systems. At $K_{\rm min}>2$, we can determine a lower bound on the stability timescale for the system. When using $K_{\rm hmn}$ and $K_{\rm avg}$, the estimated instability timescale no longer yields a good lower bound on the measured timescale, especially for $10<K_{\rm avg}<20$ where the instability timescale varies between 1 and $3.65\times10^7 \ t_0$. However, we can estimate the upper bound of the instability timescale with $K_{\rm hmn}$ or $K_{\rm avg}$. A combination of $K_{\rm min}$ and $K_{\rm hmn}$ (or $K_{\rm avg}$) would yield the variation in instability timescale. \citet{pu2015} also conduct numerical simulations on non-EMS systems, they drew the value of K from a Gaussian distribution with mean value $K_{\rm mean}$ and variance $\sigma_K$, and they found that the instability timescale is well determined by $K_{\rm mean}-0.5\sigma_K$. Since $K_{\rm mean}-0.5\sigma_K$ is close to the smaller values of K in each system, our results are consistent.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{tk1.pdf}
\caption{The instability timescale $t/t_0$ vs. $K_{\rm min}$ (left panel), $K_{\rm hmn}$ (middle panel) and $K_{\rm avg}$ (right panel). The upper, middle and lower panel shows the results of three, four and five planet systems, respectively. Systems that have close encounters are shown as dots, while systems that are stable within $3.65\times10^7$ $t_0$ are shown as triangles. Dark lines show the instability timescale (from previous works) as a function of $K$ for EMS planetary systems. For three planet systems $log_{10}t/t_0=1.65K-3.71$ \citep{chambers1996}, for four planet systems $log_{10}t/t_0=1.10K-1.75$ \citep{Rice2018}, for five planet systems $log_{10}t/t_0=0.964K-1.289$\citep{Obertas2017}. }
\label{tk1}
\end{figure}
As the $K$'s in our simulations are mostly distributed between 10 and 30 (as shown in Figure \ref{kdis}), with only a few examples of $K$'s between 2.5-10, we carry out a set of additional simulations focusing on small separations.
We simulate three groups of five-planet systems using the same distribution of orbital elements and planetary mass as those described in Section \ref{sec:set-up}, except now we change their distribution of orbital periods. We consider three different scenarios. Group 1: we adopt EMS systems where the $K$'s are uniformly distributed between 2.5 and 10. Group 2: the median values of $K$ in each planetary system ($K_m$) are uniformly distributed between 2.5 and 10 and the standard deviation of $K$ in each system is $\sigma_K=0.3 \ K_m$. Group 3: similar to Group 2, but with $\sigma_K=0.6 \ K_m$. The instability timescales of the three groups are shown in Figure \ref{tk2}.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{tk2.pdf}
\caption{The instability timescale $t/t_0$ vs. $K_{\rm min}$ (left panel), $K_{\rm hmn}$ (middle panel) and $K_{\rm avg}$ (right panel) for the five-planet systems. Systems that have close encounters are shown as dots, while systems that are stable within $3.65\times10^7$ $t_0$ are shown as triangles. Upper panels: Group 1, the EMS systems. Middle panels: Group 2, median values of $K$ ($K_m$) in each planetary system are uniformly distributed between 2.5 and 10, the standard deviation of $K$ in each system is $\sigma_K=0.3$ $K_m$. Lower panels: Group 3, similar to middle panels, but with $\sigma_K=0.6$ $K_m$. The dark lines represent the instability timescale as a function of $K$, i.e.,$log_{10}t/t_0=0.964K-1.289$ from \citet{Obertas2017}.}
\label{tk2}
\end{figure}
For Group 1, the EMS case, our results agree with those of \citet{Obertas2017}, although the planets in our simulations have different masses. For Groups 2 and 3, the scatter in the instability timescale is as large as four orders of magnitude, much larger than what is observed with the equal spacing of Group 1. Also, the scatter in instability timescale increases with the scatter of $K$ in each planetary system. Despite the large scatter, we can still approximate the lower limit of the instability timescale with $K_{\rm min}$ using the relationship between $log_{10}t/t_0$ and $K$ in EMS systems. The variation of instability timescale of one planetary system can be roughly determined with $K_{\rm min}$ and $K_{\rm hmn}$ (or $K_{\rm avg}$). Nevertheless, with the ability of determining the lower limit of the instability timescale, $K_{\rm min}$ performs better than $K_{\rm hmn}$ and $K_{\rm avg}$ as a stability criteria in combination with the analysis in previous paragraphs.
\section{Period ratio distribution}\label{sec:periodratio}
Planet pairs with small $K_{\rm min}$ likely collide with each other or are scattered, and as a consequence, the architecture of multi-planet systems are sculpted by their dynamical evolution. Here, we study the final period ratio distribution of the systems after $3.65\times10^7 \ t_0$. The initial and final period ratio distributions of all planetary systems are shown in the upper panel of Figure \ref{pr_Rl}.
We can see in that figure that planet pairs with period ratios smaller than 1.05 or larger than 1.1 remain stable. Planet pairs with period ratios near 1 are protected by the 1:1 MMR, as discussed in Section \ref{sec:11mmr}. The number of stable planetary systems increases with period ratio between 1.1 up to a value near 1.33, after which the distribution is almost flat. Additionally, we see that there are dips on the near side and peaks on the far side of the first-order MMRs, including 2:1, 3:2, 4:3, 5:4, 6:5, and 7:6. This result is similar to the observed period ratio distribution (lower panel of Figure \ref{pr_Rl}), except that the width and depth of the gap on the near side of the MMRs are smaller than those in the observation. Also, there is no significant feature at period ratio of 2.17 in the simulation. Period ratio distribution from Pu $\&$ Wu (2015) also shows asymmetry features around MMRs, but there is no obvious peaks on the far side of MMRs in their simulations. We investigate how these features were produced in the following paragraphs.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{pr_Rl.pdf}
\caption{ The upper panel is the initial (shown in gray) and final (shown in purple) period ratio distribution from our simulations described in Section \ref{sec:set-up} where the planetary mass are Rayleigh distributed with $\sigma_m=6$ $m_{\oplus}$. The middle panel is the initial (gray) and final (purple) period ratio distribution from the simulations where the planetary mass are Rayleigh distributed with $\sigma_m=30$ $m_{\oplus}$. The lower panel is the period ratio distribution of the confirmed \textit{Kepler}\ planet pairs. The vertical dashed lines indicate planet pairs near the first order MMRs and period ratio of 2.17. $t_0$ is the initial orbital period of the inner most planet. }
\label{pr_Rl}
\end{figure}
\subsection{Period ratio asymmetry near first-order MMR}
The behavior of two accreting planets near the first order MMRs 2:1 and 3:2 has been studied by \citet{petrovich2013}. They found that the period ratio distribution develops an asymmetric dip-peak structure near the resonance. In our simulations, this feature appears in both two planet systems and systems with more than two planets, although the planetary mass is fixed during the evolution. We find that planet pairs with initial period ratios near MMR are likely to have final period ratios larger than their initial values.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{pr_evo.pdf}
\caption{ Upper panel: The evolution of the period ratio (shown as the gray curve) with initial value of 2.0 (shown as the red horizontal line) in one of the two-planet systems. Lower panel: the median value of the period ratio $\mathcal{P}_m$ during $3.65\times10^7 \ t_0$ with different total planetary mass $m_1+m_2$. The total mass is randomly split between $m_1$ and $m_2$. Other orbital elements are the same with the example in the upper panel.}
\label{pr_evo}
\end{figure}
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{SP_Rl.pdf}
\caption{The average difference $\mathcal{P}_s$ between the period ratio during the evolution and the initial period ratio at different initial period ratios in different planetary systems. The vertical dashed lines indicate the position of first order MMRs. $\mathcal{P}_s>0$ indicates that the planet pair is more likely to be on the far side of the initial period ratio than the near side. }
\label{fig8}
\end{figure}
One example of a planet pair in the two-planet system with an initial period ratio of 2.0 is shown in the upper panel of Figure \ref{pr_evo}. As its period ratio evolves due to mutual interaction, the pair tends to stay on the far side of the 2:1 MMR. To better describe this property, we define the average difference between the period ratio during the evolution and the initial period ratio as $\mathcal{P}_s=\sum^{n}_{i=1}(pr_i-pr_0)/n$, where $n$ represents the number of data that is output during the simulation, $pr_i$ represents the period ratio of the $i_{\rm th}$ output from the simulation, and $pr_0$ represents the initial period ratio. If $\mathcal{P}_s>0$, then the period ratio is more likely to be larger than its initial value. We show $\mathcal{P}_s$ at different period ratios for two, three, four, and five planet systems in Figure \ref{fig8}. We find that there are significant peaks of $\mathcal{P}_s$ at period ratios 7:6, 6:5, 5:4, 4:3, 3:2, and 2:1---especially for the two planet systems. A consequence of this feature is that whenever we measure the period ratio distribution, there is excess probability that period ratios initially on the near side of the MMRs will be seen on the far side.
\citet{petrovich2013} proposed that the equivalent width of the peaks/dips is proportional to the planetary mass. To verify this conclusion, we choose the same two-planet system shown in the upper panel of Figure \ref{pr_evo} to conduct an additional set of simulations. We slowly increase the total mass of the two planets and calculate the median value of the period ratio $\mathcal{P}_m$ during the evolution. The total mass is randomly split between the two planets. We find that $\mathcal{P}_m$ does increase with the planetary mass, in agreement with their work (see the lower panel of Figure \ref{pr_evo}).
They also suggest that planetary mass should be in the range of $20-100$ $m_{\oplus}$ in order to explain the structure near 3:2 and 2:1 MMRs in the \textit{Kepler}\ observation. Such masses are much larger than the masses we use in our simulations and are larger than the planetary mass obtained for the typical \textit{Kepler}\ system as measured with transit time variations \citep{Hadden2017}. To verify that this dip-peak structure persists over a longer evolution time, we integrate the five-planet systems up to $3.65\times10^8 \ t_0$ for a comparison. We found that the two results are similar as all of the features remain (except for an additional 18 systems that go unstable).
\subsubsection{Varying the planetary mass distribution}
As mentioned above, the planetary masses in our simulations are too small to fully explain the observations with this mechanism. In this section, we used Rayleigh distributed planetary masses with $\sigma_m=30 \ m_{\oplus}$. (The average value of planetary mass is increased by a factor of five from the previous section.) The other parameter distributions remain the same. The final period ratio distribution for these simulations is shown in the middle panel of Figure \ref{pr_Rl}. We see that the widths and depths of the dips near the first order MMRs are larger than those of the smaller planetary mass with $\sigma_m=6 \ m_{\oplus}$---especially for the 2:1 and 3:2 MMRs where the gap on the near side is only slightly smaller than the observations.
Additionally, the increase of planetary mass by a factor of 5 leads to a decrease of $K$ by a factor of 1.7 from the original values, substantially reducing the instability timescale (particularly for planet pairs with period ratios between 1.1 and 1.5). Since the planetary masses observed by \textit{Kepler}\ are rarely this large, the mechanism we present here can only account for a portion of the observed asymmetry in the period ratio distribution near MMR. In addition, we note that the shallower period ratio distribution for small period ratios could be used to constrain the planetary masses observed in \textit{Kepler}\ systems---though the constraint from TTV observations is likely more stringent.
\subsubsection{Varying the eccentricity distribution}\label{ecc}
We now consider the effects of larger initial eccentricities and inclinations. In previous sections, the orbits of the planets are nearly circular and co-planar with eccentricities and inclinations $\sim 10^{-3}$. Here, we use eccentricity and inclination distributions of $\sigma_{e,i}=0.01$ and $\sigma_{e,i}=0.05$. Again, other parameter distributions remain the same with those described in Section \ref{sec:set-up}. The final period ratio distributions are shown in Figure \ref{pr_Rl_e}. The results of the simulations with $\sigma_{e,i}=0.01$ are similar to those of $\sigma_{e,i}=1\times10^{-3}$, except that the peak on the far side of the 3:2 MMR is not as strong. However, when $\sigma_{e,i}$ increases to 0.05, peaks and dips near MMRs almost disappear, as shown by \citet{Xie:2014} that the asymmetry features around MMRs will become weaker with increasing eccentricity.
Planets with higher eccentricities tend to be more unstable when their period ratios are between 1.1 and 1.7 than in the small eccentricity and inclination cases. This is both because the increased eccentricity yields a higher probability that two planets have close encounters and because the resonance width increases with eccentricity \citep{deck2013,Hadden:2018}, so resonance overlap is more likely to occur. We compare our simulations to the resonance overlap criteria from \citet{Hadden:2018} and find that our results conform to that stability criteria. Thus, distributions of eccentricity and inclination with $\sigma_{e,i}=0.05$ are too large for planet pairs to produce the observed features. Moreover, the larger eccentricities and inclinations yield a period ratio distribution that is more shallow between 1.1 and 1.5 than the observations (similar to what occurred with larger mass planets from the previous section). We investigate the constraints that can be placed on the eccentricities from this feature in a later section.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{pr_Rl_e.pdf}
\caption{The top three panels show the initial (light gray) and final (purple) period ratio distributions of our simulations with Rayleigh distributed eccentricity and inclination $\sigma_{e,i}=10^{-3}$, $\sigma_{e,i}=0.01$, $\sigma_{e,i}=0.05$, respectively. The bottom panel is the Kepler period ratio distribution. }
\label{pr_Rl_e}
\end{figure}
\subsection{The probability density function of the period ratios}\label{subsec:pdf}
\subsubsection{De-biased period ratios of the \textit{Kepler}\ planets}
We have shown that (at least a portion of) the asymmetry feature near MMRs can be produced via planetary dynamics originating from a distribution that lacks those features. In this section, we compare the probability density function (PDF) of period ratios between the observed \textit{Kepler}\ data and our simulations. As expected, \textit{Kepler}\ observations contain geometric bias and pipeline incompleteness \citep{ragozzine2010,borucki2011,lissauer2011,ciardi2013,steffen2015,coughlin2016,brakensiek2016}. \citet{steffen2015} suggests that the influence of pipeline incompleteness, compared to the geometric bias, is the smaller of the two effects so we only consider the geometric bias here.
We have a total of 583 confirmed planet pairs with period ratio $<5$ from the Q1-Q17 DR25 catalog. To avoid the influence of very long period planets, which can significantly affect the distribution if not treated correctly, we cut off the sample with $a/R_{\star}<150$. According to previous studies \citep{lissauer2011,fang2012,Tremaine2012,fabrycky2014}, \textit{Kepler}\ multi-planet systems are rather flat, so we assume the mutual inclination of planets in a system are Rayleigh distributed with $\sigma\sim$ $1.5^{\circ}$, similar to \citet{steffen2015}. We use the CORBITS algorithms from \citet{brakensiek2016} to calculate the probability of detecting the outer planet given that the inner planet is detected. The inverse of the probability is adopted as the weight of the planet pair. Finally, we construct a kernel density estimator of the period ratio distribution. For each period ratio, we use a Gaussian distribution with the median value $\mu$ equal to the period ratio $Pr$ and the standard deviation $\sigma$ to be $0.00005Pr$. The total area of the Gaussian distributions is normalized to 1.
The PDF of the observed period ratio and the de-biased period ratio distributions are shown in the lower panel of Figure \ref{pdfb|a}. After de-biasing, the peaks near 3:2 and 2:1 MMRs persist, but they are not as significant as the original ones (especially for the peak near the 3:2 MMR) which are also seen in Figure 4 of \citet{brakensiek2016}. We calculate the weight of each period ratio as the inverse of transiting probability of the outer planet given that the inner planet is transiting. Additionally, we discuss another weighting scheme---which uses the inverse probability for both planets transiting the host star (rather than the conditional probability)---in the Appendix.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{comp_001.pdf}
\includegraphics[width=\columnwidth]{e8.pdf}
\caption{Upper panel: the probability density function (PDF) of the period ratio for different kind of planetary systems described in Section \ref{sec:set-up} after $3.65\times10^7$ $t_0$. For a comparison, the PDF of all planetary systems are also shown (the red curve). Lower panel: The comparison between the observed and simulated period ratio distributions. The PDF of the original period ratio distribution from \textit{Kepler}\ observations is shown in gray. PDF of the period ratio from de-biased \textit{Kepler}\ observations is shown in light dark. Yellow represents the smoothed PDF of the de-biased \textit{Kepler}\ observations. Red represents the re-normalized PDF of the period ratio from the simulation with evolution time of $3.65\times10^7$ $t_0$, while orange represents the re-normalized PDF of the period ratio of four-planet systems with an integration time of $3.65\times10^9$ $t_0$. The gray vertical lines indicate period ratios at 4:3, 3:2, 2:1 and 2.17.}f
\label{pdfb|a}
\end{figure}
\subsubsection{Comparison of the PDF between observation and simulation}\label{sec:com}
In order to compare the observed \textit{Kepler}\ period ratio distribution with our simulation (described in Section 2) results, we smooth the PDFs of both samples with suitable bandwidth. The bandwidth is chosen as the smallest value that gives a unimodal distribution. The smoothed PDFs of the four kinds of planetary systems are shown in the upper panel of Figure \ref{pdfb|a}. For a comparison, the PDF of all planetary systems is also shown. We see that the shape of the PDF for the four-planet systems is very close to that of all planetary systems. The comparison between the smoothed PDFs of all samples in our simulations and the observation is shown in the lower panel of Figure \ref{pdfb|a}. Note that the decrease of PDF at period ratio $>2.75$ is caused by the smoothing method and is not necessarily physical. Moreover, while the period ratios in our simulations are distributed only between one and three, we can reasonably assume that period ratios larger than three will remain stable and that the final distribution will likely match the initial distribution for any simulated system in that regime. We re-normalize the PDF of our simulation such that the largest value of the PDF of the observations and our simulation coincide. (That is, we increase the height of the simulated distribution so that it matches the overall height of the observations.) Differences in the two distributions following this modification should indicate period ratios where planets are under-represented relative to what dynamical stability would otherwise allow.
We find that the PDF of our simulation and the observations (the red and yellow curves in the lower panel of Figure \ref{pdfb|a}) roughly coincide between period ratios of 1.5 and 2.1. For the deficit of planet pairs with period ratios between 1.1 and 1.5, the data show fewer systems than what our simulations suggest could survive. However, given our limited integration time, there may be some residual instabilities that have not had time to manifest. Rather than continuing to integrate all planetary systems to a longer time, we simulate a set of four-planet systems and integrate them to $3.65\times10^9$ $t_0$ (100 times longer than the previous simulations). We choose the four-planet systems for further integration because the shape of the PDF for four-planet systems roughly resembles the shape of the PDF for the whole samples. (see upper panel of Figure \ref{pdfb|a}). The new simulations contain five hundred four-planet systems with the parameter distributions described in section 2. The re-normalized PDF of the new simulations is shown as the orange curve in the lower panel of Figure \ref{pdfb|a}. We see that the shape of the new PDF changes very little when compared to that of the previous simulations (the red curve in the lower panel of Figure \ref{pdfb|a}). Therefore, we suspect the shape of the PDF at period ratio $<1.5$ in the observation is not entirely due to instability, but may also be influenced by the initial eccentricity distribution, which we will discuss later.
For planet pairs with period ratio $> 2.1$, there is an obvious deficit in the observations when compared with the prediction of planet pairs that would otherwise survive given our simulations. Since systems with period ratios this large should be stable for very long time (i.e., longer than the age of the universe), these results indicate planet pairs do not emerge from the protoplanetary disk with those period ratios to the same degree that they do with smaller period ratios, at least for systems like those observed by \textit{Kepler} . Thus, whatever formation or dynamical processes are ongoing while the protoplanetary disk is present, the frequency of planet pairs that are produced with period ratios between 2.1 and 3 is 30-50\% lower than the frequency of those produced between 1.5 and 2.1.
The sizable fraction of planet pairs that survive in the 1:1 MMR is at odds with the lack of observed planet pairs in those orbits. This discrepancy likely indicates that planets either rarely form or are rarely driven into those configurations---if they did form, a large fraction would have survived. It is possible that such planet pairs have been missed by the transit search algorithms, but the high signal-to-noise ratios of many of the \textit{Kepler}\ detections makes this explanation difficult to justify in most cases. (Though, we recommend revisiting the \textit{Kepler}\ discoveries with this in mind.)
\subsubsection{Eccentricity of multi-planet systems when gas disk dissipates}
We showed that the observed period ratios between 1.1 and 1.5 can not be explained by the effects of instability with initial orbits that are nearly circular. But We see from Figure \ref{pr_Rl_e} that orbital eccentricity drives more planet pairs with small period ratios into instability. While the eccentricities of planets are likely to grow during the dynamical evolution following the dispersion of the gas disk, we can constrain the maximum initial eccentricity by comparing the shape of the PDFs between the observations and our simulations using different values of initial eccentricity. We conduct a set of simulations with the same orbital parameters as those described in Section \ref{sec:set-up}, except for the eccentricity and the inclinations. The integration time is $3.65\times10^7$ $t_0$. The results are shown in Figure \ref{comp_e}.
We see from these simulations that systems with initial eccentricities and inclinations $\sigma_{e,i} \simeq 0.03$ are roughly consistent with the observed period ratios between 1.1 and 1.5 while larger values of initial eccentricity do not match the profile of the observed distribution. Thus, the eccentricity and inclination distributions should have typical values $\sigma_{e,i} < 0.03$ when the gas disk dissipates. \citet{xie2016} proposed $e=0.04\pm0.04$ for multi-planet systems, which places an upper limit to the initial eccentricities around 0.04. Our prediction that $\sigma_{e,i} < 0.03$ is consistent with their limit.
\begin{figure}
\vspace{0cm}\hspace{0cm}
\centering
\includegraphics[width=\columnwidth]{comp_e.pdf}
\caption{The smoothed probability density function (PDF) of the period ratio for the observation (dark) and the re-normalized PDF of the period ratio from our simulations with different initial eccentricity and inclination distributions. }
\label{comp_e}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
In this paper, we studied non-EMS multi-planet systems to investigate their stability and the evolution of their period ratio distribution. In contrast to previous works, which assume the planets have equal mutual separation after the disk dissipates \citep{chambers1996,zhou2007,Smith2009,Obertas2017}, we begin with the premise that the orbital periods between adjacent planet pairs in multi-planet systems are uniformly distributed. Thus, any differences between the observed distribution of period ratios and the results of our simulations are likely due to some physical process other than dynamical instability.
After an evolution time of $3.65\times10^7 \ t_0$, we find that surviving planet pairs with orbital period ratios $<1.1$ are protected by the 1:1 MMR (both in two-planet systems and systems with more than two planets). These planets can be stable for $3.65\times10^9 \ t_0$ or longer whether in tadpole or horseshoe orbits. Thus, the lack of co-orbital planet pairs in the observations indicates that either such planets are difficult to detect (which seems unlikely), or are rarely produced in planetary systems similar to those seen by \textit{Kepler} . If there was a viable mechanism to produce a large population of 1:1 MMR planet pairs, many would survive and should be seen.
For planets far from the 1:1 MMR, the lower limits of their stability timescales determined by $K_{\rm min}$ are consistent with what is predicted in EMS systems. While planets in our simulations are not of equal mass, the differences between them are within one order-of-magnitude, our results should be largely unchanged as the Hill radius depends only weakly on planetary mass. Of the statistical quantities we studied to characterize instability timescales, we find that $K_{\rm min}$ performs most consistently.
Our period ratio distribution shows a dip-peak asymmetry near first order MMRs, where more planets are on the far side of the resonance than near side. We find that period ratios that are initially on the near side of these resonances are observed on the far side of the resonance more often due to their orbital evolution. This result may partly explain the observed features near MMR in the \textit{Kepler}\ data. (Period ratios farther from the first order MMRs do not show such asymmetries in their orbital evolution.) This deviation of the period ratio near MMR increases with planetary mass. \citet{petrovich2013} proposed that in order to explain the observed asymmetric structure, the planetary mass should be in the range of 20 $-$ 100 $m_{\oplus}$. However, the TTV-determined masses in \citet{Hadden2017} are too small to account for the dip-peak feature of the \textit{Kepler}\ systems. We also investigate the influence that eccentricity can have on the period ratio distribution and find that the dip-peak structure depends inversely upon the eccentricity of the planetary orbits---larger eccentricities show smaller asymmetry. A non-zero initial eccentricity distribution with $\sigma_e=0.05$ is too large to produce the dip-peak structure.
Finally, we compare the probability density function of the de-biased period ratio distribution of the \textit{Kepler}\ observation to our simulations. We find that the general shape of the period ratio distribution less than $\sim2.1$ can be explained by dynamical instability of planetary systems with non-circular orbits with initial eccentricities $\lesssim 0.03$. This same eccentricity preserves the asymmetry features near MMR while larger eccentricities simultaneously alters the resulting period ratio distribution removes the asymmetries. (We note, however, that the asymmetries near MMR may not be caused by the mechanism we present here). Local features near MMR and near 2.17 \citep{steffen2015} may require unique explanations.
We also find an obvious deficit of planet pairs with period ratios $\gtrsim 2.1$ in the \textit{Kepler}\ data (the deficit is nearly 50\% of what would survive if they were initially present). Thus, we suspect that planet pairs are either not formed as often with these period ratios, or if they are produced, that interactions with the gas disk may drive them to smaller period ratios. For example, it may be that the initial distribution of period ratios is essentially flat, but that $\sim 25$\% of the planet pairs eventually converge to period ratios between 1.5 and 2.1---producing the two-plateaus shown in Figure \ref{pdfb|a}.
\textit{Kepler}\ planetary systems are often portrayed as compact since planet pairs typically have small period ratios and orbit close to their host star. However, the criteria for describing a system this way is ill defined. Dynamical processes for planetary orbits are scale invariant, where resonance or other effects occur near certain period ratios regardless of the overall size of the system. Only when some new physical scale enters the description is the invariance broken and the dynamics changed. The results from \citet{Rice2018} indicate that dynamical effects related to instability are not markedly different between systems at 0.1 AU (where most \textit{Kepler}\ planets are found) and at 1 AU where the solar system terrestrial planets are found---though more work on this issue is warranted. The only scale where planetary system architecture is seen to change in the observations of \textit{Kepler}\ planets is when the inner planet has an orbital period less than a few days \citep[$\sim 0.05$ AU][]{SteffenFarr2013,SteffenCoughlin2016}. Moreover, period ratios observed in the solar system are similar to period ratios observed in most \textit{Kepler}\ systems. With the exception of the Jupiter/Mars ratio, solar system period ratios lie between 1.5 and 3 with the majority being less than 2.5. Thus, unless the solar system is considered to be ``compact'' there is little to suggest that the typical \textit{Kepler}\ planetary system should be so described.
This work shows that instability plays a significant role in sculpting planetary system architectures for period ratios less than 1.33 (see lower panel of Figure \ref{pdfb|a}). These results suggest that a reasonable criterion for ``compactness'' could be that for a system to be considered compact, it must contain a planet pair with a period ratio less than this value. For the \textit{Kepler}\ multiplanet systems, this criterion would classify roughly 4\% of the systems as compact (or roughly 6\% of systems containing more than two planets---which may be more representative of multiplanet systems generally).
\section*{Acknowledgements}
We thank the anonymous referee who helped us to improve this paper. Thanks for useful discussion with Daniel C. Fabrycky, Ji-Wei Xie and Songhu Wang. This work is supported by the National Natural Science Foundation of China (grant No. 11503009, 11333002, 11673011, 11661161014), Technology of Space Telescope Detecting Exoplanet and Life supported by National Defense Science and Engineering Bureau civil spaceflight advanced research project (D030201), and China Scholarship Program. JHS acknowledges support from the NASA Kepler Participating Scientist Program under grant number NNX16AK32G and the NASA Exoplanet Research Program under grant number NNX17AK94G.
|
\section*{Introduction}
\label{intro}
Thermoelectrics, solid state heat enginges which generate a voltage from an applied temperature difference or vice versa, offer enticing advantages for a wide range of technological applications.\cite{twaha_zhu_yan_li_2016} However, sub-optimal performance characteristics of active materials have limited their use in power generation. The most important characteristic is the Seebeck coefficient (S) which is the ratio between the voltage and temperature gradients $\nabla V=S\nabla T$. When used in a circuit, current flows through the thermoelectric material, and the lower the electrical resistance the greater the power which can be drawn. This is reflected by the material's power factor, $S^2\sigma$, where $\sigma$ is the electrical conductivity. Finally, the power conversion efficiency can be found by dividing the electrical power by the heat flow, and this is typically expressed as a dimensionless figure of merit: $zT=\frac{S^2\sigma}{\kappa}$ where $\kappa$ is the thermal conductivity. $\kappa$ is composed of electronic thermal conductivity $\kappa_e$ and lattice thermal conductivity $\kappa_L$, $\kappa=\kappa_e+\kappa_L$. Finally, the power factor and figure of merit change drastically with temperature and a wide temperature range is used during power conversion, from ambient temperature to high temperature. Compared to other methods of generating electricity, thermoelectrics typically have low thermal conversion efficiency (low zT) but high power density (high $S^2\sigma$). Despite the drawbacks, thermoelectric power generation is still used in applications where the reliability of solid state devices is crucial such as electricity generation in remote locations.\cite{fihelly_baxter_1970} To date, the most widespread use of thermoelectrics is cooling where sub-optimal zT values do not as heavily affect device efficiency, power densities can be quite high, a compact form factor aids insulation, and a narrower temperature range is used.\cite{zhao_tan_2014}
The list of known materials with high thermoelectric efficiency, such as Bi$_{2}$Te$_{3}$,\cite{bjap_5_386} PbTe,\cite{science_321_554} Mg$_{2}B^{IV}$ ($B^{IV}$=Si, Ge, Sn)\cite{scrmat_69_606} and others is constantly increasing. The discovery of new thermoelectric materials has historically been experimentally-driven. The small bandgap of many thermoelectric materials is of the same order as the thermal energy (kT) especially at high temperatures, and this in turn is responsible for the large swings in device performance with temperature. Small bandgap semiconductors typically have the highest figure of merit because the Seebeck coefficient tends to increase with the bandgap,\cite{jem_28_869} but both $\sigma$ and $\kappa_e$ decreases. Small bandgap semiconductors are in the regime where $\kappa_e$ and $\kappa_L$ are comparable. Materials with large unit cells are also ideal because $\kappa_L$ is minimized. A material which is limited by lattice thermal conductivity can be nanostructured to increase phonon scattering and reduce $\kappa_L$. The lattice thermal conductivity can be difficult to evaluate from first principles, but can be measured accurately by experiment. Despite the fact that experimentally-measured performance characteristics are more reliable than predicted characteristics, fast measurement cycles may not realize a material's full potential because optimal performance typically requires careful tuning through doping.
The demand for better thermoelectric materials and the stringent requirements on material properties has made the use of high-throughput calculations to screen potential materials quite appealing. A very good indication of the importance of the subject is a recent publication of high-throughput calculations\cite{jmcc_4_4414} of thermoelectric properties for more than 48000 inorganic compounds from the Material Project. From the theoretical point of view, the accurate prediction of thermoelectric properties still represents a formidable task.\cite{thermoelectric_review} Not only are many complicated physical processes involved (electronic correlations, phonons, electron-phonon interaction) but also thermoelectric materials have sufficiently large unit cells (often containing about 10 atoms or more), that the application of advanced electronic structure methods is too costly in terms of computer time/memory. As a result, the majority of theoretical works use density functional theory (DFT) to get insight on the basic electronic structure of thermoelectric materials. In this respect, a good example is the above mentioned high-throughput calculations performed in Ref. [\onlinecite{jmcc_4_4414}], where the combination of DFT (as implemented in the VASP code\cite{prb_54_11169}) and the semi-classical Boltzmann equation (as implemented in the code BoltzTrap\cite{cpc_175_67}) was applied to calculate transport properties from electronic structure. Namely, authors of the Ref. [\onlinecite{jmcc_4_4414}] have compared the calculated Seebeck coefficient (S) and the thermoelectric power factor $S^{2}\sigma$. As it follows from Ref. [\onlinecite{jmcc_4_4414}], the performance of the combination DFT-band-structure plus Boltzmann equation is qualitatively good for the Seebeck coefficient with the primary source of discrepancy being the DFT underestimation of the bandgap. DFT underestimation of the bandgap is the leading error in both S and $S^{2}\sigma$, and naturally, is the principal target for adjustments. The accuracy of the power factor is worse because evaluation of the conductivity involves an additional approximation, namely a constant and universal relaxation time $\tau$. As it follows from the Ref. [\onlinecite{jmcc_4_4414}], the results, especially for the Seebeck coefficient, can be considerably improved if one uses the "scissor" operator, i.e. rigid shift of calculated unoccupied bands relative to the occupied bands making the bandgap equal to the experimental one. The scissor operator corrects the bandgap, but leaves the band dispersion and corresponding effective masses unchanged. It is, thus, interesting to see how an application of advanced electronic structure methods (which can address the bandgap issue more directly than DFT, avoiding the "scissor" operator) works in terms of the evaluation of S.
Application of many body methods to the materials of thermoelectric importance is still rather rare,\cite{prb_88_045206,prb_93_205442,prb_82_085104,prb_90_075105,prb_88_165135,prb_82_245203} and it is limited to the simplest approach, namely one-shot GW (G$_{0}$W$_{0}$). In this approach, Green's function G and screened interaction W are evaluated based on DFT one-electron spectra and then the correction to these spectra is evaluated based on the first order self energy $\Sigma=G_{0}W_{0}$. Often additional approximations are involved, such as the plasmon pole approximation, diagonal only form of the self energy, and perturbative (first order) solution of the Dyson's equation. As it was shown in Refs. [\onlinecite{prb_90_075105,prb_93_205442,prb_88_045206}], the diagonal approximation for the self energy, often associated with G$_{0}$W$_{0}$ studies, fails in the case of Bi$_{2}$Tl$_{3}$, suggesting that wave functions (which usually are considered at the DFT level in the above approximation) undergo essential changes and one has to consider off-diagonal elements of the self energy in order to obtain a meaningful result. In this respect one can speculate that this actually might be true for thermoelectric materials in general, because as a rule, they possess small (0.1-0.3eV) bandgaps and DFT usually is unable to reproduce this bandgap, resulting in a metallic spectrum. The transition from a metal (in DFT approximation) to a semiconductor (in GW based methods) may be accompanied with a reconstruction of wave functions, which should not be neglected. Developing this idea, it is natural to assume that the effects of self-consistency might be more essential in this class of materials as compared to the wide bandgap semiconductors (such as Si, GaAs, ...). Semiconductors like Si are semiconductors already at the DFT level. GW approaches corrects the bandgap without a noticeable change in the wave functions. In thermoelectric materials, on the other hand, strong restructuring in wave functions discovered in a one shot study, might continue during subsequent iterations. This revelation makes the application of self consistent GW-based methods appealing.
Additionally, the small bandgap of thermoelectric materials can cause breakdown of another widely used approximation. Namely, the temperature dependence of transport properties. Frequently, the temperature-dependence of transport quantities is considered as coming entirely from the change in carrier concentrations with temperature, however, the chemical potential can shift as well and the former approximation would implying charging of the system. For small bandgap systems this approximation breaks down.
Temperature-dependent electrical conductivity of a recently synthesized\cite{cm_30_4207} material CoAsSb matched an activated band insulator with bandgaps derived from activation energies. However there are transitions between two different bandgaps. From 325-475 K, the bandgap was 0.124 eV whereas from 575-975 K the bandgap was 0.256 eV. The peak Seebeck coefficient (-134 $\mu V/K$) occurs in the former range at 450 K. The Goldsmid-Sharp bandgap (GS gap) $E_{GS}=2*S_{max}*T_{max}$ is 0.12 eV. The actual composition which was prepared and explored experimentally was CoAsSb$_{0.883}$.
Theoretical study of this material \cite{cm_30_4207} consisted in the evaluation of the electronic structure in the local density approximation (LDA) with subsequent application of BoltzTrap to calculate transport properties assuming fixed chemical potential. LDA resulted in a metallic band structure contradicting with experimental results.\cite{cm_30_4207} Consequently there was a mismatch of the calculated transport properties in accordance with general tendencies discovered in high-throughput calculations.\cite{jmcc_4_4414} To compare ways of resolving the bandgap issue, authors of the Ref. [\onlinecite{cm_30_4207}] applied the modified Becke-Johnson functional mBJ (of semiempirical nature) and obtained a small bandgap of 0.074eV for the LDA-optimized geometry. For the experimentally-measured geometry, even with the mBJ functional, the band structure was gapless. The authors speculated that the presence of a bandgap in experiments can be explained by the influence of the defects.
Thus, a natural continuation of the study of thermoelectric properties of CoAsSb and of the metal to small bandgap semiconductor problem is to calculate the electronic structure of CoAsSb with the use of self-consistent ab-initio many body techniques. In this work, we apply self consistent linearized quasi-particle GW method LQSGW\cite{prb_85_155129,cpc_219_407} to study electronic structure and thermoelectric properties of CoAsSb. We compare the LQSGW calculations with semiempirical hybrid functional PBE0\cite{jcp_105_9982} with different admix of Hartree-Fock non-local exchange (0.1, 0.15, and standard 0.25). In the second part of our study, we apply semiclassical Boltzman transport theory\cite{Boltzman} to evaluate thermoelectric properties of CoAsSb including temperature-dependence of the chemical potential. In this study calculations were performed using both experimental and optimized geometries from Ref. [\onlinecite{cm_30_4207}]. We have also checked the optimized geometry by performing independent optimization using VASP code and concluded that the structure obtained is essentially identical to the one obtained with Wien2k.
The plan of the paper is as follows. Section \ref{meth} discusses the methods for the calculations. Section \ref{res} provides the results obtained and the discussion. The conclusions are given thereafter.
\section{Methods}
\label{meth}
Electronic structure of CoAsSb was evaluated using the FlapwMBPT software package\cite{flapwmbpt} for the experimentally-measured geometry. FlapwMBPT implements diagrammatic approaches based on Full potential Linearized Augmented Plane Wave supplemented with Local Orbitals (FLAPW+LO). For this study, four approximations were selected. (1) the Generalized Gradient Approximation (GGA) as parameterized in PBE form.\cite{prl_77_3865} (2) The self-consistent linearized quasi-particle GW method, which was first introduced in [\onlinecite{prb_85_155129}], and additional details about the approach were given in [\onlinecite{cpc_219_407}]. (3) The modified LQSGW approach wherein W is evaluated only once based on time dependent density functional theory TD-DFT.\cite{prl_52_997} (4) Hybrid functional PBE0,\cite{jcp_105_9982} as well as variations of it with different percentage of the exact exchange.
To be specific, in the modified LQSGW approach (3) W is evaluated from the reducible polarizability $\chi$ ($W=V+V\chi V$, where $V$ is the bare Coulomb interaction). $\chi$ is evaluated as the exact (in DFT approximation) functional derivative of the electronic density $\rho$ with respect to an external electric field $\phi$: $\chi=\frac{\delta \rho}{\delta \phi}$. When $\rho$ is calculated in DFT approximation, the evaluation of the functional derivative corresponds to the solution of the following equation:
\begin{equation}\label{chi}
\chi=\chi_{0}+\chi_{0}[V+\frac{\delta V_{xc}}{\delta \rho}]\chi,
\end{equation}
where $\chi_{0}$ is the non-interacting polarizability and $V_{xc}$ is the exchange-correlation potential of DFT. The polarizability $\chi$ (and correspondingly W), evaluated this way, mimics the vertex corrections and usually provides a good approximation, especially for metallic systems. It was demonstrated, for example, in the calculations of dynamical response functions in Na and Al.\cite{prb_84_075109} Thus, we believe it should be more accurate for small bandgap semiconductors than the corresponding quantity obtained in LQSGW. We will abbreviate the third approach as LQSGW$^{PBE}$ reflecting the fact that W is evaluated following its definition with PBE parameterization of the GGA approximation. Direct application of vertex-corrected GW schemes\cite{prb_94_155101,prb_95_195120} would be too time consuming to apply for CoAsSb. However, the above slightly more advanced method LQSGW$^{PBE}$ (as compared to LQSGW) still mimics true vertex-corrected GW calculations and improves the bandgap.
CoAsSb has a large unit cell (12 atoms), and its symmetry group (P2$_{1}$/c, arsenopyrite-type structure \cite{cm_30_4207}) has 4 operations. The evaluation of the band energies on sufficiently fine \textbf{k}-mesh in the Brillouin zone is necessary followed by subsequent interpolation to even finer \textbf{k}-mesh used to study transport properties. In this work, all electronic structure calculations were performed using $8\times 8\times 8$ \textbf{k}-mesh. For a 12 atom unit cell with an $8\times 8\times 8$ \textbf{k}-mesh, computational requirements of self-consistent LQSGW calculations were quite formidable. This issue was resolved by i) using real-space plus Matsubara's time implementation \cite{prb_85_155129} of polarizability and self-energy evaluation, which allows considerable time savings as compared to the traditional reciprocal-space plus Matsubara's frequency formulation and ii) extensively usage of Message Passing Interface (MPI) to distribute the computational workload. In this respect, we refer interested reader to our earlier publication\cite{cpc_219_407} where the details of our parallelization strategy for the evaluation of all principal ingredients of GW algorithm are discussed. They are too numerous to elaborate the details here, but we would like to point out about one addition (with respect to what was presented in [\onlinecite{cpc_219_407}]) which specifically was implemented in the course of our present study. Namely, the calculation of W for a specific momentum and Matsubara frequency was performed by a process group to distribute peak memory usage. This has been achieved by creating an interface with the ScaLAPACK library.\cite{scalapack} Our FLAPW+LO basis set consisted of approximately 1000 functions and included semi-core states (3s,3p of Co; 3s,3p,3d of As; 4s,4p,4d of Sb) as well as one additional high energy LO for s,p,and d orbitals of all atoms. Green's function and self energy (fermionic functions) in the LQSGW part of the calculations were expanded using all band states generated from the above FLAPW+LO basis set. In this way, the well known issue of slow convergence of GW-based methods with respect to high energy states was properly addressed. The size of the Product Basis (PB) for the bosonic functions (polarizability and screened interaction) slightly exceeded 4000, with approximately equal number of functions from the muffin-tin (MT) spheres and from the interstitial region. The radii of the MT spheres were 2.22 a.u. (Co and As), and 2.55 a.u (Sb). Inside the muffin-tin spheres the spherical harmonic expansion was terminated at $L_{max}=4$ for both the LAPW basis and the product basis. In all calculations we started with 80 iterations within GGA to seed subsequent LQSGW/hybrid iterations. After that, 20 iterations of LQSGW/hybrid functionals were sufficient to converge the bandgap better than 0.005eV.
Following the calculation of electronic structure, the temperature- and direction-dependent Seebeck coefficient was evaluated using Boltzman theory\cite{Boltzman}. The band energies obtained on the $8\times 8\times 8$ \textbf irreducible {k}-mesh were interpolated\cite{pr_178_1419,jcompp_67_253,prb_38_2721} onto finer $32\times 32\times 32$ \textbf{k}-mesh in the Brillouin zone. Comparison to control calculations with a $64\times 64\times 64$ \textbf{k}-mesh confirms that the results were sufficiently converged on the $32\times 32\times 32$ \textbf{k}-mesh. In the initial experimental study, Sb vacancies were observed but the doping level is unknown. The Seebeck coefficient was maximized (in absolute value) at 450 K $S_{max}$. Goldsmid and Sharp have shown that the temperature where the Seebeck coefficient maximizes ($T_{max}$) is a function of doping level, but their product, the Goldsmid-Sharp gap $E_{GS}=2*S_{max}T_{max}$ is invariant to the doping level. Therefore, all transport quantities were calculated at the doping level which produced $T_{max}=450 K$. The equations of the Boltzman theory for transport properties were also implemented as a part of the FlapwMBPT code. That was done (instead of using Boltztrap code) because our study deals essentially with contributions to transport properties beyond the leading bandgap-dependent terms including temperature-dependent shifts in the chemical potential. Also, taking into account possible future developments of the code, it was preferable to us to incorporate the corresponding models into FlapwMBPT.
We have used a model to address the uncertainty related to the non-stoichiometry of the compound, which we discuss here. Instead of changing the geometry, we assumed that impact of the non-stoichiometry on the transport properties can be modeled by considering the system as doped. Experimental information\cite{cm_30_4207} suggests that there should be an excess of electrons (n-type semiconductor). As it was discussed by Goldsmid and Sharp in [\onlinecite{jem_28_869}], the maximum in Seebeck coefficient as a function of temperature is formed where the excitation of extrinsic carriers is superceded by excitations of electron-holes pairs across the energy gap. Thus, in order to model the low temperature part (extrinsic carriers), some source of doping needs to be introduced. But for the modeling the high temperature regime, the host band structure is needed. Doping affects $T_{max}$, so for each method, the doping level was adjusted to achieve $T_{max}=450K$ which provides a consistent comparison between the electronic structure methods.
\section{Results and discussion}
\label{res}
\begin{figure}[t]
\centering
\includegraphics[width=8.0 cm]{fig1.eps}
\caption{Band structure of CoAsSb obtained in GGA approximation. The valence band has a maximum at point Q which is near the path between Y and H. The conduction band has minima at Y and E. Both the valence band and conduction band cross the Fermi level. } \label{b_o_gga}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=8.0 cm]{fig2.eps}
\caption{Band structure of CoAsSb obtained in LQSGW$^{PBE}$ approximation. Fourier interpolation\cite{pr_178_1419,jcompp_67_253,prb_38_2721} has been used to obtain the band energies along the path in \textbf{k}-space. The valence band has a maximum at point Q which is near the path between Y and H. The conduction band has minima at Y and E.} \label{b_e_qp}
\end{figure}
\begin{table}[t]
\caption{Bandgap (eV, second column) and Goldsmid-Sharp gap, (eV, third column)\cite{jem_28_869}. For GGA, the band overlap is presented as a negative bandgap. Doping level at $T_{max} = 450 K$ (electrons per unit cell, fourth column). Maximum (most negative) Seebeck coefficient $S_{max}$ ($\mu V/K$, fifth column). Fractions of exact exchange for hybrid functionals are shown in brackets. Calculations using the scissor operator are marked with (s).} \label{e_dif}
\begin{center}
\begin{tabular}{@{}c c c c c}
\hline \hline
Method & Bandgap & GS gap& Doping& $S_{max}$\\
\hline
GGA &-0.078 &&0.03&-74\\
Hybrid (0.1) &0.102 &0.155&0.009&-172\\
Hybrid (0.15) &0.230 &0.234&0.003&-260\\
Hybrid (0.25) &0.444 &0.396&0.0003&-440\\
LQSGW &0.279 &0.324&0.0032&-360\\
LQSGW$^{PBE}$ &0.246 &0.288&0.006&-320\\
GGA(s) &0.12 &0.188&0.015&-209\\
Hybrid (0.25)(s) &0.12 &0.155&0.007&-172\\
LQSGW (s) &0.12 &0.198&0.016&-220\\
LQSGW$^{PBE}$ (s) &0.12 &0.212&0.018&-235\\
Exp. &0.12-0.256 &0.121&&-134\\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[b]
\centering
\includegraphics[width=8.0 cm]{fig3.eps}
\caption{Seebeck coefficient obtained with LQSGW$^{PBE}$ at different doping levels. 7 doping levels from 0.0001 to 0.01 electrons/unit cell along with zero doping are presented.} \label{seeb_74a}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=8.0 cm]{fig4.eps}
\caption{Temperature-dependent Seebeck coefficient obtained using each method. Doping was applied to achieve a maximum Seebeck coefficient at $T_max = 450 K$.} \label{seeb_all}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=8.0 cm]{fig5.eps}
\caption{Temperature-dependent Seebeck coefficient for each electronic structure method followed by scissor operator to achieve a bandgap of 0.12 eV. Doping was applied to achieve a maximum Seebeck coefficient at $T_max = 450 K$.} \label{seeb_scis}
\end{figure}
Bandgaps obtained from each electronic structure method are shown in Table \ref{e_dif}. For the experimental geometry, GGA results in a metal (Figure \ref{b_o_gga}) in agreement with work [\onlinecite{cm_30_4207}]. LQSGW results in a band gap of 0.279 eV which is slightly above the experimental range 0.12-0.256 eV.\cite{cm_30_4207} The range of exact exchange in PBE0 (10\%-25\%) results in bandgaps which cover the experimental uncertainty. For the optimized geometry, we obtained a tiny gap in our DFT calculation (0.005 eV), which can be attributed to using GGA instead of LDA as in [\onlinecite{cm_30_4207}]. But with LQSGW, a noticeable overestimation of the band gap (0.441eV) is observed. We conclude that the DFT-optimized geometry of the stoichiometric compound gave a worse estimation of the band gap than the experimental geometry, and thus we will adhere to the experimental geometry henceforth.
To elicit the differences between the electronic structures computed by GGA and $LQSGW^{PBE}$, their band structures are compared. The band structure from GGA is shown in Figure \ref{b_o_gga} whilst Figure \ref{b_e_qp} shows the band structure from LQSGW$^{PBE}$. Notable differences are observed at some high symmetry points which reshape band edges or shift them more at some wave vectors than others. As it follows from the Table \ref{e_dif}, LQSGW$^{PBE}$ results in a better band gap as compared to LQSGW. Thus, we conclude that vertex corrections are essential and further improvement in the calculated band structure can be achieved with advanced diagrammatic approaches.
We now turn to the consideration of thermoelectric properties. The Seebeck coefficient calculated with LQSGW$^{PBE}$ approach is presented in Figure \ref{seeb_74a} for various doping levels. Zero doping results in a positive Seebeck coefficient at low temperatures and a negative Seebeck coefficient at high temperatures in contrast to the consistent negative Seebeck coefficient from the experiment(Figure 7a in [\onlinecite{cm_30_4207}]). From a small increase in the doping, the LQSGW$^{PBE}$ curves suddenly become much more similar to the experiment and $T_{max}$ increases. The correct position of the maximum is achieved at the doping 0.006 electrons per unit cell which we compare to the experimental composition $CoAsSb_{0.883}$. The value of the Seebeck coefficient at the maximum ($S_{max}$) is -320 $\mu V/K$ which exceeds (in amplitude) the experimental value -134 $\mu V/K$\cite{cm_30_4207} reflecting the overestimation of the calculated band gap. For each method and the experiment, $S_{max}$ and the GS gap are collected in Table \ref{e_dif} where $T_{max} = 450 K$. Interestingly, the experimental GS gap deduced from the Seebeck coefficient (0.121 eV) coincides with bandgap deduced from conductivity from 325-475 K (0.12eV). The bandgaps and the GS gap are quite similar to one another, however the deviations among methods are intriguing. For LQSGW and $LQSGW^{PBE}$, the GS gap is greater than the bandgap. For the hybrid functionals, the GS gaps are greater than the lowest bandgap hybrid (bandgap: 0.102 eV, GS gap: 0.155 eV) but lower than the highest bandgap hybrid (bandgap: 0.444 eV, GS gap: 0.396 eV). The doping level necessary to obtain $T_{max} = 450 K$ mostly decreases as the bandgap increases. When the scissor operator was applied, the GS gap was always significantly above the bandgap.
Figure \ref{seeb_all} presents Seebeck coefficient calculated as a function of temperature for each method used in this study and it is compared with the experimental function (Fig. 7a in Ref.[\onlinecite{cm_30_4207}]). Experimental values of S are: 300 K( -100 $\mu V/K$), 450 K( -134 $\mu V/K$) and 1000K (-75 $\mu V/K$). The overall ladle shape is reproduced by each method except GGA. Each method shows a smaller magnitude at high temperatures and each is less than 100 $\mu V/K$ in magnitude, however each method behaves somewhat differently at low temperature. For the hybrid functionals, the larger the gap, the larger the difference between low and high temperature coefficients and the 3 cross each other. Interestingly, although the hybrid functional which corresponds to the best calculated band gap (0.1 of exact exchange) is somewhat more negative than the experiment at 300K and 450 K, at higher temperature it is more positive than the experiment. The more modest changes in temperature found for LQSGW and $LQSGW^{PBE}$ suggest that the dynamical correlation effects which are missing in the hybrid functionals bring important features in the spectra. The large deviation from experiment at low temperatures can be understood from the fact that extrinsic carriers are the dominant contributor here and correctly assigning extrinsic carriers has proven difficult.
For GGA, PBE(0.25), LQSGW and $LQSGW^{PBE}$, we also studied the effect of a rigid shift of the conduction bands (scissor operator) which results in the experimental band gap - an approach which is quite popular when using DFT. Figure \ref{seeb_scis} presents the results. The curves are mostly parallel to one another and have similar temperature dependence. Notably, GGA gained the curved shape of other methods and more closely matches the experimental temperature dependence, but not the magnitude of the Seebeck coefficient. We find that this justifies the application of scissor operator to the DFT band structure. We note that the PBE(0.25) curve retained much of its large swing with temperature. We think, that this fact reflects the above mentioned differences in the shape of the bands, which is a testimony of the importance of dynamic correlations in studies of this material and thermoelectrics in general.
\section*{Conclusions}
\label{concl}
In conclusion, we have extended the theoretical study of CoAsSb performed in Ref. [\onlinecite{cm_30_4207}] by applying advanced (as compared to DFT) electronic structure methods LQSGW and LQSGW$^{PBE}$ along with semiempirical hybrid functional (PBE0) with a few different percentages of the exact exchange. The bandgaps from LQSGW (0.279 eV) and $LQSGW^{PBE}$ (0.246) are a significant improvement from the gapless DFT calculations. If anything the bandgap is now too large, however this is complicated by apparent differences in the experimental gap at low temperature (0.12 eV) and high temperature (0.256 eV). Boltzman semiclassical transport theory was used to estimate the Seebeck coefficient from electronic structure. A full estimation of temperature-dependence was performed by using an extrinsic carrier density which reproduced the temperature of the Seebeck coefficient maximum. Additionally, the temperature-dependent Seebeck coefficient is significantly improved especially at high temperature. The effect of the scissor operator on the Seebeck coefficient was compared to advanced electronic structure methods by equating the bandgaps. This resulted in excellent agreement between GGA, LQSGW, and $LQSGW^{PBE}$, however all 3 resulted in a Goldsmid-Sharp gap much larger than the experiment. The remaining differences between the experimentally-measured and calculated Seebeck coefficient can primarily be ascribed to i) difficulty in assigning extrinsic carrier density given the non-stoichiometry found in the experiment and ii) absence of explicit diagrammatic vertex corrections in our calculations. Which one of the two source is more important is not clear at this point. Possible future extension of this work (by including diagrammatic vertex corrections) might allow to reduce the uncertainty in this respect.
\section*{Acknowledgments}
\label{ackn}
This work was supported by the U.S. Department of Energy, Office of Science, Basic
Energy Sciences as a part of the Computational Materials Science Program.
A.R. thanks support from a NASA Space Technology Research Fellowship. We thank Chang-Jong Kang for sharing data on CoAsSb with us.
|
\section{Introduction} \label{sec:intro}
The study of fairness in machine learning is driven by an abundance of examples where learning algorithms were perceived as discriminating against protected groups~\citep{Sween13,DTD15}. Addressing this problem requires a conceptual --- perhaps even philosophical --- understanding of what fairness means in this context. In other words, the million dollar question is (arguably\footnote{Recent work takes a somewhat different view~\citep{KRPH+17}.}) this: What are the formal constraints that fairness imposes on learning algorithms? On a very high level, most of the answers proposed so far~\citep{LRT11,DHPR+12,ZWSP+13,MPS16,JKMA16,ZVGG+17}
fall into two (partially overlapping) categories: individual fairness notions, and group fairness notions.
In the former category, the best known example is the influential fair classification model of \citet{DHPR+12}. The model involves a set of individuals and a set of outcomes. It is instructive to think of financially-motivated settings where the outcomes are, say, credit card offerings or displayed advertisements, and a loss function represents the benefit (e.g., in terms of revenue) of mapping a given individual to a given outcome. The centerpiece of the model is a \emph{similarity metric} on the space of individuals; it is specific to the classification task at hand, and ideally captures the ethical ground truth about relevant attributes. For example, a man and a woman who are similar in every other way should be considered similar for the purpose of credit card offerings, but perhaps not for lingerie advertisements. Assuming such a metric is available, fairness can be naturally formalized as a Lipschitz constraint, which requires that individuals who are close according to the similarity metric be mapped to distributions over outcomes that are close according to some standard metric (such as total variation). The algorithmic problem is then to find a classifier that minimizes loss, subject to the Lipschitz constraint.
As attractive as this model is, it has one clear weakness from a practical viewpoint: the availability of a similarity metric. \citet{DHPR+12} are well aware of this issue; they write that justifying this assumption is ``one of the most challenging aspects'' of their approach. They add that ``in reality the metric used will most likely only be society's current best approximation to the truth.'' But, despite recent progress on automating ethical decisions in certain domains~\citep{NGAD+18,FSSD+18}, the task-specific nature of the similarity metric makes even a credible approximation thereof seem unrealistic. In particular, if one wanted to learn a similarity metric, it is unclear what type of examples a relevant dataset would consist of.
An alternative notion of individual fairness, therefore, is called for. And our proposal draws on an extensive body of work on rigorous approaches to fairness, which --- modulo one possible exception (see Section~\ref{subsec:rel}) --- has not been tapped by machine learning researchers: the literature on \emph{fair division}~\citep{BT96,Moul03}. The most prominent notion is that of \emph{envy-freeness}~\citep{Fol67,Var74}, which, in the context of the allocation of goods, requires that the utility of each individual for his allocation be at least as high as his utility for the allocation of any other individual; this is the gold standard of fairness for problems such as cake cutting \citep{RW98,Pro13} and rent division \citep{Su99,GMPZ17}.
Similarly, in the classification setting, envy-freeness would simply mean that the utility of each individual for his distribution over outcomes is at least as high as his utility for the distribution over outcomes assigned to any other individual. For example, it may well be the case that Bob is offered a worse credit card than that offered to Alice (in terms of, say, annual fees), but this outcome is not unfair if Bob is genuinely more interested in the card offered to him because he does not qualify for Alice's card, \edit{or because its specific rewards program better fits his needs. Such rich utility functions are also evident in the context of job advertisements~\citep{DTD15}: people generally want higher paying jobs, but would presumably have higher utility for seeing advertisements for jobs that better fit their qualifications and interests.}
Of course, as before, envy-freeness requires access to individuals' utility functions, but --- in stark contrast to the similarity metric of \citet{DHPR+12} --- \edit{we do not view this assumption as a barrier to implementation.} Indeed, there are a variety of techniques for learning utility functions~\citep{CKO01,NJ04,BCIW12}. \edit{Moreover, in our running example of advertising, one can even think of standard measures like expected click-through rate (CTR) as an excellent proxy for utility.}
It is worth noting that the classification setting is different from classic fair division problems in that the ``goods'' (outcomes) are non-excludable. In fact, one envy-free solution simply assigns each individual to his favorite outcome; but when the loss function disagrees with the utility functions, it may be possible to achieve smaller loss without violating the envy-freeness constraint.
In summary, we view envy-freeness as a compelling, well-established, and, importantly, practicable notion of individual fairness for classification tasks. Our goal is to understand its learning-theoretic properties.
\subsection{Our Results}
The technical challenge we face is that the space of individuals is potentially huge, yet we seek to provide universal envy-freeness guarantees. To this end, we are given a sample consisting of individuals drawn from an unknown distribution. We are interested in learning algorithms that minimize loss, subject to satisfying the envy-freeness constraint, \emph{on the sample}. Our primary technical question is that of generalizability, that is, \emph{given a classifier that is envy free on a sample, is it approximately envy free on the underlying distribution?} \edit{Surprisingly, \citet{DHPR+12} do not study generalizability in their model, and we are aware of only one subsequent paper that takes a learning-theoretic viewpoint on individual fairness and gives theoretical guarantees (see Section~\ref{subsec:rel}).}
In Section~\ref{sec:arb}, we do not constrain the classifier in question. Therefore, we need some strategy to extend a classifier that is defined on a sample; assigning an individual the same outcome as his \emph{nearest neighbor} in the sample is a popular choice. However, we show that \emph{any} strategy for extending a classifier from a sample, on which it is envy free, to the entire set of individuals is unlikely to be approximately envy free on the underlying distribution, unless the sample is exponentially large.
For this reason, in Section~\ref{sec:mixtures}, we focus on structured families of classifiers. On a high level, our goal is to relate the combinatorial richness of the family to generalization guarantees. One obstacle is that standard notions of dimension do not extend to the analysis of randomized classifiers, whose range is \emph{distributions} over outcomes (equivalently, real vectors). We circumvent this obstacle by considering mixtures of \emph{deterministic} classifiers that belong to a family of bounded Natarajan dimension (an extension of the well-known VC dimension to multi-class classification). Our main technical result asserts that, under this assumption, envy-freeness on a sample does generalize to the underlying distribution, even if the sample is relatively small (its size grows almost linearly in the Natarajan dimension). Finally, we discuss the implications of this result in Section~\ref{sec:disc}.
\subsection{Related Work}
\label{subsec:rel}
Conceptually, our work is most closely related to work by \citet{ZVGG+17}. They are interested in group notions of fairness, and advocate preference-based notions instead of parity-based notions. In particular, they assume that each group has a utility function for \emph{classifiers}, and define the \emph{preferred treatment} property, which requires that the utility of each group for its own classifier be at least its utility for the classifier assigned to any other group. Their model and results focus on the case of binary classification where there is a desirable outcome and an undesirable outcome, so the utility of a group for a classifier is simply the fraction of its members that are mapped to the desirable outcome. Although, at first glance, this notion seems similar to envy-freeness, it is actually fundamentally different.\footnote{On a philosophical level, the fair division literature deals exclusively with individual notions of fairness. In fact, even in group-based extensions of envy-freeness~\citep{MS17} the allocation is shared by groups, but individuals must not be envious. We subscribe to the view that group-oriented notions (such as statistical parity) are objectionable, because the outcome can be patently unfair to individuals.}
Our paper is also completely different from that of \citeauthor{ZVGG+17}~in terms of technical results; theirs are purely empirical in nature, and focus on the increase in accuracy obtained when parity-based notions of fairness are replaced with preference-based ones.
\edit{Very recent, concurrent work by \citet{RY18} provides generalization guarantees for the metric notion of individual fairness introduced by \citet{DHPR+12}, or, more precisely, for an approximate version thereof.
There are two main differences compared to our work: first, we propose envy-freeness as an alternative notion of fairness that circumvents the need for a similarity metric. Second, they focus on randomized \emph{binary} classification, which amounts to learning a real-valued function, and so are able to make use of standard Rademacher complexity results to show generalization. By contrast, standard tools do not directly apply in our setting. It is worth noting that several other papers provide generalization guarantees for notions of group fairness, but these are more distantly related to our work \citep{ZWSP+13,WGOS17,DOBS+18,KNRW18,HKRR18}.}
\section{The Model} \label{sec:problem}
We assume that there is
a space $\cX$ of individuals, a finite space $\cY$ of outcomes, and a utility function $u : \cX \times \cY \to [0,1]$ encoding the preferences
of each individual for the outcomes in $\cY$. In the advertising example, individuals are users, outcomes are advertisements, and the utility function reflects the benefit an individual derives from being shown a particular advertisement. For any distribution $p \in \Delta(\cY)$ (where $\Delta(\cY)$ is the set of distributions over $\cY$) we let
$u(x,p) = \expect_{y \sim p}[u(x,y)]$ denote individual $x$'s expected utility
for an outcome sampled from $p$. We refer to a function $h :
\cX \to \Delta(\cY)$ as a \emph{classifier}, even though it can return a distribution over outcomes.
\subsection{Envy-Freeness}
Roughly speaking, a classifier $h :
\cX \to \Delta(\cY)$ is envy free if no individual prefers the
outcome distribution of someone else over his own.
\begin{defn}
%
A classifier $h : \cX \to \Delta(\cY)$ is \emph{envy free (EF)} on a set $S$ of
individuals if $u(x,h(x)) \geq u(x,h(x'))$ for all $x,x' \in S$. Similarly,
$h$ is \emph{$(\alpha,\beta)$-EF} with respect to a distribution $P$ on $\cX$ if
\[\prob_{x,x'\sim P}\bigl( u(x,h(x)) < u(x,h(x')) - \beta \bigr) \leq \alpha.\]
Finally, $h$ is \emph{$(\alpha,\beta)$-pairwise EF} on a set of pairs of
individuals $S = \{(x_i,x'_i)\}_{i=1}^n$ if \[\frac{1}{n} \sum_{i=1}^n
\ind{u(x_i, h(x_i)) < u(x_i, h(x'_i)) - \beta} \leq \alpha.\]
%
\end{defn}
Any classifier that is EF on a sample $S$ of individuals is also
$(\alpha,\beta)$-pairwise EF on any pairing of the individuals in $S$, for any $\alpha \geq 0$ and $\beta \geq 0$. The
weaker pairwise EF condition is all that is required for our
generalization guarantees to hold.
\subsection{Optimization and Learning}
Our formal learning problem can be stated as follows. Given sample access to an
unknown distribution $P$ over individuals $\cX$ and their utility functions, and a known loss function $\ell : \cX \times \cY \to
[0,1]$, find a classifier $h : \cX \to \Delta(\cY)$ that is
$(\alpha,\beta)$-EF with respect to $P$ minimizing expected loss $\expect_{x \sim
P}[\ell(x, h(x))]$, where for $x \in \cX$ and $p \in \Delta(\cY)$, $\ell(x,p) =
\expect_{y \sim p}[\ell(x,y)]$.
We follow the empirical risk minimization (ERM)
learning approach, i.e., we collect a sample of individuals drawn i.i.d~from
$P$ and find an EF classifier with low loss on the sample. Formally, given a sample of individuals $S=\{x_1,\ldots,x_n\}$ and their utility functions $u_{x_i}(\cdot)=u(x_i,\cdot)$, we are interested in a classifier $h:S\to\Delta(\cY)$ that minimizes $\sum_{i=1}^n \ell(x_i,h(x_i))$ among all classifiers that are EF on $S$. The algorithmic problem itself is beyond the scope of the current paper; see Section~\ref{sec:disc} for further discussion.
Recall that we consider randomized classifiers that can assign a distribution over outcomes to each of the individuals. However, one might wonder whether the EF classifier that minimizes loss on a sample happens to always be deterministic. Or, at least, the optimal deterministic classifier on the sample might incur a loss that is very close to that of the optimal randomized classifier. If this were true, we could restrict ourselves to classifiers of the form $h: \cX \to \cY$, which would be much easier to analyze. Unfortunately, it turns out that this is not the case. In fact, there could be an arbitrary (multiplicative) gap between the optimal randomized EF classifier and the optimal deterministic EF classifier. The intuition behind this is as follows. A deterministic classifier that has very low loss on the sample, but is not EF, would be completely discarded in the deterministic setting. On the other hand, a randomized classifier could take this loss-minimizing deterministic classifier and mix it with a classifier with high ``negative envy'', so that the mixture ends up being EF and at the same time has low loss. This is made concrete in Example~\ref{ex:det-random-gap} in the appendix.
\section{Arbitrary Classifiers}\label{sec:arb}
An important (and typical) aspect of our learning problem is that the classifier
$h$ needs to provide an outcome distribution for every individual, not just
those in the sample. For example, if $h$ chooses advertisements for visitors of
a website, the classifier should still apply when a new visitor arrives.
Moreover, when we use the classifier for new individuals, it must continue to be
EF. In this section, we consider two-stage approaches that first choose outcome
distributions for the individuals in the sample, and then extend those decisions
to the rest of $\cX$.
In more detail, we are given a sample $S = \{x_1, \dots, x_n\}$ of individuals
and a classifier $h : S \to \Delta(\cY)$ assigning outcome distributions to each
individual. Our goal is to extend these assignments to a classifier
$\overline{h} : \cX \to \Delta(\cY)$ that can be applied to new individuals as
well. For example, $h$ could be the loss-minimizing EF classifier on the sample
$S$.
For this section, we
assume that $\cX$ is equipped with a distance metric $d$. Moreover, we assume in this section that the utility function $u$ is $L$-Lipschitz on $\cX$. That is, for
every $y \in \cY$ and for all $x, x' \in \cX$, we have $|u(x, y) - u(x', y)|
\leq L \cdot d(x,x')$.
Under the foregoing assumptions, one natural way to extend the classifier on the sample
to all of $\cX$ is to assign new individuals the same outcome distribution
as their nearest neighbor in the sample. Formally, for a set $S \subset \cX$ and any individual
$x \in \cX$, let $\NN_S(x)\in \text{arg\,min}_{x'\in S}d(x,x')$ denote the nearest neighbor of $x$ in $S$ with
respect to the metric $d$ (breaking ties arbitrarily). The following simple result (whose proof is relegated to Appendix~\ref{app:arb}) establishes that this approach preserves envy-freeness in cases where
the sample is exponentially large.
\begin{thm} \label{thm:nn-upper-bound}
%
Let $d$ be a metric on $\cX$, $P$ be a distribution on $\cX$, and $u$ be
an $L$-Lipschitz utility function. Let $S$ be a set of individuals such that
there exists $\hat \cX \subset \cX$ with $P(\hat \cX) \geq 1-\alpha$ and
$\sup_{x \in \hat \cX}(d(x, \NN_S(x)) \leq \beta/(2L)$. Then for any
classifier $h : S \to \Delta(\cY)$ that is EF on $S$, the extension
$\overline{h} : \cX \to \Delta(\cY)$ given by $\overline{h}(x) = h(\NN_S(x))$
is $(\alpha, \beta)$-EF on $P$.
%
\end{thm}
The conditions of Theorem~\ref{thm:nn-upper-bound} require that the set of
individuals $S$ is a $\beta/(2L)$-net for at least a $(1-\alpha)$-fraction of
the mass of $P$ on $\cX$. In several natural situations, an exponentially large
sample guarantees that this occurs with high probability. For example, if $\cX$
is a subset of $\reals^q$, $d(x,x') = \norm{x-x'}_2$, and $\cX$ has diameter at
most $D$, then for any distribution $P$ on $\cX$, if $S$ is an i.i.d.~sample of size
$O(\frac{1}{\alpha}(\frac{LD\sqrt{q}}{\beta})^q(q \log \frac{LD\sqrt{q}}{\beta} +
\log \frac{1}{\delta}))$, it will satisfy the conditions of
Theorem~\ref{thm:nn-upper-bound} with probability at least $1-\delta$. This
sampling result is folklore, but, for the sake of completeness, we prove it in Lemma~\ref{lem:euclideanCover}
of Appendix~\ref{app:arb}.
However, the exponential upper bound given by the nearest neighbor strategy is as far as we can go in terms of generalizing envy-freeness from a sample (without further assumptions). Specifically, our next result establishes that \emph{any} algorithm --- even randomized --- for extending classifiers from the sample to the entire space $\cX$
requires an exponentially large sample of individuals to ensure envy-freeness on
the distribution $P$. The proof of Theorem~\ref{thm:lb} can be found in Appendix~\ref{app:arb}.
\begin{thm}
\label{thm:lb}
There exists a space of individuals $\cX \subset \reals^q$, and a distribution $P$ over $\cX$ such that, for every randomized algorithm $\mathcal{A}$ that extends classifiers on a sample to $\cX$, there exists an $L$-Lipschitz utility function $u$ such that, when a sample of individuals $S$ of size $n = 4^q / 2$ is drawn from $P$ without replacement, there exists an EF classifier on $S$ for which, with probability at least $1 - 2\exp(-4^q/100) - \exp(-4^q/200)$ jointly over the randomness of $\mathcal{A}$ and $S$, its extension by $\mathcal{A}$ is not $(\alpha, \beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$.
\end{thm}
We remark that a similar result would hold even if we sampled $S$ with replacement; we sample here without replacement purely for ease of exposition.
\begin{proof}[Proof of Theorem~\ref{thm:lb}]
Let the space of individuals be $\cX = [0,1]^q$ and the outcomes be $\cY = \{0,1\}$. We partition the space $\cX$ into cubes of side length $s = 1/4$. So, the total number of cubes is $m = \left(1/s\right)^q = 4^q$. Let these cubes be denoted by $c_1, c_2, \dots c_m$, and let their centers be denoted by $\mu_1, \mu_2, \dots \mu_m$. Next, let $P$ be the uniform distribution over the centers $\mu_1, \mu_2, \dots \mu_m$. For brevity, whenever we say ``utility function'' in the rest of the proof, we mean ``$L$-Lipschitz utility function.''
To prove the theorem, we use Yao's minimax principle~\citep{Yao77}. Specifically, consider the following two-player zero sum game. Player 1 chooses a deterministic algorithm $\mathcal{D}$ that extends classifiers on a sample to $\cX$, and player 2 chooses a utility function $u$ on $\cX$. For any subset $S \subset \cX$, define the classifier $h_{u,S}:S \to \cY$ by assigning each individual in $S$ to his favorite outcome with respect to the utility function $u$, i.e. $h_{u,S}(x) = \text{arg\,max}_{y \in \cY} u(x,y)$ for each $x \in S$, breaking ties lexicographically. Define the cost of playing algorithm $\mathcal{D}$ against utility function $u$ as the probability over the sample $S$ (of size $m/2$ drawn from $P$ without replacement) that the extension of $h_{u,S}$ by $\mathcal{D}$ is not $(\alpha, \beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$. Yao's minimax principle implies that for any randomized algorithm~$\mathcal{A}$, its expected cost with respect to the worst-case utility function~$u$ is at least as high as the expected cost of any distribution over utility functions that is played against the best deterministic algorithm $\mathcal{D}$ (which is tailored for that distribution). Therefore, we establish the desired lower bound by choosing a specific distribution over utility functions, and showing that the best deterministic algorithm against it has an expected cost of at least $1 - 2\exp(-m/100) - \exp(-m/200)$.
To define this distribution over utility functions, we first sample outcomes $y_1, y_2, \dots, y_m$ i.i.d. from Bernoulli($1/2$). Then, we associate each cube center $\mu_i$ with the outcome $y_i$, and refer to this outcome as the \textit{favorite} of $\mu_i$. For brevity, let $\neg y$ denote the outcome other than $y$, i.e. $\neg y = (1-y)$. For any $x \in \cX$, we define the utility function as follows. Letting $c_j$ be the cube that $x$ belongs to,
\begin{equation}
\label{eq:util}
u(x,y_j) = L \left[\frac{s}{2} - \|x - \mu_j\|_{\infty}\right]; \quad u(x, \neg y_j) = 0.
\end{equation}
See Figure~\ref{fig:grid} for an illustration.
\begin{figure}[t]
\centering
\includegraphics[width=0.25\columnwidth]{grid.jpg}
\caption{Illustration of $\cX$ and an example utility function $u$ for $d=2$. Red shows preference for $1$, blue shows preference for $0$, and darker shades correspond to more intense preference. (The gradients are rectangular to match the $L_\infty$ norm, so, strangely enough, the misleading X pattern is an optical illusion.)}
\label{fig:grid}
\end{figure}
We claim that the utility function of Equation~\eqref{eq:util} is indeed $L$-Lipschitz with respect to any $L_p$ norm. This is because for any cube $c_i$, and for any $x, x' \in c_i$, we have
\begin{align*}
\left|u(x,y_i) - u(x',y_i)\right| &= L\left|\|x - \mu_i\|_{\infty} - \|x' - \mu_i\|_{\infty}\right|\\
&\leq L\|x - x'\|_{\infty}
\leq L\|x - x'\|_{p}.
\end{align*}
Moreover, for the other outcome, we have $u(x, \neg y_i) = u(x', \neg y_i) = 0$. It follows that $u$ is $L$-Lipschitz within every cube. At the boundary of the cubes, the utility for any outcome is $0$, and hence $u$ is also continuous throughout $\cX$. Because it is piecewise Lipschitz and continuous, $u$ must be $L$-Lipschitz throughout $\cX$, with respect to any $L_p$ norm.
Next, let $\mathcal{D}$ be an arbitrary deterministic algorithm that extends classifiers on a sample to $\cX$. We draw the sample $S$ of size $m/2$ from $P$ without replacement. Consider the distribution over favorites of individuals in $S$. Each individual in $S$ has a favorite that is sampled independently from Bernoulli$(1/2)$. Hence, by Hoeffding's inequality, the fraction of individuals in $S$ with a favorite of $0$ is between $\frac{1}{2} - \epsilon$ and $\frac{1}{2} + \epsilon$ with probability at least $1 - 2\exp(-m \epsilon^2)$. The same holds simultaneously for the fraction of individuals with favorite $1$.
Given the sample $S$ and the utility function $u$ on the sample (defined by the instantiation of their favorites), consider the classifier $h_{u,S}$, which maps each individual $\mu_i$ in the sample $S$ to his favorite $y_i$. This classifier is clearly EF on the sample. Consider the extension $h_{u,S}^\mathcal{D}$ of $h_{u,S}$ to the whole of $\cX$ as defined by algorithm $\mathcal{D}$. Define two sets $Z_0$ and $Z_1$ by letting $Z_y = \{\mu_j \notin S \ | \ h_{u,S}^\mathcal{D}(\mu_j) = y\}$, and let $y_*$ denote an outcome that is assigned to at least half of the out-of-sample centers, i.e., an outcome for which $|Z_{y_*}| \geq |Z_{\neg y_*}|$. Furthermore, let $\theta$ denote the fraction of out-of-sample centers assigned to $y_*$. Note that, since $|S| = m/2$, the number of out-of-sample centers is also exactly $m/2$. This gives us $|Z_{y_*}| = \theta \frac{m}{2}$, where $\theta \geq \frac{1}{2}$.
Consider the distribution of favorites in $Z_{y_*}$ (these are independent from the ones in the sample since $Z_{y_*}$ is disjoint from $S$). Each individual in this set has a favorite sampled independently from Bernoulli$(1/2)$. Hence, by Hoeffding's inequality, the fraction of individuals in $Z_{y_*}$ whose favorite is $\neg y_*$ is at least $\frac{1}{2} - \epsilon$ with probability at least $1 - \exp(-\frac{m}{2} \epsilon^2)$. We conclude that with a probability at least
$1 - 2\exp(-m \epsilon^2) - \exp(-\frac{m}{2} \epsilon^2)$,
the sample $S$ and favorites (which define the utility function $u$) are such that: (i)~the fraction of individuals in $S$ whose favorite is $y \in \{0,1\}$ is between $\frac{1}{2} - \epsilon$ and $\frac{1}{2} + \epsilon$, and (ii)~the fraction of individuals in $Z_{y_*}$ whose favorite is $\neg y_*$ is at least $\frac{1}{2} - \epsilon$.
We now show that for such a sample $S$ and utility function~$u$, $h_{u,S}^\mathcal{D}$ cannot be $(\alpha, \beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$. To this end, sample $x$ and $x'$ from $P$. One scenario where $x$ envies $x'$ occurs when (i)~the favorite of $x$ is $\neg y_*$, (ii)~$x$ is assigned to $y_*$, and (iii)~$x'$ is assigned to $\neg y_*$. Conditions (i) and (ii) are satisfied when $x$ is in $Z_{y_*}$ and his favorite is $\neg y_*$. We know that at least a $\frac{1}{2} - \epsilon$ fraction of the individuals in $Z_{y_*}$ have the favorite $\neg y_*$. Hence, the probability that conditions (i) and (ii) are satisfied by $x$ is at least $(\frac{1}{2} - \epsilon)|Z_{y_*}|\frac{1}{m} = (\frac{1}{2} - \epsilon)\frac{\theta}{2}$.
Condition (iii) is satisfied when $x'$ is in $S$ and has favorite $\neg y_*$ (and hence assigned $\neg y_*$), or, if $x'$ is in $Z_{\neg y_*}$. We know that at least a $\left(\frac{1}{2} - \epsilon\right)$ fraction of the individuals in $S$ have the favorite $\neg y_*$. Moreover, the size of $Z_{\neg y_*}$ is $(1-\theta)\frac{m}{2}$. So, the probability that condition (iii) is satisfied by $x'$ is at least
\[\frac{\left(\frac{1}{2} - \epsilon\right)|S| + |Z_{\neg y_*}|}{m} = \frac{1}{2}\left(\frac{1}{2} - \epsilon\right) + \frac{1}{2}(1-\theta).\]
Since $x$ and $x'$ are sampled independently, the probability that all three conditions are satisfied is at least \[\left(\frac{1}{2} - \epsilon\right)\frac{\theta}{2} \cdot \left[\frac{1}{2}\left(\frac{1}{2} - \epsilon\right) + \frac{1}{2}(1-\theta)\right].\] This expression is a quadratic function in $\theta$, that attains its minimum at $\theta = 1$ irrespective of the value of $\epsilon$. Hence, irrespective of $\mathcal{D}$, this probability is at least $\left[\frac{1}{2}\left(\frac{1}{2} - \epsilon\right)\right]^2$. For concreteness, let us choose $\epsilon$ to be $1/10$ (although it can be set to be much smaller). On doing so, we have that the three conditions are satisfied with probability at least $1/25$. And when these conditions are satisfied, we have $u(x, h_{u,S}^\mathcal{D}(x)) = 0$ and $u(x, h_{u,S}^\mathcal{D}(x')) = Ls/2$, i.e., $x$ envies $x'$ by $Ls/2 = L/8$. This shows that, when $x$ and $x'$ are sampled from $P$, with probability at least $1/25$, $x$ envies $x'$ by $L/8$. We conclude that with probability at least $1 - 2\exp(-m/100) - \exp(-m/200)$ jointly over the selection of the utility function $u$ and the sample $S$, the extension of $h_{u,S}$ by $\mathcal{D}$ is not $(\alpha,\beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$.
To convert the joint probability into expected cost in the game, note that for two discrete, independent random variables $X$ and $Y$, and for a Boolean function $\mathcal{E}(X,Y)$, it holds that
\begin{equation}
\label{eq:exp}
\operatorname{Pr}_{X,Y}(\mathcal{E}(X,Y) = 1) = \mathbb{E}_X\left[\operatorname{Pr}_Y(\mathcal{E}(X,Y) = 1)\right].
\end{equation}
Given sample $S$ and utility function $u$, let $\mathcal{E}(u, S)$ be the Boolean function that equals $1$ if and only if the extension of $h_{u,S}$ by $\mathcal{D}$ is not $(\alpha,\beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$. From Equation~\eqref{eq:exp},
$\operatorname{Pr}_{u,S}(\mathcal{E}(u,S) = 1)$ is equal to $\mathbb{E}_u\left[\operatorname{Pr}_S(\mathcal{E}(u,S) = 1)\right]$. The latter term is exactly the expected value of the cost, where the expectation is taken over the randomness of $u$. It follows that the expected cost of (any) $\mathcal{D}$ with respect to the chosen distribution over utilities is at least $1 - 2\exp(-m/100) - \exp(-m/200)$.
%
\end{proof}
\section{Low-Complexity Families of Classifiers} \label{sec:mixtures}
In this section we show that (despite Theorem~\ref{thm:lb}) generalization for envy-freeness is possible using
much smaller samples of individuals, as long as we restrict ourselves to
choosing a classifier from a family of relatively low complexity.
In more detail, two
classic complexity measures are the
VC-dimension~\citep{VC} for binary classifiers, and the Natarajan dimension~\citep{Natarajan89:Dimension} for multi-class
classifiers. However, to the best of our knowledge, there is no suitable dimension
directly applicable to functions ranging over distributions, which in our case
can be seen as $|\cY|$-dimensional real vectors. One possibility would be to restrict
ourselves to deterministic classifiers of the type $h : \cX \to \cY$. However, we have seen in Section~\ref{sec:problem} that envy-freeness is a very strong constraint on deterministic classifiers. Instead,
we will consider a family $\cH$ consisting of randomized mixtures of
deterministic classifiers belonging to a family $\cG \subset \{g : \cX \to \cY\}$ of
low Natarajan dimension. This allows us to adapt Natarajan-dimension-based
generalization results to our setting while still working with randomized
classifiers.
\subsection{Natarajan Dimension Primer}
\label{subsec:nat}
Before presenting our main result, we briefly summarize the
definition and relevant properties of the Natarajan dimension. For more details, we refer the reader to
\citep{SBD14:MLBook}.
We say that a family $\cG$ \emph{multi-class shatters} a set of
points $x_1, \dots, x_n$ if there exist labels $y_1, \dots y_n$ and $y'_1,
\dots, y'_n$ such that for every $i \in [n]$ we have $y_i \neq y'_i$, and for any
subset $C \subset[n]$ there exists $g \in \cG$ such that $g(x_i) = y_i$ if $i
\in C$ and $g(x_i)=y'_i$ otherwise. The Natarajan dimension of a family $\cG$ is the
cardinality of the largest set of points that can be multi-class shattered by
$\cG$.
For example, suppose we have a feature map $\Psi : \cX \times \cY \to \reals^q$
that maps each individual-outcome pair to a $q$-dimensional feature vector, and
consider the family of functions that can be written as $g(x) =
\text{arg\,max}_{y \in \cY} w^\tp \Psi(x,y)$ for weight vectors $w \in
\reals^q$. This family has Natarajan dimension at most $q$.
For a set $S \subset \cX$ of points, we let $\cG\resto_S$ denote the restriction
of $\cG$ to $S$, which is any subset of $\cG$ of minimal size such that for
every $g \in \cG$ there exists $g' \in \cG\resto_S$ such that $g(x) = g'(x)$
for all $x \in S$. The size of $\cG\resto_S$ is the number of different
labelings of the sample $S$ achievable by functions in $\cG$. The following
Lemma is the analogue of Sauer's lemma for binary classification.
\begin{lem}[Natarajan] \label{lem:natarajan}
%
For a family $\cG$ of Natarajan dimension $d$ and any subset $S \subset \cX$,
we have $\bigl|\cG\resto_S\bigr| \leq |S|^d |\cY|^{2d}$.
%
\end{lem}
Classes of low Natarajan dimension also enjoy the following uniform convergence
guarantee.
\begin{lem} \label{lem:natarajanAgnostic}
Let $\cG$ have Natarajan dimension $d$ and fix a loss function $\ell : \cG
\times \cX \to [0,1]$. For any distribution $P$ over $\cX$, if $S$ is an i.i.d.~sample drawn from $P$ of size $O(\frac{1}{\epsilon^2}(d \log |\cY| + \log
\frac{1}{\delta}))$, then with probability at least $1-\delta$ we have
$
\sup_{g \in \cG} \, \left|
\expect_{x \sim P}[\ell(g,x)] - \frac{1}{n}\sum_{x\in S} \ell(g,x)
\right| \leq \epsilon.
$
\end{lem}
\subsection{Main Result}
We consider the family of classifiers that can be expressed as a randomized
mixture of $m$ deterministic classifiers selected from a family $\cG \subset \{g : \cX
\to \cY\}$. Our generalization guarantees will depend on the complexity of the
family $\cG$, measured in terms of its Natarajan dimension, and the number $m$ of
functions we are mixing. More formally, let $\vec{g} = (g_1, \dots, g_m) \in
\cG^m$ be a vector of $m$ functions in $\cG$ and $\alpha \in \Delta_m$ be a
distribution over $[m]$, where $\Delta_m = \{ p \in \reals^m \,:\, p_i \geq 0,
\sum_i p_i = 1\}$ is the $m$-dimensional probability simplex. Then consider the
function $h_{\vec{g},\alpha} : \cX \to \Delta(\cY)$ with assignment
probabilities given by
$
\prob(h_{\vec{g}, \alpha}(x) = y)
=
\sum_{i=1}^m \ind{g_i(x) = y} \alpha_i.
$
Intuitively, for a given individual $x$, $h_{\vec{g}, \alpha}$ chooses one of
the $g_i$ randomly with probability $\alpha_i$, and outputs $g_i(x)$. Let
\[\cH(\cG, m) = \{ h_{\vec{g},\alpha} : \cX \to \Delta(\cY) \,:\, \vec{g} \in
\cG^m, \alpha \in \Delta_m\}\] be the family of classifiers that can be written
this way. Our main technical result shows that envy-freeness generalizes for
this class.
\begin{thm} \label{thm:mixtureGeneralize}
%
Suppose $\cG$ is a family of deterministic classifiers of Natarajan
dimension $d$, and let $\cH = \cH(\cG,m)$ for $m\in \mathbb{N}$. For any distribution
$P$ over $\cX$, $\gamma>0$, and $\delta > 0$, if $S = \{(x_i, x'_i)\}_{i=1}^n$
is an i.i.d.~sample of pairs drawn from $P$ of size \[n \geq
O\left(\frac{1}{\gamma^2}\left(dm^2 \log \frac{dm|\cY|\log(m|\cY|/\gamma)}{\gamma} + \log
\frac{1}{\gamma}\right)\right),\] then with probability at least $1-\delta$, every classifier $h
\in \cH$ that is $(\alpha,\beta)$-pairwise-EF on $S$ is also
$(\alpha+7\gamma, \beta+4\gamma)$-EF on $P$.
%
\end{thm}
\edit{Theorem~\ref{thm:mixtureGeneralize} is only effective insofar as families of classifiers of low Natarajan dimension are useful. And, indeed, several prominent families have low Natarajan dimension~\citep{DSS12}, including one vs. all (which is a special case of the example given in Section~\ref{subsec:nat}), multiclass SVM, tree-based classifiers, and error correcting output codes.
We now turn to the theorem's proof, which consists of two steps.} First,
we show that envy-freeness generalizes for finite classes. Second, we show that
$\cH(\cG,m)$ can be approximated by a finite subset.
\begin{lem} \label{lem:finiteGeneralize}
%
Let $\cH \subset \{h : \cX \to \Delta(\cY)\}$ be a finite family of
classifiers. For any $\gamma > 0$, $\delta > 0$, and $\beta \geq 0$ if $S =
\{(x_i, x'_i)\}_{i=1}^n$ is an i.i.d.~sample of pairs from $P$ of size $n \geq
\frac{1}{2\gamma^2}\ln \frac{|\cH|}{\delta}$, then with probability at least
$1-\delta$, every $h \in \cH$ that is $(\alpha,\beta)$-pairwise-EF on $S$
(for any $\alpha)$ is also $(\alpha + \gamma, \beta)$-EF on $P$.
%
\end{lem}
\begin{proof}
%
Let $f(x,x',h) = \ind{u(x,h(x)) < u(x,h(x')) - \beta}$ be the indicator that
$x$ is envious of $x'$ by at least $\beta$ under classifier $h$. Then $f(x_i,
x'_i, h)$ is a Bernoulli random variable with success probability
$\expect_{x,x'\sim P}[f(x,x',h)]$. Applying Hoeffding's inequality to any
fixed hypothesis $h \in \cH$ guarantees that $\prob_S(\expect_{x,x'\sim
P}[f(x,x',h)] \geq \frac{1}{n}\sum_{i=1}^n f(x_i, x'_i, h) + \gamma) \leq
\exp(-2n\gamma^2)$. Therefore, if $h$ is $(\alpha,\beta)$-EF on $S$,
then it is also $(\alpha+\gamma, \beta)$-EF on $P$ with probability at least $1 - \exp(-2n\gamma^2)$. Applying the union
bound over all $h \in \cH$ and using the lower bound on $n$ completes the
proof.
%
\end{proof}
Next, we show that $\cH(\cG, m)$ can be covered by a finite subset. Since each
classifier in $\cH$ is determined by the choice of $m$ functions from $\cG$ and
mixing weights $\alpha \in \Delta_m$, we will construct finite covers of $\cG$
and $\Delta_m$. Our covers $\hat \cG$ and $\hat \Delta_m$ will guarantee that
for every $g \in \cG$, there exists $\hat g \in \hat \cG$ such that $\prob_{x
\sim P}(g(x) \neq \hat g(x)) \leq \gamma/m$. Similarly, for any mixing weights
$\alpha \in \Delta_m$, there exists $\hat \alpha \in \Delta_m$ such that
$\norm{\alpha - \hat \alpha}_1 \leq \gamma$. If $h \in \cH(\cG,m)$ is the
mixture of $g_1, \dots, g_m$ with weights $\alpha$, we let $\hat h$ be the
mixture of $\hat g_1, \dots, \hat g_m$ with weights $\hat \alpha$. This
approximation has two sources of error: first, for a random individual $x \sim
P$, there is probability up to $\gamma$ that at least one $g_i(x)$ will disagree
with $\hat g_i(x)$, in which case $h$ and $\hat h$ may assign completely
different outcome distributions. Second, even in the high-probability event that
$g_i(x) = \hat g_i(x)$ for all $i \in [m]$, the mixing weights are not
identical, resulting in a small perturbation of the outcome distribution
assigned to $x$.
\begin{lem} \label{lem:finiteCover}
%
Let $\cG$ be a family of deterministic classifiers with Natarajan dimension $d$,
and let $\cH = \cH(\cG, m)$ for some $m\in\mathbb{N}$. For any $\gamma > 0$, there
exists a subset $\hat \cH \subset \cH$ of size $O\bigl(\frac{(dm|\cY|^2 \log(m
|\cY| / \gamma))^{dm}}{\gamma^{(d+1)m}}\bigr)$ such that for every $h \in \cH$
there exists $\hat h \in \cH$ satisfying:
%
\begin{enumerate}[topsep=0pt,itemsep=0pt,leftmargin=*]
%
\item $\prob_{x \sim P}( \norm{h(x) - \hat h(x)}_1 > \gamma) \leq \gamma$.
%
\item If $S$ is an i.i.d.~sample of individuals of size
$O(\frac{m^2}{\gamma^2}(d \log |\cY| + \log \frac{1}{\delta}))$ then w.p.
$\geq 1-\delta$, we have $\norm{h(x) - \hat h(x)}_1 \leq \gamma$ for all but
a $2\gamma$-fraction of $x \in S$.
%
\end{enumerate}
%
\end{lem}
\begin{proof}
%
As described above, we begin by constructing finite covers of $\Delta_m$ and
$\cG$. First, let $\hat \Delta_m \subset \Delta_m$ be the set of distributions
over $[m]$ where each coordinate is a multiple of $\gamma/m$. Then we have
$|\hat \Delta_m| = O( (\frac{m}{\gamma})^m)$ and for every $p \in \Delta_m$,
there exists $q \in \hat \Delta_m$ such that $\norm{p - q}_1 \leq \gamma$.
In order to find a small cover of $\cG$, we use the fact that it has low
Natarajan dimension. This implies that the number of effective functions in
$\cG$ when restricted to a sample $S'$ grows only polynomially in the size of
$S'$. At the same time, if two functions in $\cG$ agree on a large sample,
they will also agree with high probability on the distribution.
Formally, let $S'$ be an i.i.d.~sample drawn from $P$ of size
$O(\frac{m^2}{\gamma^2}d\log|\cY|)$, and let $\hat \cG = \cG\resto_{S'}$ be any
minimal subset of $\cG$ that realizes all possible labelings of $S'$ by
functions in $\cG$. We now argue that with probability 0.99, for every $g \in
\cG$ there exists $\hat g \in \hat \cG$ such that $\prob_{x \sim P}(g(x) \neq
\hat g(x)) \leq \gamma/m$. For any pair of functions $g,g' \in \cG$, let $(g,g') :
\cX \to \cY^2$ be the function given by $(g,g')(x) = (g(x), g'(x))$, and let
$\cG^2 = \{(g,g') \,:\, g,g' \in \cG\}$. The Natarajan dimension of $\cG^2$ is
at most $2d$ (see Lemma~\ref{lem:G2} in Appendix~\ref{app:mixtures}). Moreover, consider the loss $c : \cG^2 \times \cX \to \{0,1\}$ given by $c(g,g',x) =
\ind{g(x) \neq g'(x)}$. Applying Lemma~\ref{lem:natarajanAgnostic} with the
chosen size of $|S'|$ ensures that with probability at least $0.99$ every pair
$(g,g') \in \cG^2$ satisfies \[\left|\expect_{x \sim P}[c(g,g',x)] -
\frac{1}{|S'|} \sum_{x \in S'} c(g,g',x)\right| \leq \frac{\gamma}{m}.\] By the definition
of $\hat \cG$, for every $g \in \cG$, there exists $\hat g \in \hat \cG$ for
which $c(g,\hat g,x) = 0$ for all $x \in S'$, which implies that $\prob_{x \sim
P}(g(x) \neq \hat g(x)) \leq \gamma/m$.
Using Lemma~\ref{lem:natarajan} to
bound the size of $\hat \cG$, we have that \[|\hat \cG| \leq |S'|^d |\cY|^{2d} =
O\left(\left(\frac{m^2}{\gamma^2}d|\cY|^2\log|\cY|\right)^d\right).\] Since this
construction succeeds with non-zero probability, we are guaranteed that such a
set $\hat \cG$ exists. Finally, by an identical uniform convergence argument,
it follows that if $S$ is a fresh i.i.d.~sample of the size given in Item 2 of the lemma's
statement, then, with probability at least $1-\delta$, every $g$ and $\hat g$ will disagree on at most a
$2\gamma/m$-fraction of $S$, since they disagree with probability at most
$\gamma/m$ on $P$.
Next, let $\hat \cH = \{ h_{\vec g, \alpha} \, : \, \vec g \in \hat G^m, \alpha
\in \hat \Delta_m\}$ be the same family as $\cH$, except restricted to choosing
functions from $\hat \cG$ and mixing weights from $\hat \Delta_m$. Using the
size bounds above and the fact that ${N \choose m} = O((\frac{N}{m})^m)$, we
have that \[|\hat \cH| = {|\hat \cG| \choose m} \cdot |\hat \Delta_m| =
O\left(\frac{(dm^2|\cY|^2 \log(m |\cY| /
\gamma))^{dm}}{\gamma^{(2d+1)m}}\right).\]
Suppose that $h$ is the mixture of $g_1, \dots, g_m \in \cG$ with weights
$\alpha \in \Delta_m$. Let $\hat g_i$ be the approximation to $g_i$ for each
$i$, let $\hat \alpha \in \hat \Delta_m$ be such that $\norm{\alpha - \hat
\alpha}_1 \leq \gamma$, and let $\hat h$ be the random mixture of $\hat g_1,
\dots, \hat g_m$ with weights $\hat \alpha$. For an individual $x$ drawn from
$P$, we have $g_i(x) \neq \hat g_i(x)$ with probability at most $\gamma/m$,
and therefore they all agree with probability at least $1-\gamma$. When this
event occurs, we have $\norm{h(x) - \hat h(x)}_1 \leq \norm{\alpha - \hat
\alpha}_1 \leq \gamma$.
The second part of the claim follows by similar reasoning, using the fact that
for the given sample size $|S|$, with probability at least $1-\delta$, every
$g \in \cG$ disagrees with its approximation $\hat g \in \hat \cG$ on at most
a $2\gamma/m$-fraction of $S$. This means that $\hat g_i(x) = g_i(x)$ for all $i \in
[m]$ on at least a $(1-2\gamma)$-fraction of the individuals $x$ in $S$. For these
individuals, $\norm{h(x) - \hat h(x)}_1 \leq \norm{\alpha - \hat \alpha}_1 \leq \gamma$.
\end{proof}
Combining the generalization guarantee for finite families given in
Lemma~\ref{lem:finiteGeneralize} with the finite approximation given in
Lemma~\ref{lem:finiteCover}, we are able to show that envy-freeness also
generalizes for $\cH(\cG,m)$.
\begin{proof}[Proof of Theorem~\ref{thm:mixtureGeneralize}]
%
Let $\hat \cH$ be the finite approximation to $\cH$ constructed in
Lemma~\ref{lem:finiteCover}. If the sample is of size $|S| =
O(\frac{1}{\gamma^2}(dm \log(dm|\cY|\log|\cY|/\gamma) + \log
\frac{1}{\delta}))$, we can apply Lemma~\ref{lem:finiteGeneralize} to this
finite family, which implies that for any $\beta' \geq 0$, with probability at
least $1-\delta/2$ every $\hat h \in \hat \cH$ that is
$(\alpha',\beta')$-pairwise-EF on $S$ (for any $\alpha'$) is also
$(\alpha'+\gamma, \beta')$-EF on $P$. We apply this lemma with $\beta' = \beta +
2\gamma$. Moreover, from Lemma~\ref{lem:finiteCover}, we know that if $|S| =
O(\frac{m^2}{\gamma^2}(d \log|\cY| + \log \frac{1}{\delta}))$, then with
probability at least $1-\delta/2$, for every $h \in \cH$, there exists $\hat
h \in \hat \cH$ satisfying $\norm{h(x) - \hat h(x)}_1 \leq \gamma$ for all but a
$2\gamma$-fraction of the individuals in $S$. This implies that on all but at
most a $4\gamma$-fraction of the pairs in $S$, $h$ and $\hat h$ satisfy this
inequality for both individuals in the pair. Assume these high probability
events occur. Finally, from Item 1 of the lemma we have that $\prob_{x_1,x_2
\sim P}(\max_{i=1,2}\norm{h(x_i) - \hat h(x_i)}_1 > \gamma) \leq 2\gamma$.
Now let $h \in \cH$ be any classifier that is
$(\alpha,\beta)$-pairwise-EF on $S$. Since the utilities are in $[0,1]$ and
$\max_{x=x_i,x_i'} \norm{h(x) - \hat h(x)}_1 \leq \gamma$ for all but a
$4\gamma$-fraction of the pairs in $S$, we know that $\hat h$ is $(\alpha +
4\gamma, \beta + 2\gamma)$-pairwise-EF on $S$. Applying the envy-freeness
generalization guarantee (Lemma~\ref{lem:finiteGeneralize}) for $\hat \cH$, it follows that $\hat h$ is also
$(\alpha + 5\gamma, \beta + 2\gamma)$-EF on $P$. Finally, using the
fact that \[\prob_{x_1,x_2 \sim P}\left(\max_{i=1,2} \norm{h(x_i) - \hat h(x_i)}_1 >
\gamma\right) \leq 2\gamma,\] it follows that $h$ is $(\alpha + 7\gamma, \beta +
4\gamma)$-EF on $P$.
%
\end{proof}
It is worth noting that the (exponentially large) approximation $\hat \cH$ is
only used in the generalization analysis; importantly, an ERM algorithm need not construct
it.
\section{Discussion}\label{sec:disc}
We believe that envy-freeness gives a new, useful perspective on individual fairness in classification --- when individuals have rich utility functions, which, as we have argued in detail in Section~\ref{sec:intro}, is the case in advertising. However, in some domains there are only two possible outcomes, one of which is `good' and the other 'bad'; examples include predicting whether an individual would default on a loan, and whether an offender would recidivate. In these degenerate cases envy-freeness would require that the classifier assign each and every individual the exact same probability of obtaining the `good' outcome, which, clearly, is not a reasonable constraint.
\edit{It is also worth noting that we have not directly addressed the problem of \emph{computing} the loss-minimizing envy-free classifier from a given family on a given sample of individuals. Just like in the work of \citet{DHPR+12}, when the classifier is arbitrary, this problem can be written as a linear program of polynomial size in the number of outcomes, because envy-freeness amounts to a set of linear constraints. In both settings, though, one needs to restrict the family of classifiers to obtain good sample complexity, and, moreover, the na\"ive formulation would be intractable when dealing with a combinatorial space of outcomes. Nevertheless, the linearity of envy-freeness may enable practical mixed-integer linear programming formulations with respect to certain families. More generally, given the wealth of powerful optimization tools at the community's disposal, we do not view computational complexity as a long-term obstacle to implementing our approach.}
\bibliographystyle{aaai}
\section{Introduction} \label{sec:intro}
The study of fairness in machine learning is driven by an abundance of examples where learning algorithms were perceived as discriminating against protected groups~\citep{Sween13,DTD15}. Addressing this problem requires a conceptual --- perhaps even philosophical --- understanding of what fairness means in this context. In other words, the million dollar question is (arguably\footnote{Recent work takes a somewhat different view~\citep{KRPH+17}.}) this: What are the formal constraints that fairness imposes on learning algorithms? On a very high level, most of the answers proposed so far~\citep{LRT11,DHPR+12,ZWSP+13,MPS16,JKMA16,ZVGG+17}
fall into two (partially overlapping) categories: individual fairness notions, and group fairness notions.
In the former category, the best known example is the influential fair classification model of \citet{DHPR+12}. The model involves a set of individuals and a set of outcomes. It is instructive to think of financially-motivated settings where the outcomes are, say, credit card offerings or displayed advertisements, and a loss function represents the benefit (e.g., in terms of revenue) of mapping a given individual to a given outcome. The centerpiece of the model is a \emph{similarity metric} on the space of individuals; it is specific to the classification task at hand, and ideally captures the ethical ground truth about relevant attributes. For example, a man and a woman who are similar in every other way should be considered similar for the purpose of credit card offerings, but perhaps not for lingerie advertisements. Assuming such a metric is available, fairness can be naturally formalized as a Lipschitz constraint, which requires that individuals who are close according to the similarity metric be mapped to distributions over outcomes that are close according to some standard metric (such as total variation). The algorithmic problem is then to find a classifier that minimizes loss, subject to the Lipschitz constraint.
As attractive as this model is, it has one clear weakness from a practical viewpoint: the availability of a similarity metric. \citet{DHPR+12} are well aware of this issue; they write that justifying this assumption is ``one of the most challenging aspects'' of their approach. They add that ``in reality the metric used will most likely only be society's current best approximation to the truth.'' But, despite recent progress on automating ethical decisions in certain domains~\citep{NGAD+18,FSSD+18}, the task-specific nature of the similarity metric makes even a credible approximation thereof seem unrealistic. In particular, if one wanted to learn a similarity metric, it is unclear what type of examples a relevant dataset would consist of.
An alternative notion of individual fairness, therefore, is called for. And our proposal draws on an extensive body of work on rigorous approaches to fairness, which --- modulo one possible exception (see Section~\ref{subsec:rel}) --- has not been tapped by machine learning researchers: the literature on \emph{fair division}~\citep{BT96,Moul03}. The most prominent notion is that of \emph{envy-freeness}~\citep{Fol67,Var74}, which, in the context of the allocation of goods, requires that the utility of each individual for his allocation be at least as high as his utility for the allocation of any other individual; this is the gold standard of fairness for problems such as cake cutting \citep{RW98,Pro13} and rent division \citep{Su99,GMPZ17}.
Similarly, in the classification setting, envy-freeness would simply mean that the utility of each individual for his distribution over outcomes is at least as high as his utility for the distribution over outcomes assigned to any other individual. For example, it may well be the case that Bob is offered a worse credit card than that offered to Alice (in terms of, say, annual fees), but this outcome is not unfair if Bob is genuinely more interested in the card offered to him because he does not qualify for Alice's card, \edit{or because its specific rewards program better fits his needs. Such rich utility functions are also evident in the context of job advertisements~\citep{DTD15}: people generally want higher paying jobs, but would presumably have higher utility for seeing advertisements for jobs that better fit their qualifications and interests.}
Of course, as before, envy-freeness requires access to individuals' utility functions, but --- in stark contrast to the similarity metric of \citet{DHPR+12} --- \edit{we do not view this assumption as a barrier to implementation.} Indeed, there are a variety of techniques for learning utility functions~\citep{CKO01,NJ04,BCIW12}. \edit{Moreover, in our running example of advertising, one can even think of standard measures like expected click-through rate (CTR) as an excellent proxy for utility.}
It is worth noting that the classification setting is different from classic fair division problems in that the ``goods'' (outcomes) are non-excludable. In fact, one envy-free solution simply assigns each individual to his favorite outcome; but when the loss function disagrees with the utility functions, it may be possible to achieve smaller loss without violating the envy-freeness constraint.
In summary, we view envy-freeness as a compelling, well-established, and, importantly, practicable notion of individual fairness for classification tasks. Our goal is to understand its learning-theoretic properties.
\subsection{Our Results}
The technical challenge we face is that the space of individuals is potentially huge, yet we seek to provide universal envy-freeness guarantees. To this end, we are given a sample consisting of individuals drawn from an unknown distribution. We are interested in learning algorithms that minimize loss, subject to satisfying the envy-freeness constraint, \emph{on the sample}. Our primary technical question is that of generalizability, that is, \emph{given a classifier that is envy free on a sample, is it approximately envy free on the underlying distribution?} \edit{Surprisingly, \citet{DHPR+12} do not study generalizability in their model, and we are aware of only one subsequent paper that takes a learning-theoretic viewpoint on individual fairness and gives theoretical guarantees (see Section~\ref{subsec:rel}).}
In Section~\ref{sec:arb}, we do not constrain the classifier in question. Therefore, we need some strategy to extend a classifier that is defined on a sample; assigning an individual the same outcome as his \emph{nearest neighbor} in the sample is a popular choice. However, we show that \emph{any} strategy for extending a classifier from a sample, on which it is envy free, to the entire set of individuals is unlikely to be approximately envy free on the underlying distribution, unless the sample is exponentially large.
For this reason, in Section~\ref{sec:mixtures}, we focus on structured families of classifiers. On a high level, our goal is to relate the combinatorial richness of the family to generalization guarantees. One obstacle is that standard notions of dimension do not extend to the analysis of randomized classifiers, whose range is \emph{distributions} over outcomes (equivalently, real vectors). We circumvent this obstacle by considering mixtures of \emph{deterministic} classifiers that belong to a family of bounded Natarajan dimension (an extension of the well-known VC dimension to multi-class classification). Our main technical result asserts that, under this assumption, envy-freeness on a sample does generalize to the underlying distribution, even if the sample is relatively small (its size grows almost linearly in the Natarajan dimension). Finally, we discuss the implications of this result in Section~\ref{sec:disc}.
\subsection{Related Work}
\label{subsec:rel}
Conceptually, our work is most closely related to work by \citet{ZVGG+17}. They are interested in group notions of fairness, and advocate preference-based notions instead of parity-based notions. In particular, they assume that each group has a utility function for \emph{classifiers}, and define the \emph{preferred treatment} property, which requires that the utility of each group for its own classifier be at least its utility for the classifier assigned to any other group. Their model and results focus on the case of binary classification where there is a desirable outcome and an undesirable outcome, so the utility of a group for a classifier is simply the fraction of its members that are mapped to the desirable outcome. Although, at first glance, this notion seems similar to envy-freeness, it is actually fundamentally different.\footnote{On a philosophical level, the fair division literature deals exclusively with individual notions of fairness. In fact, even in group-based extensions of envy-freeness~\citep{MS17} the allocation is shared by groups, but individuals must not be envious. We subscribe to the view that group-oriented notions (such as statistical parity) are objectionable, because the outcome can be patently unfair to individuals.}
Our paper is also completely different from that of \citeauthor{ZVGG+17}~in terms of technical results; theirs are purely empirical in nature, and focus on the increase in accuracy obtained when parity-based notions of fairness are replaced with preference-based ones.
\edit{Very recent, concurrent work by \citet{RY18} provides generalization guarantees for the metric notion of individual fairness introduced by \citet{DHPR+12}, or, more precisely, for an approximate version thereof.
There are two main differences compared to our work: first, we propose envy-freeness as an alternative notion of fairness that circumvents the need for a similarity metric. Second, they focus on randomized \emph{binary} classification, which amounts to learning a real-valued function, and so are able to make use of standard Rademacher complexity results to show generalization. By contrast, standard tools do not directly apply in our setting. It is worth noting that several other papers provide generalization guarantees for notions of group fairness, but these are more distantly related to our work \citep{ZWSP+13,WGOS17,DOBS+18,KNRW18,HKRR18}.}
\section{The Model} \label{sec:problem}
We assume that there is
a space $\cX$ of individuals, a finite space $\cY$ of outcomes, and a utility function $u : \cX \times \cY \to [0,1]$ encoding the preferences
of each individual for the outcomes in $\cY$. In the advertising example, individuals are users, outcomes are advertisements, and the utility function reflects the benefit an individual derives from being shown a particular advertisement. For any distribution $p \in \Delta(\cY)$ (where $\Delta(\cY)$ is the set of distributions over $\cY$) we let
$u(x,p) = \expect_{y \sim p}[u(x,y)]$ denote individual $x$'s expected utility
for an outcome sampled from $p$. We refer to a function $h :
\cX \to \Delta(\cY)$ as a \emph{classifier}, even though it can return a distribution over outcomes.
\subsection{Envy-Freeness}
Roughly speaking, a classifier $h :
\cX \to \Delta(\cY)$ is envy free if no individual prefers the
outcome distribution of someone else over his own.
\begin{defn}
%
A classifier $h : \cX \to \Delta(\cY)$ is \emph{envy free (EF)} on a set $S$ of
individuals if $u(x,h(x)) \geq u(x,h(x'))$ for all $x,x' \in S$. Similarly,
$h$ is \emph{$(\alpha,\beta)$-EF} with respect to a distribution $P$ on $\cX$ if
\[\prob_{x,x'\sim P}\bigl( u(x,h(x)) < u(x,h(x')) - \beta \bigr) \leq \alpha.\]
Finally, $h$ is \emph{$(\alpha,\beta)$-pairwise EF} on a set of pairs of
individuals $S = \{(x_i,x'_i)\}_{i=1}^n$ if \[\frac{1}{n} \sum_{i=1}^n
\ind{u(x_i, h(x_i)) < u(x_i, h(x'_i)) - \beta} \leq \alpha.\]
%
\end{defn}
Any classifier that is EF on a sample $S$ of individuals is also
$(\alpha,\beta)$-pairwise EF on any pairing of the individuals in $S$, for any $\alpha \geq 0$ and $\beta \geq 0$. The
weaker pairwise EF condition is all that is required for our
generalization guarantees to hold.
\subsection{Optimization and Learning}
Our formal learning problem can be stated as follows. Given sample access to an
unknown distribution $P$ over individuals $\cX$ and their utility functions, and a known loss function $\ell : \cX \times \cY \to
[0,1]$, find a classifier $h : \cX \to \Delta(\cY)$ that is
$(\alpha,\beta)$-EF with respect to $P$ minimizing expected loss $\expect_{x \sim
P}[\ell(x, h(x))]$, where for $x \in \cX$ and $p \in \Delta(\cY)$, $\ell(x,p) =
\expect_{y \sim p}[\ell(x,y)]$.
We follow the empirical risk minimization (ERM)
learning approach, i.e., we collect a sample of individuals drawn i.i.d~from
$P$ and find an EF classifier with low loss on the sample. Formally, given a sample of individuals $S=\{x_1,\ldots,x_n\}$ and their utility functions $u_{x_i}(\cdot)=u(x_i,\cdot)$, we are interested in a classifier $h:S\to\Delta(\cY)$ that minimizes $\sum_{i=1}^n \ell(x_i,h(x_i))$ among all classifiers that are EF on $S$. The algorithmic problem itself is beyond the scope of the current paper; see Section~\ref{sec:disc} for further discussion.
Recall that we consider randomized classifiers that can assign a distribution over outcomes to each of the individuals. However, one might wonder whether the EF classifier that minimizes loss on a sample happens to always be deterministic. Or, at least, the optimal deterministic classifier on the sample might incur a loss that is very close to that of the optimal randomized classifier. If this were true, we could restrict ourselves to classifiers of the form $h: \cX \to \cY$, which would be much easier to analyze. Unfortunately, it turns out that this is not the case. In fact, there could be an arbitrary (multiplicative) gap between the optimal randomized EF classifier and the optimal deterministic EF classifier. The intuition behind this is as follows. A deterministic classifier that has very low loss on the sample, but is not EF, would be completely discarded in the deterministic setting. On the other hand, a randomized classifier could take this loss-minimizing deterministic classifier and mix it with a classifier with high ``negative envy'', so that the mixture ends up being EF and at the same time has low loss. This is made concrete in Example~\ref{ex:det-random-gap} in the appendix.
\section{Arbitrary Classifiers}\label{sec:arb}
An important (and typical) aspect of our learning problem is that the classifier
$h$ needs to provide an outcome distribution for every individual, not just
those in the sample. For example, if $h$ chooses advertisements for visitors of
a website, the classifier should still apply when a new visitor arrives.
Moreover, when we use the classifier for new individuals, it must continue to be
EF. In this section, we consider two-stage approaches that first choose outcome
distributions for the individuals in the sample, and then extend those decisions
to the rest of $\cX$.
In more detail, we are given a sample $S = \{x_1, \dots, x_n\}$ of individuals
and a classifier $h : S \to \Delta(\cY)$ assigning outcome distributions to each
individual. Our goal is to extend these assignments to a classifier
$\overline{h} : \cX \to \Delta(\cY)$ that can be applied to new individuals as
well. For example, $h$ could be the loss-minimizing EF classifier on the sample
$S$.
For this section, we
assume that $\cX$ is equipped with a distance metric $d$. Moreover, we assume in this section that the utility function $u$ is $L$-Lipschitz on $\cX$. That is, for
every $y \in \cY$ and for all $x, x' \in \cX$, we have $|u(x, y) - u(x', y)|
\leq L \cdot d(x,x')$.
Under the foregoing assumptions, one natural way to extend the classifier on the sample
to all of $\cX$ is to assign new individuals the same outcome distribution
as their nearest neighbor in the sample. Formally, for a set $S \subset \cX$ and any individual
$x \in \cX$, let $\NN_S(x)\in \text{arg\,min}_{x'\in S}d(x,x')$ denote the nearest neighbor of $x$ in $S$ with
respect to the metric $d$ (breaking ties arbitrarily). The following simple result (whose proof is relegated to Appendix~\ref{app:arb}) establishes that this approach preserves envy-freeness in cases where
the sample is exponentially large.
\begin{thm} \label{thm:nn-upper-bound}
%
Let $d$ be a metric on $\cX$, $P$ be a distribution on $\cX$, and $u$ be
an $L$-Lipschitz utility function. Let $S$ be a set of individuals such that
there exists $\hat \cX \subset \cX$ with $P(\hat \cX) \geq 1-\alpha$ and
$\sup_{x \in \hat \cX}(d(x, \NN_S(x)) \leq \beta/(2L)$. Then for any
classifier $h : S \to \Delta(\cY)$ that is EF on $S$, the extension
$\overline{h} : \cX \to \Delta(\cY)$ given by $\overline{h}(x) = h(\NN_S(x))$
is $(\alpha, \beta)$-EF on $P$.
%
\end{thm}
The conditions of Theorem~\ref{thm:nn-upper-bound} require that the set of
individuals $S$ is a $\beta/(2L)$-net for at least a $(1-\alpha)$-fraction of
the mass of $P$ on $\cX$. In several natural situations, an exponentially large
sample guarantees that this occurs with high probability. For example, if $\cX$
is a subset of $\reals^q$, $d(x,x') = \norm{x-x'}_2$, and $\cX$ has diameter at
most $D$, then for any distribution $P$ on $\cX$, if $S$ is an i.i.d.~sample of size
$O(\frac{1}{\alpha}(\frac{LD\sqrt{q}}{\beta})^q(q \log \frac{LD\sqrt{q}}{\beta} +
\log \frac{1}{\delta}))$, it will satisfy the conditions of
Theorem~\ref{thm:nn-upper-bound} with probability at least $1-\delta$. This
sampling result is folklore, but, for the sake of completeness, we prove it in Lemma~\ref{lem:euclideanCover}
of Appendix~\ref{app:arb}.
However, the exponential upper bound given by the nearest neighbor strategy is as far as we can go in terms of generalizing envy-freeness from a sample (without further assumptions). Specifically, our next result establishes that \emph{any} algorithm --- even randomized --- for extending classifiers from the sample to the entire space $\cX$
requires an exponentially large sample of individuals to ensure envy-freeness on
the distribution $P$. The proof of Theorem~\ref{thm:lb} can be found in Appendix~\ref{app:arb}.
\begin{thm}
\label{thm:lb}
There exists a space of individuals $\cX \subset \reals^q$, and a distribution $P$ over $\cX$ such that, for every randomized algorithm $\mathcal{A}$ that extends classifiers on a sample to $\cX$, there exists an $L$-Lipschitz utility function $u$ such that, when a sample of individuals $S$ of size $n = 4^q / 2$ is drawn from $P$ without replacement, there exists an EF classifier on $S$ for which, with probability at least $1 - 2\exp(-4^q/100) - \exp(-4^q/200)$ jointly over the randomness of $\mathcal{A}$ and $S$, its extension by $\mathcal{A}$ is not $(\alpha, \beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$.
\end{thm}
We remark that a similar result would hold even if we sampled $S$ with replacement; we sample here without replacement purely for ease of exposition.
\begin{proof}[Proof of Theorem~\ref{thm:lb}]
Let the space of individuals be $\cX = [0,1]^q$ and the outcomes be $\cY = \{0,1\}$. We partition the space $\cX$ into cubes of side length $s = 1/4$. So, the total number of cubes is $m = \left(1/s\right)^q = 4^q$. Let these cubes be denoted by $c_1, c_2, \dots c_m$, and let their centers be denoted by $\mu_1, \mu_2, \dots \mu_m$. Next, let $P$ be the uniform distribution over the centers $\mu_1, \mu_2, \dots \mu_m$. For brevity, whenever we say ``utility function'' in the rest of the proof, we mean ``$L$-Lipschitz utility function.''
To prove the theorem, we use Yao's minimax principle~\citep{Yao77}. Specifically, consider the following two-player zero sum game. Player 1 chooses a deterministic algorithm $\mathcal{D}$ that extends classifiers on a sample to $\cX$, and player 2 chooses a utility function $u$ on $\cX$. For any subset $S \subset \cX$, define the classifier $h_{u,S}:S \to \cY$ by assigning each individual in $S$ to his favorite outcome with respect to the utility function $u$, i.e. $h_{u,S}(x) = \text{arg\,max}_{y \in \cY} u(x,y)$ for each $x \in S$, breaking ties lexicographically. Define the cost of playing algorithm $\mathcal{D}$ against utility function $u$ as the probability over the sample $S$ (of size $m/2$ drawn from $P$ without replacement) that the extension of $h_{u,S}$ by $\mathcal{D}$ is not $(\alpha, \beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$. Yao's minimax principle implies that for any randomized algorithm~$\mathcal{A}$, its expected cost with respect to the worst-case utility function~$u$ is at least as high as the expected cost of any distribution over utility functions that is played against the best deterministic algorithm $\mathcal{D}$ (which is tailored for that distribution). Therefore, we establish the desired lower bound by choosing a specific distribution over utility functions, and showing that the best deterministic algorithm against it has an expected cost of at least $1 - 2\exp(-m/100) - \exp(-m/200)$.
To define this distribution over utility functions, we first sample outcomes $y_1, y_2, \dots, y_m$ i.i.d. from Bernoulli($1/2$). Then, we associate each cube center $\mu_i$ with the outcome $y_i$, and refer to this outcome as the \textit{favorite} of $\mu_i$. For brevity, let $\neg y$ denote the outcome other than $y$, i.e. $\neg y = (1-y)$. For any $x \in \cX$, we define the utility function as follows. Letting $c_j$ be the cube that $x$ belongs to,
\begin{equation}
\label{eq:util}
u(x,y_j) = L \left[\frac{s}{2} - \|x - \mu_j\|_{\infty}\right]; \quad u(x, \neg y_j) = 0.
\end{equation}
See Figure~\ref{fig:grid} for an illustration.
\begin{figure}[t]
\centering
\includegraphics[width=0.25\columnwidth]{grid.jpg}
\caption{Illustration of $\cX$ and an example utility function $u$ for $d=2$. Red shows preference for $1$, blue shows preference for $0$, and darker shades correspond to more intense preference. (The gradients are rectangular to match the $L_\infty$ norm, so, strangely enough, the misleading X pattern is an optical illusion.)}
\label{fig:grid}
\end{figure}
We claim that the utility function of Equation~\eqref{eq:util} is indeed $L$-Lipschitz with respect to any $L_p$ norm. This is because for any cube $c_i$, and for any $x, x' \in c_i$, we have
\begin{align*}
\left|u(x,y_i) - u(x',y_i)\right| &= L\left|\|x - \mu_i\|_{\infty} - \|x' - \mu_i\|_{\infty}\right|\\
&\leq L\|x - x'\|_{\infty}
\leq L\|x - x'\|_{p}.
\end{align*}
Moreover, for the other outcome, we have $u(x, \neg y_i) = u(x', \neg y_i) = 0$. It follows that $u$ is $L$-Lipschitz within every cube. At the boundary of the cubes, the utility for any outcome is $0$, and hence $u$ is also continuous throughout $\cX$. Because it is piecewise Lipschitz and continuous, $u$ must be $L$-Lipschitz throughout $\cX$, with respect to any $L_p$ norm.
Next, let $\mathcal{D}$ be an arbitrary deterministic algorithm that extends classifiers on a sample to $\cX$. We draw the sample $S$ of size $m/2$ from $P$ without replacement. Consider the distribution over favorites of individuals in $S$. Each individual in $S$ has a favorite that is sampled independently from Bernoulli$(1/2)$. Hence, by Hoeffding's inequality, the fraction of individuals in $S$ with a favorite of $0$ is between $\frac{1}{2} - \epsilon$ and $\frac{1}{2} + \epsilon$ with probability at least $1 - 2\exp(-m \epsilon^2)$. The same holds simultaneously for the fraction of individuals with favorite $1$.
Given the sample $S$ and the utility function $u$ on the sample (defined by the instantiation of their favorites), consider the classifier $h_{u,S}$, which maps each individual $\mu_i$ in the sample $S$ to his favorite $y_i$. This classifier is clearly EF on the sample. Consider the extension $h_{u,S}^\mathcal{D}$ of $h_{u,S}$ to the whole of $\cX$ as defined by algorithm $\mathcal{D}$. Define two sets $Z_0$ and $Z_1$ by letting $Z_y = \{\mu_j \notin S \ | \ h_{u,S}^\mathcal{D}(\mu_j) = y\}$, and let $y_*$ denote an outcome that is assigned to at least half of the out-of-sample centers, i.e., an outcome for which $|Z_{y_*}| \geq |Z_{\neg y_*}|$. Furthermore, let $\theta$ denote the fraction of out-of-sample centers assigned to $y_*$. Note that, since $|S| = m/2$, the number of out-of-sample centers is also exactly $m/2$. This gives us $|Z_{y_*}| = \theta \frac{m}{2}$, where $\theta \geq \frac{1}{2}$.
Consider the distribution of favorites in $Z_{y_*}$ (these are independent from the ones in the sample since $Z_{y_*}$ is disjoint from $S$). Each individual in this set has a favorite sampled independently from Bernoulli$(1/2)$. Hence, by Hoeffding's inequality, the fraction of individuals in $Z_{y_*}$ whose favorite is $\neg y_*$ is at least $\frac{1}{2} - \epsilon$ with probability at least $1 - \exp(-\frac{m}{2} \epsilon^2)$. We conclude that with a probability at least
$1 - 2\exp(-m \epsilon^2) - \exp(-\frac{m}{2} \epsilon^2)$,
the sample $S$ and favorites (which define the utility function $u$) are such that: (i)~the fraction of individuals in $S$ whose favorite is $y \in \{0,1\}$ is between $\frac{1}{2} - \epsilon$ and $\frac{1}{2} + \epsilon$, and (ii)~the fraction of individuals in $Z_{y_*}$ whose favorite is $\neg y_*$ is at least $\frac{1}{2} - \epsilon$.
We now show that for such a sample $S$ and utility function~$u$, $h_{u,S}^\mathcal{D}$ cannot be $(\alpha, \beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$. To this end, sample $x$ and $x'$ from $P$. One scenario where $x$ envies $x'$ occurs when (i)~the favorite of $x$ is $\neg y_*$, (ii)~$x$ is assigned to $y_*$, and (iii)~$x'$ is assigned to $\neg y_*$. Conditions (i) and (ii) are satisfied when $x$ is in $Z_{y_*}$ and his favorite is $\neg y_*$. We know that at least a $\frac{1}{2} - \epsilon$ fraction of the individuals in $Z_{y_*}$ have the favorite $\neg y_*$. Hence, the probability that conditions (i) and (ii) are satisfied by $x$ is at least $(\frac{1}{2} - \epsilon)|Z_{y_*}|\frac{1}{m} = (\frac{1}{2} - \epsilon)\frac{\theta}{2}$.
Condition (iii) is satisfied when $x'$ is in $S$ and has favorite $\neg y_*$ (and hence assigned $\neg y_*$), or, if $x'$ is in $Z_{\neg y_*}$. We know that at least a $\left(\frac{1}{2} - \epsilon\right)$ fraction of the individuals in $S$ have the favorite $\neg y_*$. Moreover, the size of $Z_{\neg y_*}$ is $(1-\theta)\frac{m}{2}$. So, the probability that condition (iii) is satisfied by $x'$ is at least
\[\frac{\left(\frac{1}{2} - \epsilon\right)|S| + |Z_{\neg y_*}|}{m} = \frac{1}{2}\left(\frac{1}{2} - \epsilon\right) + \frac{1}{2}(1-\theta).\]
Since $x$ and $x'$ are sampled independently, the probability that all three conditions are satisfied is at least \[\left(\frac{1}{2} - \epsilon\right)\frac{\theta}{2} \cdot \left[\frac{1}{2}\left(\frac{1}{2} - \epsilon\right) + \frac{1}{2}(1-\theta)\right].\] This expression is a quadratic function in $\theta$, that attains its minimum at $\theta = 1$ irrespective of the value of $\epsilon$. Hence, irrespective of $\mathcal{D}$, this probability is at least $\left[\frac{1}{2}\left(\frac{1}{2} - \epsilon\right)\right]^2$. For concreteness, let us choose $\epsilon$ to be $1/10$ (although it can be set to be much smaller). On doing so, we have that the three conditions are satisfied with probability at least $1/25$. And when these conditions are satisfied, we have $u(x, h_{u,S}^\mathcal{D}(x)) = 0$ and $u(x, h_{u,S}^\mathcal{D}(x')) = Ls/2$, i.e., $x$ envies $x'$ by $Ls/2 = L/8$. This shows that, when $x$ and $x'$ are sampled from $P$, with probability at least $1/25$, $x$ envies $x'$ by $L/8$. We conclude that with probability at least $1 - 2\exp(-m/100) - \exp(-m/200)$ jointly over the selection of the utility function $u$ and the sample $S$, the extension of $h_{u,S}$ by $\mathcal{D}$ is not $(\alpha,\beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$.
To convert the joint probability into expected cost in the game, note that for two discrete, independent random variables $X$ and $Y$, and for a Boolean function $\mathcal{E}(X,Y)$, it holds that
\begin{equation}
\label{eq:exp}
\operatorname{Pr}_{X,Y}(\mathcal{E}(X,Y) = 1) = \mathbb{E}_X\left[\operatorname{Pr}_Y(\mathcal{E}(X,Y) = 1)\right].
\end{equation}
Given sample $S$ and utility function $u$, let $\mathcal{E}(u, S)$ be the Boolean function that equals $1$ if and only if the extension of $h_{u,S}$ by $\mathcal{D}$ is not $(\alpha,\beta)$-EF with respect to $P$ for any $\alpha < 1/25$ and $\beta < L/8$. From Equation~\eqref{eq:exp},
$\operatorname{Pr}_{u,S}(\mathcal{E}(u,S) = 1)$ is equal to $\mathbb{E}_u\left[\operatorname{Pr}_S(\mathcal{E}(u,S) = 1)\right]$. The latter term is exactly the expected value of the cost, where the expectation is taken over the randomness of $u$. It follows that the expected cost of (any) $\mathcal{D}$ with respect to the chosen distribution over utilities is at least $1 - 2\exp(-m/100) - \exp(-m/200)$.
%
\end{proof}
\section{Low-Complexity Families of Classifiers} \label{sec:mixtures}
In this section we show that (despite Theorem~\ref{thm:lb}) generalization for envy-freeness is possible using
much smaller samples of individuals, as long as we restrict ourselves to
choosing a classifier from a family of relatively low complexity.
In more detail, two
classic complexity measures are the
VC-dimension~\citep{VC} for binary classifiers, and the Natarajan dimension~\citep{Natarajan89:Dimension} for multi-class
classifiers. However, to the best of our knowledge, there is no suitable dimension
directly applicable to functions ranging over distributions, which in our case
can be seen as $|\cY|$-dimensional real vectors. One possibility would be to restrict
ourselves to deterministic classifiers of the type $h : \cX \to \cY$. However, we have seen in Section~\ref{sec:problem} that envy-freeness is a very strong constraint on deterministic classifiers. Instead,
we will consider a family $\cH$ consisting of randomized mixtures of
deterministic classifiers belonging to a family $\cG \subset \{g : \cX \to \cY\}$ of
low Natarajan dimension. This allows us to adapt Natarajan-dimension-based
generalization results to our setting while still working with randomized
classifiers.
\subsection{Natarajan Dimension Primer}
\label{subsec:nat}
Before presenting our main result, we briefly summarize the
definition and relevant properties of the Natarajan dimension. For more details, we refer the reader to
\citep{SBD14:MLBook}.
We say that a family $\cG$ \emph{multi-class shatters} a set of
points $x_1, \dots, x_n$ if there exist labels $y_1, \dots y_n$ and $y'_1,
\dots, y'_n$ such that for every $i \in [n]$ we have $y_i \neq y'_i$, and for any
subset $C \subset[n]$ there exists $g \in \cG$ such that $g(x_i) = y_i$ if $i
\in C$ and $g(x_i)=y'_i$ otherwise. The Natarajan dimension of a family $\cG$ is the
cardinality of the largest set of points that can be multi-class shattered by
$\cG$.
For example, suppose we have a feature map $\Psi : \cX \times \cY \to \reals^q$
that maps each individual-outcome pair to a $q$-dimensional feature vector, and
consider the family of functions that can be written as $g(x) =
\text{arg\,max}_{y \in \cY} w^\tp \Psi(x,y)$ for weight vectors $w \in
\reals^q$. This family has Natarajan dimension at most $q$.
For a set $S \subset \cX$ of points, we let $\cG\resto_S$ denote the restriction
of $\cG$ to $S$, which is any subset of $\cG$ of minimal size such that for
every $g \in \cG$ there exists $g' \in \cG\resto_S$ such that $g(x) = g'(x)$
for all $x \in S$. The size of $\cG\resto_S$ is the number of different
labelings of the sample $S$ achievable by functions in $\cG$. The following
Lemma is the analogue of Sauer's lemma for binary classification.
\begin{lem}[Natarajan] \label{lem:natarajan}
%
For a family $\cG$ of Natarajan dimension $d$ and any subset $S \subset \cX$,
we have $\bigl|\cG\resto_S\bigr| \leq |S|^d |\cY|^{2d}$.
%
\end{lem}
Classes of low Natarajan dimension also enjoy the following uniform convergence
guarantee.
\begin{lem} \label{lem:natarajanAgnostic}
Let $\cG$ have Natarajan dimension $d$ and fix a loss function $\ell : \cG
\times \cX \to [0,1]$. For any distribution $P$ over $\cX$, if $S$ is an i.i.d.~sample drawn from $P$ of size $O(\frac{1}{\epsilon^2}(d \log |\cY| + \log
\frac{1}{\delta}))$, then with probability at least $1-\delta$ we have
$
\sup_{g \in \cG} \, \left|
\expect_{x \sim P}[\ell(g,x)] - \frac{1}{n}\sum_{x\in S} \ell(g,x)
\right| \leq \epsilon.
$
\end{lem}
\subsection{Main Result}
We consider the family of classifiers that can be expressed as a randomized
mixture of $m$ deterministic classifiers selected from a family $\cG \subset \{g : \cX
\to \cY\}$. Our generalization guarantees will depend on the complexity of the
family $\cG$, measured in terms of its Natarajan dimension, and the number $m$ of
functions we are mixing. More formally, let $\vec{g} = (g_1, \dots, g_m) \in
\cG^m$ be a vector of $m$ functions in $\cG$ and $\alpha \in \Delta_m$ be a
distribution over $[m]$, where $\Delta_m = \{ p \in \reals^m \,:\, p_i \geq 0,
\sum_i p_i = 1\}$ is the $m$-dimensional probability simplex. Then consider the
function $h_{\vec{g},\alpha} : \cX \to \Delta(\cY)$ with assignment
probabilities given by
$
\prob(h_{\vec{g}, \alpha}(x) = y)
=
\sum_{i=1}^m \ind{g_i(x) = y} \alpha_i.
$
Intuitively, for a given individual $x$, $h_{\vec{g}, \alpha}$ chooses one of
the $g_i$ randomly with probability $\alpha_i$, and outputs $g_i(x)$. Let
\[\cH(\cG, m) = \{ h_{\vec{g},\alpha} : \cX \to \Delta(\cY) \,:\, \vec{g} \in
\cG^m, \alpha \in \Delta_m\}\] be the family of classifiers that can be written
this way. Our main technical result shows that envy-freeness generalizes for
this class.
\begin{thm} \label{thm:mixtureGeneralize}
%
Suppose $\cG$ is a family of deterministic classifiers of Natarajan
dimension $d$, and let $\cH = \cH(\cG,m)$ for $m\in \mathbb{N}$. For any distribution
$P$ over $\cX$, $\gamma>0$, and $\delta > 0$, if $S = \{(x_i, x'_i)\}_{i=1}^n$
is an i.i.d.~sample of pairs drawn from $P$ of size \[n \geq
O\left(\frac{1}{\gamma^2}\left(dm^2 \log \frac{dm|\cY|\log(m|\cY|/\gamma)}{\gamma} + \log
\frac{1}{\gamma}\right)\right),\] then with probability at least $1-\delta$, every classifier $h
\in \cH$ that is $(\alpha,\beta)$-pairwise-EF on $S$ is also
$(\alpha+7\gamma, \beta+4\gamma)$-EF on $P$.
%
\end{thm}
\edit{Theorem~\ref{thm:mixtureGeneralize} is only effective insofar as families of classifiers of low Natarajan dimension are useful. And, indeed, several prominent families have low Natarajan dimension~\citep{DSS12}, including one vs. all (which is a special case of the example given in Section~\ref{subsec:nat}), multiclass SVM, tree-based classifiers, and error correcting output codes.
We now turn to the theorem's proof, which consists of two steps.} First,
we show that envy-freeness generalizes for finite classes. Second, we show that
$\cH(\cG,m)$ can be approximated by a finite subset.
\begin{lem} \label{lem:finiteGeneralize}
%
Let $\cH \subset \{h : \cX \to \Delta(\cY)\}$ be a finite family of
classifiers. For any $\gamma > 0$, $\delta > 0$, and $\beta \geq 0$ if $S =
\{(x_i, x'_i)\}_{i=1}^n$ is an i.i.d.~sample of pairs from $P$ of size $n \geq
\frac{1}{2\gamma^2}\ln \frac{|\cH|}{\delta}$, then with probability at least
$1-\delta$, every $h \in \cH$ that is $(\alpha,\beta)$-pairwise-EF on $S$
(for any $\alpha)$ is also $(\alpha + \gamma, \beta)$-EF on $P$.
%
\end{lem}
\begin{proof}
%
Let $f(x,x',h) = \ind{u(x,h(x)) < u(x,h(x')) - \beta}$ be the indicator that
$x$ is envious of $x'$ by at least $\beta$ under classifier $h$. Then $f(x_i,
x'_i, h)$ is a Bernoulli random variable with success probability
$\expect_{x,x'\sim P}[f(x,x',h)]$. Applying Hoeffding's inequality to any
fixed hypothesis $h \in \cH$ guarantees that $\prob_S(\expect_{x,x'\sim
P}[f(x,x',h)] \geq \frac{1}{n}\sum_{i=1}^n f(x_i, x'_i, h) + \gamma) \leq
\exp(-2n\gamma^2)$. Therefore, if $h$ is $(\alpha,\beta)$-EF on $S$,
then it is also $(\alpha+\gamma, \beta)$-EF on $P$ with probability at least $1 - \exp(-2n\gamma^2)$. Applying the union
bound over all $h \in \cH$ and using the lower bound on $n$ completes the
proof.
%
\end{proof}
Next, we show that $\cH(\cG, m)$ can be covered by a finite subset. Since each
classifier in $\cH$ is determined by the choice of $m$ functions from $\cG$ and
mixing weights $\alpha \in \Delta_m$, we will construct finite covers of $\cG$
and $\Delta_m$. Our covers $\hat \cG$ and $\hat \Delta_m$ will guarantee that
for every $g \in \cG$, there exists $\hat g \in \hat \cG$ such that $\prob_{x
\sim P}(g(x) \neq \hat g(x)) \leq \gamma/m$. Similarly, for any mixing weights
$\alpha \in \Delta_m$, there exists $\hat \alpha \in \Delta_m$ such that
$\norm{\alpha - \hat \alpha}_1 \leq \gamma$. If $h \in \cH(\cG,m)$ is the
mixture of $g_1, \dots, g_m$ with weights $\alpha$, we let $\hat h$ be the
mixture of $\hat g_1, \dots, \hat g_m$ with weights $\hat \alpha$. This
approximation has two sources of error: first, for a random individual $x \sim
P$, there is probability up to $\gamma$ that at least one $g_i(x)$ will disagree
with $\hat g_i(x)$, in which case $h$ and $\hat h$ may assign completely
different outcome distributions. Second, even in the high-probability event that
$g_i(x) = \hat g_i(x)$ for all $i \in [m]$, the mixing weights are not
identical, resulting in a small perturbation of the outcome distribution
assigned to $x$.
\begin{lem} \label{lem:finiteCover}
%
Let $\cG$ be a family of deterministic classifiers with Natarajan dimension $d$,
and let $\cH = \cH(\cG, m)$ for some $m\in\mathbb{N}$. For any $\gamma > 0$, there
exists a subset $\hat \cH \subset \cH$ of size $O\bigl(\frac{(dm|\cY|^2 \log(m
|\cY| / \gamma))^{dm}}{\gamma^{(d+1)m}}\bigr)$ such that for every $h \in \cH$
there exists $\hat h \in \cH$ satisfying:
%
\begin{enumerate}[topsep=0pt,itemsep=0pt,leftmargin=*]
%
\item $\prob_{x \sim P}( \norm{h(x) - \hat h(x)}_1 > \gamma) \leq \gamma$.
%
\item If $S$ is an i.i.d.~sample of individuals of size
$O(\frac{m^2}{\gamma^2}(d \log |\cY| + \log \frac{1}{\delta}))$ then w.p.
$\geq 1-\delta$, we have $\norm{h(x) - \hat h(x)}_1 \leq \gamma$ for all but
a $2\gamma$-fraction of $x \in S$.
%
\end{enumerate}
%
\end{lem}
\begin{proof}
%
As described above, we begin by constructing finite covers of $\Delta_m$ and
$\cG$. First, let $\hat \Delta_m \subset \Delta_m$ be the set of distributions
over $[m]$ where each coordinate is a multiple of $\gamma/m$. Then we have
$|\hat \Delta_m| = O( (\frac{m}{\gamma})^m)$ and for every $p \in \Delta_m$,
there exists $q \in \hat \Delta_m$ such that $\norm{p - q}_1 \leq \gamma$.
In order to find a small cover of $\cG$, we use the fact that it has low
Natarajan dimension. This implies that the number of effective functions in
$\cG$ when restricted to a sample $S'$ grows only polynomially in the size of
$S'$. At the same time, if two functions in $\cG$ agree on a large sample,
they will also agree with high probability on the distribution.
Formally, let $S'$ be an i.i.d.~sample drawn from $P$ of size
$O(\frac{m^2}{\gamma^2}d\log|\cY|)$, and let $\hat \cG = \cG\resto_{S'}$ be any
minimal subset of $\cG$ that realizes all possible labelings of $S'$ by
functions in $\cG$. We now argue that with probability 0.99, for every $g \in
\cG$ there exists $\hat g \in \hat \cG$ such that $\prob_{x \sim P}(g(x) \neq
\hat g(x)) \leq \gamma/m$. For any pair of functions $g,g' \in \cG$, let $(g,g') :
\cX \to \cY^2$ be the function given by $(g,g')(x) = (g(x), g'(x))$, and let
$\cG^2 = \{(g,g') \,:\, g,g' \in \cG\}$. The Natarajan dimension of $\cG^2$ is
at most $2d$ (see Lemma~\ref{lem:G2} in Appendix~\ref{app:mixtures}). Moreover, consider the loss $c : \cG^2 \times \cX \to \{0,1\}$ given by $c(g,g',x) =
\ind{g(x) \neq g'(x)}$. Applying Lemma~\ref{lem:natarajanAgnostic} with the
chosen size of $|S'|$ ensures that with probability at least $0.99$ every pair
$(g,g') \in \cG^2$ satisfies \[\left|\expect_{x \sim P}[c(g,g',x)] -
\frac{1}{|S'|} \sum_{x \in S'} c(g,g',x)\right| \leq \frac{\gamma}{m}.\] By the definition
of $\hat \cG$, for every $g \in \cG$, there exists $\hat g \in \hat \cG$ for
which $c(g,\hat g,x) = 0$ for all $x \in S'$, which implies that $\prob_{x \sim
P}(g(x) \neq \hat g(x)) \leq \gamma/m$.
Using Lemma~\ref{lem:natarajan} to
bound the size of $\hat \cG$, we have that \[|\hat \cG| \leq |S'|^d |\cY|^{2d} =
O\left(\left(\frac{m^2}{\gamma^2}d|\cY|^2\log|\cY|\right)^d\right).\] Since this
construction succeeds with non-zero probability, we are guaranteed that such a
set $\hat \cG$ exists. Finally, by an identical uniform convergence argument,
it follows that if $S$ is a fresh i.i.d.~sample of the size given in Item 2 of the lemma's
statement, then, with probability at least $1-\delta$, every $g$ and $\hat g$ will disagree on at most a
$2\gamma/m$-fraction of $S$, since they disagree with probability at most
$\gamma/m$ on $P$.
Next, let $\hat \cH = \{ h_{\vec g, \alpha} \, : \, \vec g \in \hat G^m, \alpha
\in \hat \Delta_m\}$ be the same family as $\cH$, except restricted to choosing
functions from $\hat \cG$ and mixing weights from $\hat \Delta_m$. Using the
size bounds above and the fact that ${N \choose m} = O((\frac{N}{m})^m)$, we
have that \[|\hat \cH| = {|\hat \cG| \choose m} \cdot |\hat \Delta_m| =
O\left(\frac{(dm^2|\cY|^2 \log(m |\cY| /
\gamma))^{dm}}{\gamma^{(2d+1)m}}\right).\]
Suppose that $h$ is the mixture of $g_1, \dots, g_m \in \cG$ with weights
$\alpha \in \Delta_m$. Let $\hat g_i$ be the approximation to $g_i$ for each
$i$, let $\hat \alpha \in \hat \Delta_m$ be such that $\norm{\alpha - \hat
\alpha}_1 \leq \gamma$, and let $\hat h$ be the random mixture of $\hat g_1,
\dots, \hat g_m$ with weights $\hat \alpha$. For an individual $x$ drawn from
$P$, we have $g_i(x) \neq \hat g_i(x)$ with probability at most $\gamma/m$,
and therefore they all agree with probability at least $1-\gamma$. When this
event occurs, we have $\norm{h(x) - \hat h(x)}_1 \leq \norm{\alpha - \hat
\alpha}_1 \leq \gamma$.
The second part of the claim follows by similar reasoning, using the fact that
for the given sample size $|S|$, with probability at least $1-\delta$, every
$g \in \cG$ disagrees with its approximation $\hat g \in \hat \cG$ on at most
a $2\gamma/m$-fraction of $S$. This means that $\hat g_i(x) = g_i(x)$ for all $i \in
[m]$ on at least a $(1-2\gamma)$-fraction of the individuals $x$ in $S$. For these
individuals, $\norm{h(x) - \hat h(x)}_1 \leq \norm{\alpha - \hat \alpha}_1 \leq \gamma$.
\end{proof}
Combining the generalization guarantee for finite families given in
Lemma~\ref{lem:finiteGeneralize} with the finite approximation given in
Lemma~\ref{lem:finiteCover}, we are able to show that envy-freeness also
generalizes for $\cH(\cG,m)$.
\begin{proof}[Proof of Theorem~\ref{thm:mixtureGeneralize}]
%
Let $\hat \cH$ be the finite approximation to $\cH$ constructed in
Lemma~\ref{lem:finiteCover}. If the sample is of size $|S| =
O(\frac{1}{\gamma^2}(dm \log(dm|\cY|\log|\cY|/\gamma) + \log
\frac{1}{\delta}))$, we can apply Lemma~\ref{lem:finiteGeneralize} to this
finite family, which implies that for any $\beta' \geq 0$, with probability at
least $1-\delta/2$ every $\hat h \in \hat \cH$ that is
$(\alpha',\beta')$-pairwise-EF on $S$ (for any $\alpha'$) is also
$(\alpha'+\gamma, \beta')$-EF on $P$. We apply this lemma with $\beta' = \beta +
2\gamma$. Moreover, from Lemma~\ref{lem:finiteCover}, we know that if $|S| =
O(\frac{m^2}{\gamma^2}(d \log|\cY| + \log \frac{1}{\delta}))$, then with
probability at least $1-\delta/2$, for every $h \in \cH$, there exists $\hat
h \in \hat \cH$ satisfying $\norm{h(x) - \hat h(x)}_1 \leq \gamma$ for all but a
$2\gamma$-fraction of the individuals in $S$. This implies that on all but at
most a $4\gamma$-fraction of the pairs in $S$, $h$ and $\hat h$ satisfy this
inequality for both individuals in the pair. Assume these high probability
events occur. Finally, from Item 1 of the lemma we have that $\prob_{x_1,x_2
\sim P}(\max_{i=1,2}\norm{h(x_i) - \hat h(x_i)}_1 > \gamma) \leq 2\gamma$.
Now let $h \in \cH$ be any classifier that is
$(\alpha,\beta)$-pairwise-EF on $S$. Since the utilities are in $[0,1]$ and
$\max_{x=x_i,x_i'} \norm{h(x) - \hat h(x)}_1 \leq \gamma$ for all but a
$4\gamma$-fraction of the pairs in $S$, we know that $\hat h$ is $(\alpha +
4\gamma, \beta + 2\gamma)$-pairwise-EF on $S$. Applying the envy-freeness
generalization guarantee (Lemma~\ref{lem:finiteGeneralize}) for $\hat \cH$, it follows that $\hat h$ is also
$(\alpha + 5\gamma, \beta + 2\gamma)$-EF on $P$. Finally, using the
fact that \[\prob_{x_1,x_2 \sim P}\left(\max_{i=1,2} \norm{h(x_i) - \hat h(x_i)}_1 >
\gamma\right) \leq 2\gamma,\] it follows that $h$ is $(\alpha + 7\gamma, \beta +
4\gamma)$-EF on $P$.
%
\end{proof}
It is worth noting that the (exponentially large) approximation $\hat \cH$ is
only used in the generalization analysis; importantly, an ERM algorithm need not construct
it.
\section{Discussion}\label{sec:disc}
We believe that envy-freeness gives a new, useful perspective on individual fairness in classification --- when individuals have rich utility functions, which, as we have argued in detail in Section~\ref{sec:intro}, is the case in advertising. However, in some domains there are only two possible outcomes, one of which is `good' and the other 'bad'; examples include predicting whether an individual would default on a loan, and whether an offender would recidivate. In these degenerate cases envy-freeness would require that the classifier assign each and every individual the exact same probability of obtaining the `good' outcome, which, clearly, is not a reasonable constraint.
\edit{It is also worth noting that we have not directly addressed the problem of \emph{computing} the loss-minimizing envy-free classifier from a given family on a given sample of individuals. Just like in the work of \citet{DHPR+12}, when the classifier is arbitrary, this problem can be written as a linear program of polynomial size in the number of outcomes, because envy-freeness amounts to a set of linear constraints. In both settings, though, one needs to restrict the family of classifiers to obtain good sample complexity, and, moreover, the na\"ive formulation would be intractable when dealing with a combinatorial space of outcomes. Nevertheless, the linearity of envy-freeness may enable practical mixed-integer linear programming formulations with respect to certain families. More generally, given the wealth of powerful optimization tools at the community's disposal, we do not view computational complexity as a long-term obstacle to implementing our approach.}
\bibliographystyle{aaai}
|
\section{\@startsection{section}{1}%
\z@{.7\linespacing\@plus\linespacing}{.5\linespacing}%
{\normalfont\scshape}
\patchcmd{\@settitle}{\uppercasenonmath\@title}{\Large\boldmath}{}{}
\patchcmd{\@settitle}{\begin{center}}{\begin{flushleft}}{}{}
\patchcmd{\@settitle}{\end{center}}{\end{flushleft}}{}{}
\patchcmd{\@setauthors}{\MakeUppercase}{\normalsize}{}{}
\patchcmd{\@setauthors}{\centering}{\raggedright}{}{}
\patchcmd{\section}{\scshape}{\large\bfseries\boldmath}{}{}
\patchcmd{\subsection}{\bfseries}{\bfseries\boldmath}{}{}
\renewcommand{\@secnumfont}{\bfseries}
\patchcmd{\@startsection}{\@afterindenttrue}{\@afterindentfalse}{}{}
\patchcmd{\abstract}{\leftmargin3pc}{\leftmargin1pc}{}{}
\def\maketitle{\par
\@topnum\z@
\@setcopyright
\thispagestyle{empty
\ifx\@empty\shortauthors \let\shortauthors\shorttitle
\else \andify\shortauthors
\fi
\@maketitle@hook
\begingroup
\@maketitle
\toks@\@xp{\shortauthors}\@temptokena\@xp{\shorttitle}%
\toks4{\def\\{ \ignorespaces}
\edef\@tempa{%
\@nx\markboth{\the\toks4
\@nx\MakeUppercase{\the\toks@}}{\the\@temptokena}}%
\@tempa
\endgroup
\c@footnote\z@
\@cleartopmattertags
}
\def\maketag@@@#1{\hbox{\m@th\normalfont\normalsize#1}}
\makeatother
\newcommand{\mathscr{P}}{\mathscr{P}}
\newcommand{\mathscr{I}}{\mathscr{I}}
\newcommand{\mathrm{Tail}}{\mathrm{Tail}}
\newcommand{\tilde{P}}{\tilde{P}}
\newcommand{\tilde{p}}{\tilde{p}}
\lstset{
basicstyle=\ttfamily,
xleftmargin=2em,xrightmargin=1em,
breaklines=true
}
\newcommand{\mcode}[1]{
\lstinline[language=Mathematica]|#1|
}
\title{Linked partition ideals and Kanade--Russell conjectures}
\author[S. Chern]{Shane Chern}
\address[S. Chern]{Department of Mathematics, The Pennsylvania State University, University Park, PA 16802, USA}
\email{<EMAIL>}
\author[Z. Li]{Zhitai Li}
\address[Z. Li]{Department of Mathematics, The Pennsylvania State University, University Park, PA 16802, USA}
\email{<EMAIL>}
\date{}
\begin{document}
\maketitle
\begin{abstract}
This paper will primarily present a method of proving generating function identities for partitions from linked partition ideals. The method we introduce is built on a conjecture by George Andrews and that those generating functions satisfy some $q$-difference equations. We will come up with the generating functions of partitions in the Kanade--Russell conjectures to illustrate the effectiveness of this method.
\Keywords{Partitions, Kanade--Russell conjectures, linked partition ideals, $q$-difference equations, computer algebra.}
\MSC{Primary 05A17; Secondary 11P84.}
\end{abstract}
\section{Introduction}
\subsection{Background}
As usual, a \textit{partition} $\lambda$ of a positive integer $n$ is a weakly decreasing sequence of positive integers $\lambda_1\ge \lambda_2\ge\cdots\ge \lambda_{\ell}$ whose sum equals $n$. By convention, we may assume that $0$ has one partition, which is called the empty partition $\emptyset$.
In the theory of partitions, generating function identities are of great interest as they encapsulate considerable information of the partitions in question. In a series of papers \cite{And1972,And1974,And1975} dated back to the 1970s, George Andrews initiated a general theory of partition identities. These papers were later included in Chapter 8 of Andrews' monograph ``\textit{The theory of partitions}'' \cite{And1976}. In particular, Andrews introduced the concept of linked partition ideal. Recently, Andrews further communicated the idea that linked partition ideals deserve some more attention for their generating functions can be elegantly formulated.
The following conjecture by Andrews provides us a basis of ``guessing'' the generating function:
\begin{conjecture}[Andrews]
Every linked partition ideal has a bivariate generating function of the form
\begin{equation}\label{eq:And-conj-0}
\sum_{n_1,\ldots,n_r\ge 0}\frac{(-1)^{L_1(n_1,\ldots,n_r)}q^{Q(n_1,\ldots,n_r)+L_2(n_1,\ldots,n_r)}x^{L_3(n_1,\ldots,n_r)}}{(q^{B_1};q^{A_1})_{n_1}\cdots (q^{B_r};q^{A_r})_{n_r}},
\end{equation}
in which $L_1$, $L_2$ and $L_3$ are linear forms in $n_1,\ldots,n_r$ and $Q$ is a quadratic form in $n_1,\ldots,n_r$. Here the coefficient of the $x^mq^n$ term is the number of partitions of $n$ in this linked partition ideal with exactly $m$ parts.
\end{conjecture}
This conjecture has numerous pieces of empirical evidence:
\begin{enumerate}[label*=\arabic*.,leftmargin=*]
\item The first Rogers--Ramanujan identity (cf.~Corollary 7.6 in \cite{And1976}) states that the number of partitions of a nonnegative integer $n$ into parts congruent to $\pm 1$ modulo $5$ is the same as the number of partitions of $n$ such that each two consecutive parts have difference at least $2$. We know that the generating function of partitions under the above difference-at-a-distance theme is
\[
\sum_{n\ge 0}\frac{q^{n^2}}{(q;q)_n}.
\]
A generalization of the Rogers--Ramanujan identities is due to Gordon (cf.~Theorem 7.5 in \cite{And1976}). In a special case of Gordon's generalization, we encounter partitions of the form $\lambda_1+ \lambda_2+\cdots+ \lambda_{\ell}$, where for all $j$, $\lambda_j - \lambda_{j+k-1}\ge 2$ with $k\ge 2$ fixed. It can be shown that the generating function is
\[
\sum_{n_1,n_2,\ldots,n_{k-1}\ge 0}\frac{q^{N_1^2+N_2^2+\cdots+N_{k-1}^2}}{(q;q)_{n_1}(q;q)_{n_2}\cdots(q;q)_{n_{k-1}}},
\]
where $N_j=n_j+n_{j+1}+\cdots+n_{k-1}$. Andrews showed that this partition set is a linked partition ideal; see \cite[Chapter 8]{And1976}.
\item In the first G\"{o}llnitz--Gordon identity, one studies partitions of the form $\lambda_1+ \lambda_2+\cdots+ \lambda_{\ell}$, in which no odd part is repeated, $\lambda_j-\lambda_{j-1}\ge 2$ if $\lambda_j$ odd and $\lambda_j-\lambda_{j-1}> 2$ if $\lambda_j$ even. It can be shown that the generating function is
\[
\sum_{n_1,n_2,n_3\ge 0}\frac{(-1)^{n_2}q^{n_1^2+n_3^2+2n_1n_2+n_2}}{(q^2;q^2)_{n_1}(q^2;q^2)_{n_2}(q^2;q^2)_{n_3}}.
\]
This partition set is also a linked partition ideal as claimed by Andrews in \cite[Chapter 8]{And1976}.
\end{enumerate}
With the aid of the above conjecture and necessary computer algebra assistance, if we want to find a generating function identity for a linked partition ideal, we are able to single out the promising candidates by running through a number of multi-summations in the above fashion and comparing the series expansions.
\subsection{Kanade--Russell conjectures}
As we have already seen, many linked partition ideals consist of partitions under certain difference-at-a-distance theme.
\begin{definition}[cf.~\cite{KR2015}]
We say a partition $\lambda=\lambda_1+ \lambda_2+\cdots+ \lambda_{\ell}$ satisfies the \textit{difference at least $d$ at distance $k$} condition if, for all $j$, $\lambda_j - \lambda_{j+k}\ge d$.
\end{definition}
In this setting, we may restate the corresponding partition set in the first Rogers--Ramanujan identity as ``the set of partitions with difference at least $2$ at distance $1$''.
In 2014, Kanade and Russell \cite{KR2015} proposed six challenging conjectures, labeled as $I_1$--$I_6$, on partition identities of Rogers--Ramanujan type. For example, the first of their conjectures reads as follows.
\begin{conjecture}[Kanade--Russell Conjecture $I_1$]
The number of partitions of a nonnegative integer $n$ into parts congruent to $1$, $3$, $6$ or $8$ modulo $9$ is the same as the number of partitions of $n$ with difference at least $3$ at distance $2$ such that if two consecutive parts differ by at most $1$, then their sum is divisible by $3$.
\end{conjecture}
It should be remarked that these conjectures are intriguingly related to the representation theory of affine Lie algebra. For a detailed description of the idea behind them, one may refer to Kanade's Ph.D. Thesis \cite{Kan2015}.
On the other hand, in Russell's Ph.D. Thesis \cite{Rus2016}, companions to the Kanade--Russell conjectures $I_4$--$I_6$ were considered. Further, several more conjectures of the same flavor were proposed in \cite{KR2018}. In particular, among these conjectures (including the six conjectures in \cite{KR2015}), there are eleven of them involving the modulus $12$. It is notable that in a very recent paper of Bringmann, Jennings-Shaffer and Mahlburg \cite{BJM2018}, seven of the modulo $12$ conjectures were proved, while the rest were, although not completely proved, simplified to a great extent.
One major difficulty of proving the Kanade--Russell conjectures is that it is not always easy to find generating functions for partitions under certain difference-at-a-distance themes. Fortunately, this problem was settled in two recent papers of Kanade and Russell \cite{KR2018}, and Kur\c{s}ung\"{o}z \cite{Kur2018}, in which different sets of identities (but with some overlap) were demonstrated, respectively. However, their proofs, although different, are both purely combinatorial.
However, if we notice that the partition sets under difference-at-a-distance themes in the six conjectures $I_1$--$I_6$ are either a linked partition ideal or a subset of a linked partition ideal, a more algebraic approach can be provided. The resulting generating functions, in turn, give us more evidence for Andrews' conjecture.
\subsection{Outline of the paper}
In Section \ref{sect:linked-partition-ideal}, we will give a detailed account of linked partition ideals and their generating functions. Among those, the most consequential property is that those generating functions satisfy some $q$-difference equations. Such a $q$-difference equation is obtained by solving a $q$-difference system. To do so, we reformulate in Section \ref{sec:q-diff} an algorithm due to Andrews (cf.~\cite[Lemma 8.10]{And1976}) into the matrix form to make it easier to manipulate in computer algebra systems like \textit{Mathematica}.
In the next three sections, as experiments, we apply our method to not only reprove the six generating function identities involved in the Kanade--Russell conjectures but also present six more new identities. It is notable that our method is also applicable to the cases where the partitions in question are from a nice subset of a linked partition ideal.
We end our paper with several interesting transformation formulas motivated by a recent paper of Bringmann, Jennings-Shaffer and Mahlburg \cite{BJM2018}.
\begin{remark}
The proof of Andrews' conjecture should be regarded as a worthy objective: if we could successfully prove this conjecture, we essentially obtain a universal and robust method of deducing generating functions for partitions from linked partition ideals.
\end{remark}
\section{Linked partition ideals}\label{sect:linked-partition-ideal}
We now give a brief review of linked partition ideals. Note that we shall restate some definitions in \cite[Chapter 8]{And1976}.
Let $\mathscr{P}$ be the set of partitions. Given a partition $\lambda\in\mathscr{P}$, let $|\lambda|$ denote the sum of all parts of $\lambda$, let $\sharp(\lambda)$ denote the number of parts in $\lambda$ and let $\sharp_k(\lambda)$ denote the number of occurrences of parts of size $k$ in $\lambda$. For example, if $\lambda=3+3+2+1+1+1$, then $\sharp(\lambda)=6$, $\sharp_1(\lambda)=3$, $\sharp_2(\lambda)=1$, $\sharp_3(\lambda)=2$ and $\sharp_k(\lambda)=0$ for $k\ge 4$. From the definition of partitions, one can see that only finitely many of the $\sharp_k(\lambda)$ are nonzero.
We define a partial order ``$\le$'' by asserting that, for any two partitions $\lambda$ and $\pi$, $\pi\le \lambda$ if $\sharp_k(\pi)\le \sharp_k(\lambda)$ for all $k$. Andrews also defined the ``meet'' and ``join'' operations for $\lambda$ and $\pi$ by treating $\mathscr{P}$ as a lattice:
\begin{enumerate}
\item $\lambda\cap\pi$ satisfies $\sharp_k(\lambda\cap\pi)=\min(\sharp_k(\lambda), \sharp_k(\pi))$ for all $k$;
\item $\lambda\cup\pi$ satisfies $\sharp_k(\lambda\cup\pi)=\max(\sharp_k(\lambda), \sharp_k(\pi))$ for all $k$.
\end{enumerate}
\begin{definition}
A subset $\mathscr{I}$ of $\mathscr{P}$ is called a \textit{partition ideal} if for any $\lambda$ in $\mathscr{I}$, $\pi$ is also in $\mathscr{I}$ whenever $\pi\le\lambda$.
\end{definition}
\begin{remark}
Andrews further asserted that a partition ideal is indeed a semi-ideal in the notation of lattice theory.
\end{remark}
We next define the modulus of a partition ideal. To do so, we need the following notation.
Let $\mathscr{I}$ be a partition ideal. We define $\mathscr{I}^{(m)}$ by the collection of partitions in $\mathscr{I}$ whose smallest part is $>m$. We also include the empty partition $\emptyset$ in $\mathscr{I}^{(m)}$.
We then define a map $\phi$ by sending a partition $\lambda=\lambda_1+\lambda_2+\cdots+\lambda_\ell$ to $(\lambda_1+1)+(\lambda_2+1)+\cdots+(\lambda_\ell+1)$ and the empty partition to itself.
\begin{definition}
We say that a partition ideal $\mathscr{I}$ has \textit{modulus} $m$ if $m$ is a positive integer such that $\phi^m \mathscr{I}=\mathscr{I}^{(m)}$.
\end{definition}
\begin{remark}
A partition ideal might have more than one modulus.
\end{remark}
For two partitions $\lambda$ and $\pi$ in $\mathscr{I}$, their sum $\lambda\oplus \pi$ is defined by collecting their parts in weakly decreasing order. Lemma 8.9 in \cite{And1976} gives a unique decomposition for each $\lambda\in\mathscr{I}$ if $\mathscr{I}$ has modulus $m$.
\begin{lemma}
Let $\mathscr{I}$ be a partition ideal of modulus $m$. For each $\lambda\in\mathscr{I}$, we uniquely have
$$\lambda=\lambda_{(1)}\oplus(\phi^m\lambda_{(2)})\oplus(\phi^{2m}\lambda_{(3)})\oplus\cdots$$
where $\lambda_{(1)}$, $\lambda_{(2)}$, $\lambda_{(3)}$, $\ldots$ are in $\mathscr{I}$, all satisfying the property that the largest part $\le m$.
\end{lemma}
\begin{definition}
We define, for each partition ideal $\mathscr{I}$ of modulus $m$,
$$L_{\mathscr{I},m}:=\{\lambda\in \mathscr{I}: \text{the largest part of $\lambda\le m$}\}.$$
Here, again, the empty partition is included in $L_{\mathscr{I},m}$. Further, if the value of $m$ is clear from the context, then the $m$ in the subscript could be omitted and we simply write $L_{\mathscr{I}}:=L_{\mathscr{I},m}$.
\end{definition}
\begin{definition}
For any partition $\pi\in\mathscr{P}$, its \textit{$m$-tail} $\mathrm{Tail}_m(\pi)$ is defined to be the collection of parts of $\pi$ which are at most $m$. For example,
$$\mathrm{Tail}_2(3+3+2+1+1+1)=2+1+1+1.$$
\end{definition}
Now we are ready to give the definition of linked partition ideals.
\begin{definition}\label{def:linked}
We say that a partition ideal $\mathscr{I}$ is a \textit{linked partition ideal} if
\begin{enumerate}[label=(\roman*)]
\item $\mathscr{I}$ has a modulus, say $m$;
\item the $L_{\mathscr{I}}$ corresponding to $m$ is a finite set;
\item for each $\pi\in L_{\mathscr{I}}$, there corresponds a minimal subset $\mathcal{L}_{\mathscr{I}}(\pi)\subseteq L_{\mathscr{I}}$ (called the linking set of $\pi$) and a positive integer $l(\pi)$ (called the span of $\pi$) such that for any partition $\lambda$, it belongs to $\mathscr{I}$ with $\mathrm{Tail}_m(\lambda)=\pi$ if and only if we can find a partition $\tilde{\pi}$ with $\mathrm{Tail}_m(\tilde{\pi})\in \mathcal{L}_{\mathscr{I}}(\pi)$ such that
$$\lambda=\pi\oplus \left(\phi^{l(\pi)m} \tilde{\pi}\right).$$
\end{enumerate}
\end{definition}
\begin{example}
Let $\mathscr{D}$ denote the set of partitions into distinct parts. A trivial observation is that $\mathscr{D}$ is a partition ideal of modulus $1$. Further, $L_{\mathscr{D}}=L_{\mathscr{D},1}=\{\emptyset,1\}$. We also have
\begin{gather*}
\mathcal{L}_{\mathscr{D}}(\emptyset)=\{\emptyset,1\},\quad l(\emptyset)=1,\\
\intertext{and}
\mathcal{L}_{\mathscr{D}}(1)=\{\emptyset,1\},\quad l(1)=1.
\end{gather*}
\end{example}
\begin{example}
Let $\mathscr{R}$ denote the set of partitions in which each two consecutive parts have difference at least $2$. We can find that $\mathscr{R}$ is a partition ideal of modulus $2$. Further, $L_{\mathscr{R}}=L_{\mathscr{R},2}=\{\emptyset,1,2\}$. We also have
\begin{gather*}
\mathcal{L}_{\mathscr{R}}(\emptyset)=\{\emptyset,1,2\},\quad l(\emptyset)=1,\\
\mathcal{L}_{\mathscr{R}}(1)=\{\emptyset,1,2\},\quad l(1)=1,\\
\intertext{and}
\mathcal{L}_{\mathscr{R}}(2)=\{\emptyset,2\},\quad l(2)=1.
\end{gather*}
\end{example}
Finally, we consider a bivariate generating function for any subset $\mathscr{S}$ of $\mathscr{P}$:
\begin{equation}
G_{\mathscr{S}}(x)=G_{\mathscr{S}}(x,q):=\sum_{\lambda\in\mathscr{S}}x^{\sharp(\lambda)}q^{|\lambda|}.
\end{equation}
In the setting of Definition \ref{def:linked}, if we further define $\mathscr{I}_{\pi}$ by the set of partitions in $\mathscr{I}$ whose $m$-tail is $\pi\in L_{\mathscr{I}}$, then (8.4.13) in \cite{And1976} tells us that
\begin{equation}\label{eq:important-gf-1}
\sum_{\mu\in \mathscr{I}_{\pi}} x^{\sharp(\mu)} q^{|\mu|}= x^{\sharp(\pi)}q^{|\pi|} \sum_{\varpi\in \mathcal{L}_{\mathscr{I}}(\pi)}\sum_{\nu\in \mathscr{I}_{\varpi}} \left(xq^{l(\pi)m}\right)^{\sharp(\nu)} q^{|\nu|}.
\end{equation}
In other words,
\begin{equation}\label{eq:important-gf-2}
G_{\mathscr{I}_{\pi}}(x)= x^{\sharp(\pi)}q^{|\pi|} \sum_{\varpi\in \mathcal{L}_{\mathscr{I}}(\pi)}G_{\mathscr{I}_{\varpi}}(xq^{l(\pi)m}).
\end{equation}
\section{Systems of $q$-difference equations}\label{sec:q-diff}
As we will see in the next section, a crucial point there can be summarized as the following question:
\begin{question}
Suppose we have a system of $q$-difference equations, say,
\begin{equation}\label{eq:system-gen}
\left\{\!\!\begin{array}{c}
F_1(x)=p_{1,1}(x)F_1(xq^m)+p_{1,2}(x)F_2(xq^m)+\cdots+p_{1,k}(x)F_k(xq^m)\\[0.5em]
F_2(x)=p_{2,1}(x)F_1(xq^m)+p_{2,2}(x)F_2(xq^m)+\cdots+p_{2,k}(x)F_k(xq^m)\\[0.5em]
\vdots\\[0.5em]
F_k(x)=p_{k,1}(x)F_1(xq^m)+p_{k,2}(x)F_2(xq^m)+\cdots+p_{k,k}(x)F_k(xq^m)
\end{array}\right.,
\end{equation}
where the $F$'s and $p$'s are in $x$ and $q$, is it possible to deduce a $q$-difference equation merely involving $F_1$?
\end{question}
Fortunately, an affirmative algorithm is provided by Andrews in the proof of \cite[Lemma 8.10]{And1976}. We would like to translate Andrews' algorithm to the matrix form to make it more transparent.
\medskip
At first, the system \eqref{eq:system-gen} can be written in the matrix form
\begin{equation}\label{eq:step-0}
\begin{pmatrix}
F_1(x)\\
F_2(x)\\
\vdots\\
F_k(x)
\end{pmatrix}
=
\begin{pmatrix}
p_{1,1}(x) & p_{1,2}(x) & \cdots & p_{1,k}(x)\\
p_{2,1}(x) & p_{2,2}(x) & \cdots & p_{2,k}(x)\\
\vdots & \vdots & \ddots & \vdots\\
p_{k,1}(x) & p_{k,2}(x) & \cdots & p_{k,k}(x)\\
\end{pmatrix}
\begin{pmatrix}
F_1(xq^m)\\
F_2(xq^m)\\
\vdots\\
F_k(xq^m)
\end{pmatrix}.
\end{equation}
\medskip
\textbf{Step (1).} We put $u_1(x)=F_1(x)$. Then \eqref{eq:step-0} becomes
\begin{equation}\label{eq:step-1}
\begin{pmatrix}
u_1(x)\\
F_2(x)\\
\vdots\\
F_k(x)
\end{pmatrix}
=
\begin{pmatrix}
p_{1,1}(x) & p_{1,2}(x) & \cdots & p_{1,k}(x)\\
p_{2,1}(x) & p_{2,2}(x) & \cdots & p_{2,k}(x)\\
\vdots & \vdots & \ddots & \vdots\\
p_{k,1}(x) & p_{k,2}(x) & \cdots & p_{k,k}(x)\\
\end{pmatrix}
\begin{pmatrix}
u_1(xq^m)\\
F_2(xq^m)\\
\vdots\\
F_k(xq^m)
\end{pmatrix}.
\end{equation}
If $p_{1,2}(x)=p_{1,3}(x)=\cdots=p_{1,k}(x)=0$, then we shall terminate at this place by noticing that
\[
u_1(x)=p_{1,1}(x)u_1(xq^m).
\]
\medskip
For Steps ($\mathrm{s}$) with $2\le s\le k$, we proceed iteratively as follows.
\medskip
\textbf{Step ($\mathbf{s}$).} Supposing that in the $(s-1)$-th Step, we obtain
\begin{equation}\label{eq:step-i-1}
\begin{pmatrix}
u_1(x)\\
\vdots\\
u_{s-1}(x)\\
F_s(x)\\
\vdots\\
F_k(x)
\end{pmatrix}
=
\tilde{P}_{s-1}
\begin{pmatrix}
u_1(xq^m)\\
\vdots\\
u_{s-1}(xq^m)\\
F_s(xq^m)\\
\vdots\\
F_k(xq^m)
\end{pmatrix},
\end{equation}
where $\tilde{P}_{s-1}$ is a $k\times k$ matrix with the $(i,j)$-th entry being $\tilde{p}_{i,j}(x)$.
Since we have arrived at the $s$th Step, we know that at least one of the $\tilde{p}_{s-1,s}(x)$, $\tilde{p}_{s-1,s+1}(x)$, \ldots, $\tilde{p}_{s-1,k}(x)$ is not identically zero. Otherwise, the program should be terminated at the $(s-1)$-th Step. Further, if $\tilde{p}_{s-1,s}(x)$ is identically zero and $\tilde{p}_{s-1,t}(x)$ (for some $t$ with $s+1\le t\le k$) is not identically zero, \eqref{eq:step-i-1} can be rewritten by swapping $F_s$ and $F_t$. In such a case, $\tilde{P}_{s-1}$ should be rewritten by swapping $\tilde{p}_{s,s}(x)$ and $\tilde{p}_{t,t}(x)$, swapping $\tilde{p}_{s,t}(x)$ and $\tilde{p}_{t,s}(x)$, swapping $\tilde{p}_{i,s}(x)$ and $\tilde{p}_{i,t}(x)$ for $i\ne s,t$, and swapping $\tilde{p}_{s,j}(x)$ and $\tilde{p}_{t,j}(x)$ for $j\ne s,t$. For notational convenience, we simply rename $F_s$ by $F_t$ and $F_t$ by $F_s$ so that the new relation is still of the form \eqref{eq:step-i-1} while $\tilde{p}_{s-1,s}(x)$ is not identically zero.
We then make the following substitution
\begin{equation}
u_{s}(xq^m)=\tilde{p}_{s-1,s}(x)F_s(xq^m)+\tilde{p}_{s-1,s+1}(x)F_{s+1}(xq^m)+\cdots+\tilde{p}_{s-1,k}(x)F_k(xq^m).
\end{equation}
Written in the matrix form, we have
\begin{equation}\label{eq:step-i-trans}
\begin{pmatrix}
u_1(xq^m)\\
u_2(xq^m)\\
\vdots\\
u_{s-1}(xq^m)\\
u_{s}(xq^m)\\
F_{s+1}(xq^m)\\
\vdots\\
F_k(xq^m)
\end{pmatrix}
=
T(x)
\begin{pmatrix}
u_1(xq^m)\\
u_2(xq^m)\\
\vdots\\
u_{s-1}(xq^m)\\
F_s(xq^m)\\
F_{s+1}(xq^m)\\
\vdots\\
F_k(xq^m)
\end{pmatrix},
\end{equation}
where
$$T(x)=\begin{pmatrix}
1 & 0 & \cdots & 0 & 0 & 0 & \cdots & 0\\
0 & 1 & \cdots & 0 & 0 & 0 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots\\
0 & 0 & \cdots & 1 & 0 & 0 & \cdots & 0\\
0 & 0 & \cdots & 0 & \tilde{p}_{s-1,s}(x) & \tilde{p}_{s-1,s+1}(x) & \cdots & \tilde{p}_{s-1,k}(x)\\
0 & 0 & \cdots & 0 & 0 & 1 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots\\
0 & 0 & \cdots & 0 & 0 & 0 & \cdots & 1\\
\end{pmatrix}.$$
Here all diagonal entries in the $k\times k$ matrix $T(x)$ are $1$ except for the $s$th diagonal entry. In the $s$th row of $T(x)$, for $s\le t\le k$, the $(s,t)$-th entry is $\tilde{p}_{s-1,t}(x)$. All remaining entries in $T(x)$ are $0$.
Since $\tilde{p}_{s-1,s}(x)$ is not identically zero, the matrix $T(x)$ is invertible. In particular, we have
$$T(x)^{-1}=
\begin{pmatrix}
1 & 0 & \cdots & 0 & 0 & 0 & \cdots & 0\\
0 & 1 & \cdots & 0 & 0 & 0 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots\\
0 & 0 & \cdots & 1 & 0 & 0 & \cdots & 0\\
0 & 0 & \cdots & 0 & \frac{1}{\tilde{p}_{s-1,s}(x)} & -\frac{\tilde{p}_{s-1,s+1}(x)}{\tilde{p}_{s-1,s}(x)} & \cdots & -\frac{\tilde{p}_{s-1,k}(x)}{\tilde{p}_{s-1,s}(x)}\\
0 & 0 & \cdots & 0 & 0 & 1 & \cdots & 0\\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots\\
0 & 0 & \cdots & 0 & 0 & 0 & \cdots & 1\\
\end{pmatrix}.$$
It follows from \eqref{eq:step-i-1} and \eqref{eq:step-i-trans} that
\begin{equation}\label{eq:step-i-final}
\begin{pmatrix}
u_1(x)\\
\vdots\\
u_{s}(x)\\
F_{s+1}(x)\\
\vdots\\
F_k(x)
\end{pmatrix}
=
\tilde{P}_{s}
\begin{pmatrix}
u_1(xq^m)\\
\vdots\\
u_{s}(xq^m)\\
F_{s+1}(xq^m)\\
\vdots\\
F_k(xq^m)
\end{pmatrix},
\end{equation}
where
$$\tilde{P}_{s}=T(xq^{-m}) \tilde{P}_{s-1} T(x)^{-1}.$$
\begin{claim}\label{claim:q-diff}
The matrix $\tilde{P}_{s}$ obtained above is of the form
\[
\begin{blockarray}{rccccccccc}
& 1 & 2 & 3 & 4 & \cdots & s & s+1 & \cdots & k \\
\begin{block}{r(ccccccccc)}
1 & \star & 1 & 0 & 0 & \cdots & 0 & 0 & \cdots & 0 \\
2 & \star & \star & 1 & 0 & \cdots & 0 & 0 & \cdots & 0 \\
\vdots & \vdots & \vdots &\vdots &\vdots &\ddots &\vdots &\vdots &\ddots&\vdots \\
s-1 & \star & \star & \star & \star & \cdots & 1 & 0 & \cdots & 0 \\
s & \star & \star & \star & \star & \cdots & \star & \star & \cdots & \star \\
\vdots & \vdots & \vdots &\vdots &\vdots &\ddots &\vdots &\vdots &\ddots&\vdots \\
k & \star & \star & \star & \star & \cdots & \star & \star & \cdots & \star \\
\end{block}
\end{blockarray}\ .
\]
More precisely, in row $r$ $(1\le r\le s-1)$ of $\tilde{P}_{s}$, the $(r,r+1)$-th entry is $1$ and the $(r,c)$-th entries are $0$ for all $c>r+1$.
\end{claim}
\begin{proof}
We argue by induction on $s$. When $s=1$, there is nothing to prove. Assuming that the result is true for some $s-1$ and noticing that $\tilde{P}_{s-1}$ is such a matrix obtained in the $(s-1)$-th Step, we know that $\tilde{p}_{r,r+1}(x)=1$ for all $r\le s-2$ and that $\tilde{p}_{r,c}(x)=0$ for all $r\le s-2$ and $c>r+1$.
It is obvious that the first $s-1$ rows of $T(xq^{-m}) \tilde{P}_{s-1}$ are identical to the first $s-1$ rows of $\tilde{P}_{s-1}$. Let the $(j,c)$-th entry of $T(x)^{-1}$ be $T_{j,c}^{(-1)}(x)$.
For $r\le s-1$, the $(r,c)$-th entry of $\tilde{P}_{s}=T(xq^{-m}) \tilde{P}_{s-1} T(x)^{-1}$ is given by
$$\sum_{j=1}^k \tilde{p}_{r,j}(x)T_{j,c}^{(-1)}(x).$$
If $c=r+1$, then the only non-zero contribution in the above summation is
\begin{align*}
\tilde{p}_{r,r+1}(x)T_{r+1,r+1}^{(-1)}(x)&=\begin{cases}
1\cdot 1 & \text{if $r\le s-2$}\\
\tilde{p}_{s-1,s}(x)\cdot \frac{1}{\tilde{p}_{s-1,s}(x)} & \text{if $r=s-1$}
\end{cases}\\[0.5em]
&=1.
\end{align*}
If $c>r+1$, then we first treat the $r=s-1$ case. One has
\begin{align*}
\sum_{j=1}^k \tilde{p}_{s-1,j}(x)T_{j,c}^{(-1)}(x)&=\tilde{p}_{s-1,s}(x)T_{s,c}^{(-1)}(x)+\tilde{p}_{s-1,c}(x)T_{c,c}^{(-1)}(x)\\
&=\tilde{p}_{s-1,s}(x)\cdot \left(-\frac{\tilde{p}_{s-1,c}(x)}{\tilde{p}_{s-1,s}(x)}\right)+\tilde{p}_{s-1,c}(x)\cdot 1\\
&=0.
\end{align*}
For $r\le s-2$, we simply notice that $\tilde{p}_{r,j}(x)=0$ for $j>r+1$ from our assumption and that $T_{j,c}^{(-1)}(x)=0$ for $j\le r+1$ since $j\le s-1$ and $j\ne c$.
\end{proof}
Let $\tilde{p}_{i,j}^{\text{New}}(x)$ be the $(i,j)$-th entry of $\tilde{P}_{s}$. If $\tilde{p}_{s,t}^{\text{New}}(x)=0$ for all $t\ge s+1$, then we shall stop at this place by noticing with the help of Claim \ref{claim:q-diff} that
\begin{align*}
u_1(x)&=\tilde{p}_{1,1}^{\text{New}}(x)u_1(xq^m)+u_2(xq^m),\\[0.5em]
u_2(x)&=\tilde{p}_{2,1}^{\text{New}}(x)u_1(xq^m)+\tilde{p}_{2,2}^{\text{New}}(x)u_2(xq^m)+u_3(xq^m),\\[0.5em]
&\;\;\vdots\\[0.5em]
u_{s-1}(x)&=\tilde{p}_{s-1,1}^{\text{New}}(x)u_1(xq^m)+\tilde{p}_{s-1,2}^{\text{New}}(x)u_2(xq^m)+\cdots+u_{s}(xq^m),\\[0.5em]
u_{s}(x)&=\tilde{p}_{s,1}^{\text{New}}(x)u_1(xq^m)+\tilde{p}_{s,2}^{\text{New}}(x)u_2(xq^m)+\cdots+\tilde{p}_{s,s}^{\text{New}}(x)u_{s}(xq^m).
\end{align*}
\textbf{Final setup.} Assuming that the above program is terminated after $\ell$ ($\le k$) steps, we obtain a new system of $q$-difference equations
{\small\begin{align*}
u_1(x)&=r_{1,1}(x)u_1(xq^m)+u_2(xq^m),\\[0.5em]
u_2(x)&=r_{2,1}(x)u_1(xq^m)+r_{2,2}(x)u_2(xq^m)+u_3(xq^m),\\[0.5em]
&\;\;\vdots\\[0.5em]
u_{\ell-1}(x)&=r_{\ell-1,1}(x)u_1(xq^m)+r_{\ell-1,2}(x)u_2(xq^m)+\cdots+r_{\ell-1,\ell-1}(x)u_{\ell-1}(x)+u_{\ell}(xq^m),\\[0.5em]
u_{\ell}(x)&=r_{\ell,1}(x)u_1(xq^m)+r_{\ell,2}(x)u_2(xq^m)+\cdots+r_{\ell,\ell-1}(x)u_{\ell-1}(xq^m)+r_{\ell,\ell}(x)u_{\ell}(xq^m),
\end{align*}}
where the $r$'s are in $x$ and $q$.
With this new system, a $q$-difference equation involving merely $u_1$ can be obtained by simple eliminations. Finally, we recall that $F_1(x)$ is set to be $u_1(x)$ in Step (1).
\section{Kanade--Russell conjectures}
We may summarize the following four types of partition sets under difference-at-a-distance themes from the Kanade--Russell conjectures.
\smallskip
\noindent\textbullet{~}\;\textsc{Type I}:
Partitions with difference at least $3$ at distance $2$ such that if two consecutive parts differ by at most $1$, then their sum is divisible by $3$.
\smallskip
\noindent\textbullet{~}\;\textsc{Type II}:
Partitions with difference at least $3$ at distance $2$ such that if two consecutive parts differ by at most $1$, then their sum is congruent to $2$ modulo $3$.
\smallskip
\noindent\textbullet{~}\;\textsc{Type III}:
Partitions with difference at least $3$ at distance $3$ such that if parts at distance $2$ differ by at most $1$, then the sum of the two parts and their intermediate part is congruent to $1$ modulo $3$.
\smallskip
\noindent\textbullet{~}\;\textsc{Type IV}:
Partitions with difference at least $3$ at distance $3$ such that if parts at distance $2$ differ by at most $1$, then the sum of the two parts and their intermediate part is congruent to $2$ modulo $3$.
\medskip
In this section, we investigate partition sets of types I, II, III and IV in the setting of linked partition ideals.
\subsection{Partition set of type I}
Recall that the partition set of type I is the set of partitions with difference at least $3$ at distance $2$ such that if two consecutive parts differ by at most $1$, then their sum is divisible by $3$. In other words, if $\lambda=\lambda_1+\lambda_2+\cdots+\lambda_\ell$ is in this partition set, then
\begin{enumerate}[label=(\roman*)]
\item $\lambda_i-\lambda_{i+2}\ge 3$;
\item $\lambda_i-\lambda_{i+1}\le 1$ implies $\lambda_i+\lambda_{i+1}\equiv 0 \pmod{3}$.
\end{enumerate}
Let $\mathscr{P}_{T_{\mathrm{I}}}$ denote the partition set of type I.
\begin{claim}
$\mathscr{P}_{T_{\mathrm{I}}}$ is a partition ideal of modulus $3$.
\end{claim}
\begin{proof}
We first prove that $\mathscr{P}_{T_{\mathrm{I}}}$ is a partition ideal. It suffices to show that for any $\lambda=\lambda_1+\lambda_2+\cdots+\lambda_\ell$ in $\mathscr{P}_{T_{\mathrm{I}}}$, if we delete a part from $\lambda$, the resulting partition $\tilde{\lambda}$ is still in $\mathscr{P}_{T_{\mathrm{I}}}$. Obviously, if the deleted part is $\lambda_1$ or $\lambda_\ell$, then $\tilde{\lambda}\in\mathscr{P}_{T_{\mathrm{I}}}$. Hence, we may assume that $\tilde{\lambda}=\lambda_1+\cdots+\lambda_{k-1}+\lambda_{k+1}+\cdots+\lambda_\ell$ with $\lambda_k$ not being $\lambda_1$ or $\lambda_\ell$. To see that $\tilde{\lambda}$ satisfies the first condition, it suffices to check that $\lambda_{k-2}-\lambda_{k+1}\ge \lambda_{k-1}-\lambda_{k+1}\ge 3$ and $\lambda_{k-1}-\lambda_{k+2}\ge \lambda_{k-1}-\lambda_{k+1}\ge 3$. On the other hand, $\lambda_{k-1}-\lambda_{k+1}\ge 3$ and the fact that $\lambda\in\mathscr{P}_{T_{\mathrm{I}}}$ ensure the second condition. Hence $\tilde{\lambda}\in\mathscr{P}_{T_{\mathrm{I}}}$, as desired. Finally, the fact that $\mathscr{P}_{T_{\mathrm{I}}}$ has modulus $3$ is trivial.
\end{proof}
From the definition of $\mathscr{P}_{T_{\mathrm{I}}}$, it is straightforward to observe the following facts.
\begin{claim}
The $L_{\mathscr{P}_{T_{\mathrm{I}}}}$ corresponding to modulus $3$ equals
$$\{\emptyset,\;1,\;2+1,\;3+1,\;2,\;3,\;3+3\}.$$
\end{claim}
\begin{claim}\label{claim:T1-3}
The span and linking set of partitions in $L_{\mathscr{P}_{T_{\mathrm{I}}}}$ are given as follows.
\begin{equation*}
\begin{array}{lp{0.5cm}cp{0.5cm}l}
&& \text{span} && \quad\quad\quad\quad\text{linking set}\\
\pi_0 = \emptyset && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_1 = 1 && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_2 = 2+1 && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_3 = 3+1 && 1 && \{\pi_0,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_4 = 2 && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_5 = 3 && 1 && \{\pi_0,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_6 = 3+3 && 1 && \{\pi_0,\; \pi_5,\; \pi_6\}
\end{array}
\end{equation*}
\end{claim}
Let us denote by $H_i(x)=H_i(x,q)$ the generating function of partitions $\lambda$ in $\mathscr{P}_{T_{\mathrm{I}}}$ with $\mathrm{Tail}_m(\lambda)=\pi_i$ for $i=0,1,\ldots, 6$ where the $\pi_i$'s are as defined in Claim \ref{claim:T1-3}.
Following \eqref{eq:important-gf-2}, we have
{\footnotesize\begin{align}
H_0(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T1H0}\\
x^{-1}q^{-1}H_1(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T1H1}\\
x^{-2}q^{-3}H_2(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T1H2}\\
x^{-2}q^{-4}H_3(x)&=H_0(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T1H3}\\
x^{-1}q^{-2}H_4(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T1H4}\\
x^{-1}q^{-3}H_5(x)&=H_0(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T1H5}\\
x^{-2}q^{-6}H_6(x)&=H_0(xq^3)+H_5(xq^3)+H_6(xq^3).\notag\label{eq:T1H6}\\
\end{align}}
Let $G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)=G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x,q)$ (resp.~$G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x)$, $G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x)$) denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{I}}}$ whose smallest part is at least $1$ (resp.~$2$, $3$).
It follows that
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)&=H_0(x)+H_1(x)+H_2(x)+H_3(x)+H_4(x)+H_5(x)+H_6(x)\nonumber\\
&=H_0(xq^{-3}),\label{eq:GT11}\\
G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x)&=H_0(x)+H_4(x)+H_5(x)+H_6(x)\nonumber\\
&=x^{-1}H_5(xq^{-3}),\label{eq:GT12}\\
G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x)&=H_0(x)+H_5(x)+H_6(x)\nonumber\\
&=x^{-2}H_6(xq^{-3}).\label{eq:GT13}
\end{align}
Hence, to determine $q$-difference equations satisfied by $G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)$, $G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x)$ and $G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x)$, it suffices to find $q$-difference equations for $H_0(x)$, $H_5(x)$ and $H_6(x)$, respectively.
We now deduce from \eqref{eq:T1H0}, \eqref{eq:T1H1}, \eqref{eq:T1H2} and \eqref{eq:T1H4} that
\begin{align}
H_1(x)&=xqH_0(x),\\
H_2(x)&=x^2 q^3 H_0(x),\\
H_4(x)&=x q^2 H_0(x),\\
\intertext{and likewise from \eqref{eq:T1H3} and \eqref{eq:T1H5} that}
H_3(x)&=xqH_5(x).
\end{align}
As a result, the system \eqref{eq:T1H0}--\eqref{eq:T1H6} can be rewritten as
\begin{framed}
{\small\begin{align}
H_0(x)&=(1+xq^4+x^2q^9+xq^5)H_0(xq^3)+(1+xq^4)H_5(xq^3)+H_6(xq^3),\label{eq:T1H0New}\\
H_5(x)&=(xq^3+x^2q^8)H_0(xq^3)+xq^3H_5(xq^3)+xq^3H_6(xq^3),\label{eq:T1H5New}\\
H_6(x)&=x^2q^6H_0(xq^3)+x^2q^6H_5(xq^3)+x^2q^6H_6(xq^3).\label{eq:T1H6New}
\end{align}}
\end{framed}
We first use the algorithm in Section \ref{sec:q-diff} to deduce the $q$-difference equation satisfied by $H_0(x)$ and accordingly $G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)$.
\medskip
\textbf{Step (1).} We put $u_0(x)=H_0(x)$. Then
\begin{equation}
\begin{pmatrix}
u_0(x)\\
H_5(x)\\
H_6(x)
\end{pmatrix}
=\tilde{P}_1
\begin{pmatrix}
u_0(xq^3)\\
H_5(xq^3)\\
H_6(xq^3)
\end{pmatrix},
\end{equation}
where
\[
\tilde{P}_1=\begin{pmatrix}
1+xq^4+x^2q^9+xq^5 & 1+xq^4 & 1\\
xq^3+x^2q^8 & xq^3 & xq^3\\
x^2q^6 & x^2q^6 & x^2q^6
\end{pmatrix}.
\]
\textbf{Step (2).} We put $u_5(x)=(1+xq^4)H_5(xq^3)+H_6(xq^3)$. Then
\begin{equation}
\begin{pmatrix}
u_0(x)\\
u_5(x)\\
H_6(x)
\end{pmatrix}
=
\tilde{P}_2
\begin{pmatrix}
u_0(xq^3)\\
u_5(xq^3)\\
H_6(xq^3)
\end{pmatrix},
\end{equation}
where
\[
\tilde{P}_2=\begin{pmatrix}
1+xq^4+xq^5+x^2q^9 & 1 & 0\\[0.5em]
xq^3(1+xq+xq^3+xq^5+x^2q^6) & \frac{xq^3(1+xq+xq^3)}{1+xq^4} & \frac{x^2q^7(1+xq+xq^3)}{1+xq^4}\\[0.5em]
x^2q^6 & \frac{x^2q^6}{1+xq^4} & \frac{x^3q^{10}}{1+xq^4}
\end{pmatrix}.
\]
\textbf{Step (3).} We put $u_6(x)=\frac{x^2q^7(1+xq+xq^3)}{1+xq^4}H_6(xq^3)$. Then
\begin{equation}
\begin{pmatrix}
u_0(x)\\
u_5(x)\\
u_6(x)
\end{pmatrix}
=
\tilde{P}_3
\begin{pmatrix}
u_0(xq^3)\\
u_5(xq^3)\\
u_6(xq^3)
\end{pmatrix},
\end{equation}
where
\[
\tilde{P}_3=\begin{pmatrix}
1+xq^4+xq^5+x^2q^9 & 1 & 0\\[0.5em]
xq^3(1+xq+xq^3+xq^5+x^2q^6) & \frac{xq^3(1+xq+xq^3)}{1+xq^4} & 1\\[0.5em]
\frac{x^4q^7(1+x+xq^{-2})}{1+xq} & \frac{x^4q^7(1+x+xq^{-2})}{(1+xq)(1+xq^4)} & \frac{x^3q^4(1+x+xq^{-2})}{(1+xq)(1+xq+xq^3)}
\end{pmatrix}.
\]
\medskip
For convenience, we write
\begin{align}
u_0(x)&=r_{0,0}(x)u_0(xq^3)+u_5(xq^3),\label{eq:T1NewNewNew0}\\
u_5(x)&=r_{5,0}(x)u_0(xq^3)+r_{5,5}(x)u_5(xq^3)+u_6(xq^3),\label{eq:T1NewNewNew5}\\
u_6(x)&=r_{6,0}(x)u_0(xq^3)+r_{6,5}(x)u_5(xq^3)+r_{6,6}u_6(xq^3),\label{eq:T1NewNewNew6}
\end{align}
where the coefficients are rational functions in $x$ and $q$ given by $\tilde{P}_3$.
Noting from \eqref{eq:GT11} that
\begin{align}\label{eq:u0-eli}
G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)=H_0(xq^{-3})=u_0(xq^{-3}),
\end{align}
we may eliminate $u_5(x)$ by \eqref{eq:T1NewNewNew0},
\begin{align}\label{eq:u5-eli}
u_5(x)=G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)-r_{0,0}(xq^{-3}) G_{\mathscr{P}_{T_{\mathrm{I},1}}}(xq^3).
\end{align}
Substituting \eqref{eq:u5-eli} into \eqref{eq:T1NewNewNew5}, we may eliminate $u_6(x)$,
\begin{align}
u_6(x)&=G_{\mathscr{P}_{T_{\mathrm{I},1}}}(xq^{-3})-\left(r_{0,0}(xq^{-6})+r_{5,5}(xq^{-3})\right) G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)\nonumber\\
&\quad+ \left(r_{0,0}(xq^{-3})r_{5,5}(xq^{-3})-r_{5,0}(xq^{-3})\right)G_{\mathscr{P}_{T_{\mathrm{I},1}}}(xq^3).\label{eq:u6-eli}
\end{align}
Substituting \eqref{eq:u0-eli}, \eqref{eq:u5-eli} and \eqref{eq:u6-eli} into \eqref{eq:T1NewNewNew6}, we arrive at, after simplification, the following $q$-difference equation for $G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)$.
\begin{theorem}\label{th:q-diff-T11}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x)+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{I},1}}}(xq^3)&+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{I},1}}}(xq^9)=0,\label{eq:q-diff-T11}
\end{align}
where
\begin{align*}
p_0(x,q)&=1+x(q^4+q^6),\\
p_3(x,q)&=-1-x(q + q^2 + q^3 + q^4 + q^6)-x^2(q^3 + q^4 + q^5 + 2 q^6 + q^7 + q^8 + q^9)\\
&\quad-x^3(q^7 + q^9 + q^{10} + q^{12}),\\
p_6(x,q)&=x^3(q^{11} + q^{13})+x^4(q^{14} + q^{15} + q^{16} + q^{17} + q^{18})+x^5(q^{19} + q^{21}),\\
\intertext{and}
p_9(x,q)&=x^5 q^{27}+x^6(q^{28}+q^{30}).
\end{align*}
\end{theorem}
In the same manner, we may find the $q$-difference equations for $H_5(x)$ and $H_6(x)$, and accordingly $G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x)$ and $G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x)$.
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x)+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{I},2}}}(xq^3)&+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{I},2}}}(xq^9)=0,
\end{align}
where
\begin{align*}
p_0(x,q)&=1+x(q^5+q^8),\\
p_3(x,q)&=-1-x(q^2 + q^3 + q^4 + q^5 + q^8)\nonumber\\
&\quad-x^2(2q^6 + q^7 + q^8 + q^9 + q^{10} + q^{11} + q^{12})-x^3(q^{11} + 2 q^{14} + q^{17}),\\
p_6(x,q)&=x^3(q^{16} + q^{17})+x^4(-q^{17} + q^{18} + q^{19} + q^{21} + q^{22} + q^{23} + q^{24})\\
&\quad+x^5(q^{26} + q^{29}),\\
\intertext{and}
p_9(x,q)&=x^5 q^{33}+x^6(q^{35}+q^{38}).
\end{align*}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x)+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{I},3}}}(xq^3)&+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{I},3}}}(xq^9)=0,
\end{align}
where
\begin{align*}
p_0(x,q)&=1+x(q^6+q^7),\\
p_3(x,q)&=-1-x(q^3 + q^4 + q^5 + q^6 + q^7)\\
&\quad-x^2(q^6 + q^8 + 2q^9 + 2q^{10} + q^{11} + q^{12})-x^3(q^{12} + q^{13} + q^{15} + q^{16}),\\
p_6(x,q)&=x^3(q^{16} + q^{17})+x^4(q^{20} + q^{21} + q^{22} + q^{23} + q^{24})+x^5(q^{27} + q^{28}),\\
\intertext{and}
p_9(x,q)&=x^5 q^{36}+x^6(q^{39}+q^{40}).
\end{align*}
\end{theorem}
\begin{remark}
A simple \textit{Mathematica} program is implemented to compute these recurrences. We are happy to supply our codes upon request.
\end{remark}
\subsection{Partition set of type II}
Recall that the partition set of type II is the set of partitions with difference at least $3$ at distance $2$ such that if two consecutive parts differ by at most $1$, then their sum is congruent to $2$ modulo $3$. In other words, if $\lambda=\lambda_1+\lambda_2+\cdots+\lambda_\ell$ is in this partition set, then
\begin{enumerate}[label=(\roman*)]
\item $\lambda_i-\lambda_{i+2}\ge 3$;
\item $\lambda_i-\lambda_{i+1}\le 1$ implies $\lambda_i+\lambda_{i+1}\equiv 2 \pmod{3}$.
\end{enumerate}
Let $\mathscr{P}_{T_{\mathrm{II}}}$ denote the partition set of type II.
\begin{claim}
$\mathscr{P}_{T_{\mathrm{II}}}$ is a partition ideal of modulus $3$.
\end{claim}
\begin{claim}
The $L_{\mathscr{P}_{T_{\mathrm{II}}}}$ corresponding to modulus $3$ equals
$$\{\emptyset,\;1,\;1+1,\;3+1,\;2,\;3+2,\;3\}.$$
\end{claim}
\begin{claim}\label{claim:T2-3}
The span and linking set of partitions in $L_{\mathscr{P}_{T_{\mathrm{II}}}}$ are given as follows.
\begin{equation*}
\begin{array}{lp{0.5cm}cp{0.5cm}l}
&& \text{span} && \quad\quad\quad\quad\text{linking set}\\
\pi_0 = \emptyset && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_1 = 1 && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_2 = 1+1 && 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_3 = 3+1 && 1 && \{\pi_0,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_4 = 2 && 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_5 = 3+2 && 1 && \{\pi_0,\; \pi_4,\; \pi_5,\; \pi_6\}\\
\pi_6 = 3 && 1 && \{\pi_0,\; \pi_4,\; \pi_5,\; \pi_6\}
\end{array}
\end{equation*}
\end{claim}
Similarly, let us denote by $H_i(x)=H_i(x,q)$ the generating function of partitions $\lambda$ in $\mathscr{P}_{T_{\mathrm{II}}}$ with $\mathrm{Tail}_m(\lambda)=\pi_i$ for $i=0,1,\ldots, 6$ where the $\pi_i$'s are as defined in Claim \ref{claim:T2-3}.
Following \eqref{eq:important-gf-2}, we have
{\footnotesize\begin{align}
H_0(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T2H0}\\
x^{-1}q^{-1}H_1(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T2H1}\\
x^{-2}q^{-2}H_2(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T2H2}\\
x^{-2}q^{-4}H_3(x)&=H_0(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T2H3}\\
x^{-1}q^{-2}H_4(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T2H4}\\
x^{-2}q^{-5}H_5(x)&=H_0(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3),\notag\\
\label{eq:T2H5}\\
x^{-1}q^{-3}H_6(x)&=H_0(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3).\notag\label{eq:T2H6}\\
\end{align}}
Let $G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x)=G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x,q)$ (resp.~$G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x)$) denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{II}}}$ whose smallest part is at least $1$ (resp.~$2$).
Let $G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x)$ denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{II}}}$ where $1$ appears at most once.
It follows that
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x)&=H_0(x)+H_1(x)+H_2(x)+H_3(x)+H_4(x)+H_5(x)+H_6(x)\nonumber\\
&=H_0(xq^{-3}),\label{eq:GT21}\\
G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x)&=H_0(x)+H_4(x)+H_5(x)+H_6(x)\nonumber\\
&=x^{-1}H_6(xq^{-3}),\label{eq:GT22}\\
G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x)&=H_0(x)+H_1(x)+H_3(x)+H_4(x)+H_5(x)+H_6(x)\nonumber\\
&=x^{-1}qH_4(xq^{-3}).\label{eq:GT23}
\end{align}
We may deduce from \eqref{eq:T2H0}, \eqref{eq:T2H1} and \eqref{eq:T2H2} that
\begin{align}
H_1(x)&=xqH_0(x),\\
H_2(x)&=x^2 q^2 H_0(x),\\
\intertext{and likewise from \eqref{eq:T2H3}, \eqref{eq:T2H5} and \eqref{eq:T2H6} that}
H_3(x)&=xqH_6(x),\\
H_5(x)&=xq^2H_6(x).
\end{align}
Hence, the system \eqref{eq:T2H0}--\eqref{eq:T2H6} can be rewritten as
\begin{framed}
{\small\begin{align}
H_0(x)&=(1+xq^4+x^2q^8)H_0(xq^3)+H_4(xq^3)+(1+xq^4+xq^5)H_6(xq^3),\label{eq:T2H0New}\\
H_4(x)&=(x q^2+x^2 q^6)H_0(xq^3)+x q^2 H_4(xq^3)+(x q^2+x^2q^6+x^2 q^7)H_6(xq^3),\label{eq:T2H4New}\\
H_6(x)&=x q^3H_0(xq^3)+x q^3H_4(xq^3)+(x q^3+x^2q^8)H_6(xq^3).\label{eq:T2H6New}
\end{align}}
\end{framed}
Using the algorithm in Section \ref{sec:q-diff}, we are able to prove the following $q$-difference equations for $G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x)$, $G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x)$ and $G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x)$, respectively.
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x)+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{II},1}}}(xq^3)&+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{II},1}}}(xq^9)=0,
\end{align}
where
\begin{align*}
p_0(x,q)&=1+x(q^4+q^5),\\
p_3(x,q)&=-1-x(q + q^2 + q^3 + q^4 + q^5)-x^2(q^2 + q^4 + 2 q^5 + 2 q^6 + q^7 + q^8)\\
&\quad-x^3(q^6 + q^7 + q^9 + q^{10}),\\
p_6(x,q)&=x^3(q^{10} + q^{11})+x^4(q^{12} + q^{13} + q^{14} + q^{15} + q^{16})+x^5(q^{17} + q^{18}),\\
\intertext{and}
p_9(x,q)&=x^5 q^{26}+x^6(q^{27}+q^{28}).
\end{align*}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x)+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{II},2}}}(xq^3)&+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{II},2}}}(xq^9)=0,
\end{align}
where
\begin{align*}
p_0(x,q)&=1+x(q^5+q^7),\\
p_3(x,q)&=-1-x(q^2 + q^3 + q^4 + q^5 + q^7)\\
&\quad-x^2(q^5 + q^6 + q^7 + 2 q^8 + q^9 + q^{10} + q^{11}) - x^3(q^{10} + q^{12} + q^{13} + q^{15}),\\
p_6(x,q)&=x^3(q^{14} + q^{16})+x^4(q^{18} + q^{19} + q^{20} + q^{21} + q^{22})+x^5(q^{24} + q^{26}),\\
\intertext{and}
p_9(x,q)&=x^5 q^{32}+x^6(q^{34}+q^{36}).
\end{align*}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x)+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{II},a}}}(xq^3)&+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{II},a}}}(xq^9)=0,
\end{align}
where
\begin{align*}
p_0(x,q)&=1+x(q^4+q^8),\\
p_3(x,q)&=-1-x(q + q^2 + q^3 + q^4 + q^8)\\
&\quad-x^2(q^4 + 2 q^5 + q^6 + q^8 + q^9 + q^{10} + q^{11})-x^3(q^9 + q^{12} + q^{13} + q^{16}),\\
p_6(x,q)&=x^3(-q^{12} + q^{13} + q^{14} + q^{15})\\
&\quad+x^4(-q^{13} + q^{15} + q^{16} + q^{19} + q^{20} + q^{21} + q^{22})+x^5(q^{23} + q^{27}),\\
\intertext{and}
p_9(x,q)&=x^5 q^{29}+x^6(q^{30}+q^{34}).
\end{align*}
\end{theorem}
\subsection{Partition set of type III}
Recall that the partition set of type III is the set of partitions with difference at least $3$ at distance $3$ such that if parts at distance $2$ differ by at most $1$, then the sum of the two parts and their intermediate part is congruent to $1$ modulo $3$. In other words, if $\lambda=\lambda_1+\lambda_2+\cdots+\lambda_\ell$ is in this partition set, then
\begin{enumerate}[label=(\roman*)]
\item $\lambda_i-\lambda_{i+3}\ge 3$;
\item $\lambda_i-\lambda_{i+2}\le 1$ implies $\lambda_i+\lambda_{i+1}+\lambda_{i+2}\equiv 1 \pmod{3}$.
\end{enumerate}
Let $\mathscr{P}_{T_{\mathrm{III}}}$ denote the partition set of type III.
\begin{claim}
$\mathscr{P}_{T_{\mathrm{III}}}$ is a partition ideal of modulus $3$.
\end{claim}
\begin{claim}
The $L_{\mathscr{P}_{T_{\mathrm{III}}}}$ corresponding to modulus $3$ equals
\begin{align*}
\{&\emptyset,\;1,\;1+1,\;2,\;2+1,\;2+1+1,\;2+2,\;3,\;3+1,\\
&3+1+1,\;3+2,\;3+2+1,\;3+2+2,\;3+3,\;3+3+1\}.
\end{align*}
\end{claim}
\begin{claim}\label{claim:T3-3}
The span and linking set of partitions in $L_{\mathscr{P}_{T_{\mathrm{III}}}}$ are given as follows.
{\footnotesize\begin{equation*}
\begin{array}{lcp{0.05cm}l}
& \text{span} && \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\text{linking set}\\
\pi_0 = \emptyset & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_1 = 1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_2 = 1+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_3 = 2 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_4 = 2+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_5 = 2+1+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_6 = 2+2 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_7 = 3 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_8 = 3+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_9 = 3+1+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_{10} = 3+2 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_{11} = 3+2+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_{12} = 3+2+2 & 1 && \{\pi_0,\; \pi_3,\; \pi_6,\; \pi_7,\; \pi_{10},\; \pi_{12},\; \pi_{13}\}\\
\pi_{13} = 3+3 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{13},\; \pi_{14}\}\\
\pi_{14} = 3+3+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{13},\; \pi_{14}\}\\
\end{array}
\end{equation*}}
\end{claim}
Let us denote by $H_i(x)=H_i(x,q)$ the generating function of partitions $\lambda$ in $\mathscr{P}_{T_{\mathrm{III}}}$ with $\mathrm{Tail}_m(\lambda)=\pi_i$ for $i=0,1,\ldots, 14$ where the $\pi_i$'s are as defined in Claim \ref{claim:T3-3}.
Following \eqref{eq:important-gf-2}, we have
{\footnotesize\begin{align}
H_0(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-1}q^{-1}H_1(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-2}H_2(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-1}q^{-2}H_3(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-3}H_4(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-3}q^{-4}H_5(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-4}H_6(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_6(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-1}q^{-3}H_7(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_6(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-2}q^{-4}H_8(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_6(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-3}q^{-5}H_9(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_6(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-2}q^{-5}H_{10}(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_6(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-3}q^{-6}H_{11}(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_6(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-3}q^{-7}H_{12}(x)&=H_0(xq^3)+H_3(xq^3)+H_6(xq^3)+H_7(xq^3)+H_{10}(xq^3)+H_{12}(xq^3)\notag\\
&\quad+H_{13}(xq^3),\\
x^{-2}q^{-6}H_{13}(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_7(xq^3)+H_8(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-3}q^{-7}H_{14}(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_7(xq^3)+H_8(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{13}(xq^3)+H_{14}(xq^3).
\end{align}}
This system may be simplified as
\begin{framed}
{\begin{align}
H_0(x)&=(1+xq^4+x^2q^8+xq^5+x^2q^9+x^3q^{13})H_0(xq^3)\notag\\
&\quad+(xq^4+1+xq^4+x^2q^8+xq^5+x^2q^9)H_7(xq^3)\notag\\
&\quad+H_{12}(xq^3)+(1+xq^4)H_{13}(xq^3),\\
H_7(x)&=(xq^3+x^2q^7+x^2q^8+x^3q^{12})H_0(xq^3)\notag\\
&\quad+(x^2q^7+xq^3+x^2q^7+x^2q^8+x^3q^{12})H_7(xq^3)\notag\\
&\quad+xq^3H_{12}(xq^3)+(xq^3+x^2q^7)H_{13}(xq^3),\\
H_{12}(x)&=(x^3q^7+x^4q^{12})H_0(xq^3)+(x^4q^{11}+x^3q^7+x^4q^{12})H_7(xq^3)\notag\\
&\quad+x^3q^7H_{12}(xq^3)+x^3q^7H_{13}(xq^3),\\
H_{13}(x)&=(x^2q^6+x^3q^{10}+x^3q^{11})H_0(xq^3)+(x^2q^6+x^3q^{10}+x^3q^{11})H_7(xq^3)\notag\\
&\quad+(x^2q^6+x^3q^{10})H_{13}(xq^3).
\end{align}}
\end{framed}
Let $G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x)=G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x,q)$ (resp.~$G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x)$) denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{III}}}$ whose smallest part is at least $1$ (resp.~$2$).
Let $G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x)$ denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{III}}}$ where $1$ appears at most once.
It follows that
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x)&=H_0(x)+H_1(x)+H_2(x)+H_3(x)+H_4(x)+H_5(x)\notag\\
&\quad+H_6(x)+H_7(x)+H_8(x)+H_9(x)+H_{10}(x)\notag\\
&\quad+H_{11}(x)+H_{12}(x)+H_{13}(x)+H_{14}(x)\notag\\
&=H_0(xq^{-3}),\\
G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x)&=H_0(x)+H_3(x)+H_6(x)+H_7(x)+H_{10}(x)+H_{12}(x)\notag\\
&\quad+H_{13}(x)\notag\\
&=x^{-3}q^2H_{12}(xq^{-3}),\\
G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x)&=H_0(x)+H_1(x)+H_3(x)+H_4(x)+H_6(x)+H_7(x)\notag\\
&\quad+H_8(x)+H_{10}(x)+H_{11}(x)+H_{12}(x)+H_{13}(x)\notag\\
&\quad+H_{14}(x)\notag\\
&=x^{-1}H_7(xq^{-3}).
\end{align}
Analogously, we can use the algorithm in Section \ref{sec:q-diff} to deduce the following $q$-difference equations for $G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x)$, $G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x)$ and $G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x)$, respectively.
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x)&+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{III},1}}}(xq^3)+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{III},1}}}(xq^9)+p_{12}(x,q)G_{\mathscr{P}_{T_{\mathrm{III},1}}}(xq^{12})=0,
\end{align}
where
{\footnotesize\begin{align*}
p_0(x,q)&=1+x(q^{4}+q^{5}+2 q^{7}+q^{9}+q^{10})\\
&\quad+ x^2(q^{9}+2 q^{11}+q^{12}+q^{13}+2 q^{14}+q^{15}+q^{16}+2 q^{17}+q^{19})\\
&\quad+ x^3(q^{16}+q^{18}+q^{19}+2 q^{21}+q^{23}+q^{24}+q^{26}),\\
p_3(x,q)&=-1-x(q+q^{2}+q^{3}+q^{4}+q^{5}+2 q^{7}+q^{9}+q^{10})\\
&\quad
-x^2(q^{2}+q^{3}+2 q^{4}+2 q^{5}+3 q^{6}+2 q^{7}+3 q^{8}+3 q^{9}+3 q^{10}+4 q^{11}+3 q^{12}+2 q^{13}\\
&\quad\quad\quad\quad+2 q^{14}+q^{15}+q^{16}+2 q^{17}+q^{19})\\
&\quad
- x^3(q^{4}+q^{5}+2 q^{6}+4 q^{7}+3 q^{8}+5 q^{9}+5 q^{10}+6 q^{11}+7 q^{12}+8 q^{13}+7 q^{14}+6 q^{15}\\
&\quad\quad\quad\quad+6 q^{16}+4 q^{17}+5 q^{18}+4 q^{19}+3 q^{20}+3 q^{21}+q^{22}+q^{23}+q^{24}+q^{26})\\
&\quad
-x^4(q^{8}+2 q^{9}+2 q^{10}+5 q^{11}+4 q^{12}+6 q^{13}+10 q^{14}+8 q^{15}+10 q^{16}+11 q^{17}+8 q^{18}\\
&\quad\quad\quad\quad+10 q^{19}+9 q^{20}+8 q^{21}+7 q^{22}+6 q^{23}+4 q^{24}+3 q^{25}+2 q^{26}+2 q^{27}+q^{28}\\
&\quad\quad\quad\quad+q^{29})\\
&\quad
-x^5(q^{13}+3 q^{15}+3 q^{16}+4 q^{17}+7 q^{18}+6 q^{19}+7 q^{20}+10 q^{21}+7 q^{22}+9 q^{23}+9 q^{24}\\
&\quad\quad\quad\quad+6 q^{25}+7 q^{26}+5 q^{27}+4 q^{28}+3 q^{29}+3 q^{30}+q^{31}+q^{32})\\
&\quad
- x^6(q^{20}+2 q^{22}+2 q^{23}+q^{24}+4 q^{25}+2 q^{26}+3 q^{27}+4 q^{28}+2 q^{29}+3 q^{30}+3 q^{31}\\
&\quad\quad\quad\quad+q^{32}+2 q^{33}+q^{34}+q^{36}),\\
p_6(x,q)&=x^4(q^{12}+q^{14}+2 q^{16}+q^{18}+q^{20})\\
&\quad
+x^5(q^{13}+2 q^{15}+q^{16}+4 q^{17}+3 q^{18}+4 q^{19}+4 q^{20}+5 q^{21}+5 q^{22}+4 q^{23}+4 q^{24}\\
&\quad\quad\quad\quad+3 q^{25}+4 q^{26}+q^{27}+2 q^{28}+q^{30})\\
&\quad
+x^6(q^{17}+2 q^{18}+3 q^{19}+5 q^{20}+5 q^{21}+9 q^{22}+9 q^{23}+10 q^{24}+12 q^{25}+12 q^{26}\\
&\quad\quad\quad\quad+14 q^{27}+12 q^{28}+12 q^{29}+10 q^{30}+9 q^{31}+9 q^{32}+5 q^{33}+5 q^{34}+3 q^{35}\\
&\quad\quad\quad\quad+2 q^{36}+q^{37})\\
&\quad
+x^7(q^{22}+3 q^{23}+4 q^{24}+6 q^{25}+7 q^{26}+12 q^{27}+12 q^{28}+16 q^{29}+18 q^{30}+16 q^{31}\\
&\quad\quad\quad\quad+19 q^{32}+19 q^{33}+16 q^{34}+18 q^{35}+16 q^{36}+12 q^{37}+12 q^{38}+7 q^{39}+6 q^{40}\\
&\quad\quad\quad\quad+4 q^{41}+3 q^{42}+q^{43})\\
&\quad
+x^8(q^{28}+2 q^{29}+3 q^{30}+6 q^{31}+6 q^{32}+9 q^{33}+11 q^{34}+11 q^{35}+13 q^{36}+16 q^{37}\\
&\quad\quad\quad\quad+12 q^{38}+16 q^{39}+13 q^{40}+11 q^{41}+11 q^{42}+9 q^{43}+6 q^{44}+6 q^{45}+3 q^{46}\\
&\quad\quad\quad\quad+2 q^{47}+q^{48})\\
&\quad
+x^9(q^{35}+q^{36}+q^{37}+3 q^{38}+2 q^{39}+3 q^{40}+5 q^{41}+3 q^{42}+5 q^{43}+5 q^{44}+3 q^{45}\\
&\quad\quad\quad\quad+5 q^{46}+3 q^{47}+2 q^{48}+3 q^{49}+q^{50}+q^{51}+q^{52}),\\
p_{9}(x,q)&=-x^6q^{30}
-x^7(q^{31}+q^{32}+2 q^{34}+q^{36}+q^{37}+q^{38}+q^{39}+q^{40})\\
&\quad
-x^8(q^{33}+2 q^{35}+q^{36}+q^{37}+2 q^{38}+2 q^{39}+3 q^{40}+4 q^{41}+3 q^{42}+3 q^{43}+3 q^{44}\\
&\quad\quad\quad\quad+2 q^{45}+3 q^{46}+2 q^{47}+2 q^{48}+q^{49}+q^{50})\\
&\quad
-x^9(q^{37}+q^{39}+q^{40}+q^{41}+3 q^{42}+3 q^{43}+4 q^{44}+5 q^{45}+4 q^{46}+6 q^{47}+6 q^{48}\\
&\quad\quad\quad\quad+7 q^{49}+8 q^{50}+7 q^{51}+6 q^{52}+5 q^{53}+5 q^{54}+3 q^{55}+4 q^{56}+2 q^{57}+q^{58}\\
&\quad\quad\quad\quad+q^{59})\\
&\quad
-x^{10}(q^{45}+q^{46}+2 q^{47}+2 q^{48}+3 q^{49}+4 q^{50}+6 q^{51}+7 q^{52}+8 q^{53}+9 q^{54}+10 q^{55}\\
&\quad\quad\quad\quad+8 q^{56}+11 q^{57}+10 q^{58}+8 q^{59}+10 q^{60}+6 q^{61}+4 q^{62}+5 q^{63}+2 q^{64}+2 q^{65}\\
&\quad\quad\quad\quad+q^{66})\\
&\quad
-x^{11}(q^{53}+q^{54}+3 q^{55}+3 q^{56}+4 q^{57}+5 q^{58}+7 q^{59}+6 q^{60}+9 q^{61}+9 q^{62}+7 q^{63}\\
&\quad\quad\quad\quad+10 q^{64}+7 q^{65}+6 q^{66}+7 q^{67}+4 q^{68}+3 q^{69}+3 q^{70}+q^{72})\\
&\quad
-x^{12}(q^{60}+q^{62}+2 q^{63}+q^{64}+3 q^{65}+3 q^{66}+2 q^{67}+4 q^{68}+3 q^{69}+2 q^{70}+4 q^{71}\\
&\quad\quad\quad\quad+q^{72}+2 q^{73}+2 q^{74}+q^{76}),\\
\intertext{and}
p_{12}(x,q)&=x^{12}q^{90} + x^{13}(q^{91}+q^{92}+2 q^{94}+q^{96}+q^{97})\\
&\quad+ x^{14}(q^{93}+2 q^{95}+q^{96}+q^{97}+2 q^{98}+q^{99}+q^{100}+2 q^{101}+q^{103})\\
&\quad+x^{15}(q^{97}+q^{99}+q^{100}+2 q^{102}+q^{104}+q^{105}+q^{107}).
\end{align*}}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x)&+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{III},2}}}(xq^3)+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{III},2}}}(xq^9)+p_{12}(x,q)G_{\mathscr{P}_{T_{\mathrm{III},2}}}(xq^{12})=0,
\end{align}
where
{\footnotesize\begin{align*}
p_0(x,q)&=1+x(q^{5}+q^{6}+2 q^{8}+q^{10}+q^{11})\\
&\quad+ x^2(q^{11}+2 q^{13}+q^{14}+q^{15}+2 q^{16}+q^{17}+q^{18}+2 q^{19}+q^{21})\\
&\quad+ x^3(q^{19}+q^{21}+q^{22}+2 q^{24}+q^{26}+q^{27}+q^{29}),\\
p_3(x,q)&=-1-x(q^{2}+q^{3}+q^{4}+q^{5}+q^{6}+2 q^{8}+q^{10}+q^{11})\\
&\quad
-x^2(q^{4}+q^{5}+2 q^{6}+2 q^{7}+3 q^{8}+2 q^{9}+3 q^{10}+3 q^{11}+3 q^{12}+4 q^{13}+3 q^{14}+2 q^{15}\\
&\quad\quad\quad\quad+2 q^{16}+q^{17}+q^{18}+2 q^{19}+q^{21})\\
&\quad
- x^3(q^{7}+q^{8}+2 q^{9}+4 q^{10}+3 q^{11}+5 q^{12}+5 q^{13}+6 q^{14}+7 q^{15}+8 q^{16}+7 q^{17}+6 q^{18}\\
&\quad\quad\quad\quad+6 q^{19}+4 q^{20}+5 q^{21}+4 q^{22}+3 q^{23}+3 q^{24}+q^{25}+q^{26}+q^{27}+q^{29})\\
&\quad
-x^4(q^{12}+2 q^{13}+2 q^{14}+5 q^{15}+4 q^{16}+6 q^{17}+10 q^{18}+8 q^{19}+10 q^{20}+11 q^{21}+8 q^{22}\\
&\quad\quad\quad\quad+10 q^{23}+9 q^{24}+8 q^{25}+7 q^{26}+6 q^{27}+4 q^{28}+3 q^{29}+2 q^{30}+2 q^{31}+q^{32}+q^{33})\\
&\quad
-x^5(q^{18}+3 q^{20}+3 q^{21}+4 q^{22}+7 q^{23}+6 q^{24}+7 q^{25}+10 q^{26}+7 q^{27}+9 q^{28}+9 q^{29}\\
&\quad\quad\quad\quad+6 q^{30}+7 q^{31}+5 q^{32}+4 q^{33}+3 q^{34}+3 q^{35}+q^{36}+q^{37})\\
&\quad
- x^6(q^{26}+2 q^{28}+2 q^{29}+q^{30}+4 q^{31}+2 q^{32}+3 q^{33}+4 q^{34}+2 q^{35}+3 q^{36}+3 q^{37}\\
&\quad\quad\quad\quad+q^{38}+2 q^{39}+q^{40}+q^{42}),\\
p_6(x,q)&=x^4(q^{16}+q^{18}+2 q^{20}+q^{22}+q^{24})\\
&\quad
+x^5(q^{18}+2 q^{20}+q^{21}+4 q^{22}+3 q^{23}+4 q^{24}+4 q^{25}+5 q^{26}+5 q^{27}+4 q^{28}+4 q^{29}\\
&\quad\quad\quad\quad+3 q^{30}+4 q^{31}+q^{32}+2 q^{33}+q^{35})\\
&\quad
+x^6(q^{23}+2 q^{24}+3 q^{25}+5 q^{26}+5 q^{27}+9 q^{28}+9 q^{29}+10 q^{30}+12 q^{31}+12 q^{32}\\
&\quad\quad\quad\quad+14 q^{33}+12 q^{34}+12 q^{35}+10 q^{36}+9 q^{37}+9 q^{38}+5 q^{39}+5 q^{40}+3 q^{41}\\
&\quad\quad\quad\quad+2 q^{42}+q^{43})\\
&\quad
+x^7(q^{29}+3 q^{30}+4 q^{31}+6 q^{32}+7 q^{33}+12 q^{34}+12 q^{35}+16 q^{36}+18 q^{37}+16 q^{38}\\
&\quad\quad\quad\quad+19 q^{39}+19 q^{40}+16 q^{41}+18 q^{42}+16 q^{43}+12 q^{44}+12 q^{45}+7 q^{46}+6 q^{47}\\
&\quad\quad\quad\quad+4 q^{48}+3 q^{49}+q^{50})\\
&\quad
+x^8(q^{36}+2 q^{37}+3 q^{38}+6 q^{39}+6 q^{40}+9 q^{41}+11 q^{42}+11 q^{43}+13 q^{44}+16 q^{45}\\
&\quad\quad\quad\quad+12 q^{46}+16 q^{47}+13 q^{48}+11 q^{49}+11 q^{50}+9 q^{51}+6 q^{52}+6 q^{53}+3 q^{54}\\
&\quad\quad\quad\quad+2 q^{55}+q^{56})\\
&\quad
+x^9(q^{44}+q^{45}+q^{46}+3 q^{47}+2 q^{48}+3 q^{49}+5 q^{50}+3 q^{51}+5 q^{52}+5 q^{53}+3 q^{54}\\
&\quad\quad\quad\quad+5 q^{55}+3 q^{56}+2 q^{57}+3 q^{58}+q^{59}+q^{60}+q^{61}),\\
p_{9}(x,q)&=-x^6q^{36}
-x^7(q^{38}+q^{39}+2 q^{41}+q^{43}+q^{44}+q^{45}+q^{46}+q^{47})\\
&\quad
-x^8(q^{41}+2 q^{43}+q^{44}+q^{45}+2 q^{46}+2 q^{47}+3 q^{48}+4 q^{49}+3 q^{50}+3 q^{51}+3 q^{52}\\
&\quad\quad\quad\quad+2 q^{53}+3 q^{54}+2 q^{55}+2 q^{56}+q^{57}+q^{58})\\
&\quad
-x^9(q^{46}+q^{48}+q^{49}+q^{50}+3 q^{51}+3 q^{52}+4 q^{53}+5 q^{54}+4 q^{55}+6 q^{56}+6 q^{57}\\
&\quad\quad\quad\quad+7 q^{58}+8 q^{59}+7 q^{60}+6 q^{61}+5 q^{62}+5 q^{63}+3 q^{64}+4 q^{65}+2 q^{66}+q^{67}\\
&\quad\quad\quad\quad+q^{68})\\
&\quad
-x^{10}(q^{55}+q^{56}+2 q^{57}+2 q^{58}+3 q^{59}+4 q^{60}+6 q^{61}+7 q^{62}+8 q^{63}+9 q^{64}+10 q^{65}\\
&\quad\quad\quad\quad+8 q^{66}+11 q^{67}+10 q^{68}+8 q^{69}+10 q^{70}+6 q^{71}+4 q^{72}+5 q^{73}+2 q^{74}+2 q^{75}\\
&\quad\quad\quad\quad+q^{76})\\
&\quad
-x^{11}(q^{64}+q^{65}+3 q^{66}+3 q^{67}+4 q^{68}+5 q^{69}+7 q^{70}+6 q^{71}+9 q^{72}+9 q^{73}+7 q^{74}\\
&\quad\quad\quad\quad+10 q^{75}+7 q^{76}+6 q^{77}+7 q^{78}+4 q^{79}+3 q^{80}+3 q^{81}+q^{83})\\
&\quad
-x^{12}(q^{72}+q^{74}+2 q^{75}+q^{76}+3 q^{77}+3 q^{78}+2 q^{79}+4 q^{80}+3 q^{81}+2 q^{82}+4 q^{83}\\
&\quad\quad\quad\quad+q^{84}+2 q^{85}+2 q^{86}+q^{88}),\\
\intertext{and}
p_{12}(x,q)&=x^{12}q^{102} + x^{13}(q^{104}+q^{105}+2 q^{107}+q^{109}+q^{110})\\
&\quad+ x^{14}(q^{107}+2 q^{109}+q^{110}+q^{111}+2 q^{112}+q^{113}+q^{114}+2 q^{115}+q^{117})\\
&\quad+x^{15}(q^{112}+q^{114}+q^{115}+2 q^{117}+q^{119}+q^{120}+q^{122}).
\end{align*}}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x)&+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{III},a}}}(xq^3)+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{III},a}}}(xq^9)+p_{12}(x,q)G_{\mathscr{P}_{T_{\mathrm{III},a}}}(xq^{12})=0,
\end{align}
where
{\footnotesize\begin{align*}
p_0(x,q)&=1+x(q^4+q^5+q^7+q^8+q^{10}+q^{11})\\
&\quad+ x^2(q^9+q^{11}+2 q^{12}+q^{14}+2 q^{15}+q^{16}+2 q^{18}+q^{19}+q^{21})\\
&\quad+ x^3(q^{16}+q^{19}+q^{20}+q^{22}+q^{23}+q^{25}+q^{26}+q^{29}),\\
p_3(x,q)&=-1-x(q+q^2+q^3+q^4+q^5+q^7+q^8+q^{10}+q^{11})\\
&\quad
-x^2(q^{3}+2 q^{4}+2 q^{5}+3 q^{6}+2 q^{7}+3 q^{8}+3 q^{9}+2 q^{10}+3 q^{11}+4 q^{12}+2 q^{13}+2 q^{14}\\
&\quad\quad\quad\quad+2 q^{15}+q^{16}+2 q^{18}+q^{19}+q^{21})\\
&\quad
- x^3(q^{6}+3 q^{7}+3 q^{8}+4 q^{9}+5 q^{10}+5 q^{11}+5 q^{12}+7 q^{13}+7 q^{14}+7 q^{15}+7 q^{16}+5 q^{17}\\
&\quad\quad\quad\quad+4 q^{18}+5 q^{19}+4 q^{20}+3 q^{21}+3 q^{22}+2 q^{23}+q^{24}+q^{25}+q^{26}+q^{29})\\
&\quad
-x^4(q^{10}+3 q^{11}+3 q^{12}+3 q^{13}+6 q^{14}+7 q^{15}+7 q^{16}+10 q^{17}+10 q^{18}+8 q^{19}+9 q^{20}\\
&\quad\quad\quad\quad+9 q^{21}+8 q^{22}+8 q^{23}+7 q^{24}+5 q^{25}+5 q^{26}+4 q^{27}+2 q^{28}+2 q^{29}+q^{30}\\
&\quad\quad\quad\quad+q^{31}+q^{32})\\
&\quad
-x^5(q^{15}+q^{16}+q^{17}+4 q^{18}+5 q^{19}+3 q^{20}+6 q^{21}+8 q^{22}+5 q^{23}+7 q^{24}+10 q^{25}\\
&\quad\quad\quad\quad+6 q^{26}+6 q^{27}+9 q^{28}+5 q^{29}+4 q^{30}+5 q^{31}+3 q^{32}+2 q^{33}+3 q^{34}+q^{35}\\
&\quad\quad\quad\quad+q^{37})\\
&\quad
- x^6(q^{22}+q^{23}+q^{25}+3 q^{26}+q^{27}+q^{28}+4 q^{29}+2 q^{30}+q^{31}+4 q^{32}+3 q^{33}+3 q^{35}\\
&\quad\quad\quad\quad+3 q^{36}+q^{38}+2 q^{39}+q^{42}),\\
p_6(x,q)&=x^4(q^{16}+q^{18}+q^{19}+q^{20}+q^{21}+q^{23})\\
&\quad
+x^5(q^{17}+q^{19}+3 q^{20}+3 q^{21}+3 q^{22}+4 q^{23}+5 q^{24}+4 q^{25}+4 q^{26}+5 q^{27}+4 q^{28}\\
&\quad\quad\quad\quad+3 q^{29}+3 q^{30}+3 q^{31}+q^{32}+q^{34})\\
&\quad
+x^6(q^{21}+2 q^{22}+2 q^{23}+4 q^{24}+6 q^{25}+8 q^{26}+8 q^{27}+10 q^{28}+10 q^{29}+11 q^{30}\\
&\quad\quad\quad\quad+13 q^{31}+13 q^{32}+11 q^{33}+10 q^{34}+10 q^{35}+8 q^{36}+8 q^{37}+6 q^{38}\\
&\quad\quad\quad\quad+4 q^{39}+2 q^{40}+2 q^{41}+q^{42})\\
&\quad
+x^7(q^{26}+3 q^{27}+3 q^{28}+4 q^{29}+7 q^{30}+8 q^{31}+11 q^{32}+14 q^{33}+14 q^{34}+15 q^{35}\\
&\quad\quad\quad\quad+17 q^{36}+17 q^{37}+17 q^{38}+17 q^{39}+15 q^{40}+14 q^{41}+14 q^{42}+11 q^{43}\\
&\quad\quad\quad\quad+8 q^{44}+7 q^{45}+4 q^{46}+3 q^{47}+3 q^{48}+q^{49})\\
&\quad
+x^8(q^{31}+q^{33}+4 q^{34}+3 q^{35}+4 q^{36}+8 q^{37}+8 q^{38}+7 q^{39}+12 q^{40}+12 q^{41}+10 q^{42}\\
&\quad\quad\quad\quad+14 q^{43}+14 q^{44}+10 q^{45}+12 q^{46}+12 q^{47}+7 q^{48}+8 q^{49}+8 q^{50}+4 q^{51}\\
&\quad\quad\quad\quad+3 q^{52}+4 q^{53}+q^{54}+q^{56})\\
&\quad
+x^9(q^{38}+2 q^{41}+2 q^{42}+3 q^{44}+4 q^{45}+q^{46}+3 q^{47}+6 q^{48}+2 q^{49}+2 q^{50}+6 q^{51}\\
&\quad\quad\quad\quad+3 q^{52}+q^{53}+4 q^{54}+3 q^{55}+2 q^{57}+2 q^{58}+q^{61}),\\
p_{9}(x,q)&=-x^6q^{36}
-x^7(q^{37}+q^{38}+q^{40}+q^{41}+q^{43}+q^{44}+q^{45}+q^{46}+q^{47})\\
&\quad
-x^8(q^{39}+q^{41}+2 q^{42}+q^{44}+2 q^{45}+2 q^{46}+2 q^{47}+4 q^{48}+3 q^{49}+2 q^{50}+3 q^{51}\\
&\quad\quad\quad\quad+3 q^{52}+2 q^{53}+3 q^{54}+2 q^{55}+2 q^{56}+q^{57})\\
&\quad
-x^9(q^{43}+q^{46}+q^{47}+q^{48}+2 q^{49}+3 q^{50}+3 q^{51}+4 q^{52}+5 q^{53}+4 q^{54}+5 q^{55}\\
&\quad\quad\quad\quad+7 q^{56}+7 q^{57}+7 q^{58}+7 q^{59}+5 q^{60}+5 q^{61}+5 q^{62}+4 q^{63}+3 q^{64}\\
&\quad\quad\quad\quad+3 q^{65}+q^{66})\\
&\quad
-x^{10}(q^{52}+q^{53}+q^{54}+2 q^{55}+2 q^{56}+4 q^{57}+5 q^{58}+5 q^{59}+7 q^{60}+8 q^{61}+8 q^{62}\\
&\quad\quad\quad\quad+9 q^{63}+9 q^{64}+8 q^{65}+10 q^{66}+10 q^{67}+7 q^{68}+7 q^{69}+6 q^{70}+3 q^{71}\\
&\quad\quad\quad\quad+3 q^{72}+3 q^{73}+q^{74})\\
&\quad
-x^{11}(q^{59}+q^{61}+3 q^{62}+2 q^{63}+3 q^{64}+5 q^{65}+4 q^{66}+5 q^{67}+9 q^{68}+6 q^{69}+6 q^{70}\\
&\quad\quad\quad\quad+10 q^{71}+7 q^{72}+5 q^{73}+8 q^{74}+6 q^{75}+3 q^{76}+5 q^{77}+4 q^{78}+q^{79}+q^{80}\\
&\quad\quad\quad\quad+q^{81})\\
&\quad
-x^{12}(q^{66}+2 q^{69}+q^{70}+3 q^{72}+3 q^{73}+3 q^{75}+4 q^{76}+q^{77}+2 q^{78}+4 q^{79}+q^{80}\\
&\quad\quad\quad\quad+q^{81}+3 q^{82}+q^{83}+q^{85}+q^{86}),\\
\intertext{and}
p_{12}(x,q)&=x^{12}q^{99} + x^{13}(q^{100}+q^{101}+q^{103}+q^{104}+q^{106}+q^{107})\\
&\quad+ x^{14}(q^{102}+q^{104}+2 q^{105}+q^{107}+2 q^{108}+q^{109}+2 q^{111}+q^{112}+q^{114})\\
&\quad+x^{15}(q^{106}+q^{109}+q^{110}+q^{112}+q^{113}+q^{115}+q^{116}+q^{119}).
\end{align*}}
\end{theorem}
\subsection{Partition set of type IV}
Recall that the partition set of type IV is the set of partitions with difference at least $3$ at distance $3$ such that if parts at distance $2$ differ by at most $1$, then the sum of the two parts and their intermediate part is congruent to $2$ modulo $3$. In other words, if $\lambda=\lambda_1+\lambda_2+\cdots+\lambda_\ell$ is in this partition set, then
\begin{enumerate}[label=(\roman*)]
\item $\lambda_i-\lambda_{i+3}\ge 3$;
\item $\lambda_i-\lambda_{i+2}\le 1$ implies $\lambda_i+\lambda_{i+1}+\lambda_{i+2}\equiv 2 \pmod{3}$.
\end{enumerate}
Let $\mathscr{P}_{T_{\mathrm{IV}}}$ denote the partition set of type IV.
\begin{claim}
$\mathscr{P}_{T_{\mathrm{IV}}}$ is a partition ideal of modulus $3$.
\end{claim}
\begin{claim}
The $L_{\mathscr{P}_{T_{\mathrm{IV}}}}$ corresponding to modulus $3$ equals
\begin{align*}
\{&\emptyset,\;1,\;1+1,\;2,\;2+1,\;2+2,\;2+2+1,\;3,\;3+1,\\
&3+1+1,\;3+2,\;3+2+1,\;3+3,\;3+3+1,\;3+3+2\}.
\end{align*}
\end{claim}
\begin{claim}\label{claim:T4-3}
The span and linking set of partitions in $L_{\mathscr{P}_{T_{\mathrm{IV}}}}$ are given as follows.
{\footnotesize\begin{equation*}
\begin{array}{lcp{0.05cm}l}
& \text{span} && \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\text{linking set}\\
\pi_0 = \emptyset & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_1 = 1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_2 = 1+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_3 = 2 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_4 = 2+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_5 = 2+2 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_6 = 2+2+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_6,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_7 = 3 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_8 = 3+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_9 = 3+1+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_2,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_7,\; \pi_8,\; \pi_9,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_{10} = 3+2 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_{11} = 3+2+1 & 1 && \{\pi_0,\; \pi_1,\; \pi_3,\; \pi_4,\; \pi_5,\; \pi_7,\; \pi_8,\; \pi_{10},\; \pi_{11},\; \pi_{12},\; \pi_{13},\; \pi_{14}\}\\
\pi_{12} = 3+3 & 1 && \{\pi_0,\; \pi_3,\; \pi_7,\; \pi_{10},\; \pi_{12},\; \pi_{14}\}\\
\pi_{13} = 3+3+1 & 1 && \{\pi_0,\; \pi_3,\; \pi_7,\; \pi_{10},\; \pi_{12},\; \pi_{14}\}\\
\pi_{14} = 3+3+2 & 1 && \{\pi_0,\; \pi_3,\; \pi_7,\; \pi_{10},\; \pi_{12},\; \pi_{14}\}\\
\end{array}
\end{equation*}}
\end{claim}
Let us denote by $H_i(x)=H_i(x,q)$ the generating function of partitions $\lambda$ in $\mathscr{P}_{T_{\mathrm{IV}}}$ with $\mathrm{Tail}_m(\lambda)=\pi_i$ for $i=0,1,\ldots, 14$ where the $\pi_i$'s are as defined in Claim \ref{claim:T4-3}.
Following \eqref{eq:important-gf-2}, we have
{\footnotesize\begin{align}
H_0(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-1}q^{-1}H_1(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-2}H_2(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-1}q^{-2}H_3(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-3}H_4(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_6(xq^3)+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)\notag\\
&\quad+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-4}H_5(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3)\notag\\
&\quad+H_7(xq^3)+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)\notag\\
&\quad+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-3}q^{-5}H_6(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_6(xq^3)\notag\\
&\quad+H_7(xq^3)+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)\notag\\
&\quad+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-1}q^{-3}H_7(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)\notag\\
&\quad+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-4}H_8(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)\notag\\
&\quad+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-3}q^{-5}H_9(x)&=H_0(xq^3)+H_1(xq^3)+H_2(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)\notag\\
&\quad+H_7(xq^3)+H_8(xq^3)+H_9(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)\notag\\
&\quad+H_{12}(xq^3)+H_{13}(xq^3)+H_{14}(xq^3),\\
x^{-2}q^{-5}H_{10}(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-3}q^{-6}H_{11}(x)&=H_0(xq^3)+H_1(xq^3)+H_3(xq^3)+H_4(xq^3)+H_5(xq^3)+H_7(xq^3)\notag\\
&\quad+H_8(xq^3)+H_{10}(xq^3)+H_{11}(xq^3)+H_{12}(xq^3)+H_{13}(xq^3)\notag\\
&\quad+H_{14}(xq^3),\\
x^{-2}q^{-6}H_{12}(x)&=H_0(xq^3)+H_3(xq^3)+H_7(xq^3)+H_{10}(xq^3)+H_{12}(xq^3)+H_{14}(xq^3),\\
x^{-3}q^{-7}H_{13}(x)&=H_0(xq^3)+H_3(xq^3)+H_7(xq^3)+H_{10}(xq^3)+H_{12}(xq^3)+H_{14}(xq^3),\\
x^{-3}q^{-8}H_{14}(x)&=H_0(xq^3)+H_3(xq^3)+H_7(xq^3)+H_{10}(xq^3)+H_{12}(xq^3)+H_{14}(xq^3).
\end{align}}
This system may be simplified as
\begin{framed}
{\begin{align}
H_0(x)&=(1+xq^4+x^2q^8+xq^5+x^2q^9)H_0(xq^3)+(1+xq^4)H_5(xq^3)\notag\\
&\quad+(1+xq^4+x^2q^8)H_7(xq^3)+(1+xq^4)H_{10}(xq^3)\notag\\
&\quad+(1+xq^4+xq^5)H_{12}(xq^3),\\
H_5(x)&=(x^2q^4+x^3q^8+x^3q^9+x^4q^{13})H_0(xq^3)+(x^2q^4+x^3q^8)H_5(xq^3)\notag\\
&\quad+(x^2q^4+x^3q^8)H_7(xq^3)+(x^2q^4+x^3q^8)H_{10}(xq^3)\notag\\
&\quad+(x^2q^4+x^3q^8+x^3q^9)H_{12}(xq^3),\\
H_7(x)&=(xq^3+x^2q^7+x^3q^{11}+x^2q^8+x^3q^{12})H_0(xq^3)+xq^3H_5(xq^3)\notag\\
&\quad+(xq^3+x^2q^7+x^3q^{11})H_7(xq^3)+(xq^3+x^2q^7)H_{10}(xq^3)\notag\\
&\quad+(xq^3+x^2q^7+x^2q^8)H_{12}(xq^3),\\
H_{10}(x)&=(x^2q^5+x^3q^9+x^3q^{10}+x^4q^{14})H_0(xq^3)+x^2q^5H_5(xq^3)\notag\\
&\quad+(x^2q^5+x^3q^9)H_7(xq^3)+(x^2q^5+x^3q^9)H_{10}(xq^3)\notag\\
&\quad+(x^2q^5+x^3q^9+x^3q^{10})H_{12}(xq^3),\\
H_{12}(x)&=(x^2q^6+x^3q^{11})H_0(xq^3)+x^2q^6H_7(xq^3)+x^2q^6H_{10}(xq^3)\notag\\
&\quad+(x^2q^6+x^3q^{11})H_{12}(xq^3).
\end{align}}
\end{framed}
Let $G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x)=G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x,q)$ denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{IV}}}$ whose smallest part is at least $1$.
Let $G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x)$ denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{IV}}}$ where $1$ appears at most once.
Let $G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x)$ denote the generating function of partitions in $\mathscr{P}_{T_{\mathrm{IV}}}$ where the smallest part is at least $2$ with $2$ appearing at most once.
It follows that
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x)&=H_0(x)+H_1(x)+H_2(x)+H_3(x)+H_4(x)+H_5(x)\notag\\
&\quad+H_6(x)+H_7(x)+H_8(x)+H_9(x)+H_{10}(x)\notag\\
&\quad+H_{11}(x)+H_{12}(x)+H_{13}(x)+H_{14}(x)\notag\\
&=H_0(xq^{-3}),\\
G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x)&=H_0(x)+H_1(x)+H_3(x)+H_4(x)+H_5(x)+H_6(x)\notag\\
&\quad+H_7(x)+H_8(x)+H_{10}(x)+H_{11}(x)+H_{12}(x)\notag\\
&\quad+H_{13}(x)+H_{14}(x)\notag\\
&=x^{-2}q^2H_5(xq^{-3}),\\
G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x)&=H_0(x)+H_3(x)+H_7(x)+H_{10}(x)+H_{12}(x)+H_{14}(x)\notag\\
&=x^{-2}H_{12}(xq^{-3}).
\end{align}
Analogously, we can use the algorithm in Section \ref{sec:q-diff} to deduce the following $q$-difference equations for $G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x)$, $G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x)$ and $G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x)$, respectively.
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x)&+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(xq^3)+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(xq^9)+p_{12}(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(xq^{12})=0,
\end{align}
where
{\footnotesize\begin{align*}
p_0(x,q)&=1+x(q^{4}+q^{5}+q^{7}+q^{8}+q^{10}+q^{11})\\
&\quad+ x^2(q^{9}+q^{11}+2 q^{12}+q^{14}+2 q^{15}+q^{16}+2 q^{18}+q^{19}+q^{21})\\
&\quad+ x^3(q^{16}+q^{19}+q^{20}+q^{22}+q^{23}+q^{25}+q^{26}+q^{29}),\\
p_3(x,q)&=-1-x(q+q^{2}+q^{3}+q^{4}+q^{5}+q^{7}+q^{8}+q^{10}+q^{11})\\
&\quad
-x^2(q^{2}+q^{3}+2 q^{4}+2 q^{5}+3 q^{6}+2 q^{7}+2 q^{8}+3 q^{9}+2 q^{10}+3 q^{11}+4 q^{12}+2 q^{13}\\
&\quad\quad\quad\quad+2 q^{14}+2 q^{15}+q^{16}+2 q^{18}+q^{19}+q^{21})\\
&\quad
- x^3(2 q^{5}+2 q^{6}+3 q^{7}+4 q^{8}+4 q^{9}+4 q^{10}+5 q^{11}+6 q^{12}+7 q^{13}+7 q^{14}+6 q^{15}\\
&\quad\quad\quad\quad+6 q^{16}+5 q^{17}+3 q^{18}+4 q^{19}+4 q^{20}+3 q^{21}+3 q^{22}+2 q^{23}+q^{24}+q^{25}\\
&\quad\quad\quad\quad+q^{26}+q^{29})\\
&\quad
-x^4(2 q^{9}+3 q^{10}+2 q^{11}+4 q^{12}+6 q^{13}+6 q^{14}+8 q^{15}+10 q^{16}+8 q^{17}+9 q^{18}+9 q^{19}\\
&\quad\quad\quad\quad+7 q^{20}+8 q^{21}+8 q^{22}+7 q^{23}+6 q^{24}+5 q^{25}+3 q^{26}+3 q^{27}+2 q^{28}+q^{29}\\
&\quad\quad\quad\quad+q^{30}+q^{31}+q^{32})\\
&\quad
-x^5(q^{14}+q^{15}+2 q^{16}+5 q^{17}+3 q^{18}+4 q^{19}+8 q^{20}+6 q^{21}+5 q^{22}+9 q^{23}+8 q^{24}\\
&\quad\quad\quad\quad+6 q^{25}+9 q^{26}+6 q^{27}+4 q^{28}+6 q^{29}+4 q^{30}+2 q^{31}+3 q^{32}+2 q^{33}+q^{34}\\
&\quad\quad\quad\quad+q^{35})\\
&\quad
- x^6(q^{21}+q^{22}+2 q^{24}+2 q^{25}+q^{26}+3 q^{27}+3 q^{28}+q^{29}+3 q^{30}+4 q^{31}+q^{32}+2 q^{33}\\
&\quad\quad\quad\quad+3 q^{34}+q^{35}+q^{36}+2 q^{37}+q^{40}),\\
p_6(x,q)&=x^4(q^{12}+q^{14}+q^{16}+q^{17}+q^{19}+q^{21})\\
&\quad
+x^5(q^{13}+q^{15}+2 q^{16}+2 q^{17}+3 q^{18}+3 q^{19}+4 q^{20}+4 q^{21}+4 q^{22}+4 q^{23}+4 q^{24}\\
&\quad\quad\quad\quad+4 q^{25}+3 q^{26}+3 q^{27}+2 q^{28}+2 q^{29}+q^{30}+q^{32})\\
&\quad
+x^6(q^{17}+q^{18}+2 q^{19}+3 q^{20}+4 q^{21}+7 q^{22}+7 q^{23}+8 q^{24}+8 q^{25}+11 q^{26}+11 q^{27}\\
&\quad\quad\quad\quad+12 q^{28}+12 q^{29}+11 q^{30}+11 q^{31}+8 q^{32}+8 q^{33}+7 q^{34}+7 q^{35}+4 q^{36}\\
&\quad\quad\quad\quad+3 q^{37}+2 q^{38}+q^{39}+q^{40})\\
&\quad
+x^7(2 q^{23}+3 q^{24}+3 q^{25}+4 q^{26}+7 q^{27}+9 q^{28}+10 q^{29}+14 q^{30}+14 q^{31}+16 q^{32}\\
&\quad\quad\quad\quad+17 q^{33}+15 q^{34}+15 q^{35}+17 q^{36}+16 q^{37}+14 q^{38}+14 q^{39}+10 q^{40}+9 q^{41}\\
&\quad\quad\quad\quad+7 q^{42}+4 q^{43}+3 q^{44}+3 q^{45}+2 q^{46})\\
&\quad
+x^8(q^{29}+q^{30}+3 q^{31}+4 q^{32}+4 q^{33}+8 q^{34}+8 q^{35}+7 q^{36}+11 q^{37}+13 q^{38}+10 q^{39}\\
&\quad\quad\quad\quad+14 q^{40}+14 q^{41}+10 q^{42}+13 q^{43}+11 q^{44}+7 q^{45}+8 q^{46}+8 q^{47}+4 q^{48}\\
&\quad\quad\quad\quad+4 q^{49}+3 q^{50}+q^{51}+q^{52})\\
&\quad
+x^9(q^{36}+q^{38}+2 q^{39}+q^{40}+2 q^{41}+4 q^{42}+2 q^{43}+3 q^{44}+5 q^{45}+3 q^{46}+3 q^{47}\\
&\quad\quad\quad\quad+5 q^{48}+3 q^{49}+2 q^{50}+4 q^{51}+2 q^{52}+q^{53}+2 q^{54}+q^{55}+q^{57}),\\
p_{9}(x,q)&=-x^6q^{30}
-x^7(q^{31}+q^{32}+q^{34}+q^{35}+q^{37}+q^{38}+q^{39}+q^{40}+q^{41})\\
&\quad
-x^8(q^{33}+q^{35}+2 q^{36}+q^{38}+2 q^{39}+2 q^{40}+2 q^{41}+4 q^{42}+3 q^{43}+2 q^{44}+3 q^{45}\\
&\quad\quad\quad\quad+2 q^{46}+2 q^{47}+3 q^{48}+2 q^{49}+2 q^{50}+q^{51}+q^{52})\\
&\quad
-x^9(q^{37}+q^{40}+q^{41}+q^{42}+2 q^{43}+3 q^{44}+3 q^{45}+4 q^{46}+4 q^{47}+3 q^{48}+5 q^{49}\\
&\quad\quad\quad\quad+6 q^{50}+6 q^{51}+7 q^{52}+7 q^{53}+6 q^{54}+5 q^{55}+4 q^{56}+4 q^{57}+4 q^{58}+3 q^{59}\\
&\quad\quad\quad\quad+2 q^{60}+2 q^{61})\\
&\quad
-x^{10}(q^{46}+q^{47}+q^{48}+q^{49}+2 q^{50}+3 q^{51}+3 q^{52}+5 q^{53}+6 q^{54}+7 q^{55}+8 q^{56}\\
&\quad\quad\quad\quad+8 q^{57}+7 q^{58}+9 q^{59}+9 q^{60}+8 q^{61}+10 q^{62}+8 q^{63}+6 q^{64}+6 q^{65}+4 q^{66}\\
&\quad\quad\quad\quad+2 q^{67}+3 q^{68}+2 q^{69})\\
&\quad
-x^{11}(q^{55}+q^{56}+2 q^{57}+3 q^{58}+2 q^{59}+4 q^{60}+6 q^{61}+4 q^{62}+6 q^{63}+9 q^{64}+6 q^{65}\\
&\quad\quad\quad\quad+8 q^{66}+9 q^{67}+5 q^{68}+6 q^{69}+8 q^{70}+4 q^{71}+3 q^{72}+5 q^{73}+2 q^{74}+q^{75}+q^{76})\\
&\quad
-x^{12}(q^{62}+2 q^{65}+q^{66}+q^{67}+3 q^{68}+2 q^{69}+q^{70}+4 q^{71}+3 q^{72}+q^{73}+3 q^{74}+3 q^{75}\\
&\quad\quad\quad\quad+q^{76}+2 q^{77}+2 q^{78}+q^{80}+q^{81}),\\
\intertext{and}
p_{12}(x,q)&=x^{12}q^{93} + x^{13}(q^{94}+q^{95}+q^{97}+q^{98}+q^{100}+q^{101})\\
&\quad+ x^{14}(q^{96}+q^{98}+2 q^{99}+q^{101}+2 q^{102}+q^{103}+2 q^{105}+q^{106}+q^{108})\\
&\quad+x^{15}(q^{100}+q^{103}+q^{104}+q^{106}+q^{107}+q^{109}+q^{110}+q^{113}).
\end{align*}}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x)&+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(xq^3)+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(xq^9)+p_{12}(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(xq^{12})=0,
\end{align}
where
{\footnotesize\begin{align*}
p_0(x,q)&=1+x(q^{4}+q^{5}+q^{6}+q^{7}+q^{8}+q^{9})\\
&\quad+ x^2(q^{9}+q^{10}+q^{11}+2 q^{12}+2 q^{13}+2 q^{14}+q^{15}+q^{16}+q^{17})\\
&\quad+ x^3(q^{16}+q^{17}+q^{18}+q^{19}+q^{20}+q^{21}+q^{22}+q^{23}),\\
p_3(x,q)&=-1-x(q+q^{2}+q^{3}+q^{4}+q^{5}+q^{6}+q^{7}+q^{8}+q^{9})\\
&\quad
-x^2(q^{3}+2 q^{4}+2 q^{5}+3 q^{6}+3 q^{7}+4 q^{8}+4 q^{9}+4 q^{10}+3 q^{11}+3 q^{12}+2 q^{13}+2 q^{14}\\
&\quad\quad\quad\quad+q^{15}+q^{16}+q^{17})\\
&\quad
- x^3(q^{5}+q^{6}+2 q^{7}+4 q^{8}+5 q^{9}+6 q^{10}+8 q^{11}+8 q^{12}+8 q^{13}+8 q^{14}+8 q^{15}+7 q^{16}\\
&\quad\quad\quad\quad+6 q^{17}+4 q^{18}+3 q^{19}+2 q^{20}+q^{21}+q^{22}+q^{23})\\
&\quad
-x^4(q^{9}+2 q^{10}+3 q^{11}+5 q^{12}+7 q^{13}+8 q^{14}+9 q^{15}+11 q^{16}+13 q^{17}+13 q^{18}\\
&\quad\quad\quad\quad+11 q^{19}+10 q^{20}+8 q^{21}+6 q^{22}+5 q^{23}+4 q^{24}+3 q^{25}+q^{26})\\
&\quad
-x^5(q^{14}+2 q^{15}+2 q^{16}+4 q^{17}+6 q^{18}+8 q^{19}+10 q^{20}+10 q^{21}+10 q^{22}+9 q^{23}\\
&\quad\quad\quad\quad+8 q^{24}+8 q^{25}+7 q^{26}+5 q^{27}+3 q^{28}+q^{29}+q^{30}+q^{31})\\
&\quad
- x^6(q^{21}+2 q^{22}+2 q^{23}+3 q^{24}+3 q^{25}+3 q^{26}+4 q^{27}+4 q^{28}+3 q^{29}+2 q^{30}+2 q^{31}\\
&\quad\quad\quad\quad+q^{32}+q^{33}+q^{34}),\\
p_6(x,q)&=x^4(q^{16}+q^{17}+q^{18}+q^{19}+q^{20}+q^{21})\\
&\quad
+x^5(q^{17}+2 q^{18}+2 q^{19}+3 q^{20}+5 q^{21}+6 q^{22}+5 q^{23}+5 q^{24}+6 q^{25}+5 q^{26}+3 q^{27}\\
&\quad\quad\quad\quad+2 q^{28}+2 q^{29}+q^{30})\\
&\quad
+x^6(q^{20}+2 q^{21}+5 q^{22}+7 q^{23}+8 q^{24}+10 q^{25}+13 q^{26}+14 q^{27}+15 q^{28}+15 q^{29}\\
&\quad\quad\quad\quad+14 q^{30}+13 q^{31}+10 q^{32}+8 q^{33}+7 q^{34}+5 q^{35}+2 q^{36}+q^{37})\\
&\quad
+x^7(q^{24}+2 q^{25}+4 q^{26}+7 q^{27}+10 q^{28}+13 q^{29}+16 q^{30}+19 q^{31}+21 q^{32}+21 q^{33}\\
&\quad\quad\quad\quad+21 q^{34}+21 q^{35}+19 q^{36}+16 q^{37}+13 q^{38}+10 q^{39}+7 q^{40}+4 q^{41}+2 q^{42}\\
&\quad\quad\quad\quad+q^{43})\\
&\quad
+x^8(q^{29}+q^{30}+2 q^{31}+5 q^{32}+7 q^{33}+10 q^{34}+12 q^{35}+14 q^{36}+16 q^{37}+16 q^{38}\\
&\quad\quad\quad\quad+16 q^{39}+16 q^{40}+14 q^{41}+12 q^{42}+10 q^{43}+7 q^{44}+5 q^{45}+2 q^{46}+q^{47}\\
&\quad\quad\quad\quad+q^{48})\\
&\quad
+x^9(q^{36}+q^{37}+2 q^{38}+3 q^{39}+3 q^{40}+4 q^{41}+5 q^{42}+5 q^{43}+5 q^{44}+5 q^{45}+4 q^{46}\\
&\quad\quad\quad\quad+3 q^{47}+3 q^{48}+2 q^{49}+q^{50}+q^{51}),\\
p_{9}(x,q)&=-x^6q^{36}
-x^7(q^{37}+q^{38}+q^{39}+q^{40}+q^{41}+q^{42}+q^{43}+q^{44}+q^{45})\\
&\quad
-x^8(q^{39}+q^{40}+q^{41}+2 q^{42}+2 q^{43}+3 q^{44}+3 q^{45}+4 q^{46}+4 q^{47}+4 q^{48}+3 q^{49}\\
&\quad\quad\quad\quad+3 q^{50}+2 q^{51}+2 q^{52}+q^{53})\\
&\quad
-x^9(q^{43}+q^{44}+q^{45}+2 q^{46}+3 q^{47}+4 q^{48}+6 q^{49}+7 q^{50}+8 q^{51}+8 q^{52}+8 q^{53}\\
&\quad\quad\quad\quad+8 q^{54}+8 q^{55}+6 q^{56}+5 q^{57}+4 q^{58}+2 q^{59}+q^{60}+q^{61})\\
&\quad
-x^{10}(q^{50}+3 q^{51}+4 q^{52}+5 q^{53}+6 q^{54}+8 q^{55}+10 q^{56}+11 q^{57}+13 q^{58}+13 q^{59}\\
&\quad\quad\quad\quad+11 q^{60}+9 q^{61}+8 q^{62}+7 q^{63}+5 q^{64}+3 q^{65}+2 q^{66}+q^{67})\\
&\quad
-x^{11}(q^{55}+q^{56}+q^{57}+3 q^{58}+5 q^{59}+7 q^{60}+8 q^{61}+8 q^{62}+9 q^{63}+10 q^{64}+10 q^{65}\\
&\quad\quad\quad\quad+10 q^{66}+8 q^{67}+6 q^{68}+4 q^{69}+2 q^{70}+2 q^{71}+q^{72})\\
&\quad
-x^{12}(q^{62}+q^{63}+q^{64}+2 q^{65}+2 q^{66}+3 q^{67}+4 q^{68}+4 q^{69}+3 q^{70}+3 q^{71}+3 q^{72}\\
&\quad\quad\quad\quad+2 q^{73}+2 q^{74}+q^{75}),\\
\intertext{and}
p_{12}(x,q)&=x^{12}q^{93} + x^{13}(q^{94}+q^{95}+q^{96}+q^{97}+q^{98}+q^{99})\\
&\quad+ x^{14}(q^{96}+q^{97}+q^{98}+2 q^{99}+2 q^{100}+2 q^{101}+q^{102}+q^{103}+q^{104})\\
&\quad+x^{15}(q^{100}+q^{101}+q^{102}+q^{103}+q^{104}+q^{105}+q^{106}+q^{107}).
\end{align*}}
\end{theorem}
\begin{theorem}
It holds that
\begin{align}
p_0(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x)&+p_3(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(xq^3)+p_6(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x q^6)\nonumber\\
&+p_9(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(xq^9)+p_{12}(x,q)G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(xq^{12})=0,
\end{align}
where
{\footnotesize\begin{align*}
p_0(x,q)&=1+x(q^{5}+q^{6}+q^{7}+q^{8}+q^{9}+q^{10})\\
&\quad+ x^2(q^{11}+q^{12}+q^{13}+2 q^{14}+2 q^{15}+2 q^{16}+q^{17}+q^{18}+q^{19})\\
&\quad+ x^3(q^{19}+q^{20}+q^{21}+q^{22}+q^{23}+q^{24}+q^{25}+q^{26}),\\
p_3(x,q)&=-1-x(q^{2}+q^{3}+q^{4}+q^{5}+q^{6}+q^{7}+q^{8}+q^{9}+q^{10})\\
&\quad
-x^2(q^{5}+2 q^{6}+2 q^{7}+3 q^{8}+3 q^{9}+4 q^{10}+4 q^{11}+4 q^{12}+3 q^{13}+3 q^{14}+2 q^{15}+2 q^{16}\\
&\quad\quad\quad\quad+q^{17}+q^{18}+q^{19})\\
&\quad
- x^3(q^{8}+q^{9}+2 q^{10}+4 q^{11}+5 q^{12}+6 q^{13}+8 q^{14}+8 q^{15}+8 q^{16}+8 q^{17}+8 q^{18}\\
&\quad\quad\quad\quad+7 q^{19}+6 q^{20}+4 q^{21}+3 q^{22}+2 q^{23}+q^{24}+q^{25}+q^{26})\\
&\quad
-x^4(q^{13}+2 q^{14}+3 q^{15}+5 q^{16}+7 q^{17}+8 q^{18}+9 q^{19}+11 q^{20}+13 q^{21}+13 q^{22}\\
&\quad\quad\quad\quad+11 q^{23}+10 q^{24}+8 q^{25}+6 q^{26}+5 q^{27}+4 q^{28}+3 q^{29}+q^{30})\\
&\quad
-x^5(q^{19}+2 q^{20}+2 q^{21}+4 q^{22}+6 q^{23}+8 q^{24}+10 q^{25}+10 q^{26}+10 q^{27}+9 q^{28}\\
&\quad\quad\quad\quad+8 q^{29}+8 q^{30}+7 q^{31}+5 q^{32}+3 q^{33}+q^{34}+q^{35}+q^{36})\\
&\quad
- x^6(q^{27}+2 q^{28}+2 q^{29}+3 q^{30}+3 q^{31}+3 q^{32}+4 q^{33}+4 q^{34}+3 q^{35}+2 q^{36}+2 q^{37}\\
&\quad\quad\quad\quad+q^{38}+q^{39}+q^{40}),\\
p_6(x,q)&=x^4(q^{20}+q^{21}+q^{22}+q^{23}+q^{24}+q^{25})\\
&\quad
+x^5(q^{22}+2 q^{23}+2 q^{24}+3 q^{25}+5 q^{26}+6 q^{27}+5 q^{28}+5 q^{29}+6 q^{30}+5 q^{31}+3 q^{32}\\
&\quad\quad\quad\quad+2 q^{33}+2 q^{34}+q^{35})\\
&\quad
+x^6(q^{26}+2 q^{27}+5 q^{28}+7 q^{29}+8 q^{30}+10 q^{31}+13 q^{32}+14 q^{33}+15 q^{34}+15 q^{35}\\
&\quad\quad\quad\quad+14 q^{36}+13 q^{37}+10 q^{38}+8 q^{39}+7 q^{40}+5 q^{41}+2 q^{42}+q^{43})\\
&\quad
+x^7(q^{31}+2 q^{32}+4 q^{33}+7 q^{34}+10 q^{35}+13 q^{36}+16 q^{37}+19 q^{38}+21 q^{39}+21 q^{40}\\
&\quad\quad\quad\quad+21 q^{41}+21 q^{42}+19 q^{43}+16 q^{44}+13 q^{45}+10 q^{46}+7 q^{47}+4 q^{48}+2 q^{49}\\
&\quad\quad\quad\quad+q^{50})\\
&\quad
+x^8(q^{37}+q^{38}+2 q^{39}+5 q^{40}+7 q^{41}+10 q^{42}+12 q^{43}+14 q^{44}+16 q^{45}+16 q^{46}\\
&\quad\quad\quad\quad+16 q^{47}+16 q^{48}+14 q^{49}+12 q^{50}+10 q^{51}+7 q^{52}+5 q^{53}+2 q^{54}+q^{55}\\
&\quad\quad\quad\quad+q^{56})\\
&\quad
+x^9(q^{45}+q^{46}+2 q^{47}+3 q^{48}+3 q^{49}+4 q^{50}+5 q^{51}+5 q^{52}+5 q^{53}+5 q^{54}+4 q^{55}\\
&\quad\quad\quad\quad+3 q^{56}+3 q^{57}+2 q^{58}+q^{59}+q^{60}),\\
p_{9}(x,q)&=-x^6q^{42}
-x^7(q^{44}+q^{45}+q^{46}+q^{47}+q^{48}+q^{49}+q^{50}+q^{51}+q^{52})\\
&\quad
-x^8(q^{47}+q^{48}+q^{49}+2 q^{50}+2 q^{51}+3 q^{52}+3 q^{53}+4 q^{54}+4 q^{55}+4 q^{56}+3 q^{57}\\
&\quad\quad\quad\quad+3 q^{58}+2 q^{59}+2 q^{60}+q^{61})\\
&\quad
-x^9(q^{52}+q^{53}+q^{54}+2 q^{55}+3 q^{56}+4 q^{57}+6 q^{58}+7 q^{59}+8 q^{60}+8 q^{61}+8 q^{62}\\
&\quad\quad\quad\quad+8 q^{63}+8 q^{64}+6 q^{65}+5 q^{66}+4 q^{67}+2 q^{68}+q^{69}+q^{70})\\
&\quad
-x^{10}(q^{60}+3 q^{61}+4 q^{62}+5 q^{63}+6 q^{64}+8 q^{65}+10 q^{66}+11 q^{67}+13 q^{68}+13 q^{69}\\
&\quad\quad\quad\quad+11 q^{70}+9 q^{71}+8 q^{72}+7 q^{73}+5 q^{74}+3 q^{75}+2 q^{76}+q^{77})\\
&\quad
-x^{11}(q^{66}+q^{67}+q^{68}+3 q^{69}+5 q^{70}+7 q^{71}+8 q^{72}+8 q^{73}+9 q^{74}+10 q^{75}+10 q^{76}\\
&\quad\quad\quad\quad+10 q^{77}+8 q^{78}+6 q^{79}+4 q^{80}+2 q^{81}+2 q^{82}+q^{83})\\
&\quad
-x^{12}(q^{74}+q^{75}+q^{76}+2 q^{77}+2 q^{78}+3 q^{79}+4 q^{80}+4 q^{81}+3 q^{82}+3 q^{83}+3 q^{84}\\
&\quad\quad\quad\quad+2 q^{85}+2 q^{86}+q^{87}),\\
\intertext{and}
p_{12}(x,q)&=x^{12}q^{105} + x^{13}(q^{107}+q^{108}+q^{109}+q^{110}+q^{111}+q^{112})\\
&\quad+ x^{14}(q^{110}+q^{111}+q^{112}+2 q^{113}+2 q^{114}+2 q^{115}+q^{116}+q^{117}+q^{118})\\
&\quad+x^{15}(q^{115}+q^{116}+q^{117}+q^{118}+q^{119}+q^{120}+q^{121}+q^{122}).
\end{align*}}
\end{theorem}
\section{``Guessing'' the generating functions}
It is, of course, not easy to discover a closed form for each generating function directly from $q$-difference equations obtained in the previous section. However, Andrews' conjecture presented in the introduction shall give us enough clues.
Recall that Andrews' conjecture states as follows.
\begin{conjecture}
Every linked partition ideal $\mathscr{I}$ has a bivariate generating function $G_{\mathscr{I}}(x,q)$ of the form
\begin{equation}\label{eq:And-conj}
\sum_{n_1,\ldots,n_r\ge 0}\frac{(-1)^{L_1(n_1,\ldots,n_r)}q^{Q(n_1,\ldots,n_r)+L_2(n_1,\ldots,n_r)}x^{L_3(n_1,\ldots,n_r)}}{(q^{B_1};q^{A_1})_{n_1}\cdots (q^{B_r};q^{A_r})_{n_r}},
\end{equation}
where $L_1$, $L_2$ and $L_3$ are linear forms in $n_1,\ldots,n_r$ and $Q$ is a quadratic form in $n_1,\ldots,n_r$.
\end{conjecture}
It also appears to be true that some ``nice'' subsets of a linked partition ideal enjoy a generating function of the form \eqref{eq:And-conj}. One may investigate the second Rogers--Ramanujan identity as an example.
Hence, we may search from a number of multi-summations of the form \eqref{eq:And-conj} and compare the series expansions to find suitable candidates.
\begin{theorem}
Let $G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x,q)$ (resp.~$G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x,q)$, $G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x,q)$) denote the generating function of partitions of type I whose smallest part is at least $1$ (resp.~$2$, $3$). We have
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x,q)&=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}},\label{eq:real-gf-t11}\\
G_{\mathscr{P}_{T_{\mathrm{I},2}}}(x,q)&=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2+n_1+3n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}},\label{eq:real-gf-t12}\\
G_{\mathscr{P}_{T_{\mathrm{I},3}}}(x,q)&=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2+2n_1+3n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}}.\label{eq:real-gf-t13}
\end{align}
\end{theorem}
\begin{remark}
Here \eqref{eq:real-gf-t11}, \eqref{eq:real-gf-t12} and \eqref{eq:real-gf-t13} are (3.1), (3.10) and (3.14) in \cite{Kur2018}. They correspond to the Kanade--Russell conjectures $I_1$, $I_2$ and $I_3$, respectively.
\end{remark}
\begin{theorem}
Let $G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x,q)$ (resp.~$G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x,q)$) denote the generating function of partitions of type II whose smallest part is at least $1$ (resp.~$2$) and let $G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x,q)$ denote the generating function of partitions of type II where $1$ appears at most once. We have
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{II},1}}}(x,q)&=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2-n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}},\label{eq:real-gf-t21}\\
G_{\mathscr{P}_{T_{\mathrm{II},2}}}(x,q)&=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2+n_1+2n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}},\label{eq:real-gf-t22}\\
G_{\mathscr{P}_{T_{\mathrm{II},a}}}(x,q)&=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2+2n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}}.\label{eq:real-gf-t23}
\end{align}
\end{theorem}
\begin{remark}
Here \eqref{eq:real-gf-t22} is (3.15) in \cite{Kur2018}. It corresponds to the Kanade--Russell conjecture $I_4$.
\end{remark}
\begin{theorem}
Let $G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x,q)$ (resp.~$G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x,q)$) denote the generating function of partitions of type III whose smallest part is at least $1$ (resp.~$2$) and let $G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x,q)$ denote the generating function of partitions of type III where $1$ appears at most once. We have
\begin{align}
&G_{\mathscr{P}_{T_{\mathrm{III},1}}}(x,q)\nonumber\\
&\quad=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}-n_2-\frac{n_3}{2}}x^{n_1+2n_2+3n_3}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}},\label{eq:real-gf-t31}\\
&G_{\mathscr{P}_{T_{\mathrm{III},2}}}(x,q)\nonumber\\
&\quad=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{3n_1}{2}+n_2+\frac{5n_3}{2}}x^{n_1+2n_2+3n_3}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}},\label{eq:real-gf-t32}\\
&G_{\mathscr{P}_{T_{\mathrm{III},a}}}(x,q)\nonumber\\
&\quad=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}+n_2+\frac{5n_3}{2}}x^{n_1+2n_2+3n_3}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}.\label{eq:real-gf-t33}
\end{align}
\end{theorem}
\begin{remark}
Here \eqref{eq:real-gf-t33} is (47) (corrected: in its numerator, the last term of the exponent of $q$ should read $4k$ instead of $3k$) in \cite{KR2018}. It corresponds to the Kanade--Russell conjecture $I_5$.
\end{remark}
\begin{theorem}
Let $G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x,q)$ denote the generating function of partitions of type IV whose smallest part is at least $1$, let $G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x,q)$ denote the generating function of partitions of type IV where $1$ appears at most once and let $G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x,q)$ denote the generating function of partitions of type IV where the smallest part is at least $2$ with $2$ appearing at most once. We have
\begin{align}
&G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(x,q)\nonumber\\
&\quad=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}-n_2+\frac{n_3}{2}}x^{n_1+2n_2+3n_3}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}},\label{eq:real-gf-t41}\\
&G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(x,q)\nonumber\\
&\quad=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}+n_2+\frac{n_3}{2}}x^{n_1+2n_2+3n_3}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}},\label{eq:real-gf-t42}\\
&G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(x,q)\nonumber\\
&\quad=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{3n_1}{2}+3n_2+\frac{7n_3}{2}}x^{n_1+2n_2+3n_3}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}.\label{eq:real-gf-t43}
\end{align}
\end{theorem}
\begin{remark}
Here \eqref{eq:real-gf-t43} is (51) in \cite{KR2018}. It corresponds to the Kanade--Russell conjecture $I_6$.
\end{remark}
In the above theorems, we rediscover six generating function identities proved in \cite{KR2018} and \cite{Kur2018} and obtain six new identities. We will provide an approach to prove these identities in the next section with the help of computer algebra.
\begin{remark}
It is, of course, fine to discover the above sum-like generating functions by trial and error with this tedious work left to a computer. But sometimes human observation might reduce the workload. Let us use \eqref{eq:real-gf-t11} as an example. If we write
\begin{equation*}
G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x,q)=\sum_{M\ge 0}g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)x^M,
\end{equation*}
then the $q$-difference equation in Theorem \ref{th:q-diff-T11} gives the first several expressions of $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$:
\begin{align*}
g_{\mathscr{P}_{T_{\mathrm{I},1}}}(0)&=1,\\
g_{\mathscr{P}_{T_{\mathrm{I},1}}}(1)&=\frac{q}{1-q},\\
g_{\mathscr{P}_{T_{\mathrm{I},1}}}(2)&=\frac{q^3+q^4+q^6}{(1-q^2)(1-q^3)},\\
g_{\mathscr{P}_{T_{\mathrm{I},1}}}(3)&=\frac{q^7}{(1-q)(1-q^2)(1-q^3)},\\
g_{\mathscr{P}_{T_{\mathrm{I},1}}}(4)&=\frac{q^{12}+q^{15}+q^{17}+q^{18}-q^{19}+q^{20}-q^{21}}{(1-q)(1-q^3)(1-q^4)(1-q^6)}.
\end{align*}
Recall that the sum-like generating function is
\begin{equation}\label{eq:and-gf}
\sum_{n_1,\ldots,n_r\ge 0}\frac{(-1)^{L_1(n_1,\ldots,n_r)}q^{Q(n_1,\ldots,n_r)+L_2(n_1,\ldots,n_r)}x^{L_3(n_1,\ldots,n_r)}}{(q^{B_1};q^{A_1})_{n_1}\cdots (q^{B_r};q^{A_r})_{n_r}}.
\end{equation}
We observe that the numerator of $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(2)$ has more than one term. Hence, the linear equation $L_3(n_1,\ldots,n_r)=2$ might have multiple nonnegative solutions $(n_1,\ldots,n_r)$. It is fair to guess that $L_3$ looks like $n_1+n_2+\cdots$ or $n_1+2n_2+\cdots$. Also, the denominators of $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ indicate that there might be terms like $(q;q)_{n}$ and $(q^3;q^3)_{n}$ in the denominator of the summand in \eqref{eq:and-gf}. Hence, one may first try multi-summations like
$$\sum_{n_1,n_2\ge 0}\frac{(-1)^{L_1(n_1,n_2)}q^{Q(n_1,n_2)+L_2(n_1,n_2)}x^{n_1+n_2}}{(q;q)_{n_1} (q^{3};q^{3})_{n_2}}$$
or
$$\sum_{n_1,n_2\ge 0}\frac{(-1)^{L_1(n_1,n_2)}q^{Q(n_1,n_2)+L_2(n_1,n_2)}x^{n_1+2n_2}}{(q;q)_{n_1} (q^{3};q^{3})_{n_2}}.$$
If these expressions fail to be a candidate, then one could continue to modify them and carry on the searching procedure. However, it should be emphasized that in this remark we do not intend to assert that the sum-like generating function must contain some particular ``magical'' exponents and bases.
\end{remark}
\section{Computer algebra assistance}
Proofs of the generating function identities in the previous section can be carried out by the same procedure. We only demonstrate \eqref{eq:real-gf-t11} as an instance.
\subsection{The main idea}
Let us write
\begin{equation}\label{eq:T11-new}
G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x,q)=\sum_{M\ge 0}g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)x^M,
\end{equation}
where $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)\in\mathbb{Q}(q)$. We can translate the $q$-difference equation in Theorem \ref{th:q-diff-T11} to a recurrence of $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$.
\begin{definition}
Let $\mathbb{K}=\mathbb{Q}(q)$ with $q$ transcendental. A sequence $(a_n)$ in $\mathbb{K}$ is called \textit{$q$-holonomic} if there exist $p,p_0,\ldots,p_r\in\mathbb{K}[x]$, not all zero, such that
$$p_0(q^n)a_n+p_1(q^n)a_{n+1}+\cdots+p_r(q^n)a_{n+r}=p(q^n).$$
\end{definition}
Hence, the sequence $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ is $q$-holonomic.
On the other hand, if we write
\begin{equation}\label{eq:T11-NewNew}
\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}}=\sum_{M\ge 0}\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)x^M,
\end{equation}
we may also find a recurrence relation satisfied by $\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$. Hence, $\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ is also $q$-holonomic.
A result of Kauers and Koutschan \cite{KK2009} states that if two sequences $(a_n)$ and $(b_n)$ are $q$-holonomic, so is their linear combination $(\alpha a_n+\beta b_n)$. Hence, we may find a recurrence relation satisfied by $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)-\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$. As long as $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)-\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)=0$ for enough initial cases, we are safe to say that this difference is identical to $0$ for all $M$ and hence arrive at the desired generating function identity.
\subsection{Two \textit{Mathematica} packages}
To proceed with our proof, we require two \textit{Mathematica} packages: \texttt{qMultiSum} \cite{Rie2003} and \texttt{qGeneratingFunctions} \cite{KK2009}. These packages along with their instructions can be found on the webpage of Research Institute for Symbolic Computation (RISC) of Johannes Kepler University.\footnote{See \url{https://www3.risc.jku.at/research/combinat/software/ergosum/index.html}.}
To begin with, we load the two packages after installing them.
\begin{lstlisting}[language=Mathematica]
<<RISC`qMultiSum`
<<RISC`qGeneratingFunctions`
\end{lstlisting}
\subsection{Recurrence for $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$}
For the polynomials $p_{3i}(x,q)$ ($i=0,\ldots,3$) defined in Theorem \ref{th:q-diff-T11}, we write
$$p_{3i}(x,q)=\sum_{j=0}^{J_{3i}}p_{3i,j}(q)x^j.$$
Then with \eqref{eq:T11-new}, one may rewrite \eqref{eq:q-diff-T11} as
\begin{align*}
0&=\sum_{i=0}^3 p_{3i}(x,q)G_{\mathscr{P}_{T_{\mathrm{I},1}}}(xq^{3i})\\
&=\sum_{i=0}^3\sum_{j=0}^{J_{3i}}\sum_{m\ge 0}p_{3i,j}(q)g_{\mathscr{P}_{T_{\mathrm{I},1}}}(m) q^{3im}x^{m+j}\\
&=\sum_{M\ge 0}\sum_{i=0}^3 \sum_{m=\max(0,M-J_{3i})}^M q^{3im}p_{3i,M-m}(x,q)g_{\mathscr{P}_{T_{\mathrm{I},1}}}(m) x^M.
\end{align*}
Hence, for all $M\ge 0$,
\begin{align}
\sum_{i=0}^3 \sum_{m=\max(0,M-J_{3i})}^M q^{3im}p_{3i,M-m}(x,q)g_{\mathscr{P}_{T_{\mathrm{I},1}}}(m) = 0,
\end{align}
from which we see that $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ ($M\ge 1$) is uniquely determined by $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(0)$. It is also trivial that $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(0)=1$.
In particular, for $M\ge 0$, we have the following recurrence
\begin{align}
0&=g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)\left((q^{28}+q^{30})q^{9M}\right)\nonumber\\
&\quad+g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+1)\left((q^{19}+q^{21})q^{6(M+1)}+q^{27}q^{9(M+1)}\right)\nonumber\\
&\quad+g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+2)\left((q^{14}+q^{15}+q^{16}+q^{17}+q^{18})q^{6(M+2)}\right)\nonumber\\
&\quad+g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+3)\left(-(q^7+q^9+q^{10}+q^{12})q^{3(M+3)}+(q^{11}+q^{13})q^{6(M+3)}\right)\nonumber\\
&\quad+g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+4)\left(-(q^3+q^4+q^5+2q^6+q^7+q^8+q^9)q^{3(M+4)}\right)\nonumber\\
&\quad+g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+5)\left((q^4+q^6)-(q+q^2+q^3+q^4+q^6)q^{3(M+5)}\right)\nonumber\\
&\quad+g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+6)\left(1-q^{3(M+6)}\right).\label{eq:T11-rec1}
\end{align}
\subsection{Recurrence for $\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$}
Notice that for $M\ge 0$,
$$\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)=\sum_{n\le \frac{M}{2}}\frac{q^{(M-2n)^2+3n^2+3n(M-2n)}}{(q;q)_{M-2n}(q^3;q^3)_n}.$$
The recurrence satisfied by $\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ can be computed automatically by the \texttt{qMultiSum} package with the following codes:
\begin{lstlisting}[language=Mathematica]
ClearAll[M];
summand = q^(3n^2+(M-2n)^2+3n(M-2 n))/(qPochhammer[q,q,M-2n] qPochhammer[q^3,q^3,n]);
stru = qFindStructureSet[summand, {M}, {n}, {1}, {2}, {2}, qProtocol -> True]
rec = qFindRecurrence[summand, {M}, {n}, {1}, {2}, {2}, qProtocol -> True, StructSet -> stru[[1]]]
sumrec = qSumRecurrence[rec]
\end{lstlisting}
This gives us, for $M\ge 0$,
{\footnotesize\begin{align}
0&=\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)q^{9M+24}(1+2q^2+q^4+q^{3M+14})\nonumber\\
&\quad+\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+1)q^{6M+21}(1+2q^2+q^4-q^{3M+8}-q^{3M+10}+q^{3M+11}+q^{3M+13}+q^{3M+14})\nonumber\\
&\quad+\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+2)q^{6M+22}(1+q^2)(1+q^2+q^3+q^4+q^{3M+12})\nonumber\\
&\quad-\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+3)q^{3M+12}(1+q^2)(1-q+q^2)(1+q+q^2+q^3+q^{3M+12})\nonumber\\
&\quad-\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+4)q^{3M+12}(1-q+q^2)(1+q+q^2)(1+q+q^2+q^3+q^{3M+13})\nonumber\\
&\quad+\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M+5)(1-q^{3M+15})(1+2q^2+q^4+q^{3M+11}).\label{eq:T11-rec2}
\end{align}}
\subsection{Recurrence for $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)-\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$}
Finally, we deduce the recurrence for $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)-\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ from \eqref{eq:T11-rec1} and \eqref{eq:T11-rec2}. This can be accomplished by the \texttt{QREPlus} function of the \texttt{qGeneratingFunctions} package.
We need the following codes, in which \texttt{sumrec1} records the recurrence relation for $g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$ and \texttt{sumrec2} records the recurrence relation for $\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$.
\begin{lstlisting}[language=Mathematica]
ClearAll[M];
sumrec1 = {SUM[M] ((q^(28)+q^(30))q^(9M))
+ SUM[M+1] ((q^(19)+q^(21))q^(6(M+1))+q^(27)q^(9(M+1)))
+ SUM[M+2] ((q^(14)+q^(15)+q^(16)+q^(17)+q^(18))q^(6(M+2)))
+ SUM[M+3] (-(q^7+q^9+q^(10)+q^(12))q^(3(M+3))+(q^(11)+q^(13))q^(6(M+3)))
+ SUM[M+4] (-(q^3+q^4+q^5+2q^6+q^7+q^8+q^9)q^(3(M+4)))
+ SUM[M+5] ((q^4+q^6)-(q+q^2+q^3+q^4+q^6)q^(3(M+5)))
+ SUM[M+6] (1-q^(3(M+6)))
== 0};
sumrec2 = {SUM[M] q^(9M+24) (1+2q^2+q^4+q^(3M+14))
+ SUM[M+1] q^(6M+21) (1+2q^2+q^4-q^(3M+8)-q^(3M+10)+q^(3M+11)+q^(3M+13)+q^(3M+14))
+ SUM[M+2] q^(6M+22) (1+q^2) (1+q^2+q^3+q^4+q^(3M+12))
- SUM[M+3] q^(3M+12) (1+q^2) (1-q+q^2) (1+q+q^2+q^3+q^(3M+12))
- SUM[M+4] q^(3M+12) (1-q+q^2) (1+q+q^2) (1+q+q^2+q^3+q^(3M+13))
+ SUM[M+5] (1-q^(3M+15)) (1+2q^2+q^4+q^(3M+11))
== 0};
QREPlus[sumrec1, sumrec2, SUM[M]]
\end{lstlisting}
The output gives us an order six recurrence. Hence, to show
$$g_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)=\tilde{g}_{\mathscr{P}_{T_{\mathrm{I},1}}}(M)$$
for all $M\ge 0$, it suffices to show that the equality holds for $M=0,\ldots,5$. This can be checked easily.
We therefore arrive at
$$G_{\mathscr{P}_{T_{\mathrm{I},1}}}(x,q)=\sum_{n_1,n_2\ge 0}\frac{q^{n_1^2+3n_2^2+3n_1n_2}x^{n_1+2n_2}}{(q;q)_{n_1} (q^3;q^3)_{n_2}}.$$
\subsection{Other identities}
Similar to \eqref{eq:T11-new} and \eqref{eq:T11-NewNew}, let us write
$$G_{\mathscr{P}_{T_{*}}}(x,q)=\sum_{M\ge 0}g_{\mathscr{P}_{T_{*}}}(M)x^M$$
and the multi-summations on the right hand sides of \eqref{eq:real-gf-t12}--\eqref{eq:real-gf-t43} as
$$\sum_{M\ge 0}\tilde{g}_{\mathscr{P}_{T_{*}}}(M)x^M,$$
where ``$*$'' may be ``$\mathrm{I},2$'', ``$\mathrm{I},3$'', etc. We list the orders of recurrences satisfied by $g_{\mathscr{P}_{T_{*}}}(M)$, $\tilde{g}_{\mathscr{P}_{T_{*}}}(M)$ and $g_{\mathscr{P}_{T_{*}}}(M)-\tilde{g}_{\mathscr{P}_{T_{*}}}(M)$ in Table \ref{ta:order} for the reader's convenience.
{\footnotesize\begin{table}[ht]\caption{Orders of recurrences satisfied by $g_{\mathscr{P}_{T_{*}}}(M)$, $\tilde{g}_{\mathscr{P}_{T_{*}}}(M)$ and $g_{\mathscr{P}_{T_{*}}}(M)-\tilde{g}_{\mathscr{P}_{T_{*}}}(M)$}\label{ta:order}
\centering
\begin{tabular}{cccccccccccc}
\hline
$*$ & $\mathrm{I},2$ & $\mathrm{I},3$ & $\mathrm{II},1$ & $\mathrm{II},2$ & $\mathrm{II},a$ & $\mathrm{III},1$ & $\mathrm{III},2$ & $\mathrm{III},a$ & $\mathrm{IV},1$ & $\mathrm{IV},a$ & $\mathrm{IV},b$\\
$g_{\mathscr{P}_{T_{*}}}$ & 6 & 6 & 6 & 6 & 6 & 15 & 15 & 15 & 15 & 15 & 15\\
$\tilde{g}_{\mathscr{P}_{T_{*}}}$ & 5 & 5 & 5 & 5 & 5 & 4 & 4 & 4 & 4 & 4 & 4\\
$g_{\mathscr{P}_{T_{*}}}-\tilde{g}_{\mathscr{P}_{T_{*}}}$ & 6 & 6 & 6 & 6 & 6 & 15 & 15 & 15 & 15 & 15 & 15\\
\hline
\end{tabular}
\end{table}}
\section{Closing remarks}
In a very recent paper of Bringmann, Jennings-Shaffer and Mahlburg \cite{BJM2018}, along with other results, the Kanade--Russell conjectures $I_5$ and $I_6$ were proved. Here the analytic forms of $I_5$ and $I_6$ read respectively as
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{III},a}}}(1,q)&=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}+n_2+\frac{5n_3}{2}}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}\nonumber\\
&=\frac{1}{ ( q, q^3, q^4, q^6, q^7, q^{10}, q^{11} ; q^{12})_\infty },\\[10pt]
G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(1,q)&=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{3n_1}{2}+3n_2+\frac{7n_3}{2}}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}\nonumber\\
&=\frac{1}{ ( q^2, q^3, q^5, q^6, q^7, q^8, q^{11} ; q^{12})_\infty }.
\end{align}
The authors of \cite{BJM2018} cleverly reformulated $G_{\mathscr{P}_{T_{\mathrm{III},a}}}(1,q)$ and $G_{\mathscr{P}_{T_{\mathrm{IV},b}}}(1,q)$ and then added a new parameter so that the new bivariate generating functions satisfy simpler $q$-difference equations, from which the authors deduced the above identities.
Recall the standard notation for basic hypergeometric series:
$${}_{r+1}\phi_r\left(\begin{matrix} a_0,a_1,a_2\ldots,a_r\\ b_1,b_2,\ldots,b_r \end{matrix}; q, z\right):=\sum_{n\ge 0}\frac{(a_0;q)_n(a_1;q)_n\cdots(a_r;q)_n}{(q;q)_n(b_1;q)_n\cdots (b_r;q)_n} z^n.$$
Following the proofs of (1.15) and (1.16) in \cite{BJM2018}, one may prove the following identities with no difficulty.
\begin{theorem}
We have
\begin{align}
G_{\mathscr{P}_{T_{\mathrm{III},1}}}(1,q)&=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}-n_2-\frac{n_3}{2}}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}\nonumber\\
&=(-q;q)_\infty (-q^3;q^6)_\infty\; {}_{2}\phi_1\left(\begin{matrix} q^{-1},q\\ q^2 \end{matrix}; q^6, -q^3\right),\\[10pt]
G_{\mathscr{P}_{T_{\mathrm{III},2}}}(1,q)&=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{3n_1}{2}+n_2+\frac{5n_3}{2}}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}\nonumber\\
&=(-q^2;q)_\infty (-q^3;q^6)_\infty\; {}_{2}\phi_1\left(\begin{matrix} q,q^5\\ q^8 \end{matrix}; q^6, -q^3\right),\\[10pt]
G_{\mathscr{P}_{T_{\mathrm{IV},1}}}(1,q)&=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}-n_2+\frac{n_3}{2}}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}\nonumber\\
&=(-q;q)_\infty (-q^3;q^6)_\infty\; {}_{2}\phi_1\left(\begin{matrix} q^{-1},q\\ q^4 \end{matrix}; q^6, -q^3\right),\\[10pt]
G_{\mathscr{P}_{T_{\mathrm{IV},a}}}(1,q)&=\sum_{n_1,n_2,n_3\ge 0}\frac{q^{\frac{n_1^2}{2}+3n_2^2+\frac{9n_3^2}{2}+2 n_1 n_2+6n_2 n_3+3n_3n_1+\frac{n_1}{2}+n_2+\frac{n_3}{2}}}{(q;q)_{n_1} (q^2;q^2)_{n_2} (q^3;q^3)_{n_3}}\nonumber\\
&=(-q;q)_\infty (-q^3;q^6)_\infty\; {}_{2}\phi_1\left(\begin{matrix} q,q^5\\ q^4 \end{matrix}; q^6, -q^3\right).
\end{align}
\end{theorem}
Note that we shall use a refinement of Proposition 2.4 in \cite{BJM2018}, the proof of which comes from a slight modification of the original proof of Bringmann, Jennings-Shaffer and Mahlburg.
\begin{proposition}
Suppose that $A(x)=\sum_{n\ge0}\alpha_nx^n$ has positive radius of convergence
and $A(x)$ satisfies
\begin{align*}
A(x)&=\left(1+q^a+x^2q^b+x^2q^c\right)A\left(xq^3\right)\nonumber\\
&\quad -q^a\left(1+x^2q^{b+c-a-d+6}\right)\left(1+x^2q^{d}\right)A\left(xq^6\right),
\end{align*}
where $a\not\in 3\mathbb{Z}$ if $a\le-6$.
Then
\begin{align}
A(x)&=\alpha_0\left(-x^2q^{d-6};q^6\right)_\infty\;\sum_{n\ge0}\frac{ \left( q^{b-d+6},q^{c-d+6};q^6 \right)_n (-1)^nq^{(d-6)n} }{\left( q^{6},q^{a+6};q^6 \right)_n}x^{2n}\nonumber\\
&\quad+\alpha_1 \left(-x^2q^{d-6};q^6\right)_\infty\;\sum_{n\ge0}\frac{ \left( q^{b-d+9},q^{c-d+9};q^6 \right)_n (-1)^nq^{(d-6)n} }{\left( q^{9}, q^{a+9};q^6 \right)_n}x^{2n+1}.
\end{align}
\end{proposition}
\subsection*{Acknowledgements}
We would like to thank George Andrews for helpful discussions and for sharing his conjecture. We also want to thank the referees for their detailed comments on an earlier version of this paper.
\bibliographystyle{amsplain}
|
\section{Introduction
Large scale object detection datasets like ImageNet~\cite{imagenet}, MS-COCO~\cite{coco} and CrowdHuman~\cite{crowdhuman} try to define the ground truth bounding boxes as clear as possible.
However, we observe that the ground-truth bounding boxes are inherently ambiguous in some cases. The ambiguities\ makes it hard to label and hard to learn the bounding box regression function.
Some inaccurately labeled bounding boxes from MS-COCO are shown in Figure~\ref{fig:dataset} (a)(c). When the object is partially occluded, the bounding box boundaries are even more unclear, shown in Figure~\ref{fig:dataset} (d) from YouTube-BoundingBoxes~\cite{youtube}.
Object detection is a multi-task learning problem consisting of object localization and object classification. Current state-of-the-art object detectors (\eg, Faster R-CNN~\cite{faster}, Cascade R-CNN~\cite{cascade} and Mask R-CNN~\cite{mask}) rely on bounding box regression to localize objects.
However, the traditional bounding box regression loss (\ie, the smooth L1 loss~\cite{rcnn}) does not take such the ambiguities of the ground truth bounding boxes into account.
Besides, bounding box regression is assumed to be accurate when the classification score is high, which is not always the case, illustrated in Figure~\ref{fig:intro}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{assets/intro.pdf}
\caption{Illustration of failure cases of VGG-16 Faster R-CNN on MS-COCO. (a) both candidate boxes are inaccurate in a certain coordinate. (b) the left boundary of the bounding box which has the higher classification score is inaccurate. (\textit{better viewed in color})
}
\label{fig:intro}
\end{center}
\end{figure}
To address these problems, we propose a novel bounding box regression loss, namely KL Loss, for learning bounding box regression and localization uncertainty\ at the same time. Specifically, to capture the uncertainties\ of bounding box prediction, we first model the bounding box prediction and ground-truth bounding box as Gaussian distribution and Dirac delta function respectively. Then the new bounding box regression loss is defined as the KL divergence of the predicted distribution and ground-truth distribution. Learning with KL Loss\ has three benefits: (1) The ambiguities\ in a dataset can be successfully captured. The bounding box regressor gets smaller loss from ambiguous bounding boxes. (2) The learned variance is useful during post-processing. We propose var voting\ (variance voting) to vote the location of a candidate box using its neighbors' locations weighted by the predicted variances during non-maximum suppression (NMS). (3) The learned probability distribution is interpretable. Since it reflects the level of uncertainty\ of the bounding box prediction, it can potentially be helpful in down-stream applications like self-driving cars and robotics~\cite{djuric2018motion,gualtieri2018learning,he2017vehicle}.
To demonstrate the generality of KL Loss\ and var voting, we evaluate various CNN-based object detectors on both PASCAL VOC 2007 and MS-COCO including VGG-CNN-M-1024, VGG-16, ResNet-50-FPN, and Mask R-CNN. Our experiments suggest that our approach offers better object localization accuracy for CNN-based object detectors. For VGG-16 Faster R-CNN on MS-COCO, we improve the AP from 23.6\% to \textbf{29.1\%}, with only 2ms increased inference latency on the GPU (GTX 1080 Ti). Furthermore, we apply this pipeline to ResNet-50-FPN Mask R-CNN and improve the AP and AP$^{90}$ by \textbf{1.8\%} and \textbf{6.2\%} respectively, which outperforms the previous state-of-the-art bounding box refinement algorithm~\cite{iounet}.
\section{Related Work
\paragraph{Two-stage Detectors:} Although one-stage detection algorithms~\cite{ssd,yolo,cornernet,fsaf} are efficient, state-of-the-art object detectors are based on two-stage, proposal-driven mechanism~\cite{faster,rfcn,dcn,mask,light,cascade}. Two-stage detectors generate cluttered object proposals, which result in a large number of duplicate bounding boxes. However, during the standard NMS procedure, bounding boxes with lower classification scores will be discarded even if their locations are accurate. Our var voting\ tries to utilize neighboring bounding boxes based on localization confidence for better localization of the selected boxes.
\paragraph{Object Detection Loss:} To better learn object detection, different kind of losses have been proposed. UnitBox~\cite{unitbox} introduced an Intersection over Union (IoU) loss function for bounding box prediction. Focal Loss~\cite{fl} deals with the class imbalance by changing the standard cross entropy loss such that well-classified examples are assigned lower weights. \cite{rao2018learning} optimizes for the mAP via policy gradient for learning globally optimized object detector. \cite{uncertainties} introduces uncertainties for depth estimation. The idea is further extended to the 3D object detection~\cite{feng2018leveraging,feng2018towards}. \cite{kendall2017multi} proposes to weight multi-task loss for scene understanding by considering the uncertainty of each task.
With KL Loss, our model can adaptively adjust variances for the boundaries of every object during training, which can help to learn more discriminative features.
\paragraph{Non-Maximum Suppression:} NMS has been an essential part of computer vision for many decades. It is widely used in edge detection~\cite{rosenfeld1971edge}, feature point detection~\cite{lowe2004distinctive} and objection detection~\cite{rcnn,fast,faster,rothe2014non}.
Recently, soft NMS and learning NMS~\cite{softnms,nmslearn} are proposed to improve NMS results. Instead of eliminating all lower scored surrounding bounding boxes, soft-NMS~\cite{softnms} decays the detection scores of all other neighbors as a continuous function of their overlap with the higher scored bounding box. Learning NMS~\cite{nmslearn} proposed to learn a new neural network to perform NMS using only boxes and their classification scores.
\paragraph{Bounding Box Refinement:} MR-CNN~\cite{nmsavg} is first proposed to merge boxes during iterative localization. Relation network~\cite{relationnet} proposes to learn the relation between bounding boxes. Recently, IoU-Net~\cite{iounet} proposes to learn the IoU between the predicted bounding box and the ground-truth bounding box. IoU-NMS is then applied to the detection boxes, guided by the learned IoU. Different from IoU-Net, we propose to learn the localization variance from a probabilistic perspective. It enables us to learn the variances for the four coordinates of a predicted bounding box separately instead of only IoU.
Our var voting\ determine the new location of a selected box based on the variances of neighboring bounding boxes learned by KL Loss, which can work together with soft-NMS (Table~\ref{tab:cocovgg} and Table \ref{tab:cocofpn}).
\section{Approach
In this section, we first introduce our bounding box parameterization. Then we propose KL Loss\ for training detection network with localization confidence. Finally, a new NMS approach is introduced for improving localization accuracy with our confidence estimation.
\subsection{Bounding Box Parameterization}
Based on a two-stage object detector Faster R-CNN or Mask R-CNN~\cite{faster,mask} shown in Figure~\ref{fig:arch}, we propose to regress the boundaries of a bounding box separately.
Let $(x_1, y_1, x_2, y_2) \in \mathcal{R}^4$ be the bounding box representation as a 4-dimensional vector, where each dimension is the box boundary location. We adopt the parameterizations of the $(x_1, y_1 ,x_2, y_2)$ coordinates instead of the $(x, y, w, h)$ coordinates used by R-CNN~\cite{rcnn}:
\begin{equation} \label{eq:box_encoding}
\begin{aligned}
& t_{x_1} =\frac{x_1 - x_{1a}}{w_a}, t_{x_2} =\frac{x_2-x_{2a}}{w_a}\\
& t_{y_1} =\frac{y_1 - y_{1a}}{h_a}, t_{y_2} =\frac{y_2-y_{2a}}{h_a}\\
& t_{x_1}^* =\frac{x_1^* - x_{1a}}{w_a}, t_{x_2}^* =\frac{x_2^*-x_{2a}}{w_a}\\
& t_{y_1}^* =\frac{y_1^* - y_{1a}}{h_a}, t_{y_2}^* =\frac{y_2^*-y_{2a}}{h_a}\\
\end{aligned}
\end{equation}
where $t_{x_1}$, $t_{y_1}$, $t_{x_2}$, $t_{y_2}$ are the predicted offsets. $t_{x_1}^*$, $t_{y_1}^*$, $t_{x_2}^*$, $t_{y_2}^*$ are the ground-truth offsets. $x_{1a}$, $x_{2a}$, $y_{1a}$, $y_{2a}$, $w_a$, $h_a$ are from the anchor box. $x_1$, $y_1$, $x_2$, $y_2$ are from the predicted box.
In the following discussions, a bounding box coordinate is denoted as $x$ for simplicity because we can optimize each coordinate independently.
We aim to estimate the localization confidence along with the location. Formally, our network predicts a probability distribution instead of only bounding box location. Though the distribution could be more complex ones like multivariate Gaussian or a mixture of Gaussians, in this paper we assume the coordinates are independent and use single variate gaussian for simplicity:
\begin{equation}\label{eq:gaussian}
P_\Theta(x) = \frac{1}{\sqrt{2\pi\sigma^2}} e^{-\frac{(x-x_e)^2}{2\sigma^2}}
\end{equation}
where $\Theta$ is the set of learnable parameters. $x_e$ is the estimated bounding box location. Standard deviation $\sigma$ measures uncertainty of the estimation. When $\sigma \rightarrow 0$, it means our network is extremely confident about estimated bounding box location. It is produced by a fully-connected layer on top of the fast R-CNN head (\verb|fc7|). Figure~\ref{fig:arch} illustrates the fast R-CNN head of our network architecture for object detection.
The ground-truth bounding box can also be formulated as a Gaussian distribution, with $\sigma \rightarrow 0$, which is a Dirac delta function:
\begin{equation}\label{eq:delta}
P_D(x) = \delta(x - x_g)
\end{equation}
where $x_g$ is the ground-truth bounding box location.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{assets/arch.pdf}
\caption{Our network architecture for estimating localization confidence. Different from standard fast R-CNN head of a two stage detection network, our network esitmates standard deviations along with bounding box locations, which are taken into account in our regression loss KL Loss}
\label{fig:arch}
\end{center}
\end{figure}
\subsection{Bounding Box Regression with KL Loss
The goal of object localization in our context is to estimate $\hat{\Theta}$ that minimize the KL-Divergence between $P_\Theta(x)$ and $P_D(x)$~\cite{ml} over $N$ samples:
\begin{equation}
\hat{\Theta} = \argmin_\Theta \frac{1}{N} \sum D_{KL}(P_D(x)||P_\Theta(x))
\end{equation}
We use the KL-Divergence as the loss function $L_{reg}$ for bounding box regression. The classification loss $L_{cls}$ remains the same. For a single sample:
\begin{equation}
\begin{aligned}
L_{reg} & = D_{KL}(P_D(x)||P_\Theta(x)) \\
& = \int P_D(x) \log P_D(x) \text{d}x - \int P_D(x) \log P_\Theta(x) \text{d}x\\
& = \frac{(x_g-x_e)^2}{2\sigma^2} + \frac{\log(\sigma^2)}{2} + \frac{\log(2\pi)}{2} - H(P_D(x))
\end{aligned}
\end{equation}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{assets/dist.pdf}
\caption{The Gaussian distributions in {\color{blue} blue} and {\color{Gray} gray} are our estimations. The Dirac delta function in {\color{Orange} orange} is the distribution of the ground-truth bounding box. When the location $x_e$ is estimated inaccurately, we expect the network to be able to predict larger variance $\sigma^2$ so that $L_{reg}$ will be lower (blue)}
\label{fig:dist}
\end{center}
\end{figure}
Shown in Figure~\ref{fig:dist}, when the location $x_e$ is estimated inaccurately, we expect the network to be able to predict larger variance $\sigma^2$ so that $L_{reg}$ will be lower.
$\log(2\pi)/2 $ and $ H(P_D(x)) $ do not depend on the estimated parameters $\Theta$, hence:
\begin{equation}
\begin{aligned}
& L_{reg} \propto \frac{(x_g-x_e)^2}{2\sigma^2} + \frac{1}{2}\log(\sigma^2)
\end{aligned}
\end{equation}
When $\sigma = 1$, KL Loss\ degenerates to the standard Euclidean loss:
\begin{equation}\label{eq:degenerate}
\begin{aligned}
L_{reg} \propto \frac{(x_g-x_e)^2}{2}
\end{aligned}
\end{equation}
The loss is differentiable w.r.t location estimation $x_e$ and localization standard deviation $\sigma$:
\begin{equation}\label{eq:gradient}
\begin{aligned}
\frac{\text{d}}{\text{d}x_e}L_{reg} &= \frac{x_e - x_g}{\sigma^2}\\
\frac{\text{d}}{\text{d}\sigma}L_{reg} &= -\frac{(x_e - x_g)^2}{\sigma^{3}} + \frac{1}{\sigma}
\end{aligned}
\end{equation}
However, since $\sigma$ is in the denominators, the gradient sometimes can explode at the beginning of training. To avoid gradient exploding, our network predicts $\alpha = \log(\sigma^2)$ instead of $\sigma$ in practice:
\begin{equation}\label{eq:l11}
\begin{aligned}
& L_{reg} \propto \frac{e^{-\alpha}}{2}(x_g-x_e)^2 + \frac{1}{2}\alpha\\
\end{aligned}
\end{equation}
We convert $\alpha$ back to $\sigma$ during testing.
For $|x_g-x_e| > 1$, we adopt a loss similar to the smooth $L_1$ loss defined in Fast R-CNN~\cite{fast}:
\begin{equation}\label{eq:l12}
L_{reg} = e^{-\alpha}(|x_g-x_e|-\frac{1}{2}) + \frac{1}{2}\alpha
\end{equation}
We initialize the weights of the FC layer for $\alpha$ prediction with random Gaussian initialization. The standard deviation and mean are set to 0.0001 and 0 respectively, so that KL Loss\ will be similar to the standard smooth L1 loss at the beginning of training. (Equation~\ref{eq:l11} and Equation~\ref{eq:l12}).
\subsection{Variance Voting}\label{sec:ournms
\begin{figure*}[t]
\begin{center}
\includegraphics[width=\linewidth]{assets/softer.pdf}
\caption{Results of var voting\ with VGG-16 Faster R-CNN on MS-COCO. The {\color{ForestGreen} green} textbox in the middle of each boundary is the corresponding standard deviation $\sigma$ we predicted (Equation~\ref{eq:gaussian}). Two failure situations corresponding to Figure~\ref{fig:intro} that can be improved by var voting: (a) When each candidate bounding box is inaccurate in some coordinates (women on the right), our var voting\ can incorporate their localization confidence and produce better boxes. (b) The bounding box with a higher classification score (train 0.99) actually has lower localization confidence than the bounding box with a lower classification score (train 0.35). After var voting, the box scored 0.99 moves towards the correct location. (\textit{better viewed in color})}
\label{fig:img}
\end{center}
\end{figure*}
After we obtain the variance of predicted location, it is intuitive to vote candidate bounding box location according to the learned variances of neighboring bounding boxes.
Shown in Algorithm~\ref{alg:nms}, we change NMS with three lines of code:
\begin{algorithm}[H]
\caption {var voting}{$\mathcal{B}$ is $N\times4$ matrix of initial detection boxes.
$\mathcal{S}$ contains corresponding detection scores.
$\mathcal{C}$ is $N\times4$ matrix of corresponding variances. $\mathcal{D}$ is the final set of detections.
$\sigma_t$ is a tunable parameter of var voting. The lines in {\color{blue}blue} and in {\color{ForestGreen}green} are soft-NMS and var voting\ respectively.}\label{alg:nms}
\begin{algorithmic}
\State $\mathcal{B} = \{b_1, ..,b_N\} $, $\mathcal{S} = \{s_1, ..,s_N\}$, $\mathcal{C} = \{\sigma^2_1, ..,\sigma^2_N\}$
\State$\mathcal{D} \leftarrow \{\}$
\State$\mathcal{T} \leftarrow \mathcal{B}$
\While {$\mathcal{T} \ne$ empty}
\State$m \leftarrow$ argmax $\mathcal{S} $
\State $ \mathcal{T} \leftarrow \mathcal{T} - b_m$
\State\textcolor{blue}{$ \mathcal{S} \leftarrow \mathcal{S} f(IoU(b_m, T)) $} \Comment{soft-NMS}
\textcolor{ForestGreen}{
\State$idx \leftarrow IoU(b_m, B) > 0$} \Comment{var voting
\textcolor{ForestGreen}{
\State$p \leftarrow exp(-(1-IoU(b_m, \mathcal{B}[idx]))^2/\sigma_t)$}
\textcolor{ForestGreen}{
\State $b_m \leftarrow p(\mathcal{B}[idx]/\mathcal{C}[idx])/p(1/\mathcal{C}[idx])$}
\State$\mathcal{D} \leftarrow \mathcal{D} \bigcup b_m
\EndWhile
\State\textbf{return} {$\mathcal{D}, \mathcal{S}$}
\end{algorithmic}
\end{algorithm}
We vote the location of selected boxes within the loop of standard NMS or soft-NMS~\cite{softnms}. After selecting the detection with maximum score $b$,
$\{x_1, y_1, x_2, y_2, s, \sigma_{x_1}, \sigma_{y_1}, \sigma_{x_2}, \sigma_{y_2}\}$, its new location is computed according to itself and its neighboring bounding boxes.
Inspired by soft-NMS, we assign higher weights for boxes that are closer and have lower uncertainties. Formally, let $x$ be a coordinate (\eg, $x_1$) and $x_i$ be the coordinate of $i$th box. The new coordinate is computed as follow:
\begin{equation}
\begin{aligned}
p_i &= e^{-(1-IoU(b_i,b))^2/\sigma_t}\\
x &= \frac{\sum_i p_ix_i / \sigma_{x,i}^2}{\sum_i p_i/\sigma_{x,i}^2}\\
& \text{subject\ to } IoU(b_i, b) > 0
\end{aligned}
\end{equation}
$\sigma_t$ is a tunable parameter of var voting. Two types of neighboring bounding boxes will get lower weights during voting: (1)
Boxes with high variances. (2) Boxes that have small IoU with the selected box. Classification score is not involved in the voting, since lower scored boxes may have higher localization confidence. In Figure~\ref{fig:img}, we provide a visual illustration of var voting. With var voting, the two situations as mentioned earlier in Figure~\ref{fig:intro} that lead to detection failure can sometimes be avoided.
\section{Experiments
To demonstrate our method for accurate object detection, we use two datasets: MS-COCO~\cite{coco} and PASCAL VOC 2007~\cite{pascal-voc-2007}.
We use four GPUs for our experiments. The training schedule and batch size are adjusted according to the linear scaling rule~\cite{hour}.
For VGG-CNN-M-1024 and VGG-16 Net~\cite{vgg}, our implementation is based on Caffe~\cite{caffe}.
For ResNet-50 FPN~\cite{resnet,fpn} and Mask R-CNN~\cite{mask}, our implementation is based on Detectron~\cite{detectron}.
For VGG-16~\cite{vgg} Faster R-CNN, following \verb|py-faster-rcnn|\footnote{github.com/rbgirshick/py-faster-rcnn}, we train on \verb|train2014| and test on \verb|val2014|. For other object detection networks, we train and test on the newly defined \verb|train2017| and \verb|val2017| respectively. We set $\sigma_t$ to 0.02. Unless specified, all hyper-parameters are set to default\footnote{github.com/facebookresearch/Detectron}.
\begin{table*}[t]
\begin{center}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c|c}
\hline\hline
KL Loss & soft-NMS & var voting & AP & AP$^{50}$ & AP$^{75}$ & AP$^S$ & AP$^M$ & AP$^L$ & AR$^{1}$ & AR$^{10}$ & AR$^{100}$ \\ \hline\hline
& & &23.6 & 44.6& 22.8 &6.7 &25.9 &36.3 &23.3 &33.6 & 34.3 \\ \hline
&\checkmark & & 24.8 & 45.6 & 24.6 & 7.6 & 27.2 & 37.6 & 23.4 & 39.2 & 42.2 \\ \hline
\checkmark & & & 26.4 & 47.9 & 26.4 & 7.4 & 29.3 & 41.2 & 25.2 & 36.1 & 36.9 \\ \hline
\checkmark & &\checkmark & 27.8 & 48.0 & 28.9 & 8.1 & 31.4 & 42.6 & 26.2 & 37.5 & 38.3 \\ \hline
\checkmark &\checkmark & & 27.8 & 49.0 & 28.5 & 8.4 & 30.9 & 42.7 & 25.3 & 41.7 & 44.9 \\ \hline
\checkmark &\checkmark &\checkmark &\textbf{29.1} &\textbf{49.1} &\textbf{30.4} &\textbf{8.7} &\textbf{32.7} &\textbf{44.3} &\textbf{26.2} &\textbf{42.5}&\textbf{45.5} \\ \hline
\end{tabular}
\caption{The contribution of each element in our detection pipeline on MS-COCO. The baseline model is VGG-16 Faster R-CNN}
\label{tab:cocovgg}
\end{center}
\end{table*}
\begin{table}[t]
\begin{center}
\begin{tabular}{c|c}
\hline\hline
method & latency (ms) \\ \hline
baseline & 99 \\ \hline
ours & 101 \\ \hline
\end{tabular}
\end{center}
\caption{Inference time comparison on MS-COCO with VGG-16 Faster R-CNN on a GTX 1080 Ti GPU, CUDA 8~\cite{cuda} and CUDNN 6~\cite{cudnn}}
\label{tab:cocotime}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{l|c|c|l}
\hline\hline
fast R-CNN head & backbone & KL Loss & \multicolumn{1}{c}{AP} \\ \hline
\multirow{2}{*}{2mlp head} & \multirow{2}{*}{FPN} & & 37.9 \\
& & \checkmark & 38.5$^{+0.6}$ \\ \hline
\multirow{2}{*}{2mlp head + mask} & \multirow{2}{*}{FPN} & & 38.6 \\
& & \checkmark & 39.5$^{+\textbf{0.9}}$ \\ \hline
\multirow{2}{*}{conv5 head} & \multirow{2}{*}{RPN} & & 36.5 \\
& & \checkmark & 38.0$^{+\textbf{1.5}}$ \\ \hline
\end{tabular}
\caption{Comparison of different fast R-CNN heads. The model is ResNet-50 Faster R-CNN}
\label{tab:c5}
\end{center}
\end{table}
\subsection{Ablation Study}
We evaluate the contribution of each element in our detection pipeline:
KL Loss, soft-NMS and var voting\ with VGG-16 Faster R-CNN. The detailed results are shown in Table~\ref{tab:cocovgg}.
\paragraph{KL Loss:} Surprisingly, simply training with KL Loss\ greatly improves the AP by \textbf{2.8\%}, which is also observed on ResNet-50 Faster R-CNN and Mask R-CNN (\textbf{1.5\%} and \textbf{0.9\%} improvement respectively, shown in Table~\ref{tab:c5} and Table~\ref{tab:mask}). First, by learning to predict high variances for samples with high uncertainties during training, the network can learn more from useful samples. Second, the gradient for localization can be adaptively controlled by the network during training (Equation~\ref{eq:gradient}), which encourages the network to learn more accurate object localization. Third, KL Loss\ incorporates learning localization confidence which can potentially help the network to learn more discriminative features.
The learned variances through our KL Loss\ are interpretable. Our network will output higher variances for challenging object boundaries, which can be useful in vision applications like self-driving cars and robotics. The first row of Figure~\ref{fig:img} shows some qualitative examples of the standard deviation learned through our KL Loss.
\paragraph{Soft-NMS:} As expected, soft-NMS performs consistently on both baseline and our network trained with KL Loss. It improves the AP by 1.2\% and 1.4\% on the baseline and our network respectively, shown in Table~\ref{tab:cocovgg}.
\paragraph{Variance Voting:} Finally, with var voting, the AP is further improved to \textbf{29.1\%}. We made the observation that improvement mainly comes from the more accurate localization. Notice that the AP$^{50}$ is only improved by 0.1\%. However, AP$^{75}$, AP$^{M}$ and AP$^{L}$ are improved by 1.8\%, 1.8\%, and 1.6\% respectively, shown in Table~\ref{tab:cocovgg}. This indicates that classification confidence is not always associated with localization confidence. Therefore, learning localization confidence apart from classification confidence is important for more accurate object localization.
We also found that var voting\ and soft-NMS can work well together. Applying var voting\ with the standard NMS improves the AP by 1.4\%. Applying var voting\ after soft-NMS still can improve the AP by 1.3\%. We argue that soft-NMS is good at scoring candidate bounding boxes which improve overall performance, whereas var voting\ is good at refining those selected bounding boxes for more accurate object localization. The second row of Figure~\ref{fig:img} shows some qualitative examples of our var voting.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\linewidth]{assets/sigma.pdf}
\caption{Varying $\sigma_t$ for var voting\ with ResNet-50 Faster R-CNN. (\textit{better viewed in color})}
\label{fig:sigma}
\end{center}
\end{figure}
Shown in Figure~\ref{fig:sigma}, we test the sensitivity of the tunable parameter $\sigma_t$ for var voting. When $\sigma_t=0$, var voting\ is not activated. We observe that the AP$^{75}$, AP$^{80}$ and AP$^{90}$ can be significantly affected by $\sigma_t$, while AP$^{50}$ is less sensitive to $\sigma_t$. Acceptable values of $\sigma_t$ varies from around $0.005 \sim 0.05$. We use $\sigma_t=0.02$ in all experiments.
\paragraph{Inference Latency:}
We also evaluate the inference time of our improved VGG-16 Faster R-CNN on a single GTX 1080 Ti GPU with CUDA 8 and CUDNN 6, as it is crucial for resource-limited applications~\cite{shufflenet,addressnet,cp,amc,lightweight}.
Shown in Table~\ref{tab:cocotime}, our approach only increases \textbf{2ms} latency on GPU. Different from IoUNet~\cite{iounet} which uses \verb|2mlp| head for IoU prediction, our approach only requires a $4096\times324$ fully-connected layer for the localization confidence prediction.
\paragraph{RoI Box Head:}
We test the effectiveness of KL Loss\ with different RoI box heads on a deeper backbone: ResNet-50. \verb|res5/conv5| head consists of 9 convolutional layers which can be applied to each RoI as fast R-CNN head. \verb|2mlp| head consists of two fully connected layers. \verb|res5| head can learn more complex representation than the commonly used \verb|2mlp| head. Shown in Table~\ref{tab:c5}, KL Loss\ can improve the AP by \textbf{0.9\%} with mask. KL Loss\ can further improve the AP by \textbf{1.5\%} with \verb|conv5| head. We hypothesize that the localization variance is much more challenging to learn than localization, therefore KL Loss\ can benefit more from the expressiveness of \verb|conv5| head. Since \verb|conv5| head is not commonly used in recent state-of-the-art detectors, we still adopt the \verb|2mlp| head in the following experiments.
\begin{table*}[t]
\begin{center}
\begin{tabular}{l|l|lllll}
\hline\hline
& \multicolumn{1}{c|}{AP} & \multicolumn{1}{c}{AP$^{50}$} & \multicolumn{1}{c}{AP$^{60}$} & \multicolumn{1}{c}{AP$^{70}$} & \multicolumn{1}{c}{AP$^{80}$} & \multicolumn{1}{c}{AP$^{90}$} \\ \hline
baseline~\cite{detectron} & 38.6 & \textbf{59.8} & 55.3 & 47.7 & 34.4 & 11.3 \\
MR-CNN~\cite{nmsavg} & 38.9 & \textbf{59.8} & 55.5 & 48.1 & 34.8$^{+0.4}$ & 11.9$^{+0.6}$ \\
soft-NMS~\cite{softnms} & 39.3 & 59.7 & \textbf{55.6} & \textbf{48.9} &35.9$^{+1.5}$ &12.0$^{+0.7}$ \\
IoU-NMS+Refine~\cite{iounet} & 39.2 & 57.9 & 53.6 & 47.4 & 36.5$^{+2.1}$ & 16.4$^{+5.1}$ \\ \hline
KL Loss & 39.5$^{+0.9}$ & 58.9 & 54.4 & 47.6 & 36.0$^{+1.6}$ & 15.8$^{+4.5}$ \\
KL Loss+var voting & 39.9$^{+1.3}$ & 58.9 & 54.4 & 47.7 & 36.4$^{+2.0}$ & 17.0$^{+5.7}$ \\
KL Loss+var voting+soft-NMS & \textbf{40.4}$^{+1.8}$ & 58.7 & 54.6 & 48.5 & \textbf{37.5}$^{+3.3}$ & \textbf{17.5}$^{+6.2}$ \\ \hline
\end{tabular}
\caption{Comparisons of different methods for accurate object detection on MS-COCO. The baseline model is ResNet-50-FPN Mask R-CNN. We improve the baseline by $\approx$ 2\% in AP}
\label{tab:mask}
\end{center}
\end{table*}
\subsection{Accurate Object Detection}
Table~\ref{tab:mask} summarizes the performance of different methods for accurate object detection on ResNet-50-FPN Mask R-CNN. With KL Loss, the network can learn to adjust the gradient for ambiguous bounding boxes during training. As a result, Mask R-CNN trained with KL Loss\ performs significantly better than the baseline for high overlap metrics like AP$^{90}$. Variance Voting\ improves the localization results by voting the location according to the localization confidences of neighboring bounding boxes. AP$^{80}$ and AP$^{90}$ are further improved by 0.4\% and 1.2\% respectively. Variance Voting\ is also compatible with soft-NMS. Variance Voting\ combined with soft-NMS improves the AP$^{90}$ and the overall AP of the final model by \textbf{6.2\%} and \textbf{1.8\%} respectively. Compared with IoUNet~\cite{iounet}: (1) our variance and localization are learned together with KL Loss, which improves the performance. (2) KL Loss\ does not require a separate 2mlp head for learning localization confidence, which introduces nearly no additional computation. (3) var voting\ does not require iterative refinement, which is much faster.
\begin{table}[t]
\begin{center}
\begin{tabular}{c|l|c}
\hline\hline
backbone & \multicolumn{1}{c|}{method} & mAP \\ \hline
\multirow{4}{*}{} & baseline & 60.4 \\ \cline{2-3}
VGG-CNN- & KL Loss & 62.0 \\
M-1024 & KL Loss+var voting & 62.8 \\
& KL Loss+var voting+soft-NMS & \textbf{63.6} \\ \hline
\multirow{7}{*}{VGG-16}
& baseline & 68.7 \\
& QUBO (tabu)~\cite{qubo} & 60.6 \\
& QUBO (greedy)~\cite{qubo} & 61.9 \\
& soft-NMS~\cite{softnms} & 70.1 \\ \cline{2-3}
& KL Loss & 69.7 \\
& KL Loss+var voting & 70.2 \\
& KL Loss+var voting+soft-NMS & \textbf{71.6} \\ \hline
\end{tabular}
\caption{Comparisons of different approaches on PASCAL VOC 2007 with Faster R-CNN.}
\label{tab:vocvgg}
\end{center}
\end{table}
\begin{table*}[t]
\begin{center}
\begin{tabular}{c|l|c|ccccc}\hline\hline
type & \multicolumn{1}{c|}{method} & AP & AP$^{50}$ & AP$^{75}$ & AP$^S$ & AP$^M$& AP$^L$ \\ \hline
\multirow{7}{*}{fast R-CNN} & baseline (1x schedule)~\cite{detectron} &36.4 &\textbf{58.4} &39.3 &\textbf{20.3} &39.8 &48.1\\
& baseline (2x schedule)~\cite{detectron} &36.8 &\textbf{58.4} &39.5 &19.8 &39.5 &49.5 \\
& IoU-NMS~\cite{iounet} & 37.3 & 56.0 &- &- &- &- \\
& soft-NMS~\cite{softnms} &37.4 &58.2 &41.0 &\textbf{20.3} &40.2 &50.1\\ \cline{2-8}
& KL Loss &37.2 &57.2 &39.9 &19.8 &39.7 &50.1\\
& KL Loss+var voting &37.5 &56.5 &40.1 &19.4 &40.2 &51.6\\
& KL Loss+var voting+soft-NMS &\textbf{38.0} &56.4 &\textbf{41.2} &19.8 &\textbf{40.6} &\textbf{52.3}\\ \hline\hline
\multirow{9}{*}{Faster R-CNN} &baseline (1x schedule)~\cite{detectron} &36.7 &58.4 &39.6 &21.1 &39.8 &48.1\\
& IoU-Net~\cite{iounet} & 37.0 &58.3 & - & -&-&- \\
& IoU-Net+IoU-NMS~\cite{iounet} & 37.6& 56.2 &-&-&- &- \\
& baseline (2x schedule)~\cite{detectron} &37.9 &59.2 &41.1 &21.5 &41.1 &49.9 \\
& IoU-Net+IoU-NMS+Refine~\cite{iounet} &38.1 & 56.3&-&-&-&- \\
& soft-NMS\cite{softnms} &38.6 &\textbf{59.3} &42.4 &21.9 &\textbf{41.9} &50.7 \\ \cline{2-8}
& KL Loss & 38.5 & 57.8& 41.2 & 20.9 & 41.2 & 51.5\\
& KL Loss +var voting & 38.8 & 57.8& 41.6 & 21.0 & 41.5 & 52.0\\
& KL Loss +var voting+soft-NMS & \textbf{39.2} & 57.6& \textbf{42.5} & 21.2 & 41.8 & \textbf{52.5}\\ \hline
\end{tabular}
\caption{Performance comparison with FPN ResNet-50 on MS-COCO}
\label{tab:cocofpn}
\end{center}
\end{table*}
We further evaluate our approach on the feature pyramid network (ResNet-50 FPN)~\cite{fpn,resnet}, shown in Table~\ref{tab:cocofpn}.
For fast R-CNN version, training with KL Loss\ increases the baseline by 0.4\%. After applying var voting\ along with soft-NMS, our model achieves \textbf{38.0\%} in AP, which outperforms both IoU-NMS and soft-NMS baselines. Training end-to-end with KL Loss\ can help the network learn more discriminative features, which improves the baseline AP by \textbf{0.6\%}. The final model achieves 39.2\% in AP, which improves the baseline by 1.3\%.
\subsection{Experiments on PASCAL VOC 2007
Even though our approach is designed for large scale object detection, it could also generalize well on small datasets. We perform experiments with Faster R-CNN~\cite{faster} on PASCAL VOC 2007, which consists of about 5k \verb|voc_2007_trainval| images and 5k \verb|voc_2007_test| images over 20 object categories. Backbone networks: VGG-CNN-M-1024 and VGG-16 Net~\cite{vgg} are tested.
Shown in Table~\ref{tab:vocvgg}, we compare our approach with soft-NMS and quadratic unconstrained binary optimization (QUBO~\cite{qubo}).
For QUBO, we test both greedy and classical tabu solver (we manually tuned the penalty term for both solvers to get better performance). We observe that it is much worse than the standard NMS, though it was reported to be better for pedestrian detection. We hypothesize that QUBO is better at pedestrian detection since there are more occluded bounding boxes~\cite{crowdhuman}.
For VGG-CNN-M-1024, training with var voting\ improves the mAP by \textbf{1.6\%}. var voting\ further improves the mAP by \textbf{0.8\%}.
For VGG-16, our approach improves the mAP by \textbf{2.9\%}, combined with soft-NMS.
We notice that var voting\ could still improve performance even after soft-NMS is applied to the initial detection boxes. This observation is consistent with our experiments on MS-COCO (Table~\ref{tab:cocovgg}).
\section{Conclusion
To conclude, the uncertainties in large-scale object detection datasets can hinder the performance of state-of-the-art object detectors. Classification confidence is not always strongly related to localization confidence. In this paper, a novel bounding box regression loss with uncertainty is proposed for learning more accurate object localization. By training with KL Loss, the network learns to predict localization variance for each coordinate. The resulting variances empower var voting, which can refine the selected bounding boxes via voting. Compelling results are demonstrated for VGG-16 Faster R-CNN, ResNet-50 FPN and Mask R-CNN on both MS-COCO and PASCAL VOC 2007.
\section*{Acknowledgement
This research was partially supported by National Key R\&D Program of China (No.~2017YFA0700800).
We would love to express our appreciation to Prof.~Kris Kitani and Dr.~Jian Sun for the useful discussions during this research. \nocite{softernms,pts}
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
Sophisticated signal-processing often requires that Lorentz reciprocity---the scattering symmetry of an electromagnetic system under exchange of source and detector---be broken. In particular, directionally-routing propagating electromagnetic modes without adding noise or incurring loss is vital for quantum information processing with superconducting circuits.
Although Maxwell's equations place no restrictions on the bandwidth over which lossless non-reciprocity can occur, synthesis of low-loss on-chip circulators and gyrators has mostly been limited to narrow-band devices~\cite{sliwa:2015,lecocq:2017,chapman:2017b,abdo:2017} (fractional bandwidth $<1\%$). Ferrite circulators~\cite{fay:1965} provide octave bandwidths, but
their large magnets make them difficult to miniaturize or integrate with superconducting circuits.
Given the large number of qubits required for fault-tolerant quantum computation~\cite{fowler:2012}, and the standard use of non-reciprocity in qubit readout, circulator performance metrics such as loss, added noise, bandwidth, dynamic range, and size must be simultaneously optimized.
To address this optimization problem, we propose the superconducting implementation of a recent concept for broadband circulation based on active modulation and delay~\cite{galland:2013,yang:2014,chapman:2017c,lu:2018a,biedka:2018}.
The approach requires that the delay be commensurate with the modulation frequency, but imposes no constraints on the frequency of the input signal. This aspect is responsible for the concept's bandwidth, allowing for operation all the way down to dc~\cite{biedka:2018}. It improves on previous designs~\cite{reiskarimian:2016,biedka:2017,dinc:2017,rosenthal:2017} for broadband active circulators by engineering the desired circulation in a hardware efficient manner, which may in principle be lossless. Such schemes are complementary to superconducting travelling-wave isolators, which provide an alternate path to compact broadband non-reciprocity~\cite{ranzani:2017}.
When modeled as a superconducting integrated circuit, simulations with realistic parameters show our design allows for low-loss circulation ($<0.35$ dB) with octave bandwidth and nW scale power-handling. With respect to bandwidth and linearity, this represents a 50-fold improvement over other near-lossless superconducting circulators~\cite{sliwa:2015,lecocq:2017,chapman:2017b}, well-suited for integration with broadband quantum-limited amplifiers~\cite{macklin:2015,roy:2015,naaman:2017}.
\section{Theory of operation}
\label{sec:theory}
\begin{figure*}[hbt]
\begin{center}
\includegraphics[width=2.0\columnwidth]{CirculatorConceptCompact-01.eps}
\caption {\textbf{Circulation with actively modulated transfer switches and delays.} (a) A four-port circulator is constructed from two dynamically modulated transfer switches and a pair of delay lines. The left (blue) switch is shown in the crossed state, and the right (red) switch is shown in the through. The delay is set to one quarter of the modulation period $\tau = T/4$. (b) The left and right switches are modulated with period $T$ and a relative phase of $\pi/2$. For visual clarity, the modulation period is divided into quarter-period intervals indicated by the pastel background colors. (c) The schematics depict the way the switches will route a right-propagating signal, depending on its arrival time. Signals incident on port 1 are always routed to port 2, and signals incident on port 3 are routed to port 4. Note that these are not ``snap-shots'' of the switches' configurations, as the signal is delayed by a quarter-period of the modulation as it propagates through the network. (d) The same depiction as (c), but for left-travelling waves incident on ports 2 or 4. To keep the device's orientation the same as in (b), time flows from right to left in this diagram.
}
\label{fig:fig1}
\end{center}
\end{figure*}
The proposed circulator is composed of two transfer switches connected by a pair of delay lines (Fig.~\ref{fig:fig1}a). The transfer switches have four ports, and are dynamically modulated with period $T$, toggling between their ``through'' and ``crossed'' states. Figure~\ref{fig:fig1}b shows how the two switches are periodically actuated by square waves with a relative phase of $\theta = \pi/2$. The delays are chosen to have duration $\tau = T/4$.
To see how the circulation is created, consider how a right-propagating wave packet (e.g. a signal-burst incident on ports 1 or 3) will be scattered by the network. Figure~\ref{fig:fig1}c shows how the two switches will route the incident signal, depending on the portion of the modulation period in which it arrives. For example, if the signal arrives in the first quarter-period of the modulation (the yellow time-window), the left (blue) transfer switch will be in its through state. But by the time the signal exits the delay line, the right (red) transfer switch will be in its crossed state. Signals incident on port 1 are therefore routed to port 2, and signals incident on port 3 are routed to port 4.
This process may be continued for the entire modulation period. One observes that the port-to-port signal-routing is independent of the incident signal's arrival time, despite the fact that the signal may travel along different paths before exiting the network.
The non-reciprocal character of the network may be verified by repeating this analysis for left-travelling wave packets (e.g. signals incident on ports 2 or 4), as depicted in Fig.~\ref{fig:fig1}d. Again, the port-to-port signal-routing is independent of the signal's arrival time: signals incident on port 2 are always routed to port 3, and signals incident on port 4 are always routed to port 1. The scattering of the network is not invariant upon exchange of source and detector. It forms a four-port circulator.
\begin{comment}
This graphical representation of the device's circulation may be complemented by a simple quantitative analysis. Let $H_l(t)$ and $H_r(t)$ represent the periodically modulated states of the left and right transfer switches, which take values 1 and -1 when the switches are in their through or crossed states, respectively. For this reason, $H_l(t)^2 = H_r(t)^2 = 1$. In addition, due to the $\theta = \pi/2$ relative phase difference between the two modulation profiles, we have $H_l(t-T/4) = H_r(t)$.
For left-travelling waves, signals are delayed by $\tau = T/4$ before encountering the right switch. A signal that impinges on the network's 1$^\textrm{st}$ or 3$^\textrm{rd}$ ports at time $t$ is therefore routed by the two switches according to the product $H_l(t) \times H_r(t - \tau) = -H_l(t)^2 = -1$. The negative value of this product indicates the odd-parity of the switches: one and only one of the switches is in its crossed state, so port 1 routes to port 4 and port 3 routes to port 2.
In contrast, right travelling waves are delayed before they encounter the left switch. Signals incident on the 2$^\textrm{nd}$ or 4$^\textrm{th}$ port of the network are therefore routed according to the product $H_r(t) \times H_l(t - \tau) = H_l(t)^2 = 1$. The positive value of this product indicates the even parity of the switches: either zero or both of the switches are in their crossed states, so port 4 routes to port 3 and port 2 routes to port 1.
This kind of time-domain representation provides a simple explanation for the circulator's operation, and shows how incident signals are circulated by the network without altering their spectral content. It is worth noting, however, that the switches' square wave modulation means the signal is spectrally dispersed while it propagates through the delay lines---the second switch has not yet ``unraveled'' the frequency modulation enacted by the first switch.
\end{comment}
This graphical time-domain representation of the circulator may be complemented by a quantitative frequency-domain analysis in which all $m$ sidebands (detuned by integer multiples of the modulation frequency $\Omega = 2 \pi/T$)
are treated as ports of the network.
The scattering of a signal incident on the first physical port may then be expressed as
\begin{eqnarray}
\label{circulatorS}
S_{11}^m &=& 0, \\ \nonumber
S_{21}^m &=& \frac{1}{2}(\delta_{m,0} -r_m), \\ \nonumber
S_{31}^m &=& 0, \\ \nonumber
S_{41}^m &=& \frac{1}{2}(\delta_{m,0} +r_m), \\ \nonumber
r_m &\equiv& -\frac{4}{\pi^2} \sum_k \frac{e^{j k \Omega \tau} e^{j(m-k)\theta}}{(m-k)k},
\end{eqnarray}
where the sum on $k$ runs over all odd integers, $\delta$ is the Kronecker-delta, and $m$ is an index denoting effective ports at different frequencies, which is taken to be zero at the frequency of the input signal (e.g. $m = -2$ denotes the scattering of signals detuned from the input by $-2\Omega$). The term $r_m$ describes the transmission of an odd-excitation of the network's first and third ports: its interference with the transmission of an even excitation creates the desired circulation.
Such a representation is useful in quantifying the effect of imperfections and in evaluating the device's linearity.
(Absent nonidealities, Eq.~(\ref{circulatorS}) yields perfect four-port circulation when $\Omega \tau = \pi/2$, $\theta = \pm \pi/2$, for which $r_m = \pm\delta_{m,0}$).
Appendix~\ref{sec:derivation} includes the full multi-frequency scattering matrix derivation, and extension of Eq.~(\ref{circulatorS}) to include the effects of group delay dispersion, finite modulation bandwidth, and loss.
The device's sensitivity to these non-idealities depends on the modulation rate $\Omega$: Equation~(\ref{circulatorS}) shows
that group delay dispersion becomes problematic an $\Omega/(2\pi)$ shift in frequency causes a delay variation on the scale of $\tau$. Similarly,
a finite bandwidth $\Omega_b$ of the square-wave modulation truncates the sum in Eq.~(\ref{circulatorS}) at $k_\textrm{max} = \Omega_b/\Omega$.
Both of these considerations prejudice the design in favor of longer delays and modulation rates which are much smaller than the frequencies of the circulated signals. Slow modulation also facilitates filtering, allowing suppression of spurious coupling between the modulation and signal lines.
The benefits of slow modulation, however, do not come without a price. Slow modulation is only possible with long delays, which in the context of integrated circuits typically entails greater loss and circuit size and/or reduced bandwidth.
In a superconducting design space, though, vanishing conductive losses make long, low-loss, broadband delays possibl
~\cite{hohenwarter:1993,wang:2005,su:2008,zhong:2018}, and circuit size is the only remaining penalty. We leverage this fact in the proposed superconducting implementation of a broadband, low-loss, on-chip circulator, the topic of the following section.
\section{Proposed superconducting implementation}
We set design goals for our circulator by anticipating its use in superconducting qubit readout. In particular, we design for integrability with broadband near-quantum-limited amplifiers such as Josephson parametric amplifiers~\cite{roy:2015} or Josephson traveling-wave parametric amplifiers~\cite{macklin:2015}, in the frequency range of 4 to 8 GHz.
To allow for robust, wafer-scale production, we work in the design space allowed by optical lithographic fabrication
The modular nature of the proposed circulator reduces the challenge of designing a low-loss, broadband circulator to the problem of designing low-loss, broadband transfer switches and delays. We address these in turn, and then present simulations of the full device.
\subsection{Transfer switch}
To construct a fast superconducting transfer switch, we arrange two pairs of nominally identical and tunable inductors in a symmetric lattice (bridge) topology (see Fig.~\ref{fig:symlattice}a). The pairs of inductors tune in a coordinated fashion, allowing the switch to toggle between its through and crossed states (inset of Fig.~\ref{fig:symlattice}a).
\begin{figure}
\begin{center}
\includegraphics[width=1.0\columnwidth]{SymmetricLatticeWithMatchV3.eps}
\caption {\textbf{Transfer switch with tunable inductors.} A symmetric lattice of tunable inductors forms a transfer switch. (a) When the inductors are tuned such that $l_t \ll l_c$, the switch is in its through state. Tuning $l_c \ll l_t$ realizes the switch's crossed state. Parallel capacitances $c$ are used to match the switch to a characteristic impedance $Z_0$. (b) Calculated scattering parameters of the network in (a) parametrized by Eq.~(\ref{epsilon}), with $l_0 = 0.94$ nH, $c = 270$ fF, $\epsilon = 2.5 \times 10^{-2}$. Solid lines are exact solutions. Dashed lines are the approximations to first-order in $\epsilon$ given in Eq.~(\ref{epsexpansion}). Insertion loss is $<0.03$ dB between 4 and 8 GHz.
Gray rectangle shows the Bode-Fano criterion (Eq.~(\ref{BF})) between 2 and 10 GHz. Its proximity to $S_{11}$ attests to the efficacy of the single-pole matching network.
}
\label{fig:symlattice}
\end{center}
\end{figure}
Constructing the switch with purely reactive elements allows it to be completely free of dissipation. Limiting insertion los
, however, requires impedance matching.
Synthesis of a matching network is simplified if the bridge can be strongly imbalanced---that is, if the impedance of the through elements may be made much larger than the impedance of the crossed elements, or vice versa. In this limit, the Bode-Fano criterion~\cite{bode:1945,fano:1950} for the switch is approximately that of a series $RL$ circuit~\cite{pozar:2011}:
\begin{eqnarray}
-\int_0^\infty \ln{|\Gamma(\omega)|} d\omega < \pi \frac{R}{L}.
\label{BF}
\end{eqnarray}
Here $R$ is taken to be the characteristic impedance of the surrounding lines $Z_0 = 50$ Ohms, $\Gamma(\omega)$ is the reflection coefficient, and $L$ is the small inductance in the network ($l_t$ in the switch's through state, $l_c$ in the crossed state). A broadband match may then be obtained for sufficiently small $L$. For example, limiting reflections to $\Gamma = -20$ dB over an 8 GHz band requires that $L$ be less than approximately 1 nH. We make this match in a simple, symmetrical way, with shunting capacitors $c$ sandwiching the inductor $L$ (Fig.~\ref{fig:symlattice}a).
\begin{comment}
In that limit, the matching problem is essentially reduced to matching a series $RL$ circuit, and can be solved in a simple, symmetrical way, with shunting capacitors $c$ sandwiching the inductor $L$ (Fig.~\ref{fig:symlattice}a).
The efficacy of such a ``single-pole'' matching filter is not far from
the Bode-Fano criterion~\cite{bode:1945,fano:1950} for the switch is approximately that of a series $RL$ circuit~\cite{pozar:2011}:
\begin{eqnarray}
-\int_0^\infty \ln{|\Gamma(\omega)|} d\omega < \pi \frac{R}{L}.
\label{BF}
\end{eqnarray}
Here $R$ is taken to be the characteristic impedance of the surrounding lines $Z_0 = 50$ Ohms, $\Gamma(\omega)$ is the reflection coefficient, and $L$ is the small inductance in the network ($l_t$ in the switch's through state, $l_c$ in the crossed state). A broadband match may then be obtained for sufficiently small $L$. For example, limiting reflections to $\Gamma = -20$ dB over an 8 GHz band requires that $L$ be less than approximately 1 nH.
\end{comment}
To quantify the needed imbalance, we analyze the switch in its ``through'' position and parametrize the pairs of tunable inductors as
\begin{eqnarray}
l_{t} &=& l_0, \nonumber \\
l_{c} &=& \frac{l_0}{\epsilon}.
\label{epsilon}
\end{eqnarray}
Here $\epsilon$ is a small parameter that describes the degree to which the bridge can be imbalanced. A standard circuit analysis (see Appendix~\ref{sec:switchS}) allows expansion of the switch's scattering parameters in powers of $\epsilon$. We find that the amplitude of both reflection and unwanted transmission (e.g. transmission to isolated ports of the switch) scale to leading order as $\epsilon$, provided the impedance of the resonator $Z_r = \sqrt{\frac{l_0}{2c}}$ is equal to $Z_0$, and the
capacitive corner-frequency $f_c \equiv (Z_0 c)^{-1}$ is much larger than the maximum frequency of operation ($\approx10$ GHz). When $l_0 \approx 1$ nH, this may be accomplished with $c \approx 200$ fF. Fine tuning of the inductance and capacitance may then be used to adjust the passband of the switch around the center frequency $\omega_0 = 1/\sqrt{2 l_0 c}$. Made in this way, the imbalanced switch is effectively an RLC oscillator with quality factor of order one.
As the switch's isolation and return loss depend directly on $\epsilon$, implementing inductors which may be tuned over a wide range is a critical task.
We suggest two ways in which small $\epsilon$ values may be realized.
The first uses laddered arrays of Josephson junctions which are flux-biased close to frustration. Devices made in this way have demonstrated $\epsilon \leq 2\times 10^{-2}$~\cite{bell:2012,naaman:2016}.
The second uses dc-SQUIDs with small, high-aspect ratio loops ($\simeq5$ $\mu$m $\times$ $\simeq100$ nm). These may be fabricated with optical lithography by orienting the loop such that its normal vector is parallel to the chip's surface. We expect $\epsilon<2.5\times10^{-2}$ to be possible with such a design. (For more details, see Appendix~\ref{sec:SQUIDs}).
Implemented in either fashion, the tunable inductors may then be arranged in a bridge geometry and tuned with a pair of flux controls, as in Refs.~\cite{kerckhoff:2015,chapman:2016,chapman:2017,chapman:2017b}.
Figure~\ref{fig:symlattice}b shows the magnitudes of the calculated scattering parameters of our switch design, optimized for the 4 to 8 GHz band given $\epsilon = 2.5 \times 10^{-2}$. Between 4 and 8 GHz, insertion loss is $<0.03$ dB, isolation exceeds $25$ dB, and reflections are less than $-26$ dB. The impedance match is therefore comparable to the Bode-Fano limit for an $RL$ circuit (shaded gray rectangle). The switch's group delay dispersion is non-zero, but as the total delay provided by the switch is approximately the inverse of its bandwidth (and $Q \approx 1$), it imparts a delay which is small relative to that of the delay lines.
Imbalancing an inductive bridge to this degree puts constraints on the junction's plasma (self-resonance) frequency $\omega_p$. When the tunable inductors are flux-biased into their high-inductance state, their Josephson inductance changes by a factor of $1/\epsilon$, decreasing the plasma frequency of the unbiased junction by a factor of $\sqrt{\epsilon}$. To ensure that this reduced plasma frequency is always greater than the circulator's operation band, we require $\omega_p > 2 \pi \times \sqrt{\epsilon} \times 10$ GHz in the unbiased junctions.
For a niobium trilayer process where $\omega_p$ is a known function of the junction critical current density $J_c$~\cite{maezawa:1995,tolpygo:2017}, the above constraint defines a minimum $J_c$. Given a minimum feature size of $2~\mu$m, the minimum junction critical current is then set at $I_0 = 9$ $\mu$A. Finally, the target array inductance of $l_0 \approx 1$ nH can be achieved by cascading $N \approx 40$ of the SQUIDs in series.
An auxiliary benefit of this large plasma frequency requirement comes in the device's linearity. Similar broadband switches fabricated with $35$ $\mu$A junctions have reported $1$-dB compression points at $-53$ dBm~\cite{naaman:2016}. Assuming power-handling scales with the square of junction critical currents, the design proposed here would saturate around $-65$ dBm, ample for integration with Josephson travelling-wave amplifiers~\cite{macklin:2015}.
For use in the proposed active circulator, it is not enough for the transfer switches to perform statically: they must also be fast. In principal, the only physical constraint on the switching speed is the junction self-resonance frequency. In practice, however, further restriction of the modulation bandwidth can limit the extent to which photons in the flux control lines couple into the circuit and emerge as noise in the operation band.
In particular, in the regime where the modulation rate $\Omega$ is much less than the 4 to 8 GHz operating frequencies, it is advantageous to limit the modulation bandwidth to half of the minimum operating frequency. This provides ample separation between the modulation band ($1/T$ to 2 GHz) and the signal band (4 to 8 GHz), which facilitates filtering. It also ensures that modulation tones which spuriously couple into the signal lines will be below the signal band, even if they are mixed up in frequency by the second transfer switch. Further discussion of the control waveforms' effect on the circulator's output spectrum is given in Appendix~\ref{sec:noise}.
\subsection{Delay lines and optimal modulation rate}
Modulating the transfer switches slowly requires a broadband and low-loss delay on the scale of ns, many microwave periods long. Furthermore, this delay must be relatively free of dispersion. These considerations suggest the use of a transmission line delay which supports a TEM or quasi-TEM mode, such as a coplanar waveguide (CPW). Such a line may be tightly meandered to realize the desired broadband delay in a compact manner.
The primary sources of dispersion in such a structure are the resonant modes associated with the ground plane and the bends in the path of the CPW. To account for these, we conducted numerical simulations of meandered CPWs using a planar method-of-moments solver. These show 1 ns delays may be achieved over the entire 2 to 10 GHz band, with return loss better than $30$ dB and fractional group velocity dispersion $<3\times10^{-4}$ per $80$ MHz. Dispersion at this level would cause $0.05$ dB of circulator insertion loss (see Appendix~\ref{sec:derivation}). (Dispersion from the quasi-TEM nature of the CPW mode, and the frequency-dependence of niobium's surface inductance, are irrelevant on this scale~\cite{hasnain:1986,gao:2008}.)
At temperatures and frequencies where the thermal and photon energies are much less than the superconducting gap, dissipation in a CPW is dominated by dielectric loss. For a signal with frequency $\omega$ propagating in a TEM mode for a time $\tau= \pi/(2\Omega)$, the attenuation $a$ from the surrounding dielectric is~\cite{orfanidis:2002}
\begin{eqnarray}
a = \textrm{exp}\left(-\frac{\omega \tau}{2} \tan{\delta}\right).
\label{dielectricloss}
\end{eqnarray}
Here $\tan{\delta}$ is the dielectric loss tangent. Coplanar waveguides fabricated on silicon substrates routinely demonstrate quality factors $>10^{5}$~\cite{oconnell:2008}, making the dissipation from a $1$ ns delay less than 3 mdB for signals in our band.
To estimate the physical size of such a delay and the resulting circuit, we note that for a CPW on silicon where microwaves travel more slowly than the vacuum speed of light $c_0$,
the needed delay line has length $d \approx \pi c_0/(5 \Omega)$.
Assuming a minimum lithographic feature size of $2$~$\mu$m, a 50 Ohm CPW may be as narrow as 10 $\mu$m, and a reasonable value for the meander pitch is $p = 60$~$\mu$m.
The switches themselves require less than 1 mm$^2$, so the entire layout may fit on a chip with area $2dp$ + 2 mm$^2$.
Selecting a precise value for the modulation frequency $\Omega$ therefore amounts to balancing competing effects: slower modulation minimizes the loss associated with dispersion and finite modulation bandwidth, but increases dielectric losses and circuit size. Figure~\ref{fig:optimal} summarizes these trade-offs. The blue trace shows circuit size (left axis), while the red traces shows the various loss channels (right axis), calculated with Eq.~(\ref{dielectricloss}) and the methods described in Appendix~\ref{sec:derivation}. A modulation rate of $\Omega = 2\pi \times 80$ MHz ($d\approx 37.5$ cm) is estimated to keep the circulator's insertion loss below $0.1$ dB while allowing the device to fit on a $7$ mm by $7$ mm chip.
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=1.0\columnwidth]{OptimalModulationRateV3-01.eps}
\caption {\textbf{Optimal modulation rate.} Estimated chip area (left axis, blue trace) and total insertion loss (right axis, thick red trace) as a function of the modulation frequency. Contributions to the insertion loss from dispersive, dielectric, and finite modulation bandwidth effects are also shown. These assume a fractional group velocity dispersion of $3\times10^{-4}$ per $80$ MHz, a dielectric loss tangent of $\tan{\delta} = 10^{-5}$, and an $\Omega_b/(2\pi) = 2$ GHz modulation bandwidth.
}
\label{fig:optimal}
\end{center}
\end{figure}
\subsection{Simulated Performance}
To predict the performance of the proposed superconducting circulator we conduct time-domain numerical simulations of the design. The delays are modeled as ideal transmission lines which provide a delay of $\tau = \pi/(2 \Omega) = 3.125$ ns (commensurate with $\Omega = 2 \pi \times 80$ MHz). The transfer switches are modeled by the lumped-element network depicted in Fig.~\ref{fig:symlattice}a, with several non-idealities included:
First, each inductor is replaced by an array of $N = 46$ inductors in series, each of which is assumed to be tunable by an external signal $\Phi_e$, via the relation
\begin{eqnarray}
l(\Phi_e) = \frac{l_0}{\sqrt{\epsilon^2 + (1-2\epsilon) \cos^2{\left(\Phi_e\right)}}}+l_g.
\label{TunableInductance}
\end{eqnarray}
This functional form is chosen to replicate the flux-dependence of an asymmetrical dc-SQUID's inductance~\cite{vanduzer:1981}, and make contact with Eq.~(\ref{epsilon}) in the limits $\Phi_e \to 0$ and $\Phi_e \to \pi/2$. We fix $N l_0 = 1.02$ nH and $\epsilon = 2.5 \times 10^{-2}$, and the geometric inductance is chosen to be $N l_g = 207$ pH. Second, alongside each of these tunable inductors we place a parallel capacitance of $c_J = 180$ fF to mimic the self-capacitance of the junctions. Lastly, the tunable inductors are modulated by two approximately square-wave signals with frequency $\Omega = 2 \pi \times 80$ MHz. We model the finite modulation bandwidth by synthesizing the bias signals from the sum of the first 25 Fourier components of a square-wave, such that the bias waveform has no spectral weight above $\Omega_b/(2 \pi) = 25 \Omega/(2 \pi) = 2$ GHz
Figure~\ref{fig:simulink}a shows the first column of the simulated network's scattering parameters.
\begin{figure}
\begin{center}
\includegraphics[width=1.0\columnwidth]{SIMULINKsparamsV4-01.eps}
\caption {\textbf{Simulated performance.} Time-domain numerical simulations of the circulator, with transfer switches formed from symmetric lattices of tunable inductors modulated at $\Omega = 2\pi\times 80$ MHz. (a) First column of the simulated scattering matrix, when the device is configured for counter-clockwise circulation. The horizontal line at $-20$ dB is a guide to the eye. (b) Power spectrum of the signal emerging from the circulator's second port, when the circulator is driven by a 6 GHz tone incident on its first port. Power is normalized to the incident signal power $P_{\textrm{inc}}$. The thin black trace shows the same quantity in a simulation with infinite modulation bandwidth.}
\label{fig:simulink}
\end{center}
\end{figure}
Between 5.9 and 8.2 GHz, the circulator's isolation exceeds 20 dB. Insertion loss in the same band is less than $0.3$ dB. Accounting for loss caused by disperion in the CPW delays (which is not present in the time-domain simulation), the total loss of the device is expected to be less than $0.35$ dB across this band.
The primary performance limitation is the finite modulation bandwidth: if the simulation is repeated with infinite modulation bandwidth and all other parameters held fixed, isolation is greater than $20$ dB between $4$ and $8.5$ GHz, and insertion loss is improved to better than $0.1$ dB. Power emerging out of the circulator's other three ports (at the input frequency) accounts for $0.03$ dB of this loss. We attribute the final $0.05$ dB to the dispersion of the switches.
The simulated $~0.2$ dB of insertion loss that results from an $\Omega_b = 2\pi \times 2$ GHz modulation bandwidth is greater than the $0.07$ dB expected from the analysis in Appendix~\ref{sec:derivation}.
The discrepancy arises from the sensitive dependence of the inductance on the control signal near the $\Phi_e = \pi/2$ operating point (at $\Phi_e = \pi/2$, $dl/d\Phi_e$ diverges).
While the rising-edge of the finite bandwidth flux signal approaches (and then overshoots) its target value of $\pi/2$, the efficacy of the switches is diminished until the control flux settles.
The ripple in the scattering parameters has a free spectral range of $2\Omega$, and is caused by small reflections at the second transfer switch. As these reflections take two extra passes through the delay lines, they experience an additional delay of $2 \tau = \pi/\Omega$ before interfering with the un-reflected signal.
Despite the coordinated modulation used to create circulation, the majority of the transmitted power is unchanged in frequency. The orange trace in Fig.~\ref{fig:simulink}b shows the spectral content of the signal exiting the circulator's second port, when the first port is driven by a $6$ GHz tone. Most of the power ($-0.28$ dB) emerges at 6 GHz. Sidebands detuned from $6$ GHz by multiples of $2 \Omega/(2\pi)$ are also visible. The largest of these sidebands is less than $-27$ dB. For reference, the thin black trace also shows the same simulation with infinite modulation bandwidth. Although they are suppressed by over $45$ dB, modulation sidebands are still present in the spectrum. They are caused by the finite isolation of the transfer switches
\section{Conclusion}
We have described the theory of operation for a broadband and low-loss circulator based on active modulation and delay. The device may be understood as a network of two simple components: transfer switches and delays. Design of a superconducting implementation was presented and discussed in terms of the optimal modulation rate $\Omega$, a key design parameter. The device was then simulated numerically: isolation is greater than $20$ dB while insertion loss is less than $0.35$ dB over a bandwidth of $2.3$ GHz. Based on comparison with demonstrated Josephson junction based switches, its $1$-dB compression point is expected to be $-65$ dBm.
The proposed circulator therefore represents a major advance in on-chip non-reciprocity, improving on the bandwidth and power-handling of other near-lossless devices~\cite{sliwa:2015,lecocq:2017,chapman:2017b} by more than a factor of $50$. These gains become especially powerful when the broadband circulator is combined with a linear and broadband near-quantum limited amplifier~\cite{macklin:2015,naaman:2017}.
Such a quantum microwave receiver, which would be linear at powers $10^6$ times greater than a typical qubit readout tone, and operable over a bandwidth $10^3$ times greater than a typical readout-cavity's coupling rate, could allow for the simultaneous readout of many qubits in an on-chip package.
\vspace{0.1in}
\noindent{\emph {Acknowledgment}} The authors thank Jiansong Gao, Bradley Moores, and Xian Wu for helpful discussions. E.I.R acknowledges support from the ARO QuaCGR fellowship. This work is supported by the ARO under contract W911NF-14-1-0079 and the National Science Foundation under Grant Number 1125844.
|
\section{Introduction}
\label{SEC:intro}
Cities are the examples of complex systems.
However, the functioning of this type of complex system depends upon social, economic, and environmental factors.
In spite of the diversity of the physical forms of the cities, there is a pattern and order in the expansion of cities.
The size and shape of the cities are known to play a fundamental role in social and economic life \cite{pnas, batty}.
The study of the size distribution of urban areas in various parts of the world helps us to understand
their dynamics and also plays a key role to manage their growth and environmental impacts.
City size distribution for various countries has been extensively studied for several decades.
In general it is believed that the city sizes obey a remarkably simple law, known as Zipf’s law \cite {zipf}
(alternatively known as Pareto distribution or simply power law) for cities with heavy population.
Although a considerable amount of research has been devoted to power law behaviors, in various contexts, however,
when real data \cite{usa, brazil, kgbb, kgbb2013} is analyzed,
in most of the cases the power law trend holds only for an intermediate range of values;
there is a power law breakdown in the lower distribution tails \cite{stanley2000, newman2005}.
This has motivated the studies on the size distribution of smaller cities in different countries
\cite{12kgbb, 13kgbb, 14kgbb, 15kgbb, 16kgbb, 17kgbb, 18kgbb, 19kgbb,luck17I,luck17U}.
Recently, it has been shown that the data for the Indian city size distribution exhibit
a strong reverse Pareto in the lower tail, log-normal in the mid-range and a Pareto in the upper tail \cite{luck17I}.
A relatively simpler mixture of lognormal in the lower tail and Pareto in the upper tail has also been fitted
for US city size distributions \cite{luck17U}.
The existing literature on modeling the sizes of cities having small to extremely large populations
are mostly, up to the knowledge of the authors, if not all, confined to the Pareto law in the upper tail;
hence use of a more complex mixture distribution is required
(with another appropriate distribution in the lower or middle ranges).
Such mixture distributions are known to be difficult to fit statistically (estimation process is often complex)
and often difficult to interpret. There is also no evidence of one such mixture distribution providing significantly good fit
to different types of city-size distributions in socio-economically different countries across the world.
The goal of the present paper is to search for a global pattern of city size distributions along with their discrimination and evolution, so that one can compare all the cities in a universal framework and better understand the dynamics of urbanization in society.
In this work, we focus on a particular two-parameter Rank-Ordered (RO) distribution,
namely the discrete generalized beta distribution (DGBD) \cite{Martinez-Mekler09}, for cities
with different sizes in a given country in any continent of the world, including the important economies like India, China, USA, Brazil
as well as the recent distribution of the sizes of World Cities with heavy population.
This particular RO distribution is a discrete generalization of the continuous beta distribution,
where a given property of a system of $N$ items is ordered according to its importance (rank)
so that the domain is the natural numbers 1 to $N$ instead of the interval $(0,1)$ of usual continuous beta distribution.
Several RO distributions including the DGBD and its further generalizations have already been studied to provide good fits
in the context of different count or rank-size data from arts and sciences \cite{Martinez-Mekler09,Ausloos/Cerqueti:2016,Alvarez-Martinez/etc:2011, Alvarez-Martinez/etc:2014, Oscar2017,
Alvarez-Martinez/etc:2018, Lalit/etc:2018}.
In the present paper, our main contribution is to propose and illustrate that the DGBD can be successfully
applied to provide a simple, yet excellent, universal fit to the city size data across different socio-economic countries.
Our analyses show that the data for all countries under consideration follow, with a wonderful goodness-of-fit,
the DGB distribution, a two parameter RO distribution, incorporating the product of two power laws
defined over the complete data set --
one measured from left to right and another in the opposite direction.
For the vanishing value of one particular parameter,
the DGB distribution simplifies to the usual Pareto law for the cities with heavy population (see Section \ref{SEC:ROD}).
To understand the underlying process of the urban morphology, the Shannon entropy of the proposed DGB distribution has also been studied.
Based on the two parameters of the distribution and the corresponding entropy,
one can indeed discriminate and study the evolution of city distributions in different countries across the world.
The rest of the paper is organized as follows. Sec.~II deals with the mathematical formulation of the two-parameter DGBD.
In the respective subsections A, B and C, we define our particular RO distribution and the corresponding Shannon entropy
and provide the methodology for the estimation of the parameters.
In Sec.~II.D we detail the prediction and the goodness-of-fit test for examining the performance of our proposal.
Sec.~III describes the data used for our empirical analyses and also demonstrates the RO modeling of these data of different countries across the world,
including India, China, USA, Brazil, Australia, 3 European countries and 3 African countries at different points of time.
This section also presents the analysis of the world-wide size distributions of cities and countries in the year 2018 (Sec.~IV.H).
Sec.~V narrates the discrimination and evolution of the city-size distributions across the countries
and the advantage of our proposal in this context.
Finally, we conclude the paper with some discussions about the future works in Sec.~VI.
\section{Modeling City-sizes using the Rank-Order distribution }
\subsection{The Rank-Ordering (RO) distribution}
\label{SEC:ROD}
Let us first recapitulate the mathematical formulation of rank-ordered distribution, in particular the DGB distribution.
Suppose we have data on the sizes of $N$ items (e.g., cities) arranged in decreasing order of ``importance" (e.g., size)
so that the $i$-th item (or city) has size $n_i$ with rank $r_i$.
Note that $(r_1, \ldots, r_N)$ is a permutation of $(1, \ldots, N)$.
The most widely used hyperbolic Pareto (Zipf's) law fits the data through the probability of rank $r$ being given by
\begin{equation}
f_P(r) = A\cdot\frac{1}{r^\nu}, ~~~~r=1, \ldots, N,
\label{EQ:Pareto}
\end{equation}
where $\nu>0$ is a model parameter, known as the exponent, and
$A$ is the normalizing constant defined by the formula
$ A=\left[ \sum_{r=1}^N r^{-\nu}\right]^{-1}$.
As already noted in Section \ref{SEC:intro}, this Pareto or power law provides good fit to the data
only at low-ranks (large sizes).
This is mostly because of the fact that the empirical data distribution often have an inflection point,
which is not captured by the single power law (\ref{EQ:Pareto}).
In order to accommodate the inflection as well as the shape of the two ends of the hierarchy,
one can use the simple two-parameter RO distribution, namely the DGBD \cite{Martinez-Mekler09}, defined by
\begin{equation}
f_{(a,b)}(r) = A\frac{(N+1-r)^b}{r^a}, ~~~~r=1, \ldots, N,
\label{EQ:RO}
\end{equation}
where $a, b$ are two real valued model parameters and $A$ is again the normalizing constant.
Given $N$ and $(a,b)$, one can compute $A$ as
\begin{equation}
A=\left[ \sum_{r=1}^N \frac {(N+1-r)^b}{r^a}\right]^{-1},
\label{EQ:RO_A}
\end{equation}
so that $f_{(a,b)}(\cdot) $ is a proper probability distribution, i.e. $\sum_{r=1}^N f_{(a,b)}(r) =1$.
\begin{figure}[!h]
\centering
\subfloat[\large $b=0$]{
\includegraphics[width=0.34\textwidth]{RO_b00_2000.png}
\label{FIG:b0ap}}
~
\subfloat[\large $b=0.5$]{
\includegraphics[width=0.34\textwidth]{RO_b05_2000.png}
\label{FIG:loglogCumHaz_SmallCell}}
\\
\subfloat[\large $b=0.3$]{
\includegraphics[width=0.34\textwidth]{RO_b03_2000.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[\large $b=-0.3$]{
\includegraphics[width=0.34\textwidth]{RO_b03m_2000.png}
\label{FIG:RO_negB}}
\\
\subfloat[\large $b=1$]{
\includegraphics[width=0.34\textwidth]{RO_b10_2000.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[\large $b=-1$]{
\includegraphics[width=0.34\textwidth]{RO_b10m_2000.png}
\label{FIG:loglogCumHaz_SmallCell}}
\caption{The form of the DGBD $f\equiv f_{(a,b)}$, in log-log scale,
for different parameter values with $a\geq 0$ and $N=2000$
[Dash-dotted line: $a=0$, Dashed: $a=0.5$, Dotted: $a=1$, Solid: $a=1.5$].}
\label{FIG:RO_posA}
\end{figure}
The two parameters $a$ and $b$ indeed characterize the shape of the rank-size distribution
at two-ends of low and high ranks, respectively, in respect to the inflection point.
Figure \ref{FIG:RO_posA}, representing the shape of the DGBD for different values of $a, b\in\mathbb{R}$
and $N=2000$ in the log-log scale,
clearly indicates its potentiality in modeling a wide variety of rank-size data structures;
see also \cite{Ausloos/Cerqueti:2016}.
In particular, it is interesting to note that the case $b=0$ yields back
the Zipf's law in (\ref{EQ:Pareto}) with $\nu = a$ having linear shape in the log-log scale
(Figures \ref{FIG:b0ap});
these special cases indicate that there is no inflection point in the empirical data
so that only one parameter ($a=\nu$) is sufficient to model the slope of the data (in log-log scale).
Further, the shape of the DGBD becomes more flat as $a$ and $b$ decreases close to zero,
just like the effect of $\nu$ on the power law.
For any fixed $a$, the parameter $b$ in the DGBD allows a wide variety of different distributional graphs;
furthermore, for some exceptional data a non-monotonic DGBD may overall be the best fit although the data is monotonic.
This will be commented later on.
In terms of physical interpretation, the range of the distribution becomes more broad
(spread over a larger interval and is less uniform) with increasing values of $a,b$,
leading to easier discrimination between ranks in the opposite extremes.
On the other hand, as the values of either $a$ or $b$ decreases,
the sizes of the individual ranks become more uniform (lesser spread)
and hence more difficult is their discrimination;
the case $a=0=b$ coincides with the uniform distribution which is most uncertain and less broad.
This phenomenon can be well connected and characterized by the concept of entropy,
as detailed in the following subsection.
\subsection{Shannon Entropy of the distribution }
A city is not and cannot in any sense be treated as an isolated system, since it always exchanges
materials, energy, information and people with its surroundings.
When related to a probability, the concept of the Shannon-Gibb's entropy can be used in analyzing
the uncertainty in a city-size distribution.
For the DGBD, the Shannon entropy \cite{Ausloos/Cerqueti:2016} is given by,
\begin{eqnarray}\label{entropy1}
S (a,b) &=& -\sum_{r=1}^N f_{(a,b)}(r) \log f_{(a,b)}(r)
= -\log A
-A\sum_{r=1}^N \frac{(N+1-R)^b}{r^a} [b \log (N+1-r)-a \log r].
\end{eqnarray}
Note that, since $f_{(a,b)}(r) $ is assumed to be normalized to $1$ function, following \cite{Ausloos/Cerqueti:2016},
it is absolutely meaningful to use the Shannon entropy as defined in (\ref{entropy1}).
The entropy measure in Eq.~(\ref{entropy1}) lies at the basis of information theory \cite{jaynes} and
is applied successfully across several disciplines of arts and sciences \cite{kapurbook}.
\begin{figure}[!b]
\centering
\subfloat[$N=2000$]{
\includegraphics[width=0.45\textwidth]{ROD_Entropy_N2000.png}
\label{FIG:loglogCumHaz_Veteran}}
\subfloat[$N=200,000$]{
\includegraphics[width=0.45\textwidth]{ROD_Entropy_N200000.png}
\label{FIG:loglogCumHaz_SmallCell}}
\\
\subfloat[Plot over varying $N$]{
\includegraphics[width=0.7\textwidth]{ROD_Entropy_3D1.png}
\label{FIG:loglogCumHaz_SmallCell}}
\caption{Entropy of the DGBD over $(a, b)$ for different $N$}
\label{FIG:Entropy_N20}
\end{figure}
Figures \ref{FIG:Entropy_N20} (a) and (b) respectively display the contour plots for the exact values of the entropy corresponding to the DGBD with $N=2000$ and $200,000$ over the parameters $(a,b)$.
Also, increasing one dimension in the plot, in Fig. 2(c) we have represented the entropy for different values of $N$ along with the variation of $(a,b)$.
Clearly, the maximum entropy corresponding to the case $a=b=0$ leads to the uniform distribution.
The case $b=0$, with any $a >0 $, corresponds to the entropy for the power law distribution.
Further, the pattern of the entropy is almost indifferent for different $N$ although the absolute value of $S$ changes.
Hence, the city-size distributions can be compared in terms of their entropy and distributional structure
only through the values of the two parameters $a,$ and $b$;
we will illustrate it further in Section \ref{SEC:discrimination}.
\subsection{Parameter Estimation through Maximum Likelihood}
\label{SEC:est}
We now encapsulate how the DGB distribution can be used for universal modeling of the city-size distributions
and explain the methodology for the estimation of the parameters $(a,b)$ from the empirical data.
Suppose we have the data for a population of $N$ cities in order of decreasing sizes
and $x_i$, $r_i$ respectively denote the actual size and rank of the $i$-th city for $i=1, \ldots, N$.
We can fit the DGB distribution to these data by means of estimating the parameters $(a, b)$
which leads the assumed
probabilities $f_{(a,b)} (r)$ to be $closest$ to the empirical (normalized) sizes in an appropriate sense.
The best (asymptotically) statistical estimation technique is the maximum likelihood estimator (MLE)
which measures the closeness in terms of Kullback-Leibler divergence measure
and maximizes the probability of the observed (sample) data under the assumed model.
Since the Kullback-Leibler divergence is directly linked with the Shannon entropy
$S$ in (\ref{entropy1}) and the corresponding cross-entropy,
the MLE is also the same as the corresponding minimum cross-entropy estimator,
popular in information sciences and statistical physics.
The MLE also has minimum standard error in estimation, asymptotically for larger sample sizes,
among a wide class of useful estimators.
Given the ranked city-size data $(x_i, r_i)$, the MLE of $(a, b)$ in the best fitted RO model (DGBD)
is to be obtained by maximizing the likelihood function given by
\begin{equation}
L(a, b) = \prod_{i=1}^N f_{(a,b)}(r_i)^{{x}_i} = \prod_{i=1}^N \frac{(N+1-r_i)^{b{x}_i}}{r_i^{a{x}_i}}A^{{x}_i}.
\end{equation}
Equivalently, one can also maximize the log-likelihood function given by
\begin{eqnarray}
&&\ell(a, b) = \log L(a, b)
= b \sum_{i=1}^N {{x}_i}\log(N+1-r_i)-
a\sum_{i=1}^n{{x}_i}\log(r_i)+ \log(A)\sum_{i=1}^n{{x}_i},
\nonumber
\end{eqnarray}
or solve the estimating equations
$$\frac{\partial}{\partial a}\ell(a, b) = 0 = \frac{\partial}{\partial b}\ell(a, b).$$
Noting from Eq.~(\ref{EQ:RO}) that $A$ is a function of $(a, b)$,
we can numerically maximize $\ell(a, b)$ or solve the above estimating equations, with respect to $a, b\in \mathbb{R}$,
to obtain their MLE, which we will denote as $(\widehat{a}, \widehat{b})$. In all our data analyses we have numerically minimized the negative of log likelihood function $\ell (a,b)$, in the software MATLAB (version14a)
using the in-built function \textit{'fminsearch'}.
\subsection{Prediction and Goodness-of-fit}
\label{SEC:goodness}
Once we have obtained the MLE $(\widehat{a}, \widehat{b})$ of the parameters $(a, b)$ of
the DGBD fitted to a given set of empirical data, we can predict the rank-sizes from this fitted model.
The predicted normalized size of rank $r_i$ is given by $f_{(\widehat{a}, \widehat{b})}(r_i)$
and hence, multiplying it by the total size $\sum\limits_{i=1}^N x_i$, one gets the predicted size of $r_i$
that can be denoted as
$$p_i = \left(\sum_{i=1}^N x_i\right)f_{(\widehat{a}, \widehat{b})}(r_i).$$
If the estimated RO model is a good fit to the given data, the predicted values $p_i$
should be close to the corresponding observed size $x_i$ and the corresponding plot gives us a visual indication of the goodness-of-fit of our RO model (via DGBD) to any dataset.
A quantitative measure of fit can also be obtained by summing the error in prediction;
we will here consider the measure defined in terms of the Kolmogorov-Smirnov distance between
the predicted and observed cumulative frequencies of ranks as given by
\begin{eqnarray}
KS = \max_{1\leq i \leq N} \left|\left(\sum_{j: r_j\leq r_i} p_j\right) - \left(\sum_{j: r_j\leq r_i} x_j\right)\right|.
\label{EQ:KS}
\end{eqnarray}
The use of cumulative sizes instead of actual sizes in the definition of KS
provides better stability of this goodness-of-fit measure.
Note that, this KS measure becomes zero if and only if $p_i=x_i$ for all $i=1, \ldots, N$,
i.e., all the predicted sizes coincide exactly with the observed sizes.
For a given dataset, the KS measures closer to zero indicate better goodness-of-fit
and the model corresponding to minimum KS will be the $best$ as a predictive model of the underlying rank-size distribution.
\section{Data Description}
\label{SEC:data}
In the last few decades the human population is increasingly clustered in urban areas. The urban population of the world has grown rapidly from 746 million in 1950 to 3.9 billion in 2014. Asia, despite its lower level of urbanization, is home to 53\% of the world's urban population,
followed by Europe with 14\% and Latin America and the Caribbean with 13\% \cite{desa}.
Today, 55\% of the world's population lives in urban areas, a proportion that is expected to increase to 68\% by 2050.
Projections show that urbanization, the gradual shift in residence of the human population from rural to urban areas,
combined with the overall growth of the world's population could add another 2.5 billion people to urban areas by 2050,
with close to 90\% of this increase taking place in Asia and Africa, according to a new United Nations data set launched in May, 2018 \cite{un}.
Together, India, China and Nigeria will account for 35\% of the projected growth of the world's urban population between 2018 and 2050.
By 2050, it is projected that India will have added 416 million urban dwellers, China 255 million and Nigeria 189 million.
In the present day context these data motivate us to study the size-distribution of {\it{all}} cities, or urban settlements
in different countries, in particular India and China, once again.
As an effect of urbanization, it is now difficult to distinguish city, suburb or town. All definitions of $city$ has shortcomings, either it is defined as incorporate area, an urban agglomeration, or a human settlement with population density larger than some threshold value \cite{plos2007}.
For our analysis, we use the term $city$ to mean any human settlement that is denser than its surroundings;
we consider towns, cities, counties, urban agglomerations according to the availability of the data for various countries
as described below; these datasets are public and freely available.
\begin{itemize}
\item Indian city-size data from the censuses at the years 1991, 2001 and 2011 \cite{india data},
with minimum inhabitant of 5,000.
\item Chinese city and county sizes for three years, namely 1990, 2000 and 2010, as defined in \cite{china data}.
\item USA data for each years between 2010 to 2017, with minimum size of 5,000,
where the city populations are estimated based on 2010 census \cite{usa data}.
\item Brazil city populations for the years 1991, 2000, 2010 and 2017 (estimated),
with a minimum of 20,000 inhabitants \cite{brazilcity}.
\item City sizes of three African Countries, namely Uganda, Sudan and Algeria for the years 2014, 2008 and 2008, respectively,
with minimum sizes of 15,000, 20,000 and 13,000 \cite{africa}.
\item Italian cities for the years 1981, 991, 2001, 2011 and 2017 (estimated) \cite{italy}, with minimum population of 50,000.
\item Sweden city size estimates for each fifth year from 1990 to 2015 and 2017 \cite{sweden},
having population greater than 10,000.
\item Switzerland city-sizes for the years 1980, 1990, 2000 and
the corresponding estimates for 2010 and 2017 \cite{switzerland},
with minimum inhabitants of 10,000.
\item Australian urban agglomerations (UA) (all) for the years 2011 and 2016 \cite{australia}.
\item We also have considered the population data for major cities around the world in the year 2018,
as obtained from \cite{world2018}.
\item Finally, we have also tested our model with the data on the population ($> 5000 $) of 226 countries
and dependent territories across the world, as defined and detailed in \cite{world2018}.
\end{itemize}
\section{Empirical Illustrations: Universality in space and time }
In this section, we present the results obtained by the applications of the DGBD
for different countries' city-size data described in the previous section.
For comparison, the popular Pareto model, corresponding to the Zipf's law, has also been fitted
to the same sets of data through a linear regression between logarithms of size and rank,
and the superiority of our proposal has been illustrated in all the cases through the KS goodness-of-fit measure (\ref{EQ:KS}).
The highly significant fits of the DGBD for all such cases indicate its universality over time
as well as its global nature to model different kinds of city size distributions
from different economies and geographies across the world.
\subsection{Size distribution for Indian cities}
Let us start our demonstration with Indian cities,
using the census data for the years 1991, 2001 and 2011, and also remark on the viability of our prediction.
According to the most recent census in 2011 \cite{census2011},
India is the second largest country in the world with more than 1.2 billion people,
and it is predominantly rural country with about 67\% population living in rural areas \cite{world2018}.
The country has a total of 39 cities with more than one million residents each;
among them, Mumbai and Delhi have populations exceeding 10 million.
However, the country also has smaller but still very populated cities,
including 388 having populations over 100,000, and a whopping 2,483 cities with populations over 10,000.
As noted earlier in Section \ref{SEC:data}, we consider Indian cities to be defined
as human settlements with population more than 5,000.
The whole data is divided into different classes (I-V) according to its population
as follows: Class (I) is above 100,000 population, Class (II) is from 50,001 to 100,000,
Class (III) is from 20,001 to 50,000, Class (IV) is from 10,001 to 20,000 and Class (V) is from 5,000 to 10,000.
We have fitted the DGBD for all these city-sizes (Class I-V) and
also for some subset of cities starting with a larger minimum size.
In particular, we consider the subset of cities with population at least 100,000 (i.e., Class I cities),
or at least 50,000 (i.e., Class I and II cities) or at least 10,000 (i.e., Class I-IV cities).
The estimated parameter values and the goodness-of-fit measures (KS) for all four cases
and three years are reported in Table \ref{TAB:Single_2001}, along with total number of cities ($N$)
and the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively, in each cases.
The corresponding fitted (and actual) sizes are plotted in Figure \ref{FIG:Single_2011} for the most recent year 2011;
the plots for other two years are very similar and not presented here to avoid repetition
The fitted values and goodness-of-fit measures corresponding to the Pareto modeling (along with the estimated exponent $\nu$),
for the same sets of data, are also shown in Figure \ref{FIG:Single_2011} and Table \ref{TAB:Single_2001}, respectively.
\begin{figure}[!h]
\centering
\subfloat[Class I-V cities]{
\includegraphics[width=0.5\textwidth]{Single_Class_1_5.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[Class I-IV cities]{
\includegraphics[width=0.5\textwidth]{Single_Class_1_4.png}
\label{FIG:loglogCumHaz_SmallCell}}
\\
\subfloat[Class I-II cities]{
\includegraphics[width=0.5\textwidth]{Single_Class_1_2.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[Class I cities]{
\includegraphics[width=0.5\textwidth]{Single_Class_1_1.png}
\label{FIG:loglogCumHaz_SmallCell}}
\caption{Plots of the actual and the predicted sizes over ranks for different sets of Indian cities in 2011
(Black square: observed sizes, blue circle: RO fits, red star: Pareto fits)}
\label{FIG:Single_2011}
\end{figure}
\begin{table}[h]
\caption{Parameter estimates for Indian Cities in different years, along with total number of cities ($N$) \\
and the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively.}
\centering
\begin{tabular}{l|rrr|rrr|rr}\hline
City & \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Classes & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
\multicolumn{9}{l}{2011}\\\hline
I & 468 & 100039 & 18394912 & 0.8552 & 0.1613 & 0.0308 & 0.9684 & 0.1252 \\
I-II & 942 & 50087 & 18394912 & 0.8791 & 0.1641 & 0.0313 & 0.9890 & 0.1470 \\
I-IV & 4001 & 10001 & 18394912 & 0.9122 & 0.1899 & 0.0310 & 1.0190 & 0.1923 \\
I-V & 5749 & 5001 & 18394912 & 0.9144 & 0.3051 & 0.0304 & 1.0856 & 0.4418 \\\hline
\multicolumn{9}{l}{2001}\\\hline
I & 394 & 100065 & 16434386 & 0.8842 & 0.0671 & 0.0275 & 0.9328 & 0.0506 \\
I-II & 798 & 50057 & 16434386 & 0.8940 & 0.0610 & 0.0256 & 0.9333 & 0.0462 \\
I-IV & 3307 & 10005 & 16434386 & 0.9079 & 0.1349 & 0.0229 & 0.9923 & 0.1560 \\
I-V & 4186 & 5002 & 16434386 & 0.9056 & 0.2411 & 0.0220 & 1.0535 & 0.3605 \\
\hline
\multicolumn{9}{l}{1991} \\\hline
I & 289 & 100235 & 9925891 & 0.8072 & 0.1374 & 0.0204 & 0.9129 & 0.0912 \\
I-II & 614 & 50048 & 9925891 & 0.8347 & 0.1166 & 0.0233 & 0.9182 & 0.0838 \\
I-IV & 2533 & 10005 & 9925891 & 0.8638 & 0.1823 & 0.0226 & 0.976 & 0.1731 \\
I-V & 3138 & 5004 & 9925891 & 0.8614 & 0.2964 & 0.0216 & 1.038 & 0.3642 \\
\hline
\end{tabular}
\label{TAB:Single_2001}
\end{table}
\vspace{1in}
The superiority of the goodness-of-fit in the proposed RO distribution, the DGBD, over the usual Pareto law
is clearly visible from Figure \ref{FIG:Single_2011} and Table \ref{TAB:Single_2001} in all cases.
In particular, when we consider a wider range of cities having sizes significantly small as well as significantly large
(i.e., enough observations and variety in both the high and low end of the ranks leading to larger $N$),
the Pareto model fails remarkably as also noted earlier
\cite{12kgbb, 13kgbb, 14kgbb, 15kgbb, 16kgbb, 17kgbb, 18kgbb, 19kgbb,luck17I,luck17U}.
In such cases, our proposal through rank ordering can provide excellent fit with very little prediction error (KS)
that is almost the same as that obtained with only large sizes (low-ends of the rank) of cities.
This observation, along with establishing the significance of the proposed approach over the existing Zipf's law
of city size distributions, also indicate the universality of the RO distribution
in predicting the city distributions with wider ranges (in both extremes) equally well.
\subsection{Chinese City-size Distribution}
China has an importance in urban planing and city-size analysis
as it is the most populous country in the world with 1,415,045,928 people according to a 2018 estimate \cite{world2018}.
According to \cite{world2018}, Shanghai is the most populous city with an estimate of 25,582, 138 in 2018.
We consider the sizes of all Chinese \textit{Cities} and \textit{Counties},
as defined in \cite{china data}, for the years 1990, 2000 and 2010.
Both the DGBD and the usual Pareto distribution is fitted for the sizes of these cities and counties separately
as well as to their combined pool.
The fitted sizes obtained in both the approaches are plotted in Figure \ref{FIG:China},
whereas the estimated parameters and KS measures are reported in Table \ref{TAB:China}.
\begin{figure}[!th]
\centering
\subfloat[1990, City]{
\includegraphics[width=0.33\textwidth]{1990China_Class_4_4.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[2000, City]{
\includegraphics[width=0.33\textwidth]{2000China_Class_4_4.png}
\label{FIG:loglogCumHaz_SmallCell}}
~
\subfloat[2010, City]{
\includegraphics[width=0.33\textwidth]{2010China_Class_4_4.png}
\label{FIG:loglogCumHaz_SmallCell}}
\\
\subfloat[1990, County]{
\includegraphics[width=0.33\textwidth]{1990China_Class_5_5.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[2000, County]{
\includegraphics[width=0.33\textwidth]{2000China_Class_5_5.png}
\label{FIG:loglogCumHaz_SmallCell}}
~
\subfloat[2010, County]{
\includegraphics[width=0.33\textwidth]{2010China_Class_5_5.png}
\label{FIG:loglogCumHaz_SmallCell}}
\\
\subfloat[1990, City+County]{
\includegraphics[width=0.33\textwidth]{1990China_Class_4_5.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[2000, City+County]{
\includegraphics[width=0.33\textwidth]{2000China_Class_4_5.png}
\label{FIG:loglogCumHaz_SmallCell}}
~
\subfloat[2010, City+County]{
\includegraphics[width=0.33\textwidth]{2010China_Class_4_5.png}
\label{FIG:loglogCumHaz_SmallCell}}
\caption{Plots of the actual and the predicted sizes over ranks for Chinese data
(Black square: observed sizes, blue circle: RO fits, red star: Pareto fits)}
\label{FIG:China}
\end{figure}
\begin{table}[h]
\caption{Estimated measures for Chinese Cities, , along with total number of cities ($N$)
and \\ the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively.}
\centering
\begin{tabular}{l|rrr|rrr|rr}\hline
& \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Year & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
\multicolumn{9}{l}{City Only}\\\hline
1990 & 731 & 7761 & 10429722 & 0.6939 & 0.5314 & 0.0081 & 1.037 & 0.4959 \\
2000 & 735 & 21293 & 16704306 & 0.6786 & 0.3695 & 0.0106 & 0.9212 & 0.2665 \\
2010 & 687 & 28783 & 20217748 & 0.6899 & 0.3786 & 0.0207 & 0.9346 & 0.2632 \\
\hline
\multicolumn{9}{l}{County Only}\\\hline
1990 & 1057 & 1006 & 266907 & 0.3073 & 0.4435 & 0.007 & 0.6005 & 0.1303 \\
2000 & 1074 & 4987 & 522369 & 0.2932 & 0.4016 & 0.0104 & 0.5603 & 0.1033 \\
2010 & 1028 & 5070 & 730810 & 0.2685 & 0.2801 & 0.0132 & 0.4627 & 0.0622 \\
\hline
\multicolumn{9}{l}{City + County Combined}\\\hline
1990 & 1788 & 1006 & 10429722 & 0.7382 & 0.5271 & 0.0188 & 1.0637 & 0.6371 \\
2000 & 1809 & 4987 & 16704306 & 0.7086 & 0.47 & 0.0106 & 1.0061 & 0.5065 \\
2010 & 1715 & 5070 & 20217748 & 0.7176 & 0.2333 & 0.0159 & 0.8675 & 0.1737 \\
\hline
\end{tabular}
\label{TAB:China}
\end{table}
One can note that, in all the cases the DGBD gives much improved fit to the city sizes
compared to the Zipfs law. This is in-line with the findings of \cite{kgbb}
where the authors have shown that the Pareto modeling works well for Chinese data
only if we consider the restricted sets of cities having sizes greater than 150000 or 200000 in 1990 and 2000, respectively.
On the other hand, the proposed RO modeling fits the whole data of all cities or counties
(as well as their combined pool) very accurately leading to a very small error value (KS).
Figure \ref{FIG:China} clearly demonstrates that, in all the cases,
the fits of the DGBD are also much better compared to the Pareto distribution in the observed data.
\subsection{Yearly City-size Distribution of USA}
USA is the 3rd most populous country in the world according to the 2018 ranking \cite{world2018}.
New York City, is the most populous city in USA (world city ranking 2)
with a present population of 8,550,405 and expected to reach 9 million by 2040.
However, USA has a developed economy in contrast to the developing ones in India and China; one may expect that USA city size distribution will be different from those of India and China if the cities with lower population are included in the study of city size distribution.
To illustrate the universality and global nature of the DGBD
in modeling city sizes of any kind of economy, let us now apply our proposal for USA city size data (estimated),
as already described in Section \ref{SEC:data}.
We consider only those entries having size greater than or equal 5,000 as a city.
The parameter estimates and goodness-of-fit measures (KS) obtained by
both the Pareto and proposed RO modeling are given in Table \ref{TAB:USA} for each years from 2010 to 2017;
results for 2010 are obtained both from the true census data as well as the estimated data for comparison purpose.
However, due to the resemblance of all these results,
only the plots of the corresponding fitted sizes for the years 2010 and 2017 (both with estimated data ) are shown in Figure \ref{FIG:USA}.
\begin{table}[h]
\caption{Estimated measures for USA Cities, , along with total number of cities ($N$)
and \\ the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively
(2010* denotes the census data; others are estimated data).}
\centering
\begin{tabular}{l|rrr|rrr|rr}\hline
& \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Year & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
2010* & 16330 & 5000 & 37253956 & 0.8667 & 0.3474 & 0.0276 & 1.0741 & 0.7011 \\\hline
2010 & 16397 & 5001 & 37327690 & 0.8663 & 0.3466 & 0.0274 & 1.0736 & 0.7020 \\
2011 & 16412 & 5000 & 37672654 & 0.8665 & 0.3504 & 0.0275 & 1.0757 & 0.7128 \\
2012 & 16418 & 5000 & 38019006 & 0.8667 & 0.3544 & 0.0277 & 1.0778 & 0.7235 \\
2013 & 16443 & 5000 & 38347383 & 0.8669 & 0.3586 & 0.0278 & 1.0800 & 0.7359 \\
2014 & 16436 & 5000 & 38701278 & 0.8670 & 0.3614 & 0.0279 & 1.0816 & 0.7436 \\
2015 & 16449 & 5000 & 39032444 & 0.8672 & 0.3650 & 0.0280 & 1.0836 & 0.7545 \\
2016 & 16456 & 5001 & 39296476 & 0.8673 & 0.3685 & 0.0281 & 1.0853 & 0.7640 \\
2017 & 16459 & 5002 & 39536653 & 0.8673 & 0.3709 & 0.0282 & 1.0865 & 0.7708 \\
\hline
\end{tabular}
\label{TAB:USA}
\end{table}
\begin{figure}[!h]
\centering
\subfloat[2010]{
\includegraphics[width=0.47\textwidth]{USA_POPESTIMATE2010.png}
\label{FIG:loglogCumHaz_Veteran}}
~
\subfloat[2017]{
\includegraphics[width=0.47\textwidth]{USA_POPESTIMATE2017.png}
\label{FIG:loglogCumHaz_SmallCell}}
\caption{Plot of the actual and the predicted sizes over ranks for USA city sizes (estimated) in 2010 and 2017
(Black square: observed sizes, blue circle: RO fits, red star: Pareto fits)}
\label{FIG:USA}
\end{figure}
Once again our RO approach fits the USA data much more accurately than the existing Pareto law.
For each year, the accuracy obtained by the RO approach is extremely small and quite stable,
but the accuracy of Pareto modeling is about 25-30 times lower
and also increases over the year
due to the increase in the variation of city sizes.
\subsection{City-size Distribution of Brazil}
We now consider another large developing country, namely Brazil, from South American continent.
In the population wise global ranking of 2018 estimate,
Brazil is in 5th position with a population of 211,179,426 \cite{brazilcity}.
We use the census data of the Brazilian city sizes as described in Section \ref{SEC:data}
and apply the RO modeling along with the usual Pareto modeling
for the cities having at least 20,000 inhabitants (as per the availability of data).
Resulting estimates are shown in Table \ref{TAB:Brazil},
which again show appreciably improved fit by our proposal with the DGBD
as compared to the usual Pareto law.
The fitted values obtained by the DGBD
is again seen to be much closer to the original city-sizes compared to the power law fits. The plots are not presented here for brevity.
\begin{table}[h]
\caption{Estimated measures for Brazilian Cities with at least 20000 inhabitants, along with total number of cities ($N$)
and \\ the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively.
(* denotes estimated data)}
\centering
\begin{tabular}{l|rrr|rrr|rr}\hline
& \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Year & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
1991 & 709 & 20002 & 9412894 & 0.886 & 0.079 & 0.021 & 0.947 & 0.096 \\
2000 & 922 & 20022 & 9813187 & 0.845 & 0.157 & 0.012 & 0.962 & 0.176 \\
2010 & 1096 & 20002 & 11152344 & 0.837 & 0.170 & 0.011 & 0.964 & 0.198 \\
2017* & 1204 & 20030 & 11998090 & 0.833 & 0.190 & 0.013 & 0.974 & 0.228 \\
\hline
\end{tabular}
\label{TAB:Brazil}
\end{table}
\subsection{Distribution Cities in Three African Countries}
Next we try to test if the proposed approach fits well even for the African countries with underdeveloped economy.
In March 2013, Africa was identified as the world's poorest inhabited continent.
Although, in 2017 the African Development Bank reported Africa to be the world's second-fastest growing economy,
and estimated that average growth will rebound to 3.4\% in 2017,
while the growth is expected to increase to 4.3\% in 2018 \cite{afr2017}.
In this respect, it is interesting to study the city size distribution of some countries in Africa;
however, in most such countries, proper census data are rarely available.
We here present the analysis for three such countries, namely Uganda, Algeria and Sudan,
where the latest census data have been found for the years 2014, 2008 and 2008, respectively, from \cite{africa}.
Based on the availability of these data, we could only consider the cities having at least 15,000,
13,000 and 20,000 inhabitants, respectively, for the three countries.
We would like to make a note here that,
although Nigeria is going to become a good choice for the urban dwellers in coming years \cite{desa},
due to the unavailability of the data to the authors,
the city-size distribution of Nigerian cities could not studied here.
Algeria is Africa's largest country and has several cities with population of over a million residents.
The country's capital and largest city is Algiers with a population of around 3.5 million in 2008;
the city of Oran is the second-largest having around 650,000 residents.
Uganda is the 81st largest country by area in the world, but 36th in population.
The largest city in Uganda is Kampala, with a population of 1,353,189 people in 2014.
Uganda has one city with more than a million people, two cities with people between 100,000 to 1 million,
but most cities (62) have smaller sizes between 10,000 to 100,000.
Sudan has two cities with more than a million people, 19 cities of sizes between 100,000 to 1 million,
and 42 cities of sizes between 10,000 to 100,000.
The largest city in Sudan is Khartoum, with a population of 1,974,647 people in 2008.
Applying both the RO and the Pareto modeling to the city-size data of these three countries,
we report the resulting estimates and fitness measures in Table \ref{TAB:Africa}.
For Uganda and Algeria, the proposed RO modeling using the DGBD performs as a close contender of the Pareto law
which also gives reasonably well fit; note that the estimated $b$ parameter in the DGBD is close to zero
in both the cases indicating the good behavior of the Pareto modeling.
But, for Sudan, the estimate of $b$ is slightly away from zero and the resulting DGBD
provides a better fit with almost half the KS error as compared to the power law approach.
\begin{table}[h]
\caption{Estimated measures for three African Countries, along with total number of cities ($N$)
and \\ the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively.}
\centering
\begin{tabular}{ll|rrr|rrr|rr}\hline
& & \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Country & Year & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
Uganda & 2014 & 105 & 15056 & 1507080 & 0.917 & 0.029 & 0.048 & 0.936 & 0.046 \\
Algeria & 2008 & 180 & 13029 & 2364230 & 0.800 & 0.024 & 0.044 & 0.798 & 0.046 \\
Sudan & 2008 & 63 & 20302 & 1849659 & 1.033 & 0.187 & 0.043 & 1.157 & 0.094 \\
\hline
\end{tabular}
\label{TAB:Africa}
\end{table}
\subsection{Distribution of Australian Agglomerations}
We next focus our attention on the 6th largest nation in the world, with very low population density;
we test the RO model, through DGBD, to characterize the sizes of Urban Agglomeration (UA) of Australia.
Australia has one of the most urbanized societies in the world.
As of 2018, Australia has an estimated population of 24.77 million, up from the official 2011 census results of 21.5 million.
It is the most populous country in Oceania, three times more populous than its neighbor Papua New Guinea (8.2 million)
and 5 times more populous than New Zealand (4.5 million), yet ranks only 54th in the world in terms of its population.
As an illustration, let us consider the data on the sizes of 101 UA in the years 2011 and 2016, obtained from \cite{australia}.
When these Australian UA sizes are modeled by the DGBD and the Pareto distributions,
as before, the DGBD provides slightly improved fit compared to the Pareto fit;
see Table \ref{TAB:Australia}. All parameter estimates are surprisingly stable in both the year
indicating the stability in the urban dynamics of Australia even in the 5 year span.
\begin{table}[h]
\caption{Estimated measures for Australian UA, along with total numbers ($N$)
and their minimum and \\ maximum sizes $x_{\min}, x_{\max}$, respectively.}
\centering
\begin{tabular}{l|rrr|rrr|rr}\hline
& \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Year & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
2011 & 101 & 6616 & 4034911 & 1.259 & 0.418 & 0.104 & 1.420 & 0.158\\
2016 & 101 & 10288 & 4446807 & 1.267 & 0.434 & 0.106 & 1.434 & 0.163\\
\hline
\end{tabular}
\label{TAB:Australia}
\end{table}
\subsection{City-size distributions of Three European Countries}
\label{SEC:Europe}
We now consider three smaller European countries, namely Italy, Sweden and Switzerland, all having strong developed economy.
The current population of Italy is 59,277,324, based on the latest United Nations estimates \cite{un},
which is equivalent to 0.78\% of the total world population (ranks 23rd).
But, as large as 71.8\% of the population of Italy is urban (42,587,390 people in 2018 estimate);
the most populated city Roma (Rome) has population 2,648,843, followed by Milano (Milan) with 1,305,591 \cite{italy}.
In the Scandinavian country Sweden, almost all of the population live in urban areas;
it has one city with more than a million people (Stockholm, with population 1,515,017 people),
9 cities of sizes between 100,000 to 1 million,
and as many as 141 small cities of population between 10,000 to 100,000 \cite{sweden}.
Our third choice Switzerland, however, has no city with more than a million people,
5 cities having 100,000 to 1 million people each, and 145 cities having 10,000 to 100,000 people;
the largest city Zurich has a population of only 341,730 residents \cite{switzerland}.
We repeat our analysis
with the census data of Italy for the years 1981, 1991, 2001 and 2011,
and the corresponding estimated city sizes for the year 2017,
restricting our attention to the human settlements having at least 50,000 inhabitants.
We also apply the same for the yearly official estimate of city sizes of Sweden,
from the year 1990 to 2015 and 2017, having at least 10,000 inhabitants.
For Switzerland we fit the distribution for the census data of the year 1980, 1990 and 2000,
and the official estimates for the year 2010 and 2017,
where again we restrict only to the settlements having at least 10,000 human inhabitants.
The resulting estimates and the KS measure of fits are provided in Table \ref{TAB:Italy}
for all the three countries.
\begin{table}[h]
\caption{Estimated measures for Cities of Italy, Sweden and Switzerland, along with total number of cities ($N$) \\
and the sizes $x_{\min}, x_{\max}$ of the smallest and the largest cities, respectively.
(* denotes estimated data)}
\centering
\begin{tabular}{l|rrr|rrr|rr}\hline
& \multicolumn{3}{c}{Summary} & \multicolumn{3}{|c|}{RO Fit} & \multicolumn{2}{c}{Pareto Fit} \\
Year & $N$ & $x_{\min}$ & $x_{\max}$ & $\widehat{a}$ & $\widehat{b}$ & KS & $\nu$ & KS \\\hline
\multicolumn{9}{l}{Italy, having at least 50000 inhabitants}\\\hline
1981 & 121 & 50666 & 2840259 & 0.898 & $-$0.085 & 0.018 & 0.820 & 0.072 \\
1991 & 128 & 50018 & 2775250 & 0.886 & $-$0.122 & 0.017 & 0.777 & 0.086 \\
2001 & 134 & 50032 & 2546804 & 0.865 & $-$0.130 & 0.018 & 0.749 & 0.088 \\
2011 & 141 & 50013 & 2617175 & 0.857 & $-$0.140 & 0.017 & 0.731 & 0.091 \\
2017* & 144 & 50645 & 2872800 & 0.872 & $-$0.152 & 0.018 & 0.733 & 0.097 \\
\hline
\multicolumn{9}{l}{Sweden, having at least 10000 inhabitants}\\\hline
1990* & 107 & 10041 & 1209557 & 1.061 & $-$0.202 & 0.052 & 0.863 & 0.160 \\
1995* & 111 & 10000 & 1275492 & 1.066 & $-$0.204 & 0.051 & 0.870 & 0.158 \\
2000* & 110 & 10267 & 1344798 & 1.085 & $-$0.213 & 0.050 & 0.884 & 0.162 \\
2005* & 112 & 10091 & 1388306 & 1.085 & $-$0.204 & 0.050 & 0.892 & 0.159 \\
2010* & 117 & 10037 & 1517006 & 1.096 & $-$0.209 & 0.049 & 0.901 & 0.161 \\
2015* & 123 & 10023 & 1654623 & 1.102 & $-$0.210 & 0.051 & 0.908 & 0.163 \\
2017* & 125 & 10028 & 1705718 & 1.103 & $-$0.209 & 0.051 & 0.912 & 0.162 \\\hline
\multicolumn{9}{l}{Switzerland, having at least 10000 inhabitants}\\\hline
1980 & 103 & 10001 & 369522 & 0.846 & $-$0.128 & 0.022 & 0.729 & 0.093 \\
1990 & 118 & 10180 & 365043 & 0.827 & $-$0.142 & 0.021 & 0.698 & 0.096 \\
2000 & 125 & 10142 & 363273 & 0.807 & $-$0.138 & 0.018 & 0.680 & 0.090 \\
2010* & 143 & 10003 & 372857 & 0.778 & $-$0.125 & 0.019 & 0.659 & 0.083 \\
2017* & 155 & 10007 & 409241 & 0.772 & $-$0.117 & 0.020 & 0.659 & 0.081 \\
\hline
\end{tabular}
\label{TAB:Italy}
\end{table}
Although Table \ref{TAB:Italy} illustrates significantly improved goodness of the RO modeling over the usual Pareto law, but an interesting observation of negative $\hat{b}$ values, unlike the $\hat{b}$ values of other countries, entails some explanation.
Among various possible functional forms of the RO distributional assumption (Figure \ref{FIG:RO_posA}),
in case the rank is defined in decreasing order of importance (as in the case of city sizes here),
it is intuitive to consider those RO distributions (in particular, those DGBDs)
providing a monotonous non-increasing structure
which indeed restricts the parameter choice as $(a, b)\geq 0$.
However, the DGBD with $a>0$ but very small negative value for $b$,
as obtained for these three European countries, indeed has an almost monotone structure
except for the extreme higher end (smaller sizes) where it increases slightly; see Figure \ref{FIG:RO_negB}.
This suggests that to get the best fit for the city size distributions of these 3 European countries we are indeed loosing the monotonicity requirement while considering
the cities with
smaller sizes. The physical reasoning behind this observation lies within the population data; the data exhibits that the extremely small cities of these 3 countries are of almost equal sizes and it is difficult to individually
rank them ( only in the higher rank end).
Note that, the major part of the city sizes including the low-rank end has a strictly monotone structure
allowing their easy discrimination from the smaller cities.
So, accommodating the negative values of the parameters, we can very easily discriminate such countries having
more uniform city size distributions at the lower end from other countries
having clearly distinct city sizes across its full spectrum
(like India, China, USA, Brazil).
On the contrary, if one wants to strictly retain
the monotone structure in the full spectrum of the city-size distribution,
then the best fit for all three European countries is given by the power law (restricting $b$ to be non-negative). It may be noted that, unlike other countries studied above, the KS error in Pareto fit is reasonably small
for Italy, Sweden and Switzerland.
\subsection{World-wide Size distributions of Cities and Countries in the year 2018}
If we focus on the word-wide city size ranking, it is noted that at present,
Tokyo is the world’s largest city with an agglomeration of 37 million inhabitants,
followed by New Delhi with 29 million, Shanghai with 26 million,
and Mexico City and Sao Paulo, each with around 22 million inhabitants \cite{un}.
Today, Cairo, Mumbai, Beijing and Dhaka all have close to 20 million inhabitants.
By 2020, Tokyo’s population is projected to begin to decline,
while Delhi is projected to continue growing and to become the most populous city in the world around 2028.
By 2030, the world is projected to have 43 megacities with more than 10 million inhabitants,
most of them in developing regions.
However, some of the fastest-growing urban agglomerations are cities with fewer than 1 million inhabitants,
many of them located in Asia and Africa. While one in eight people live in 33 megacities worldwide,
close to half of the world’s urban dwellers reside in much smaller settlements with fewer than 500,000 inhabitants.
As the world continues to urbanize, sustainable development depends increasingly on
the successful management of urban growth, especially in low-income and lower-middle-income countries
where the pace of urbanization is projected to be the fastest.
Understanding the key trends in urbanization likely to unfold over the coming years is very crucial
and this motivated us to analyze the size distributions of the top 20 cities of 230 countries
and dependent territories around the world for the year 2018; see \cite{world2018} for their details.
As before, we have restricted our modeling to the cities with a minimum number of 5000 habitants,
but still the filtered data of 3371 cities have a very wider spectrum of population varying from 5017 to 37,468,302.
These are the estimated sizes projected from the actual census data collected in the last census years
(which is different for different countries).
The complete listing shows that, although Tokyo tops the rank-wise list of top cities around the world,
there are many cities from India, China and USA.
We combine the population of the above mentioned 3371 cities across the world
and apply the proposed RO fitting along with the usual Pareto law.
Since there are wide varieties of city sizes, as expected,
the Pareto modeling fails miserably (KS measure 7.474; $\nu=1.504$).
But, the universal RO approach, using the DGBD, again provides much better fit for the distribution of world cities
with an KS error measure of only 0.04 (Figure \ref{FIG:WorldA}); the parameters of the fitted DGBD are
$a=0.632$ and $b=1.963$. Note that, here we have a significantly larger values of $b$
indicating wider spectrum of the city sizes.
\begin{figure}[!h]
\centering
\subfloat[Top 20 cities within each Countries]{
\includegraphics[width=0.5\textwidth]{World_City.png}
\label{FIG:WorldA}}
\subfloat[Cities with size greater than 3 Lakhs]{
\includegraphics[width=0.5\textwidth]{World_City3lakh_2018.png}
\label{FIG:WorldB}}
\caption{Plot of the actual and the predicted sizes over ranks for
Worldwide city sizes (estimated) in 2018 (Black square: observed sizes, blue circle: RO fits, red star: Pareto fits)}
\label{FIG:World}
\end{figure}
Due to the unavailability of any global pattern for city size distribution, a commonly used practice is to study the distribution of city sizes, within the framework for Pareto law, considering the cities with sizes larger than a specific limit.
Let us now consider all cities (171) around the world having populations of at least 300000 in the year 2018 and
sizes of these 171 cities are fitted by the DGBD and the Pareto distributions.
As expected, Pareto modeling gives a reasonable fit to these high-end city sizes (KS=0.703, $\nu=1.178$)
but the proposed RO modeling, via DGBD, yields even a better fit than Pareto with a KS value of 0.041
($\widehat{a}=0.426$, $\widehat{b}=1.302$);
see Figure \ref{FIG:WorldB} for the corresponding fitted sizes along with their original population.
This further supports our claim on the universality of the rank order modeling, using DGBD, in analyzing any kind of city size distributions.
Finally, we end our illustrations with a remark on the modeling of the country sizes.
We note that while analyzing the populations of 226 major countries ( population $>$ 5000)
across the world as per the 2018 estimate given in \cite{world2018},
the Pareto law again fails tremendously (KS=14.091, $\nu=2.257$) due to their wider spectrum.
On the other hand, the RO fit is again much better leading to the KS error of only $0.061$
($\widehat{a}=0.976$, $\widehat{b}=1.499$). This result justifies further the universal nature of the DGBD.
Probably, the countries around the world have a similar distributional structure as that of the cities within a country
and we should be able to explain them as well with a city size law.
\section{Discrimination and Evolution of the City Size Distributions}
\label{SEC:discrimination}
An important usefulness of the particular RO distribution, namely the DGBD,
is its ability to characterize the city size distribution
through the two parameters $(a, b)$.
It is noted in Sec.~\ref{SEC:ROD} that these two parameters $(a,b)$ completely define the shape of
the RO distribution (Fig.~\ref{FIG:RO_posA})
and hence the characteristics of the city sizes along with the entropy of the underlying process (Fig.~\ref{FIG:Entropy_N20}).
Therefore, their estimated values in the cases of different countries
enable us to characterize and discriminate the corresponding city-size distributions
and the underlying uncertainty through its entropy.
Similarly, the changes observed in these estimated parameters at different time points for any fixed country
also indicate the evolution of the corresponding process within that country
in terms of the shape and the uncertainty of its city-size distribution.
For illustration, in Figure \ref{FIG:AB}(top),
we have plotted the estimated values of $(a, b)$ for all our illustrative examples
with 11 countries across the seven continents at different years,
which provides a vividly clear comparison of their city-size distributions and their evolution.
The heat-map of the entropy values of the corresponding fitted DGBDs
are also presented in Figure \ref{FIG:AB}(bottom) for entropic analyses.
\begin{figure}[!h]
\centering
\includegraphics[width=18cm]{all_ab.png}
\includegraphics[width=18cm]{All_entropy.png}
\caption{(Top) Estimated parameters $(a,b)$ of the fitted DGBDs for different countries and different years and (Bottom) the corresponding entropy values. }
\label{FIG:AB}
\end{figure}
One can clearly see from Figure \ref{FIG:AB} that the large countries like India, China, USA or Brazil
have positive $b$ values indicating a completely different class from the smaller but developed countries like
Italy, Sweden or Switzerland with slightly negative values of $b$.
Moreover, the first group of countries clearly belong to a class of
higher entropy and hence have more uniform city size distributions compared to that of the second group;
but the spectrum of city sizes within the first group is much wider than those in the second group due to their larger number of cities.
Further, within the first group of countries, China has a clearly different city-size distribution
compared to those of India and US; the second group is structurally more similar.
China has the lowest entropy among these three with USA having the highest entropy value;
this is because the urban population of USA is well distributed (most uniformly) within several cities. This pattern becomes less uniform for India and China,
where the city sizes are
more distinct.
In terms of evolution through the years, both India and China have a drastic change in its city-size distribution
between the twentieth and twenty-first centuries, which can be seen from differences in their $(a, b)$ estimates.
In the twenty-first century, both China and USA become quite stable in terms of their city size distributions,
but that of India is still changing significantly in shape from the year 2001 to 2011 with a very little increase in entropy.
The city-size distribution of Brazil is found to be in the same entropy range as China over the years with a small increasing trend,
but structurally coming closer to the India-USA cluster after an abrupt change in the last decades of twentieth century.
The city sizes of European countries like Italy, Sweden and Switzerland, on the contrary,
have much smaller distribution spectrum but greater variability in their city sizes.
They are completely separated from the group of above mentioned countries through its estimated values of $b$
which are (equal or) very close to zero from the negative side due to the uniformity among the smaller
cities within these countries (as already discussed in Section \ref{SEC:Europe}).
Among them, the city-size distribution of Italy and Switzerland are pretty similar
although Switzerland has slightly greater entropy value due to greater uniformity in its cities' sizes.
For both of these countries, the entropy is increasing with time indicating a possible growth mechanism of the cities of these countries --
sizes of smaller cities are growing at a faster rate than those of the larger cities.
Such a change is more prominent in Switzerland than in Italy where big cities like Milan and Rome is still dominating.
Sweden, on the other hand, has quite a different city size distribution with comparatively lower entropy as
handful number of large cities in Sweden accommodate majority of its urban population.
African countries are slightly in between these two groups,
with Sudan having significantly lower entropy and randomness in its city size distribution followed by Uganda and Algeria, respectively.
The urban populations of Algeria is comparatively well distributed between several cities (higher entropy),
whereas the majority of the urban population of Sudan is concentrated in few largest cities
(45\% in top 3 cities, and 70\% in top 10 cities among a total of 63 cities).
The distribution of Australian Agglomeration, although structurally completely different from all other countries considered,
are similar to the city size distribution of Sweden and Sudan; a majority of its population is concentrated only in a few large UA
(57\% in top 6 UA among the total of 400 UAs). This leads to the lowest possible entropy for their size distribution
characterized with the wide range of the spectrum of Australian human settlements having significantly distinct sizes.
This exhaustive study helps us to conclude that the DGBD, along with its entropy formulation,
can characterize the evolution of the urbanization and can also discriminate
between the city size distributions of different countries.
The RO approach, with its two parameters, provides greater clarity for the characterization and discrimination between the city size distributions of different countries than the one parameter power law.
\section{Conclusions and Discussions}
There are many competing criteria in the study of cities. This has made the science of city planning such a challenging one.
Despite the enormous complexity and diversity of human behavior and extraordinary geographic variability,
we have shown that the size distribution of cities around the world follows a universal law.
A two-parameter distribution, with the parameter values varying within a very short range,
can illustrate the empirical data for the city sizes across the world.
Moreover, the entropy analysis presented here is used to quantify the variation in the distribution of city sizes.
The underlying city size distribution of the countries with a wider spread of city sizes, i.e. with a few extremely populous cities along with several smaller cities, can be characterized with low entropy.
This is possibly connected with the lesser uncertainty in those country's human settlements
focusing more on
larger cities (e.g., Swedish cities, Australian UA).
On the other hand, the uncertainty and internal movement within cities are
expected to be higher for countries with almost equal city sizes
or with lower spread of city sizes; they further lead to comparatively higher entropy of the underlying DGBD
(e.g., Switzerland, Italy or China).
Among several possible extensions of our work,
the immediate one should be an in-depth analysis of the drivers of the city dynamics
within different countries. Here, we have already studied the discrimination and evolution of countries' city size distribution, through the entropy analysis and two fitted parameter values of the proposed RO distribution, the DGBD.
These two parameters can be thought of as two $sensors$
of the city dynamics, those could be further associated with observable economic, social or environmental factors;
these observables can then be controlled for any suitable planning of the city-sizes in a country as per the requirements.
Furthermore, the possible linking of these two sensor parameters of city populations
with the underlying entropy and the maximum entropy principle might shed further light on their physical significance.
Such an analysis may be explored with the inverse problem of characterizing the DGBD
as a maximum entropy distribution under appropriate constraints.
We hope to pursue some of these interesting and useful extensions in future
and believe that the proposed universal rank-order modeling for the city sizes
(comprising the whole data set, starting from smaller human settlements to the most populous city in the country)
will open up new avenues of research in this area
\nopagebreak
\begin{acknowledgements}
The authors wish to thank the Editor and two anonymous referees for careful reading of the manuscript
and several valuable constructive suggestions which have significantly improved the paper.
The work of the first author, AG, is supported by the INSPIRE Faculty Research Grant from
the Department of Science and Technology, Government of India, India.
\end{acknowledgements}
|
\section{Main result}\label{main-result}
Let $X$ be a locally compact space with countable base, let $\mathcal B$ denote the $\sigma$-algebra of all
Borel sets in $X$, and let~$\mathcal B(X)$, $\C(X)$ respectively be the set of all numerical functions on $X$
which are Borel measurable, continuous and real respectively. As usual, given a set $\mathcal F$ of functions
on $X$, a superscript ``+'', a subscript ``$b$'' respectively will indicate that we consider functions in $\mathcal F$
which are positive, bounded respectively.
Let $\mathcal M(X)$ denote the set of all positive (Radon) measures on $X$.
Let $\mathfrak X=(\Omega, \mathfrak M, \mathfrak M_t, X_t,\theta_t, P^x)$ be a
Hunt process on $X$ (see \cite[p.\,45]{BG}).
Let $ P=(P_t)_{t>0}$ denote the transition semigroup of $\mathfrak X$, that is,
$P_tf(x)=E^x(f\circ X_t) $ for all $t>0$, $f\in \mathcal B^+(X)$ and $x\in X$.
We assume that the Hunt process $\mathfrak X$ is \emph{nice} in the following sense.
Its set
\begin{equation}\label{Hunt}
\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}:=\{w\in \mathcal B^+(X)\colon \sup_{t>0}P_tw=w\}
\end{equation}
of (Borel measurable) excessive functions has the following properties:
\begin{itemize}
\item[\rm (C)] Continuity:
Every $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ is the supremum of its minorants in $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\cap\C(X)$.
\item[\rm (S)] Separation:
$\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ is linearly separating, that is, for all $x\ne y$ and $\gamma>0$, there exists a function $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$
such that $w(x)\ne \gamma w(y)$.
\item[\rm (T)] Transience:
There are strictly positive functions $v,w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\cap \C(X)$ such that the quotient $v/w$
tends to $0$ at~infinity.
\end{itemize}
Let us observe that (C) trivially holds if the kernels $P_t$, $t>0$, or at least
the corresponding resolvent kernels $V_\lambda:=\int_0^\infty e^{-\lambda t}P_t\,dt$, $\lambda>0$,
are strong Feller, that is,
map $\mathcal B_b(X)$ into $\C_b(X)$.
For every set $A$ in $X$, the first entry time $D_A$ and the first hitting time $T_A$
are defined for $\omega\in\Omega$ by
\begin{equation*}
D_A(\omega):=\inf\{s\ge 0\colon X_s(\omega)\in A\} \quad\mbox{ and }\quad
T_A(\omega):=\inf\{s> 0\colon X_s(\omega)\in A\}.
\end{equation*}
Let ${\mathcal U}_c$ be the set of all relatively compact open sets $V$ in $X$, $V^c:=X\setminus V$.
A~numerical function $u\ge 0$ is called \emph{nearly hyperharmonic} if
\begin{equation}\label{def-nearly}
\int^\ast u\circ X_{D_{V^c}} \,dP^x\le u(x) \quad \mbox { for all $ x\in X$ and
neighborhoods $V\in {\mathcal U}_c$ of $x$}.
\end{equation}
Clearly, the set $\N$ (denoted by $\N^+$ in \cite{HN-mertens}) of such functions
is a convex cone which contains $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ and is stable under increasing limits and
\emph{arbitrary} infima. Moreover, it contains \emph{every} numerical function $u\ge 0$
which vanishes outside a set~$E$ which is polar, that is, satisfies $T_E=\infty$ almost surely.
For space-time Brownian motion on ${\real^d}\times \mathbbm{R}$,
every function $u\colon {\real^d}\times \mathbbm{R}\to [0,\infty]$ satisfying $u(x,t)\le u(x',t')$,
whenever $t\le t'$, is nearly hyperharmonic.
The purpose of this paper is to give a short, complete and understandable proof
for the following statement (where the implications (3)\,$\Rightarrow$\,(2)\,$\Rightarrow$\,(1)
hold trivially).
\begin{theorem}\label{main}
For every $u\in \mathcal B^+(X)$ the following statements are equivalent:
\begin{itemize}
\item[\rm (1)]
The function $u$ is nearly hyperharmonic.
\item[\rm (2)]
The function $u$ is the infimum of its excessive majorants.
\item[\rm (3)] For all $\mu\in \mathcal M(X)$
such that $\mu(A) +\int_{X\setminus A} w\,d\mu<\infty$
for some $A\in\mathcal B$ and majorant $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ of $u$,
\begin{equation}\label{int-inf}
\int u\,d\mu=\inf\{\int w\,d\mu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}, \ w\ge u\}.
\end{equation}
\end{itemize}
In particular, for every $\varphi\in\mathcal B^+(X)$,
the function $R_\varphi:=\inf\{w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge \varphi\}$
is the smallest nearly hyperharmonic majorant of $\varphi$.
\end{theorem}
\begin{remark}\label{KK}
{\rm
Of course, (\ref{int-inf}) trivially holds if $\int u\,d\mu=\infty$ (we take $w=\infty $).
Since $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ is $\wedge$-stable, the set of all $\mu\in \mathcal M(X)$
satisfying (\ref{int-inf}) is a convex cone.
If~$\varepsilon>0$
and $A\in \mathcal B$ such that
$\mu(A)+\int_{X\setminus A} w\,d\mu<\infty$,
there is a union $A'$ of $A$ with a compact in $X\setminus A$ such that $\mu(A')<\infty$ and $\int_{X\setminus A'} w\,d\mu<\varepsilon$.
}
\end{remark}
In fact, we shall finally prove that, for functions $u\colon X\to [0,\infty]$,
the equivalence (1)\,$\Leftrightarrow$\,(3)
already holds if $u$ is nearly Borel measurable (Theorem \ref{nearly-Borel}) and that
(1)\,$\Leftrightarrow$\,(2) even holds if $u$ is only supposed
to be equal to a Borel measurable function
outside a polar set (Theorem \ref{tilde-Borel}).
Moreover, assuming that $u$ is nearly hyperharmonic and equal to a universally
measurable function outside a polar set,
we characterize the validity of (2) in various ways
(Corollary~\ref{general-2}).
Analogous statements can be found for different settings and functions, which there are called
\emph{strongly supermedian},
in \cite{mertens,feyel-rep,
feyel-fine, beznea-boboc-feyel, beznea-boboc-book},
but the proofs given therein seem to be either incomprehensible
or incomplete (see \cite{mertens, feyel-rep,feyel-fine}) or, as in \cite{beznea-boboc-feyel} and
\hbox{\cite[Section 4] {beznea-boboc-book}},
very long and delicate.
The main novelties of our approach are
\begin{itemize}
\item
the insight that for a proof of inequalities $u\ge \eta \inf\{w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge u\} $ for
$\eta\in (0,1)$ it suffices to consider the
special case, where starting in~$\{u<\infty\}$
the expected number of times
the process $\mathfrak X$ visits
the set~of points~$y\in X$, where $\hat u(y):=\liminf_{z\to y} u(z)<u(y)$,
is finite.
\item
the consequent use of (only) the strong Markov property,
\item
the verification of the equalities in (2) and (3) first for nearly hyperharmonic functions
$u\in\mathcal B^+(X)$ having a certain finiteness property, which then implies~(!) that
\emph{every} nearly hyperharmonic $u\in\mathcal B^+(X)$
has this property,
\item
the equality (\ref{int-inf}) not only for
measures $\mu$ satisfying $\int w\,d\mu<\infty$ for some excessive majorant $w$ of $u$,
but also for all finite measures~$\mu$.
\end{itemize}
Let us observe that the additional statement
in Theorem \ref{main} is not only of interest
in its own right,
but also because of the following consequence (see \cite[Propositions~2.4, 2.5 and Theorem 3.1]{HN-mertens}).
\begin{corollary}\label{cons}
Let $\varphi\in\mathcal B^+(X)$. Then $R_\varphi=\varphi\vee \widehat{R_\varphi}\in \mathcal B^+(X)$ and
\begin{equation*}
R_\varphi(x)=\sup\{\int \varphi\circ X_{D_{V^c}}\,dP^x\colon x\in V\in {\mathcal U}_c\}, \qquad x\in X.
\end{equation*}
\end{corollary}
In Section \ref{sec-prelim} we discuss the close relationship between nice Hunt processes and balayage spaces
and establish a~crucial inequality for nearly hyperharmonic functions (Lemma \ref{PFuh}).
In Section \ref{sec-special}
we treat the special case indicated above.
In Section~\ref{sec-general} we shall see very quickly that the equality $R_u=u$ for arbitrary nearly hyperharmonic functions
$u\in \mathcal B^+(X)$ is a~consequence of our result for the special case and yields the additional statement in Theorem \ref{main}.
The implication (1)\,$\Rightarrow$\,(3)
is derived in Section~\ref{remaining}, and
in Section \ref{weaker} we present our results under weaker measurability assumptions.
In Section \ref{standard} we briefly indicate the use of our approach in the
general setting of right processes.
\section{Preliminaries}\label{sec-prelim}
Let us first recall the following. Let $\W$ be any
convex cone of positive numerical functions on $X$ having the properties stated in (C), (S)
and (T) for $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ (so that every function
in $\W$ is lower semicontinuous).
The ($\W$-)\emph{fine topology} on $X$ is the coarsest topology on $X$ which is at least as fine
as the initial topology and such that every function in $\W$ is continuous. Given
$\varphi\colon X\to [0,\infty] $,
let $\hat \varphi$, $\hat \varphi^f$ resp.\ denote the largest lower semicontinuous,
finely lower semicontinuous resp.\ minorant of~$\varphi$.
Then $(X,\W)$ is called a \emph{balayage space} provided the following hold
(see \cite{BH,H-course} and \cite[Appendix 8.1]{HN-unavoidable-hameasures}):
\begin{itemize}
\item[\rm (i)]
If $v_n\in \W$, $v_n\uparrow v$, then $v\in \W$.
\item[\rm (ii)]
If $\V\subset \W$, then $\widehat{\inf \V}^f\in\W$.
\item[\rm (iii)]
If $u,v',v''\in\W$, $u\le v'+v''$, then there exist $u',u''\in\W$ such that $u=u'+u''$ and $u'\le v'$, $u''\le v''$.
\end{itemize}
By \cite[II.4.9]{BH} (see also \cite[Corollary 2.3.8]{H-course}), for our nice Hunt process
$\mathfrak X$, the pair $(X,\mathcal E_{\mathfrak X}} % {E_{\mathbbm P})$ is a balayage space (of course, $\lim_{t\to 0} P_tf=f$ for
every $f\in \C_b(X)$ by right continuity of the paths).
So we may use results obtained in \cite{BH} and in the recent paper \cite{HN-mertens}.
\begin{remarks} {\rm
1. We note that, conversely, for every balayage space $(X,\W)$ with $1\in\W$, there exists a~corresponding
nice Hunt process (see \cite[IV.8.1]{BH}). For that matter, the condition \hbox{$1\in\W$} is not really restrictive
since, given any balayage space $(X,\W)$, the standard normalization $\widetilde \W:=(1/\tilde w)\W$ with
any strictly positive $\tilde w\in \W\cap \C(X)$ leads to a~balayage space~$(X,\widetilde \W)$
with $1\in \widetilde \W$.
2. A characterization by harmonic kernels reveals that the notion of a balayage space generalizes the
notion of a $\mathcal P$-harmonic space. Therefore the theory of balayage spaces is known to
cover the potential theory for very general partial differential operators of second order
(see, for instance, \cite{GH1}).
}
\end{remarks}
Of course, for any numerical function $\varphi\ge 0$ on $X$,
\begin{equation*}
R_\varphi:=\inf\{w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge \varphi\} \in \N,
\end{equation*}
and, by property (ii) of balayage spaces,
\begin{equation}\label{red-es}
\hat R_\varphi:=\widehat{R_\varphi}=\widehat{R_\varphi}^f\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}.
\end{equation}
By \cite[Proposition 2.2 and p.\,6]{HN-mertens}, we know even that $\hat u=\hat u^f\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ for all $u\in \N$.
We recall that, for arbitrary subsets $A$ of $X$ and $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$,
\begin{equation*}
R_w^A:=R_{1_Aw}=\inf\{v \in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon v\ge w \mbox{ on }A\} \quad\mbox{ and }\quad
\hat R_w^A:=\widehat{R_w^A},
\end{equation*}
leading to measures $\varepsilon_x^A$ and $\hat\varepsilon_x^A$ on $X$ which are characterized by
\begin{equation*}
\int w \,d\varepsilon_x^A=R_w^A (x) \quad\mbox{ and }\quad \int w \,d\hat \varepsilon_x^A=\hat R_w^A (x), \qquad w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P};
\end{equation*}
see \cite[II.4.3, II.5.4 and VI.2.1]{BH}
(in \cite{BH} these measures are denoted by~$\overset\circ \varepsilon{}_x^A$ and $\ve_x^A$).
For later use, let us observe that trivially
\begin{equation}\label{Rvv}
R_v=v, \quad\mbox{whenever $v=R_w^A$, $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$}.
\end{equation}
Given a stopping time $T$, we define
as usual\footnote{We tacitly assume
that we have an isolated point $\Delta$ added to $X$, that functions on $X$ are
identified with functions on $X_\Delta:=X\cup\{\Delta\}$ vanishing at $\Delta$ and that $X_t\colon
[0,\infty] \to X_\Delta$ with $X_\infty=\Delta$ and $X_t(\omega)=\Delta$, whenever $t\ge s$ and
$X_s(\omega)=\Delta$.}
\begin{equation*}
P_Tf(x):=E^x(f\circ X_T) \quad\mbox{ for all $f\in \mathcal B^+(X)$ and $x\in X$}.
\end{equation*}
Suppose for the moment that $A\in \mathcal B$. Then, by \cite[VI.3.14]{BH},
both $D_A$ and $T_A$ are stopping times
and,
writing $P_A$ instead of $P_{D_A}$ and $\hat P_A$ instead of $P_{T_A}$,
\begin{equation}\label{PPhat}
P_Aw =R_w^A \quad\mbox{ and }\quad
\hat P_A w =\hat R_w^A
\end{equation}
for every $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ (cf. \cite[6.12]{BG}),
where obviously
\begin{equation}\label{PDT}
\mbox{$P_Aw=w$ on $A$} \quad\mbox{ and }\quad \mbox{$P_Aw=\hat P_Aw$ on $X\setminus A$.}
\end{equation}
By (\ref{PPhat}) and \cite[Section II]{BH},
$P_Af, \hat P_Af \in \mathcal B(X) $ for all $f\in \mathcal B^+(X)$.
Hence $P_A$ and~$\hat P_A$ are kernels on $X$.
Moreover,
\begin{equation*}
P_A(x,B)=\varepsilon_x^A(B) \quad\mbox{ and }\quad \hat P_A(x,B)=\hat \varepsilon_x^A(B)
\end{equation*}
for all $x\in X$ and $B\in \mathcal B$. In particular, our definition of nearly hyperharmonic functions by (\ref{def-nearly})
coincides with the definition given by \cite[(2.2)]{HN-mertens}.
The following simple stability result will be useful.
\begin{lemma}\label{sums}
For every $\mu\in\mathcal M(X)$, the set $\mathcal F$ of all functions $f\in\mathcal B^+(X)$
such that $\int f\,d\mu=\inf\{\int w\,d\mu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P},\, w\ge f\}$
is a convex cone which is closed under countable sums.
\end{lemma}
\begin{proof} Of course, $0\in \mathcal F$ and $a f\in\mathcal F$ for all $a> 0$ and $f\in \mathcal F$.
Let $(f_n)$ be a~sequence in~$\mathcal F$ and $f:=\sum_{n\ge 1} f_n$ such that $\int f\,d\mu<\infty$.
Given $\varepsilon>0$, we may choose $w_n\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$,
$n\in\mathbbm{N}$, such that $\int w_n\,d\mu<\int f_n\,d\mu+ 2^{-n}\varepsilon$. Then $w:=\sum_{n\ge 1} w_n\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$,
$w\ge f$ and $\int w\,d\mu< \int f\,d\mu+\varepsilon$.
\end{proof}
We recall that an arbitrary set $A$ in $X$ is
called \emph{thin} at a point $x\in X$ if $\hat \varepsilon_x^A\ne \ve_x$.
By definition, the \emph{base} $b(A)$ of $A$ is the set of all $x\in X$ such that $A$ is not thin at~$x$, that is, $\hat
\varepsilon_x^A=\ve_x$. By \cite[VI.4.8]{BH},
\begin{equation}\label{base-prob}
b(A)=\{x\in X\colon T_A=0\mbox{ $P^x$-almost surely}\}, \quad \mbox{ if } A\in\mathcal B.
\end{equation}
By \cite[VI.4.1 and VI.4.4]{BH}, the base of every set $A$ in $X$ is a finely closed $G_\delta$-set
containing the fine interior of $A$, and $A\cup b(A)$ is the fine closure of $A$.
Moreover, for every $x\in X$, the measure $\hat\varepsilon_x^A$ is supported by the fine closure of $A$,
that is, the inner measure of its complement is zero; see \cite[VI.4.6]{BH}.
A set $F$ in $X$ is called \emph{totally thin} if $b(F)=\emptyset$ so that, in particular, $F$ is finely closed.
A \emph{semipolar set} is a countable union of totally thin sets.
We know that, for any infimum $u$ of functions in $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$, the set $\{\hat u<u\} $ is
semipolar; see \cite[VI.5.11] {BH}.
\begin{example}\label{hyperplanes}{\rm
For space-time Brownian motion on ${\real^d}\times \mathbbm{R}$, every hyperplane $H_t:=\mathbbm{R}^d\times \{t\}$ is totally thin.
}\end{example}
For the remainder of this section let us fix a function $u\in \mathcal B^+(X)$ which is nearly hyperharmonic.
By \cite[Proposition 2.5]{HN-mertens}, for every $A\in \mathcal B$,
\begin{equation}\label{nearly-used}
P_Au\le u \quad\mbox{ and }\quad \hat P_A u\le u.
\end{equation}
\begin{remark} {\rm
Let $S,T$ be stopping times for $\mathfrak X$, $S\le T$. Then
$P_Tw\le P_Sw\le w$ for every $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$; see \cite[VI.3.4]{BH}.
By \cite[Corollary 2.6]{HN-mertens} (which uses that, for $x\in X$,
the extreme points in the weak$^\ast$-compact convex set $\mathcal M_x(\mathcal E_{\mathfrak X}} % {E_{\mathbbm P})$
of all measures~$\mu$ on~$X$ satisfying \hbox{$\int w\,d\mu\le w(x)$} for every $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$
are the measures $\ve_x^A$, $A\in \mathcal B$), this implies that $P_Su\le u$. So the nearly hyperharmonic
function~$u$ is strongly
supermedian in the sense of
\hbox{\cite{beznea-boboc-feyel, beznea-boboc-book, feyel-rep, feyel-fine, moko-ens-compacts}}.
By~\cite[Proposition 2.7]{HN-mertens},
we even get the inequality $P_Tu\le P_Su$ (and hence, by a standard argument,
$ E^x(u\circ X_T|\mathfrak M_S)\le u\circ X_S$ $P^x$-a.s.\ for every $x\in X$).
Since we shall not use these facts in the sequel, they may also be viewed
as \emph{consequences} of Theorem \ref{main}
(to obtain $P_T(u\wedge n) \le P_S(u\wedge n) $, $n\in\mathbbm{N}$,
we consider the finite measures $\mu:=P_T(x,\cdot)+P_S(x,\cdot)$, $x\in X$).
}
\end{remark}
Let us note that (\ref{nearly-used}) implies the following.
\begin{lemma}\label{fusc}
The function $u$ is finely upper semicontinuous and, starting in the finely open set $U:=\{u<\infty\}$,
the process $\mathfrak X$ does not leave $U$, that is,
\begin{equation}\label{U-stable}
P^x[T_{X\setminus U}<\infty]=0 \qquad\mbox{ for every $x\in U$.}
\end{equation}
\end{lemma}
\begin{proof}
Let $a\in [0,\infty]$, $A:=\{u\ge a\}$ and $x\in X\setminus A$.
By \cite[VI.3.14]{BH},
there exists an increasing sequence $(K_n)$ of compacts in the Borel set $A$
such that $T_{K_n}\downarrow T_A$~$P^x$-a.s. Since $X_{T_{K_n}}\in K_n$ on
$[T_{K_n}<\infty]$, the inequalities $P_{K_n}u\le u$ yield that
\begin{equation*}
aP^x[T_A<\infty] =\lim_{n \to \infty} aP^x[T_{K_n}<\infty]\le u(x)<a.
\end{equation*}
Hence $P^x[T_A<\infty]<1$, $x\notin b(A)$. So $A$ is finely closed
showing that $u$ is finely upper semicontinuous.
Finally, taking $a=\infty$, we see that (\ref{U-stable}) holds.
\end{proof}
For every $V\in {\mathcal U}_c$, due to the lower semicontinuity
of $P_{V^c} u$ on $V$, we know that
\begin{equation}\label{nearly-used-l}
\hat P_{V^c} u=P_{V^c} u\le \hat u \quad\text{ on } V
\end{equation}
(see \cite[(2.3)]{HN-mertens}).
The following more general estimate will be crucial in Section~\ref{sec-special}.
\begin{lemma}\label{PFuh}
Let $A\in\mathcal B$ and $x\in X\setminus b(A)$ such that $x$ is not finely isolated. Then
\begin{equation*}
\hat P_Au (x) =E^x(u\circ X_{T_A}) \le \hat u (x).
\end{equation*}
\end{lemma}
\begin{proof}
Since $u_n:=u\wedge n$ is nearly hyperharmonic for every $n\in\mathbbm{N}$ and
$\hat u_n\uparrow \hat u$ by \cite[Proposition 2.3]{HN-mertens},
we may assume without loss of generality that $u$ is bounded, say
$u< M<\infty$.
Let $W_n\in {\mathcal U}_c$ such that $W_n\downarrow \{x\}$ as $n\to\infty$.
By assumption, $x\in b(X\setminus \{x\})$,
and hence $D_{W_n^c}\downarrow
D_{X\setminus \{x\}}=0$ $P^x$-a.s., whereas $T_A>0$ $P^x$-a.s.
So there exists $n\in\mathbbm{N}$ such that $\tau:=D_{W_n^c}$ satisfies
$ P^x[T_A\le \tau ]<\varepsilon/M$, and therefore
\begin{equation*}
E^x(u\circ X_{T_A}) \le \varepsilon + E^x(u\circ X_{T_A}; T_A>\tau).
\end{equation*}
Let us note that $\tau>0$ $P^x$-a.s. and that
obviously, on the set $[T_A>\tau>0]$ we have $ T_A=\tau + D_A\circ \theta_\tau$
whence $X_{T_A}=X_{D_A}\circ \theta_\tau$. Thus
we conclude that
\begin{eqnarray*}
E^x(u\circ X_{T_A}; T_A>\tau)&\le& E^x(u\circ X_{D_A}\circ
\theta_\tau) =E^x(E^{X_\tau}(u\circ X_{D_A}))\\[2pt]
&=&P_\tau P_Au(x)\le P_\tau u(x) \le \hat u(x),
\end{eqnarray*}
by the strong Markov property and (\ref{nearly-used-l}).
\end{proof}
Finally, let us recursively define stopping times $S_n^A$, $A\in \mathcal B$, $n\ge 0$, by
\begin{equation}\label{def-Sn}
S_0^A:=D_A \quad\mbox{ and }\quad S_{n+1}^A:=S_n^A+T_A\circ \theta_{S_n^A}.
\end{equation}
So $S_{n+1}^A$ is the time of the first hitting of $A$ after the time $S_n$,
$X_{S_{n+1}^A}=X_{T_A}\circ \theta_{S_n^A}$; see \cite[Section IV.6]{BH}.
\begin{proposition} \label{Proposition-PAn}
Let $A\in \mathcal B$ and $n\ge 0$. Then
\begin{equation} \label{PAn}
P_n^Af(x):=E^x(f\circ X_{S_n^A}) = P_A(\hat P_A)^n f(x) , \qquad f\in\mathcal B^+(X), \ x\in X.
\end{equation}
\end{proposition}
\begin{proof}
Obviously, (\ref{PAn}) holds for $n=0$.
Suppose that it is true for some $n\ge 0$,
and let $f\in \mathcal B^+(X)$, $x\in X$, $S_n:=S_n^A$.
By the strong Markov property,
\begin{eqnarray*}
E^x( f\circ X_{S_{n+1}})= E^x( f\circ X_{T_A}\circ \theta_{S_n}) & =&
E^x (E^{X_{S_n}}(f\circ T_A))\\ &=&E^x(( \hat P_A f)\circ X_{S_n})
= P_A(\hat P_A)^{n+1} f(x).
\end{eqnarray*}
\end{proof}
Given $A\in \mathcal B $, let
\begin{equation*}
P^A:=\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} P_n^A
\end{equation*}
so that, by Proposition \ref{Proposition-PAn},
\begin{equation*}
P^A1(x)= \sum_{n\ge 0}} % {\sum_{n=0}^{\infty} P^x[S_n^A<\infty]=\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} P_A(\hat P_A)^n1(x)
\end{equation*}
is the expected number of times
the process~$\mathfrak X$ visits $A$ when starting in~$x$.
Clearly, $ P^A1=P_Aw$ with $w:=\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} (\hat P_A)^n1\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ and $\hat P_A w+1=w$.
Thus, by~(\ref{Rvv}), (\ref{PPhat}) and (\ref{PDT}), we obtain the following
(where we may note that $\hat v$ is the expected number of visits
of $A$ at \emph{strictly positive} times).
\begin{proposition}\label{Rvav}
For every $A\in \mathcal B$, $v:=P^A1$ satisfies $ R_v=v$ and $\hat v+1_A=v$.
\end{proposition}
\section{A special case}\label{sec-special}
Throughout this section we fix a
nearly hyperharmonic function $u\in\mathcal B^+(X)$ and suppose the following.
\begin{assumption}\label{ass}
Defining $U:=\{u<\infty\}$,
and $F:=\{\hat u<u\} $, we have
\begin{equation}\label{ass-formula}
P^F 1<\infty \quad\text{ on } U,
\end{equation}
that is, starting in $U$, the expected number of times the process $\mathfrak X$ visits~$F$ is finite.
\end{assumption}
If $x\in b(F)$, then, by (\ref{base-prob}), $S_n=0$ $P^x$-a.s.\ for every $n\ge 0$, $P^F1(x)=\infty$.
Hence (\ref{ass-formula}) implies that $b(F)\cap U=\emptyset$. So $F$ is totally thin
if $u<\infty$.
By Lemma \ref{fusc} and a~straightforward induction, we see that, for every $x\in U$ and
$P^x$-a.e.\ $\omega\in [S_n<\infty]$,
\begin{equation}\label{FcapU}
X_{S_n}(\omega)\in F\cap U \quad and \quad S_n(\omega)<S_{n+1}(\omega),
\end{equation}
and, for $P^x$-a.e.\ $\omega\in\Omega$,
\begin{equation}\label{sge0}
\{S_n(\omega) \colon n\ge 0, \, S_n (\omega)<\infty\}=\{s\ge 0\colon X_s(\omega)\in F\}.
\end{equation}
\begin{example}{\rm
Let us consider space-time Brownian motion
on ${\real^d}\times \mathbbm{R}$ and fix a~sequence $(t_n)$ in $\mathbbm{R}$.
For $n\in\mathbbm{N}$, let $A_n$ be an arbitrary subset of~$H_{t_n}={\real^d}\times \{t_n\}$ and let
\begin{equation*}
v_n:=1_{A_n}+1_{{\real^d}\times (t_n,\infty)}.
\end{equation*}
Then $v:=\sum_{n\ge 1} 2^{-n}v_n\in\N$, $0\le v\le 1$, and $\{\hat v<v\}$ is the union of all $A_n$, $n\in\mathbbm{N}$.
If $\{t_n\colon n\in\mathbbm{N}\} $ is dense in~$\mathbbm{R}$
and $A_n=H_{t_n}$, $n\in\mathbbm{N}$, then
$\{\hat v <v\}$ is finely dense in~${\real^d}\times \mathbbm{R}$,
$P^{\{\hat v<v\}}1=\infty$.
So Assumption \ref{ass} is very restrictive.
If $t_n=-1/n$ and $A_n=H_{t_n}$, $n\in\mathbbm{N}$, then $v:=\infty\cdot 1_{{\real^d}\times [0,\infty] }
+\sum_{n\ge 1} 2^{-n}v_n \in \N$ and $v$ satisfies Assumption \ref{ass} with $b(\{\hat v<v\})=H_0$.
}
\end{example}
We define functions $g$ and $u_0$ on $X$ by
\begin{equation*}
g(x):=
P^F(1_Uu-1_U\hat u)\quad\mbox{ and }\quad u_0(x):=u(x)-g(x), \qquad x\in U,
\end{equation*}
and $g(x)=u_0(x)=\infty$, if $ x\in X\setminus U$.
\begin{definition}
For $A\in \mathcal B$, let $\R(A)$ be the set
of sums of a function in $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$
and countably many functions~$P_Bw$,
where $B\in \mathcal B$, $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$, $B\subset A$ and $w$ is~bounded.
\end{definition}
Clearly, $\R(A)\subset \R(X)\subset \N\cap\mathcal B^+(X)$ and,
by (\ref{Rvv}) and Lemma~\ref{sums},
\begin{equation}\label{NR}
v=R_v=\inf\{w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge v\} \quad\mbox{ for every }v\in \R(X).
\end{equation}
In this section we shall establish the following result.
\begin{theorem}\label{main-special}
The function $g$ is a minorant of $u$, both $g$ and $u_0$ are nearly hyperharmonic, $g-\hat g=u-\hat u $ on $U$
and $\hat u_0=u_0$ on $U$.
Further, there are functions $g_1, u_1\in \R(F)$
such that $g_1=g$ on~$U$ and $u_1=u$ on~$U$. In particular, if $u<\infty$, then $u_0\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ and $u=R_u$.
\end{theorem}
We prepare its proof with a lemma leading to estimates by telescoping series.
\begin{lemma}\label{g-tau}
Let $V\subset X$ be open,
$x\in V\cap U$, $\tau:=D_{V^c}$ and $\Omega_n:=\{S_n<\tau\}$.
Then, for all $n\ge 0$,
\begin{equation}\label{tel-formula}
E^x(\hat u\circ X_{S_n} ; \Omega_n)\ge E^x( u\circ X_{S_{n+1}} ; \Omega_{n+1})
+E^x(u\circ X_\tau;\Omega_n\setminus \Omega_{n+1}).
\end{equation}
\end{lemma}
\begin{proof} Let $T:= T_{F\cup {V^c}} $.
By Lemma~\ref{PFuh},
$E^y(u\circ X_T)\le \hat u(y)$ for every $y\in F\cap U$,
and hence, by (\ref{FcapU}) and the strong Markov property,
\begin{equation*}
E^x(\hat u\circ X_{S_n}; \Omega_n) \ge E^x(E^{X_{S_n}} (u\circ X_T); \Omega_n)
=E^x(u\circ X_T\circ \theta_{S_n}; \Omega_n)
\end{equation*}
for every $n\ge 0$.
The proof is completed observing
that $S_n+T\circ \theta_{S_n}=S_{n+1} $ on $\Omega_{n+1}$
and $S_n+T\circ \theta_{S_n}= \tau$ on~$\Omega_n\setminus \Omega_{n+1}$.
\end{proof}
\begin{proposition}\label{g-function}
The function $g$ is a minorant of $u$.
\end{proposition}
\begin{proof}
Let $x\in U$. By Lemma \ref{g-tau} with $V:=X$, $\tau=\infty$,
\begin{equation*}
g(x)\le \sum_{n\ge 0}} % {\sum_{n=0}^{\infty}\bigl( E^x(u\circ X_{S_n})-E^x(u\circ X_{S_{n+1}})\bigr) \le E^x(u\circ X_{S_0})= P_F u(x),
\end{equation*}
where $P_F u(x)\le u(x)$, by (\ref{nearly-used}).
\end{proof}
Finally, we need the following.
\begin{lemma}\label{tech}
Let $E$ be a Borel subset of $F$. Then $v:= P^E1=P^F1_E$ and
\begin{equation}\label{markov}
P_Bv(x) = E^x(\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} 1_{E}\circ X_{S_n}\, 1_{[S_n\ge D_B]})\quad\mbox{for all $B\in \mathcal B$ and $x\in X$. }
\end{equation}
\end{lemma}
\begin{proof}
Let $T_n:=S_n^E$, $n\ge 0$. By Proposition \ref{PAn} and the strong Markov property,
\begin{eqnarray*}
&&P_Bv(x)=E^x(E^{X_{D_B}}(\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} 1_{E}\circ X_{T_n}))\\
& =&E^x(\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} 1_{E}\circ X_{T_n}\circ \theta_{D_B})
=E^x(\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} 1_{E}\circ X_{D_B+{T_n\circ \theta_{D_B}}}),
\end{eqnarray*}
where, for $P^x$-almost every $\omega\in \Omega$, the last sum
is the number of all $s\ge D_B(\omega)$ such that $X_s(\omega)\in E$, which
in turn is the sum in (\ref{markov}); see (\ref{sge0}).
The equality $v=P^F1_E$ follows taking $B=X$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{main-special}]
There are Borel sets $F_k$ in $F $ and $a_k\in (0,\infty)$, $k\in\mathbbm{N}$, with
\begin{equation}\label{choice-Fk}
1_Uu-1_U\hat u=\sum_{k\ge 1} a_k 1_{F_k}.
\end{equation}
(Indeed, choosing step functions $f_n=\sum_{j=2}^{m_n} a_{nj} 1_{F_{nj}}\in \mathcal B(X)$
with $a_{nj}>0$ which are increasing to $f:=1_Uu-1_U\hat u$, it suffices to observe that
$f$ is the sum of $f_1$ and the step functions $f_{n+1}-f_n$, $n\in\mathbbm{N}$.)
By Lemma \ref{tech}, $v_k:=P^{F_k}1= P^E1_{F_k}$ for every $k\in\mathbbm{N}$.
Hence, by Proposition \ref{Rvav},
\begin{equation}\label{gg1}
g_1:=\sum_{k\ge 1} a_k v_k \in \R(F) \quad\mbox{ and }\quad g_1=g \quad\text{ on } U.
\end{equation}
So $g\in \N\cap \mathcal B(X)$;
see (\ref{FcapU}).
Further, using \cite[(2.4) and Proposition~2.3]{HN-mertens},
\begin{equation*}
g=g_1=\sum_{k\ge 1} a_k(1_{F_k}+ \hat v_k^f) =\sum_{k\ge 1} a_k1_{F_k}+\hat g_1^f=(u-\hat u)+\hat g \quad\text{ on } U.
\end{equation*}
Next let $V\in {\mathcal U}_c$, $x\in V\cap U$ and $\tau:=D_{V^c}$.
By Lemma \ref{tech},
\begin{equation*}
g(x)-P_\tau g(x)=g_1(x)-P_\tau g_1(x)= \sum_{n\ge 0}} % {\sum_{n=0}^{\infty} E^x((u-\hat u)\circ X_{S_n}; S_n<\tau)\ge 0
\end{equation*}
(showing once more that $g$ is nearly hyperharmonic) and,
using Lemma~\ref{g-tau},
\begin{equation*}
g(x)-P_\tau g(x)\le E^x(u\circ X_{S_0}; S_0<\tau)-\sum_{n\ge 0}} % {\sum_{n=0}^{\infty} E^x(u\circ X_\tau; S_n<\tau\le S_{n+1}),
\end{equation*}
where the sum is equal to $P_\tau u(x)-E^x(u\circ X_\tau;\tau\le S_0)$. Therefore
\begin{equation*}
u_0(x)-P_\tau u_0(x)=u(x)-P_\tau u(x) -(g(x)-P_\tau g(x))
\ge u(x)- E^x(u\circ X_{S_0\wedge \tau}).
\end{equation*}
Since $S_0\wedge \tau=D_{F\cup {V^c}}$ and $P_{F\cup{V^c}}u\le u$,
by (\ref{nearly-used}), we see that
$u_0(x)-P_\tau u_0(x)\ge 0$, that is, $u_0\in \N$.
Since $\hat g+\hat u_0=\widehat{g+u_0}=\hat u$,
by \cite[Proposition 2.3]{HN-mertens}, we finally conclude that
$\hat u_0=u_0$ on $U$
and $u_1:=g_1+\hat u_0\in \R(F)$, $u_1=g+u_0=u$ on~$U$.
\end{proof}
\begin{corollary}\label{corollary-special}
If $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ such that $w=\infty$ on $\{u=\infty\}$, then $u+w\in \R(F)$.
\end{corollary}
\begin{proof}
By Theorem \ref{main-special}, $u+w=u_1+g_1+w\in \R(F)$.
\end{proof}
\section{The general case}\label{sec-general}
We first reduce the general case of a nearly hyperharmonic function $u\in \mathcal B^+(X)$
to the special one considered in the previous section.
\begin{proposition}\label{key-reduction}
Let $u\in\N\cap \mathcal B(X)$ with $\inf u(X) >0$.
Further, let $\eta\in (0,1)$,
\begin{equation*}
F:=\{\hat u< \eta u\} \quad\mbox{ and }\quad v:=1_Fu+1_{X\setminus F} \hat u.
\end{equation*}
Then $\eta u\le v\le u$, $v\in \N\cap \mathcal B(X)$, $F=\{\hat v<v\}$
and {\bfseries $P^F1< \infty$ on $\{\hat v<\infty\}$. }
\end{proposition}
\begin{proof} Of course, $\hat u\le v\le u$, hence $\hat v=\hat u$
and $v\in \N$, by~\cite[Proposition 2.2]{HN-mertens}.
Clearly, $F=\{\hat v<v\}$
and $a:=\inf \hat v(X)=\inf u(X)>0$.
The set $ A:=\{\hat u\le \eta u\}$ containing $F$
is finely closed, since $u$ is finely upper semicontinuous, by Lemma \ref{fusc}. So, for~every $x\in X$,
the measure~$P_F(x,\cdot)$ is supported by $A$, and hence
\begin{equation*}
P_F\hat u(x) \le \eta P_F u(x)\le \eta u(x).
\end{equation*}
By regularization,
$ \hat P_F\hat u\le \eta\hat u$ (see (\ref{PPhat})), that is, $\hat P_F\hat v\le \eta \hat v$.
By induction, for every $n\ge 1$, $(\hat P_F)^n\hat v\le \eta^n \hat v$, and therefore
\begin{equation*}
a P_F(\hat P_F)^n1\le P_F(\hat P_F)^n\hat v\le \eta^n P_F\hat v\le \eta^n\hat v.
\end{equation*}
Thus $P^F1<\infty$ on $\{\hat v<\infty\}$.
\end{proof}
Let us say that a function $u\in \N$ has the \emph{finiteness property} (FP) if, for
every $x\in X$ with $u(x)<\infty$, there exists a~function $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ such that $w=\infty$ on
$\{u=\infty\}$ and $w(x)<\infty$.
Trivially, every $u\in\N$ with $u<\infty$ has this property (take $w=0$).
\begin{theorem}\label{general-1}
Let $u\in \N$ be Borel measurable satisfying {\rm (FP)}. Then $u=R_u$.
\end{theorem}
\begin{proof}
Let $x\in X$, $\eta\in (0,1)$ and $\varepsilon>0$. Of course, $u+\varepsilon\in\N$ and $u+\varepsilon$ satisfies~(FP).
By Proposition \ref{key-reduction}, there exists $v\in \N\cap \mathcal B(X)$ satisfying Assumption \ref{ass}
and such that
$\eta(u+\varepsilon)\le v\le u+\varepsilon$.
We choose $w_1\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ such that $w_1=\infty$ on~$\{u=\infty\}$ and $w_1(x)<\varepsilon$.
Then $v+w_1\in \R(\{\hat v<v\})$, by Corollary \ref{corollary-special}. Thus, by (\ref{NR}),
$\eta R_u\le R_{v+w_1}=v+w_1\le u+\varepsilon+w_1$. In particular, $\eta R_u(x)\le u(x)+2\varepsilon$.
\end{proof}
The following consequence of Theorem \ref{general-1} may be surprising.
Its combination with Theorem \ref{general-1}
establishes the implication (1)\,$\Rightarrow$\,(2) in Theorem \ref{main}.
\begin{corollary}\label{u-infty-infty}
Every Borel measurable $u\in\N$ has the property {\rm (FP)}.
\end{corollary}
\begin{proof}
Let $u\in \N$ be Borel measurable,
$x\in X$ with $u(x)<\infty$ and $E:=\{u=\infty\}$.
Clearly, $1_E=1\wedge \inf_{n\in\mathbbm{N}} (u/n)\in \N$.
By Theorem~\ref{general-1}, there are $w_n\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$, $n\in\mathbbm{N}$, such that
\begin{equation*}
w_n\ge 1_E \quad\mbox{ and }\quad w_n(x)<2^{-n}.
\end{equation*}
Then $w:=\sum_{n\ge 1} w_n\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$, $w=\infty$ on $E$ and $w(x)<1$.
\end{proof}
To prove the additional statement in Theorem \ref{main}, let us consider $\varphi\in \mathcal B^+(X)$
and recall that $N_\varphi:=\inf\{u\in \N\colon u\ge \varphi\}$ is the smallest nearly hyperharmonic
majorant of $\varphi$, $N_\varphi=\varphi\vee \hat N_\varphi\in \mathcal B^+(X)$ (see \cite[Proposition 2.4]{HN-mertens}),
and hence
\begin{equation*}
N_\varphi=\inf\{w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge N_\varphi\} \ge \inf \{w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge \varphi\}=R_\varphi\ge N_\varphi.
\end{equation*}
\section{The remaining part of Theorem \ref{main}}\label{remaining}
For a proof of the implication (1)\,$\Rightarrow$\,(3) in Theorem \ref{main} we note that,
for every $u\in\N\cap \mathcal B(X)$, the set $\{\hat u<u\}$ is semipolar
(which, for example, follows from $u=\inf\{w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge u\}$)
and that every semipolar Borel set is the union of compacts $K_n$, $n\in\mathbbm{N}$,
and a~polar set $E\in \mathcal B$; see~\cite[Proposition 5.2]{HN-mertens} in connection with Remark \ref{erratum} below.
We start with a lemma on~compact sets which will quickly lead to the basic approximation result in Corollary \ref{NRmu}.
\begin{lemma}\label{mu-K-proposition}
Let $K$ be a compact in $X$ and let $w$ be a bounded function in $\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$.
Then there exists a decreasing sequence $(V_n)$ of finely open Borel sets containing~$K$ such that
$P_Kw=\inf_{n\in\mathbbm{N}} P_{V_n}w$.
\end{lemma}
\begin{proof}
By \cite[VI.1.2 (and its proof)]{BH}, $P_Kw$ is the infimum of all functions
$P_Vw$, where~$V$ is a finely open Borel set containing $K$. By a topological lemma of Choquet
(see \cite[I.1.8]{BH}), there is
a sequence $(V_n)$ of such sets satisfying
\begin{equation*}
\hat P_Kw= \widehat{\inf_{n\in\mathbbm{N}} P_{V_n}w}.
\end{equation*}
Fixing a decreasing sequence $(U_n)$ of open sets in $X$ with $\bigcap_{n\in\mathbbm{N}} \overline U_n=K$,
we may assume without loss of generality that $V_{n+1}\subset V_n\subset U_n$ for every $n\in\mathbbm{N}$.
Then, by \cite[VI.2.6]{BH}, every function $P_{V_n}w$, $n\in\mathbbm{N}$, is harmonic on~$X\setminus \overline U_n$.
Hence $\inf_{n\in\mathbbm{N}} P_{V_n}w$ is harmonic on $X\setminus K$, by \cite[III.3.1]{BH}, and we obtain that
\begin{equation*}
P_Kw=\hat P_Kw = \inf P_{V_n}w \quad\text{ on } X\setminus K.
\end{equation*}
The proof is completed observing that $P_Kw=w=\inf P_{V_n}w$ on $K$.
\end{proof}
\begin{proposition}\label{mu-proposition}
Let $A$ be the union of compacts $K_n$ in $X$, $n\in\mathbbm{N}$, and a~polar set $E\in \mathcal B$.
Further, let $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ and $\mu\in \mathcal M(X)$ be such that $w $ is bounded and $\mu(X)<\infty$.
Then
\begin{equation*}
\int P_Aw\, d\mu= \inf\{\int P_Vw\,d\mu\colon A\subset V,\ V \mbox{ finely open Borel\,}\} .
\end{equation*}
\end{proposition}
\begin{proof} Of course, we may suppose that the sequence $(K_n)$ is increasing.
Moreover, we may assume that $\mu(A)=0$, since $P_Aw=w=P_Vw$, whenever $A\subset V\subset X$.
Let us fix $\varepsilon>0$.
Since $T_E=\infty$ a.s., we have $P_E1=0$ on $X\setminus E$. Hence, by \cite[VI.1.9]{BH},
there exists an open neighborhood $U$ of $E$ such that $\int P_U1\,d\mu<\varepsilon$.
Moreover, by~Lemma~\ref{mu-K-proposition}, there exist finely open $V_n\in\mathcal B$, $n\in\mathbbm{N}$,
such that
\begin{equation*}
K_n\subset V_n \quad\mbox{ and }\quad \int P_{V_n}w\,d\mu<\int P_{K_n}w\,d\mu+2^{-n} \varepsilon.
\end{equation*}
Defining $W_n:=V_1\cup \dots\cup V_n$ and proceeding as in the proof of \cite[VI.1.4]{BH}
we get
\begin{equation*}
\int P_{W_n}w\,d\mu \le \int P_Kw\,d\mu+(1-2^{-n})\varepsilon \qquad\mbox{ for every }n\in\mathbbm{N}.
\end{equation*}
Let $W:=\bigcup_{n\ge 1}W_n$ and $V:=W\cup U$. Then $P_Vw\le P_Ww+P_Uw$ and $P_{W_n}w\uparrow P_Ww$,
by \cite[VI.1.7]{BH}. Thus we finally conclude that $\int P_Vw\,d\mu\le \int P_Kw\,d\mu+2\varepsilon$.
\end{proof}
\begin{remark}{\rm
In \cite[p.\,138]{mertens} such a result is shown for sets $A$ which are \emph{strictly thin}, that is,
satisfy $\hat P_A1< \eta$ on~$A $ for some $\eta\in (0,1)$, giving first a (delicate) proof for
the following stunning approximation of the hitting time $T_A$: For every probability measure $\nu$
on $X$ not charging $A$, there exists a decreasing sequence $(V_n)$ of finely open sets containing $A$
such that $\lim_{n \to \infty} P^\nu[T_{V_n}<T_A]=0$.
}
\end{remark}
\begin{corollary}\label{NRmu}
Let $u\in \N\cap \mathcal B(X)$, $v\in\R(\{\hat u<u\})$ and let $\mu\in\mathcal M(X)$ be a~finite measure. Then
\begin{equation}\label{NRmu-formula}
\int v\,d\mu:=\inf\{\int w\,d\mu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P},\,w\ge v\}.
\end{equation}
\end{corollary}
\begin{proof} By (\ref{red-es}), $P_Vw\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ for all finely open sets $V\in \mathcal B$ and $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$.
Thus (\ref{NRmu-formula}) follows immediately from Proposition \ref{mu-proposition} and Lemma \ref{sums}.
\end{proof}
\begin{definition}
For every Borel measurable $u\in\N$, let $\mathcal M_u(X)$ denote the set of all $\mu\in \mathcal M(X)$
such that $u$ is $\mu$-integrable and $\mu(A) +\int_{X\setminus A} w\,d\mu<\infty$
for some Borel set $A$ in $X$ and majorant $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ of $u$.
\end{definition}
Let us say that a Borel measurable $u\in \N$ has the \emph{finiteness property} (FP$'$) if, for
every $\mu\in\mathcal M_u$, there exists a~function $w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ with $w=\infty$ on the
set $\{u=\infty\}$ and~$\int w\,d\mu<\infty$.
Trivially, every $u\in\N$ with $u<\infty$ has this property (take $w=0$).
\begin{theorem}\label{Borel-mu}
Let $u\in \N $ be Borel measurable and $\mu\in\mathcal M_u(X)$.
If $u<\infty$ or, more generally, if $u$ has the property~{\rm (FP$'$)}, then
\begin{equation*}
\int u\,d\mu=\inf \{\int w\,d\mu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P},\,w\ge u\} .
\end{equation*}
\end{theorem}
\begin{proof} Let $\eta\in (0,1)$ and $\varepsilon>0$.
Assuming that $u$ has the property (FP$'$), we choose $w_1\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$
such that $w_1=\infty$ on $\{u=\infty\}$ and $\int w_1\,d\mu<\varepsilon$.
By Remark~\ref{KK}, there exists $A\in \mathcal B$ and $w_0\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ such that
$w_0\ge u$ and $\mu(A)<\infty$, $\int_{X\setminus A} w_0\,d\mu<\varepsilon$.
We fix $\delta>0$ such that $\delta \mu(A) <\varepsilon$, and define
\begin{equation*}
\nu:=1_A \mu \quad\mbox{ and }\quad u_1:=u+\delta.
\end{equation*}
By Proposition \ref{key-reduction} and Corollary \ref{corollary-special},
there exists a~Borel measurable $v\in \N$ such that
$\eta u_1\le v\le u_1$ and $v+w_1\in\R(\{\hat v<v\})$. Then, by Corollary \ref{NRmu},
\begin{eqnarray*}
& & \int u\,d\mu +2\varepsilon >\int (u_1+w_1)\,d\nu \ge \int (v+w_1)\,d\nu\\
&=&\inf\{\int w\,d\nu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P},\ w\ge v+w_1\}
\ge \eta \inf\{\int w\,d\nu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P},\ w\ge u\}.
\end{eqnarray*}
So there exists $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ such that $w\ge u$ and $\eta \int w\,d\nu<\int u\,d\mu+2\varepsilon$.
We may assume without loss of generality that $w\le w_0$. Then
\begin{equation*}
\eta \int w\,d\mu< \eta \int w\,d\nu+\varepsilon<\int u\,d\mu+3\varepsilon.
\end{equation*}
Letting $\varepsilon$ tend to $0$ and $\eta$ tend to $1$ the proof is completed.
\end{proof}
\begin{corollary}\label{u-infty}
Every Borel measurable $u\in\N$ has the property {\rm (FP$'$)}.
\end{corollary}
\begin{proof}
Let $u\in \N$ be Borel measurable, $\mu\in\mathcal M_u(X)$ and $E:=\{u=\infty\}$.
In particular, $\int u\,d\mu<\infty$, and hence $\mu(E)=0$.
Obviously, $1_E=1\wedge \inf_{n\in\mathbbm{N}} (u/n)\in \N$.
Hence, by Theorem~\ref{Borel-mu}, there exist functions
$w_n\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$, $n\in\mathbbm{N}$, such that
\begin{equation*}
w_n\ge 1_E \quad\mbox{ and }\quad \int w_n\, d\mu<\int 1_E\,d\mu+ 2^{-n}=2^{-n}.
\end{equation*}
Then~$w:=\sum_{n\ge 1} w_n\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$, $w=\infty$ on $E$ and $\int w \,d\mu<1$.
\end{proof}
Combining Corollary \ref{u-infty} with Theorem \ref{Borel-mu} we obtain
the implication (1)\,$\Rightarrow$\,(3) in Theorem \ref{main}.
\section{Weakening of the measurability assumption}\label{weaker}
In this section we shall consider nearly hyperharmonic functions which may not be Borel measurable.
Let $\B^\ast$ be the $\sigma$-algebra of all ($\mathcal B$-)universally measurable sets and,
as in \cite{HN-mertens}, let~$\tilde \mathcal B,\widetilde{\B^\ast}$ respectively denote the $\sigma$-algebra of all
sets $A$ in $X$ for which there exists a~set $B$ in $\mathcal B,\B^\ast$ respectively such that the
symmetric difference $A \triangle B$ is polar, that is, $\hat P_{A \triangle B} 1=0$.
Let us observe that, for functions $u\ge 0$ which are $\widetilde{\B^\ast}$-measurable, the upper
integral in (\ref{def-nearly}) may be replaced by the integral, since the measures
$P_{V^c}(x,\cdot)=\ve_x^{V^c}$, $x\in V\in {\mathcal U}_c$, do not charge polar sets; see \cite[VI.5.6]{BH}.
Let $f$ be a positive function on $X$ which is $\tilde \mathcal B$-measurable.
Since every polar set is contained in a~Borel polar set and every countable union of polar sets is polar,
there exist $f_1\in \mathcal B^+(X)$ and $f_2\in \mathcal B^+(X)$ such that the set $\{f_2>0\}$
is polar and $f_1\le f\le f_1+f_2$.
The following result extends the implication
(1)\,$\Rightarrow$\,(3) of Theorem~\ref{main} to functions which are nearly Borel
measurable, that is, $\tilde \mathcal B\cap \mathcal B^\ast$-measurable.
\begin{theorem}\label{nearly-Borel}
Let $u$ be a nearly Borel measurable function in $\N$. Then
\begin{equation*}
\int u\,d\mu=\inf \{\int w\,d\mu\colon w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P},\ w\ge u\}
\end{equation*}
for every $\mu\in \mathcal M(X)$ such that $\mu(A)+ \int_{X\setminus A} w\,d\mu<\infty$ for some $A\in \mathcal B$
and some majorant $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ of $u$.
\end{theorem}
\begin{proof} There exist $u_1,u_2\in \mathcal B^+(X)$ such that $ u_1\le u\le u_1+u_2$ and the set $\{u_2>0\}$ is polar.
Of course, we may assume that $\hat u\le u_1$.
Let us fix $\mu\in \mathcal M(X)$
such that $\mu(A)+\int_{X\setminus A} w\,d\mu<\infty$ for some $A\in \mathcal B$
and majorant $w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}$ of $u$.
Choosing $v_1, v_2\in \mathcal B^+(X)$ such that
$v_1\le u\le v_1+v_2$ and $v_2=0$ $\mu$-a.e., we may
assume that $u_1\le v_1$ and $v_2\le u_2$.
Then $v_1\in \N$, by \cite[Proposition 2.2]{HN-mertens}.
Since $\{v_2>0\}$ is polar, we know that $v_2\in \N$, and hence $v:=v_1+v_2\in \N$. Thus, by Theorem \ref{main},
\begin{eqnarray*}
\int u\,d\mu=\int v \,d\mu&=&\inf\{\int w\,d\mu\colon w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}, w\ge v\}\\
&\ge& \inf\{\int w\,d\mu\colon w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}, w\ge u\}\ge \int u\,d\mu.
\end{eqnarray*}
\end{proof}
For the implication (1)\,$\Rightarrow$\,(2) we even have the following.
\begin{theorem}\label{tilde-Borel}
Let $u$ be a $\tilde \mathcal B$-measurable function in $\N$. Then
\begin{equation*}
u=\inf \{w\in\mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge u\}.
\end{equation*}
\end{theorem}
\begin{proof}
Let us fix $x\in X$.
There exist $v_1,v_2\in \mathcal B^+(X)$ such that $v_1\le u\le v_1+v_2$ and
the set $\{v_2>0\}$ is polar whence $v_2\in\N$. Of course, we may assume that $\hat u\le v_1$,
$v_1(x)=u(x)$ and $v_2(x)=0$. By \cite[Proposition 2.2]{HN-mertens}, $v_1\in \N$.
So $v:=v_1+v_2\in \N$ and $R_v=v$, by Theorem~\ref{main}.
Thus $R_u(x)\le R_v(x)=v(x)= u(x)\le R_u(x)$.
\end{proof}
Using results of \cite[Section 4]{HN-mertens} this
leads to a characterization of the equality $R_u=u$ for nearly hyperharmonic
$\widetilde{\B^\ast}$-measurable functions.
To this end we recall that the
$\sigma$-algebra of all finely Borel subsets of $X$ (that is, the smallest $\sigma$-algebra
on~$X$ containing all finely open sets) is the smallest $\sigma$-algebra containing~$\mathcal B$ and all
semipolar sets; see \cite[Section 5]{HN-mertens}. In particular, $\tilde \mathcal B\subset \mathcal B^f$.
\begin{theorem}\label{general-2}Let $u\in\N$ and suppose that $u$ is $\widetilde{\B^\ast}$-measurable.
Then the following statements are equivalent:
\begin{itemize}
\item[\rm (i)]
$u=\inf \{w\in \mathcal E_{\mathfrak X}} % {E_{\mathbbm P}\colon w\ge u\}$.
\item[\rm (ii)]
$u$ is finely upper semicontinuous.
\item[\rm (iii)]
$u$ is finely Borel measurable.
\item[\rm (iv)]
$u$ is $\tilde \mathcal B$-measurable.
\item[\rm (v)]
The set $\{\hat u<u\}$ is semipolar.
\end{itemize}
\end{theorem}
\begin{proof} Trivially, (i)\,$\Rightarrow$\,(ii)\,$\Rightarrow$\,(iii).
Moreover, (iii)\,$\Leftrightarrow$\,(v) and (iii)\,$\Rightarrow$\,(iv),
by \cite[Proposition 5.3 and Corollary 5.4]{HN-mertens}. By Theorem \ref{tilde-Borel},
(iv)\,$\Rightarrow$\,(i).
\end{proof}
Clearly, previous statements on reduced functions $R_\varphi$ can now
be extended to functions $\varphi\ge 0$ on $X$ which are only supposed to be $\mathcal B^f\cap \widetilde{\B^\ast}$-measurable.
\begin{remark}\label{erratum}{\rm
The result \cite[Corollary 5.4]{HN-mertens}
relies on \cite[Proposition 5.2]{HN-mertens}
the proof of which uses \cite[Theorem 1.5]{S-negligible} stating that, given a semipolar set $S$,
there exists a~measure $\mu$ on $X$ such that $\mu^\ast(B)>0$ for every non-polar subset $B$
of $S$.
This is correct; its proof, however, is not, since \cite[Lemma 1.3]{S-negligible} is wrong.
Assuming without loss of generality that $S$ is the union of totally thin Borel sets~$F_n$, $n\in\mathbbm{N}$,
we many obtain a valid proof exhausting each $F_n$ by sets
\begin{equation*}
F_{n,m}:=F_n\cap K_m\cap \{ \hat P_{F_n}q<\eta_m q\} , \qquad m\in\mathbbm{N},
\end{equation*} where $q$ is a continuous strict potential on $X$,
$(K_m)$ is a sequence of compacts in $X$ and $\eta_m\in (0,1)$ such that $K_m\uparrow X$ and $\eta_m\uparrow 1$
as $m\to\infty$.
Indeed, let us fix $x\in X$ and $m,n\in \mathbbm{N}$. Let $F:=F_{n,m}$, $\eta:=\eta_{n,m}$. We recursively define
measures $\mu_k$ on $F$ taking $\mu_0:=\ve_x$ and
\begin{equation*}
\mu_k:=\hat\mu_{k-1} ^F:=\int \hat\varepsilon_y^F\,d\mu_{k-1}(y), \qquad k\in\mathbbm{N}.
\end{equation*}
Then $\int q\,d\mu_k=\int \hat P_Fq\,d\mu_{k-1}\le \eta \int q \, d\mu_{k-1}$, and hence
$ \inf q(F)\cdot \mu_k(F) \le \eta^k q(x)$ for every $k\in\mathbbm{N}$.
So $\lim_{k\to\infty} \mu_k=0$ which leads to a proof of \cite[Theorem 1.5]{S-negligible} not using
\cite[Lemma 1.3]{S-negligible}.
}
\end{remark}
\section{Application of our method to right processes}\label{standard}
It should be clear to the experts that our approach works as well
for a general right process $\mathfrak X$ on a Radon space $X$
provided we assume that the associated potential kernel is proper.
It would then be convenient to define~$\N$ to be the set of all functions
$u\colon X\to [0,\infty]$ such that $\int^\ast u\circ X_{D_K}\,dP^x\le u(x)$
for all $x\in X$ and compacts $K$ not containing $x$. By a result of Dynkin
(see \cite[Theoem II.5.1]{BG}), nearly Borel measurable
functions in $\N$ are supermedian with respect to the resolvent $(V_\lambda)$ associated
with $\mathfrak X$ and, essentially, we only have to replace the lower semicontinuous
regularization $\hat u$ of $u$ (which for a nearly hyperharmonic function in our setting of a~nice
Hunt process is the greatest excessive minorant of~$u$)
by the excessive regularization $\tilde u:=\lim_{\lambda\to \infty } \lambda V_\lambda u$.
|
\section{Introduction}
\label{intro}
Understanding the data-generating mechanism (g.m.) has been a main theme of causal inference. To infer the causal direction between two random variables (r.v.s) $X$ and $Y$ using passive observations, most existing approaches first model the relation between them using a functional model with certain assumptions \citep{shimizu2006linear, hoyer2009nonlinear, zhang2009identifiability, janzing2012information}. Then a certain asymmetric property (usually termed \textit{cause-effect asymmetry}), which only holds in the causal direction, is derived to conduct inference. For example, the additive noise model (ANM) \citep{hoyer2009nonlinear} represents the effect as a function of the cause with an additive independent noise: $Y = f(X) + \epsilon$. It is shown in \citep{hoyer2009nonlinear} that there is no model of the form $X = g(Y) + \tilde{\epsilon}$ that admits an ANM in the anticausal direction for most combinations $\left(f, p(X), p(\epsilon)\right)$.
Similar to ANM, most causal inference approaches based on functional models, such as LiNGAM \citep{shimizu2006linear}, PNL \citep{zhang2009identifiability}, and IGCI \citep{janzing2010causal}, assume a single causal model for all observations. However, there is no such a guarantee in practice, and it could be very common that the observations are generated by a mixture of causal models due to different data sources or data collection under different conditions, rendering existing single-causal-model based approaches inapplicable in many problems (e.g. Fig. \ref{Fig:toy}). Recently, an approach was proposed for inferring the causal direction of mixtures of ANMs with discrete variables \citep{liu2016causal}. However, the inference of such mixture models with continuous variables remains a challenging problem and is not yet well studied.
Another question regarding mixture models addressed in this paper is how one could reveal causal knowledge in clustering tasks. Specifically, we aim at finding clusters consistent with the causal g.m.s of a mixture model, which is usually vital in the preliminary phase of many research. For example in the analysis of air data (see section \ref{real:exp} for detail), discovering knowledge from air data combined from several different regions (i.e. mechanisms in causal perspective) is much more difficult than from data of each region separately. Most existing clustering algorithms are weak for this perspective as they typically define similarity between observations in the form of distances in some spaces or manifolds. Most of them neglect the relation among r.v.s within a feature vector (observation), and only use those feature dimensions to calculate an overall distance metric as the clustering criterion.
\begin{figure}[t]
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{toy/dist.png}
\caption{}
\label{Fig:toy_dist}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{toy/p02.png}
\caption{}
\label{Fig:toy_p02}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{toy/p06.png}
\caption{}
\label{Fig:toy_p06}
\end{subfigure}
\caption{Example illustrating the failure of ANM on the inference of a mixture of ANMs (a) the distribution of data generated from $M_1 : Y = X^{2} + \epsilon$ (red) and $M_2 : Y = X^{5} + \epsilon$ (blue), where $X \sim U(0, 1)$ ($x$-axis) and $\epsilon \sim U(-0.1, 0.1)$ ; (b) Conditional $p(Y|X = 0.2)$; (c) Conditional $p(Y|X = 0.6)$. It is obvious that when the data is generated from a mixture of ANMs, the consistency of conditionals is likely to be violated which leads to the failure of ANM.}
\label{Fig:toy}
\end{figure}
In this paper, we focus on analyzing observations generated by a mixture of ANMs of two r.v.s and try to answer two questions: 1) \textit{causal inference:} how can we infer the causal direction between the two r.v.s? 2) \textit{mechanism clustering:} how can we cluster the observations generated from the same g.m. together? To answer these questions, first as the main result of this paper, we show that the causal direction of the mixture of ANMs is identifiable in most cases, and we propose a variant of GP-LVM \citep{lawrence2005probabilistic} named Gaussian Process Partially Observable Model (GPPOM) for model estimation, based on which we further develop the algorithms for causal inference and mechanism clustering.
The rest of the paper is organized as follows: in section 2, we formalize the model, show its identifiability and elaborate mechanism clustering; in section 3, model estimation method is proposed; we present experiments on synthetic and real world data in section 4 and conclude in section 5.
\begin{wrapfigure}[12]{r}{0.4\textwidth}
\vspace{-19mm}
\begin{center}
\begin{tikzpicture}
\node[obs] (y) {$\mathbf{y}_{n}$};
\node[latent, above=of y] (f) {$\mathbf{f}_{n}$};
\node[obs, above=of f] (X) {$X$};
\node[latent, right=1.0cm of X] (tha) {$\theta$};
\node[latent, right=1.0cm of f] (e) {$\bm{\epsilon}_{n}$};
\node[latent, right=1.0cm of e] (beta){$\beta$};
\edge {f,e} {y}; %
\edge {tha} {f};
\edge {X} {f};
\edge {beta} {e};
\plate {yx} {(y)(f)(e)} {$N$};
\end{tikzpicture}
\end{center}
\caption{ANM Mixture Model}
\label{fig:dag}
\vspace{-2mm}
\end{wrapfigure}
\section{ANM Mixture Model}
\subsection{Model definition}
Each observation is assumed to be generated from an ANM and the entire data set is generated by a finite number of related ANMs. They are called the ANM Mixture Model (ANM-MM) and formally defined as:
\begin{dfnt}[ANM Mixture Model]
\label{def:anmmm}
An ANM Mixture Model is a set of causal models of the same causal direction between two continuous r.v.s $X$ and $Y$. All causal models share the same form given by the following ANM:
\begin{align}
Y = f(X;\theta) + \epsilon,
\label{mix_anm}
\end{align}
where $X$ denotes the cause, $Y$ denotes the effect, $f$ is a nonlinear function parameterized by $\theta$ and the noise $\epsilon \perp\!\!\!\perp X$. The differences between causal models in an ANM-MM stem only from different values of function parameter $\theta$. In ANM-MM, $\theta$ is assumed to be drawn from a discrete distribution on a finite set $\Theta = \{\theta_1,\cdots,\theta_{C}\}$, i.e. $\theta \sim p_{\theta}(\theta)= \sum_{c=1}^{C} a_{c} \mathbf{1}_{\theta_{c}}(\cdot)$, where $a_{c}>0$, $\sum_{c=1}^{C} a_c =1$ and $\mathbf{1}_{\theta_{c}}(\cdot)$ is the indicator function of a single value $\theta_{c}$.
\end{dfnt}
Obviously in ANM-MM, all observations are generated by a set of g.m.s, which share the same function form ($f$) but differ in parameter values ($\theta$). This model is inspired by commonly encountered cases where the data-generating process is slightly different in each independent trial due to the influence of certain external factors that one can hardly control. In addition, these factors are usually believed to be independent of the observed variables. The data-generating process of ANM-MM can be represented by a directed graph in Fig. \ref{fig:dag}.
\subsection{Causal inference: identifiability of ANM-MM}
Let $X$ be the cause and $Y$ be the effect ($X \to Y$) without loss of generality. As most recently proposed causal inference approaches, following postulate, which was originally proposed in \citep{daniusis2012inferring}, is adopted in the analysis of ANM-MM.
\begin{post}[Independence of input and function]
\label{indp_post}
If $X \to Y$, the distribution of $X$ and the function $f$ mapping $X$ to $Y$ are independent since they correspond to independent mechanisms of nature.
\end{post}
In a general perspective, postulate 1 essentially claims the independence
between the cause ($X$) and mechanism mapping the cause to effect \citep{janzing2010causal}. In ANM-MM, we interpret the independence between the cause and mechanism in an intuitive way: $\theta$, as the function parameter, captures all variability of mechanisms $f$ so it should be independent of the cause $X$ according to postulate \ref{indp_post}. Based on the independence between $X$ and $\theta$, cause-effect asymmetry could be derived to infer the causal direction.
Since ANM-MM consists of a set of ANMs, the identifiability result of ANM-MM can be a simple corollary of that in \citep{hoyer2009nonlinear} when the number of ANMs ($C$) is equal and there is a one-to-one correspondence between mechanisms in the forward and backward ANM-MM. In this case the condition of ANM-MM being unidentifiable is to fulfill $C$ ordinary differential equations given in \citep{hoyer2009nonlinear} simultaneously which can hardly happen in a generic case. However, $C$ in ANM-MM in both directions may not necessarily be equal and there may also exist many-to-one correspondence between ANMs in both directions. In this case, the identifiability result can not be derived as a simple corollary of \citep{hoyer2009nonlinear}. To analyze the identifiability result of ANM-MM, we first derive lemma \ref{Thm:no_backward_anm} to find the condition of existence of many-to-one correspondence (which is a generalization of the condition given in \citep{hoyer2009nonlinear}), then conclude the identifiability result of ANM-MM (theorem \ref{Thm:no_backward_anm_mm}) based on the condition in lemma \ref{Thm:no_backward_anm}. The condition that there exists one backward ANM for a forward ANM-MM is:
\begin{lema}\label{Thm:no_backward_anm}
Let $X \to Y$ and they follow an ANM-MM. If there exists a backward ANM in the anti-causal direction, i.e.
\[
X = g(Y) + \tilde{\epsilon},
\]
the cause distribution ($p_{X}$), the noise distribution ($p_{\epsilon}$), the nonlinear function ($f$) and its parameter distribution ($p_{\theta}$) should jointly fulfill the following ordinary differential equation (ODE)
\begin{equation}\label{Eq:ode}
\xi''' - \frac{G(X,Y)}{H(X,Y)} \xi'' = \frac{G(X,Y)V(X,Y)}{U(X,Y)} - H(X,Y),
\end{equation}
where $\xi := \log p_{X}$, and the definitions of $G(X, Y)$, $H(X,Y)$, $V(X,Y)$ and $U(X,Y)$ are provided in supplementary due to the page limitation.
\end{lema}
\textit{Sketch of proof.} Since $X$ and $Y$ follow an ANM-MM, their joint density is factorized in the causal direction by $p(X,Y)= \sum_{c=1}^{C} p(Y|X,\theta_c) p_{X}(X) p_{\theta}(\theta_c)
= p_{X}(X) \sum_{c=1}^{C} a_c p_{\epsilon}(Y-f(X; \theta_{c}))$. If there exists a backward ANM in the anti-causal direction, i.e. $X=g(Y) + \tilde{\epsilon}$, then $p(X,Y) = p_{\tilde{\epsilon}}(X-g(Y))p_{Y}(Y)$ and $\frac{\partial}{\partial X} \left( \frac{\partial^2 \pi / \partial X \partial Y}{\partial^2 \pi / \partial X^2} \right) = 0$ holds, where $\pi = \log \left[ p_{\tilde{\epsilon}}(X-g(Y))p_{Y}(Y) \right]$, in the backward ANM. Since $p(X,Y)$ should be the same, by substituting $p(X,Y)= p_{X}(X) \sum_{c=1}^{C} a_c p_{\epsilon}(Y-f(X; \theta_{c}))$ into $\frac{\partial}{\partial X} \left( \frac{\partial^2 \pi / \partial X \partial Y}{\partial^2 \pi / \partial X^2} \right) = 0$, the condition shown in \eqref{Eq:ode} is obtained.
The proof of lemma \ref{Thm:no_backward_anm} follows the idea of the identifiability of ANM in \citep{hoyer2009nonlinear} and is provided in the supplementary. Since the condition that one backward ANM exists for an forward ANM-MM (mixture of ANMs) is more restrictive than that for a single forward ANM, which is the identifiability in \citep{hoyer2009nonlinear}, lemma \ref{Thm:no_backward_anm} indicates that a backward ANM is unlikely to exist in the anticausal direction if 1) $X$ and $Y$ follow an ANM-MM; 2) postulate \ref{indp_post} holds. Based on lemma \ref{Thm:no_backward_anm}, it is reasonable to hypothesize that a stronger result, which is justified in theorem \ref{Thm:no_backward_anm_mm}, is valid, i.e. if the g.m. follows an ANM-MM, then it is almost impossible to have a backward ANM-MM in the anticausal direction.
\begin{thrm}\label{Thm:no_backward_anm_mm}
Let $X \to Y$ and they follow an ANM-MM. If there exists a backward ANM-MM,
\[
X = g(Y;\omega) + \tilde{\epsilon},
\]
where $\omega\sim p_{\omega}(\omega)=\sum_{\tilde{c}=1}^{\tilde{C}} b_{{\tilde{c}}} \mathbf{1}_{\omega_{\tilde{c}}}(\cdot)$, $b_{\tilde{c}}>0$, $\sum_{\tilde{c}=1}^{\tilde{C}} b_{\tilde{c}} =1$ and $\tilde{\epsilon} \perp\!\!\!\perp Y$, in the anticausal direction, then ($p_{X}$, $p_{\epsilon}$, $f$, $p_{\theta}$) should fulfill $\tilde{C}$ ordinary differential equations similar to \eqref{Eq:ode}, i.e.,
\begin{equation}\label{Eq:ode_t}
\xi''' - \frac{G^{(\tilde{c})}(X,Y)}{H^{(\tilde{c})}(X,Y)} \xi'' = \frac{G^{(\tilde{c})}(X,Y)V^{(\tilde{c})}(X,Y)}{U^{(\tilde{c})}(X,Y)} - H^{(\tilde{c})}(X,Y), ~ \tilde{c} = 1,2,\cdots,\tilde{C},
\end{equation}
where $\xi := \log p_{X}$, $G^{(\tilde{c})}(X,Y)$, $H^{(\tilde{c})}(X,Y)$, $U^{(\tilde{c})}(X,Y)$ and $V^{(\tilde{c})}(X,Y)$ are defined similarly to those in lemma \ref{Thm:no_backward_anm}.
\end{thrm}
\begin{proof}
Assume that there exists ANM-MM in both directions. Then there exists a non overlapping partition of the entire data $\bm{\mathcal{D}}\coloneqq \lbrace (x_{n},y_{n})\rbrace_{n=1}^{N} = \bm{\mathcal{D}}_{1} \cup \cdots \cup \bm{\mathcal{D}}_{\tilde{C}}$ such that in each data block $\bm{\mathcal{D}}_{\tilde{c}}$, there is an ANM-MM in the causal direction $Y = f(X;\theta) + \epsilon$, where $\theta \sim p^{(\tilde{c})}_{\theta}(\theta)$ is a discrete distribution on a finite set $\Theta^{(\tilde{c})} \subseteq \Theta$, and an ANM in the anti-causal direction $X = g(Y;\omega = \omega_{\tilde{c}}) + \tilde{\epsilon}$. According to lemma \ref{Thm:no_backward_anm}, for each data block, to ensure the existence of an ANM-MM in the causal direction and an ANM in the anti-causal direction, ($p_{X}$, $p_{\epsilon}$, $f$, $p_{\theta}$) should fulfill an ordinary differential equation in the form of \eqref{Eq:ode}. Then the existence of backward ANM-MM requires $\tilde{C}$ ordinary differential equations to be fulfilled simultaneously which yields \eqref{Eq:ode_t}.
\end{proof}
Then the causal direction in ANM-MM can be inferred by investigating the independence between the hypothetical cause and the corresponding function parameter. According to theorem \ref{Thm:no_backward_anm_mm}, if they are independent in the causal direction, then it is highly likely they are dependent in the anticausal direction. Therefore in practice, the inferred direction is the one that shows more evidence of independence between them.
\subsection{Mechanism clustering of ANM-MM}
In ANM-MM, $\theta$, which represents function parameters, can be directly used to identify different g.m.s since each parameter value corresponds to one mechanism. In other words, observations generated by the same g.m. would have the same $\theta$ if the imposed statistical model is identifiable with respect to $\theta$.
Denote the parameter associated with each observation $(x_{n}, y_{n})$ by $\bm{\theta}_{n}$, we suppose a more practical inherent clustering structure behind hidden ${\bm \theta}_n$. Formally, there is a grouping indicator of integers ${\mathbf z}\in\{1,\ldots, C\}^N$ that assign each ${\bm\theta}_n$ to one of the $C$ clusters, through the $n$th element of $\mathbf z$, e.g. ${\bm \theta}_n$ belongs to cluster $c$ if $[\mathbf{z}]_n=c, c\in \{1,\ldots, C\}.$ Following ANM-MM, we may assume each ${\bm \theta}_n$ belong to one of $C$ components and each component follows $\mathcal{N}(\mu_{c}, \sigma^2)$. A likelihood-based clustering scheme suggests minimizing $-\ell$ jointly with respect to all means and $\mathbf z$
\begin{eqnarray*}
\ell(\bm{\mathcal{M}}, {\mathbf z}) = \log \prod_{n=1}^N \prod_{c=1}^C \left\{ \frac{1}{\sqrt{2\pi}\sigma} \exp\left( -\frac{1}{2\sigma^2} ({\bm \theta}_n - \mu_c)^2 \right) \right\}^{\mathbf{1}([\mathbf{z}]_{n} = c)},
\label{eq:clustloglike}
\end{eqnarray*}
where $\bm{\mathcal{M}} = \{ \mu_{c}\}_{c=1}^{C}$ and $\mathbf{1}(\cdot)$ is the indicator function. To simplify further let's ignore the known $\sigma^2$ and minimize $-\ell$ using coordinate descent iteratively
\begin{eqnarray}
\hat {\bm{\mathcal{M}}} \mid {\mathbf z} &=& \argmin_{\bm{\mathcal{M}}} \sum_{c=1}^C \sum_{\{n \mid [\mathbf{z}]_n=c\}} ({\bm \theta}_n -\mu_c)^2 \label{eq:km_center}\\
\hat {\mathbf z}\mid {\bm{\mathcal{M}}} &=& \argmin_{\mathbf z} \sum_{c=1}^C \sum_{\{n \mid [\mathbf{z}]_n=c\}} ({\bm \theta}_n -\mu_c)^2 \label{eq:km_group}.
\end{eqnarray}
The minimizer of \eqref{eq:km_center} is the mean
$\hat{\mu_c} = \frac{1}{n_c} \sum_{\{n \mid [\mathbf{z}]_n=c\}} {\bm \theta_n},$
where $n_c$ is the size of the $c$th cluster $n_c = \sum_{n=1}^N \mathbf{1}([\mathbf{z}]_n=c).$ The minimizer of \eqref{eq:km_group} is group assignment through minimum Euclidean distance. Therefore, iterating between \eqref{eq:km_center} and \eqref{eq:km_group} coincides with applying $k$-means algorithm on all ${\bm \theta}_n$ and the goal of finding clusters consistent with the g.m.s for data from ANM-MM can be achieved by firstly estimating parameters associated with each observation and then conducting $k$-means directly on parameters.
\section{ANM-MM Estimation by GPPOM}
We propose Gaussian process partially observable model (GPPOM) and incorporate Hilbert-Schmidt independence criterion (HSIC) \citep{gretton2005measuring} enforcement into GPPOM to estimate the model parameter $\theta$. Then we summarize algorithms for causal inference and mechanism clustering of ANM-MM.
\subsection{Preliminaries}
\textbf{Dual PPCA.} Dual PPCA \citep{lawrence2004gaussian} is a latent variable model in which maximum likelihood solution for the latent variables is found by marginalizing out the parameters. Given a set of $N$ centered $D$-dimensional data $\mathbf{Y}=\left[\bm{y}_{1}, \dots, \bm{y}_{N}\right]^{T}$, dual PPCA learns the $q$-dimensional latent representation $\bm{x}_{n}$ associated with each observation $\bm{y}_{n}$. The relation between $\bm{x}_{n}$ and $\bm{y}_{n}$ in dual PPCA is $\bm{y}_{n} = \mathbf{W}\bm{x}_{n} + \bm{\epsilon}_{n}$, where the matrix $\mathbf{W}$ specifies the linear relation between $\bm{y}_{n}$ and $\bm{x}_{n}$ and noise $\bm{\epsilon}_{n} \sim\mathcal{N}(\mathbf{0}, \beta^{-1}\mathbf{I})$. Then by placing a standard Gaussian prior on each row of $\mathbf{W}$, one obtains the marginal likelihood of all observations and the objective function of dual PPCA is the log-likelihood $\mathcal{L} = -\frac{DN}{2}\ln(2\pi) - \frac{D}{2}\ln\left( |\mathbf{K}| \right) - \frac{1}{2} \tr\left( \mathbf{K}^{-1} \mathbf{YY}^{T} \right)$, where $\mathbf{K}=\mathbf{XX}^{T}+\beta^{-1}\mathbf{I}$ and $\mathbf{X}=\left[\bm{x}_{1}, \dots, \bm{x}_{N}\right]^{T}$.
\textbf{GP-LVM.} GP-LVM \citep{lawrence2005probabilistic} generalizes dual PPCA to cases of nonlinear relation between $\bm{y}_{n}$ and $\bm{x}_{n}$ by mapping latent representations in $\mathbf{X}$ to a feature space, i.e. $\bm{\Phi} = \left[\phi(\bm{x}_{1}), \dots, \phi(\bm{x}_{N})\right]^{T}$, where $\phi(\cdot)$ denotes the canonical feature map. Then $\mathbf{K} = \bm{\Phi} \bm{\Phi}^{T} + \beta^{-1}\mathbf{I}$ and $\bm{\Phi} \bm{\Phi}^{T}$ can be computed using kernel trick. GP-LVM can also be interpreted as a new class of models which consists of $D$ independent Gaussian processes \citep{williams1998prediction} mapping from a latent space to an observed data space \citep{lawrence2005probabilistic}.
\textbf{HSIC.}
HSIC \citep{gretton2005measuring}, which is based on reproducing kernel Hilbert space (RKHS) theory, is widely used to measure the dependence between r.v.s. Let $\bm{\mathcal{D}}\coloneqq \lbrace (\bm{x}_{n}, \bm{y}_{n}) \rbrace_{n=1}^{N}$ be a sample of size $N$ draw independently and identically distributed according to $P(X,Y)$, HSIC answers the query whether $X \perp\!\!\!\perp Y$. Formally, denote by $\mathcal{F}$ and $\mathcal{G}$ RKHSs with universal kernel $k$, $l$ on the compact domains $\mathcal{X}$ and $\mathcal{Y}$, HSIC is the measure defined as $\text{HSIC}(P(X,Y), \mathcal{F}, \mathcal{G}) \coloneqq \Vert \mathcal{C}_{xy} \Vert^{2}_{\text{HS}}$, which is essentially the squared Hilbert Schmidt norm \citep{gretton2005measuring} of the cross-covariance operator $\mathcal{C}_{xy}$ from RKHS $\mathcal{G}$ to $\mathcal{F}$ \citep{fukumizu2004dimensionality}. It is proved in \citep{gretton2005measuring} that, under conditions specified in \citep{gretton2005kernel}, $\text{HSIC}(P(X,Y), \mathcal{F}, \mathcal{G})= 0$ if and only if $X \perp\!\!\!\perp Y$. In practice, a biased empirical estimator of HSIC based on the sample $\bm{\mathcal{D}}$ is often adopted:
\begin{align}
\text{HSIC}_{b}(\bm{\mathcal{D}}) = \frac{1}{N^{2}} \tr\left(\mathbf{KHLH} \right),
\label{eq:emp_hsic}
\end{align}
where $\left[\mathbf{K} \right]_{ij} = k(\bm{x}_{i}, \bm{x}_{j})$, $\left[\mathbf{L} \right]_{ij} = l(\bm{y}_{i}, \bm{y}_{j})$, $\mathbf{H} = \mathbf{I} - \frac{1}{N}\vec{\mathbf{1}}\vec{\mathbf{1}}^{T}$, and $\vec{\mathbf{1}}$ is a $N \times 1$ vector of ones.
\subsection{Gaussian process partially observable model}
\textbf{Partially observable dual PPCA.} Dual PPCA is not directly applicable to model ANM-MM since: 1) part of the r.v. that maps to the effect is visible (i.e. $X$); 2) the relation (i.e. $f$) is nonlinear; 3) r.v.s that contribute to the effect should be independent ($X \perp\!\!\!\perp \theta$) in ANM-MM. To tackle 1), a latent r.v. $\theta$ is brought in dual PPCA.
Denote the observed effect by $\mathbf{Y}=\left[\bm{y}_{1}, \dots, \bm{y}_{N}\right]^{T}$, observed cause by $\mathbf{X}=\left[\bm{x}_{1}, \dots, \bm{x}_{N}\right]^{T}$, the matrix collecting function parameters associated with each observation by $\bm{\Theta}=\left[\bm{\theta}_{1}, \dots, \bm{\theta}_{N}\right]^{T}$ and the r.v. that contribute to the effect by $\tilde{X} = [X, \theta]$. Similar to dual PPCA, the relation between the latent representation and the observation is given by
\begin{align}
\bm{y}_{n} = \tilde{\mathbf{W}} \tilde{\bm{x}}_{n} + \bm{\epsilon}_{n}, \quad n=1,\dots,N \nonumber
\end{align}
where $\tilde{\bm{x}}_{n} = \left[ \bm{x}^{T}_{n}, \bm{\theta}^{T}_{n}\right]^{T}$, $\tilde{\mathbf{W}}$ is the matrix specifies the relation between $\bm{y}_{n}$ and $\tilde{\bm{x}}_{n}$, $\bm{\epsilon}_{n}\sim \mathcal{N}(\mathbf{0}, \beta^{-1}\mathbf{I})$ is the additive noise. Then by placing a standard Gaussian prior on $\tilde{\mathbf{W}}$, i.e. $p(\tilde{\mathbf{W}})=\prod_{i=1}^{D}\mathcal{N} (\tilde{\mathbf{w}}_{i,:}|\mathbf{0}, \mathbf{I})$, where $\tilde{\mathbf{w}}_{i,:}$ is the $i$th row of the matrix $\tilde{\mathbf{W}}$, the log-likelihood of the observations is given by
\begin{align}
\mathcal{L}(\bm{\Theta} |\mathbf{X}, \mathbf{Y}, \beta) = -\frac{DN}{2}\ln(2\pi) - \frac{D}{2}\ln\left( |\mathbf{\tilde{K}}| \right) - \frac{1}{2} \tr\left( \mathbf{\tilde{K}}^{-1}\mathbf{YY}^{T} \right),
\label{our_lik}
\end{align}
where $\mathbf{\tilde{K}} = \mathbf{\tilde{X}}\mathbf{\tilde{X}}^{T} + \bm{\beta}^{-1}\mathbf{I} = \left[ \mathbf{X}, \bm{\Theta} \right] \left[ \mathbf{X}, \bm{\Theta} \right]^{T} + \bm{\beta}^{-1}\mathbf{I} = \mathbf{XX^{T}} + \mathbf{\Theta\Theta^{T}} + \bm{\beta}^{-1}\mathbf{I}$ is the covariance matrix after bringing in $\theta$.
\begin{wrapfigure}[20]{R}{0.5\textwidth}
\IncMargin{1.5em}
\begin{algorithm}[H]
\SetKwData{Left}{left}
\SetKwData{This}{this}
\SetKwData{Up}{up}
\SetKwFunction{Union}{Union}
\SetKwFunction{FindCompress}{FindCompress}
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\Input{$\bm{\mathcal{D}} = \{(\bm{x}_{n}, \bm{y}_{n})\}_{n=1}^{N}$ - the set of observations of two r.v.s;\\ $\lambda$ - parameter of independence}
\Output{The causal direction}
\BlankLine
Standardize observations of each r.v.\;
Initialize $\beta$ and kernel parameters\;
Optimize \eqref{eq:opt_obj} in both directions, denote the the value of HSIC term by $\text{HSIC}_{X\to Y}$ and $\text{HSIC}_{Y\to X}$, respectively\;
\uIf{$\text{HSIC}_{X\to Y}$ < $\text{HSIC}_{Y\to X}$}{The causal direction is $X\to Y$\;}
\uElseIf{$\text{HSIC}_{X\to Y}$ > $\text{HSIC}_{Y\to X}$}{The causal direction is $Y\to X$\;}
\Else{No decision made.}
\caption{Causal Inference}\label{alg_cd}
\end{algorithm}
\DecMargin{1.5em}
\end{wrapfigure}
\textbf{General nonlinear cases (GPPOM).} Similar to the generalization from dual PPCA to GP-LVM, the dual PPCA with observable $X$ and latent $\theta$ can be easily generalized to nonlinear cases. Denote the feature map by $\phi(\cdot)$ and $\bm{\Phi} = \left[\phi(\tilde{\bm{x}}_{1}), \dots, \phi(\tilde{\bm{x}}_{N})\right]^{T}$, then the covariance matrix is given by $\tilde{\mathbf{K}} = \bm{\Phi} \bm{\Phi}^{T} + \bm{\beta}^{-1}\mathbf{I}$. The entries of $\bm{\Phi} \bm{\Phi}^{T}$ can be computed using kernel trick given a selected kernel $k(\cdot,\cdot)$. In this paper, we adopt the radial basis function (RBF) kernel, which reads $k(\bm{x}_{i}, \bm{x}_{j}) = \exp\left( - \sum_{d=1}^{D_{x}} \gamma_{d} ( \bm{x}_{id} - \bm{x}_{jd})^{2} \right)$, where $\gamma_{d}, \text{for}~ d=1,\dots, D_{x}$, are free parameters and $D_{x}$ is the dimension of the input. As a result of adopting RBF kernel, the covariance matrix $\tilde{\mathbf{K}}$ in \eqref{our_lik} can be computed as
\begin{align}
\tilde{\mathbf{K}}= \bm{\Phi} \bm{\Phi}^{T} + \bm{\beta}^{-1}\mathbf{I} = \mathbf{K}_{X} \circ \mathbf{K}_{\theta} + \bm{\beta}^{-1}\mathbf{I}, \nonumber
\end{align}
where $\circ$ denotes the Hadamard product, the entries on $i$th row and $j$th column of $\mathbf{K}_{X}$ and $\mathbf{K}_{\theta}$ are given by $\left[\mathbf{K}_{X}\right]_{ij}=k(\bm{x}_{i}, \bm{x}_{j})$ and $\left[\mathbf{K}_{\theta}\right]_{ij}=k(\bm{\theta}_{i}, \bm{\theta}_{j})$, respectively. After the nonlinear generalization, the relation between $Y$ and $\tilde{X}$ reads $Y = f(\tilde{X}) + \epsilon = f(X, \theta) + \epsilon$. This variant of GP-LVM with partially observable latent space is named GPPOM in this paper. Like GP-LVM, $\tilde{X}$ is mapped to $Y$ by the same set of Gaussian processes in GPPOM so the differences in the g.m.s is captured by $\bm{\theta}_{n}$, the latent representations associated with each observation.
\subsection{Model estimation by independence enforcement}
Both dual PPCA and GP-LVM finds the latent representations through log-likelihood maximization using scaled conjugate gradient \citep{moller1993scaled}. However, the $\bm{\theta}$ can not be found by directly conducting likelihood maximization since the ANM-MM requires additionally the independence between $X$ and $\bm{\theta}$. To this end, we include HSIC \citep{gretton2005measuring} in the objective. By incorporating HSIC term into the negative log-likelihood of GPPOM, the optimization objective reads
\begin{align}
\argmin_{ \bm{\Theta}, \Omega }{ \mathcal{J}(\bm{\Theta}) } = \argmin_{ \bm{\Theta}, \Omega }{ \left[- \mathcal{L}(\bm{\Theta} |\mathbf{X}, \mathbf{Y}, \Omega) + \lambda \log \text{HSIC}_{\text{b}}(\mathbf{X}, \bm{\Theta})\right] },
\label{eq:opt_obj}
\end{align}
where $\lambda$ is the parameter which controls the importance of the HSIC term and $\Omega$ is the set of all hyper parameters including $\beta$ and all kernel parameters $\gamma_{d}$, $d = 1, \dots, D_{x}$.
To find $\bm{\Theta}$, we resort to the gradient descant methods. The gradient of the objective $\mathcal{J}$ with respect to latent points in $\bm{\Theta}$ is given by
\begin{align}
\frac{\partial \mathcal{J}}{\partial \left[\bm{\Theta}\right]_{ij}} = \tr\left[ \left( \frac{\partial \mathcal{J}}{\partial \mathbf{K}_{\theta}} \right)^{T} \frac{\partial \mathbf{K}_{\theta}}{\partial \left[\bm{\Theta}\right]_{ij}} \right].
\label{eq:grad}
\end{align}
The first part on the right hand side of \eqref{eq:grad}, which is the gradient of $\mathcal{J}$ with respect to the kernel matrix $\mathbf{K}_{\theta}$, can be computed as
\begin{align}
\frac{\partial \mathcal{J}}{\partial \mathbf{K}_{\theta}} = - \tr \left[ \left(\mathbf{\tilde{K}}^{-1}\mathbf{YY}^{T}\mathbf{\tilde{K}}^{-1} - D\mathbf{\tilde{K}}^{-1}\right)^{T} \left( \mathbf{K}_{X} \circ \mathbf{J}^{ij} \right) \right] + \lambda \frac{1}{\tr\left(\mathbf{K}_{X}\mathbf{HK}_{\theta}\mathbf{H} \right))} \mathbf{HK}_{X}\mathbf{H},
\label{grad_wrt_kz}
\end{align}
where $\mathbf{J}^{ij}$ is the single-entry matrix, 1 at $(i,j)$ and 0 elsewhere and $\mathbf{H} = \mathbf{I} - \frac{1}{N}\vec{\mathbf{1}}\vec{\mathbf{1}}^{T}$. Combining $\frac{\partial \mathcal{L}}{\partial \mathbf{K}_{\theta}}$ with $\frac{\partial \mathbf{K}_{\theta}}{\partial \left[\bm{\Theta}\right]_{ij}}$, whose entry on the $m$th row and $n$th column is given by $\frac{\partial \left[\mathbf{K}_{\Theta}\right]_{mn}}{\partial \left[\bm{\Theta}\right]_{ij}} = \frac{\partial k(\bm{\theta}_{m}, \bm{\theta}_{n})}{\partial \left[\bm{\Theta}\right]_{ij}}$, through the chain rule, all latent points in $\bm{\Theta}$ can be optimized. With $\bm{\Theta}$, one can conduct causal inference and mechanism clustering of ANM-MM. The detailed steps are given in Algorithm \ref{alg_cd} and \ref{alg_clu}.
\begin{wrapfigure}[12]{R}{0.5\textwidth}
\IncMargin{1.5em}
\begin{algorithm}[H]
\SetKwData{Left}{left}
\SetKwData{This}{this}
\SetKwData{Up}{up}
\SetKwFunction{Union}{Union}
\SetKwFunction{FindCompress}{FindCompress}
\SetKwInOut{Input}{input}
\SetKwInOut{Output}{output}
\Input{$\bm{\mathcal{D}} = \{(\bm{x}_{n}, \bm{y}_{n})\}_{n=1}^{N}$ - the set of observations of two r.v.s;\\ $\lambda$ - parameter of independence; \\$C$ - Number of clusters}
\Output{The cluster labels}
\BlankLine
Standardize observations of each r.v.\;
Initialize $\beta$ and kernel parameters\;
Find $\bm{\Theta}$ by optimizing \eqref{eq:opt_obj} in causal direction\;
Apply $k$-means on $\bm{\theta}_{n}$, $n=1,\dots, N$\;
\Return the cluster labels.
\caption{Mechanism clustering}\label{alg_clu}
\end{algorithm}
\DecMargin{1.5em}
\end{wrapfigure}
\section{Experiments}
In this section, experimental results on both synthetic and real data are given to show the performance of ANM-MM on causal inference and mechanism clustering tasks. The Python code of ANM-MM is available online at \url{https://github.com/amber0309/ANM-MM}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth, height=4cm]{figs/re_1.png}
\caption{Accuracy ($y$-axis) versus sample size ($x$-axis) on $Y=f(X; \theta_{c}) + \epsilon$ with different mechanisms. (a) $f_{1}$, (b) $f_{2}$, (c) $f_{3}$, (d) $f_{4}$.}
\label{Fig:cd1}
\end{figure}
\subsection{Synthetic data}
In experiments of causal inference, ANM-MM is compared with ANM \citep{hoyer2009nonlinear}, PNL \citep{zhang2009identifiability}, IGCI \citep{janzing2012information}, ECP \cite{zhang2015discovery} and LiNGAM \citep{shimizu2006linear}. The results are evaluated using accuracy, which is the percentage of correct causal direction estimation of 50 independent experiments. Note that ANM-MM was applied using different parameter $\lambda \in \{0.001, 0.01, 0.1, 1, 10\}$ and IGCI was applied using different reference measures and estimators. Their highest accuracy is reported.
In experiments of clustering, ANM-MM is compared with well-known $k$-means \citep{macqueen1967some} (similarity-based) on both raw data ($k$-means) and its PCA component (PCA-$k$m), Gaussian mixture clustering (GMM) \citep{rasmussen2000infinite} (model-based), spectral clustering (SpeClu) \citep{shi2000normalized} (spectral graph theory-based) and DBSCAN \citep{ester1996density} (density-based). Clustering performance is evaluated using average adjusted Rand index \citep{hubert1985comparing} (avgARI), which is the mean ARI over 100 experiments. High ARI ($\in [-1, 1]$) indicates good match between the clustering results and the ground truth. Sample size ($N$) is 100 in all synthetic clustering experiments. Clustering results are visualized in the supplementary\footnote{The results of PCA-$k$m are not visualized since they are similar to and worse than those of $k$-means.}.
\textbf{Different g.m.s and sample sizes.} We examine the performance on different g.m.s ($f$) and sample sizes ($N$). The mechanisms adopted are the following elementary functions: 1) $f_{1}=\frac{1}{1.5 + \theta_{c} X^{2}}$; 2) $f_{2}=2 \times X ^{\theta_{c} - 0.25} $; 3) $f_{3}= \exp(-\theta_{c} X)$; 4) $f_{4}= \tanh(\theta_{c} X)$. We tested sample size $N =$ 50, 100 and 200 for each mechanism. Given $f$ and $N$, the cause $X$ is sampled from a uniform distribution $U(0,1)$ and then mapped to the effect by $Y=f(X; \theta_{c}) + \epsilon, c\in \{1,2\}$, where $\theta_{1}\sim U(1,1.1)$, $\theta_{2}\sim U(3,3.1)$ and $\epsilon \sim \mathcal{N}(0, 0.05^{2})$. Each mechanism generates half of the observations.
\textit{Causal Inference.} The results are shown in Fig. \ref{Fig:cd1}. ANM-MM and ECP outperforms others based on a single causal model, which is consistent with our anticipation. Compared with ECP, ANM-MM shows slight advantages in 3 out of 4 settings. \textit{Clustering.} The avgARI values are summarized in (i) of Table \ref{tb:clu}. ANM-MM significantly outperforms other approaches in all mechanism settings.
\begin{table}
\centering
\caption{avgARI of synthetic clustering experiments}
\label{tb:clu}
\begin{tabular}{c|cccc|cc|cc|cc}
\toprule
\multirow{2}{*}{avgARI} & \multicolumn{4}{c}{(i) $f$} & \multicolumn{2}{c}{(ii) $C$} & \multicolumn{2}{c}{(iii) $\sigma$} & \multicolumn{2}{c}{(iv) $a_{1}$} \\ \cline{2-11} \\[-0.65em]
& $f_{1}$ & $f_{2}$ & $f_{3}$ & $f_{4}$ & $3$ & $4$ & $0.01$ & $0.1$ & $0.25$ & $0.75$ \\
\midrule
ANM-MM & \textbf{0.393} & \textbf{0.660} & \textbf{0.777} & \textbf{0.682} & \textbf{0.610} & \textbf{0.447} & \textbf{0.798} & \textbf{0.608} & \textbf{0.604} & \textbf{0.867} \\
$k$-means & 0.014 & 0.039 & 0.046 & 0.046 & 0.194 & 0.165 & 0.049 & 0.042 & 0.047 & 0.013 \\
PCA-$k$m & 0.013 & 0.037 & 0.044 & 0.048 & 0.056 & 0.041 & 0.047 & 0.040 & 0.052 & 0.014 \\
GMM & 0.015 & 0.340 & 0.073 & 0.208 & 0.237 & 0.202 & 0.191 & 0.025 & 0.048 & 0.381 \\
SpeClu & 0.003 & 0.129 & 0.295 & 0.192 & 0.285 & 0.175 & 0.595 & 0.048 & 0.044 & -0.008\\
DBSCAN & 0.055 & 0.265 & 0.342 & 0.358 & 0.257 & 0.106 & 0.527 & 0.110 & 0.521 & 0.718 \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[t]
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/multiple_cd.png}
\caption{}
\label{Fig:acc_nmech}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/diff_noise.png}
\caption{}
\label{Fig:acc_noise}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{figs/diff_prop.png}
\caption{}
\label{Fig:acc_prop}
\end{subfigure}
\caption{Accuracy (y-axis) versus (a) number of mechanisms; (b) noise standard deviation; (c) mixing proportion; on $f_{3}$ with $N=100$.}
\label{Fig:syn_cd_2}
\end{figure}
\textbf{Different number of g.m.s.}\footnote{From this part on, g.m. is fixed to be $f_{3}$.}
We examine the performance on different number of g.m.s ($C$ in Definition \ref{def:anmmm}). $\theta_{1}$, $\theta_{2}$ and $\epsilon$ are the same as in previous experiments. In the setting of three mechanisms, $\theta_{3}\sim U(0.5, 0.6)$. In the setting of four, $\theta_{3}\sim U(0.5, 0.6)$ and $\theta_{4}\sim U(2, 2.1)$. Again, the numbers of observations from each mechanism are the same.
\textit{Causal Inference.} The results are given in Fig. \ref{Fig:acc_nmech} which shows decreasing trend for all approaches. However, ANM-MM keeps 100\% when the number of mechanisms increases from 2 to 3. \textit{Clustering.} The avgARI values are given in (ii) and (i)$f_3$ of Table \ref{tb:clu}. The performance of different approaches show different trends which is probably due to the clustering principle they are based on. Although ANM-MM is heavily influenced by $C$, its performance is still much better than others.
\begin{wrapfigure}[15]{r}{0.5\textwidth}
\vspace{-5mm}
\begin{center}
\includegraphics[width=\linewidth]{figs/real_pairs.png}
\end{center}
\vspace{-5mm}
\caption{Accuracy on real cause-effect pairs.}
\label{fig:real_pairs}
\end{wrapfigure}
\textbf{Different noise standard deviations.} We examine the performance on different noise standard deviations $\sigma$. $\theta_{1}$, $\theta_{2}$ are the same as in the first part of experiments. Three different cases where $\sigma=0.01, 0.05$ and 0.1 are tested.
\textit{Causal Inference.} The results are given in Fig. \ref{Fig:acc_noise}. The change in $\sigma$ in this range does not significantly influence the performance of most causal inference approaches. ANM-MM keeps 100\% accuracy for all choice of $\sigma$. \textit{Clustering.} The avgARI values are given in (iii) and (i)$f_3$ of Table \ref{tb:clu}. As our anticipation, the clustering results heavily rely on $\sigma$ and all approaches show a decreasing trend in avgARI as $\sigma$ increases. However, ANM-MM is the most robust against large $\sigma$.
\textbf{Different mixing proportions.} We examine the performance on different mixing proportions ($a_{c}$ in Definition \ref{def:anmmm}). $\theta_{1}$, $\theta_{2}$ and $\sigma$ are the same as in the first part of experiments. Cases where $a_{1} = 0.25, 0.5$ and 0.75 (corresponding $a_{2} = 0.75, 0.5$ and 0.25) are tested.
\textit{Causal Inference.} The results on different $a_{1}$ are given in Fig. \ref{Fig:acc_prop}. Approaches based on a single causal model are sensitive to the change in $a_{1}$ whereas ECP and ANM-MM are more robust and outperform others. \textit{Clustering.} The avgARI values of experiments on different $a_{1}$ are given in (iv) and (i)$f_3$ of Table \ref{tb:clu}. The results of comparing approaches are significantly affected by $a_{1}$ and ANM-MM shows best robustness against the change in $a_{1}$.
\subsection{Real data}\label{real:exp}
\textbf{Causal inference on T\"{u}ebingen cause-effect pairs.} We evaluate the causal inference performance of ANM-MM on real world benchmark cause-effect pairs\footnote{https://webdav.tuebingen.mpg.de/cause-effect/.} \citep{mooij2016distinguishing}. Nine out of 41 data sets are excluded in our experiment because either they consists of multivariate or categorical data (pair 47, 52, 53, 54, 55, 70, 71, 101 and 105) or the estimated latent representations are extremely close\footnote{close in the sense that $|\theta_{i} - \theta_{j}|<0.001$.} (pair 12 and 17). Fifty independent experiments are repeated for each pair, and the percentage of correct inference of different approaches are recorded. Then average percentage of pairs from the same data set is computed as the accuracy of the corresponding data set. In each independent experiment, different inference approaches are applied on 90 points randomly sampled from raw data without replacement.
The results are summarized in Fig. \ref{fig:real_pairs} with blue solid line indicating median accuracy and red dashed line indicating mean accuracy. It shows that the performance of ANM-MM is satisfactory, with highest median accuracy of about 82\%. IGCI also performs quite well, especially in terms of median, followed by PNL.
\begin{figure}[t]
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{apdx/real_pair_gt.png}
\caption{Ground truth}
\label{Fig:real_1_1}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{apdx/real_pair_gp.png}
\caption{ANM-MM}
\label{Fig:real_1_2}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{apdx/real_pair_km.png}
\caption{$k$-means}
\label{Fig:real_1_3}
\end{subfigure} \\
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{apdx/real_pair_gmm.png}
\caption{GMM}
\label{Fig:real_2_1}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{apdx/real_pair_spec.png}
\caption{SpeClu}
\label{Fig:real_2_2}
\end{subfigure}
\begin{subfigure}{.32\linewidth}
\centering
\includegraphics[width=\linewidth]{apdx/real_pair_db.png}
\caption{DBSCAN}
\label{Fig:real_2_3}
\end{subfigure}
\caption{Ground truth and clustering results of different approaches on BAFU air data.}
\label{fig:real_clu1}
\end{figure}
\textbf{Clustering on BAFU air data.} We evaluate the clustering performance of ANM-MM on real air data obtained online\footnote{https://www.bafu.admin.ch/bafu/en/home/topics/air.html}. This data consists of daily mean values of ozone ($\mu g / m^{3}$) and temperature ($^\circ$) of 2009 from two distinct locations in Switzerland. In our experiment, we regard the data as generating from two mechanisms (each corresponds to a location). The clustering results are visualized in Fig. \ref{fig:real_clu1}. The ARI values of ANM-MM is 0.503, whereas $k$-means, GMM, spectral clustering and DBSCAN could only obtain ARI of -0.001, 0.003, 0.078 and 0.003, respectively. ANM-MM is the only one that could reveal the property related to the location of the data g.m..
\section{Conclusion}
In this paper, we extend the ANM to a more general model (ANM-MM) in which there are a finite number of ANMs of the same function form and differ only in parameter values. The condition of identifiability of ANM-MM is analyzed. To estimate ANM-MM, we adopt the GP-LVM framework and propose a variant of it called GPPOM to find the optimized latent representations and further conduct causal inference and mechanism clustering. Results on both synthetic and real world data verify the effectiveness of our proposed approach.
\subsubsection*{Acknowledgments}
This work is partially supported by the Hong Kong Research Grants Council.
\bibliographystyle{apalike}
|
\section{Introduction}
\label{sectionintro}
\begin{quote}
``You have fifteen seconds. Using standard math notation, English words, or both, name a single whole number--not an infinity--on a blank index card. Be precise enough for any reasonable modern mathematician to determine exactly what number you've named, by consulting only your card and, if necessary, the published literature. Are you ready? Get set. Go.'' -- Scott Aaronson, ``Who can Name the Bigger Number?'' \footnote{\url{http://www.scottaaronson.com/writings/bignumbers.html}}\\~\\
``The game of describing the largest integer, when played by experts, lapses into a hopeless argument over legitimacy.'' -- Joel Spencer, ``Large numbers and unprovable theorems'' \cite{Spencer}
\end{quote}~
This paper addresses the question: what is the best strategy in a ``name the largest number'' contest such as the one described above? Of course, the answer depends on what the rules of the contest are. Scott Aaronson's idea of letting a ``reasonable modern mathematician'' judge the entries might sound fair at first, but the answer to the question of whether a given entry is a precise description of a number may depend on exactly which mathematician you ask. It turns out that this dependence is not incidental but fundamental, and slightly different standards for what counts as a valid entry give rise to radically different strategies for winning the game.
One way to make precise the notion of a ``precise description of a number'' is to fix a \emph{precise language}, i.e. a language whose syntax and semantics are both precisely defined, for describing numbers.\Footnote{I\ warn the reader that the phrase `precise language' is not standard terminology. While I\ will attempt to use standard terminology wherever possible, in some cases this is not possible; for instance, the reason I\ use the phrase `precise language' is that the more standard phrase `formal language' tends to refer only to syntax and not to semantics. The concept of a formal language is amenable to mathematical definition, whereas the concept of a precise language is not unless it is restricted in some way.}
A major goal of mathematics in the nineteenth and early twentieth centuries was to create precise languages capable of expressing mathematical thought. According to one view, which I will call the standard mathematicians' view, this goal was accomplished by the introduction in the early twentieth century of the language of (first-order,\Footnote{In this paper I\ may seem to have an unjustified preference for first-order languages, as opposed to things like second-order languages and languages that allow plural quantification. However, any second-order language can be interpreted as a first-order language by the expedient of viewing the ``properties'' in the original language as a type of object that one is allowed to quantify over. Although statements involving plural quantification are more difficult to straightforwardly translate into first-order form, it seems to me that in all cases where the original sentence is not ambiguous, it is possible to do so, and that moreover the issues that concern philosophers regarding plural quantification have little do to with the issues considered in this paper.} classical,\Footnote{Here `classical' means ``based on classical logic'' -- in particular, every statement in a classical language is either true or false.} Zermelo--Fraenkel) \emph{set theory}, which is allegedly capable of codifying all mathematical facts. Another important language developed in the late nineteenth century is (first-order, classical, Peano) \emph{arithmetic} or \emph{number theory}. Roughly speaking, set theory is the language in which the fundamental objects of discourse are sets and the fundamental relation is membership, and number theory is the language in which the fundamental objects of discourse are the natural numbers $0,1,2,3,\ldots$ (which we will hereafter refer to simply as ``numbers'') and the fundamental operations are addition and multiplication. The syntaxes of both set theory and number theory have precise definitions, which are given in any good textbook on mathematical logic. However, it is open to philosophical debate whether the languages can be assigned precise semantics; see Sections \ref{subsectionplan}, \ref{sectionnumberrealism}, and \ref{sectionsetrealism} for further discussion. Here I\ will attempt to take a neutral position with respect to this debate, and merely analyze the consequences of any particular set of beliefs regarding the preciseness of mathematical languages. In what follows, I\ will denote the languages of set theory and number theory by the labels \textbf{\textup{ST}}\ and \textbf{\textup{NT}}, respectively.\Footnote{The languages of set theory and number theory are sometimes denoted by the labels \textbf{\textup{ZFC}}\ and \textbf{\textup{PA}}, respectively, as shorthand for `Zermelo--Fraenkel set theory with Choice' and `Peano Arithmetic'. These labels are also used to denote standard axiom systems associated with these languages. However, in order to emphasize the distinction between a language and an axiom system associated with that language, I\ reserve the labels \textbf{\textup{ZFC}}\ and \textbf{\textup{PA}}\ for the axiom systems and I\ use the alternate labels \textbf{\textup{ST}}\ and \textbf{\textup{NT}}\ for the languages. I\ chose new labels not based on mathematicians' names in order to suggest that the languages are somehow less ``historically contingent'' than the axiom systems, which I\ believe to be the case.} It is well-known that \textbf{\textup{NT}}\ has significantly less expressive power than \textbf{\textup{ST}}\ in the sense that there is an algorithm for translating sentences\ of \textbf{\textup{NT}}\ into \textbf{\textup{ST}}\ (which claims to preserve meaning or at least truth-values), but not vice-versa.
If we have in mind a precise language, then we can define a precise largest number contest by declaring that an entry to the contest is valid if and only if it is a term\Footnote{The word `term' is a technical term in mathematical logic whose closest non-technical analogue is something like ``a singular noun phrase that includes a definite article''. The key thing about terms is that each term is intended to refer to a unique object.} of the language that names a number. We can write this in semi-formal mathematical language as follows:
\begin{lncs}
\label{lnc1}
Let $L$ be a precise language. The \emph{largest number contest of $L$}, denoted $\textbf{\textup{LNC}}(L)$ is the contest in which a valid entry is a term of $L$ that succeeds at referring to a number, and the score of the entry is the number it refers to. (In all of our contests, the winner is the valid entry with the highest score.) We will denote the score of an entry $\phi$ by $[\phi]$.
\end{lncs}
To many modern mathematicians, the largest number contest of \textbf{\textup{ST}}\ will seem like the obvious ``first choice'' for a largest number contest. On the other hand, many people are uncertain as to whether set theory is really a precise language, but are more confident in number theory; the largest number contest of \textbf{\textup{NT}}\ may be a more natural ``first choice'' for such people.\Footnote{There is the minor technical issue that the terms of these languages have much less expressive power than the languages themselves; this can be remedied by adding new rules for constructing terms that bring the language's terms on a par with the rest of the language in terms of expressive power, without increasing the expressive power of the language as a whole. See Appendix \ref{appendixPT} for details.} Each of these contests has the disadvantage that it is not possible to determine whether any given entry is valid in an algorithmic way. In fact, due to Tarski's theorem on the undefinability of truth (see Section \ref{sectionphilosophical}) it is not even possible to formalize the notion of a ``valid entry'' in the precise language being used for the contest (namely \textbf{\textup{ST}}\ or \textbf{\textup{NT}}). Despite this lack of formalizability, it appears that the question of which entries are valid is still legitimate and precise, assuming that the language used to define the contest has in fact been given precise semantics.\Footnote{Some mathematicians trained under the set theory framework may object that any question that cannot be translated into \textbf{\textup{ST}}\ should be suspected of being imprecise. However, this view is subject to the counterargument that any reason to believe that \textbf{\textup{ST}}\ is a precise language is also a reason to believe that the question of which entries in \textbf{\textup{LNC}}(\textbf{\textup{ST}}) are valid is precise; on the other hand, if \textbf{\textup{ST}}\ is not precise, then it is not clear why one would use it as a standard of precision. Some further discussion may be found in Section \ref{sectionsetrealism}.} So it makes sense to ask what the best strategy to use is, and in fact it is not too hard to find a strategy for the contest which is close to being optimal. This strategy is based on the following theorem:\Footnote{The proofs of all theorems and corollaries, as well as the explanation of the axiom system they are understood to be theorems of, are given in Appendices \ref{appendixPT}-\ref{appendixproofs}, where more general versions of the theorems are also stated.}
\begin{theorem}[Cf. Theorem \ref{theoremrecstratB}]
\label{theoremrecstrat}
Let $L$ be either set theory or number theory. Then for each number $n$, the phrase `the largest number that can be represented in $L$ by a term consisting of $n$ symbols or fewer'\Footnote{Cf. \cite[Figs. 9-10]{Spencer}, \url{http://web.mit.edu/arayo/www/bignums.html}} can be translated into $L$ as a term consisting of $O(n)$ symbols.\Footnote{Recall that $O(n)$ is shorthand for `at most a constant times $n$'.} This term, which we will denote by $\philang{n}$, can be written explicitly.
\end{theorem}
Our strategy for the largest number contest of $L$ can now be given as follows (assuming that $L$ is precise): first compute the largest number $n$ such that the term $\philang{n}$ can fit onto the ``index card'' Scott Aaronson gives us to write our entry on (small handwriting will come in handy here), and then write $\philang{n}$ on the card.
This strategy, which we will call the \emph{brute force strategy} of $L$, is not necessarily optimal, but it is \emph{nearly optimal} in the sense that no strategy can beat it too badly. To be precise, by the definition of $\philang{n}$, any entry that scores higher than $\philang{n}$ necessarily uses at least $n+1$ symbols to do so, which is not too much less than the $O(n)$ symbols needed by the brute force strategy. So if there is a better strategy, the difference is not overwhelming in the sense that if one player plays with the brute force strategy and another player plays with a better strategy, who wins may be determined more by who has more resources to express their answer (for example, small handwriting) than it is by who is using the better strategy. (This is the reason for the name ``brute force strategy''.) The question of who wins also starts to depend on the specific details of the formal definition of $L$; for example, augmenting \textbf{\textup{ST}}\ with the primitive ability to write `$x\notin y$' (``the set $x$ is not a member of the set $y$'') as shorthand for `$\neg x\in y$' (``it is not true that $x$ is a member of $y$''), thus saving a symbol, may make a crucial difference to the outcome. I\ will take the point of view that any improvement that depends on such details is not a significant improvement, and thus I\ will consider the ``near-optimality'' possessed by the brute force strategy to be essentially the same as optimality. (This is a standard practice in many areas of mathematics and computer science.)
An obvious objection is that the largest number contest of $L$, though precise, does not stay true to the ``spirit'' of being a ``name the largest number'' contest. For example, a description of the term $\philang{n}$ in terms of $n$ could perhaps be made to fit on an index card,\Footnote{If this paper counts as ``published literature'' then one could simply write `$\philang{10^{100}}$' together with the appropriate citation.} and then one could simply write `and my number is the number denoted by the term $\philang{10^{100}}$'.\Footnote{Alternatively, one could simply write the original English phrase `the largest number that can be represented in $L$ by a term consisting of $10^{100}$ symbols or fewer' on the index card, as is done for example in \cite{Spencer}. However, the point is that the fact that Theorem \ref{theoremrecstrat} specifies a procedure for (eventually) translating this phrase into $L$ might be thought of as adding some ``legitimacy'' to the indirect strategy, as described in the next paragraph.} We will call this the \emph{indirect strategy}, and we note that the entry it produces is not a valid entry to $\textbf{\textup{LNC}}(L)$, since it is an English description of a number rather than a term of the language $L$. Since $10^{100}$ is much larger than the number of symbols that any physically realistic entry could have, in a largest number contest in which the entry produced by the indirect strategy is considered valid, it wins against any entry which is valid in $\textbf{\textup{LNC}}(L)$. However, it seems that the indirect strategy should be allowed if we follow Aaronson's principle that a player only needs to submit an entry that is ``precise enough for any reasonable modern mathematician to determine exactly what number [the player has] named'' (assuming, of course, that $L$ is a precise language to begin with).
The objection becomes more poignant if we consider the fact that the hypothetical entry of the previous paragraph is not only perfectly precise, but also has the property that it can be converted into a valid entry to $\textbf{\textup{LNC}}(L)$ given a sufficient (albeit physically unrealistically large) amount of time and effort. Concentrating on this part of the objection, we can imagine a second contest in which the entries are not required to be terms of the language $L$, but are allowed to be merely \emph{descriptions} of terms. The crucial question now becomes what sorts of descriptions of these terms are allowed. One possibility would be to allow these descriptions to be themselves terms of the language $L$, or perhaps terms of a different precise language that has the ability to talk about strings of symbols.\Footnote{It is well-known that the theory of strings of symbols can be translated into both \textbf{\textup{ST}}\ and \textbf{\textup{NT}}.} However, it is not hard to see that the ``brute force strategy'' described above can be trivially modified to deal with this new game: instead of writing $\philang{n}$, write `$\philang{m}$, where $m =$'$\philang{n}$ for an appropriate value of $n$. Things get more interesting if we look at other ways of describing strings. For now at least, let us take the conservative viewpoint that in order to count for the purposes of the contest, a description of a string should contain enough information that it is possible to use it to algorithmically generate the string it describes, although this may take an arbitrarily long time. Semi-formally:
\begin{lncs}
\label{lnc1.1}
Let $L$ be a precise language. The \emph{largest number contest of $L$ allowing indirect entries}, denoted $\textbf{\textup{LNC}}^*(L)$, is the contest in which a valid entry is an algorithm (written in some fixed programming language the details of which are irrelevant\Footnote{Note that a programming language is a particular type of precise language: a language for precisely specifying mathematical algorithms. In this paper a \emph{program} is understood to be a description of an algorithm in a programming language.}) that succeeds at (eventually) computing a term of $L$ that refers to a number. If the algorithm fails to terminate by returning a string as its output, or if the string it returns is not a term of $L$ referring to a number, then the entry is considered invalid. The score of a valid entry is the number that its output refers to.
\end{lncs}
Taking our cue from the strategy that ``won'' $\textbf{\textup{LNC}}(L)$, a natural strategy for $\textbf{\textup{LNC}}^*(L)$ is to first find an algorithm that computes a very large number $n$, and then to combine it with a second algorithm that computes the string $\philang{n}$ given $n$ as input. However, this leaves open the question of how to write an algorithm that computes a very large number. It is worth listing this as a separate largest number contest:
\begin{lnc}
\label{lnc2}
The \emph{Busy Beaver contest}\Footnote{The Busy Beaver contest is inspired by the \emph{Busy Beaver function}. Given a number $n$ as input, the Busy Beaver function returns the largest number that is the output of some algorithm that can be written in at most $n$ symbols (in some fixed programming language). Equivalently, the value of the Busy Beaver function at $n$ is the largest number computed by any possible entry to the Busy Beaver contest written in at most $n$ symbols. (Technical note: The classical Busy Beaver function is defined not in terms of the number of symbols in a description of an algorithm, but rather in terms of the \emph{number of states} in a (one-tape, two-symbol) \emph{Turing machine representation} of the algorithm. However, the length of a program encoding an algorithm is a much more natural metric of the algorithm's complexity than the number of states in a Turing machine representation, and I\ have no qualms with saying that the ``true'' Busy Beaver function is the one described above.)} is the contest in which a valid entry is an algorithm that succeeds at computing a number, and the score of the entry is its output.
\end{lnc}
It is clear that any strategy for the Busy Beaver contest whose score is $n$ can be converted into a strategy for $\textbf{\textup{LNC}}^*(L)$ whose score is $[\philang{n}]$, and that if a strategy is optimal or nearly optimal for the Busy Beaver contest, then its converted version is nearly optimal for $\textbf{\textup{LNC}}^*(L)$. Thus, we can avoid dealing with $\textbf{\textup{LNC}}^*(L)$ directly and instead analyze the Busy Beaver contest in its place.
\begin{remark*}
Any strategy for the Busy Beaver contest can also be converted into a strategy for $\textbf{\textup{LNC}}(\textbf{\textup{NT}})$ or $\textbf{\textup{LNC}}^*(\textbf{\textup{ST}})$ with the same score as the original strategy, via the fact that phrases of the form `the output of the algorithm $\alpha$, if it exists' can be translated into \textbf{\textup{NT}}\ and \textbf{\textup{ST}}. However, \textbf{\textup{NT}}\ and \textbf{\textup{ST}}\ are capable of naming much larger numbers, and scores in the Busy Beaver contest will in general be much lower than scores in the largest number contests of \textbf{\textup{NT}}\ and \textbf{\textup{ST}}.
\end{remark*}
It turns out that the best strategy for the Busy Beaver contest is very different from the best strategy for $\textbf{\textup{LNC}}(\textbf{\textup{ST}})$ or $\textbf{\textup{LNC}}(\textbf{\textup{NT}})$. To describe this strategy, we need the notion of an \emph{axiom system} for a precise language. An axiom system $A$ for a precise language $L$, also known as an \emph{axiomatization} of $L$, is a list of \emph{axioms} and \emph{inference rules}. An axiom is a sentence\ of $L$ which is either considered obviously true, or merely hypothesized to be true,\Footnote{A major difference between the mathematical and philosophical communities is that mathematicians generally view axioms as ``hypotheses'' while philosophers usually view axioms as ``obvious or self-evident truths''. Furthermore, while some mathematical hypotheses are considered likely to be true, others are introduced purely for the sake of argument.} while an inference rule is a syntactic rule for creating new sentences\ of $L$ from previous ones, with the understanding that the inference rule is supposed to formalize some method of reasoning, in the sense that the previous sentences\ represent ``prior beliefs'' and the new sentences\ represent ``deductions'' or ``conclusions formed using the reasoning method''. A list of claims of $L$ is said to be a \emph{syntactic proof in $A$} if each claim on the list is either an axiom or can be created from the previous ones using some inference rule. We use the phrase `syntactic proof' to distinguish between a syntactic proof and a \emph{semantic proof} of a claim, which is an argument that convinces us, or would convince us, that the claim is true. An axiom system aspires for its syntactic proofs to be also semantic proofs; when this is the case we call the axiom system \emph{valid}. The standard axiomatizations of \textbf{\textup{ST}}\ and \textbf{\textup{NT}}\ are the Zermelo--Fraenkel axiom system for set theory and the Peano axiom system for number theory, which we will denote by \textbf{\textup{ZFC}}\ and \textbf{\textup{PA}}, respectively (for ``Zermelo--Fraenkel set theory with Choice'' and ``Peano Arithmetic'', respectively)
In what follows we consider a pair $(L,A)$ where $L$ is a precise language and $A$ is a valid axiom system for $L$. In order to prove results about $(L,A)$ we must make some additional assumptions about the structure of $L$; specifically, we will assume that $L$ is a \emph{proposition/term language} as defined in Appendix \ref{appendixPT}. Note that both \textbf{\textup{ST}}\ and \textbf{\textup{NT}}\ are proposition/term languages in this sense.
\begin{theorem}[Cf. Theorem \ref{theoremaxiomstratB}]
\label{theoremaxiomstrat}
Let $A$ be a valid axiomatization of a precise proposition/term language $L$ such that all sentences\ of the form `the algorithm $\alpha$ eventually halts', where $\alpha$ is a concrete algorithm, can be translated into $L$. Then for each $n$, there is a program of length $O(\log(n))$ encoding an algorithm that succeeds at computing $\fA{n}$, the largest number which is the output of an algorithm that can be proven to halt via a proof in $A$ of length at most $n$. This program, which we will denote by $\alphaA{n}$, can be written explicitly.
\end{theorem}
\begin{remark*}
Note that the length of the program $\alphaA{n}$ is only $O(\log(n))$ rather than $O(n)$. The reason for this is that $\alphaA{n}$ only needs to describe $n$ in sufficient detail that it can be computed (for example by giving its binary or decimal representation), rather than needing to contain $n$ symbols to start with.
\end{remark*}
Our strategy for the Busy Beaver contest can now be given succinctly as follows: submit the program $\alphaA{10^{100}}$. This is feasible because $\log\double{\psi_{L,A}}{n}$ is small enough that a program of length $O(\log(10^{100}))$ is ordinarily short enough that it can be physically implemented.\Footnote{Of course, this depends on the constant implied by the $O(\cdot)$ notation, but it appears not to be too large.} Although theoretically one may submit the program $\alphaA{n}$ for much larger values of $n$, it turns out that it is not worth it to do so because gains from this sort of change are overwhelmed by much larger easy improvements elsewhere, such as replacing $(L,A)$ by the metasystem $(L+1,A+1)$ described below.
Unlike the brute force strategy, the above strategy, which we will call the \emph{axiomatic strategy} of $A$, is not nearly optimal. In fact, there is no humanly understandable nearly optimal strategy, since any nearly optimal entry will appear to us as a nearly random sequence of symbols; see Appendix \ref{appendixmoreBBC} for details. However, if the language $L$ and the axiom system $A$ are chosen wisely, the strategy can have a different type of optimality, namely \emph{relative epistemological optimality}. The idea is that even though there may be many possible valid entries that would win against the axiomatic strategy if they were submitted as competitors, humanity may never be able to determine which of these entries are in fact valid. So if you are playing against human opponents that prefer to only submit entries that they believe are valid, then with the axiomatic strategy you may have a reasonable shot at winning.
Suppose that Alice and Bob are human players, and that Alice has just submitted an entry $\alphaA{10^{100}}$ using the axiomatic strategy, while Bob is going to submit an entry $\alpha_0$ that he believes is valid. Presumably, this means that Bob has executed (either internally or externally\Footnote{An example of external execution of reasoning is a computer-assisted proof. Alternatively, we could say that a human who comes to believe the conclusion of a computer-assisted proof is applying internal reasoning using the belief that the external world (and in particular the computer performing the calculations used in the proof) operates according to the laws of physics. However, according to this second way of looking at things the chain of reasoning cannot properly be described as mathematical or strictly deductive. Thus the second way of looking at things makes the situation more difficult to analyze, which is why I\ stick to considering a computed-assisted proof as external reasoning.}) some chain of reasoning whose conclusion is that $\alpha_0$ is valid. Although it is theoretically possible that this chain of reasoning cannot be formalized in any axiom system, it seems plausible that in practice it usually can be, and that in fact there is usually an axiom system capable of formalizing Bob's reasoning which is not too much more powerful than an axiom system that Bob explicitly endorses. Moreover, arguments that can be formalized in one axiom system also tend to be able to be formalized in more powerful axiom systems, so as long as the axiom system that Bob's reasoning can be formalized in is less powerful than Alice's preselected axiom system $A$, then Bob's reasoning can probably also be formalized in $A$ as well. Finally, if we formalize Bob's reasoning in the axiom system $A$, then we may expect that the formalization of Bob's reasoning will not be too much longer than a formalization
of his reasoning in English (or another natural language); all that is needed is that the ratio between the two lengths will not be more than (for example) $10^{50}$. On the other hand, due to physical limitations that apply regardless of whether Bob's reasoning is internal or external,
it is reasonable to expect that we can formalize his reasoning in English using no more than (for example) $10^{50}$ symbols, and thus that we can formalize his reasoning in the axiom system $A$ using no more than $10^{100}$ symbols. If this turns out to be the case, then $\alpha_0$ can be proven to halt in $A$ using reasoning encoded in no more than $10^{100}$ symbols. Thus the output of $\alpha_0$ is at most $\fA{10^{100}}$, which is the output of Alice's entry $\alphaA{10^{100}}$. So Alice will win.\Footnote{The philosophical argument given in this paragraph is significant not for its rigor, which is practically non-existent, but because it seems to mesh well with what we expect. It seems to me\ that if largest number contests were actually held, many of them would work in exactly this way.} \Footnote{It is also interesting to ask what might happen if Bob bases his beliefs on probabilistic or heuristic reasoning. For example, many of the arguments for accepting the validity of strong axiomatizations of set theory (i.e. large cardinal axioms) are based on such reasoning. Recently there has been much interest in the difficult question of how to formalize probabilistic reasoning about mathematical concepts (see e.g. \cite{GBCST} and the references therein), but it is not clear to me that any of the frameworks proposed so far are capable of modeling (for example) the standard heuristic arguments in favor of measurable cardinals (see \cite{Maddy1,Maddy2}). Thus, if Bob bases his belief that his entry is valid on probabilistic reasoning, it may be necessary to find an ``intermediate axiom'' which Bob would also accept on the basis of the same probabilistic reasoning, which implies that Bob's entry is valid (in the axiomatic framework Bob already accepts), and which is a consequence of the axiom system $A$ chosen by Alice, in order for the argument to apply.}
In order to make the analysis fair, I\ should also address the question of why Alice might believe that her entry is valid. Its validity cannot be proven in the axiom system $A$ in a physically realistic timeframe, since if it could then Alice would win if she played against herself, which makes no sense. However, it turns out that there is a precise language and corresponding axiom system not too much more powerful than the pair $(L,A)$ capable of formalizing the reasoning used in the proof of Theorem \ref{theoremaxiomstrat}: namely, the pair $(L+1,A+1)$, where
\begin{itemize}
\item $L+1$ is the \emph{metalanguage} of $L$, a precise proposition/term language capable of expressing the notions of truth and reference with respect to $L$ in the sense that claims of the form `$\phi$ is a true sentence\ of $L$' and terms of the form `the object denoted by the $L$-term $t$' can be straightforwardly translated into $L+1$; and
\item $A+1$ is the \emph{metasystem} of $A$, a natural axiomatization of $L+1$ which formalizes the assumption that the axiom system $A$ is valid as well as standard beliefs and methods of reasoning about strings of symbols.
\end{itemize}
(See Appendices \ref{appendixPT} and \ref{appendixaxiomatizations} for more details.) Thus, Alice may come to believe that her entry is valid using reasoning that can be formalized in the axiom system $A+1$.
In what follows, we will metonymically refer to a pair $(L,A)$, where $L$ is a precise language and $A$ is an axiomatization of it, as an axiom system, and to the pair $(L+1,A+1)$ as the metasystem of $(L,A)$.
\begin{corollary}[Cf. Corollary \ref{corollaryaxiomstratB}]
\label{corollaryaxiomstrat}
Let $(L_1,A_1)$ and $(L_2,A_2)$ be valid axiom systems with the translation property hypothesized in Theorem \ref{theoremaxiomstrat}, and suppose that $(L_1,A_1)$ is at least as powerful as the metasystem $(L_2+1,A_2+1)$ in the sense that all proofs of $(L_2+1,A_2+1)$ can be translated into $(L_1,A_1)$. Then in the Busy Beaver contest, the axiomatic strategy of $(L_1,A_1)$ wins agains the axiomatic strategy of $(L_2,A_2)$ (assuming both axiom systems as well as the algorithm for translating one into the other are small enough that they could be physically implemented).
\end{corollary}
Consequently, the key question to ask about the contestants to the Busy Beaver contest is: what is the weakest axiom system capable of formalizing their reasoning? If Alice uses reasoning incapable of being formalized in an axiom system $A$ capable of formalizing Bob's reasoning, this may lead her to believe that the axiom system $A$ is valid, in which case she can win simply by playing its axiomatic strategy. If both players' logic is best described in terms of the same axiomatic framework, then we are led to the consideration of another type of largest number contest.
A common defect in all of the largest number contests considered above is that there is no clear way for human judges to determine whether any given entry is valid or not. Even if a term $\phi$ refers to a number (and is therefore a valid entry to the first contest), there may be no proof that $\phi$ refers to a number in any axiom system that the judges accept as valid. However, the optimal strategies for the largest number contests of \textbf{\textup{ST}}\ and \textbf{\textup{NT}}, i.e. the brute force strategies of \textbf{\textup{ST}}\ and \textbf{\textup{NT}}, are relatively uncontroversial (since the proof of Theorem \ref{theoremrecstrat} may be considered valid by the judges) so the issue does not matter too much for those contests. In the Busy Beaver contest the situation is different. Theoretically a player could run the algorithm described in his\ entry to show that it halts, but by design this would take an extremely large amount of time. The contest may be modified by adding the requirement that \emph{contestants must submit, as an addendum to their entry, a proof that their entry is valid}. Then whatever reasoning convinces the contestants of the validity of their entry could also be shown to the judges
This new requirement brings us back around almost full circle: like the question of what constitutes a precise description of a number, the question of what constitutes a valid proof that an entry is valid is subject to philosophical debate, unless we describe a precise criterion, such as validity in a fixed axiom system, for what counts as a proof. But if we introduce such a structure, then as before it turns out that the best strategy is a brute force strategy:
\begin{lncs}
\label{lnc3}
Let $(L,A)$ be as in Theorem \ref{theoremaxiomstrat}. The \emph{Busy Beaver proof contest of $(L,A)$} is the contest in which a valid entry is a pair of the form $(\alpha,\rho)$, where $\alpha$ is an algorithm and $\rho$ is a proof in $A$ that $\alpha$ succeeds at computing a number, and the score of the entry $(\alpha,\rho)$ is the output of $\alpha$.
\end{lncs}
\begin{theorem}[Cf. Theorem \ref{theoremrecstrat2B}]
\label{theoremrecstrat2}
Let $L$ be either set theory or number theory, and correspondingly let $A$ be either \textbf{\textup{ZFC}}\ or \textbf{\textup{PA}}.\Footnote{The proof of this theorem is also valid for a much broader class of languages and axiom systems; see Appendices \ref{appendixPT}-\ref{appendixproofs} for details.} Then for each $n$, the $n$th case of Theorem \ref{theoremaxiomstrat} can be (syntactically\Footnote{If the axiom system $(L,A)$ is valid, then this syntactic proof is also a semantic proof.}) proven in $A$ using reasoning encoded in at most $O(n^2\log(n))$ characters.\Footnote{This number is possibly somewhat sensitive to the exact definition of a ``proof in the axiom system $A$''. Though many definitions of the word ``proof'' exist in the literature, this theorem is based on the definition of ``proof'' described in Appendix \ref{appendixaxiomatizations}, which is intended to closely model the notion of ``proof'' in standard mathematical practice, including automated proof checkers.} The proof in $A$, which we will denote by $\double{\rho_A}{n}$, can be written explicitly.
\end{theorem}
The brute force strategy for the Busy Beaver proof contest of $(L,A)$ can then be given as follows: First compute the largest number $n$ such that the $\double{\rho_A}{n}$ is short enough to be submitted as an addendum to the entry $\alphaA{n}$, and then submit the pair $(\alphaA{n},\double{\rho_A}{n})$ as an entry. (Note that the length of $\double{\rho_A}{n}$ is the limiting factor here; $\alphaA{n}$ is much shorter.) Like the brute force strategies of the largest number contests of \textbf{\textup{ST}}\ and \textbf{\textup{NT}}, this strategy is nearly optimal in the sense that no other valid strategy can beat it too badly. However, in this case the near-optimality is a little bit worse in the sense that the number of symbols needed is $O(n^2\log(n))$ rather than $O(n)$. It turns out that if we consider proofs that incorporate ``shorthand'' corresponding to the common mathematical practice of giving ``definitions'', then the number of symbols needed can be reduced to $O(n\log(n))$.
Like the largest number contests of \textbf{\textup{ST}}\ and \textbf{\textup{NT}}, an objection can be raised to the various Busy Beaver proof contests on the grounds that they are not ``real'' largest number contests: for any such contest, it is easy to describe entries that ``should be'' accepted as valid since their corresponding proofs use methods of reasoning just a little bit outside of the axiomatic framework that has been fixed for the judging, which are easily seen to win against all valid entries by a large margin. However, we seem to have exhausted all our ``possible moves'' for getting away from the problem (though we will see later that this is not quite true), so perhaps it is best to just admit defeat: there is no good way to formalize the largest number contest.
To be specific, an important thesis of this paper is the following:
\begin{thesis}
\label{mainthesis}
For any precise largest number contest, the best strategy is either a brute force strategy or an axiomatic strategy.
\end{thesis}
I\ will not argue for this thesis other than by pointing to the lack of counterexamples despite encountering a broad range of precise largest number contests. The real test will be to see whether a counterexample will be found in the future.
In contests where the optimal strategy is a brute force strategy, there is always the feeling that the spirit of the contest has been lost since it is easy to give descriptions of numbers that are just as precise as the implicit descriptions of numbers provided by valid entries in the contest, that nevertheless win against all legitimate entries. Moreover, the fact that victory in the contest is tied to merely being able to physically write down a certain sequence of symbols, when we all know in advance exactly what those symbols will be, makes a mockery of what was supposed to be a test of wits. On the other hand, in contests where the best strategy is axiomatic, it is impossible to determine which entries are valid in an algorithmic way, and this problem is most conspicuous exactly when we look at the strategies that appear to have the best chance of winning, namely the axiomatic strategies of the most powerful axiom systems for the most expressive languages. In general the plausibility of an axiom system is inversely proportional to its power, so this leads us to the uncomfortable situation where the strategies which are the best if valid are also those which are the least likely to be valid, and the most difficult to justify.\\
{\bf Acknowledgements.} The author thanks Lance and Joseph Simmons for comments on an earlier version of this paper, and Scott Aaronson for writing about the question from which this paper arose. The author was supported in part by EPSRC Programme Grant EP/J018260/1.
\tableofcontents
\section{Philosophical largest number contests}
\label{sectionphilosophical}
One reaction to the difficulties of formalizing the largest number problem is to eschew formalization and instead consider a different, more ``philosophical'', approach to the challenge of implementing a largest number contest. Namely, we can take literally the idea of judging the entries according to the standard of whether a ``reasonable modern mathematician'' would find them to be precise. The fact that different mathematicians may give different answers to this question can be dealt with by specifying in advance who will judge the contest. Alternatively, we may divide mathematicians into broad classes based on their philosophy of mathematics, and allow each philosophical school to elect a panel of judges to represent it. Each entry could then spark a philosophical debate within the panel as to whether or not it is valid. Semi-formally:
\begin{lncs}
\label{lnc4}
Let $X$ be a school of thought in the philosophy of mathematics. Then the \emph{largest number contest run by adherents of $X$} is the contest in which an entry consists of a natural language description of a number together with a philosophical argument defending the premise that the description in fact uniquely identifies a number. The entry is valid if the adherents of $X$ agree with the argument, and the winner is whichever entry the adherents of $X$ believe describes the largest number.
\end{lncs}
\begin{remark*}
In practice, the matter of deciding which entry describes the largest number is much more straightforward than the matter of deciding which entries are valid.
\end{remark*}
Obviously, largest number contests run by philosophers are not precise in the same way that the contests of the previous section are. Thus, instead of immediately describing what the optimal strategy is, I\ will discuss possible strategies for this contest for various schools of thought over the next few sections. In particular, I\ will argue that in most cases Thesis \ref{mainthesis} applies to this contest as well, in the following sense: With enough ``nudging'', it is possible to convince the adherents of any school of thought that their largest number contest can be reduced to a largest number contest in which either a brute force strategy or an axiomatic strategy is optimal. The idea is that there are only a few ``free variables'' in the interpretation of a philosophy that need to be pinned down before its largest number contest becomes a precise largest number contest.
Actually, most philosophers will find the idea that a philosophical largest number contest can be won with a brute force strategy to be abhorrent, since it appears to mean that our powers of description are dependent on our ability to physically instantiate large descriptions of numbers, rather than merely to precisely describe such descriptions. This means that in practice philosophical largest number contests are usually won by axiomatic strategies. Of course, an axiom system is also the sort of thing that it is possible to give philosophical arguments for and against. This leads to the following family of contests:
\begin{lncs}
\label{lnc5}
Let $X$ be a school of thought in the philosophy of mathematics, and let $L$ be a language that adherents of $X$ believe is precise. Then the \emph{strongest axiom contest for $L$ run by adherents of $X$} is the contest in which an entry consists of an axiomatization of $L$ (given either informally or as a list of axioms and inference rules) together with a philosophical argument defending the premise that the axiom system is valid. The entry is valid if the adherents of $X$ agree with the argument, and the winner is whichever entry the adherents of $X$ believe describes the most powerful axiom system.
\end{lncs}
Again, I\ will describe possible strategies for various strongest axiom contests over the next few sections. Note that the largest number and strongest axiom contests will have to be treated as different contests at first, even though I\ will argue that usually the largest number contest reduces to the strongest axiom contest.
Without any information about what school of thought is running the contest, it is not possible to be very precise about what constitutes a good strategy for either of these types of contests. Nevertheless, it turns out that there is a broad \emph{class} of strategies that tend to be good strategies for the first type of contest. Namely, one type of possible entry to a philosophical largest number contest is a description of a (hopefully) precise language $L$ for naming numbers together with a term of $L$ which is supposed to refer to a number. Once the language $L$ is fixed, then the philosophical largest number contest reduces to the largest number contest of $L$, a precise largest number contest. However, the appropriate strategic response to this depends on what kind of language $L$ is. In Appendix \ref{appendixPT} I\ will introduce a class of languages called \emph{proposition/term languages}, or \emph{PT languages}, which have a regular structure designed to be compatible with the ideas in this paper. This structure is somewhat different from, and more general than, the usual structure of the languages considered in mathematical logic. It appears to more closely resemble the structure of natural language.
\subsection{Classical languages}
For the purposes of this paper, a \emph{classical} language is a proposition/term language that is assumed to satisfy\Footnote{Here, I\ use the phrase `assumed to satisfy' to mark assumptions made by the users of a language without passing judgement on whether those assumptions are accurate or not. In general, those who disagree with the implicit premises of a language will not be considered to be valid users of the language, though they can still talk about the language.} the \emph{law of excluded middle}, which essentially claims that every sentence\ of the language is either true or false (more precisely, every sentence\ whose terms all have referents). We will call languages that in fact satisfy the law of excluded middle \emph{definite languages}; thus, a classical language is one which speakers of the language treat as definite. The notion of definiteness is closely related to the concept of preciseness, but in what follows I\ will argue that it is important to treat them as separate notions, by describing indefinite languages that are in certain respects just as precise as classical definite ones, but which nevertheless cannot be understood sensibly as definite languages.
If $L$ is a classical language, then by an appropriate generalization of Theorem \ref{theoremrecstrat} (see Theorem \ref{theoremrecstratB}), it has a brute force strategy. Now if $L$ is part of an entry to a philosophical largest number contest, then the obvious way to complete the entry is by using the brute force strategy of $L$. The following theorem shows that this \emph{linguistic strategy} has the property that it increases in strength as the expressive power of the language increases:
\begin{theorem}[Cf. Theorem \ref{theoremlingstratB}]
\label{theoremlingstrat}
Let $L_1$ and $L_2$ be two precise proposition/term languages, at least one of which is classical, and suppose that $L_1$ can reproduce the metalanguage $L_2 + 1$, i.e. that the metalanguage $L_2 + 1$ can be translated into $L_1$. Then for any $n$, the phrase `the largest number that can be represented in $L_2$ using at most $n$ symbols' can be translated into $L_1$ as a term of length $O(\log(n))$ which succeeds at naming a number. Thus, in any philosophical largest number contest in which both strategies are considered valid, the linguistic strategy of $L_1$ will win against the linguistic strategy of $L_2$.
\end{theorem}
For example, it follows from well-known facts\Footnote{Specifically, the fact that set theory asserts that there is a set of all numbers, together with standard methods for formalizing in \textbf{\textup{ST}}\ the notions of truth and reference for a language whose domain of quantification is a fixed set.} that set theory \textbf{\textup{ST}}\ can reproduce the metalanguage of number theory \textbf{\textup{NT}}, so Theorem \ref{theoremlingstrat} shows that the brute force strategy of \textbf{\textup{ST}}\ wins against the brute force strategy of \textbf{\textup{NT}}. As another example, one way to phrase Tarski's theorem on the undefinability of truth (see \cite{Tarski} and Theorem \ref{theoremtarski}) is to say that no classical language is capable of reproducing its own metalanguage, and this is reflected in Theorem \ref{theoremlingstrat} in the fact that no language can describe a larger number than the largest number it describes.
\subsection{Selfmeta languages}
Despite Tarski's theorem, it turns out to be possible to create proposition/term languages that are capable of reproducing their own metalanguages, as long as they are not required to be classical. We will call a proposition/term language which is capable of reproducing its own metalanguage \emph{selfmeta}. The most well-known examples of selfmeta languages are \emph{Kripke extensions} of classical languages (see Section \ref{sectionkripke}), which are essentially selfmeta ``by definition'' in the sense that they have primitive notation for expressing their own concept of truth. However, we will see in the next two sections that there are other natural examples of selfmeta languages.
This leads to the question: what is the best strategy for the largest number contest of a selfmeta language (in the sense of Definition \ref{lnc1})? It turns out to be an axiomatic strategy, due to the following theorem:
\begin{theorem}[Cf. Theorem \ref{theoremselfmetaB}]
\label{theoremselfmeta}
Let $L_*$ be a precise selfmeta language, and let $A$ be a valid axiomatization of a precise proposition/term language $L$ with at least as much expressive power as $L_*$ ($L = L_*$ is allowed). Then for each $n$, the phrase `the largest number named by any term $t$ such that $A$ proves that $t$ succeeds at naming a number using reasoning encoded in less than $n$ symbols' can be translated into $L_*$ as a term of length $O(\log(n))$ which succeeds at naming a number. This term, which we will denote by $\double{\psi_{L_*,A}}{10^{100}}$, can be written explicitly. The proof of this theorem (for fixed $L_*,L,A$) can be formalized in the axiom system $A+1$.
\end{theorem}
For any axiom system $(L,A)$, we define the \emph{axiomatic strategy} of $A$ for the largest number contest of $L_*$ to be to submit the term $\double{\psi_{L_*,A}}{10^{100}}$ given in the above theorem. A player using the axiomatic strategy of $A$ strategy will win against any player submitting an entry whose validity can be proven in $A$, and by Theorem \ref{theoremselfmeta}, the validity of the axiomatic strategy of $A$ can be proven in $A+1$. This leads to the following corollary:
\begin{corollary}[Cf. Corollary \ref{corollaryselfmetaB}]
\label{corollaryselfmeta}
Let $L_*$ be a selfmeta precise language, and let $(L_1,A_1)$ and $(L_2,A_2)$ be valid axiom systems with at least as much expressive power as $L_*$. Suppose in addition that the axiom system $(L_1,A_1)$ is at least as powerful as the metasystem $(L_2+1,A_2+1)$. Then in the largest number contest of $L_*$, the axiomatic strategy of $A_1$ wins against the axiomatic strategy of $A_2$.
\end{corollary}
Note that Theorem \ref{theoremselfmeta} shows that a philosophical largest number contest can be ``hijacked'' if the judges of the contest are lax enough. Namely, the adherents of a school of thought $X$ may decide that they believe that a certain selfmeta language $L_*$ is precise, and then they may decide to declare that an entry to their contest is valid if and only if it is a term of $L_*$ that succeeds at naming a number. If the judges do not impose an additional requirement that the entry must be accompanied by a semantic proof that the entry names a number, then the contest can be won by adherents of other schools of thought that admit more expressive languages and correspondingly more powerful axiom systems. If these other philosophers' languages are valid, or at least valid enough that their corresponding axiomatic strategies are valid, then they will win the contest.
It is natural for the adherents of $X$ to prevent this ``hijacking'' by stipulating that entries to their largest number contest must be accompanied by semantic proofs that they name numbers. If they impose this requirement, then their largest number contest reduces to the strongest axiom contest of the selfmeta language $L_*$ run by adherents of $X$. In particular, this method of judging entries does not trivialize the largest number contest, so philosophers may find it acceptable. In fact, I\ will argue that any coherent position on the philosophy of mathematics can eventually be formalized into a stable selfmeta language, causing its largest number contest to reduce to the strongest axiom contest of that language:
\begin{thesis}
\label{thesis2}
Any coherent position on the philosophy of mathematics can be formalized into a ``maximal'' selfmeta language which adherents of the philosophy should accept as precise.
\end{thesis}
By contrast, the largest \emph{axiom} contest can never stabilize in this way. There are two reasons for this. First of all, anyone who accepts the validity of an axiom system $A$ starts to use reasoning that cannot be formalized in $A$, but can be reasonably formalized in the metasystem $A+1$. Thus it is natural to ask him\ whether he\ accepts the validity of $A+1$. If not, then it becomes difficult for him\ to justify his\ own reasoning as valid, and if so, then the axiomatic strategy of $A+1$ beats the axiomatic strategy of $A$ and thus the contest has not stabilized.\Footnote{This is analogous to the fact that any strategy in a largest number contest can be beaten by a modification of itself which says to add 1 to the number described by the original strategy, if such a modification is valid.} Secondly, the largest number \emph{proof} contest of a selfmeta language with respect to any fixed axiomatization has a ``brute force'' strategy, so it will be avoided by philosophers who believe in ``mind over matter''
\begin{lncs}
The \emph{largest number proof contest} of an axiomatization $A$ of a precise proposition/term language $L$ is the contest in which a valid entry is a pair of the form $(\psi,\rho)$, where $\psi$ is a term in $L$ and $\rho$ is a proof in $A$ that $\psi$ succeeds at naming a number, and the score of the entry $(\psi,\rho)$ is the number named by $\psi$.
\end{lncs}
\begin{theorem}
\label{theoremrecstrat3}
Let $A$ be an axiomatization of a selfmeta language $L = L_*$. Then for each $n$, the $n$th case of Theorem \ref{theoremselfmeta} can be proven in $A$ using reasoning encoded in at most $O(n)$ symbols. The proof in $A$, which we will denote by $\double{\rho_{L,A}}{n}$, can be given explicitly.
\end{theorem}
The brute force strategy for the largest number proof contest of $(L,A)$ can now be given as follows: First compute the largest number $n$ such that $\double{\rho_{L,A}}{n}$ can be submitted as an addendum to the entry $\double{\psi_{L,A}}{n}$ (without exceeding physical or imposed limits), and then submit the pair $(\double{\psi_{L,A}}{n},\double{\rho_{L,A}}{n})$ as an entry. As usual, this strategy is nearly optimal in the sense that no other valid strategy can beat it too badly.
Note that if $L$ is a selfmeta language and $A$ is an axiomatization of $L$, then $A+1$ can be viewed either as an axiomatization of $L+1$ or an axiomatization of $L$. Since $L$ is capable of reproducing $L+1$, both languages have equal expressive power and thus axiomatizations of either one can be viewed as axiomatizations of the other.\\
\subsection{Plan for the next few sections}
\label{subsectionplan}
In what follows, I\ will attempt to describe entries to the largest number and/or strongest axiom contests corresponding to various schools of thought in the philosophy of mathematics. For clarity I\ introduce the following terms:
\begin{itemize}
\item \emph{Set realism} is the view that it is possible to talk precisely about sets, understood in the classical sense according to which some sets are undescribable even in principle.
\item \emph{Descriptionalism} is the opposite view, that all mathematical objects should be describable at least in principle, even if it would take an extremely large (but finite) amount of time or resources to describe them.
\item \emph{Number realism} is the view that it is possible to talk precisely about natural numbers, understood in the classical sense according to which large numbers are inherently no different from small ones.
\item \emph{Ultrafinitism} is the opposite view, that the classical concept of a ``natural number'' is incoherent from a philosophical point of view.
\end{itemize}
Obviously, these labels are very broad, and it is possible to for schools of thought to get much more specific about their beliefs. In particular, the category of ``descriptionalist number realism'' includes the schools of thought traditionally known as ``constructivism'', ``intuitionism'', and ``predicativism''. However, I\ will find it convenient to avoid talking too much about about the relationship between the philosophical stances that one might take in response to a largest number contest and pre-existing philosophical schools, though I\ will mention similarities where they exist.
\begin{remark*}
The label `realism' should not be taken to indicate that adherents of either set realism or number realism are required to subscribe to \emph{philosophical realism}, the view that words and phrases like `the number 5', `the set of odd numbers', and `greenness' refer to ``real things'' that ``exist'' ``independently of the mind''.\Footnote{The scare quotes may allow the reader to guess what my\ view on the subject is: I\ am not sure what it would mean to say that a mathematical entity ``exists'' or ``does not exist''. Certainly mathematical entities are not typical examples of the category of ``things that exist'', but this appears to be a very loosely defined category (even more loose than the category of things that can be precisely talked about), and is there any point in trying to define precise boundaries for it? In the absence of any motivating question, it appears that the answer is no.} A mathematical realist may or may not believe that the entities he talks about ``actually exist''; he only needs to believe that it is possible to talk about them precisely.
Out of the two types of mathematical realism, set realism is more closely connected with philosophical realism, since arguments in favor of set realism often presuppose philosophical realism. On the other hand, there is a tension between the two types of realism in that philosophical realism asserts that e.g. ``greenness'' exists \emph{as greenness}, while set realism asserts that it exists \emph{as the set of all things that are green}.\Footnote{This way of putting things is in some ways an oversimplification, since a set realist may deny that the notion of greenness is well-defined enough for there to be a set of all things that are green. This observation highlights the differences between the set-realist and philosophical-realist perspectives.} This tension is highlighted by the fact that set realists believe that some sets are undefinable and undescribable, whereas it is not clear that a ``property'' or ``universal'' in the sense of philosophical realism could ever be in-principle undescribable. These kinds of considerations may lead a philosophical realist to reject set realism in favor of descriptionalism.
\end{remark*}
Before delving into the details of how I\ think set realists and number realists might be able to judge their largest number and strongest axiom contests in a philosophically defensible way, I\ will say a few words regarding some general objections to the project I\ am undertaking:\\
{\it The pragmatist/formalist/ultrafinitist objection.}
Some mathematicians may object that it does not make sense to talk about the truth or falsity of mathematical sentences\ unless we specify an axiom system according to which alleged proofs or disproofs of these sentences\ are to be judged. There are a few natural objections to this point of view. First of all, the question arises of whether the truth that can be talked about once we have specified an axiom system is conceptually independent of provability in that axiom system, or whether to say that a sentence\ is true simply \emph{means} that it can be proven in the axiom system. If it is conceptually independent, then it is not clear why it cannot be talked about independently of the axiom system. On the other hand, if truth is nothing more than syntactic provability (a view known as \emph{formalism}), then it is not immediately clear what it means to say that a sentence\ is syntactically provable, since the claim that a given sentence\ is syntactically provable is usually considered to be itself a mathematical statement\ (asserting the existence of a syntactic proof), and so it is not immediately clear how such a statement\ could be meaningful unless there is some theory according to which at least some mathematical statements\ are meaningful independent of their provability in any axiom system. However, the notion of syntactic provability can be instead interpreted in the social sense: as the ability of mathematicians to construct a syntactic proof. (In practice, this ability is almost completely hypothetical since mathematicians rely nearly exclusively on informal descriptions of syntactic proofs. However, there appears to be no reason that this should constitute a philosophical problem for formalism.) This raises the question of what constitutes a syntactic proof, and it can again be understood in a social sense: as the question of a proof-checker ``following convention'' would (hypothetically) find an ``error'' or ``non sequitur'' in the alleged proof.
The objection to this view is that our intuition suggests that the question of whether an alleged syntactic proof is in fact a syntactic proof seems to be just as precise as the simpler questions of which it is composed, such as the question of whether or not two given symbols in the alleged syntactic proof are the same symbol, and moreover these simpler questions seem to be perfectly precise. If so, then this principle of ``conservation of precision under finite quantification'' is enough to get us a very rudimentary version of number realism along with a corresponding largest number contest (see Section \ref{subsectionGNR}). However, one may object that the new question is not \emph{quite} as precise as the old one, since if we create machines for checking syntactic proofs, then the probability that two machines will give different answers gets larger as the alleged proof gets longer. Eventually the probability will be large enough that the question of what the ``right'' answer is will no longer be precise (or so the argument goes). The obvious objection to this is that we ordinarily think that the reason that two machines usually give the same answer is \emph{because} there is a ``right'' answer for both of them to give, not the other way around. It is not obvious how to explain this fact other than by thinking of the machine calculations as approximations of an abstract calculation which remains precise regardless of physical implementability.
If even the principle of conservation of precision under finite quantification is rejected, then there is not much sense in talking about largest number contests, because on such a view it is not clear that the concept of a ``number'' is coherent enough to be worth talking about. Namely, one can talk about various ways of storing information that are classically thought of as ``representing'' numbers, and the stored information can itself be called a ``number'', but there appears to be no principled way to distinguish between valid and invalid representations of numbers, since every ``number'' uncontroversially represents itself but it is difficult to argue that it represents any other kind of number. For example, binary numbers cannot in general be thought of as ``representing'' numbers stored as tally marks, since it is easy (and in fact common practice) to store binary numbers so large that they cannot feasibly be implemented as tally marks. This ``radically relativist'' view about numbers is known as \emph{actualism} or \emph{ultrafinitism}, and it is hard to argue against, though it goes against many people's intuitions.\Footnote{In particular, ultrafinitism denies the intuitively plausible claim that the validity of the standard (feasible) algorithm for comparing the sizes of binary numbers can be thought of in terms of a second (unfeasible) algorithm that converts both numbers to a unary representation (i.e. tally marks) and then compares their raw sizes. Namely, it is intuitively plausible that \emph{what we mean} by saying that the standard algorithm is valid is that it gives the same answer as the second algorithm would. But ultrafinitism denies the coherence of talking about what the second algorithm would do, once the numbers involved are large enough.} It also seems to be difficult to explain the success of mathematical methods using ultrafinitism, though perhaps it is not impossible.\\
{\it The imprecision objection.} Another common objection, perhaps related to the formalist/pragmatist one, is that the question ``is this entry a precise description of a number?'' is not itself a precise question. This is true, but it misses the point that attempts to convince someone that a certain description is precise are really attempts to expand their worldview by introducing new intuitive concepts. Of course there is no precise way of characterizing when we have successfully expanded our worldview (rather than introducing erroneous thinking), but anyone who is not a pure formalist has already expanded their worldview by allowing at least some intuitively intelligible precise concepts, so the idea of further expansion should not be viewed with so much suspicion.\Footnote{One can use the imprecision objection to justify not taking too seriously the question of which philosophical school of thought is ``correct'', on the grounds that this question is not precise, and I\ mostly endorse this position.}
\section{Classical vs selfmeta languages}
\label{sectionkripke}
One of the most fundamental questions in the philosophy of language is: what does it mean for a statement\ to be \emph{true}? In general the answer to this question may depend on many things, such as the context in which the statement\ was made.\Footnote{For a simple example, whether or not a statement\ of the form `John did X' is true depends on which ``John'' the speaker is referring to.}
Additionally, some statements\ are so vague that we would be uncomfortable with saying that the question of whether or not they are true is a precise one. However, the whole point of creating a precise language is that it should be possible to give a precise, context-independent answer to the question of what it would mean for any given sentence\ in the language to be true. Thus, any coherent philosophy should be able to explain precisely what it would mean for any given sentence\ in a language considered precise to be true.
The theory of mathematical logic contains a partial answer to this question in the case of classical languages. Specifically, classical model theory attempts to make precise the notion of truth with respect to a classical language by defining it recursively. For example, a sentence\ of the form `$\forall x \; \phi(x)$' (short for ``for all $x$ [in the category of objects under consideration], $\phi(x)$ is true'') is defined to be true of a given model $M$ if and only if for every object $x$ in $M$, $\phi(x)$ is true of $M$.\Footnote{Strictly speaking, we should say that `$\forall x \; \phi(x)$' is true of $M$ if for every object $a\in M$, $\phi(x)$ is true of $M$ relative to the variable assignment $\lb x=a\rb$.} This definition is \emph{recursive} in the sense that a concept (``truth'') is defined in terms of itself.
The question now becomes: how can a recursive definition be considered to be precise? It certainly gives some information about the concept being defined, namely a way in which it relates to itself, but there is conceivably some imprecision remaining, since the concept has not been cashed out in terms of fundamental precise concepts. We can make recursive definitions more precise by instituting the convention that they should only be considered meaningful in cases where the meaning ultimately derives from concepts outside the recursion. To formalize this convention, we introduce the notion of a ``category generated by production rules'', which is perhaps best illustrated by example. Namely, suppose we have a two-letter alphabet $E$ consisting of the letters `a' and `b'. Then the notion of a \emph{string in $E$} can be defined by the following production rules: the letters `a' and `b' both constitute strings in $E$, and whenever we have a string in $E$, then concatenating it with either `a' or `b' forms another string. These two rules generate the category; any object that cannot be eventually created by these production rules is not a string in $E$.
The concept of production rules can be used to define the concept of a recursive definition as follows: each case of a recursive definition corresponds to a different production rule for the category of ``facts about the recursively defined concept''. For example, the recursive definition of truth considered above could be described in terms of production rules for the category of true sentences\ and the category of false sentences: there would be a production rule stipulating that ``if the sentences\ $\phi(x)$ $(x\in M)$ are all members of the category of true sentences, then `$\forall x \; \phi(x)$' should be added to the category of true sentences'', and another production rule stipulating that ``if at least one of the sentences\ $\phi(x)$ $(x\in M)$ is a member of the category of false sentences, then `$\forall x \; \phi(x)$' should be added to the category of false sentences''.
Does the concept of production rules need to be defined, or is it an irreducible concept? One way we can try to define it is by using the notion of \emph{indexical recursion}, where a sequence of objects $x_0, x_1, x_2,\ldots$ is defined by specifying how to construct each object $x_k$ in terms of the previous objects $x_0,\ldots,x_{k-1}$. Indexical recursion may be viewed as less problematic than general recursion for two reasons. First, the structure of indexically recursive definitions makes it clear that the definition is a reductive one, reducing more complicated concepts to simpler ones as measured by the indices of the objects being considered. Second, with indexically recursive definitions it is clear that the meaning always ultimately derives from outside the recursion, so there is no need for a convention handling the case where this does not happen.
We could attempt to define the notion of a category generated by production rules in terms of indexical recursion as follows: for each number $k$ there is the concept of the ``category generated by stage $k$'', denoted $C_k$. It is defined by indexical recursion as the set of all things that can be created using a single instance of the production rules from elements of $C_j$, $j < k$. We could then define the final generated category to be the union $C = \bigcup_k C_k$. However, this definition does not capture the intuitive notion of a category generated by production rules (at least in the sense we mean here), because the category $C$ is not necessarily closed under the production rules, if there are ``infinitary'' production rules such as the rule considered above for adding universally quantified sentences\ to the collection of true sentences. Thus, it is conceivably possible to add more objects to $C$ using the same production rules, and what we really want to consider is the category that results from continuing to do this indefinitely.
Thus, it is problematic to reduce the notion of a category generated by production rules to the notion of indexical recursion. However, a more mathematically sophisticated variant on this idea based on transfinite induction is arguably more successful (we leave the details to the reader). Nevertheless, for the remainder of this section we will treat production rules as an irreducible concept.
It is worth mentioning another well-known and popular attempt to reduce the concept of production rules to other fundamental concepts: the ``intersective definition'' due to Frege. They consider the scenario where the objects generated by a list of production rules $R$ are known in advance to lie in a given set $S$, and they define the set generated by the list $R$ to be the intersection of all subsets of $S$ that are closed under the production rules of $R$.\Footnote{A set is \emph{closed} under a list of production rules if it is not possible to add new members to the set by applying the production rules.} However, in my\ view this reduction is neither conceptually valid nor appropriate for the present context, for several reasons summarized in Appendix \ref{appendixintersectivedefinition}, and consequently I\ will ignore it in the subsequent discussion.
To summarize, using the concept of recursion, possibly supplemented by the concept of production rules, it is possible to precisely define the notion of truth with respect to a classical language such as \textbf{\textup{NT}}\ or \textbf{\textup{ST}}, assuming that the relevant category (of numbers or sets) is precise enough to be quantified over. This fact is surely a relief to any philosopher who wishes to maintain that these languages are precise. But there is no need to stop there. The argument from recursion can be modified to handle the notion of truth not only for classical languages, but also for certain selfmeta extensions which we now define:
\begin{definition}
\label{definitionkripke}
The \emph{Kripke extension} of a classical language $L$, which we will denote by $\K(L)$, is simply the language $L$ augmented with a primitive ability to write phrases of the form `$\phi$ is a true sentence\ of $\K(L)$', where $\phi$ is a term of $\K(L)$, intended to denote a string.
\end{definition}
Let us be clear that the Kripke extension of $L$ is very different from the metalanguage of $L$. The metalanguage of $L$ is denoted $L+1$ and has the primitive ability to write phrases of the form `$\phi$ is a true sentence\ of $L$', where $\phi$ is a term of $L+1$. However, $L+1$ does not have the ability to write phrases of the form `$\phi$ is a true sentence\ of $L+1$'. The metalanguage of $L$ is a classical language, whereas the Kripke extension of $L$ is a selfmeta language.
The crucial question about Kripke extensions is: Is the Kripke extension of a precise classical language itself a precise language? Equivalently, is there a precise characterization of what it means for any given sentence\ of a Kripke extension to be true?
Fortunately this question has a fairly simple answer: \emph{If the notion of truth for the original classical language can be defined precisely in terms of production rules, then so can the notion of truth for the Kripke extension}. The reason for this is that one simply has to add the following rules to the list of production rules:
\begin{itemize}
\item ``If $\phi$ is in the category of true sentences, then add `$\phi$ is true' to the category of true sentences, and add `$\phi$ is false' to the category of false sentences.''
\item ``If $\phi$ is in the category of false sentences, then add `$\phi$ is false' to the category of true sentences, and add `$\phi$ is true' to the category of false sentences.''
\end{itemize}
There is nothing unusual about these production rules, in that they have the same sort of structure as the production rules used in the recursive definitions of truth for \textbf{\textup{NT}}\ and \textbf{\textup{ST}}, so adding them does not appear to lead to any further philosophical problems.
Thus, it appears that if $L$ is a classical language which can be viewed as precise according to the standard picture of the meaning of languages provided by mathematical logic, then the same is true for the Kripke extension $\K(L)$, which is a selfmeta language. It is natural to ask whether the same principle works in reverse, namely whether every precise selfmeta language has a precise classical extension.
In favor of an affirmative answer, one could argue that if $L$ is a precise selfmeta language, then the fact that $L$ is precise already means that any sentence of $L$ either has a truth-value or does not. If so, then we could define the classicization $\C(L)$ to be \textbf{\textup{NT}}\ augmented with the ability to talk about whether sentences in $L$ are true, false, or indeterminate. (Note that this presupposes that \textbf{\textup{NT}}\ is precise.) It would seem that $\C(L)$ would be a classical language with at least as much expressive power as $L$.
To get a clear picture of the challenges this view faces, it is helpful to analyze why a Kripke extension $\K(L)$ is not already capable of reproducing its classicization $\C(\K(L))$. After all, the expressive power of $\C(\K(L))$ comes from the fact that it can express the claim that a given sentence $\phi$ of $\K(L)$ is true, false, or indeterminate, and $\K(L)$ can already express the claim that $\phi$ is true or false. So the advantage of $\C(\K(L))$ is that it can express the claim that $\phi$ is indeterminate.
But as a matter of fact $\K(L)$ can already express this claim in a roundabout way: as `$\phi$ is neither true nor false'. However, if we think of this as a sentence of $\K(L)$, then it can never be true. According to the definition of the semantics of $\K(L)$ via production rules, `$\phi$ is neither true nor false' could only be true if `$\phi$ is true' and `$\phi$ is false' are both false, which in turn could only be true if $\phi$ was both false and true, which is impossible.
The paradox here is that while no sentence\ of $\K(L)$ can be truthfully asserted to be indeterminate within $\K(L)$, nevertheless there are sentences\ of $\K(L)$ that are clearly neither true nor false. The standard example is the liar's paradoxical sentence `This sentence is false'. This sentence can be translated into the language \K(\textbf{\textup{NT}}) (the Kripke extension of classical number theory) using the standard method of quining. To be precise, there exists a sentence\ $\phi$ of \K(\textbf{\textup{NT}}) which is provably equivalent to `$\phi$ is false'. Thus if $\phi$ is true, then $\phi$ is false, and if $\phi$ is false, then $\phi$ is true. Since $\phi$ cannot be both true and false, it follows that $\phi$ is neither true nor false. Moreover, it is easy to formalize this argument to show that according to the production rules $\phi$ will never be added to the category of true sentences\ nor to the category of false sentences.
The resolution of the paradox is that the phrase ``not true'' has a slightly different meaning when we use it inside $\K(L)$ and when we use it outside $\K(L)$. Inside $\K(L)$, asserting that a sentence is not true is the same as asserting that it is false, which is the same as asserting that it is in the category of false sentences. But outside $\K(L)$, asserting that a sentence is not true means only that it will never be added to the category of true sentences.
Thus, the classicization $\C(\K(L))$ has a hidden power that $\K(L)$ does not: it can make ``never'' claims, rather than merely claims about what \emph{is} in a given category. Now it is not clear that the former type of claim is anywhere near as precise as the latter. For one thing, if we view the generation of a category as an ongoing process, then claims that something will never get added to a category are claims about the future, while claims that something is already in a given category are claims about the present.
But the more serious objection is that it is not clear how to define the semantics for a language in which one can make assertions of indeterminacy interpreted in a weak sense: if we proceed in the usual way via production rules for the categories of true and false sentences, then it is not clear what sort of production rule should lead to adding a sentence\ of the form `$\phi$ is indeterminate' to the category of true sentences.\Footnote{More precisely, it is not clear what sort of production rule should lead to adding a sentence\ of the form `$\phi$ is not true' to the category of true sentences; the natural candidate leads to the notion of indeterminacy being interpreted in the strong/impossible sense, not the weak sense.}
In conclusion, selfmeta languages appear to be favored over classical languages from a philosophical point of view, in the sense that they make more natural ``stopping points'' for increasing the expressive power of a language. Beyond this, it is natural to ask specifically which selfmeta languages best correspond to which philosophical positions. For this purpose, it is necessary to consider separately the set and number realist points of view.
\section{Number realism}
\label{sectionnumberrealism}
There are many different points of view that a number realist could take, leading to many different selfmeta languages for our Philosophical Largest Number Contest. Below, they are listed in order of increasing expressive power. We show that each language is capable of reproducing the classicization of the previous language, and thus the languages are ordered in terms of increasing power. Each language has various possible axiomatizations and various possible classical sublanguages whose definitions we leave to the reader.
\subsection{Gradualist number realism}
\label{subsectionGNR}
Though there are many ways to analyze the philosophical divisions that exist between number realists, a natural ``litmus test'' is the following \emph{Principle of Limited Omniscience}, which some number realists find to be intuitively obvious, while others deny its validity:
\begin{principle}[Principle of Limited Omniscience]
Let $\phi(0),\phi(1),\ldots$ be a sequence of sentences\ that are precise enough to have a definite truth-value. Then the sentence\ $\Phi = ` \forall n \; \phi(n)$' (i.e. `all of the sentences\ $\phi(0),\phi(1),\ldots$ are true') is precise enough to have a definite truth-value.
\end{principle}
The name comes from constructivist philosophy: some philosophers known as ``constructivists'' take the point of view that to claim that a mathematical sentence\ has a definite truth-value is the same as to claim the (hypothetical) ability to discover what the truth-value is.\Footnote{Some constructivists may instead take the weaker point of view that mathematics is best purified by labelling claims of the form `$X$ has a definite truth-value, but I don't know how to find it' as ``non-mathematical''.} On this view, a mathematician who uses the Principle of Limited Omniscience to conclude that the Goldbach conjecture (the claim that every even number $\geq 4$ can be written as the sum of two primes) is either true or false is claiming the ability to figure out (given enough time and effort, but with a precise plan from the start) whether the Goldbach conjecture is true or false. However, no algorithm is known to be able to succeed at this task. Thus, the mathematician appears to be claiming a kind of limited omniscience, hence the name.
An obvious criticism of this view is that in ordinary reasoning, when one claims that a sentence\ has a definite truth-value, one is not ordinarily claiming to know what the truth-value is, even in a speculative way. For example, after flipping a coin and not looking at the outcome, one may reasonably assert that the claim that the coin has landed on heads has a definite truth-value, even though one may not know which it is. It is not clear why mathematical reasoning should be any different.
Nevertheless, there are still good reasons why someone might not accept the Principle of Limited Omniscience. One reason is the principle that a claim cannot be true unless there is/are some thing/things that makes/make it true. The obvious answer to the question of what kind of things could make the claim $\Phi$ true is that the truths of the claims $\phi(0),\phi(1),\ldots$ could make $\Phi$ true. But it is not clear that it makes sense to talk about infinitely many things at the same time, and if it is impossible to do so, then the previous sentence does not make sense and thus $\Phi$ cannot be made true in this way. (There may be other ways to make $\Phi$ true, such as the existence of valid reasoning of one sort or another demonstrating that $\phi(n)$ is true where $n$ is a free variable used in the reasoning, but it is not at all clear that these other ways will be enough to make up for the loss of power from the ``main'' or ``canonical'' way of making $\Phi$ true. See Appendix \ref{appendiximplication} for further discussion.)
There are a couple of different ways to phrase the same point. Recall that in the previous section, I\ argued that future-tense claims about ongoing processes are inherently less precise than present-tense claims. Now if we view the assignment of truth-values to the claims $\phi(0),\phi(1),\ldots$ as an ongoing process, such that at any given point in time only finitely many of the claims have been assigned truth-values, then the claim that all of the sentences\ $\phi(0),\phi(1),\ldots$ will eventually be labelled as true is a future-tense claim, and thus we can coherently reject it on the grounds of imprecision.
Alternatively, if we take the point of view that what really matters is our ability to write an algorithm computing the truth-value of any sentence, then we can reject universal quantification on the grounds that our algorithm could never safely return `true' as the truth-value for a universally quantified claim, since at any point in time the algorithm will only have computed finitely many of the truth-values of the sentences\ $\phi(0),\phi(1),\ldots$. Again, this defect could be partially fixed if we allowed the algorithm to return `true' if there is a proof that a universally quantified sentence\ is true, but this solution is inelegant and in any case many true sentences\ cannot be proven.
These arguments point not only towards rejecting the Principle of Limited Omniscience, but towards rejecting unbounded universal quantification over the category of numbers as a precise concept. I\ call \emph{gradualist number realism} the view that unbounded universal quantification should be rejected for these reasons, but bounded universal and unbounded existential quantification should be allowed. The corresponding mathematical theory will be called \emph{gradualist number theory} and denoted \textbf{\textup{GNT}}.\Footnote{Gradualist number theory may be compared with \emph{Heyting arithmetic}, a theory which allows unbounded universal quantification but requires its axiomatizations to use it in a more restricted way. To me the fundamental problem with Heyting arithmetic is that its semantics are fundamentally unclear due to the way that implication is treated; see Appendix \ref{appendiximplication} for a more sensible treatment of the notion of implication.} The term ``gradualist'' refers to the fact that a gradualist number realist may hold that the creation of the category of numbers is an ongoing process.
\subsection{Logicist number realism}
If a number realist accepts the Principle of Limited Omniscience, as well as the concept of recursion via production rules described in the previous section, then he\ must also accept the language \K(\textbf{\textup{NT}}), the Kripke extension of classical number theory. One option is to go no farther, and we call this view \emph{logicist number realism}
\subsection{The language of generated categories}
\label{subsectionGC}
One reason to go beyond \K(\textbf{\textup{NT}}) is based on analyzing the intuitions behind the Principle of Limited Omniscience. The fundamental idea is that it works because the category of numbers is ``well-defined'': it includes everything that you get if you start with $0$ and repeatedly apply the successor operation. But this intuition can be extended to other categories with precisely defined generating processes, leading to the following generalization of the Principle of Limited Omniscience:
\begin{principle}[Principle of Precise Categories]
\label{principleGC}
Let $C$ be a category generated by production rules, and suppose that we have a method for associating to each member $x$ of $C$ a sentence\ $\phi(x)$ which is precise enough that it is either true or false. Then the sentence\ $\Phi = `\forall x\in C \; \phi(x)$' (``for all $x$ in $C$, $\phi(x)$ is true'') is either true or false.
\end{principle}
Given Principle \ref{principleGC}, it follows from a trivial argument\Footnote{If $\phi$ is a sentence\ of \K(\textbf{\textup{NT}}), then for every $\psi$ in the category of true sentences\ of \K(\textbf{\textup{NT}}), either $\phi = \psi$ or $\phi \neq \psi$. Thus by Principle \ref{principleGC}, either $\phi = \psi$ for some such $\psi$, or $\phi \neq \psi$ for all such $\psi$. But in the first case $\phi$ is true, while in the second case $\phi$ is not true.} that every sentence\ of \K(\textbf{\textup{NT}}) is either true or not. Thus the classicization \C(\K(\textbf{\textup{NT}})) is also a precise language, and in particular Principle \ref{principleGC} implies that \K(\textbf{\textup{NT}}) is not maximal. This raises the question: What kind of language is best capable of formalizing the reasoning of a philosopher who accepts Principle \ref{principleGC}?
Well, it is clear that if we accept Principle \ref{principleGC}, then the notion of a category generated by production rules becomes crucially important. Thus, it becomes useful to develop a convention for describing such categories. For simplicity we consider the case where the categories to be described are subsets of the natural numbers, such as the set of all true sentences of \K(\textbf{\textup{NT}}) or similarly constructed categoies. Our convention in this case is as follows: if $\Phi(n,C)$ is a formula with one first-order free variable `$n$' and one second-order free variable `$C$' that only appears positively (i.e. `$m\in C$' can appear but not `$m\notin C$'), then the category $C_\Phi$ is generated according to the family of production rules: ``If $\Phi(n,C_\Phi)$ is true, then add $n$ to $C$''. For example, if $C_\Phi$ is the category of true statements of \textbf{\textup{NT}}\ then $\Phi(n,C)$ would be a disjunct of several statements including ``$n=`\forall m \; \phi(m)
RQ$ and $\forall m \; `\phi(m){\textrm'}\in C$''.
Using this convention, we define the \emph{language of generated categories} \GC, which is \textbf{\textup{NT}}\ augmented with the ability to refer to categories of the form $C_\Phi$, where $\Phi(n,C)$ is a formula in \GC. We need to allow $\Phi$ to range over \GC, and not just \textbf{\textup{NT}}, in order to make \GC\ a selfmeta language (we leave the proof that it is selfmeta as an exercise\Footnote{The basic idea is to recursively define a map $\phi(x_1,\ldots,x_n) \mapsto C_\Phi$ from formulas with an arbitrary number of free variables into categories, such that $\phi(a_1,\ldots,a_n)$ is true if and only if $\lb a_1,\ldots,a_n\rb \in C_\Phi$, where $\lb \cdot\rb$ is the standard $n$-tupling function on $\N$.}). However, it has the effect that not all categories $C_\Phi$ that can be written in \GC\ are well-defined. If $\Phi$ refers to \textbf{\textup{NT}}, then $C_\Phi$ is well-defined, and if $\Phi$ refers only to categories $C_\Psi$ such that $C_\Psi$ is well-defined, then $C_\Phi$ is well-defined. However, there is no way of knowing in advance whether this is the case or not. The strength of an axiomatization of \GC\ could be measured by how many categories $C_\Phi$ it can prove are well-defined
\begin{example}
The set of true statements in \K(\textbf{\textup{NT}}) is a generated category in the sense described here (we leave the details as an exercise to the reader).
\end{example}
\begin{example}
The set of objects in the minimal model of Kripke--Platek set theory, together with true non-quantified statements about this minimal model, is a generated category.
\end{example}
\begin{example}
The set of computable ordinals is a generated category. It follows that the expression $f(n) = BB_{\alpha(n)}(n)$ defined by Scott Aaronson in his MathOverflow post ``Succinctly naming big numbers: ZFC versus Busy-Beaver''\Footnote{\url{https://mathoverflow.net/questions/34710/succinctly-naming-big-numbers-zfc-versus-busy-beaver?rq=1}} is definable in \GC. It also follows that the Feferman--Schutte ordinal $\Gamma_0$ can be expressed in \GC, meaning that \GC\ is to be considered ``impredicative''.\Footnote{The ordinal $\Gamma_0$ is supposed to be the exact limit of predicative reasoning, which means that predicative reasoning will be best described by a classical language, whose largest number contest therefore has a brute force strategy. Similar remarks apply to this case as to the case of classical set theory discussed in the next section.}
\end{example}
\subsection{Transfinite time and constructive set theory}
\label{subsectionDST}
Having reduced the concept of recursion to the concept of production rules, we can ask whether we can reduce the concept of production rules to something more fundamental. The answer becomes clearer if we ask: in what order do objects get added to a generated category? One could answer that this question is nonsensical, in which case the language of generated categories would be the most fundamental language. However, if we think that the question is meaningful, then we are led to a notion of time more general than finite time, i.e. infinite or ordinal time. This is because for an infinitary production rule like ``if $\phi(n)$ is true for every $n$, then $\forall n \; \phi(n)$ is true'', if each $\phi(n)$ is added to the category of true statements at a different point in time, then $\forall n \; \phi(n)$ can only be added after an infinite amount of time has passed. Ordinal arithmetic gives a way of quantifying infinite amounts of time
So, adding objects to categories according to production rules is a computation that takes place over ordinal time. It is then perhaps natural to allow set theory based on transfinite time, as long as it does not break the descriptionalist principle that every mathematical object should be describable in principle. We define the universe of constructible sets $\LL$ as follows:
\begin{definition}
\label{definitionconstructibleuniverse}
~
\begin{itemize}
\item At the start of time $\LL$ is the empty set $\emptyset$.
\item At any given ordinal time $\alpha$, we denote the current value of $\LL$ by $L_\alpha$. The successor $L_{\alpha+1}$ is defined as
\[
L_{\alpha+1} = L_\alpha \cup \{ \{x\in L_\alpha : \phi^\alpha(x)\} : \phi\in \textbf{\textup{ST}}\},
\]
where $\phi^\alpha(x)$ denotes $\phi(x)$ with its predicates restricted to $L_\alpha$ (so e.g. if $\phi(x)$ is ``$\forall y \; y\in x$'', then $\phi^\alpha(x)$ would be ``$\forall y\in L_\alpha \; y\in x$'').
\item The universe continues expanding without limit as long as the principle that every object must be describable can be adhered to.
\end{itemize}
\end{definition}
Note that each set $\{x\in L_\alpha : \phi^\alpha(x)\}$ can be described perfectly in terms of the formula $\phi$ and the ordinal $\alpha$. Thus, the universe will continue to expand for as long as it is possible to precisely describe ordinals.
It does not make sense to universally quantify over the entire universe of constructible sets, because the amount to which it can expand is defined vaguely. However, it makes sense to existentially quantify over it, since asserting the existence of an ordinal is the same as claiming to have a way of identifying the ordinal. Thus, we define the language of gradualist descriptivist set theory \textbf{\textup{GDST}}\ in an analogous manner to gradualist number theory \textbf{\textup{GNT}}\ (see Appendix \ref{appendixPT} for details), but quantifying over sets rather than over numbers.
The strength of the language \textbf{\textup{GDST}}\ depends on exactly how it is interpreted: which ordinals are describable? One answer is ``the ordinals that can be proven to exist in Kripke--Platek set theory'', and then \textbf{\textup{GDST}}\ is weaker than the language of generated categories \GC.
However, with a more powerful axiom system, one can prove that \textbf{\textup{GDST}}\ is capable of reproducing the classicization of the language of generated categories \C(\GC). For example, the axiom ``there is an ordinal that cannot be referred to by a term of \textbf{\textup{GDST}}'' is sufficiently strong.\Footnote{Indeed, if $f$ is any term in \textbf{\textup{GDST}}\ that refers to a method of associating to an ordinal a subset of $\N$, then for each $n$ we can consider the term `$\text{fst}\{\alpha : n\in f(\alpha+1)\}$': by hypothesis, the ordinal $\beta$ coming from our axiom is not equal to this ordinal for any $n$. It follows that for all $n$, either $n\notin f(\alpha+1)$ or $n\in f(\beta)$ for some $\beta\leq \alpha$. So if $f$ is monotone increasing, then $f(\alpha) = f(\alpha+1)$. This implies that \textbf{\textup{GDST}}\ is capable of reproducing \C(\GC).}
\section{Set realism}
\label{sectionsetrealism}
The ``official'' philosophy of the mathematical community, to the extent that there is one, is \emph{set realism}, the view that it is possible to talk precisely about ``sets'', where the word `set' is understood in at least one of its two classical senses. This philosophy is enshrined in the convention that a properly ``mathematical'' proof is one that can be formalized in the axiom system \textbf{\textup{ZFC}}\ (Zermelo--Fraenkel set theory with Choice), which is the standard axiomatization of the language of classical set theory \textbf{\textup{ST}}. However, most mathematicians are not very aware of the philosophical arguments for and against set realism, and perhaps the primary reason for the widespread use of \textbf{\textup{ZFC}}\ is simply the largeness of its class of syntactically provable sentences\ -- many mathematicians are pragmatists or formalists, and if we are not interested in philosophical truth and we are just manipulating symbols, then we might as well allow the broadest possible rules for symbol manipulation, as long as the system doesn't become uninteresting due to the discovery of an inconsistency.
There are two different classical ways of interpreting the word `set', corresponding two different types of set realism:\\
{\it The naive view.} One point of view is that the intuitive notion of a ``set'' or ``collection'' is already precise enough that the phrase `the category of all sets whose members are sets, members of members are sets, and so on' uniquely identifies a category of objects, the ``universe of (hereditary) sets''. According to this view, sentences\ in \textbf{\textup{ST}}\ are thought of as talking about and quantifying over the members of this category.
A key point brought up in favor of this view is the fact that we ordinarily consider two different lists of objects to be equivalent if they agree on the question of which objects are members of the list. For example, the list `Alice and Bob' is considered to be equivalent to the list `Bob and Alice' even though the word order is different. It is then claimed that this is evidence that we believe that there is a pre-existing object (a ``collection'') that the lists `Alice and Bob' and `Bob and Alice' both identify.
A criticism of this view is that the fact that two terms are considered equivalent does not necessarily imply that there is some pre-existing object that we are thinking of them as identifying. In fact, the fact that we think of their equivalence in terms of the members of the lists rather than in terms of the two lists' identifying a single common object suggests that perhaps there is not such an object.
Another criticism of the naive view is the well-known Russell's paradox. Although it is usually conceived of as a mathematical argument yielding a contradiction in a certain axiom system (known as \emph{naive set theory}), its relevance to philosophy is that there is no obvious reason why `the set of all sets that do not contain themselves' is not a valid description of a set under the intuitive notion of a set. Thus the paradox appears to show that the intuitive notion of a set as a pre-existing object (assuming that there is such an intuitive notion) is inconsistent, and in particular is not a precise notion.
Two ways a set realist can respond to this are (a) to claim that their intuition does not suggest that there is a set of all sets that do not contain themselves or (b) to suggest that there should be a ``limiting principle'' used in conjunction with intuition, for example John von Neumann's criterion that ``A [class] is too large [to be a set] [if and only if] it can be put in one-to-one correspondence with the [class] of all sets'' \cite[footnote 6]{Maddy1}. It seems that there is not much to say about (a); regarding (b), it seems problematic that this ``limiting principle'' was not suggested by intuition but merely by the discovery of paradoxes. In fact, the idea that infinite sets and ``classes'' have ``sizes'' that can be compared by checking for the existence of one-to-one correspondences is itself a modern idea that appears to have only a meager basis in intuition.
Another objection is that the distinction between sets and ``classes'' is conceptually incoherent since ``classes'' seem to have all of the properties we would intuitively expect sets to have. One response to this is that the notion of a ``class'' corresponds to the notion of a \emph{logical} set rather than the notion of a classical set; that is, it is simply a description of a criterion for distinguishing between sets abstracted away from the details of that description. Another possible response is to deny that the notion of ``classes'' is legitimate and to use other heuristics for thinking about the ``size'' of sets rather than von Neumann's.
A final objection is that in order to precisely talk about a category of sets, it is first necessary to precisely define what sorts of objects can be members of those sets. The classical way of answering this is to restrict attention to the class of \emph{hereditary} sets, i.e. sets whose members are sets, whose members are sets, all the way down. However, one might respond that this leads to circularity: the concept of a hereditary set is only well-defined to the extent that the category of possible members of a hereditary set, i.e. the category of hereditary sets, is well-defined.\\
{\it The iterative picture.} In part due to the difficulties outlined above, many set realists today prefer a different conception of the universe of sets, which can be described as follows:
\begin{intpict}[Iterative universe of sets, classical conception]
\label{intclassST}
~
\begin{itemize}
\item[1.] The construction of the universe of sets occurs in ``stages''. At each stage, there is a well-defined collection of sets that exist at that stage, and each set that exists at a stage also exists at all later stages. The collection of all sets that exist at stage $\alpha$ is usually denoted $V_\alpha$.
\item[2.] In the first (or 0th) stage, no sets exist, i.e. $V_0$ is defined to be the empty collection.
\item[3.] After we have reached stage $\alpha$ and consequently defined a collection $V_\alpha$, when passing to the next stage $\alpha+1$ we add to $V_\alpha$ all subsets of $V_\alpha$, i.e. $V_{\alpha+1}$ is defined to be the collection of all subsets of $V_\alpha$.
\item[4.] The process is endless in a strong way: after infinitely many stages have passed, there is another stage after that, and so on: no phenomenon (such as ``infinitely many stages having passed'') is enough to make the process stop. (A process which continues after infinitely many stages have passed is called a \emph{transfinite process}.) The collection of all sets that exist at a ``limit stage'' $\alpha$ is simply the union of the collection of sets that have existed before that stage, i.e. $V_\alpha$ is defined to be the collection of sets that are in some collection of the form $V_\beta$, where $\beta$ is a stage preceding $\alpha$.
\end{itemize}
\end{intpict}
This picture depends on two key concepts: the notion of a ``subset'' and the notion of an ``endless transfinite process''. It may be compared with the picture of the constructible universe given in \6\ref{subsectionDST}, in which each of these concepts is modified somewhat (specifically, ``subset'' becomes ``logical subset'' and ``endless transfinite process'' becomes ``transfinite process that goes on as long as it can remain descriptionalist'').
The notion of a ``subset'' appears to be at least somewhat more coherent than the naive notion of a ``set''. In particular, it is exempt from criticism based on Russell's paradox, on the incoherence of the distinction between sets and classes, or on the necessity to precisely define a category of objects before one can precisely talk about sets that contain objects from this category. However, one may still object that the intuitive notion of a set or subset is usually formed by example from things like lists of objects or descriptions of categories of objects, but according to standard set realism neither of those things forms a prototypical example of a set.\Footnote{For example, a set realist would normally hold that only ``countably many'' sets can be described.} To put it another way, one might define a subset of a category to be a way of choosing some members of the category for inclusion and others for exclusion, but it is not clear that we can talk about ``choosing'' independent of our methods for making choices, and our methods for making precise choices generally involve descriptions or algorithms of some sort.
Regarding the notion of an ``endless transfinite process'', the obvious criticism is that the intuitive notion of a ``process'' clearly excludes the idea that there could be a new stage after infinitely many stages have passed, so that the notion of a ``transfinite process'' already does not make sense. After all, there is no point in time when infinitely many stages will have passed; infinity is only an abstraction. The main responses to this criticism seem to be: firstly, that dealing with transfinite processes becomes intuitive after you deal with them for a while; and secondly, that many parts of the theory of transfinite processes can be encoded in mathematics which is less controversial (see in particular \6\ref{subsectionDST} above). Note that it is possible to accept the notion of a transfinite process while rejecting the notion of an abstract subset, see \6\ref{subsectionDST}.
In addition, in this paper I\ will present a further criticism of the idea of an \emph{endless} process: not that such a process is nonsensical, but that \emph{universal quantification over the outputs of such a process is imprecise or nonsensical}. The reason for this is that at any given point in time, the process is only partially complete and so it makes no sense to quantify over all objects that the process will eventually produce, only the ones that it has already produced. It is possible for a set realist to accept this criticism while remaning a set realist, leading to a view I\ will call \emph{gradualist set realism}. However, accepting the criticism leads to the conclusion that the language of classical set theory \textbf{\textup{ST}}\ (with unbounded quantifiers) does not have precise semantics.
Finally, let us briefly mention one argument used in favor of set realism which does not itself specify a semantics to be used in the interpretation of \textbf{\textup{ST}}: the \emph{indispensability argument} which alleges that since modern mathematics is indispensable for modern science, which can be empirically verified, modern mathematics and in particular set theory has been empirically verified. However, there is little reason to think that this is anything more than a historical accident. Namely, the fact that mathematicians make sure that their arguments can be formalized in \textbf{\textup{ZFC}}\ is not necessarily the basis of the conceptual validity of their arguments, which are ordinarily conceived through intuitive reasoning. Verifying that arguments can be formalized in \textbf{\textup{ZFC}}, though useful, may be nothing more than a double-check. In particular, there seems to be a consensus among mathematical logicians that a great deal of mathematical reasoning can be formalized in languages and axiom systems much weaker than \textbf{\textup{ST}}\ and \textbf{\textup{ZFC}}. For example, Harvey Friedman's ``grand conjecture'' claims that ``Every theorem published in the Annals of Mathematics [that can be translated into \textbf{\textup{NT}}] can be proved in \textbf{\textup{EFA}}''.\Footnote{\url{http://cs.nyu.edu/pipermail/fom/1999-April/003014.html}. In support of his conjecture, Friedman claims that ``in every case where a logician has taken the time to learn the proofs, that logician also [has seen] how to prove the theorem in Peano Arithmetic.''} Here \textbf{\textup{EFA}}\ denotes ``Elementary Function Arithmetic'', an axiomatization of \textbf{\textup{NT}}\ that is much weaker than the standard axiomatization \textbf{\textup{PA}}. Simlarly, Solomon Feferman has claimed that ``surprisingly meager (in the proof-theoretical sense) predicatively justified systems suffice for the direct formalization of almost all, if not all, scientifically applicable mathematics'' \cite[p.443]{Feferman5}, and Geoffrey Hellman has said that ``whether, indeed, there are \emph{any} serious limitations to predicativist mathematics' power to recover scientifically applicable mathematics remains an important open question'' \cite[p.462]{Hellman}. Here, \emph{predicativism} is a type of number realism that allows for fairly powerful languages and axiom systems.\\
{\it Strongest axiom contest for the language of set theory \textbf{\textup{ST}}.} The obvious starting point for the strongest axiom contest for \textbf{\textup{ST}}\ is its standard axiomatization \textbf{\textup{ZFC}}. For an example of a philosophical defense of \textbf{\textup{ZFC}}\ from a set realist perspective, see \cite{Maddy1}. Essentially, \textbf{\textup{ZFC}}\ formalizes many common intuitions about the ``universe of sets'', as described either by the naive view or in the iterative picture. However, \textbf{\textup{ZFC}}\ does not capture all possible intuitions about set theory, and there is no particular reason for set realists to limit themselves to \textbf{\textup{ZFC}}\ rather than trying to axiomatize other intuitions. In particular, in the iterative picture the phrase `all subsets' in (3), and the phrase `no phenomenon is enough to make the process stop' in (4), can always generate more intuitions, which we can then attempt to formalize.
A straightforward example is the \emph{axiom of inaccessible cardinals}. Roughly speaking, this axiom says that there is some stage $\alpha$ we can stop at such that the resulting universe $V_\alpha$ is closed under the rules for producing sets described by the axiom system \textbf{\textup{ZFC}}.\Footnote{The phrase `rules for producing sets described by the axiom system \textbf{\textup{ZFC}}' is not taken literally, since the axiom schemas of \textbf{\textup{ZFC}}\ (separation and replacement) can only be instantiated for definable classes, and the production rules used to define inaccessible cardinals are taken to apply to all classes (i.e. subsets of $V_\alpha$), regardless of whether or not they are definable.} This axiom can be thought of as a formalization of the principle that ``no phenomenon is enough to make the process stop'', since it hypothesizes that the phenomenon of having used all of the production rules of \textbf{\textup{ZFC}}\ as much as we possibly can is not enough to make the process stop.
Though there are more \emph{large cardinal axioms} for which it is possible to give a strong intuitive argument or ``intrinsic justification'' using set realism as a starting assumption,\Footnote{Opinion varies on exactly which axioms can count as ``intrinsically justified''. Some say that axioms as strong as the axiom of remarkable cardinals are justified by intuition \cite[Theorem 3.3]{McCallum}, while I\ do not really see how even the axiom of Mahlo cardinals can be made plausible by appealing to intuition (the argument given in \cite{McCallum} is not very convincing since the number of possible values of a second-order variable in $V_\kappa$ increases dramatically as $\kappa$ increases, so there is no clear reason why we would expect anything to stabilize over all possible values of such a variable).} eventually we get to axioms which philosophers have attempted to give ``extrinsic justifications'' based on data other than intuition.\Footnote{Though popular, the labels `intrinsic' and `extrinsic' seem wrong to me. Namely, human intuition is not something that is intrinsic to any language; it is conceptually independent. In other words, to define a language, it is not necessary to define all possible intuitions that one may have about the language.} For example, when attempting to justify an axiom philosophers may cite as evidence ``confirmation by instances, prediction, providing new proofs of old theorems, unifying new results with old, extending patterns begun in weaker theories, providing powerful new ways of solving old problems, filling a gap in a previously conjectured `false, but natural' proof, explanatory power, and intertheoretic connections'' \cite[\6VII]{Maddy2}.\Footnote{It is interesting to note that almost all of the items on this list are the kind of evidence used by \emph{abductive reasoning}, or reasoning from consequences. If we view intuitive arguments as \emph{deductive reasoning}, then it appears that what is missing is \emph{inductive reasoning}, or reasoning based on expectations. While inductive reasoning is often used in mathematics in the form of probabilistic heuristic arguments (such as the standard argument justifying the conjecture that $\sqrt 2$ has equal proportions of all digits in its decimal expansion), it is interesting to ask whether there are any large cardinal axioms that can be made more plausible by inductive reasoning.}
While such considerations are certainly relevant in any discussion of the epistemology of mathematics, there does seem to be a relatively clear conceptual distinction between intuitively plausible axioms and hypotheses made plausible in other ways, though in any given case it may be unclear where the line lies. Note that large cardinal axioms whose standard justifications are ``extrinsic'' tend to be far more powerful than those whose standard justifications are intuitive. It appears to be an interesting question whether the same will hold for number realist languages
Large cardinal axioms, i.e. axioms that hypothesize the existence of stages in the iterative picture with certain properties, are conventionally ordered according to their \emph{consistency strength}. An axiom system $A$ is said to be \emph{consistent} if no claim of the form $P\wedge\neg P$ (i.e. ``the claim $P$ is both true and false'', a \emph{contradiction}) is syntactically provable in $A$. The notion of consistency is particularly important in mathematics because it can be formalized in fairly weak languages such as \textbf{\textup{NT}}.\Footnote{From listening to some mathematicians, one could get the impression that this formalizability property means that syntactic consistency is a more primitive notion than truth. One of the goals of this paper is to argue the opposite: that languages and their semantics are more conceptually primitive than axiom systems and their corresponding notions of syntactic provability.} Since no sentence\ can be both true and false, every valid axiom system is consistent, but the converse is not true. The axiom system $A_1$ is said to have \emph{greater\Footnote{In mathematics the word `greater' is often used in an inclusive sense, i.e. as shorthand for `greater or equal'.} consistency strength} than the axiom system $A_2$ if the consistency of $A_2$ can be proven in some ``neutral'' axiom system (known as the \emph{metatheory}) using the hypothesis that $A_1$ is consistent. The selection of a metatheory is important because otherwise all consistent axiom systems would be equally consistent.\Footnote{Axiom systems could still be ranked according to how plausible the philosophical arguments for their consistency are, but the idea of such a ranking has a decidedly non-mathematical flavor. Reference to a metatheory is also important because of my\ view that sentences of the form `if $P$, then $Q$' can only be thought of as claims or sentences\ (rather than merely intelligible utterances) to the extent that modes of reasoning with respect to which it is being claimed that $Q$ can be inferred from $P$ can be understood from context. This view will be explained further in Appendix \ref{appendiximplication}.} There is nothing to rule out choosing one of the systems $A_1$ or $A_2$ as the metatheory, but usually a ``weaker'' axiom system such as \textbf{\textup{PA}}\ or even \textbf{\textup{PRA}}\ (Primitive Recursive Arithmetic) is chosen. Note that the ordering of axiom systems by consistency strength is essentially the reverse of the ordering by plausibility: an axiom system $A_1$ with greater consistency strength tends to be less plausible than an axiom system $A_2$ with lesser consistency strength, since if both $A_1$ and the ``neutral'' axiom system are valid, then $A_2$ is consistent and therefore plausibly valid; so the plausibility of $A_1$ feeds into the plausibility of $A_2$ but not vice-versa
However, the ordering of axiom systems by consistency strength is not the appropriate one to make comparisons between their corresponding axiomatic strategies in the Busy Beaver contest and the largest number contests of selfmeta languages, according to Corollaries \ref{corollaryaxiomstrat} and \ref{corollaryselfmeta}. In fact, no axiomatization of \textbf{\textup{ST}}\ is stronger than any other according to the criterion specified there, since the language \textbf{\textup{ST}}\ is incapable of reproducing its own metalanguage, and thus if $A_1$ and $A_2$ are axiomatizations of \textbf{\textup{ST}}, it is not even possible to ask the question of whether $A_1$ is stronger than the metasystem $A_2 + 1$, since these axiom systems are not formulated in the same language. This is a hint that the classical language \textbf{\textup{ST}}\ is not a good language for metamathematics for any form of set realism, a point which I\ will argue further below. For now, let us note that a partial response based on classical mathematics is that a slightly stronger notion than consistency suffices to compare axiom systems well enough to determine which axiomatic strategies will do better at the Busy Beaver contest, namely the notion of \emph{$\omega$-consistency}. The notion of $\omega$-consistency depends on the notion of an $\omega$-syntactic proof, which is essentially a syntactic proof in which ``infinite but countable'' chains of reasoning are allowed. An axiom system is said to be $\omega$-consistent if no contradiction is $\omega$-syntactically provable. We have the following:
\begin{theorem}
\label{theoremaxiomstrat3}
Let $A_1$ and $A_2$ be two $\omega$-consistent axiomatizations of \textbf{\textup{ST}}, and suppose that $A_1$ can prove that $A_2$ is $\omega$-consistent. Then the axiomatic strategy of $A_1$ in the Busy Beaver contest wins against the axiomatic strategy of $A_2$.
\end{theorem}
In my\ opinion, this theorem is not very natural because the notion of $\omega$-consistency is not conceptually simple (in comparison to the notion of a language that has the ability to talk about another language's semantics), and because it only applies to the Busy Beaver contest and not to the largest number contests of selfmeta languages, for which other concepts would be needed. However, it does allow the comparison of the axiomatic strategies of various large cardinal axioms. I\ leave as an exercise to mathematical logicians the question of whether the standard ordering of large cardinal axioms according to consistency strength is the same as the ordering according to the strength of their axiomatic strategies in the Busy Beaver contest as well as in other contests possessing axiomatic strategies.\\
{\it Largest number contest run by set realists.} Even if we are satisfied with the view that in the end, $\omega$-consistency or some other constructed criterion is to be used for judging the strength of axiom systems rather than their raw ability to formalize reasoning about each other, when we consider the largest \emph{number} contest rather than the largest \emph{axiom} contest, things get more difficult for a semi-naive version of set realism. This is because the most obvious way for a semi-naive set realist to make precise his\ largest number contest is to reduce it to the largest number contest of the language \textbf{\textup{ST}}. But this contest has a brute force strategy! To highlight the absurdity, we ask the set realist if he\ is really going to reject the entry `the number referred to by the term $\double{\phi_\textbf{\textup{ST}}}{10^{100}}$, where $\phi_\textbf{\textup{ST}}$ is as in Theorem \ref{theoremrecstrat}' to their contest on the grounds that it is not technically a term of \textbf{\textup{ST}}. This is true, but surely the entry is almost as good as a term of \textbf{\textup{ST}}, since it can be converted into one given enough time and computational power. Moreover, ordinary mathematical discourse is not directly framed in \textbf{\textup{ST}}, but the idea is that there is some method, often partly algorithmic, via which this discourse can be translated into \textbf{\textup{ST}}. So the idea that an algorithm that can generate a term of \textbf{\textup{ST}}\ is almost as good as a term of \textbf{\textup{ST}}\ seems to be embedded in the very idea of interpreting standard mathematical reasoning as reasoning in \textbf{\textup{ST}}
Once the set realist admits that the informal description `my entry is the term $\double{\phi_\textbf{\textup{ST}}}{10^{100}}$' is a valid entry, things really start to get interesting. If he\ allows the use of an algorithm for generating terms of \textbf{\textup{ST}}, then he\ has to address the question of what kind of algorithms are allowed. He\ could of course require the algorithms to belong to a small class like the class of primitive recursive algorithms, but this would be somewhat arbitrary and would also yield a degenerate largest number contest. Similar remarks apply to the class of algorithms that can be proven to terminate in some fixed axiom system like \textbf{\textup{ZFC}}. A better option is to require only a semantic proof that the algorithm described in an entry halts, such as a philosophical argument for some large cardinal axiom together with the stipulation that the axiomatic strategy for the Busy Beaver contest should be used. This would at least gets him\ away from the realm of degenerate largest number contests and into the realm where axiomatic strategies are appropriate
However, stopping at algorithms seems to indicate that the key thing about them is the fact that they \emph{can be used to compute} strings in \textbf{\textup{ST}}, rather than just the fact that they \emph{uniquely identify} strings in \textbf{\textup{ST}}. Such an emphasis seems a little inappropriate for set realists, who are usually uninterested in whether objects can be computed as long as they can be proven to exist. Thus, it seems likely that many set realists will take the point of view that any description that uniquely identifies a string in \textbf{\textup{ST}}\ is almost as good as the string it identifies, and that talking about ``the number named by the \textbf{\textup{ST}}-term named by the \textbf{\textup{ST}}-term $t$'' (where $t$ is a concrete \textbf{\textup{ST}}-term) is just as precise as talking about ``the number named by the \textbf{\textup{ST}}-term $t$''.
These considerations suggest that a set realist who accepts \textbf{\textup{ST}}\ should also accept its metalanguage $\textbf{\textup{ST}}+1$, which is capable of transcribing the above terms. Applying the same argument, the set realist should also accept $\textbf{\textup{ST}}+2$, and so on, all the way up to the Kripke extension $\K(\textbf{\textup{ST}})$. This reinforces our earlier conclusion in Section \ref{sectionkripke} that any philosopher who accepts the language \textbf{\textup{ST}}\ should also accept the language \K(\textbf{\textup{ST}}).
However, a set realist need not accept \textbf{\textup{ST}}\ as a precise language at all. He\ may instead say that this analysis shows that the idea of allowing arbitrary quantification over the outputs of a so-called ``endless'' process is incoherent. In other words, whenever we start to talk about the objects that have been created by the process up to some point, then our ability to meaningfully talk about these objects depends on the existence of larger objects delineating truths from falsehoods, meaning that the process continues past the collection of objects that we are currently talking about. In this view, which we will call \emph{gradualist set realism}, it is incoherent to claim that ``for all sets $x$ and $y$, there is a set $\{x,y\}$'', since the truth of such a sentence\ would depend on facts about all sets, and at no time do all sets exist, so at no time can something be true about all of them. Instead, the Axiom of Pair should be thought of merely as a description of a process for creating sets from other sets, but not as a claim that would have been precise if we didn't already know that the process exists.
It turns out to be possible use the principles of gradualist set realism to define a language similar to \textbf{\textup{ST}}, which we will denote by \textbf{\textup{GST}}\ (for \emph{gradualist set theory}), which does not allow unbounded universal quantification; see Appendix \ref{appendixPT} for details. Bounded universal quantification is allowed, because the quantification ranges over the members of a set which already exists, so that it makes sense to talk about all of the members of the set simultaneously. The language \textbf{\textup{GST}}\ has the property that it is \emph{selfmeta}, meaning that it is capable of reproducing its own metalanguage. Note that classical languages cannot be selfmeta due to Tarski's theorem on the undefinability of truth; the key here is that the language is not fundamentally based on the classical (unbounded) predicate calculus.
Although the language \textbf{\textup{GST}}\ initially appears to be incapable of formalizing classical mathematics (which uses reasoning that is supposed to be formalizable in \textbf{\textup{ZFC}}), it can in fact reproduce a language with respect to which \textbf{\textup{ZFC}}\ is usually presumed to be valid: the language with the syntax of \textbf{\textup{ST}}\ whose semantics is given by interpreting the word `set' to range over all members of $V_\kappa$, where $\kappa$ is a concretely specified inaccessible cardinal (for example, the first inaccessible cardinal). we will denote this language by $\textbf{\textup{ST}}_\kappa$.\Footnote{Note that gradualist set realism appears to be incompatible with classical set realism, since the languages $\textbf{\textup{ST}}_\kappa$ all have different semantics and so it appears that the language \textbf{\textup{ST}}\ does not have any canonical semantics.} Note that \textbf{\textup{GST}}\ must have primitive language for talking about large cardinals with a specified property, such as the first inaccessible cardinal, in order for this to work.
Since \textbf{\textup{GST}}\ is selfmeta, there are no obvious philosophical arguments showing that someone who accepts that \textbf{\textup{GST}}\ has precise semantics should also accept that some stronger language has precise semantics. Namely, the argument that anyone who accepts a language should also accept its metalanguage is irrelevant, because someone who accepts \textbf{\textup{GST}}\ has already accepted the metalanguage of \textbf{\textup{GST}}, since it is reproducible in \textbf{\textup{GST}}. Thus, \textbf{\textup{GST}}\ can be viewed as an ``ultimate ontology'', and can thereby used to convert the imprecise ``philosophical largest number contest'' into a precise largest number contest, namely the largest number contest of \textbf{\textup{GST}}\ (cf. Contest definitionPT \ref{lnc1}).
The obvious strategy for this largest number contest is the \emph{linguistic strategy} mentioned just before Theorem \ref{theoremlingstrat}: first describe a large cardinal $\kappa$ using \textbf{\textup{GST}}, and then use the brute force strategy on $\textbf{\textup{ST}}_\kappa$. In order for this to work we need to make sure that our description of $\kappa$ in \textbf{\textup{GST}}\ is in fact a valid description, and for this purpose we need to make some kind of argument. Arguments tend to be best made in some axiom system, which we can expect to be an axiomatization of \textbf{\textup{GST}}
Theorem \ref{theoremselfmeta} shows that the Philosophical Largest Number Contest for gradualist set realists reduces to the Philosophical Strongest Axiom Contest. At this point it will be useful to revisit our analysis of the strongest axiom contest for set realists, this time under the gradualist assumption. Since \textbf{\textup{GST}}\ is selfmeta, it is legitimate to compare large cardinal axioms on the basis of whether one axiom system is more powerful than the metasystem of another axiom system. However, there is the issue that not all large cardinal axioms can be translated into \textbf{\textup{GST}}, since some are stated using unbounded quantifiers. Though more powerful large cardinal axioms tend to be more likely to use unbounded quantifiers, this relationship is not straightforward and the strongest large cardinal axiom not currently known to be inconsistent with \textbf{\textup{ZFC}}, i.e. the rank-into-rank Axiom {\bf I0}, is formulated using only bounded quantifiers.
If $\phi$ is a large cardinal axiom formulated using unbounded quantifiers, then it is still possible to translate into \textbf{\textup{GST}}\ the claim that $\phi$ is true in some universe: namely, this claim can be translated as `$\exists \kappa \text{ inaccessible such that $\phi^\kappa$ holds}$', where $\phi^\kappa$ denotes the result of replacing all quantifiers in $\phi$ by quantifiers over the universe $V_\kappa$. We will call this claim the ``existence of a universe satisfying $\phi$''. Thus, in addition to comparing the strength of various large cardinal axioms with bounded quantifiers, one may also compare the strength of the assumption that there exists a universe satisfying a given large cardinal axiom with unbounded quantifiers. Note that if $\phi$ is an axiom formulated in \textbf{\textup{GST}}, then the existence of a universe satisfying $\phi$ is a stronger claim than $\phi$; in fact, it is stronger than the metasystem of $\phi$, since \textbf{\textup{GST}}\ can talk about the notion of truth within any fixed universe. So to show that an axiom system $A_1$ is stronger than the metasystem of another axiom system $A_2$, it suffices to show that $A_1$ can prove the existence of a universe satisfying $A_2$.
Again, we leave the details of exactly which large cardinal axioms are the strongest according to this criterion as an exercise for the interested mathematician. (We expect that the axiom systems we give below for dealing with truth should suffice to answer this question, so the question can be attacked from a purely formalist perspective.)
\ignore{
\begin{remark}
This way of thinking makes it completely unsurprising that, in the words of Wikipedia, ``There is no generally agreed precise definition of what a large cardinal property is'' -- if there was such a precise definition, it could probably be made into a large cardinal property itself.
\end{remark}
\begin{example}
Consider the following program to output a really big number:\\
/\;\; Let $\alpha = \emptyset$;\\
/\;\; LABEL 1;\\
/\;\; If $|\alpha| < 10^{10}$(\\
/\;\;\;\; Let $\alpha = \text{``Compute $T(\alpha)$, then add $1$''}$;\\
/\;\;\;\; GOTO 1;)\\
/\;\; Else(\\
/\;\;\;\; For $n = 1,\ldots,10^{10} - 10,000$(\\
/\;\;\;\;\;\; For $\beta$ of length $n$(\\
/\;\;\;\;\;\;\;\; Allocate Graham's number of cycles to compute $T(\beta)$, and return the result as $\gamma$;\\
/\;\;\;\;\;\;\;\; If $\gamma = \alpha$(\\
/\;\;\;\;\;\;\;\;\;\; Let $\alpha = \text{``Compute $T(T(\beta))$''}$;\\
/\;\;\;\;\;\;\;\;\;\; GOTO 1;)\\
/\;\;\;\;\;\;\;\; )));\\
/\;\;Return $T(\alpha)$\\
Let's call this program ``Solomonoff induction''. It might do really well or really badly, depending on circumstances -- it's not clear. But the funny thing is that \emph{computations like \eqref{solomonoff} seem to be more or less an attempt to imitate Solomonoff induction ourselves} -- with a bit more cleverness in that we have the idea of ``a trick which worked already will work again'' that Solomonoff induction doesn't have -- nevertheless, this trick idea is not actually always true, so it makes sense that a computer should not implement it.
Also, it seems to make a big difference whether the computer thinks of the solution
\[
\text{Add $10^10$ ones, then add $10^{10} - 100$ ones, then add $10^{10} - 200$ ones, then add...}
\]
or
\[
\text{Add ones until the string length is $10^{10}$, then add ones until the string length is $10^{10}$, then add...}
\]
first, and there doesn't seem to be much opportunity to correct later: suppose that each ``add $X$ ones'' has a different way of expressing $X$, as an arbitrary consequence of how fast the code can be collapsed. E.g. $10^{10}$ can be written just ``$10^{10}$, but other numbers will need explicit decimal or binary expansions. The way we set it up, Solomonoff basically has to predict what the pattern behind these are in order to get anywhere, which is... backwards.
Are we subject to the same limitation somewhere, in a way which we do not see?
Conclusion: Although Solomonoff induction sounds good theoretically, it doesn't capture all parts of human intuition, such as: There are abstract ``numbers'', and we can write them however we want, in particular we can write them in whichever way will make there be a pattern. And: The information needs to be stored in a way that doesn't depend too sensitively on the arbitrary cutoff. Is it possible to solve these problems? It seems sort of like creating an AI. One suggestion would be to tell the computer that it has to use a specific symbol to represent the number $10^{10}$, and that when the program is executed on slightly different values of that symbol, then slightly different length strings result (describing the ``rate of change'' of $\alpha$ in the last few steps before we started the induction step). Does this work? It's a technical question... One big problem is that you may get something like:
\begin{quote}
``Iterate $(x x)$ a thousand times, then iterate $(x x)$ 1200 times, then iterate $(x x)$ 991 times, then apply $(x x)$ $(x x)$ $(x x)$ \ldots (a thousand times) $(x x)$ $xxxxxx\ldots$ (to $10^{10}$)''
\end{quote}
\end{example}
\ignore{
\section{Formalizing the entire paper}
Sorts used in this paper:
\begin{itemize}
\item categories
\item members of categories (objects)
\item languages
\item axiom systems
\item algorithms
\item utterances (terms, sentences, predicates, definitions, utterances)
\item numbers
\end{itemize}
Non-mathematical:
\begin{itemize}
\item humans
\item thought processes / reasoning
\item concepts
\end{itemize}
Things that they can do:
\begin{itemize}
\item terms can denote objects
\item humans can say, give, write
\end{itemize}
Things that they can be:
\begin{itemize}
\item sentences\ can be true or false, precise or imprecise, coherent or incoherent
\item languages and utterances can be precise or imprecise
\item categories can be complete or incomplete
\item reasoning can be valid/precise or invalid/imprecise
\end{itemize}
This is not very informative... More useful are the axioms:
\begin{itemize}
\item Let $C$ be a complete category and let $P$ be a property. Then the phrase `all members of $C$ have property $P$' is true if and only if all members of $C$ have property $P$.
\item Let $C$ be a category and let $P$ be a property. Then the phrase `there exists a member of $C$ with property $P$' is true if and only if there exists a member of $C$ with property $P$.
\end{itemize}
\newpage
|
\section{Introduction}
An interesting problem, with many practical applications, is to study limit
theorems for processes conditioned to start from a fixed past trajectory. This
problem is difficult, since the stationary processes started from a fixed past
trajectory, or from a point, are no longer stationary. Furthermore, the
validity of a limit theorem is not enough to assure that the convergence still
holds when the process is not started from its equilibrium. This type of
convergence is also known under the name of almost sure conditional limit
theorem or the quenched limit theorem.
The issue of the quenched CLT for stationary processes has been widely
explored for the last few decades.
Among many others, we mention papers by Derriennic and Lin (2001), Cuny and
Peligrad (2012), Cuny and Voln\'{y} (2013), Cuny and Merlev\`{e}de (2014),
Voln\'{y} and Woodroofe (2014), Barrera et al. (2016). Some of these results
were surveyed in Peligrad (2015).\newline
A random field consists of multi-indexed random variables $(X_{\mathbf{u
})_{\mathbf{u}\in Z^{d}},$ where $d$ is a positive integer. The main
difficulty when analyzing the asymptotic properties of random fields, is the
fact that the future and the past do not have a unique interpretation. To
compensate for the lack of ordering of the filtration, it is customary to use
the notion of commuting filtrations. Traditionally, this kind of filtration is
constructed based on random fields which are functions of independent and
identically distributed random variables. Alternatively, commuting filtrations
can be induced by stationary random fields with independent columns or rows.
See for example, El Machkouri et al. (2013) and Peligrad and Zhang (2018a). As
in the case of random processes, a fruitful approach for proving limit
theorems for random fields is via the martingale approximation method, which
was started by Rosenblatt (1972) and its development is still in progress.
Recently, the interest is in the approximation by ortho-martingales which were
introduced by Cairoli (1969).
We would like to mention several important recent contributions in this
direction by Gordin (2009), Voln\'{y} and Wang (2014), Voln\'{y} (2015), Cuny
et al. (2015), Peligrad and Zhang (2018a), Giraudo (2017) and Peligrad and
Zhang (2018b).
However, the corresponding quenched version of these results have rarely been
explored. To the best of our knowledge, so far, the only quenched invariance
principle for random fields is due to Peligrad and Voln\'{y} (2018). Their
paper contains a quenched functional CLT for ortho-martingales and a quenched
functional CLT for random fields via co-boundary decomposition. By
constructing an example of an ortho-martingale which satisfies the CLT but not
its quenched form, Peligrad and Voln\'{y} (2018) showed that, contrary with
the one dimensional index set, the finite second moment condition is not
enough for the quenched CLT. For the validity of this type of results, they
provided a minimal moment condition, that is: $E\left( X_{\mathbf{0}}^{2}
\log^{d-1}(1+|X_{\mathbf{0}}|)\right) <\infty$, where $\mathbf{0
=(0,\cdots,0)\in Z^{d}$ and $d$ is the dimension. \newline
Here, we aim to establish sufficient conditions in terms of projective
criteria such that a quenched CLT holds. One of the results of this paper is a
natural extension of the quenched CLT for ortho-martingales in Peligrad and
Voln\'{y} (2018) to more general random fields under the generalized Hannan
projective condition (1973). Our result is also a quenched version of the main
theorem in Peligrad and Zhang (2018a). The functional form of a quenched CLT
that we shall use in our applications will also be explored in this paper. The
tools for proving these results consist of ortho-martingale approximations,
projective decompositions and ergodic theorems for Dunford-Schwartz
operators.\newline
Our paper is organized as follows. In the next section, we introduce the
preliminaries and our main results for double-indexed random fields. In
Section 3, we prove the quench CLT's for double-indexed random fields.
Extensions to general indexed random fields and their proofs are given in
Section 4. Section 5 contains a functional CLT\ which will be used in
applications. In Section 6, we apply our results to linear and Volterra random
fields with independent innovations, which are often encountered in economics.
For the convenience of the reader, in the Appendix, we provide a well-known
inequality for martingales and an important theorem in decoupling theory which
will be of great importance for the proof of our main results.
\section{Preliminaries and Results}
For the sake of clarity, especially due to the complicated notation, in this
section, we shall only talk about the double-indexed random fields. After
obtaining results for double-indexed random fields, we will extend them to
random fields indexed by $Z^{d},d>2$. We shall introduce first a stationary
random field adapted to a stationary filtration. In order to construct a
flexible filtration it is customary to start with a stationary real valued
random field $(\xi_{n,m})_{n,m\in Z}$ defined on a probability space
$(\Omega,\mathcal{K},P)$ and\ define the filtration
\begin{equation}
\mathcal{F}_{k,\ell}=\sigma(\xi_{j,u}:j\leq k,\text{ }u\leq\ell).
\label{def fitration
\end{equation}
For all $i,j\in Z$, we also define the following sigma algebras generated by
the union of sigma algebras: $\mathcal{F}_{\infty, j}=\vee_{n\in Z
\mathcal{F}_{n,j}$, $\mathcal{F}_{i,\infty}=\vee_{m\in Z}\mathcal{F}_{i,m}$
and $\mathcal{F}_{\infty,\infty}=\vee_{i, j\in Z}\mathcal{F}_{i,j}$. \newline
To ease the notation, sometimes the conditional expectation will be denoted b
\[
E_{a,b}X=E(X|\mathcal{F}_{a,b}).
\]
In addition we consider that the filtration is commuting in the sense that
\begin{equation}
E_{u,v}E_{a,b}X=E_{a\wedge u,b\wedge v}X, \mathbb{\ } \label{pcf
\end{equation}
where the symbol $a\wedge b$ stands for the minimum between $a$ and $b$. As we
mentioned before, this type of filtration is induced, for instance, by an
initial random field $(\xi_{n,m})_{n,m\in Z}$ of independent random variables
or more generally can be induced by stationary random fields $(\xi
_{n,m})_{n,m\in Z}$ where only the columns are independent, i.e. $\bar{\eta
}_{m}=(\xi_{n,m})_{n\in Z}$ are independent. This model often appears in
statistical applications when one deals with repeated realizations of a
stationary sequence.
It is interesting to point out that commuting filtrations can be described by
the equivalent formulation: for $a\geq u$ we have
\[
E_{u,v}E_{a,b}X=E_{u,b\wedge v}X.
\]
This follows from the Markovian-type property (see for instance Problem 34.11
in Billingsley, 1995).\newline
Without restricting the generality we shall define $(\mathbf{\xi}_{\mathbf{u
})_{\mathbf{u}\in Z^{2}}$ in a canonical way on the probability space $\Omega$
$=R^{Z^{2}}$, endowed with the $\sigma-$field, $\mathcal{B}(\Omega),$
generated by cylinders.
Now on $R^{Z^{2}}$ we shall introduce the operator
\[
T^{\mathbf{u}}((x_{\mathbf{v}})_{\mathbf{v}\in Z^{2}})=(x_{\mathbf{v+u
})_{\mathbf{v}\in Z^{2}}.
\]
Two of them will play an important role in our paper namely, when
$\mathbf{u=}(1,0)$ and when $\mathbf{u=}(0,1).$ By interpreting the indexes as
notations for the lines and columns of a matrix, we shall cal
\[
T((x_{u,v})_{(u,v)\in Z^{2}})=(x_{u+1,v})_{(u,v)\in Z^{2}
\]
the vertical shift an
\[
S((x_{u,v})_{(u,v)\in Z^{2}})=(x_{u,v+1})_{(u,v)\in Z^{2}
\]
the horizontal shift.\newline
Now we introduce the stationary random field $(X_{\mathbf{m}})_{\mathbf{m}\in
Z^{2}}$ in the following way. For a real-valued measurable function $f$ on
$R^{N^{2}}$, we define
\begin{equation}
X_{j,k}=f(T^{j}S^{k}(\mathbf{\xi}_{a,b})_{a\leq0, b\leq0}). \label{defXfield
\end{equation}
The variable $X_{0,0}$ will be assumed to be square integrable (in $L^{2}$)
and with mean $0.$ We notice that the variables $(X_{n,m})_{n,m\in Z}$ are
adapted to the filtration $(\mathcal{F}_{n,m})_{n,m\in Z}$. \newline
Let $\phi:[0,\infty)\rightarrow\lbrack0,\infty)$ be a Young function, that is,
a convex function satisfying
\[
\lim_{x\rightarrow0}\frac{\phi(x)}{x}=0\text{ and }\lim_{x\rightarrow\infty
}\frac{\phi(x)}{x}=\infty.
\]
We shall define the Luxemburg norm associated with $\phi$ which will be needed
in the sequel. For any measurable function $f$ from $\Omega$ to $R$, the
Luxemburg norm of $f$ is defined by (see relation 9.18 and 9.19 on page 79 of
Krasnosel'skii and Rutitskii (1961))
\begin{equation}
\lVert f \rVert_{\phi}=\inf\{{k\in(0,\infty)}:E\phi(|f|/k)\leq1\}.
\label{phi norm
\end{equation}
In the sequel, we use the notation
\[
S_{k,j}=\sum\nolimits_{u,v=1}^{k,j}X_{u,v},\ P^{\omega}(\cdot)=P(\cdot
|\mathcal{F}_{0,0})({\omega}) \text{ for any } \omega\in\Omega.
\]
Also, we shall denote by $E^{\omega}$ the expectation corresponding to
$P^{\omega}$ and $\ \Rightarrow$ the convergence in distribution.
For an integrable random variable $X$, we introduce the projection operators
defined b
\[
P_{\tilde{0},0}(X):=(E_{0,0}-E_{-1,0})(X)
\
\[
P_{0,\tilde{0}}(X):=(E_{0,0}-E_{0,-1})(X).
\]
Note that, by (\ref{pcf}), we hav
\[
{\mathcal{P}}_{{\mathbf{0}}}(X):=P_{\tilde{0},0}\circ P_{0,\tilde{0
}(X)=P_{0,\tilde{0}}\circ P_{\tilde{0},0}(X)=(E_{0,0}-E_{0,-1}-E_{-1,0
+E_{-1,-1})(X).
\]
Then for $(u,v)\in Z^{2}$, we can define the projections $\mathcal{P}_{u,v}$
as follows
\[
\mathcal{P}_{u,v}(\cdot):=(E_{u,v}-E_{u,v-1}-E_{u-1,v}+E_{u-1,v-1})(\cdot)
\]
We shall introduce the definition of an ortho-martingale, which will be
referred to as a martingale with multiple indexes or simply martingale.
\begin{definition}
Let $d$ be a function and define
\begin{equation}
D_{n,m}=d(\xi_{i,j},i\leq n,j\leq m). \label{defD
\end{equation}
Assume integrability. We say that $(D_{n,m})_{n,m\in Z}$ is a field of
martingale differences if $E_{a,b}(D_{n,m})=0$ if either $a<n$ or $b<m.$
\end{definition}
Se
\[
M_{k,j}=\sum\nolimits_{u,v=1}^{k,j}D_{u,v}.
\]
\begin{definition}
We say that a random field $(X_{n,m})_{n,m\in Z}$ defined by (\ref{defXfield})
admits a martingale approximation if there is a field of martingale
differences $(D_{n,m})_{n,m\in Z}$ defined by (\ref{defD})\ such that
\begin{equation}
\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}E^{\omega}(S_{n,m}-M_{n,m})^{2}=0
\text{ for almost all } \ \omega\in\Omega. \label{martapprx
\end{equation}
\end{definition}
The following theorem is an extension of the quenched CLT for
ortho-martingales in Peligrad and Voln\'{y} (2018) to stationary random fields
satisfying the generalized Hannan condition (1973). It also can be viewed as a
random field version of Proposition 11 in Cuny and Peligrad (2012) (see also
Voln\'{y} and Woodroofe (2014)).\newline
Throughout the paper we shall assume the setting above namely:\newline
\textbf{Condition A.} $(X_{n,m})_{n,m\in Z}$ is defined by (\ref{defXfield}),
the filtrations are commuting and either $T$ or $S$ is ergodic.
\begin{theorem}
\label{Thm1nn}Assume Condition A and in additio
\begin{equation}
\sum_{u,v\geq0}\lVert\mathcal{P}_{0,0}(X_{u,v}) \rVert_{2}<\infty.
\label{cond
\end{equation}
Then, for almost all $\omega\in\Omega,
\[
\frac{1}{n}\bar{S}_{n,n}\Rightarrow N(0,\sigma^{2})\text{ under }P^{\omega
}\text{ when }n\rightarrow\infty.
\]
where $\bar{S}_{n,n}=S_{n,n}-R_{n,n}$ with $R_{n,n}=E_{n,0}(S_{n,n
)+E_{0,n}(S_{n,n})-E_{0,0}(S_{n,n})$ and
\[
\sigma^{2}=\lVert\sum_{u,v\geq0}\mathcal{P}_{0,0}(X_{u,v})\rVert_{2}^{2}
=\lim_{n\wedge m\rightarrow\infty}\frac{E(\bar{S}_{n,n}^{2})}{n^{2}}.
\]
$.$
\end{theorem}
In Theorem \ref{Thm1nn} the random centering $R_{n,n}$ cannot be avoided. As a
matter of fact, for $d=1,$ Voln\'{y} and Woodroofe (2010) constructed an
example showing that the CLT\ for partial sums need not be quenched. It should
also be noticed that, for a stationary ortho-martingale, the existence of
finite second moment is not enough for the validity of a quenched CLT when the
summation in taken on rectangles (see Peligrad and Voln\'{y} (2018)). In order
to assure the validity of a martingale approximation with a suitable moment
condition we shall reinforce condition (\ref{cond}) when dealing with indexes
$n$ and $m$ which converge independently to infinity.
\begin{theorem}
\label{Thm2nm} Assume now that (\ref{cond}) is reinforced t
\begin{equation}
\sum_{u,v\geq0}\lVert\mathcal{P}_{0,0}(X_{u,v})\rVert_{\phi}<\infty,
\label{main condi
\end{equation}
where $\phi(x)=x^{2}\log(1+|x|)$ and $\lVert\cdot\rVert_{\phi}$ is defined by
(\ref{phi norm}). Then, for almost all $\omega\in\Omega,$
\begin{equation}
\frac{1}{(nm)^{1/2}}\bar{S}_{n,m}\Rightarrow N(0,\sigma^{2})\text{ under
}P^{\omega}\text{ when }n\wedge m\rightarrow\infty, \label{QCLTnm
\end{equation}
where $\bar{S}_{n,m}=S_{n,m}-R_{n,m}$ with $R_{n,m}=E_{n,0}(S_{n,m
)+E_{0,m}(S_{n,m})-E_{0,0}(S_{n,m})$ and
\[
\sigma^{2}=\lVert\sum_{u,v\geq0}\mathcal{P}_{0,0}(X_{u,v})\rVert_{2}^{2}
=\lim_{n\wedge m\rightarrow\infty}\frac{E(\bar{S}_{n,m}^{2})}{nm}.
\]
\end{theorem}
The random centering is not needed if we impose two regularity conditions.
\begin{corollary}
\label{Cor} Assume that the conditions of Theorem \ref{Thm2nm} hold. If
\begin{equation}
\frac{E_{0,0}\left( E_{0,m}^{2}(S_{n,m})\right) }{nm}\rightarrow0\text{ a.s.
\ }\text{and }\ \frac{E_{0,0}\left( E_{n,0}^{2}(S_{n,m})\right)
{nm}\rightarrow0\text{ a.s. }\text{ when }n\wedge m\rightarrow\infty,
\label{regularity
\end{equation}
then for almost all $\omega\in\Omega,$
\begin{equation}
\frac{1}{(nm)^{1/2}}S_{n,m}\Rightarrow N(0,\sigma^{2})\text{ under }P^{\omega
}\text{ when }n\wedge m\rightarrow\infty. \label{quenched mn
\end{equation}
If the conditions of Theorem \ref{Thm1nn} hold and (\ref{regularity}) holds
with $m=n,$ then for almost all $\omega\in\Omega$
\begin{equation}
\frac{1}{n}S_{n,n}\Rightarrow N(0,\sigma^{2})\text{ under }P^{\omega}\text{
when }n\rightarrow\infty. \label{Quenched nn
\end{equation}
\end{corollary}
For the sake of applications, we provide a sufficient condition which takes
care of the regularity assumptions (\ref{regularity}).
\begin{theorem}
\label{Cor mn}Assume that
\begin{equation}
\sum_{u,v\geq1}\frac{\lVert E_{1,1}(X_{u,v})\rVert_{2}}{(uv)^{1/2}}<\infty.
\label{higher moment 2
\end{equation}
(a) Then for almost all $\omega\in\Omega$ (\ref{Quenched nn})
holds.\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \linebrea
(b) If in addition (\ref{main condi}) is satisfied, then for almost all
$\omega\in\Omega$ (\ref{quenched mn}) holds.
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \linebrea
(c) If for some $q>2$
\begin{equation}
\sum_{u,v\geq1}\frac{\lVert E_{1,1}(X_{u,v})\rVert_{q}}{(uv)^{1/q}}<\infty,
\label{higher moment q
\end{equation}
then the quenched convergence in (\ref{quenched mn}) holds.
\end{theorem}
\begin{remark}
\label{sigma^2 copy(1)} In Corollary \ref{Cor} and Theorem \ref{Cor mn},
$\sigma^{2}$ can be identified as $\sigma^{2}=\lim_{n\wedge m\rightarrow
\infty}E S_{n,m}^{2}/nm$ ($\lim_{n\rightarrow\infty}E S_{n,n}^{2}/n^{2}$ respectively).
\end{remark}
\begin{remark}
Theorem \ref{Cor mn} can be viewed as an extension to the random fields of
Proposition 12 in Cuny and Peligrad (2012). As we shall see, the proof for
random fields is much more involved and requires several intermediary steps
and new ideas.
\end{remark}
\section{Proofs}
Let us point out the main idea of the proof. Since Peligrad and Voln\'{y}
(2018) proved a quenched CLT for ortho-martingales, we reduce the proof to the
existence of an almost sure ortho-martingale approximation. We prove first
Theorem \ref{Thm2nm}, since the proof of Theorem \ref{Thm1nn} is similar with
the exception that we use different ergodic theorems.
Let us denote by $\hat{T}$ and $\hat{S}$ the operators on $L^{2}$ defined by
$\hat{T}f=f\circ T$, $\hat{S}f=f\circ S.$\newline
\begin{proof}
[Proof of Theorem \ref{Thm2nm}]Starting from condition (\ref{main condi}), by
triangle inequality we have that
\begin{equation}
f_{\mathbf{0}}:=\sum_{u,v\geq0}|\mathcal{P}_{0,0}(X_{u,v})|<\infty\text{ a.s.
} \label{cvg of proj
\end{equation}
and
\begin{equation}
\lVert f_{\mathbf{0}}\rVert_{\phi}\leq\sum_{u,v\geq0}\lVert\mathcal{P
_{0,0}(X_{u,v})\rVert_{\phi}<\infty. \label{momf0
\end{equation}
Note that by (\ref{cvg of proj}) $\mathcal{P}_{1,1}(S_{n,m})$ is convergent
almost surely. Denote the pointwise limit by
\[
D_{1,1}=\lim_{n\wedge m\rightarrow\infty}\mathcal{P}_{1,1}(S_{n,m
)=\sum_{u,v\geq1}\mathcal{P}_{1,1}(X_{u,v}).
\]
Meanwhile, by the triangle inequality and (\ref{main condi}), we obtain
\[
\sup_{n,m\geq1}|\mathcal{P}_{1,1}(S_{n,m})|\leq\sum_{u,v\geq1}|\mathcal{P
_{1,1}(X_{u,v})| \ \ \text{a.s.}
\]
and
\[
E\biggl( \sum_{u,v\geq1}|\mathcal{P}_{1,1}(X_{u,v})|\biggr) ^{2
\leq\biggl( \sum_{u,v\geq1}\lVert\mathcal{P}_{1,1}(X_{u,v})\rVert
_{2}\biggr) ^{2}<\infty.
\]
Thus, by the dominated convergence theorem, $\mathcal{P}_{1,1}(S_{n,m})$
converges to $D_{1,1}$ a.s. and in $L^{2}(P)$ as $n\wedge m\rightarrow\infty$.
Since $E_{0,1}(\mathcal{P}_{1,1}(S_{n,m}))=0$ a.s. and $E_{1,0}(\mathcal{P
_{1,1}(S_{n,m}))=0$ a.s., by defining for every $i,j\in Z$, $D_{i,j}=\hat
{T}^{i-1}\hat{S}^{j-1}D_{1,1}$, we conclude that $(D_{i,j})_{i,j\in Z}$ is a
field of martingale differences. By the expression of $D_{1,1}$ above,
\[
D_{i,j}=\sum_{(u,v)\geq(i,j)}\mathcal{P}_{i,j}(X_{u,v}).
\]
Now we look into the decomposition of $S_{n,m}$ (see Peligrad and Zhang
(2018b) for details)
\begin{equation}
S_{n,m}-R_{n,m}=\sum_{i=1}^{n}\sum_{j=1}^{m}{\mathcal{P}}_{i,j}({\sum
\limits_{u=i}^{n}}\sum\limits_{v=j}^{m}X_{u,v}) \label{ort dec
\end{equation}
wher
\[
R_{n,m}\ =E_{n,0}(S_{n,m})+E_{0,m}(S_{n,m})-E_{0,0}(S_{n,m}).
\]
Therefore
\[
\frac{S_{n,m}-R_{n,m}-M_{n,m}}{\sqrt{nm}}=\frac{1}{\sqrt{nm}}\sum_{i=1
^{n}\sum_{j=1}^{m}\biggl(\mathcal{P}_{i,j}({\sum\limits_{u=i}^{n}
\sum\limits_{v=j}^{m}X_{u,v})-D_{i,j}\biggr).
\]
By the orthogonality of the field of martingale differences $(\mathcal{P
_{i,j}({\sum\limits_{u=i}^{n}}\sum\limits_{v=j}^{m}X_{u,v})-D_{i,j})_{i,j\in
Z}$ and the assumption that the filtration is commuting, we hav
\[
\frac{1}{nm}E_{0,0}\left( S_{n,m}-R_{n,m}-M_{n,m}\right) ^{2}=\frac{1
{nm}\sum_{i=1}^{n}\sum_{j=1}^{m}E_{0,0}\biggl(\mathcal{P}_{i,j}({\sum
\limits_{u=i}^{n}}\sum\limits_{v=j}^{m}X_{u,v})-D_{i,j}\biggr)^{2}.
\]
From Theorem 1 in Peligrad and Voln\'{y} (2018), we know that the quenched CLT
holds for $M_{n,m}/\sqrt{nm}$. Therefore by Theorem 25.4 in Billingsley
(1995), in order to prove the conclusion of this theorem, it is enough to show
that
\begin{equation}
\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}E_{0,0}\left( S_{n,m
-R_{n,m}-M_{n,m}\right) ^{2}=0\ \ \text{a.s.} \label{negli
\end{equation}
Define the operators
\[
Q_{1}(f)=E_{0,\infty}(\hat{T}f);\ Q_{2}(f)=E_{\infty,0}(\hat{S}f)
\]
Note that $Q_{1}$ and $Q_{2}$ are commuting Dunford-Schwartz operators and we
can write
\[
E_{0,0}\left( \mathcal{P}_{i,j}(X_{u,v})\right) ^{2}=Q_{1}^{i}Q_{2
^{j}(\mathcal{P}_{0,0}(X_{u-i,v-j}))^{2}.
\]
By simple algebra we obtain
\begin{align*}
& E_{0,0}\biggl(\mathcal{P}_{i,j}({\sum\limits_{u=i}^{n}}\sum\limits_{v=j
^{m}X_{u,v})-D_{i,j}\biggr)^{2}\\
& =E_{0,0}\biggl({\sum\limits_{u=n+1}^{\infty}}\sum\limits_{v=j
^{m}\mathcal{P}_{i,j}(X_{u,v})+{\sum\limits_{u=i}^{\infty}}\sum\limits_{v=m+1
^{\infty}\mathcal{P}_{i,j}(X_{u,v})\biggr )^{2}.
\end{align*}
Therefore, by elementary inequalities we have the following bound
\begin{align*}
\frac{1}{nm}E_{0,0}\left( S_{n,m}-R_{n,m}-M_{n,m}\right) ^{2} & =\frac
{1}{nm}\sum_{i=1}^{n}\sum_{j=1}^{m}E_{0,0}\biggl(\mathcal{P}_{i,j
({\sum\limits_{u=i}^{n}}\sum\limits_{v=j}^{m}X_{u,v})-D_{i,j}\biggr)^{2}\\
& \leq2(I_{n,m}+II_{n,m}),
\end{align*}
where we have used the notations
\[
I_{n,m}=\frac{1}{nm}\sum_{i=1}^{n}\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j
\biggl(\sum_{u=n+1-i}^{\infty}\sum_{v=0}^{\infty}|\mathcal{P}_{0,0
(X_{u,v})|\biggr)^{2
\]
and
\[
II_{n,m}=\frac{1}{nm}\sum_{i=1}^{n}\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j
\biggl(\sum_{u=0}^{\infty}\sum_{v=m+1-j}^{\infty}|\mathcal{P}_{0,0
(X_{u,v})|\biggr)^{2}.
\]
The task is now to show the almost sure negligibility of each term. By
symmetry we treat only one of them. \newline Let $c$ be a fixed integer
satisfying $c<n$. We decompose $I_{n,m}$ into two part
\begin{equation}
\frac{1}{nm}\sum_{i=1}^{n-c}\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j}\biggl(\sum
_{u=n+1-i}^{\infty}\sum_{v=0}^{\infty}|\mathcal{P}_{0,0}(X_{u,v
)|\biggr)^{2}:=A_{n,m}(c) \label{A
\end{equation}
and
\begin{equation}
\frac{1}{nm}\sum_{i=n-c+1}^{n}\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j}\biggl(\sum
_{u=n+1-i}^{\infty}\sum_{v=0}^{\infty}|\mathcal{P}_{0,0}(X_{u,v
)|\biggr)^{2}:=B_{n,m}(c). \label{B
\end{equation}
Note tha
\begin{align*}
B_{n,m}(c) & \leq\frac{1}{nm}\sum_{i=n-c+1}^{n}\sum_{j=1}^{m}Q_{1}^{i
Q_{2}^{j}f_{\mathbf{0}}^{2}\\
& =\frac{1}{nm}\sum_{i=1}^{n}\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j}f_{\mathbf{0
}^{2}-\frac{1}{nm}\sum_{i=1}^{n-c}\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j
f_{\mathbf{0}}^{2},
\end{align*}
where $f_{\mathbf{0}}$ is given by (\ref{cvg of proj}).
Since $Q_{1}$ and $Q_{2}$ are commuting Dunford-Schwartz operators, and by
(\ref{momf0}) we have that $E(f_{\mathbf{0}}^{2}\log(1+|f_{\mathbf{0
}|))<\infty,$ by the ergodic theorem (Krengel (1985), Theorem 1.1, Ch. 6), for
each $c$ fixed,
\begin{equation}
\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}\sum_{i=1}^{n-c}\sum_{j=1
^{m}Q_{1}^{i}Q_{2}^{j}f_{\mathbf{0}}^{2}=g\text{ a.s. } \label{R1
\end{equation}
where
\[
g=\lim_{n\rightarrow\infty}\frac{1}{n}\sum_{i=1}^{n-c}Q_{1}^{i}\left(
\lim_{m\rightarrow\infty}\frac{1}{m}\sum_{j=1}^{m}Q_{2}^{j}(f_{\mathbf{0
})\right) .
\]
Since we assume that either $S$ or $T$ is ergodic, without loss of generality,
here we assume that $S$ is ergodic. By applying Lemma 7.1 in Dedecker et al.
(2014), we obtain tha
\[
\lim_{m\rightarrow\infty}\frac{1}{m}\sum_{j=1}^{m}Q_{2}^{j}(f_{\mathbf{0
})=E\left( f_{\mathbf{0}}^{2}\right) \text{ a.s. },
\]
which implies that $g$ in (\ref{R1}) is a constant almost surely and
$g=E\left( f_{\mathbf{0}}^{2}\right) $.
Therefore, for all $c>0
\[
\lim_{n\wedge m\rightarrow\infty}B_{n,m}(c)=0\text{ a.s.
\]
In order to treat the first term in the decomposition of $I_{n,m}$, note that
\[
A_{n,m}(c)\leq\frac{1}{nm}\sum_{i=1}^{n-c}\sum_{j=1}^{m}Q_{1}^{i}Q_{2
^{j}f_{\mathbf{0}}^{2}(c)\ \text{ where }f_{\mathbf{0}}(c)=\sum_{u=c}^{\infty
}\sum_{v=0}^{\infty}|\mathcal{P}_{0,0}(X_{u,v})|.
\]
Again, by the ergodic theorem for Dunford-Schwartz operators (Krengel (1985),
Theorem 1.1, Ch. 6) and Lemma 7.1 in Dedecker et al. (2014), for each $c$ fixe
\begin{equation}
\label{R2}\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}\sum_{i=1}^{n-c
\sum_{j=1}^{m}Q_{1}^{i}Q_{2}^{j}f_{\mathbf{0}}^{2}(c)=E\left( f_{\mathbf{0
}^{2}(c)\right) \text{ a.s.
\end{equation}
In addition, by (\ref{cvg of proj}), we know that $\lim_{c\rightarrow\infty
}|f_{\mathbf{0}}(c)|=0.$ So, by the dominated convergence theorem, we have
\[
\lim_{c\rightarrow\infty}\lim_{n\wedge m\rightarrow\infty} A_{n,m}(c)\leq
\lim_{c\rightarrow\infty}E(f_{\mathbf{0}}^{2}(c))=0\text{ a.s.
\]
\ The proof of the theorem is now complete. \
\end{proof}
\bigskip
The proof of Theorem \ref{Thm1nn} requires only a slight modification of the
proof of Theorem \ref{Thm2nm}. Indeed, instead of Theorem 1.1 in Ch. 6 in
Krengel (1985), we shall use Theorem 2.8 in Ch. 6 in the same book.\newline
\begin{proof}
[Proof of Corollary \ref{Cor}]By Theorem \ref{Thm2nm} together with Theorem
25.4 in Billingsley (1995), it suffices to show that (\ref{regularity})
implies tha
\begin{equation}
\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}E_{0,0}(R_{n,m}^{2})=0\ \text{
a.s. } \label{CR
\end{equation}
Simple computations, involving the fact that the filtration is commuting,
gives that
\begin{equation}
E_{0,0}(R_{n,m}^{2})=E_{0,0}\left( E_{n,0}^{2}(S_{n,m})\right)
+E_{0,0}\left( E_{0,m}^{2}(S_{n,m})\right) -E_{0,0}^{2}(S_{n,m})
\label{estimate
\end{equation}
and since $E_{0,0}^{2}(S_{n,m})\leq E_{0,0}\left( E_{0,m}^{2}(S_{n,m
)\right) $ a.s.$,$ we have
\[
\text{ }\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}E_{0,0}(R_{n,m
^{2})=0\ \text{ a.s. }\ \text{by condition (\ref{regularity})}.
\]
\end{proof}
\bigskip
We give next the proof of Theorem \ref{Cor mn}. Before proving this theorems,
we shall first establish several preliminary facts presented as three lemmas.
\begin{lemma}
\label{fact1}Let $q\geq2.$ Condition (\ref{higher moment q}) implie
\begin{equation}
\sum_{u\geq1}\frac{1}{u^{1/q}}\sum_{v\geq0}\Vert{P_{0,\tilde{0}}(X_{u,v
)}\Vert_{q}<\infty. \label{nm fact
\end{equation}
\end{lemma}
\begin{proof}
Throughout the proof, denote by $C_{q}>0$ a generic constant depending on $q$
which may take different values from line to line. By the H\"{o}lder
inequality and the Rosenthal inequality for martingales (see Theorem
\ref{Rosenthal ineq} in the Appendix), we hav
\begin{gather*}
\sum_{v\geq1}\lVert P_{0,\tilde{0}}(X_{u,v})\rVert_{q}=\sum_{v\geq1}\lVert
P_{-u,-\tilde{v}}(X_{0,0})\rVert_{q}\leq\sum_{n\geq0}(2^{n})^{\frac{q-1}{q
}\left( \sum_{v=2^{n}}^{2^{n+1}-1}\lVert P_{-u,-\tilde{v}}(X_{0,0})\rVert
_{q}^{q}\right) ^{\frac{1}{q}}\\
\leq C_{q}\sum_{n\geq0}(2^{n})^{\frac{q-1}{q}}\lVert\sum_{v=2^{n}}^{2^{n+1
-1}P_{-u,-\tilde{v}}(X_{0,0})\rVert_{q}\leq2C_{q}\sum_{n\geq0}(2^{n
)^{\frac{q-1}{q}}\lVert E_{-u,-2^{n}}(X_{0,0})\rVert_{q}.\text{
\end{gather*}
Since the sequence $(\Vert{E_{-u,-n}(X_{0,0})}\Vert_{q})_{n\geq1}$ is
non-increasing in $n$, it follows that
\[
(2^{n})^{\frac{q-1}{q}}\Vert{E_{-u,-{2^{n}}}(X_{0,0})}\Vert_{q}\leq
2\sum_{k=2^{n-1}}^{2^{n}-1}\frac{\Vert{E_{-u,-k}(X_{0,0})}\Vert_{q}}{k^{1/q
}.
\]
S
\begin{equation}
\sum_{v=1}^{\infty}\Vert{P_{0,\tilde{0}}(X_{u,v})}\Vert_{q}\leq C_{q
\sum_{k\geq1}\frac{\Vert{E_{-u,-k}(X_{0,0})}\Vert_{q}}{k^{1/q}}.
\label{nm fact0
\end{equation}
Thus relation (\ref{nm fact}) holds by (\ref{higher moment q}),
(\ref{nm fact0}) and stationarity.
In addition, for any $u\geq0,$ we also hav
\begin{equation}
\sum_{v=1}^{\infty}\lVert P_{0,\tilde{0}}(X_{u,v})\rVert_{q}<\infty.
\label{nm other fact1
\end{equation}
By the symmetric roles of $m$ and $n$, for any $v\geq0,$ we have
\begin{equation}
\sum_{u=1}^{\infty}\lVert P_{\tilde{0},0}(X_{u,v})\rVert_{q}<\infty.
\label{nm other fact2
\end{equation}
\end{proof}
\begin{lemma}
\label{fact2}Condition (\ref{higher moment 2}) implies
\begin{equation}
\lim_{n\wedge m\rightarrow\infty}\frac{1}{nm}E_{0,0}(R_{n,m}^{2})=0\text{ a.s.
} \label{negliR
\end{equation}
\end{lemma}
\begin{proof}
First we show that (\ref{higher moment 2}) implies tha
\[
\frac{E_{0,0}^{2}(S_{n,m})}{nm}\rightarrow0\text{ a.s. when }n\wedge
m\rightarrow\infty.
\]
We bound this term in the following wa
\begin{gather*}
\frac{|E_{0,0}(S_{n,m})|}{\sqrt{nm}}\leq\frac{1}{\sqrt{nm}}\sum_{u=1}^{n
\sum_{v=1}^{m}|E_{0,0}(X_{u,v})|\\
\leq\frac{1}{\sqrt{n}}\sum_{u=1}^{c}\sum_{v=1}^{\infty}\frac{|E_{0,0
(X_{u,v})|}{\sqrt{v}}+\sum_{u=c+1}^{\infty}\sum_{v=1}^{\infty}\frac
{|E_{0,0}(X_{u,v})|}{\sqrt{uv}}\\
\leq\frac{c}{\sqrt{n}}\sup_{1\leq u\leq c}\sum_{v=1}^{\infty}\frac
{|E_{0,0}(X_{u,v})|}{\sqrt{v}}+\sum_{u=c+1}^{\infty}\sum_{v=1}^{\infty
\frac{|E_{0,0}(X_{u,v})|}{\sqrt{uv}}.
\end{gather*}
Now, (\ref{higher moment 2}) implies that
\[
\sum_{u=1}^{\infty}\sum_{v=1}^{\infty}\frac{|E_{0,0}(X_{u,v})|}{\sqrt{uv
}<\infty\text{ a.s.
\]
Therefore,
\begin{equation}
\frac{|E_{0,0}(S_{n,m})|}{\sqrt{nm}}\rightarrow0\text{ a.s.} \label{negl1
\end{equation}
by letting $n\rightarrow\infty$ followed by $c\rightarrow\infty.$
By (\ref{estimate}) and the symmetric roles of $m$ and $n$, the theorem will
follow if we can show that
\[
E_{0,0}\frac{(E_{0,m}^{2}(S_{n,m}))}{nm}\rightarrow0\text{ a.s. when }n\wedge
m\rightarrow\infty.
\]
By (\ref{negl1})\ this is equivalent to showing tha
\[
\frac{1}{nm}E_{0,0}\left( E_{0,m}(S_{n,m})-E_{0,0}(S_{n,m})\right)
^{2}\rightarrow0\text{ a.s. when }n\wedge m\rightarrow\infty.
\]
We start from the representatio
\begin{align*}
E_{0,0}\left( E_{0,m}(S_{n,m})-E_{0,0}(S_{n,m})\right) ^{2} & =\sum
_{j=1}^{m}E_{0,0}\biggl[P_{0,\tilde{j}}\biggl(\sum_{u=1}^{n}\sum_{v=j
^{m}X_{u,v}\biggr)\biggr]^{2}\\
& =\sum_{j=1}^{m}E_{0,0}\biggl[\widehat{S}^{j}\biggl(P_{0,\tilde{0}
(\sum_{u=1}^{n}\sum_{v=0}^{m-j}X_{u,v})\biggr)^{2}\biggr].
\end{align*}
So,
\begin{gather*}
\frac{1}{nm}E_{0,0}\left( E_{0,m}(S_{n,m})-E_{0,0}(S_{n,m})\right)
^{2}=\frac{1}{mn}\sum_{j=1}^{m}E_{0,0}\biggl[\widehat{S}^{j}\biggl(\sum
_{u=1}^{n}\sum_{v=0}^{m-j}P_{0,\tilde{0}}(X_{u,v})\biggr)^{2}\biggr]\\
\leq\frac{2}{mn}\sum_{j=1}^{m}E_{0,0}\biggl[\widehat{S}^{j}\biggl(\sum
_{u=1}^{c}\sum_{v=0}^{m-j}|P_{0,\tilde{0}}(X_{u,v})|\biggr)^{2}\biggr]\\
+\frac{2}{m}\sum_{j=1}^{m}E_{0,0}\biggl[\widehat{S}^{j}\biggl(\sum_{u=c+1
^{n}\frac{1}{\sqrt{u}}\sum_{v=0}^{m-j}|P_{0,\tilde{0}}(X_{u,v})|\biggr)^{2
\biggr]\\
=I_{n,m,c}+II_{n,m,c}.
\end{gather*}
Let us introduce the operato
\[
Q_{0}(f)=E_{0,0}(\widehat{S}f).
\]
We treat first the term $I_{n,m,c}.\ $For $c$ fixe
\begin{align*}
I_{n,m,c} & \leq\frac{2c^{2}}{mn}\sup_{1\leq u\leq c}\sum_{j=1}^{m
E_{0,0}\biggl[\widehat{S}^{j}\biggl(\sum_{v=0}^{\infty}|P_{0,\tilde{0
}(X_{u,v})|\biggr)^{2}\biggr]\\
& =\frac{2c^{2}}{mn}\sup_{1\leq u\leq c}\sum_{j=1}^{m}Q_{0}^{j
\biggl[\biggl(\sum_{v=0}^{\infty}|P_{0,\tilde{0}}(X_{u,v})|\biggr)^{2}\biggr].
\end{align*}
By (\ref{nm other fact1}), the function
\[
g(u)=\sum_{v=0}^{\infty}|P_{0,\tilde{0}}(X_{u,v})|
\]
is square integrable. By the ergodic theorem for Dunford-Schwartz operators
(see Theorem 11.4 in Eisner et al., 2015 or Corollary 3.8 in Ch. 3, Krengel,
1985) and Lemma 7.1 in Dedecker et al. (2014),
\[
\frac{1}{m}\sum_{j=1}^{m}Q_{0}^{j}\left[ g^{2}(u)\right] \rightarrow
E(g^{2}(u))\text{ a.s.
\]
and therefore, since $c$ is fixed,
\[
\lim_{n\wedge m\rightarrow\infty}I_{n,m,c}=0\text{ a.s.
\]
In order to treat the second term, note that
\[
II_{n,m,c}\leq\frac{2}{m}\sum_{j=1}^{m}Q_{0}^{j}\biggl[\biggl(\sum
_{u=c}^{\infty}\frac{1}{\sqrt{u}}\sum_{v=0}^{\infty}|P_{0,\tilde{0}
(X_{u,v})|\biggr)^{2}\biggr].
\]
Denot
\[
h(c)=\sum_{u=c}^{\infty}\frac{1}{\sqrt{u}}\sum_{v=0}^{\infty}|P_{0,\tilde{0
}(X_{u,v})|.
\]
By (\ref{nm fact}), we know tha
\begin{equation}
\sum_{u=1}^{\infty}\frac{1}{\sqrt{u}}\sum_{v=0}^{\infty}\lVert P_{0,\tilde{0
}(X_{u,v})\rVert_{2}<\infty. \label{nn fact
\end{equation}
So, $E(h^{2}(c))<\infty.$ Again, by the ergodic theorem for the
Dunford-Schwartz operators (see Theorem 11.4 in Eisner et al., 2015 or
Corollary 3.8 in Ch. 3, Krengel, 1985), we obtain
\[
\frac{1}{m}\sum_{j=1}^{m}Q_{0}^{j}(h^{2}(c))\rightarrow E\left(
h^{2}(c)\right) \leq\left( \sum_{u=c}^{\infty}\frac{1}{\sqrt{u}}\sum
_{v=0}^{\infty}\lVert P_{0,\tilde{1}}(X_{u,v})\rVert_{2}\right) ^{2}.
\]
So, by (\ref{nn fact})
\[
\lim_{c\rightarrow\infty}\lim_{m\rightarrow\infty}\frac{1}{m}\sum_{j=1
^{m}Q_{0}^{j}(h^{2}(c))=0\text{ a.s.
\]
\end{proof}
\begin{lemma}
\label{fact3} Let $q\geq2.$ Condition (\ref{higher moment q}) implie
\begin{equation}
\sum_{u,v\geq0}\lVert\mathcal{P}_{0,0}(X_{u,v})\rVert_{q}<\infty,
\label{nm fact 0
\end{equation}
which clearly implies (\ref{main condi}).
\end{lemma}
\begin{proof}
By applying twice the Rosenthal inequality for martingales (see Theorem
\ref{Rosenthal ineq} in the Appendix), for any integers $a\leq b$ and $c\leq
d$, we hav
\begin{equation}
\sum_{k=a}^{b}\sum_{k^{\prime}=c}^{d}\left\Vert \mathcal{P}_{-k,-k^{\prime
}(X_{0,0})\right\Vert _{q}^{q}\leq C_{q}\Vert{\sum_{k=a}^{b}\sum_{k^{\prime
}=c}^{d}\mathcal{P}_{-k,-k^{\prime}}(X_{0,0})}\Vert_{q}^{q}.
\label{Rosenthal bound
\end{equation}
In addition, note that for any integers $a\leq b$ and $c\leq d$, we hav
\begin{equation}
\Vert{\sum_{k=a}^{b}\sum_{k^{\prime}=c}^{d}\mathcal{P}_{-k,-k^{\prime
}(X_{0,0})}\Vert_{q}^{q}\leq4^{q}\Vert{E_{-a,-c}(X_{0,0})}\Vert_{q}^{q}.
\label{Fact
\end{equation}
Then by the H\"{o}lder's inequality together with (\ref{Rosenthal bound}) and
(\ref{Fact}), we obtain
\begin{gather*}
\sum_{u,v\geq1}\left\Vert \mathcal{P}_{-u,-v}(X_{0,0})\right\Vert _{q}\leq
\sum_{n,m\geq0}(2^{n}2^{m})^{\frac{q-1}{q}}\left( \sum_{k=2^{n}}^{2^{n+1
-1}\sum_{k^{\prime}=2^{m}}^{2^{m+1}-1}\left\Vert \mathcal{P}_{-k,-k^{\prime
}(X_{0,0})\right\Vert _{q}^{q}\right) ^{\frac{1}{q}}\\
\leq4C_{q}\sum_{n,m\geq0}(2^{n}2^{m})^{\frac{q-1}{q}}\left\Vert E_{-2^{n
,-2^{m}}(X_{0,0})\right\Vert _{q}.
\end{gather*}
Since $\lVert E_{-2^{n},-2^{m}}(X_{0,0})\rVert$ is non-increasing in $n$ and
$m$, it follows that
\[
(2^{n}2^{m})^{\frac{q-1}{q}}\left\Vert E_{-2^{n},-2^{m}}(X_{0,0})\right\Vert
_{q}\leq4\sum_{u=2^{n-1}}^{2^{n}-1}\sum_{v=2^{m-1}}^{2^{m}-1}\frac{\lVert
E_{-u,-v}(X_{0,0})\rVert_{q}}{(uv)^{1/q}}.
\]
Therefore, by the relations above, we have proved that (\ref{higher moment q})
implies
\[
\sum_{u,v\geq1}\lVert\mathcal{P}_{-u,-v}(X_{0,0})\rVert_{q}<\infty.
\]
Similarly we have
\[
\sum_{u=1}^{\infty}\lVert\mathcal{P}_{-u,0}(X_{0,0})\rVert_{q}<\infty\text{
and }\sum_{v=1}^{\infty}\lVert\mathcal{P}_{0,-v}(X_{0,0})\rVert_{q}<\infty.
\]
Thus by stationarity (\ref{nm fact 0}) holds.
\end{proof}
\bigskip
\begin{proof}
[Proof of Theorem \ref{Cor mn}]The item (a) of the theorem follows as a
combination of Theorem \ref{Thm1nn} with Lemmas \ref{fact2} and \ref{fact3},
applied with $q=2.$
To prove item (b) of this theorem we combine Theorem \ref{Thm2nm} with Lemmas
\ref{fact2}.
Finaly, the conclusion (c) is a consequence of \ref{Thm2nm} combined with
Lemmas \ref{fact2} and \ref{fact3}, applied with $q>2.$
\end{proof}
\section{Random fields with multidimensional index sets}
In this section we extend our results to random fields indexed by $Z^{d}$,
$d>2$. By $\mathbf{u\leq n}$ we understand $\mathbf{u=}(u_{1},...,u_{d})$,
$\mathbf{n=}(n_{1},...,n_{d})$ and $1\leq u_{1}\mathbf{\leq}n_{1}$,...,$1\leq
u_{d}\mathbf{\leq}n_{d}\mathbf{.}$ We shall start with a strictly stationary
real-valued random field $\mathbf{\xi}=(\xi_{\mathbf{u}})_{\mathbf{u}\in
Z^{d}}$, defined on the canonical probability space $R^{Z^{d}}$ and\ define
the filtrations $\mathcal{F}_{\mathbf{u}}=\sigma(\xi_{\mathbf{j}
:\mathbf{j}\leq\mathbf{u})$. We shall assume that the filtration is commuting
if $E_{\mathbf{u}}E_{\mathbf{a}}(X)=E_{\mathbf{u}\wedge\mathbf{a}}(X),$ where
the minimum is taken coordinate-wise and we used notation $E_{\mathbf{u
}(X)=E(X|\mathcal{F}_{\mathbf{u}})$. We define
\begin{equation}
X_{\mathbf{m}}=f((\xi_{\mathbf{j}})_{\mathbf{j}\leq\mathbf{m}})\text{ and set
}S_{\mathbf{k}}=\sum\nolimits_{\mathbf{u}=\mathbf{1}}^{\mathbf{k
}X_{\mathbf{u}}. \label{Xdef d
\end{equation}
The variable $X_{\mathbf{0}}$ is assumed to be square integrable (in $L^{2}$)
and with mean $0$. We also define $T_{i}$ the coordinate-wise translations and
then
\[
X_{\mathbf{k}}=f(T_{1}^{k_{1}}\circ...\circ T_{d}^{k_{d}}(\xi_{\mathbf{u
})_{\mathbf{u}\leq\mathbf{0}}).
\]
Let $d$ be a function and defin
\begin{equation}
D_{\mathbf{m}}=d((\xi_{\mathbf{j}})_{\mathbf{j}\leq\mathbf{m}})\text{ and set
}M_{\mathbf{k}}=\sum\nolimits_{\mathbf{u}=\mathbf{1}}^{\mathbf{k
}D_{\mathbf{u}}. \label{def Dg
\end{equation}
Assume integrability. We say that $(D_{\mathbf{m}})_{\mathbf{m}\in Z^{d}}$ is
a field of martingale differences if $E_{\mathbf{a}}(D_{\mathbf{m}})=0$ if at
least one coordinate of $\mathbf{a}$ is strictly smaller than the
corresponding coordinate of $\mathbf{m.}$ Now we introduce the $d$-dimensional
projection operator. By using the fact that the filtration is commuting, it is
convenient to define projections $\mathcal{P}_{\mathbf{u}}$ in the following wa
\[
\mathcal{P}_{\mathbf{u}}(X):=P_{\mathbf{u}(1)}\circ P_{\mathbf{u}(2)
\circ...\circ P_{\mathbf{u}(d)}(X),
\]
wher
\begin{equation}
P_{\mathbf{u}(j)}(Y):=E(Y|\mathcal{F}_{\mathbf{u}})-E(Y|\mathcal{F
_{\mathbf{u}(j)}), \label{proj def d
\end{equation}
where $\mathbf{u}(j)$ has all the coordinates of $\mathbf{u}$ with the
exception of the $j$-th coordinate, which is $u_{j}-1$. For instance when
$d=3,$ $P_{\mathbf{u}(2)}(Y)=E(Y|\mathcal{F}_{u_{1},u_{2},u_{3}
)-E(Y|\mathcal{F}_{u_{1},u_{2}-1,u_{3}}).$
We say that a random field $(X_{\mathbf{n}})_{\mathbf{n}\in Z^{d}}$ admits a
martingale approximation if there is a field of martingale differences
$(D_{\mathbf{m}})_{\mathbf{m}\in Z^{d}}$ such that for almost all $\omega
\in\Omega$
\begin{equation}
\frac{1}{|\mathbf{n|}}E^{\omega}\left( S_{\mathbf{n}}-M_{\mathbf{n}}\right)
^{2}\rightarrow0\text{ when }\min_{1\leq i\leq d}n_{i}\rightarrow\infty,
\label{martapprxd
\end{equation}
where $|\mathbf{n|=}n_{1}...n_{d}.$
Let $R_{\mathbf{n}}$ be the remainder term of the decomposition of
$S_{\mathbf{n}}$ such that
\[
S_{\mathbf{n}}=\sum_{\mathbf{u}=\mathbf{1}}^{\mathbf{n}}\mathcal{P
_{\mathbf{u}}(S_{\mathbf{n}})+R_{\mathbf{n}}.
\]
In this context we have:
\begin{theorem}
\label{Thm nn d} Assume that $(X_{\mathbf{n}})_{\mathbf{n}\in Z^{d}}$ is
defined by (\ref{Xdef d}) and there is an integer $i$, $1\leq i\leq d$, such
that $T_{i}$ is ergodic and the filtrations are commuting. In addition assume
that
\begin{equation}
\label{cond nn}\sum_{\mathbf{u\geq0}}\lVert\mathcal{P}_{\mathbf{0
}(X_{\mathbf{u}})\rVert_{2}<\infty. \ \
\end{equation}
Then, for almost all $\omega\in\Omega,
\[
(S_{n,\cdots, n}-R_{n,\cdots, n})/n^{d/2}\Rightarrow N(0,\sigma^{2})\text{
under }P^{\omega}\text{ when }n\rightarrow\infty.
\]
\end{theorem}
\begin{theorem}
\label{Thm nm d} Furthermore, assume now condition (\ref{cond nn}) is
reinforced t
\begin{equation}
\label{main condi2}\sum_{\mathbf{u\geq0}}\lVert\mathcal{P}_{\mathbf{0
}(X_{\mathbf{u}})\rVert_{\varphi}<\infty, \ \
\end{equation}
where $\varphi(x)=x^{2}\log^{d-1}(1+|x|)$ and $\lVert\cdot\rVert_{\varphi}$ is
defined by (\ref{phi norm}).\newline Then, for almost all $\omega\in\Omega,$
\[
\frac{1}{\sqrt{|\mathbf{n}|}}(S_{\mathbf{n}}-R_{\mathbf{n}}) \Rightarrow
N(0,\sigma^{2})\text{ under }P^{\omega}\text{ when }\min_{1\leq i\leq d
n_{i}\rightarrow\infty.
\]
\end{theorem}
\begin{corollary}
\label{Cor2} Assume that the conditions of Theorem \ref{Thm nm d} hold and for
all $j$, $1\leq j\leq d$ we have
\begin{equation}
\label{Reminder condi d}\frac{1}{|\mathbf{n|}}E_{\mathbf{0}
\biggl(E_{\mathbf{n}_{j}}^{2}(S_{\mathbf{n}})\biggr)\rightarrow0\text{ a.s.
when }\min_{1\leq i\leq d}n_{i}\rightarrow\infty.
\end{equation}
where $\mathbf{n}_{j}\mathbf{\in}Z^{d}$ has the $j$-th coordinate $0$ and the
other coordinates equal to the coordinates of $\mathbf{n}$. Then, for almost
all $\omega\in\Omega$,
\begin{equation}
\label{quenched nm d}S_{\mathbf{n}}/\sqrt{|\mathbf{n}|}\Rightarrow
N(0,\sigma^{2}) \text{ under } P^{\omega} \text{ when } \min_{1\leq i\leq
d}n_{i}\rightarrow\infty.
\end{equation}
If the conditions of Theorem \ref{Thm nn d} hold and (\ref{Reminder condi d})
holds with $\mathbf{n}=(n,n,\cdots,n)$, then for almost all $\omega\in\Omega
$,
\begin{equation}
\label{quenched nn d}\frac{1}{n^{d/2}}S_{n,\cdots,n}\Rightarrow N(0,\sigma
^{2}) \text{ under } P^{\omega} \text{ when } n\rightarrow\infty.
\end{equation}
\end{corollary}
\begin{theorem}
\label{Cor nn nm d} Assume that $(X_{\mathbf{n}})_{n\in Z^{d}}$ is defined by
(\ref{Xdef d}) and the filtrations are commuting. Also assume that there is an
integer $i$, $1\leq i\leq d$, such that $T_{i}$ is ergodic and in addition for
$q>2$,
\begin{equation}
\sum_{\mathbf{u}\geq\mathbf{1}}\frac{\lVert E_{\mathbf{1}}(X_{\mathbf{u
})\rVert_{q}}{{|\mathbf{u}|}^{1/q}}<\infty. \label{higher moment d
\end{equation}
(a) If $q=2$, then the quenched convergence in (\ref{quenched nn d}) holds
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \linebrea
(b) If $q>2$, then the quenched convergence in (\ref{quenched nm d}) holds.
\end{theorem}
As for the case of random fields with two indexes, we start with the proof of
Theorem \ref{Thm nm d}, since the proof of Theorem \ref{Thm nn d} is similar
with the exception that we use different ergodic theorems. \bigskip
\begin{proof}
[Proof of Theorem \ref{Thm nm d}]The proof of this theorem is straightforward
following the same lines of proofs as for a double-indexed random field. It is
easy to see that, by using the commutativity property of the filtration, the
martingale approximation argument in the proof of Theorem \ref{Thm2nm} remains
unchanged if we replace $Z^{2}$ with $Z^{d}$ for $d\geq3$. The definition of
the approximating martingale is also clear. The only difference in the proof
is that for the validation of the limit in (\ref{R1}) and (\ref{R2}) when
$\min_{1\leq i\leq d}n_{i}\rightarrow\infty$, in order to apply the ergodic
theorem for Dunford-Schwartz operators, conform to Theorem 1.1 in Ch. 6 in
Krengel (1985), we have to assume $E\left[ f^{2}_{\mathbf{0}}\log
^{d-1}(1+|f_{\mathbf{0}}|)\right] <\infty$, which is implied by
(\ref{main condi2}).\newline
More precisely, let us denote by $\hat{T_{i}}$, $1\leq i\leq d$, the operators
defined by $\hat{T_{i}}f=f\circ T_{i}$. Then for $\mathbf{i}=(i_{1
,\cdots,i_{d})\in Z^{d}$, we define $Q^{\mathbf{i}}=\Pi_{k=1}^{d}Q_{k}^{i_{k
}$ where $(Q_{i})_{1\leq i\leq d}$ are operators associated with
coordinate-wise translations $(T_{i})_{1\leq i\leq d}$ defined as follows
\[
Q_{1}(f)=E_{0,\infty,\cdots,\infty}(\hat{T_{1}}f), Q_{2}(f)=E_{\infty
,0,\infty,\cdots,\infty}(\hat{T_{2}} f), \cdots, Q_{d}(f)=E_{\infty
,\cdots,\infty,0}(\hat{T_{d}} f).
\]
Then, we bound the following quantity
\[
\frac{1}{|\mathbf{n}|}E_{\mathbf{0}}\left[ |S_{\mathbf{n}}-R_{\mathbf{n
}-M_{\mathbf{n}}|^{2}\right]
\]
by the sum of $d$ terms with the first term of them in the form
\[
I_{\mathbf{n}}=\frac{1}{|\mathbf{n}|}\sum_{\mathbf{i=1}}^{\mathbf{n
}\mathbf{Q}^{\mathbf{i}}\left( \sum_{u=n_{1}+1-i_{1}}^{\infty}\sum
_{\mathbf{v\geq0}}|\mathcal{P}_{\mathbf{0}}(X_{u,\mathbf{v}})|\right)
^{2}\text{ where }\mathbf{v}\in Z^{d-1}.
\]
By symmetry, we only need to deal with this one. Let $c$ be a fixed integer
satisfying $c<n_{1}$, we decompose $I_{\mathbf{n}}$ into two parts:
\[
\frac{1}{|\mathbf{n}|}\sum_{i_{1}=1}^{n_{1}-c}\sum_{\mathbf{i^{\prime}=1
}^{\mathbf{n^{\prime}}}\mathbf{Q}^{\mathbf{i}}\left( \sum_{u=n_{1}+1-i_{1
}^{\infty}\sum_{\mathbf{v\geq0}}|\mathcal{P}_{\mathbf{0}}(X_{u,\mathbf{v
})|\right) ^{2}:=A_{\mathbf{n}}(c)
\]
and
\[
\frac{1}{|\mathbf{n}|}\sum_{i_{1}=n_{1}-c+1}^{n_{1}}\sum_{\mathbf{i^{\prime
}=1}}^{\mathbf{n^{\prime}}}\mathbf{Q}^{\mathbf{i}}\left( \sum_{u=n_{1
+1-i_{1}}^{\infty}\sum_{\mathbf{v\geq0}}|\mathcal{P}_{\mathbf{0
}(X_{u,\mathbf{v}})|\right) ^{2}:=B_{\mathbf{n}}(c)
\]
with $\mathbf{i}^{\prime}=(i_{2},\cdots,i_{d})$ and $\mathbf{n}^{\prime
}=(n_{2},\cdots,n_{d})$. Afterwards, we just proceed by following step by step
the proof for negligibility of $A_{n,m}(c)$ and $B_{n,m}$ (see (\ref{A}) and
(\ref{B}) from the proof of Theorem \ref{Thm2nm}).
The proof of Theorem \ref{Thm nn d} follows by similar arguments, just
replacing Theorem 1.1 in Ch. 6 in Krengel (1985) by Theorem 2.8 in Ch. 6 in
the same book.
\end{proof}
\bigskip
\begin{proof}
[Proof of Corollary \ref{Cor2}]The negligibility of the reminder
$R_{\mathbf{n}}$ can be shown exactly in the same way as the negligibility of
the term $R_{n,m}$ in the proof of Corollary \ref{Cor}.
\end{proof}
\bigskip
\begin{proof}
[Proof of Theorem \ref{Cor nn nm d}]As in the proof of (\ref{fact1 d}) and
(\ref{fact2 d}) in Theorem \ref{Cor mn}, we can show that
(\ref{higher moment d}) implies the following facts
\begin{equation}
\sum_{\mathbf{u\geq1}}\frac{1}{\sqrt{|\mathbf{u}|}}\sum_{v\geq0
\Vert{P_{\mathbf{0}(d)}(X_{\mathbf{u},v})}\Vert_{q}<\infty, \label{fact d
\end{equation}
\begin{equation}
\sum_{\mathbf{v\geq0}}\Vert{P_{\mathbf{0}}(X_{u,\mathbf{v}})}\Vert_{q
<\infty\label{fact1 d
\end{equation}
and
\begin{equation}
\sum_{u\geq1}\frac{1}{\sqrt{u}}\sum_{\mathbf{v\geq0}}\Vert{P_{\mathbf{0
}(X_{u,\mathbf{v}})}\Vert_{q}<\infty, \label{fact2 d
\end{equation}
where $\mathbf{0}=(0,\cdots,0)\in Z^{d},\mathbf{u},\mathbf{v}\in Z^{d-1}$ and
$P_{\mathbf{0}}=P_{\mathbf{0}(2)}\circ P_{\mathbf{0}(3)}\circ\cdots\circ
P_{\mathbf{0}(d)}$ with $P_{\mathbf{0}(j)}$ defined by (\ref{proj def d}).
To prove the corollary, we need to show that
\begin{equation}
\label{induction}\frac{1}{|\mathbf{n}|}E_{\mathbf{0}}\left( E_{\mathbf{n
^{(k)}}^{2}(S_{\mathbf{n}})\right) \rightarrow0 \text{ a.s. when }
\min_{1\leq i\leq d}n_{i}\rightarrow\infty,
\end{equation}
where $\mathbf{n}^{(k)}\in Z^{d}$ has $k$ coordinates equal to the
corresponding coordinates of $\mathbf{n}$ and the other $n-k$ coordinates zero
for all $0\leq k\leq d-1$. We will proceed by induction.
First, we have to show that
\[
\frac{E^{2}_{\mathbf{0}}(S_{\mathbf{n}})}{|\mathbf{n}|}\rightarrow0 \text{
a.s.} \text{ and }\frac{1}{|\mathbf{n}|}E_{\mathbf{0}}\left( E_{0,\cdots, 0,
n_{d}}^{2}(S_{\mathbf{n}})\right) \rightarrow0 \text{ a.s. when } \min_{1\leq
i\leq d}n_{i}\rightarrow\infty,
\]
which are easy to establish by similar arguments as in the proof of Theorem
\ref{Cor mn}, by using (\ref{higher moment d}) and (\ref{fact d}). That is,
(\ref{induction}) holds for $k=0$ and $k=1$. Now assume that for $k<d-1$ the
result holds. The fact that the result holds for $k=d-1$ follows
straightforward by using (\ref{fact1 d}) and (\ref{fact2 d}). The proof of
this theorem is complete now.
\end{proof}
\section{Functional CLT}
In this section we give the functional CLT form for Theorem \ref{Cor mn}. It
should be noted that for $d=1$ the quenched functional CLT in the
corresponding setting is due to Cuny and Voln\'{y} (2013). Their approach is
based on an almost sure maximal martingale approximation and involves the
introduction of two new parameters. This method cannot be easily applied for
random fields since it leads to quite complicated remainder terms in the
maximal martingale approximation. Fortunately, their innovative idea of using
the maximal operator can be also applied for random fields, as we shall see in
the direct proof bellow.
For $(s,t)\in\lbrack0,1]^{2},$ we introduce the stochastic process
\[
W_{n,m}(t,s)=\frac{1}{\sqrt{nm}}S_{[nt],[ms]}.
\]
where $[x]$ denotes the integer part of $x.$ We shall denote by
$(W(t,s))_{(t,s)\in\lbrack0,1]^{2}}$ the standard $2$-dimensional Brownian
sheet and we shall investigate the weak convergence in $D([0,1]^{2})$ endowed
with the uniform topology of $(W_{n,m}(t,s))$ to $(W(t,s)).$ As usual, the
proof of this theorem involves two steps, namely the proof of the convergence
of the finite dimensional distributions to the corresponding ones of the
standard $2$-dimensional Brownian sheet and tightness.
We call the random field $(X_{k,\ell})$ defined by (\ref{defXfield}) regular
if
\begin{equation}
E(X_{0,0}|\mathcal{F}_{0,-\infty})=0\text{ a.s. and$\;$ }E(X_{0,0
|\mathcal{F}_{-\infty,0})=0\text{ a.s.} \label{regX
\end{equation}
Our first result provides a necessary condition for tightness.
\begin{prop}
\label{proptight}Assume that the random field is regular and in addition, for
$q>2,$ we have
\begin{equation}
\sum_{i,j\geq0}\left\Vert \mathcal{P}_{-i,-j}(X_{0,0})\right\Vert _{q}<\infty.
\label{projdelta
\end{equation}
Then $W_{n,m}(t,s)$ is tight in $D([0,1]^{2})$ endowed with the uniform topology.
\end{prop}
\begin{proof}
[Proof of Proposition \ref{proptight}]We shall start the proof of this theorem
by a preliminary consideration: For $2<p<q,$ let us introduce the function
\begin{align*}
f_{i,j,p}^{\ast} & =\sup_{n,v\geq1}\frac{1}{nv}\sup_{n,v}\sum_{k=1}^{n
\sum_{\ell=1}^{v}E^{\omega}(|\mathcal{P}_{k-i,\ell-j}(X_{k,\ell})|^{p})\\
& =\sup_{n,v\geq1}\frac{1}{nv}\sum_{k=1}^{n}\sum_{\ell=1}^{v}Q_{1}^{k
Q_{2}^{\ell}(|\mathcal{P}_{-i,-j}(X_{0,0})|^{p}).
\end{align*}
Let us mention first that, by Corollary 1.7 in Chapter 6 of Krengel (1985)
applied to the function $|\mathcal{P}_{-i,-j}(X_{0,0})|^{p},$ for $\lambda>1$
we hav
\[
\lambda^{p}P\left( \left( f_{i,j,p}^{\ast}\right) ^{1/p}>\lambda\right)
\leq CE\left( |\mathcal{P}_{-i,-j}(_{0,0})|^{p}\log^{+}|\mathcal{P
_{-i,-j}(_{0,0})|\right) \leq CE\mathbf{|}\mathcal{P}_{-i,-j}(X_{0,0})|^{q}.
\]
It follows that $\left( f_{i,j,p}^{\ast}\right) ^{1/p}$ belongs to the weak
space $L^{p,\mathrm{weak}}$ defined b
\[
L^{p,\mathrm{weak}}=\{f\text{ real-valued measurable function defined on
}\Omega:\text{ }\sup_{\lambda>0}\lambda^{p}P(|f|>\lambda)<\infty\}.
\]
This is a Banach space whose norm will be denoted by $\left\Vert
\cdot\right\Vert _{p,\mathrm{weak}}$ and it is equivalent to the pseudo-norm
$\left( \sup_{\lambda>0}\lambda^{p}P(|f|>\lambda)\right) ^{1/p}$. We have
tha
\[
\left\Vert \sum_{i,j\geq0}\left( f_{i,j,p}^{\ast}\right) ^{1/p}\right\Vert
_{p,\mathrm{weak}}\leq\sum_{i,j\geq0}\left\Vert \left( f_{i,j,p}^{\ast
}\right) ^{1/p}\right\Vert _{p,\mathrm{weak}}\leq\sum_{i,j\geq0}\left\Vert
\mathcal{P}_{-i,-j}(X_{0,0})\right\Vert _{q}.
\]
Therefore, if $\sum_{i,j\geq0}\left\Vert \mathcal{P}_{-i,-j}(X_{0,0
)\right\Vert _{q}<\infty$ the
\[
\sum_{i,j\geq0}\left( f_{i,j,p}^{\ast}\right) ^{1/p}<\infty\text{
\ }P-\text{a.s.
\]
For proving tightness we shall verify the moment condition given in relation
(3) in Bickel and Wichura (1971). To verify it, denote an increment of the
process $W_{n,m}(t,s)$ on the rectangle $A=[t_{1},t_{2})\times\lbrack
s_{1},s_{2})$ by
\[
\Delta(A)=\frac{1}{\sqrt{nm}}|\sum\nolimits_{k=[nt_{1}]}^{[nt_{2}]-1
\sum\nolimits_{\ell=[ms_{1}]}^{[ms_{2}]-1}X_{k,\ell}|.
\]
Let us note that by (\ref{regX}) we have the representatio
\[
X_{k,\ell}=\sum_{i,j\geq0}\mathcal{P}_{k-i,\ell-j}(X_{k,\ell})\ \text{a.s.
\]
Fix $\omega$ where this representation holds for all $k$ and $\ell$ and also
$\sum_{i,j\geq0}\left( f_{i,j,p}^{\ast}\right) ^{1/p}<\infty$. Therefore we
hav
\[
\left\Vert \Delta(A)\right\Vert _{\omega,p}\leq\frac{1}{\sqrt{nm}
\sum_{i,j\geq0}\Vert\sum\nolimits_{k=[nt_{1}]}^{[nt_{2}]-1}\sum\nolimits_{\ell
=[ms_{1}]}^{[ms_{2}]-1}\mathcal{P}_{k-i,\ell-j}(X_{k,\ell})\Vert_{\omega
,p}\text{ },
\]
where $\left\Vert \cdot\right\Vert _{\omega,p}$ denotes the norm in
$L^{p}(P^{\omega}).$ Note that, because we have to compute the $p$-th moments
of an ortho-martingale, we can use the Burkholder inequality as given in
Theorem 3.1 of Fazekas (2005) and obtai
\[
\left\Vert \Delta(A)\right\Vert _{\omega,p}\leq\frac{C_{p}}{\sqrt{nm}
\sum_{i,j\geq0}\Vert\sum\nolimits_{k=[nt_{1}]}^{[nt_{2}]-1}\sum\nolimits_{\ell
=[ms_{1}]}^{[ms_{2}]-1}\mathcal{P}_{k-i,\ell-j}^{2}(X_{k,\ell})\Vert
_{\omega,p/2}^{1/2}.
\]
By applying now twice, consecutively, the Cauchy-Schwartz inequality, we
obtain
\[
E^{\omega}(\Delta^{p}(A))\leq C_{p}[(t_{2}-t_{1})(s_{2}-s_{1})]^{p/2}\left\{
\sum_{i,j\geq0}\left( f_{i,j,p}^{\ast}(\omega)\right) ^{1/p}\right\} ^{p}.
\]
If $B$ is a neighboring rectangle of $A$, by the H\"{o}lder inequality we
have
\[
E^{\omega}(\Delta^{p/2}(A)\Delta^{p/2}(B))\leq K_{p,\omega}\left( \mu
(A)\mu(B)\right) ^{p/4},
\]
where $\mu$ is the Lebesgue measure on $[0,1]^{2}$. Therefore the moment
condition in relation (3) in Bickel and Wichura (1971) is verified with
$\gamma=p$ and $\beta=p/2.$ Since $\beta>1$ the tightness follows from Theorem
3 in Bickel and Wichura (1971).
\end{proof}
\begin{theorem}
\label{Thm2nmfunct} Assume that (\ref{projdelta}) and (\ref{higher moment 2})
are satisfied. Then for $P$-almost all $\omega,$ the sequence of processes
$(W_{n,m}(t,s))_{n,m\geq1}$ converges in distribution on $D([0,1]^{2})$
endowed with uniform topology to $\sigma W(t,s),$ as $n\wedge m\rightarrow
\infty$ under $P^{\omega}$.
\end{theorem}
\begin{proof}
[Proof of Theorem \ref{Thm2nmfunct}]The tightness follows by Proposition
\ref{proptight}. The proof of the convergence of finite dimensional
distributions is based on the following observation. By combining the
martingale approximation in (\ref{negli}) with the negligibility of $R_{n,m}$
in (\ref{negliR}), for almost all $\omega$ and all rational numbers $0\leq
s,t\leq1$ we obtain
\[
\lim_{n\wedge m\rightarrow\infty}\frac{\left\Vert S_{[nt],[ms]}-M_{[nt],[ms]
\right\Vert _{\omega,2}}{(nm)^{1/2}}=0.
\]
Whence, by using the Cram\`{e}r-Wold device and then the triangle inequality,
we deduce that the convergence of the finite dimensional
distributions\ follows from the corresponding result for ortho-martingales.
But this fact was already proved in Peligrad and Voln\'{y} (2019). The proof
is complete.
\end{proof}
Similarly we obtain the following result
\begin{theorem}
\label{Thm1functional} Assume that (\ref{higher moment q}) is satisfied with
$q>2$. Then the conclusion of Theorem \ref{Thm2nmfunct} holds.
\end{theorem}
Let us formulate the multi-indexed form of this result:
For $\mathbf{t}\ \in\lbrack0,1]^{d},$ where $d$ is fixed a positive integer,
we introduce the stochastic random field
\[
W_{\mathbf{n}}(\mathbf{t})=\frac{1}{\sqrt{|\mathbf{n}|}}S_{[n_{1
t_{1}]...[n_{d}t_{d}]
\]
and denote by $(W(\mathbf{t}))_{\mathbf{t}\in\lbrack0,1]^{d}}$ the standard
$d$-dimensional Brownian sheet. The following is the d-dimensional version of
Theorem \ref{Thm1functional}.
\begin{theorem}
\label{Th funct CLT d}Under the conditions of Theorem \ref{Cor nn nm d} with
$q>2$, for $P$-almost all $\omega,$ the sequence of processes $(W_{\mathbf{n
}(\mathbf{t}))_{\mathbf{n}\geq\mathbf{1}}$ converges in distribution to
$\sigma W(\mathbf{t}),$ as $\min_{1\leq i\leq d}n_{i}\rightarrow\infty$ under
$P^{\omega}$.
\end{theorem}
\section{Examples}
We shall give examples providing new results for linear and Volterra random
fields. The interest of considering these applications is for obtaining
functional quenched CLT\ by using more general sequences of constants than in
Peligrad and Voln\'{y} (2019), where a coboundary decomposition was used.
Let $d$ be an integer greater than $2$ and $q>2$. Throughout this section, as
before, we denote by $C_{q}>0$ a generic constant depending on $q$, which may
be different from line to line.
\begin{example}
\label{linear} (Linear field) Let $(\xi_{\mathbf{n}})_{\mathbf{n}\in Z^{d}}$
be a random field of independent, identically distributed random variables,
which are centered and $E\left( |\xi_{\mathbf{0}}|^{q}\right) <\infty$. For
$\mathbf{k}\geq\mathbf{0}$ defin
\[
X_{\mathbf{k}}=\sum_{\mathbf{j}\geq\mathbf{0}}a_{\mathbf{j}}\xi_{\mathbf{k
-\mathbf{j}}\text{ }.
\]
Assume that
\begin{equation}
\sum_{\mathbf{k}\geq\mathbf{1}}\frac{1}{|\mathbf{k}|^{1/q}}\biggl(\sum
_{\mathbf{j}\geq\mathbf{k-1}}a_{\mathbf{j}}^{2}\biggr)^{\frac{1}{2}}<\infty.
\label{cond lin
\end{equation}
Then the quenched functional CLT in Theorem \ref{Th funct CLT d} holds.
\end{example}
\begin{proof}
Since
\[
E_{\mathbf{1}}(X_{\mathbf{k}})=\sum_{\mathbf{j\geq k-1 }}a_{\mathbf{j}
\xi_{\mathbf{k-j}},
\]
by the independence of $\xi_{\mathbf{n}}$ and the Rosenthal inequality (see
Theorem \ref{Rosenthal ineq}, given in the Appendix), we obtai
\begin{align*}
\Vert{E_{\mathbf{1}}(X_{\mathbf{k}})}\Vert_{q}^{q} & =\Vert{\sum
_{\mathbf{j\geq k-1}}a_{\mathbf{j}}\xi_{\mathbf{k-j}}}\Vert_{q}^{q}\\
& \leq C_{q}\left[ \biggl(\sum_{\mathbf{j\geq k-1}}a_{\mathbf{j}}^{2
E(\xi_{\mathbf{k-j}}^{2})\biggr)^{\frac{q}{2}}+\sum_{\mathbf{j\geq k-1
}E|a_{\mathbf{j}}\xi_{\mathbf{k-j}}|^{q}\right] \\
& \leq C_{q}\left[ \biggl(\sum_{\mathbf{j\geq k-1}}a_{\mathbf{j}
^{2}\biggr)^{\frac{q}{2}}\left( E\xi_{\mathbf{0}}^{2}\right) ^{\frac{q}{2
}+\sum_{\mathbf{j\geq k-1}}|a_{\mathbf{j}}|^{q}E\left( |\xi_{\mathbf{0}
|^{q}\right) \right] .
\end{align*}
By the monotonicity of norms in $\ell_{p}$, we hav
\[
\biggl(\sum_{\mathbf{j\geq k-1}}|a_{\mathbf{j}}|^{q}\biggr)^{\frac{1}{q}
\leq\biggl(\sum_{\mathbf{j\geq k-1}}a_{\mathbf{j}}^{2}\biggr)^{\frac{1}{2}}.
\]
Therefore
\[
\Vert{E_{\mathbf{1}}(X_{\mathbf{k}})}\Vert_{q}\leq C_{q}\biggl(\sum
_{\mathbf{j\geq k-1}}a_{\mathbf{j}}^{2}\biggr)^{\frac{1}{2}}.
\]
So condition (\ref{higher moment d}) is implied by (\ref{cond lin}). Whence
the result in Theorem \ref{Th funct CLT d} holds.
\end{proof}
\begin{remark}
\label{rmdtwo}For the case $d=2$, we can assume the following condition
imposed to the coefficients:
\begin{equation}
\sum_{\mathbf{k}\geq\mathbf{1}}\frac{1}{|\mathbf{k}|^{1/2}}\biggl(\sum
_{\mathbf{j}\geq\mathbf{k-1}}a_{\mathbf{j}}^{2}\biggr)^{\frac{1}{2}}<\infty.
\label{cond linear
\end{equation}
(a) If we assume $E\left( \xi_{\mathbf{0}}^{2}\right) <\infty$, then the
quenched convergence in (\ref{Quenched nn}) holds.
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \linebreak(b) If we assume $E\left(
|\xi_{\mathbf{0}}|^{2}\log(1+|\xi_{\mathbf{0}}|)\right) <\infty$, the
quenched convergence in (\ref{quenched mn}) holds.
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \linebrea
(c) If we assume $E|\xi_{\mathbf{0}}|^{q}<\infty$, for $q>2$ then, the quench
functional CLT in Theorem \ref{Thm2nmfunct}, holds.
\end{remark}
\begin{proof}
The first part of the remark is a direct consequence of item (a) in Theorem
\ref{Cor mn}, since condition (\ref{higher moment 2}) is implied by
(\ref{cond linear}). To prove (b), notice that by algebraic manipulations
similar to those used in the proof of Lemma \ref{fact3},\ we have that
(\ref{cond linear}) implies that $\sum_{\mathbf{j}\geq\mathbf{0
}|a_{\mathbf{j}}|<\infty.$
Note that
\[
\mathcal{P}_{\mathbf{0}}(X_{\mathbf{u}})=a_{\mathbf{u}}\xi_{\mathbf{0}
\]
and therefore we have
\[
\text{ }\sum_{\mathbf{j}\geq\mathbf{0}}\Vert\mathcal{P}_{\mathbf{0
}(X_{\mathbf{j}})\Vert_{\phi}=\sum_{\mathbf{j}\geq\mathbf{0}
|a_{\mathbf{j}}|\cdot\Vert\xi_{\mathbf{0}}\Vert_{\phi}<\infty.
\]
It follows that (\ref{cond linear}) also implies (\ref{main condi}). Therefore
the second part of the remark follows by item (b) in Theorem \ref{Cor mn}.
Part (c) of the remark follows in a similar way.
\end{proof}
\bigskip
For example, note that (\ref{cond linear}) holds if we take
\[
a_{u,v}=\frac{1}{uv}\frac{1}{h(u)g(v)},
\]
with $h$, $g$ slowly varying functions at infinity satisfying
\[
\sum_{u\geq1}\frac{1}{uh(u)}<\infty\text{ and }\sum_{v\geq1}\frac{1
{vg(v)}<\infty.
\]
Also, we mention that the quenched convergence in (a) does not hold if only
$\sum_{\mathbf{j}\geq\mathbf{0}}|a_{\mathbf{j}}|<\infty$ as shown for $d=1$ in
Voln\'{y} and Woodroofe (2010).
\begin{example}
\label{Volterra}(Volterra field) Let $(\xi_{\mathbf{n}})_{\mathbf{n}\in Z^{d
}$ be a random field of independent, identically distributed, and centered
random variables satisfying $E\left( |\xi_{\mathbf{0}}|^{q}\right) <\infty$.
For $\mathbf{k}\geq\mathbf{0}$, defin
\[
X_{\mathbf{k}}=\sum_{(\mathbf{u},\mathbf{v)}\geq(\mathbf{0},\mathbf{0
)}a_{\mathbf{u},\mathbf{v}}\xi_{\mathbf{k-u}}\xi_{\mathbf{k-v}}.
\]
where $a_{\mathbf{u},\mathbf{v}}$ are real coefficients with $a_{\mathbf{u
,\mathbf{u}}=0$ and $\sum_{\mathbf{u,v}\geq\mathbf{0}}a_{\mathbf{u,v}
^{2}<\infty$. In addition, assume tha
\begin{equation}
\sum_{\mathbf{k\geq1}}\frac{1}{|\mathbf{k}|^{1/q}}\biggl(\sum
_{\substack{\mathbf{(u,v)\geq(k-1,k-1)}\\\mathbf{u\neq v}}}a_{\mathbf{u,v
}^{2}\biggr)^{1/2}<\infty. \label{cond volt
\end{equation}
Then the quenched functional CLT in Theorem \ref{Th funct CLT d} holds.
\end{example}
\begin{proof}
Note tha
\[
E_{\mathbf{1}}(X_{\mathbf{k}})=\sum_{\mathbf{(u,v)\geq(k-1,k-1)
}a_{\mathbf{u,v}}\xi_{\mathbf{k-u}}\xi_{\mathbf{k-v}}.
\]
Let $(\xi_{\mathbf{n}}^{\prime})_{\mathbf{n}\in Z^{d}}$ and $(\xi_{\mathbf{n
}^{\prime\prime})_{\mathbf{n}\in Z^{d}}$ be two independent copies of
$(\xi_{\mathbf{n}})_{\mathbf{n}\in Z^{d}}$. By independence and the fact that
$a_{\mathbf{k,k}}=0$, by applying the decoupling inequality together with the
Rosenthal inequality, both of which are given for convenience in the Appendix,
(see Theorem \ref{decoupling thm} and Theorem \ref{Rosenthal ineq} from the
Appendix), we obtai
\begin{gather*}
\Vert{E_{\mathbf{1}}(X_{\mathbf{k}})}\Vert_{q}^{q}=\Vert{\sum
_{\substack{\mathbf{(u,v)\geq(k-1,k-1)}\\\mathbf{u\neq v}}}a_{\mathbf{u,v}
\xi_{\mathbf{k-u}}\xi_{\mathbf{k-v}}}\Vert_{q}^{q}\leq C_{2}\Vert
{\sum_{\substack{\mathbf{(u,v)\geq(k-1,k-1)}\\\mathbf{u\neq v}
}a_{\mathbf{u,v}}\xi_{\mathbf{k-u}}^{\prime}\xi_{\mathbf{k-v}}^{\prime\prime
}\Vert_{q}^{q}\\
\leq C_{q}\biggl[\biggl(\sum_{\substack{\mathbf{(u,v)\geq(k-1,k-1)
\\\mathbf{u\neq v}}}a_{\mathbf{u,v}}^{2}E(\xi_{\mathbf{k-u}}^{\prime
\xi_{\mathbf{k-v}}^{\prime\prime})^{2}\biggr)^{\frac{q}{2}}+\sum
_{\substack{\mathbf{(u,v)\geq(k-1,k-1)}\\\mathbf{u\neq v}}}|a_{\mathbf{u,v
}|^{q}E\left( |\xi_{\mathbf{k-u}}^{\prime}\xi_{\mathbf{k-v}}^{\prime\prime
}|^{q}\right) \biggr]\\
\leq C_{q}\biggl[\biggl(\sum_{\substack{\mathbf{(u,v)\geq(k-1,k-1)
\\\mathbf{u\neq v}}}a_{\mathbf{u,v}}^{2}\biggr)^{\frac{q}{2}}E(\xi
_{\mathbf{0}}^{2})^{q}+\sum_{\substack{\mathbf{(u,v)\geq(k-1,k-1)
\\\mathbf{u\neq v}}}|a_{\mathbf{u,v}}|^{q}E\left( |\xi_{\mathbf{0}
|^{q}\right) ^{2}\biggr].
\end{gather*}
Above, the first inequality holds by Theorem \ref{decoupling thm} while the
second one is implied by Theorem \ref{Rosenthal ineq}.
Again by the monotonicity of norms in $\ell_{p}$, we hav
\[
\Vert{E_{\mathbf{1}}(X_{\mathbf{k}})}\Vert_{q}\leq C_{q}\biggl(\sum
_{\mathbf{u,v\geq k-1}}a_{\mathbf{u,v}}^{2}\biggr)^{\frac{1}{2}}.
\]
Thus the results of Theorem \ref{Th funct CLT d} hold.
\end{proof}
\section{Appendix}
For convenience, we mention one classical inequality for martingales, see
Theorem 2.11, p. 23, Hall and Heyde (1980) and also Theorem 6.6.7 Ch. 6, p.
322, de la Pe\~{n}a and Gin\'{e} (1999).
\begin{theorem}
[Rosenthal's Inequality]\label{Rosenthal ineq} Let $p\geq2$. Let $M_{n
=\sum_{k=1}^{n}X_{k}$ where $\{M_{n},\mathcal{F}_{n}\}$ is a martingale with
martingale differences $(X_{n})$. Then there are constants $0<c_{p},
C_{p}<\infty$ such tha
\begin{align*}
& c_{p}\biggl\{\sum_{k=1}^{n}E|X_{k}|^{p}+E\biggl[\biggl(\sum_{k=1
^{n}E(X_{k}^{2}|\mathcal{F}_{k-1})\biggr)^{p/2}\biggr]\biggr\}\\
& \leq\Vert{M_{n}}\Vert_{p}^{p}\leq C_{p}\biggl\{E\biggl[\biggl(\sum
_{k=1}^{n}E(X_{k}^{2}|\mathcal{F}_{k-1})\biggr)^{p/2}\biggr]+\sum_{k=1
^{n}E|X_{k}|^{p}\biggr\}.
\end{align*}
\end{theorem}
The following theorem is a decoupling result for U-statistics, which can be
found on p. 99, Theorem 3.1.1, de la Pe\~{n}a and Gin\'{e} (1999).
\begin{theorem}
[Decoupling inequality]\label{decoupling thm} Let $(X_{i})_{1\leq i\leq n}$ be
$n$ independent random variables and let $(X_{i}^{k})_{1\leq i\leq n}$,
$k=1,\cdots,m$, be $m$ independent copies of this sequences. For each
$(i_{1},i_{2},\cdots,i_{m})\in I_{n}^{m}$, let $h_{i_{1},\cdots,i_{m}
:R^{m}\rightarrow R$ be a measurable function such $E\lvert h_{i_{1
,\cdots,i_{m}}(X_{i_{1}},\cdots,X_{i_{m}})\rvert<\infty$. Let $f:[0,\infty
)\rightarrow\lbrack0,\infty)$ be a convex non-decreasing function such that
$Ef(\lvert h_{i_{1},\cdots,i_{m}}(X_{i_{1}},\cdots,X_{i_{m}})\rvert)<\infty$
for all $(i_{1},i_{2},\cdots,i_{m})\in I_{n}^{m}$, where $I_{n}^{m}=\left\{
(i_{1},\cdots,i_{m}):i_{j}\in\mathbb{N},1\leq i_{j}\leq n,i_{j}\neq
i_{k},\text{ if }j\neq k\right\} $. Then there exists $C_{m}>0$ such tha
\[
Ef(\lvert\sum_{I_{n}^{m}}h_{i_{1},\cdots,i_{m}}(X_{i_{1}},\cdots,X_{i_{m
})\rvert)\leq Ef(C_{m}\lvert\sum_{I_{n}^{m}}h_{i_{1},\cdots,i_{m}}(X_{i_{1
}^{1},\cdots,X_{i_{m}}^{m})\rvert).
\]
\end{theorem}
\section{Acknowledgements}
This research was supported in part by the NSF grant DMS-1811373. Also we
acknowledge a support from the Taft research center at the University of
Cincinnati and a grant provided by MIIS (Math\'{e}matiques, Information,
Ing\'{e}nierie des Syst\`{e}mes), Universit\'{e} de Rouen Normandie. The
authors are grateful to Dalibor Voln\'{y} for carefully reading the manuscript
and useful discussions, and to the referee for suggestions that contributed to
an improvement of a previous version of the paper.
|
\section{Introduction}
Observations obeying certain physical laws can often be described by a partial differential equation (PDE). Real world measurements carry statistical noise and thus do not generally exactly exhibit the idealised pattern of the PDE, but it is desirable that recovery of parameters from data is consistent with the PDE structure. In the mathematical literature on inverse problems several algorithms that incorporate such constraints have been proposed, notably optimisation based methods such as Tikhonov regularisation \cite{EHN96, BB18} and maximum a posteriori (MAP) estimates related to Bayesian inversion techniques \cite{stuart10, DS16}. In statistical terminology these methods can be viewed as penalised least squares estimators over parameter spaces of regression functions that are restricted to lie in the range of some `forward operator' $\mathscr G$ describing the solution map of the PDE. The case where $\mathscr G$ is linear is reasonably well studied in the inverse problems literature, but already in basic elliptic PDE examples, the map $\mathscr G$ is \textit{non-linear} and the analysis is more involved. The observation scheme considered here will be a natural continuous analogue of the standard Gaussian regression model
\begin{equation}\label{intro-obs}
Y_i=u_f(x_i)+\varepsilon_i, ~i=1, \dots, n; ~\{\varepsilon_i\} \sim^{i.i.d.}N(0,1),
\end{equation}
where $(x_i)_{i=1}^n$ are `equally spaced' design points on a bounded domain $\mathcal O \subset \mathbb R^d$ with smooth boundary $\partial \mathcal O$. The function $u_f: \mathcal O \to \mathbb R$ is, in our first example, the solution $u=u_f$ of the elliptic PDE (with $\nabla$ denoting the gradient and $\nabla \cdot$ the divergence operator)
\begin{equation}\label{intro-div}
\begin{cases}
\nabla\cdot (f\nabla u) =g \quad \textnormal{on } \mathcal O,\\
u=0\quad \textnormal{on } \mathcal \partial O,
\end{cases}
\end{equation}
where $g>0$ is a given source function defined on $\mathcal O$ and $f: \mathcal O \to (0,\infty)$ is an unknown \textit{conductivity (or diffusion) coefficient}. The second model example arises with solutions $u=u_f$ of the time-independent Schr\"odinger equation (with $\Delta$ equal to the standard Laplacian operator)
\begin{equation}\label{intro-sch}
\begin{cases}
\Delta u -2fu =0 \quad \textnormal{on } \mathcal O,\\
u=g\quad \textnormal{on } \mathcal \partial O,
\end{cases}
\end{equation}
corresponding to the unknown \textit{attenuation potential (or reaction coefficient)} $f: \mathcal O \to (0,\infty)$, and given positive `boundary temperatures' $g>0$. Both PDEs have a fundamental physical interpretation and feature in many application areas, see, e.g., \cite{EHN96, BHM04, HP08, BU10, stuart10, devore, DS16}, and references therein.
When $f>0$ belongs to some Sobolev space $H^\alpha(\mathcal O)$ for appropriate $\alpha>0$, unique solutions $u_f$ of the PDEs (\ref{intro-div}), (\ref{intro-sch}) exist, and the `forward' map $f \mapsto u_f$ is non-linear. [In fact, in (\ref{intro-sch}) only $f \ge 0$ is required.] A natural method to estimate $f$ is by a penalised least squares approach: one minimises over $f \in H^\alpha(\mathcal O)$ with $f>0$ the squared Euclidean distance $$Q_n(f)=\|Y-u_f\|^2$$ of the observation vector $(Y_i:i=1, \dots, n)$ to the fitted values $(u_f(x_i):i=1, \dots, n)$, and penalises too complex solutions $f$ by, for instance, an additive Sobolev norm $\|\cdot\|_{H^\alpha}$ - type penalty. The (from a PDE perspective) natural constraint $f>0$ can be incorporated by a smooth one-to-one transformation $\Phi$ of the penalty function, and a final estimator $\hat f$ minimises a criterion function of the form $$Q_n(f) + \lambda^2 \|\Phi^{-1}[f]\|_{H^\alpha}^2,$$ over $f \in H^\alpha(\mathcal O)$ with $f>0$, where $\lambda$ is a scalar regularisation parameter to be chosen. Both Tikhonov regularisers as well as Bayesian maximum a posteriori (MAP) estimates arising from suitable Gaussian priors fall into this class of estimators. We show in the present paper that suitable choices of $\lambda, \alpha, \Phi$ give rise to statistically optimal solutions of the above PDE constrained regression problems from data (\ref{intro-obs}), in prediction loss. The convergence rates obtained can be combined with `stability estimates' to obtain bounds also for the recovery of the parameter $f$ itself.
Our main results are based on a general convergence rate theorem for minimisers over $H^\alpha$ of functionals of the form $$F \mapsto \|Y-\mathscr G(F)\|^2 + \lambda^2 \|F\|_{H^\alpha}^2$$ in possibly non-linear inverse problems whose forward map $F \mapsto \mathscr G(F)$ satisfies a certain modulus of continuity assumption between Hilbert spaces. This result, which adapts $M$-estimation techniques \cite{sara2001, saramest} to the inverse problems setting, is of independent interest, and provides novel results also for linear forward maps, see Remark \ref{radon} for an application to Radon transforms.
For sake of conciseness, our theory is given in the Gaussian white noise model introduced in (\ref{data}) below -- it serves as an asymptotically equivalent (see \cite{BL96, R08}) continuous analogue of the discrete model (\ref{intro-obs}), and facilitates the application of PDE techniques in our proofs. Transferring our results to discrete regression models is possible, but the additional difficulties are mostly of a technical nature and will note be pursued here.
Recovery for non-linear inverse problems such as those mentioned above has been studied initially in the deterministic regularisation literature \cite{EKN89, N92, SEK93, EHN96, TJ02}, and the convergence rate theory developed there has been adapted to the statistical regression model (\ref{intro-obs}) in \cite{BHM04, BHMR07, HP08, LL10}. These results all assume that a suitable Fr\'echet derivative $D\mathscr G$ of the non-linear forward map $\mathscr G$ exists at the `true' parameter $F$, and moreover require that $F$ lies in the range of the adjoint operator of $D\mathscr G$ -- the so called `source condition'. Particularly for the PDE (\ref{intro-div}), such conditions are problematic and do not hold in general for rich enough classes of $F$'s (such as Sobolev balls) unless one makes very stringent additional model assumptions. Our results circumvent such source conditions. Further remarks, including a discussion of related convergence analysis of estimators obtained from Bayesian inversion techniques \cite{v13, dashti13, n17} can be found in Section \ref{sec-pde-rem}.
The article is organised as follows. The main results are stated in Sections \ref{sec-general} and \ref{sec-pderes}; their proofs are contained in Sections \ref{sec-pfs} and \ref{sec-ex-pf}. Some key auxiliary results about the elliptic PDE (\ref{intro-div})-(\ref{intro-sch}) and the `link functions' $\Phi$ used below are proved in Section \ref{sec-pde-facts} and \ref{sec-reg} respectively.
\subsection{Some preliminaries and basic notation}\label{sec-gen-pre}
Throughout, $\mathcal O\subseteq \mathbb R^d$, $d\geq 1$, denotes a bounded non-empty $C^\infty$-domain (an open bounded set with smooth boundary) with closure $\bar{\mathcal O}$. The usual space $L^2(\mathcal O)$ of square integrable functions carries a norm $\|\cdot\|_{L^2(\mathcal O)}$ induced by the inner product $$\langle h_1, h_2 \rangle_{L^2(\mathcal O)} = \int_\mathcal O h_1(x)h_2(x)dx,~~h_1,h_2 \in L^2(\mathcal O),$$ where $dx$ denotes Lebesgue measure. For any multi-index $i=(i_1,...,i_d)$ of `order' $|i|$, let $D^i$ denote the $i$-th (weak) partial derivative operator of order $|i|$. Then for integer $\alpha\geq 0$, the usual Sobolev spaces are defined as
\[H^\alpha(\mathcal O):=\left\{f\in L^2(\mathcal O)\; \middle| \;\textnormal{for all }|i|\leq \alpha, \; D^if \textnormal{ exists and } D^if\in L^2(\mathcal O) \right\},\]
normed by $\|f\|_{H^\alpha(\mathcal O)} = \sum_{|i| \le \alpha} \|D^if\|_{L^2(\mathcal O)}$. For non-integer real values $\alpha\geq 0$, we define $H^\alpha(\mathcal O)$ by interpolation, see, e.g., \cite{lionsmagenes} or \cite{T78}.
\par
The spaces of bounded and continuous functions on $\mathcal O$ and $\bar{\mathcal O}$ are denoted by $C(\mathcal O)$ and $C(\bar{\mathcal O})$, respectively, equipped with the supremum norm $\|\cdot\|_\infty$. For $\eta\in\mathbb{N}$, the space of $\eta$-times differentiable functions on $\mathcal O$ with (bounded) uniformly continuous derivatives is denoted by $C^\eta(\mathcal O)$. For $\eta > 0,\eta\notin \mathbb{N}$, we say $f \in C^\eta(\mathcal O)$ if for all multi-indices $\beta$ with $|\beta|\leq \lfloor\eta\rfloor $ (the integer part of $\eta$), $D^\beta f$ exists and is $\eta-\lfloor\eta\rfloor$-H\"older continuous. The norm on $C^\eta(\mathcal O)$ is
\[\|f\|_{C^\eta(\mathcal O)}=\sum_{\beta:|\beta|\leq\lfloor\eta\rfloor}\|D^\beta f\|_{\infty}+\sum_{\beta:|\beta|=\lfloor\eta\rfloor}\sup_{x,y\in\mathcal O, \;x\neq y}\frac{|D^\beta f(x)-D^\beta f(y)|}{|x-y|^{\eta-\lfloor \eta\rfloor}}. \] We also define the set of smooth functions as $C^\infty(\mathcal O)=\cap_{\eta>0} C^\eta(\mathcal O)$ and its subspace $C^\infty_c(\mathcal O)$ of functions compactly supported in $\mathcal O$.
The previous definitions will be used also for $\mathcal O$ replaced by $\partial \mathcal O$ or $\mathbb R^d$. When there is no ambiguity, we omit $\mathcal O$ from the notation.
\par
For any normed linear space $(X,\|\cdot\|_X)$ its topological dual space is
\[X^*:=\left\{L:X\to \mathbb R \textnormal{ linear s.t. } \exists C>0 \;\forall x\in X: \;|L(x)|\leq C\|x\|_X \right\},\]
which is a Banach space for the norm $\|L\|_{X^*}=\sup_{x\in X}|L(x)|/ \|x\|_{X}.$
We need further Sobolev-type spaces to address routine subtleties of the behaviour of functions near $\partial \mathcal O$: denote by $H^\alpha_c(\mathcal O)$ the completion of $C^\infty_c(\mathcal O)$ for the $H^\alpha(\mathcal O)$-norm, and let $\tilde H^\alpha(\mathcal O)$ denote the closed subspace of $H^\alpha(\mathbb R^d)$ consisting of functions supported in $\bar {\mathcal O}$. We have $H^\alpha_c(\mathcal O)=\tilde H^\alpha(\mathcal O)$ unless $\alpha=k+1/2, k \in \mathbb N$ (Section 4.3.2 in \cite{T78}), and one defines negative order Sobolev spaces $H^{-\kappa}(\mathcal O)=(\tilde H^\kappa(\mathcal O))^*, \kappa>0$, cf.~also Theorem 3.30 in \cite{ML00}.
We use the symbols ``$\lesssim, \gtrsim$'' for inequalities that hold up to multiplicative constants that are universal, or whose dependence on other constants will be clear from the context. We also use the standard notation $\mathbb{R}_+:= \{x|x\geq 0\}$ and $a\vee b:=\max\{a,b\}$ for $a,b\in \mathbb{R}$.
\section{A convergence rate result for general inverse problems}\label{sec-general}
\subsection{Forward map and white noise model}
Let $\mathbb H$ be a separable Hilbert space with inner product $\langle \cdot, \cdot \rangle_\mathbb H$. Suppose that $\tilde{\mathcal V}\subseteq L^2(\mathcal O)$ and that
\begin{equation*}
\mathscr G:\tilde{\mathcal V}\to \mathbb H, \qquad F\mapsto \mathscr G(F),
\end{equation*}
is a given `forward' map. For some $F\in \tilde {\mathcal V}$, and for scalar `noise level' $\varepsilon>0$, we observe a realisation of the equation
\begin{equation}\label{data}
Y^{(\varepsilon)}=\mathscr G(F)+\varepsilon \mathbb W,
\end{equation}
where $(\mathbb W(\psi):\psi\in \mathbb H)$ is a centred Gaussian white noise process indexed by the Hilbert space $\mathbb H$ (see p.19-20 in \cite{nicklgine}). Let $\mathbb E_F^{\varepsilon}, F \in \tilde{\mathcal V},$ denote the expectation operator under the law $\mathbb P_F^\varepsilon$ of $Y^{(\varepsilon)}$ from (\ref{data}). Observing (\ref{data}) means to observe a realisation of the Gaussian process $(\langle Y^{(\varepsilon)},\psi\rangle_{\mathbb H}: \psi \in \mathbb H)$ with marginal distributions $$\langle Y^{(\varepsilon)},\psi\rangle_{\mathbb H} \sim N(\langle \mathscr G(F), \psi \rangle_{\mathbb H}, \varepsilon^2\|\psi\|_{\mathbb H}^2).$$
In the case $\mathbb H=L^2(\mathcal O)$ relevant in Section \ref{sec-pderes} below, (\ref{data}) can be interpreted as a Gaussian shift experiment in the Sobolev space $H^{-\kappa}(\mathcal O), \kappa>d/2$ (see, e.g., \cite{cn1, n17}), and also serves as a theoretically convenient (and, for $\varepsilon = 1/\sqrt n$, as $n \to \infty$ asymptotically equivalent) continuous surrogate model for observing $(Y_i, x_i)_{i=1}^n$ in the standard fixed design Gaussian regression model
\begin{equation}\label{disdat}
Y_i = \mathscr G(F)(x_i) + \varepsilon_i, ~i=1, \dots, n, ~\{\varepsilon_i\} \sim^{i.i.d.} N(0,1),
\end{equation}
where the $x_i$ are `equally spaced' design points in the domain $\mathcal O$ (see \cite{BL96, R08}).
In the discrete model (\ref{disdat}) the least squares criterion can be decomposed as
$\|Y-\mathscr G(F)\|_{\mathbb R^n}^2$ $=\|Y\|_{\mathbb R^n}^2 - 2\langle Y, \mathscr G(F) \rangle_{\mathbb R^n} + \|\mathscr G(F)\|_{\mathbb R^n}^2$. The first term $\|Y\|_{\mathbb R^n}^2$ is independent of $F$ and can be neglected when optimising in $F$. In the continuous model (\ref{data}) we have $\|Y\|_{\mathbb H}=\infty$ a.s. (unless dim$(\mathbb H) <\infty$), which motivates to define a `Tikhonov-regularised' functional
\begin{equation}\label{Jdef2}
\mathscr J_{\lambda, \varepsilon}:\tilde{\mathcal V}\to \mathbb R,\quad \mathscr J_{\lambda, \varepsilon}(F):=2\langle Y^{(\varepsilon)},\mathscr G(F)\rangle_{\mathbb H}-\|\mathscr G(F)\|_{\mathbb H}^2-\lambda^2\|F\|_{H^\alpha}^2,
\end{equation}
where $\lambda>0$ is a regularisation parameter to be chosen, and where we set $\mathscr J_{\lambda, \varepsilon}(F)=-\infty$ for $F\notin H^\alpha$. Maximising $\mathscr J_{\lambda, \varepsilon}$ thus amounts to minimising the natural least squares fit with a $H^\alpha(\mathcal O)$-penalty for $F$, and we note that it also corresponds to maximising the penalised log-likehood function arising from (\ref{data}), see, e.g., \cite{n17}, Section 7.4. In all that follows $\|\cdot\|_{H^\alpha}$ could be replaced by any equivalent norm on $H^\alpha(\mathcal O)$.
We note that when $\mathscr G$ is non-linear, computation of a global maximiser of the (then non-convex) functional $\mathscr J_{\lambda, \varepsilon}$ may be infeasible in practice. Nevertheless, the convergence rates we obtain below provide a first rigorous understanding of the statistical complexity of the PDE inference problems at hand. It is an interesting open question whether algorithms that are computable in `polynomial time' can attain the same performance guarantees. This is subject of ongoing research (see, e.g., \cite{MNP19b}) and beyond the scope of the present paper.
\subsection{Results}\label{sec-gen-res}
For $F_1\in \tilde{\mathcal V} \cap H^\alpha$, $F_2\in\tilde{\mathcal V}$ and $\lambda >0$, define the functional
\begin{equation}\label{taudef}
\tau_\lambda^2(F_1,F_2):=\|\mathscr G(F_1)-\mathscr G(F_2)\|_{\mathbb H}^2+\lambda^2\|F_1\|_{H^\alpha}^2.
\end{equation}
The main result of this section, Theorem \ref{thm-gen}, proves the existence of maximisers $\hat F$ for $\mathscr J_{\lambda, \varepsilon}$ over suitable subsets $\mathcal V\subseteq \tilde{\mathcal V} \cap H^\alpha$ and concentration properties for $\tau_\lambda(\hat F, F_0)$, where $F_0$ is the `true' function generating the law $\mathbb P_{F_0}^\varepsilon$ from equation (\ref{data}). Note that bounds for $\tau_\lambda(\hat F, F_0)$ simultaneously control the `prediction error' $\|\mathscr G(\hat F)-\mathscr G(F_0)\|_{\mathbb H}$ as well as the regularity $\|\hat F\|_{H^\alpha}$ of the estimated output $\hat F$.
\smallskip
Theorem \ref{thm-gen} is proved under a general `modulus of continuity' condition on the map $\mathscr G$ which reads as follows.
\begin{defin}
Let $\alpha,\gamma, \kappa\in\mathbb{R}_{+}$ be non-negative real numbers and $\tilde {\mathcal V}\subseteq L^2(\mathcal O)$. Set $\mathcal H:=H^\alpha(\mathcal O)$ if $\kappa<1/2$, and $\mathcal H:=H^\alpha_c(\mathcal O)$ if $\kappa\geq 1/2$. A map $\mathscr G:\tilde{\mathcal V}\to \mathbb H$ is called $(\kappa,\gamma, \alpha)$\emph{-regular} if there exists a constant $C>0$ such that for all $F,H\in \tilde{\mathcal V}\cap \mathcal H$, we have
\begin{equation}\label{entrcond}
\|\mathscr G(F)-\mathscr G(H)\|_{\mathbb H}\leq C\big(1+\|F\|_{H^\alpha(\mathcal O)}^\gamma\vee \|H\|_{H^\alpha(\mathcal O)}^\gamma\big)\|F-H\|_{(H^{\kappa}(\mathcal O))^*},
\end{equation}
\end{defin}
This condition is easily checked for `$\kappa$-smoothing' \textit{linear} maps $\mathscr G$ with $\gamma=0$, see Remark \ref{radon} for an example. But (\ref{entrcond}) also allows for certain non-linearities of $\mathscr G$ on unbounded parameter spaces $\tilde {\mathcal V}$ that will be seen later on to accommodate the forward maps induced by the PDEs (\ref{intro-div}), (\ref{intro-sch}). See also Remarks \ref{nonlin}, \ref{maphack} below.
\begin{thm}\label{thm-gen}
Suppose that $\mathscr G: \tilde{\mathcal V}\to \mathbb H$ is a $(\kappa,\gamma, \alpha)$-regular map for some integer $\alpha>(d/2-\kappa) \vee (\gamma d/2-\kappa)$. Let $Y^{(\varepsilon)} \sim \mathbb P_{F_0}^\varepsilon$ from (\ref{data}) for some fixed $F_0 \in \tilde{\mathcal V}$.
Then the following holds.
\par
1. Let $\mathcal V\subseteq \tilde{\mathcal V}\cap \mathcal H$ be closed for the weak topology of the Hilbert space $\mathcal H$. Then for all $\lambda, \varepsilon>0$, almost surely under $\mathbb P_{F_0}^\varepsilon$, there exists a maximiser $\hat F=\hat F_{\lambda,\varepsilon} \in \mathcal V$ of $\mathscr J_{\lambda,\varepsilon}$ from (\ref{Jdef2}) over $\mathcal V$, satisfying
\begin{equation}\label{hatF}
\sup_{F\in \mathcal V}\mathscr J_{\lambda,\varepsilon}(F) = \mathscr J_{\lambda, \varepsilon} (\hat F).
\end{equation}
\par
2. Let $\mathcal V\subseteq \tilde{\mathcal V}\cap \mathcal H$. There exist constants $c_1,c_2,c_3>0$ such that for all
$\varepsilon,\lambda,\delta>0$ satisfying
\begin{equation}\label{delta-cond}
\varepsilon^{-1}\delta \geq c_1\big(1+\lambda^{-\frac{1}{2s}}\big(1+(\delta/\lambda)^{\frac{\gamma}{2s}}\big) \big),~s:=(\alpha+\kappa)/d,
\end{equation}
all $R\geq \delta$, any maximiser $\hat F =\hat F_{\lambda, \varepsilon} \in \mathcal V$ of $\mathscr J_{\lambda,\varepsilon}$ over $\mathcal V$ and any $F_* \in \mathcal V$, we have
\begin{equation}\label{main-thm-tau}
\mathbb P_{F_0}^\varepsilon\big(\tau_\lambda^2(\hat F,F_0)\geq 2(\tau_\lambda^2(F_*, F_0)+R^2)\big)\leq c_2\exp\Big(-\frac{R^2}{c_2^2\varepsilon^2}\Big),
\end{equation}
and also
\begin{equation}\label{main-thm-rate}
\mathbb{E}^\varepsilon_{F_0}\left[\tau_\lambda^2(\hat F,F_0)\right]\leq c_3 \left(\tau_\lambda^2(F_*, F_0)+\delta^2+ \varepsilon^2\right).
\end{equation}
\end{thm}
Various applications of Theorem \ref{thm-gen} for specific choices of $\kappa$, $\gamma$, $\mathcal V$ and $\tilde{\mathcal V}$ will be illustrated in the following - besides the main PDE applications from Section \ref{sec-pderes}, see Remarks \ref{lin}, \ref{maphack} and \ref{bdhack} as well as Example \ref{radon} below.
\par
Theorem \ref{thm-gen} does not necessarily require $F_0 \in \mathcal V$ as long as $F_0$ can be suitably approximated by some $F_* \in \mathcal V$, see Remark \ref{lin} for an instance of when this is relevant. If $F_0 \in \mathcal V$ then we can set $F_*=F_0$ in the above theorem and obtain the following convergence rates, which are well known to be optimal for $\kappa$-smoothing linear forward maps $\mathscr G$, and which will be seen to be optimal also for the non-linear inverse problems arising from the PDE models (\ref{intro-div}) and (\ref{intro-sch}).
\begin{cor}\label{cor-gen}
Under the conditions of Part 2 of Theorem \ref{thm-gen}, for all $R>0$ there exists $c<\infty$ such that for all $\varepsilon>0$ small enough, $\lambda = \varepsilon^{2(\alpha+\kappa)/(2(\alpha+\kappa)+d)}$ and any maximizer $\hat F_{\lambda,\varepsilon}$ of $\mathscr J_{\lambda,\varepsilon}$ over $\mathcal V$,
\begin{equation}\label{rate-cor}
\sup_{F_0\in\mathcal V:\|F_0\|_{H^\alpha}\leq R}\mathbb E^{\varepsilon}_{F_0}\left\|\mathscr G(\hat F_{\lambda,\varepsilon})-\mathscr G(F_0)\right\|_{\mathbb H}\leq c\varepsilon^{\frac{2(\alpha+\kappa)}{2(\alpha+\kappa)+d}}.
\end{equation}
\end{cor}
When images of $\|\cdot\|_{H^\alpha}$-bounded subsets of $\mathcal V$ under a forward map $\mathscr G: L^2(\mathcal O) \to L^2(\mathcal O)$ are bounded in $H^{\beta}(\mathcal O)$ for some $\beta>0$, then the $L^2$-bound (\ref{rate-cor}) extends (via interpolation and bounds for $\|\hat F\|_{H^\alpha}$ implied by Theorem \ref{thm-gen}) to $H^\eta$-norms, $\eta\in [0,\beta]$, which in turn can be used to obtain convergence rates also for $\hat F -F_0$ by using stability estimates. See the results in Section \ref{sec-pderes} and also Example \ref{radon} below for examples.
\begin{rem}[MAP estimates] \label{lin} \normalfont Let $\Pi$ be a Gaussian process prior measure for $F$ with reproducing kernel Hilbert space (RKHS) $\mathcal H$ and RKHS-norm $\bar \lambda \|\cdot\|_{H^\alpha}, \bar \lambda>0$. Taking note of the form of the likelihood function in the model (\ref{data}) (see, e.g., Section 7.4 in \cite{n17}), maximisers $\hat F$ of $\mathscr J_{\lambda,\varepsilon}$ over $\mathcal V=\mathcal H$ with $\lambda=\varepsilon \bar \lambda$ have a formal interpretation as maximum a posteriori (MAP) estimators for the resulting posterior distributions $\Pi(\cdot|Y^{(\varepsilon)})$, see also \cite{dashti13, HB15}. For instance, let $\alpha>d/2, \kappa\geq 0,$ and consider a \textit{linear} inverse problem where for $\beta=\alpha-d/2$ and $\tilde {\mathcal V}=H^\beta(\mathcal O)$, $\mathscr G:H^\beta(\mathcal O) \to \H$ is a linear map satisfying (\ref{entrcond}) with $\gamma=0$ for all $F,H\in H^\beta(\mathcal O)$. Then, applying Theorem \ref{thm-gen} with $\lambda= \varepsilon$ (so that $\bar \lambda =1$) and $\delta \approx \varepsilon^{(2\beta+2\kappa)/(2\beta+2\kappa+d)}$ yields
\begin{equation}\label{maprate}
\sup_{F_0 \in \tilde H^\beta(\mathcal O_0): \|F_0\|_{\tilde H^\beta}\le R}\mathbb E^{\varepsilon}_{F_0}\left\|\mathscr G(\hat F)-\mathscr G(F_0)\right\|_{\mathbb H}\lesssim \delta,~~R>0,
\end{equation}
for any fixed sub-domain $\mathcal O_0$ such that $\bar{\mathcal O_0} \subsetneq \mathcal O$. Indeed, one easily checks (\ref{delta-cond}), and given $F_0 \in \tilde H^\beta(\mathcal O_0)$ set $F_*= \zeta \mathscr F^{-1}[ (1_{[|\cdot|\le (\delta/\lambda)^{2/d}]} \mathscr F [F_0]] \in H^\alpha_c(\mathcal O)$, where $\zeta \in C^\infty_c(\mathcal O)$ is such that $\zeta=1$ on $\mathcal O_0$ and $\mathscr F$ is the Fourier transform. Then $\|F_*\|_{H^\alpha(\mathcal O)} \lesssim \delta/\lambda$ and $\|F^*-F_0\|_{(H^\kappa(\mathcal O))^*} \lesssim \|F^*-F_0\|_{H^{-\kappa}(\mathcal O)} \lesssim \delta$ in (\ref{main-thm-rate}) yield (\ref{maprate}). Similar comments apply to non-linear $\mathscr G$, with appropriate choice of $\bar \lambda$, see Remark \ref{maphack}.
\end{rem}
\begin{ex}[Rates for the Radon transform]\label{radon} \normalfont Let $\mathscr R: \tilde {\mathcal V} \equiv L^2(\mathcal O)\to \H$ be the Radon transform, where $\mathcal O=\{x\in\mathbb R^2:\|x\|<1\}$ and $\mathbb H = L^2(\Sigma), \Sigma:=(0,2\pi] \times \mathbb R$, equipped with Lebesgue measure, see p.9~in \cite{N86} for definitions. Then $\mathscr G=\mathscr R$ satisfies (\ref{entrcond}) with $\kappa=1/2, \gamma=0$ and any $\alpha \in \mathbb N$ -- see p.42 in \cite{N86} and note that our $\|\cdot\|_{(H^{1/2}(\mathcal O))^*}$-norm is the $\|\cdot\|_{H_0^{-1/2}(\mathcal O)}$-norm used in \cite{N86} (cf.~Theorem 3.30 in \cite{ML00}). Applying Corollary \ref{cor-gen} with $\alpha\geq 1$, $\mathcal V=H^\alpha_c(\mathcal O)$ and $\lambda = \varepsilon^{(2\alpha+1)/(2\alpha+3)}$ implies that for any $F_0\in H^\alpha_c(\mathcal O)$,
\begin{equation} \label{radfwdbd}
\mathbb E^\varepsilon_{F_0}\big[\|\mathscr R(\hat F_{\lambda, \varepsilon})-\mathscr R(F_0)\|^2_{L^2(\Sigma)} + \lambda^2\|\hat F_{\lambda, \varepsilon}\|^2_{H^\alpha_c(\mathcal O)}\big]\lesssim \varepsilon^{(4\alpha+2)/(2\alpha+3)}.
\end{equation}
Using again the estimates on p.42 in \cite{N86} and that H\"older's inequality implies $$\|g\|_{H^{1/2}(\Sigma)} \le \|g\|_{L^2(\Sigma)}^{2\alpha/(2\alpha+1)} \|g\|_{H^{\alpha+1/2}(\Sigma)}^{1/(2\alpha+1)}$$ for $H^\alpha(\Sigma)$ defined as in \cite{N86}, we deduce from (\ref{radfwdbd}) and Markov's inequality that as $\varepsilon \to 0$, \begin{equation*} \|\hat F_{\lambda, \varepsilon} - F_0\|_{L^2(\mathcal O)} \lesssim \|\mathscr R(\hat F)-\mathscr R(F_0)\|_{H^{1/2}(\Sigma)} =O_{\mathbb P_{f_0}^{\varepsilon}}\big(\varepsilon^{\frac{2\alpha}{2\alpha+3}} \big)\end{equation*} in probability [recall that random variables $(Z_n: n \in \mathbb N)$ are $O_{\Pr}(r_n)$ if $\forall \delta>0 ~\exists M=M_\delta$ s.t.~$\Pr(|Z_n|>Mr_n)<\delta$ for all $n \in\mathbb N$], with constants uniform in $\|F_0\|_{H_c^\alpha(\mathcal O)} \le R$ for any $R>0$.
Similarly, if one chooses $\lambda=\varepsilon$ instead, then the MAP estimate from Remark \ref{lin} satisfies
\begin{equation*}
\left\|\hat F_{\lambda, \varepsilon}-F_0\right\|_{L^2(\mathcal O)}=O_{\mathbb P_{f_0}^{\varepsilon}}\big(\varepsilon^{\frac{2\beta}{2\beta+3}} \big),~~\text{where }\beta:=\alpha-1>0,
\end{equation*}
uniformly over $\|F_0\|_{H^\beta_c(\mathcal O_0)}\leq R$ for $R> 0$.
\end{ex}
\begin{rem}[The effect of nonlinearity] \label{nonlin} \normalfont In the proof of Theorem \ref{thm-gen} we follow ideas for $M$-estimation from \cite{saramest, sara2001}, and condition (\ref{entrcond}) is needed to bound the entropy numbers of images $\{\mathscr G(F)\;| \;\|F\|_{H^\alpha}\leq R \}, 0<R<\infty,$ of Sobolev balls under $\mathscr G$, which in turn control the modulus of continuity of the Gaussian process that determines the convergence rate of $\hat F$ to $F_0$. The at most polynomial growth in $\|F\|_{H^\alpha}$ of the Lipschitz constants
\begin{equation}\label{gammafactor}
\left(1+\|F\|_{H^\alpha}^\gamma\vee\|H\|_{H^\alpha}^\gamma\right), ~~ \gamma \ge 0,
\end{equation}
in (\ref{entrcond}) turns out to be essential in the proof of Theorem \ref{thm-gen}. But even when only a `polynomial nonlinearity' is present ($\gamma>0$), the last term in the condition (\ref{delta-cond}) can become dominant if the penalisation parameter $\lambda$ is too small. The intuition is that, for non-linear problems, too little penalisation can mean that the maximisers $\hat F$ over unbounded parameter spaces behave erratically, yielding sub-optimal convergence rates.
\end{rem}
\section{Results for elliptic PDE models}\label{sec-pderes}
In this section, we apply Theorem \ref{thm-gen} to the inverse problems induced by the PDEs (\ref{intro-div}) and (\ref{intro-sch}). We also discuss the implied convergence rates for the parameter $f$.
\subsection{Basic setup and link functions}\label{sec-pde-pre}
For any integer $\alpha > d/2$ and any constant $K_{min}\in[0,1)$, and denoting the outward pointing normal vector at $x\in\partial \mathcal O$ by $n=n(x)$, define the parameter space (boundary derivatives are understood in the trace sense)
\begin{equation}\label{Fdef}
\begin{split}
\mathcal F:=\mathcal F_{\alpha, K_{min}} =\big\{&f\in H^\alpha(\mathcal O): f>K_{min}\text{ on } \mathcal O,~f=1 \text{ on }\partial \mathcal O,\\
&\;\;\frac{\partial^jf}{\partial n^j}=0\text{ on }\partial \mathcal O \text{ for }j=1,...,\alpha-1 \big\},
\end{split}
\end{equation}
and its subclasses
\begin{equation*}
\mathcal F_{\alpha,r}(R):=\big\{f\in\mathcal F:f>r\text{ on } \mathcal O,\;\|f\|_{H^\alpha}\leq R\big\},~ r \ge K_{min}, R > 0.
\end{equation*}
We note that the restrictions $K_{min}<1$ and $f=1$ on $\partial \mathcal O$ in (\ref{Fdef}) are made only for convenience, and could be replaced by any $K_{min}>0$ and $f=\tilde g$ for \emph{fixed} $\tilde g\in C^\infty(\partial \mathcal O)$ satisfying $\tilde g>K_{min}$. Moreover, for estimation over parameter spaces without prescribed boundary values for $f$, see Remark \ref{bdhack}.
\par
We will assume that the coefficient $f$ of the second order linear elliptic partial differential operators featuring in the boundary value problems (\ref{intro-div}) and (\ref{intro-sch}), respectively, belong to $\mathcal F_{\alpha, K_{min}}$ for large enough $\alpha$, and denote by
\begin{equation} \label{Gdef}
G:\mathcal F\to L^2(\mathcal O), \qquad f\mapsto G(f):=u_f,
\end{equation}
the corresponding solution maps. Following (\ref{data}) with $\mathbb H = L^2(\mathcal O)$, we then observe
\begin{equation}\label{data2}
Y^{(\varepsilon)}=G(f)+\varepsilon \mathbb W, ~\varepsilon>0,
\end{equation}
whose law will now be denoted by $\mathbb P_{f}^\varepsilon$ for $f\in\mathcal F$.
We will apply Theorem \ref{thm-gen} to a suitable bijective re-parameterisation of $\mathcal F$ for which the set $\mathcal V$ one optimises over is a linear space. This is natural for implementation purposes but also necessary to retain the Bayesian interpretation of our estimators from Remark \ref{lin}. To this end, we introduce `link functions' $\Phi$ -- the lowercase and uppercase notation for corresponding functions $f \in\mathcal F$ and $F=\Phi^{-1}\circ f$ will be used throughout.
\begin{defin}\label{def-phi}
1. A function $\Phi$ is called a \emph{link function} if $\Phi$ is a smooth, strictly increasing bijective map $\Phi: \mathbb R\to(K_{min},\infty)$ satisfying $\Phi(0)=1$ and $\Phi'>0$ on $\mathbb R$.
\par
2. A function $\Phi:(a,b)\to \mathbb R$, $-\infty\leq a<b\leq \infty$, is called \emph{regular} if all derivatives of $\Phi$ of order $k\geq 1$ are bounded, i.e.
\begin{equation}\label{phibddder}
\forall k\geq 1: \quad \sup_{x\in(a,b)} \left|\Phi^{(k)}(x)\right|<\infty.
\end{equation}
\end{defin}
In the notation of Theorem \ref{thm-gen}, throughout this section we set $\mathbb H=L^2(\mathcal O)$, $\tilde{\mathcal V}=\mathcal V:= \{\Phi^{-1}\circ f:f\in\mathcal F \}$ to be the `pulled-back' parameter space, and
\begin{equation}\label{G2}
\mathscr G: \mathcal V\to L^2(\mathcal O),\quad \mathscr G(F):=G(\Phi\circ F),
\end{equation}
For $\mathcal F$ as in (\ref{Fdef}), one easily verifies that
\begin{equation*}
\mathcal V=\left\{F\in H^\alpha: \frac{\partial^jF}{\partial n^j}=0\text{ on }\partial \mathcal O \text{ for }j=0,...,\alpha-1 \right\}=H^\alpha_c(\mathcal O),
\end{equation*}
where the second equality follows from the characterization of $H^\alpha_c(\mathcal O)$ in Theorem 11.5 of \cite{lionsmagenes}.
Given a realisation of (\ref{data2}) and a regular link function $\Phi$, we define the \emph{generalised Tikhonov regularised functional} $J_{\lambda,\varepsilon}:\mathcal F\to \mathbb R$,
\begin{equation}\label{Jdef}
J_{\lambda,\varepsilon}(f):=2\langle Y^{(\varepsilon)},G(f)\rangle_{L^2}-\|G(f)\|_{L^2}^2-\lambda^2\|\Phi^{-1}\circ f\|_{H^\alpha}^2, ~~\lambda>0.
\end{equation}
Then for all $f\in\mathcal F$, we have $\mathscr J_{\lambda, \varepsilon}(F)=J_{\lambda,\varepsilon}(f)$ in the notation (\ref{Jdef2}), and maximising $J_{\lambda, \varepsilon}$ over $\mathcal F$ is equivalent to maximising $\mathscr J_{\lambda,\varepsilon}$ over $H^\alpha_c=\mathcal V$. Any pair of maximisers will be denoted by
\begin{equation*}
\hat f\in \arg \max_{f\in \mathcal F}J_{\lambda, \varepsilon}(f), ~~ \hat F=\Phi^{-1}\circ \hat f\in \arg \max_{F\in H^\alpha_c}\mathscr J_{\lambda, \varepsilon}(F),~~G(\hat f)=\mathscr G(\hat F).
\end{equation*}
The proofs of the theorems which follow are based on an application of Theorem \ref{thm-gen}, after verifying that the map (\ref{G2}) satisfies (\ref{entrcond}) with $\mathcal V = H^\alpha_c$ and suitable values of $\kappa, \gamma, \alpha$. The verification of (\ref{entrcond}) is based on PDE estimates that control the modulus of continuity of the solution map (\ref{Gdef}), and on certain analytic properties of the link function $\Phi$. In practice often the choice $\Phi=\exp$ is made (cf.~\cite{stuart10}), but our results suggest that the use of a \emph{regular} link function might be preferable. Indeed, the polynomial growth requirement (\ref{gammafactor}) discussed above is not met if one chooses for $\Phi$ the exponential function. Before we proceed, let us give an example of a regular link function.
\begin{ex} \normalfont
Define the function $\phi: \mathbb R \to (0,\infty)$ by $\phi(x) = e^x1_{x<0} + (1+x) 1_{x \ge 0}$, let $\psi:\mathbb R\to [0,\infty)$ be a smooth, compactly supported function with $\int_\mathbb R \psi=1$, and write $\phi \ast \psi=\int_{\mathbb{R}}\phi(\cdot-y)\psi(y)dy$ for their convolution. It follows from elementary calculations that, for any $K_{min} \in \mathbb R$,
\[\Phi:\mathbb{R}\to (K_{min},\infty),~\Phi:= K_{min}+\frac{1-K_{min}}{\psi\ast\phi(0)}\psi \ast \phi, \]
is a regular link function with range $(K_{min},\infty)$.
\end{ex}
\subsection{Divergence form equation}\label{sec-div}
For a \emph{given} source function $g \in C^\infty(\mathcal O)$, we consider the Dirichlet boundary value problem
\begin{equation}\label{div}
\begin{cases}
\nabla\cdot (f\nabla u)=g\quad \textnormal{ on } \mathcal O,\\
u=0 \quad \textnormal{ on }\mathcal \partial O,
\end{cases}
\end{equation}
where $f \in \mathcal F_{\alpha, K_{min}}$ (see (\ref{Fdef})) for some $\alpha>d/2+1, K_{min}>0$. Then (\ref{h-emb}) implies $f\in C^{1+\eta}(\mathcal O)$ for some $\eta>0,$ and the Schauder theory for elliptic PDEs (Theorem 6.14 in \cite{gt}) then gives that (\ref{div}) has a unique classical solution in $C(\bar{\mathcal O})\cap C^{2+\eta}(\mathcal O)$ which we shall denote by $G(f)=u_{f}$.
\paragraph{Upper bounds} For a link function $\Phi$ and $f_1,f_2\in\mathcal F$, define (cf.~(\ref{taudef}))
\[\mu_{\lambda}(f_1,f_2):=\|G(f_1)-G(f_2)\|^2_{L^2}+\lambda^2\|\Phi^{-1}\circ f_1\|_{H^\alpha}^2=\tau_\lambda(F_1,F_2). \]
\begin{thm}[Prediction error]\label{thm-div-pred}
Let $\mathcal F$ be given by (\ref{Fdef}) for some integer $\alpha>(d/2+2) \vee (2d-1)$ and $K_{min}\in (0,1)$. Let $G(f)=u_f$ denote the unique solution of (\ref{div}) and let $Y^{(\varepsilon)} \sim \P_{f_0}^\varepsilon$ from (\ref{data2}) for some $f_0 \in \mathcal F$. Moreover, suppose that $\Phi:\mathbb R\to (K_{min}, \infty)$ is a regular link function and that $J_{\lambda_\varepsilon,\varepsilon}$ is given by (\ref{Jdef}), where
$$\lambda_\varepsilon:=\varepsilon^{\frac{2(\alpha+1)}{2(\alpha+1)+d}}.$$
Then the following holds.
\begin{enumerate}
\item For each $f_0\in\mathcal F$ and $\varepsilon>0$, almost surely under $\P_{f_0}^\varepsilon$, there exists a maximiser $\hat f_{\varepsilon} \in \mathcal F$ of $J_{\lambda_\varepsilon,\varepsilon}$ over $\mathcal F$.
\item For each $R>0$, $r > K_{min}$, there exist finite constants $c_1,c_2>0$ such that for any maximiser $\hat f_\varepsilon \in \mathcal F$ of $J_{\lambda_\varepsilon,\varepsilon}$, all $0<\varepsilon < 1$ and all $M\geq c_1$,
\begin{equation}\label{div-conc}
\sup_{f_0\in\mathcal F_{\alpha,r}(R)}\P_{f_0}^\varepsilon\Big(\mu^2_{\lambda_\varepsilon}(\hat f_\varepsilon,f_0)\geq M^2\varepsilon^{\frac{4(\alpha+1)}{2(\alpha+1)+d}}\Big)\leq \exp\Big(-\frac{M^2\lambda_\varepsilon^2}{c_2\varepsilon^2}\Big).
\end{equation}
\item For each $R>0$, $r > K_{min}$ and $\beta\in [0,\alpha+1]$, there exists a constant $c_3$ such that for any maximiser $\hat f_\varepsilon \in \mathcal F$ of $J_{\lambda_\varepsilon,\varepsilon}$ with corresponding $u_{\hat f_\varepsilon}$, for all $0<\varepsilon<1$,
\begin{equation} \label{finprat}
\sup_{f_0\in\mathcal F_{\alpha,r}(R)}\mathbb E^{\varepsilon}_{f_0}\left\|u_{\hat f_\varepsilon}-u_{f_0}\right\|_{H^\beta}\leq c_3\varepsilon^{\frac{2(\alpha+1-\beta)}{2(\alpha+1)+d}}.
\end{equation}
\end{enumerate}
\end{thm}
\paragraph{Lower bounds} We now give a minimax lower bound on the rate of estimation for $u_f$ which matches the bound in (\ref{finprat}). To facilitate the exposition we only consider the unit ball $\mathcal O=D:=\big\{x\in\mathbb R^d: \|x\|< 1 \big\}$, set $g=1$ identically on $\mathcal O$, and fix $H^\beta$-loss with $\beta=2$.
\begin{thm}\label{thm-div-lb}
For $K_{min}\in (0,1), \alpha>d/2+1$, $\mathcal O = D$ and $g=1$ on $\mathcal O$, consider solutions $u_f, f \in \mathcal F,$ to (\ref{div}). Then there exists $C<\infty$ such that for all $\varepsilon>0$ small enough,
\begin{equation}\label{div-lower-bd}
\inf_{\hat u_\varepsilon}\sup_{f_0\in \mathcal F_{\alpha, r}(R)}\mathbb E_{f_0}^\varepsilon\|\hat u_\varepsilon- u_{f_0}\|_{H^2}\geq C\varepsilon^{\frac{2(\alpha-1)}{2(\alpha+1)+d}},~r>K_{min}, R>0,
\end{equation}
where the infimum ranges over all measurable functions $\hat u_\varepsilon= \hat u(Y^{(\varepsilon)})$ of $Y^{(\varepsilon)}$ from (\ref{data2}) that take values in $H^2$.
\end{thm}
Observe that (\ref{div-lower-bd}) coincides with the lower bound for estimating $u_{f_0}$ as a regression function without PDE-constraint in $H^{\alpha+1}$ under $H^2$-loss. Note however that unconstrained `off the shelf' regression function estimators $\tilde u_\varepsilon$ for $u_f$ will not satisfy the non-linear PDE constraint $\tilde u =G(\tilde f)$ for some $\tilde f \in \mathcal F$, thus providing no recovery of the PDE coefficient $f_0$ itself.
\smallskip
\paragraph{Rates for $f$ via stability estimates}
For estimators $u_{\hat f_\varepsilon}$ that lie in the range of the forward map $G$, we can resort to `stability estimates' which allow to control the convergence rate of $\hat f_\varepsilon$ to $f_0$ by the rate of $G(\hat f_\varepsilon)=u_{\hat f_\varepsilon}$ towards $G(f_0)=u_{f_0}$, in appropriate norms. Injectivity and global stability estimates for this problem have been studied in several papers since Richter \cite{R81}, see the recent contribution \cite{devore} and the discussion therein. They require additional assumptions, a very common choice being that $g>0$ throughout $\bar {\mathcal O}$. The usefulness of these estimates depends in possibly subtle ways on the class of $f$'s one constrains the problem to. The original stability estimate given in \cite{R81} controls $\|f_1-f_2\|_\infty$ in terms of $\|u_{f_1}-u_{f_2}\|_{C^2}$ which does not combine well with the $H^\beta$- convergence rates obtained in Theorem \ref{thm-div-pred}. The results proved in \cite{devore} are designed for `low regularity' cases where $\alpha \in (0,1)$: they give at best
\begin{equation}\label{devore-stab}
\|f_1-f_2\|_{L^2}\leq C(f_1,f_2) \|u_{f_1}-u_{f_2}\|_{H^1}^{1/2}, \qquad f_1,f_2\in\mathcal F, ~d \ge 2,
\end{equation}
which via Theorem \ref{thm-div-pred} would imply a convergence rate of $\varepsilon^{\frac{\alpha}{2(\alpha+1)+d}}$ for $\|\hat f_\varepsilon-f_0\|_{L^2}$. For higher regularity $\alpha \ge 2$ relevant here, this can be improved. We prove in Lemma \ref{lem-div-stab} below a Lipschitz stability estimate for the map $u_f\mapsto f$ between the spaces $H^2$ and $L^2$, and combined with Theorem \ref{thm-div-pred} this gives the following rate bound for $\hat f_\varepsilon - f_0$.
\begin{thm}\label{thm-div-f}
Suppose that $\alpha$, $K_{min}$, $\mathcal F$, $G$, $\Phi$, $\lambda_\varepsilon$ are as in Theorem \ref{thm-div-pred} and that in addition, $\inf_{x\in\mathcal O} g(x)\ge g_{min}$ for some $g_{min}>0$. Let $\hat f_\varepsilon \in \mathcal F$ be any maximiser of $J_{\lambda_\varepsilon,\varepsilon}$. Then, for each $r >K_{min}$ and $R<\infty$, there exists a constant $C>0$ such that we have for all $0 < \varepsilon <1$,
\begin{equation}\label{div-f}
\sup_{f_0\in\mathcal F_{\alpha,r}(R)}\mathbb{E}^{\varepsilon}_{f_0}\|\hat f_\varepsilon-f_0\|_{L^2}\leq C\varepsilon^{\frac{2(\alpha-1)}{2(\alpha+1)+d}}.
\end{equation}
\end{thm}
The rate in Theorem \ref{thm-div-f} is strictly better than what can be obtained from (\ref{devore-stab}), or by estimating $\|u_{\hat f_\varepsilon}-u_{f_0}\|_{C^2}$ by $\|u_{\hat f_\varepsilon}-u_{f_0}\|_{H^{2+d/2+\eta}}, \eta>0,$ and using Richter's stability estimate. A more detailed study of the stability problem, and of the related question of optimal rates for estimating $f$, is beyond the scope of the present paper and will be pursued elsewhere.
\subsection{Schr\"odinger equation}\label{sec-schr}
We now turn to the Schr\"odinger equation
\begin{equation}\label{schroedeq}
\begin{cases}
\Delta u -2fu =0 \quad \textnormal{on } \mathcal O,\\
u=g\quad \textnormal{on } \mathcal \partial O,
\end{cases}
\end{equation}
with \emph{given} $g\in C^\infty(\partial \mathcal O)$. By standard results for elliptic PDEs (Theorem 6.14 in \cite{gt}), for $f \in \mathcal F_{\alpha, K_{min}}$ from (\ref{Fdef}) with $K_{min}\geq 0, \alpha>d/2$, a unique classical solution $u_f=G(f)$ to (\ref{intro-sch}) exists which lies in $C^{2+\eta}(\mathcal O)\cap C^0(\bar{\mathcal O})$ for some $\eta>0$.
The results for this PDE are similar to the previous section, although the convergence rates are quantitatively different due to the fact that the forward operator is now $2$-smoothing.
\begin{thm}[Prediction error]\label{thm-schr-pred}
Let $\mathcal F$ be given by (\ref{Fdef}) for some integer $\alpha>(d/2+2) \vee (2d-2)$
and $K_{min}\in[0,1)$. Let $G(f)=u_f$ denote the unique solution to (\ref{schroedeq}) and let $Y^{(\varepsilon)} \sim \P_{f_0}^\varepsilon$ from (\ref{data2}) for some $f_0 \in \mathcal F$. Moreover, suppose that $\Phi:\mathbb R\to (K_{min}, \infty)$ is a regular link function and that $J_{\lambda_\varepsilon,\varepsilon}$ is given by (\ref{Jdef}), where
$$\lambda_\varepsilon=\varepsilon^{\frac{2(\alpha+2)}{2(\alpha+2)+d}}.$$
Then the following holds.
\begin{enumerate}
\item For each $f_0\in\mathcal F$ and $\varepsilon>0$, almost surely under $\P_{f_0}^\varepsilon$, there exists a maximiser $\hat f_{\varepsilon} \in \mathcal F$ of $J_{\lambda_\varepsilon,\varepsilon}$ over $\mathcal F$.
\item For each $R>0$, $r > K_{min}$, there exist finite constants $c_1,c_2>0$ such that for any maximiser $\hat f_\varepsilon \in \mathcal F$ of $J_{\lambda_\varepsilon,\varepsilon}$, all $0<\varepsilon<1$ and all $M\geq c_1$, we have
\begin{equation}\label{schr-conc}
\sup_{f_0\in\mathcal F_{\alpha,r}(R)}\P_{f_0}^\varepsilon\Big(\mu^2_{\lambda_\varepsilon}(\hat f_\varepsilon,f_0)\geq M^2\varepsilon^{\frac{4(\alpha+2)}{2(\alpha+2)+d}}\Big)\leq \exp\Big(-\frac{M^2\lambda_\varepsilon^2}{c_2\varepsilon^2}\Big).
\end{equation}
\item For each $R>0$, $r > K_{min}$ and $\beta\in [0,\alpha+2]$, there exists a constant $c_3>0$ such that for any maximiser $\hat f_\varepsilon \in \mathcal F$ of $J_{\lambda_\varepsilon,\varepsilon}$ and all $0<\varepsilon<1$,
\begin{equation}\label{schr-rate}
\sup_{f_0\in\mathcal F_{\alpha,r}(R)}\mathbb E^{\varepsilon}_{f_0}\left\|u_{\hat f_\varepsilon}-u_{f_0}\right\|_{H^\beta}\leq C\varepsilon^{\frac{2(\alpha+2-\beta)}{2(\alpha+2)+d}}.
\end{equation}
\end{enumerate}
\end{thm}
For the PDE (\ref{schroedeq}) the stability estimate is easier to obtain than the one required in Theorem \ref{thm-div-f}, and here is the convergence rate for estimation of $f\in\mathcal F$. We note that the rates obtained in Theorems \ref{thm-schr-pred} and \ref{thm-schr-f} are minimax-optimal in view of Proposition 2 in \cite{n17} (and its proof).
\begin{thm}\label{thm-schr-f}
Assume that $\alpha,K_{min}, \mathcal F,G,\Phi,\lambda_\varepsilon$ are as in Theorem \ref{thm-schr-pred} and that in addition, $\inf_{x\in\mathcal \partial O}g(x)\geq g_{min}$ for some $g_{\min}>0$. Let $\hat f_\varepsilon \in \mathcal F$ be any maximiser of $J_{\lambda_\varepsilon,\varepsilon}$. Then for all $r > K_{min}$ and $R>0$, there exists a constant $C>0$ such that for all $\varepsilon>0$ small enough,
\[\sup_{f_0\in\mathcal F_{\alpha,r}(R)}\mathbb{E}_{f_0}^\varepsilon\left\|\hat f_\varepsilon-f_0 \right\|_{L^2}\leq C\varepsilon^{\frac{2\alpha}{2(\alpha+2)+d}}. \]
\end{thm}
\subsection{Concluding remarks and discussion}\label{sec-pde-rem}
\begin{rem}\label{dethack} \normalfont
The classical literature on `deterministic inverse problems' deals with convergence rate questions of Tikhonov and related regularisers, see the monograph \cite{EHN96}, \cite{EKN89, N92, SEK93, EHN96, TJ02, KNS08} and also \cite{BB18}. The convergence analysis conducted there is typically for observations $y_\delta = \mathscr G(F)+\delta$ where $\delta$ is a fixed perturbation vector in data space, equal to $L^2(\mathcal O)$ in the setting of the present paper. For non-linear problems, rates are obtained as $\|\delta\|\to 0$ under some invertibility assumptions on a suitable adjoint $D\mathscr G^*_{F}$ of the linearisation $D\mathscr G_{F}[\cdot]$ of the forward operator at the `true' parameter $F$ (`source conditions'), see, e.g., Section 10 in \cite{EHN96}. These results are not directly comparable since our noise $\mathbb W$ models genuine statistical error and hence is random and, in particular, almost surely \textit{not} an element of data space $L^2(\mathcal O)$. As shown in \cite{BHM04, BHMR07, HP08, LL10}, the `deterministic' analysis extends to the Gaussian regression model (\ref{data}) to a certain degree, but the results obtained there still rely, among other things, on invertibility properties of $D \mathscr G^*_{F}$. For the PDE (\ref{intro-div}) such `source conditions' are problematic as $D \mathscr G^*_{F}$ is not invertible in general (due to the fact $\nabla u_{f}$ can vanish on $\mathcal O$ unless some fairly specific further assumptions are made, see \cite{itokunisch}). Our techniques circumvent source conditions by first optimally solving the `forward problem', and then feeding this solution into a suitable stability estimate for $\mathscr G^{-1}$.
\end{rem}
\begin{rem} [Bayesian inversion] \label{bayeshack} \normalfont
The Bayesian approach \cite{stuart10, KvdVvZ11, DS16} to inverse problems has been very popular recently, but only few theoretical guarantees for such algorithms are available in non-linear settings: In \cite{v13}, convergence rates for the PDE (\ref{div}) are obtained for certain Bayes procedures that arise from priors for $f$ that concentrate on specific bounded subsets of $H^\alpha$. The main idea to combine regression results with stability estimates is related to our approach, but the rates obtained in \cite{v13} are suboptimal, and for the elliptic PDE models considered here do not apply to Gaussian priors. Bayesian inference for the PDE (\ref{schroedeq}) has been studied in \cite{n17}, where it is shown that procedures based on a uniform wavelet prior \textit{do} attain minimax optimal convergence rates for $f$ and $u_f$ (up to log-factors). The paper \cite{n17} also addresses the question of uncertainty quantification via the posterior distribution, by proving nonparametric Bernstein-von Mises theorems, whereas our results only concern the convergence rate of the MAP estimate for certain Gaussian priors (see Remarks \ref{lin}, \ref{maphack}). A related recent reference is \cite{MNP19} where the asymptotics for linear functionals of Gaussian MAP estimates are obtained in linear inverse problems involving Radon and more general $X$-ray transforms -- see also \cite{KvdVvZ11, R13} for earlier results for diagonalisable linear inverse problems. Finally, convergence rates for posterior distributions of PDE coefficients in certain non-linear parabolic (diffusion) settings have been studied in \cite{NS17, A18}.
\end{rem}
\begin{rem} [MAP estimates, non-linear $\mathscr G$] \label{maphack} \normalfont As explained in Remark \ref{nonlin}, Theorem \ref{thm-gen} does not necessarily produce optimal rates for the choice $\lambda =\varepsilon$ in the non-linear settings from this section where $\gamma>0$. For MAP estimates as discussed in Remark \ref{lin} our results then imply optimal convergence rates for $G(f)$ only if the Gaussian prior is re-scaled in an $\varepsilon$-dependent way, more specifically if its RKHS norm is $\bar \lambda \|\cdot\|_{H^\alpha}$ with $\bar \lambda= \varepsilon^{-d/(2\alpha+2\kappa+d)}$. Moreover, positivity of $f$ is enforced by a `regular link function' $\Phi$, excluding the exponential map. Whether these restrictions on admissible priors are artefacts of our proofs remains a challenging open question, however, to the best of our knowledge, these are the first convergence rate results for proper MAP-estimators in the non-linear PDE constrained inverse problems studied here, improving in particular upon the (`sub-sequential') consistency results \cite{dashti13}.
\end{rem}
\begin{rem}\label{bdhack} \normalfont
For both PDEs (\ref{div}) and (\ref{schroedeq}), one can also consider estimation over the parameter space
\begin{equation}\label{ftilde}
\tilde{\mathcal F}:= \{f\in H^\alpha(\mathcal O): \inf_{x \in \mathcal O}f(x) >K_{min}\text{ on }\mathcal O\},
\end{equation}
\textit{without} the boundary restrictions on $f$ from (\ref{Fdef}). Note that $\mathcal F \subset \tilde {\mathcal F}$. Then, with $\kappa=1/2-\eta$,$\eta\in (0,1/2)$, $\tilde{\mathcal V}=\mathcal V=H^\alpha(\mathcal O)$ in (\ref{G2}), $\alpha>(d/2+2) \vee 2d-\kappa$ and $K_{min}$ as before, Theorem \ref{thm-gen} applies as in Theorems \ref{thm-div-pred} and \ref{thm-schr-pred}, and
\[\sup_{f_0 \in \tilde{\mathcal F}_{\alpha,r}(R)} \mathbb{E}_{f_0}^\varepsilon\|u_{\hat f}-u_{f_0}\|_{H^\beta(\mathcal O)}\lesssim \varepsilon^ {\frac{2(\alpha+1/2-\eta-\beta)}{2(\alpha+1/2-\eta)+d}},\quad r>K_{min}, R>0, \]
where, respectively, $\beta\in [0,\alpha+1]$ (divergence form eq.) and $\beta\in [0,\alpha+2]$ (Schr\"odinger eq.), and $\tilde{\mathcal F}_{\alpha,r}(R):=\{f\in\tilde{\mathcal F}: f >r \text{ on }\mathcal O,~ \|f\|_{H^\alpha} \le R\}$. By the stability Lemmas \ref{lem-div-stab} and \ref{lem-schr-stab} (which apply to $\tilde {\mathcal F}$ as well) and arguing as in the proofs of Theorems \ref{thm-div-f} and \ref{thm-schr-f}, this yields the respective convergence rates
\[\varepsilon^ {\frac{2(\alpha-3/2-\eta)}{2(\alpha+1/2-\eta)+d}\cdot\frac{\alpha-1}{\alpha+1}} \;\text{ (div. form eq.)}, \quad \varepsilon^ {\frac{2(\alpha-3/2-\eta)}{2(\alpha+1/2-\eta)+d}\cdot\frac{\alpha}{\alpha+2}} \;\text{ (Schr\"odinger eq.)},\]
for $\mathbb{E}_{f_0}^\varepsilon\|\hat f-f_0\|_{L^2}$, uniform over $\tilde{\mathcal F}_{\alpha,r}(R)$.
\end{rem}
\section{Proofs of the main results}\label{sec-pfs}
\subsection{Convergence rates in $M$-estimation}
For the convenience of the reader we recall here some classical techniques for proving convergence rates for $M$-estimators (see \cite{sara2001, saramest}) -- these will form the basis for the proof of Theorem \ref{thm-gen}. In the following $\tilde{\mathcal V}\subseteq L^2(\mathcal O)$, $\mathscr G:\tilde{\mathcal V}\to \mathbb H$ is a Borel-measurable map, and the functionals $\mathscr J_{\lambda, \varepsilon}, \tau_\lambda^2(\cdot,\cdot)$ are given by (\ref{Jdef2}), (\ref{taudef}), respectively. Let $\mathcal V\subseteq \tilde{\mathcal V}\cap H^\alpha(\mathcal O)$ be a subset over which we aim to maximise $\mathscr J_{\lambda, \varepsilon}$. For any $F_*\in \mathcal V$ and $\lambda, R\geq 0$, define sets
\begin{align}\label{V*}
\mathcal V_*(\lambda,R)&:=\{F\in\mathcal V: \tau^2_{\lambda}(F,F_*)\leq R^2 \},
\end{align}
their images under $\mathscr G$,
\begin{equation}\label{D*}
\mathcal D_*(\lambda,R)= \big\{\mathscr G(F):F\in\mathcal V \text{ with } \tau^2_{\lambda}(F,F_*)\le R^2\big\},
\end{equation}
and also
\begin{equation}\label{J*}
J_*(\lambda,R):=R+\int_0^{2R} H^{1/2}\left(\rho,\mathcal D_*(\lambda,R),\|\cdot\|_{\mathbb H}\right)d\rho,
\end{equation}
where the usual metric entropy of $A \subset \mathbb H$ is denoted by $H(\rho,A,\|\cdot\|_{\mathbb H} )$ ($\rho>0$). The following theorem is, up to some modifications which adapt it to the continuum sampling scheme (\ref{data}) and the inverse problem setting considered here, a version of Theorem 2.1 in \cite{sara2001}.
\begin{thm}\label{thm-ep}
Let $F_*\in\mathcal V, \lambda >0,$ and let $\mathbb P^\varepsilon_{F_0}$ be the law of $Y^{(\varepsilon)}$ from (\ref{data}) for some fixed $F_0 \in \tilde{\mathcal V}$. Suppose $\Psi_*(\lambda, R)\geq J_*(\lambda,R)$ is some upper bound such that $R\mapsto \Psi_*(\lambda, R)/R^2$ is non-increasing. Then there exist universal constants $c_1, c_2, c_3$ such that for all $\varepsilon, \lambda, \delta>0$ satisfying
\begin{equation}\label{delta-req}
\delta^2\geq c_1\varepsilon\Psi_*(\lambda, \delta)
\end{equation}
and any $R\geq \delta$, we have that
\begin{align}\label{tau-conc}
\mathbb P^\varepsilon_{F_0}\Big(\mathscr J_{\lambda,\varepsilon} &\text{ has a maximizer } \hat F \text{ over }\mathcal V \text{ s.t. }\tau_\lambda^2(\hat F,F_0)\geq 2(\tau_\lambda^2(F_*,F_0)+R^2)\Big) \notag \\
&\leq c_2\exp\Big(-\frac{R^2}{c_1^2\varepsilon^2}\Big).
\end{align}
Moreover, for any maximiser $\hat F$ of $\mathscr J_{\lambda,\varepsilon}$ over $\mathcal V$ we have for some universal constant $c_3$
\begin{equation}\label{tau-exp}
\mathbb{E}_{F_0}^\varepsilon[\tau_\lambda^2(\hat F,F_0)]\leq c_3(\tau_\lambda^2(F_*,F_0)+\delta^2+\varepsilon^2).
\end{equation}
\end{thm}
\begin{proof}
1. Let $\hat F$ denote any maximiser of $\mathscr J_{\lambda, \varepsilon}$. By completing the square, we see that $\hat F$ also maximises
\[Q_{\lambda,\varepsilon}(F):=2\langle\varepsilon \mathbb W,\mathscr G( F)\rangle_{\mathbb H}-\|\mathscr G(F)-\mathscr G(F_0)\|_{\mathbb H}^2-\lambda^2\|F\|_{H^\alpha}^2.\]
Rewriting the inequality $Q_{\lambda,\varepsilon}(\hat F)\geq Q_{\lambda,\varepsilon}(F_*)$, we obtain
\[2\langle\varepsilon \mathbb W,\mathscr G( \hat F)-\mathscr G(F_*)\rangle_{\mathbb H}\geq \tau_\lambda^2(\hat F,F_0)- \tau_\lambda^2(F_*,F_0).\]
Elementary calculations as in \cite{sara2001}, p.3-4, give that for all $R>0$, if
\[\tau_\lambda^2(\hat F,F_0)\geq 2\left(\tau_\lambda^2(F_*,F_0)+R^2\right) \] holds
then we also have the inequalities
\begin{equation*}
\begin{split}
\tau_\lambda^2(\hat F,F_*)&\geq R^2 \qquad \text{and}\\
\tau_\lambda^2(\hat F,F_0)-\tau_\lambda^2(F_*,F_0)&\geq \frac 16 \tau_\lambda^2(\hat F,F_*).
\end{split}
\end{equation*}
It follows that for any $R>0$ and for $\mathbb P$ the law of the centred Gaussian process $(\mathbb W(\psi)=\langle \mathbb W, \psi \rangle_{\mathbb H}: \psi \in \mathbb H)$,
\begin{equation*}
\begin{split}
\P_{F_0}^\varepsilon&\left(\tau_\lambda^2(\hat F,F_*)\geq 2\left(\tau_\lambda^2(F_*,F_0)+R^2\right)\right)\\
&\le \P_{F_0}^\varepsilon \Big(\tau_\lambda^2(\hat F,F_*)\geq R^2,~2\langle\varepsilon \mathbb W,\mathscr G( \hat F)-\mathscr G(F_*)\rangle_{\mathbb H}\geq \frac 16 \tau_\lambda^2(\hat F,F_*) \Big)\\
&\leq \sum_{l=1}^{\infty}\P\left(\sup_{\psi\in \mathcal D_*(\lambda,2^lR)}\langle\varepsilon \mathbb W,\psi-\mathscr G(F_*)\rangle_{\mathbb H} \geq \frac{1}{48}2^{2l}R^2\right) =:\sum_{l=1}^\infty P_l.
\end{split}
\end{equation*}
\par
2. For all $\lambda, R\geq 0$, we have that $\sup_{\psi,\varphi\in\mathcal D_*(\lambda,R)}\|\psi-\varphi\|_{\mathbb H}\leq 2R,$
so that by Dudley's theorem (see \cite{nicklgine}, p.43),
\begin{equation*}
\begin{split}
&\mathbb{E}\left[\sup_{\psi\in\mathcal D_*(\lambda,R)}|\langle \mathbb{W},\psi-\mathscr G(F_*)\rangle_
\mathbb H|\right]\\
&\qquad \lesssim \inf_{\psi\in \mathcal D_*(\lambda,R)}\mathbb{E}|\langle \mathbb{W},\psi-\mathscr G(F_*)\rangle_\mathbb H|+\int_0^{2R}H^{1/2}\left(\rho,\mathcal D_*(\lambda,R),\|\cdot\|_{\mathbb H}\right)d\rho\\
&\qquad \lesssim R+\int_0^{2R}H^{1/2}\left(\rho,\mathcal D_*(\lambda,R),\|\cdot\|_{\mathbb H}\right)d\rho= J_*(\lambda, R)\leq \Psi_*(\lambda,R).
\end{split}
\end{equation*}
\par
3. Let us write $S_*(\lambda,R):=\sup_{\psi\in\mathcal D_*(\lambda,R)}|\langle \mathbb{W},\psi-\mathscr G(F_*)\rangle|$. By choosing $c$ large enough and $\delta$ such that (\ref{delta-req}) holds, we have that for all $R\geq \delta$, $\frac{1}{48}R^2-\varepsilon\Psi_*(\lambda,R)\geq \frac{1}{96}R^2$. Thus by the preceding display, the Borell-Sudakov-Tsirelson inequality (see Theorem 2.5.8 in \cite{nicklgine}), and possibly making $c>0$ larger, we obtain for all $R\geq \delta$ and $l=1,2,...$
\begin{equation}\label{pl-est}
\begin{split}
P_l&\leq \P\left( \varepsilon S_*(\lambda,2^lR)-\varepsilon \mathbb{E}[S_*(\lambda,2^lR)] \geq \frac{1}{48}2^{2l}R^2-\varepsilon\Psi_*(\lambda, 2^lR)\right)\\
&\leq \P\left( S_*(\lambda,2^lR)-\mathbb{E}[S_*(\lambda,2^lR)] \geq \frac{2^{2l}R^2}{96\varepsilon}\right)\\
&\leq \exp\left(-\frac 12\left(\frac{2^{2l}R^2}{96\varepsilon}\right)^22^{-2l}R^{-2}\right) \leq \exp\left(-\frac{2^{2l}R^2}{c\varepsilon^2}\right),
\end{split}
\end{equation}
where in the penultimate inequality, we have used
$$\sup_{\psi\in\mathcal D_*(\lambda, 2^lR)}\mathbb{E}[|\langle\mathbb{W},\psi-\mathscr G(F_*)\rangle_\mathbb H|^2]\leq 2^{2l}R^2.$$
The inequality (\ref{tau-conc}) now follows from summing (\ref{pl-est}), and (\ref{tau-exp}) follows from arguing as in the proof of Lemma 2.2 in \cite{sara2001}.
\end{proof}
\subsection{Proof of \ref{thm-gen}, Part 2}\label{sec-pfs-sec2}
We will apply Theorem \ref{thm-ep} and need the following lemma. For $F_* \in \mathcal V$, define $\mathcal V_*(\lambda, R),\mathcal D_*(\lambda, R)$ and $J_*(\lambda,R)$ by (\ref{V*}), (\ref{D*}) and (\ref{J*}) respectively. We also use the notation $H^\alpha(\mathcal O,r):=\{F\in H^\alpha(\mathcal O)\;|\; \|F\|_{H^\alpha}\leq r\}$ and $H^\alpha_c(\mathcal O,r):=\{F\in H^\alpha_c(\mathcal O)\;|\; \|F\|_{H^\alpha}\leq r\}, r >0$ and recall $s=(\alpha+\kappa)/d$.
\begin{lem}\label{lem-main}
Suppose that $\mathcal V$ and $\mathscr G$ are as in Part 2 of Theorem \ref{thm-gen}. Then there exists a positive constant $c$ such that for all $\lambda, R>0$ and $F_*\in \mathcal V$,
\begin{equation*}
\begin{split}
\Psi_*(\lambda,R)&:=R+c\big(R \lambda^{-\frac{1}{2s}}\big(1+(R/\lambda)^{\gamma/2s}\big)\big)
\end{split}
\end{equation*}
is an upper bound for $J_*(\lambda,R)$.
\end{lem}
\begin{proof}
Let us first assume that $\kappa \geq 1/2$. We estimate the metric entropy in $J_*(\lambda,R)$. Let $\rho, \lambda, R>0$ and define
\begin{equation*}\label{mk}
m:=C\left(1+R^\gamma\lambda^{-\gamma}\right),
\end{equation*}
where $C$ is the constant from (\ref{entrcond}). By definition of $\tau_\lambda$, we have $\mathcal V_*(\lambda,R)\subseteq H^\alpha_c(\mathcal O,R/\lambda)$. Fix some larger, bounded $C^\infty$-domain $\tilde{\mathcal O}\supset \bar{\mathcal O}$ and some function $\zeta\in C_c^\infty(\mathbb{R}^d)$ such that $0\leq \zeta\leq 1$, $\zeta=1$ on $\mathcal O$ and $\text{supp}(\zeta)\subset \tilde{\mathcal O}$. By the main theorem of Section 4.2.2 in \cite{T78}, there exists a bounded, linear extension operator $\mathcal E:H^{\kappa}(\mathcal O)\to H^\kappa(\mathbb{R}^d)$. Define the map $e:\phi\mapsto \zeta \mathcal E(\phi)$ which maps $H^\kappa(\mathcal O)$ continuously into $\tilde H^\kappa(\tilde {\mathcal O})$, and for $\phi\in L^2(\mathcal O)$, let $\tilde \phi:\mathbb{R}^d\to \mathbb{R}$ denote its extension by $0$ on $\mathbb{R}^d\setminus \mathcal O$. We then have, for some $c_1>0$,
\begin{equation}\label{neg-sob}
\|\phi\|_{(H^\kappa(\mathcal O))^*}=\sup_{\varphi\in H^\kappa(\mathcal O,1)}\left|\int_{\mathcal O}\phi\varphi \right|= \sup_{\varphi\in H^\kappa(\mathcal O,1)}\left|\int_{\tilde{\mathcal O}}\tilde{\phi}e(\varphi) \right|\le c_1\|\tilde{\phi}\|_{H^{-\kappa}(\tilde{\mathcal O})}.
\end{equation}
By Theorem 11.4 in \cite{lionsmagenes} and its proof, the zero extension $\phi\mapsto \tilde \phi$ is continuous from $H^\alpha_c(\mathcal O)$ to $H^\alpha(\tilde{\mathcal O})$ with norm $1$, so that
\begin{equation*}
\mathcal W:=\left\{\tilde F: F\in \mathcal V_*(R,\lambda) \right\}\subseteq H^\alpha_c(\tilde{\mathcal O},R/\lambda).
\end{equation*}
By Theorem 4.10.3 of \cite{T78}, we can pick $\tilde F_1,...,\tilde F_N\in \mathcal W$ with
\[N\leq \exp\Big(c_2\big(\frac{Rmc_1}{\lambda\rho}\big)^{\frac 1s}\Big)\]
for some universal constant $c_2$, such that the balls
\[\tilde B_i:=\Big\{\psi\in \mathcal W\;: \; \|\psi-\tilde F_i\|_{H^{-\kappa}(\tilde {\mathcal O})} \leq \frac{\rho}{mc_1}\Big\}, \qquad i=1,...,N,\]
form a covering of $\mathcal W$. Then it follows from (\ref{entrcond}) and $(\ref{neg-sob})$ that for all $i=1,...,N$ and $F$ with $\tilde F\in \tilde B_i$,
\[\|\mathscr G(F)-\mathscr G(F_i)\|_{\H}\leq m\|F-F_i\|_{(H^{\kappa}(\mathcal O))^*}\leq mc_1\|\tilde F-\tilde F_i\|_{H^{-\kappa}(\tilde{\mathcal O})}, \]
whence the balls
\[B_i' :=\{\psi\in \mathcal D_*(\lambda,R): \|\psi-\mathscr G(F_i)\|_{\mathbb H}\leq \rho \},\quad i=1,...,N\]
form a covering of $\mathcal D_*(\lambda,R)$. Hence we obtain the bound
\begin{equation}\label{entrbound}
H\left(\rho, \mathcal D_*(\lambda, R),\|\cdot\|_{\mathbb H}\right)\lesssim \Big(\frac{Rm}{\lambda\rho}\Big)^{\frac 1s},
\end{equation}
and hence also
\begin{equation*}
\begin{split}
\int_0^{2R} H^{1/2}\left(\rho, \mathcal D_*(\lambda, R),\|\cdot\|_{\mathbb H}\right)d\rho &\lesssim \int_0^{2R}\left(\frac{Rm}{\lambda\rho}\right)^{\frac 1{2s}}d\rho \lesssim R\lambda^{-\frac{1}{2s}}(1+(R/\lambda)^{\frac{\gamma}{2s}}),
\end{split}
\end{equation*}
which proves that $\Psi_*\geq J_*$ for the case $\kappa\geq 1/2$.
\par
For $\kappa<1/2$, by Theorem 11.1 in \cite{lionsmagenes}, we have $\tilde H^\kappa(\mathcal O)=H^\kappa_c(\mathcal O)=H^\kappa(\mathcal O)$ and hence $\|\cdot\|_{(H^\kappa(\mathcal O))^*}=\|\cdot\|_{H^{-\kappa}(\mathcal O)}$, whence we can use Theorem 4.10.3 of \cite{T78} directly to cover $\mathcal V_*(R,\lambda)\subseteq H^\alpha(\mathcal O,R/\lambda)$ by $H^{-\kappa}(\mathcal O)$-balls, and using (\ref{entrcond}) as above yields the entropy bound (\ref{entrbound}).
\end{proof}
By assumption on $\alpha$ we have $1+\frac{\gamma}{2s}< 2$ and hence the map $R\mapsto\Psi_*(\lambda,R)/R^2$, for $\Psi_*(\lambda,R)$ as defined in Lemma \ref{lem-main}, is decreasing. The bounds (\ref{main-thm-tau}) and (\ref{main-thm-rate}) then follow from Theorem \ref{thm-ep}. The proof of existence of maximisers is given in Section \ref{sec-ex-pf}. Finally, we obtain Corollary \ref{cor-gen} by taking $F_*=F_0$ and $\delta:=c\varepsilon^{2(\alpha+\kappa)/(2\alpha+2\kappa+d)}$ for which (\ref{delta-cond}) is easily verified for $\lambda$ chosen as in the corollary, so that Theorem \ref{thm-gen} applies.
\subsection{Proof of Theorems \ref{thm-div-pred}, \ref{thm-div-lb} and \ref{thm-div-f}}\label{sec-div-pfs}
\begin{proof}[Proof of Theorem \ref{thm-div-pred}]
We verify that $\mathscr G$ given by (\ref{G2}) with $G$ the solution map of (\ref{div}), satisfies (\ref{entrcond}) for $\mathcal V = H^\alpha_c, \mathbb H = L^2(\mathcal O)$, $\gamma =4, \kappa=1$, in order to apply Theorem \ref{thm-gen}. Let $F,H\in H^\alpha$, and let us write $f:=\Phi\circ F$, $h:=\Phi\circ H$. With $L_f, V_f$ introduced in Section \ref{sec-div-facts} we have by (\ref{G2}) and (\ref{div})
\begin{equation} \label{obpertu}
\begin{split}
&L_f[\mathscr G(F)-\mathscr G(H)]=L_f[u_f-u_h]\\
&\quad =L_f[u_f]-L_h[u_h]+(L_h-L_f)[u_h]=\nabla\cdot ((h-f)\nabla u_h),
\end{split}
\end{equation}
and then, by Lemma \ref{lem-div-h2} with $H^2_0$ defined in (\ref{h0}), the estimate
\begin{equation}\label{diventr1}
\begin{split}
\|\mathscr G(F)-\mathscr G(H)\|_{L^2}&=\left\|V_f\left[\nabla\cdot ((h-f)\nabla u_h)\right]\right\|_{L^2}\\
&\leq C(1+\|f\|_{C^1})\left\|\nabla\cdot ((h-f)\nabla u_h)\right\|_{(H^2_0)^*}.
\end{split}
\end{equation}
By applying the divergence theorem to the vector field $\varphi (h-f)\nabla u_h$, where $\varphi\in C^2_0$ is any $C^2$-function that vanishes at the boundary, we have
\begin{equation*}
\begin{split}
\left\|\nabla\cdot ((h-f)\nabla u_h)\right\|_{(H^2_0)^*}&=\sup_{\varphi\in C_0^2,\; \|\varphi\|_{H^2\leq 1}}\left|\int_{\mathcal O}\varphi \nabla\cdot ((h-f)\nabla u_h) \right|\\
&=\sup_{\varphi\in C_0^2,\; \|\varphi\|_{H^2\leq 1}}\left|\int_{\mathcal O}(h-f)\nabla \varphi \cdot \nabla u_h \right|\\
&\leq \|h-f\|_{(H^1)^*}\sup_{\varphi\in C_0^2,\; \|\varphi\|_{H^2\leq 1}}\|\nabla \varphi \cdot \nabla u_h\|_{H^1}\\
&\lesssim \|h-f\|_{(H^1)^*}\|u_h\|_{\mathcal C^2},
\end{split}
\end{equation*}
where we used the multiplicative inequality (\ref{c-h-mult}) in the last step. Combining this with (\ref{diventr1}) and Lemma \ref{lem-div-c} yields that
\begin{equation*}
\begin{split}
\|\mathscr G(F)-\mathscr G(H)\|_{L^2}&\lesssim (1+\|f\|_{C^1})(1+\|h\|_{C^1}^2)\|h-f\|_{(H^1)^*}.
\end{split}
\end{equation*}
Hence, by (\ref{phi-cm}), (\ref{phi-neg-sob}) and the Sobolev embedding (\ref{h-emb}), we obtain
\begin{equation*}
\begin{split}
\|\mathscr G(F)-\mathscr G(H)\|_{L^2}\lesssim (1+\|F\|_{H^\alpha}^4\vee\|H\|_{H^\alpha}^4)\|F-H\|_{(H^1)^*},
\end{split}
\end{equation*}
so $\mathscr G$ indeed fulfills (\ref{entrcond}) for $\gamma=4$ and $\kappa=1$.
\par
The existence of maximisers $\hat f_\varepsilon$ now follows from the first part of Theorem \ref{thm-gen}, and we prove (\ref{div-conc}) by applying Theorem \ref{thm-gen} with $F_*=F_0$. First, we note that for all $\hat f_\varepsilon$ and $f_0$,
\begin{equation}\label{mu-tau}
\mu_{\lambda}(\hat f_\varepsilon,f_0)= \tau_{\lambda}(\hat F_\varepsilon, F_0).
\end{equation}
For the choice $\delta_\varepsilon=c\varepsilon^{\frac{2(\alpha+1)}{2(\alpha+1)+d}}$ and $c$ large enough, the triple $(\varepsilon,\lambda_\varepsilon,\delta_\varepsilon)$ satisfies (\ref{delta-cond}) and Theorem \ref{thm-gen} and (\ref{mu-tau}) yield that for some $c'>0$ and any $m\geq \delta_\varepsilon$,
\[\P_{f_0}^\varepsilon\left(\mu^2_{\lambda_{\varepsilon}}(\hat f_\varepsilon,f_0)\geq 2(\delta_\varepsilon^2+m^2) \right)\leq \exp\left(-\frac{m^2}{c'\varepsilon^2}\right), \]
which proves (\ref{div-conc}).
\par
To show (\ref{finprat}), let now $\beta\in[0,\alpha+1]$, $R>0$ and $r > K_{min}$. By Lemma \ref{lem-div-bd}, we have that
\[M:=\sup_{f\in\mathcal F:\|f\|_{H^\alpha}\leq R}\|u_f\|_{H^{\alpha+1}}<\infty. \]
Now for any $f_0\in\mathcal F_{\alpha,r}(R)$, we can use (\ref{h-int}) to estimate
\begin{equation}\label{est-1}
\begin{split}
\|u_{\hat f}-u_{f_0}\|_{H^\beta}&\lesssim\|u_{\hat f}-u_{f_0}\|_{L^2}^{\frac{\alpha+1-\beta}{\alpha+1}}\|u_{\hat f}-u_{f_0}\|_{H^{\alpha+1}}^{\frac{\beta}{\alpha+1}}\\
&\lesssim \|u_{\hat f}-u_{f_0}\|_{L^2}^{\frac{\alpha+1-\beta}{\alpha+1}}\left(M^{\frac{\beta}{\alpha+1}}+\|u_{\hat f}\|_{H^{\alpha+1}}^{\frac{\beta}{\alpha+1}}\right).
\end{split}
\end{equation}
Further, Lemma \ref{lem-div-bd} and (\ref{phi-sob}) yield that
\begin{equation}\label{est-2}
\|u_{\hat f}\|_{H^{\alpha+1}}^{\frac{\beta}{\alpha+1}}\lesssim 1+\|\hat f\|_{H^\alpha}^{\alpha\beta}\lesssim 1+\|\hat F\|_{H^\alpha}^{\alpha^2\beta}\lesssim 1+\left(\lambda_{\varepsilon}^{-1}\mu_{\lambda_\varepsilon}(\hat f,f_0)\right)^{\alpha^2\beta}.
\end{equation}
Now set $\delta_\varepsilon:=c_1\varepsilon^{\frac{2(\alpha+1)}{2(\alpha+1)+d}}$ for $c_1$ from the second part of the theorem. We define the events
\begin{equation}\label{Aj}
\begin{cases}
A_0:=\{\mu_{\lambda_\varepsilon}(\hat f_\varepsilon,f_0)< \delta_\varepsilon\}\\
A_j:=\{\mu_{\lambda_\varepsilon}(\hat f_\varepsilon,f_0)\in (2^{j-1}\delta_\varepsilon,2^j\delta_\varepsilon]\}, \quad j\geq 1 .
\end{cases}
\end{equation}
By (\ref{div-conc}) and (\ref{est-1})-(\ref{est-2}), and writing $\hat \mu_{\lambda_\varepsilon}:=\mu_{\lambda_\varepsilon}(\hat f_\varepsilon,f_0)$, we then obtain
\begin{equation}\label{est-3}
\begin{split}
&\mathbb{E}_{F_0}^\varepsilon\left[\|u_{\hat f}-u_{f_0}\|_{H^\beta} \right]\lesssim \sum_{j=0}^\infty\mathbb{E}_{F_0}^\varepsilon\left[ 1_{A_j}\|u_{\hat f}-u_{f_0}\|_{L^2}^{\frac{\alpha+1-\beta}{\alpha+1}}\left(1+\lambda_{\varepsilon}^{-\alpha^2\beta}\hat \mu_{\lambda_\varepsilon}^{\alpha^2\beta}\right) \right]\\
&\lesssim \delta_\varepsilon^{\frac{\alpha+1-\beta}{\alpha+1}}+\sum_{j=1}^\infty (2^{j}\delta_\varepsilon)^{\frac{\alpha+1-\beta}{\alpha+1}}\left(1+\lambda_{\varepsilon}^{-\alpha^2\beta}(2^{j}\delta_\varepsilon)^{\alpha^2\beta}\right)\P_{f_0}^\varepsilon\left(A_j\right)\\
&\lesssim \delta_\varepsilon^{\frac{\alpha+1-\beta}{\alpha+1}}\Big(1+\sum_{j=1}^\infty 2^{\frac{j(\alpha+1-\beta)}{\alpha+1}}\big(1+(c2^j)^{\alpha^2\beta}\big)\exp \big(-\frac{2^{2j}\delta_\varepsilon^2}{c_2^2\varepsilon^2}\big) \Big) \\
&\lesssim \delta_\varepsilon^{\frac{\alpha+1-\beta}{\alpha+1}}(1+o(\varepsilon)),
\end{split}
\end{equation}
where $c_2$ is the constant from (\ref{div-conc}). The theorem is proved.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-div-f}]
We apply Lemma \ref{lem-div-stab} with $f_2=\hat f $ and $f_1 = f_0 \in \mathcal F_{\alpha,r}(R)$, so that $\|u_{f_1}\|_{C^1} \vee \|f_1\|_{C^1}$ is bounded by some fixed $B=B(R)$ (cf.~(\ref{h-emb}) and Lemma \ref{lem-div-c}). Thus, writing $\hat F_\varepsilon:=\Phi^{-1}\circ \hat f_\varepsilon$ and using (\ref{phi-cm}),
\begin{equation*}
\begin{split}
\mathbb{E}_{f_0}^\varepsilon\|\hat f_\varepsilon-f_0\|_{L^2}&\lesssim \mathbb{E}_{f_0}^\varepsilon\left[\|u_{\hat f_\varepsilon}-u_{f_0}\|_{H^2} \|\hat f_\varepsilon\|_{C_1}\right]\\
&\lesssim \mathbb{E}_{f_0}^\varepsilon\left[\|u_{\hat f_\varepsilon}-u_{f_0}\|_{L^2}^{\frac{(\alpha-1)}{\alpha+1}}\|u_{\hat f_\varepsilon}-u_{f_0}\|_{H^{\alpha+1}}^{\frac{2}{\alpha+1}} (1+\|\hat F_\varepsilon\|_{C_1})\right].
\end{split}
\end{equation*}
We now choose $\delta_\varepsilon:=c_1\varepsilon^{\frac{2(\alpha+1)}{2(\alpha+1)+d}}$ where $c_1$ is the constant from the second part of Theorem \ref{thm-div-pred}. Bounding $\|u_{\hat f_\varepsilon}-u_{f_0}\|_{H^{\alpha+1}}$ as in (\ref{est-1})-(\ref{est-2}), splitting the expectation into $A_j$, $j\geq 0$ as defined in (\ref{Aj}) and using the concentration inequality (\ref{div-conc}), we obtain as in (\ref{est-3}) the desired inequality
\begin{equation*}
\begin{split}
\mathbb{E}_{f_0}^\varepsilon\|\hat f_\varepsilon-f_0\|_{L^2}\lesssim \delta_\varepsilon^{\frac{\alpha-1}{\alpha+1}}(1+o(\varepsilon)).
\end{split}
\end{equation*}
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-div-lb}] We only prove the more difficult case $d \ge 2$.
\par
1. Let $f_0=1$. By direct computation, one verifies that the unique classical solution to (\ref{div}) with $g=1, \mathcal O=D$ is
$$u_{f_0}(x)=\frac{1}{2d}\left(\|x\|^2-1\right),\qquad \nabla u_{f_0}(x)=\frac{x}{d}.$$
Thus we have that for some $1/2<a<b<1$,
\begin{equation*}
[a,b]^d\subset D,\qquad \frac 1 {2d}\leq \partial_{x_i}u_{f_0}(x)\leq \frac 1d\quad \textnormal{for all} \;\; i=1,...,d\;\;\textnormal{and} \;\; x\in [a,b]^d.
\end{equation*}
\par
2. Now let $\Psi:\mathbb R\to\mathbb R$ be a $1$-dimensional, compactly supported, at least $(\alpha+1)$-regular Daubechies wavelet (see \cite{nicklgine}, Theorem 4.2.10). Then, for all integers $j\geq 1$, for suitable constants $n_j$, $c>1$ and shift vectors $v^{j,r}=(v^{j,r}_1,...,v^{j,r}_d)$ to be chosen later, we define the tensor wavelets $\Psi_{j,r}$, $r=1,...,n_j$ by
\[\Psi_{j,r}(x)=2^{\frac{jd}{2}}c^{-\frac{d-1}{2}}\Psi(2^jx_1+v^{j,r}_1)\prod_{i=2}^d\Psi\left(\frac{2^j}cx_i+v^{j,r}_i\right).\]
Note that the $\Psi_{j,r}$ are `steeper' by a fixed constant $c$ in $x_1$-direction than in any other direction. Due to the compact support of $\Psi$, there exists a constant $c_0$ which depends only on $c$ and $\Psi$ such that for all $j\geq j_0$ large enough, we can set $n_j=c_02^{jd}$ and find suitable vectors $v^{j,r}$ such that all $\Psi_{j,r}$ are supported in the interior $[a,b]^d$ with disjoint support. For some sufficiently small constant $\kappa>0$, we define
\begin{equation}\label{fm}
f_m:=f_0+\kappa 2^{-j(\alpha+d/2)}\sum_{r=1}^{n_j}\beta_{r,m}\Psi_{j,r},\qquad m=1,...,M,
\end{equation}
where $\beta_{r,m},\;m=1,...,M$ will be chosen later as a suitably separated elements of the hypercube $\beta_r\in\{-1,1\}^{n_j}$.
\par
3. We choose $\kappa$ small enough (independently of $c>1$), as follows. By the wavelet characterisation of Sobolev norms, all $f_m$ of the form (\ref{fm}) lie in a fixed $H^\alpha$-ball of radius $C\kappa$, for some universal constant $C>0$, in particular $\|f_m-f_0\|_\infty$ can be made as small as desired for $\kappa$ small enough, so that all the $f_m>K_{min}$. Arguing as in (\ref{obpertu}), using $L_{f_0}=\Delta$ (the standard Laplacian), (\ref{c2-isom}), the multiplicative inequality (\ref{c-mult}), Lemma \ref{lem-div-c} and the Sobolev embedding $H^\alpha\subseteq C^{1+\eta}$ (for some small $\eta>0$), we have (uniformly for all $f_m$)
\begin{equation*}
\begin{split}
\|u_{f_m}-u_{f_0}\|_{\mathcal C^2}&=\|V_{f_0}\left[\nabla\cdot \left((f_m-f_0)\nabla u_{f_m}\right)\right]\|_{\mathcal C^2}\\
&\lesssim \|\nabla\cdot \left((f_m-f_0)\nabla u_{f_m}\right)\|_{\mathcal C^0}\\
&\lesssim \|(f_m-f_0)\nabla u_{f_m}\|_{\mathcal C^1}\\
&\lesssim \|f_m-f_0\|_{\mathcal C^1}\|u_{f_m}\|_{\mathcal C^{2}}\\
&\lesssim \|f_m-f_0\|_{H^\alpha}(1+\|f_m\|_{\mathcal C^1}^2).
\end{split}
\end{equation*}
Therefore, $\sup_m \|u_{f_m}\|_{\mathcal C^2}<\infty$ and we can pick $\kappa$ so small that for all $f_m$ of the form (\ref{fm}),
\begin{equation}\label{derivest}
\frac 1 {4d}\leq \partial_{x_i}u_{f_m}(x)\leq \frac 2d\quad \textnormal{for all} \;\; i=1,...,d\;\;\textnormal{and} \;\; x\in [a,b]^d.
\end{equation}
\par
4. Next, we want to apply Theorem 6.3.2 from \cite{nicklgine}, for which two steps are needed: an appropriate lower bound on the $H^2$-distance between the $u_{f_m}$'s and a suitable upper bound on the KL-divergence of the laws $\mathbb P^\varepsilon_{f_m}, \mathbb P^\varepsilon_{f_0}$.
\par
5. We begin with the lower bound. By the isomorphism (\ref{lapl-isom}), for all $u\in H^2_0$ and $f\in \mathcal F$, we have that
\[\|u\|_{H^2}\gtrsim \|\Delta u\|_{L^2}=\|f^{-1}(L_fu-\nabla u\cdot \nabla f)\|_{L^2}\ge \|f\|^{-1}_{\infty}\|L_fu-\nabla u\cdot \nabla f\|_{L^2}. \]
For all $m,m'=1,...,M$, using this inequality with $f=f_m$,
\begin{equation}\label{fmfm'}
u=u_{f_m}-u_{f_{m'}}= V_{f_m}[\nabla\cdot (f_{m'} -f_m)\nabla u_{f_{m'}})]
\end{equation}
in view of (\ref{obpertu}), and $\sup_m\|f_m\|_{C^1}<\infty$,
\begin{equation}\label{lowerbound}
\begin{split}
\|u_{f_m}-u_{f_{m'}}\|_{H^2} & \gtrsim \left\|\nabla\cdot \left(\left(f_m-f_{m'}\right)\nabla u_{f_{m'}}\right)\right\|_{L^2} - \|\nabla (u_{f_m}-u_{f_{m'}}) \cdot \nabla f_m\|_{L^2} \\
&\geq \|\nabla(f_m-f_{m'}) \cdot \nabla u_{f_{m'}}\|_{L^2}-\|(f_m-f_{m'})\Delta u_{f_{m'}}\|_{L^2} \\
& \quad\quad -\|u_{f_m}-u_{f_{m'}}\|_{H^1} \|f_m\|_{C^1} =: I-II - III.
\end{split}
\end{equation}
We will later show that the second and third terms are of smaller order than the first term. Using (\ref{derivest}), we see
\begin{equation}\label{I-est}
\begin{split}
I&=\left\|\sum_{i=1}^{d}\partial_{x_i}(f_m- f_{m'})\partial_{x_i}u_{f_{m'}}\right\|_{L^2}\\
&\geq \|\partial_{x_1}(f_m- f_{m'})\partial_{x_1}u_{f_{m'}}\|_{L^2}-\sum_{i=2}^{d}\|\partial_{x_i}(f_m- f_{m'})\partial_{x_i}u_{f_{m'}}\|_{L^2}\\
&\geq \frac 1{4d}\|\partial_{x_1}(f_m- f_{m'})\|_{L^2}-\frac 2d\sum_{i=2}^{d}\|\partial_{x_i}(f_m- f_{m'})\|_{L^2}.
\end{split}
\end{equation}
To estimate this further, we calculate that for any $i=2,...,d$,
\begin{equation*}
\begin{split}
\partial_{x_i}\Psi_{j,r}(x)&=2^{\frac{jd}{2}}c^{-\frac{d-1}{2}}\Psi(2^jx_1+v^{j,r}_1)\\
&\qquad\qquad \times \Big(\prod_{k=2,\;k\neq i}^d\Psi\Big(\frac{2^j}cx_k+v^{j,r}_k\Big)\Big)\frac{2^j}{c}\Psi'\left(\frac{2^j}cx_i+v^{j,r}_i\right).
\end{split}
\end{equation*}
Similarly calculating $\partial_{x_1}\Psi_{j,r}$ and summing over $r=1,...,n_j$, we obtain
\[\|\partial_{x_i}(f_m-f_{m'})\|_{L^2}=\frac{1}{c}\|\partial_{x_1}(f_m-f_{m'})\|_{L^2}, \quad i=2,...,d. \]
Thus, choosing $c$ large enough and combining this with (\ref{I-est}), we can ensure that
\begin{equation*}
\begin{split}
I &\gtrsim \frac 1{4d}\|\partial_{x_1}(f_m- f_{m'})\|_{L^2}-\frac{2(d-1)}{cd}\|\partial_{x_1}(f_m- f_{m'})\|_{L^2}\\
&\geq \frac 1{8d}\|\partial_{x_1}(f_m- f_{m'})\|_{L^2}.
\end{split}
\end{equation*}
Moreover, as the first partial derivatives of the $\Psi_{j,r}$ still have disjoint support, they are orthonormal in $L^2$ and by Parseval's identity we have
\begin{equation}\label{I-est-2}
\begin{split}
\|\partial_{x_1}(f_m- f_{m'})\|_{L^2}^2&=\kappa^22^{-2j(\alpha+d/2)}\sum_{j=1}^{n_j}|\beta_{r,m}-\beta_{r,m'}|^2\|\partial_{x_1}\Psi_{j,r}\|_{L^2}^2\\
&=\|\partial_{x_1}\Psi_{0,1}\|_{L^2}^2\kappa^22^{-2j(\alpha-1+d/2)}\sum_{j=1}^{n_j}|\beta_{r,m}-\beta_{r,m'}|^2.
\end{split}
\end{equation}
By the Varshamov-Gilbert-bound (Example 3.1.4 in \cite{nicklgine}), for constants $c_1,c_2>0$ independent of $j$, we can find a subset $\mathcal M_j\subset \{-1,1\}^{c_02^{jd}}$ of cardinality $M_j=2^{c_12^{jd}}$ such that
$$\sum_{j=1}^{n_j}|\beta_{r,m}-\beta_{r,m'}|^2\geq c_22^{jd}$$
whenever $m\neq m'$. For such a subset $\mathcal M_j$, by (\ref{I-est-2}) we have
\begin{equation}\label{lowerbound2}
I\gtrsim\|\partial_{x_1}(f_m- f_{m'})\|_{L^2}\gtrsim 2^{-j(\alpha-1)}.
\end{equation}
\par
6. We next show that $II$ and $III$ in (\ref{lowerbound}) are of smaller order as $j \to \infty$. With the above choice of $f_m$'s, we have from Parseval's identity and (\ref{c-h-mult})
\begin{equation*}
\begin{split}
II^2&\leq \|f_m-f_{m'}\|_{L^2}^2\|u_{f_m}\|^2_{\mathcal C^2}= \kappa^22^{-2j(\alpha+d/2)}\sum_{r=1}^{n_j}|\beta_{r,m}-\beta_{r,m'}|^2 \|u_{f_m}\|^2_{\mathcal C^2}\\
&\lesssim 2^{-2j\alpha}=o (2^{-2j(\alpha-1)}),
\end{split}
\end{equation*}
and for term $III$ we have, by (\ref{fmfm'}), (\ref{h-int}), Lemma \ref{lem-div-h2} and arguing as in the first display of Step 7 to follow, that
\begin{align*}
\|u_{f_m}- & u_{f_{m'}}\|_{H^1} \lesssim \|u_{f_m}-u_{f_m'}\|^{1/2}_{H^2} \|u_{f_m}-u_{f_{m'}}\|^{1/2}_{L^2} \\
&\lesssim \|\nabla\cdot ((f_m-f_{m'})\nabla u_{f_{m'}})]\|^{1/2}_{L^2} \|[\nabla\cdot ((f_m-f_{m'})\nabla u_{f_{m'}})]\|^{1/2}_{(H^2_0)^*} \\
& \lesssim \|f_m-f_{m'}\|_{H^1}^{1/2} \|f_m-f_{m'}\|^{1/2}_{H^{-1}} \lesssim 2^{-j\alpha}=o (2^{-j(\alpha-1)}),
\end{align*}
where the first factor in the last line is bounded by $2^{-j(\alpha/2-1/2)}$ by similar arguments as in (\ref{I-est-2}). Combining the last two displayed estimates with (\ref{lowerbound}) and (\ref{lowerbound2}) gives the overall lower bound
\[\|u_{f_m}-u_{f_{m'}}\|_{H^2} \gtrsim 2^{-j(\alpha-1)} \approx \varepsilon^{\frac{2(\alpha-1)}{2(\alpha+1)+d}}\] with choice $j=j_\varepsilon$ such that $2^{j}\simeq \varepsilon^{-2/(2\alpha+2+d)}$.
\par
7. Now we show the upper bound. Arguing as in (\ref{obpertu}), using Lemma \ref{lem-div-h2}, integrating by parts and using the wavelet characterisation of the $H^{-1}(\mathbb R^d)$-norm (e.g., Section 4.3 in \cite{nicklgine} with $B^s_{2,2}=H^s, s \in \mathbb R$) as well as the interior support of the $\Psi_{j,r}$, we estimate
\begin{equation*}
\begin{split}
\|u_{f_m}-u_{f_0}\|_{L^2}^2&\lesssim \left\|\nabla \cdot \left((f_m-f_{0})\nabla u_{f_0}\right)\right\|_{(H^2_0)^*}^2\\
&=\left(\sup_{\|\psi\|_{H^2_0}\leq 1}\left|\int_{\mathbb R^d} \nabla \psi\cdot \nabla u_{f_0} (f_m-f_{0})\right|\right)^2\\
&\lesssim \|f_m-{f_{0}}\|_{H^{-1}(\mathbb{R}^d)}^2 \|u_{f_0}\|_{\mathcal C^1} \\
&\simeq \kappa^22^{-2j(\alpha+d/2+1)}\sum_{r=1}^{n_j}1 \lesssim 2^{-2j(\alpha+1)}.
\end{split}
\end{equation*}
By definition of $M_j$, using the results in Section 7.4 in \cite{n17} and arguing as in (6.16) in \cite{nicklgine} we thus bound the information distances as
$$\textnormal{KL}(\mathbb P^{\varepsilon}_{u_{f_m}},\mathbb P^{\varepsilon}_{u_{f_0}})\lesssim \varepsilon^{-2} \|u_{f_m} - u_{f_0}\|_{L^2}^2 \lesssim \varepsilon^{-2}2^{-2j(\alpha+1)}=2^{jd}\lesssim \log M_j,$$
so that the overall result now follows from Theorem 6.3.2 in \cite{nicklgine}.
\end{proof}
\subsection{Proof of Theorems \ref{thm-schr-pred} and \ref{thm-schr-f}}\label{sec-schr-pfs}
The proof of Theorem \ref{thm-schr-pred} follows the same principle as the proof of Theorem \ref{thm-div-pred}. By arguing exactly as in the first two steps of the proof of Theorem \ref{thm-div-pred}, in order to be able to apply Theorem \ref{thm-gen}, we now verify that the map
\[\mathscr G:H^\alpha_c\to L^2, \qquad \mathscr G(F):=G(\Phi\circ F), \]
satisfies (\ref{entrcond}) with $\mathbb H = L^2, \gamma =4, \kappa=2$. Let $F,H\in H^\alpha$ and $f=\Phi\circ F$, $h=\Phi\circ H\in\mathcal F$. By (\ref{schroedeq}), $u_f-u_h$ satisfies
\[(u_f-u_h)|_{\partial\mathcal O}=0, \qquad L_f[u_f-u_h]=(L_h-L_f)[u_h]=(f-h)u_h \] where $L_f$ is defined in Section \ref{sec-schr-G} below. Using this, the norm estimate (\ref{schr-h-2l2}), Lemma \ref{lem-schr-c2}, the embedding $H^\alpha\subseteq C^2(\mathcal O)$ as well as (\ref{phi-neg-sob}), we can then estimate
\begin{equation*}
\begin{split}
\|\mathscr G(F)-&\mathscr G(H) \|_{L^2}=\left\|u_f-u_h \right\|_{L^2}\\
&\lesssim \left(1+\|f\|_{\infty}\right)\left\|(f-h)u_h \right\|_{(H^2_0)^*}\\
&\leq \left(1+\|f\|_{\infty}\right)\|u_h\|_{\mathcal C^2}\left\|f-h\right\|_{(H^2_0)^*}\\
&\lesssim \left(1+\|f\|_{\infty}\right)\left(1+\|h\|_{\infty}\right)\left\|f-h\right\|_{(H^{2})^*}\\
&\lesssim \left(1+\|F\|_{\infty}^2\vee\|H\|_{\infty}^2\right)\left\|F-H \right\|_{(H^{2})^*}\left(1+ \|F\|_{C^2}^2\vee \|H\|_{C^2}^2\right) \\
&\lesssim \left(1+\|F\|_{H^\alpha}^4\vee \|H\|_{H^\alpha}^4\right)\left\|F-H \right\|_{(H^{2})^*}.
\end{split}
\end{equation*}
Thus (\ref{entrcond}) is fulfilled for $\gamma =4$ and $\kappa=2$. The existence of maximizers now follows from the first part of Theorem \ref{thm-gen}. The proof of the concentration inequality (\ref{schr-conc}) is completely analogous to the proof of (\ref{div-conc}), and the convergence rate (\ref{schr-rate}) follows from the same argument as in the proof of Theorem \ref{thm-div-pred}, utilizing Lemma \ref{lem-schr-c2} in place of Lemma \ref{lem-div-bd}.
Finally, the proof Theorem \ref{thm-schr-f} is analogous to that of Theorem \ref{thm-div-f}, but using Lemma \ref{lem-schr-stab} instead of Lemma \ref{lem-div-stab}, and is left to the reader.
\section{Some PDE facts}\label{sec-pde-facts}
In this section, we collect some key PDE facts which are needed to prove the results in Section \ref{sec-pderes}.
\subsection{Preliminaries}
Besides the classical H\"older spaces $C^\alpha(\mathcal O)$, we will also need the H\"older-Zygmund spaces $\mathcal C^\alpha(\mathcal O)$, see Section 3.4.2 in \cite{triebel} for definitions. For $\alpha\geq 0$, $\alpha\notin\mathbb N$, we have that $C^\alpha=\mathcal C^\alpha$ with equivalent norms, and we have the continuous embeddings $\mathcal C^{\alpha'}\subseteq C^\alpha\subseteq \mathcal C^\alpha$ for all $\alpha'>\alpha\geq 0$.
\par
We will repeatedly use the multiplicative inequalities
\begin{align}
\|fg\|_{H^\alpha}&\lesssim \|f\|_{H^\alpha}\|g\|_{H^\alpha},\qquad \alpha>d/2, \label{h-mult}\\
\|fg\|_{H^\alpha}&\lesssim \|f\|_{\mathcal C^\alpha}\|g\|_{H^\alpha}, \qquad \alpha\geq 0\label{c-h-mult},\\
\|fg\|_{\mathcal C^\alpha}&\lesssim\|f\|_{\mathcal C^\alpha}\|g\|_{\mathcal C^\alpha}, \qquad \alpha\geq 0 \label{c-mult}
\end{align}
for all $f,g$ in the appropriate function spaces, which follow from Remark 1 on p.143 and Theorem 2.8.3 in \cite{triebel}. For any $\alpha>d/2$ and $0\leq \eta<\alpha-d/2$, we also need the continuous embedding $H^\alpha\subseteq C^\eta$, with the norm estimate
\begin{equation}\label{h-emb}
\forall f\in H^\alpha,\;\; \|f\|_{C^\eta}\lesssim\|f\|_{H^\alpha}.
\end{equation}
Let $\text{tr}[\cdot]$ denote the usual trace operator for functions defined on $\mathcal O$ (for the definition on Sobolev spaces, see, e.g., Chapter 5.5 in \cite{evans}). In this and the next section, we will repeatedly use the fact that the standard Laplacian $\Delta$ and $\text{tr}[\cdot]$ establish topological isomorphisms between appropriate Sobolev and H\"older-Zygmund spaces. That is, for each $\alpha\geq 0$, we have the topological isomorphisms
\begin{align}
&(\Delta,\textnormal{tr}): H^{\alpha+2}(\mathcal O)\to H^\alpha(\mathcal O)\times H^{\alpha+3/2}(\partial\mathcal O), \quad u\mapsto (\Delta u, \textnormal{tr}[u]),\label{h-isom} \\
&(\Delta,\textnormal{tr}): \mathcal C^{\alpha+2}(\mathcal O)\to\mathcal C^{\alpha}(\mathcal O)\times \mathcal C^{\alpha+2}(\partial \mathcal O), \quad u\mapsto (\Delta u, \textnormal{tr}[u]),\label{c-isom}
\end{align}
which follow from Theorem II.5.4 in \cite{lionsmagenes} and Theorem 4.3.4 in \cite{triebel} respectively. Moreover, for any $\alpha\geq 1$, we will use the notation
\begin{align}\label{h0}
H^\alpha_0(\mathcal O):=\left\{f\in H^\alpha(\mathcal O)\;\middle| \; \textnormal{tr}[f]=0 \right\},\quad \mathcal C^\alpha_0(\mathcal O):=\left\{f\in \mathcal C^\alpha(\mathcal O)\;\middle| \; \textnormal{tr}[f]=0 \right\}.
\end{align}
We also need the following interpolation inequalities. For all $\beta_1,\beta_2\geq 0$ and $\theta\in [0,1]$, there exists a constant $C<\infty$ such that
\begin{align}
&\forall u\in \mathcal C^{\beta_1}\cap \mathcal C^{\beta_2}:\quad \|u\|_{\mathcal C^{\theta \beta_1+(1-\theta)\beta_2}}\leq C \|u\|_{\mathcal C^{\beta_1}}^\theta\|u\|_{\mathcal C^{\beta_2}}^{1-\theta}, \label{c-int}\\
&\forall u\in H^{\beta_1}\cap H^{\beta_2}:\quad \|u\|_{H^{\theta \beta_1+(1-\theta)\beta_2}}\leq C \|u\|_{H^{\beta_1}}^\theta\|u\|_{H^{\beta_2}}^{1-\theta}, \label{h-int}
\end{align}
see Theorems 1.3.3 and 4.3.1 in \cite{T78} (and note $\mathcal C^\beta=B^\beta_{\infty, \infty}, H^\beta=B^\beta_{2, 2}$).
\par
\subsection{Divergence form equation}\label{sec-div-facts}
\subsubsection{Estimates for $V_f$}\label{vfop}
For each $f\in C^1(\bar{\mathcal O})$ with $f\geq K_{min}>0$, we define the differential operator
$$L_f:H^2_0(\mathcal O)\to L^2(\mathcal O),\qquad L_f[u]= \nabla\cdot (f\nabla u).$$
By standard theory for elliptic PDEs, $L_f$ has a linear, continuous inverse operator, which we denote by
$$V_f:L^2(\mathcal O)\to H^2_0(\mathcal O),\qquad \psi\mapsto V_f\left[\psi\right],$$
see \cite{evans}, Theorem 4 in Chapter 6.3. In other words, for each right hand side $\psi\in L^2$, there exists a unique function $w_{f,\psi}:=V_f\left[\psi\right]\in H^2_0$ solving the Dirichlet problem
\begin{equation}\label{div2}
\begin{cases}
L_f[w_{f,\psi}]=\psi \quad \textnormal{on }\mathcal O,\\
w_{f,\psi}=0 \quad \textnormal{on }\partial \mathcal O
\end{cases}
\end{equation}
weakly, i.e. in the sense that the identity
\begin{equation}\label{div-wk}
-\int_{\mathcal O}\sum_{i=1}^{d}f D_iw_{f,\psi}D_iv \;=\int_{\mathcal O}\psi v \;
\end{equation}
holds for all test functions $v\in H^1_0(\mathcal O)$ (cf. \cite{evans}, Chapter 6). By the zero boundary conditions of (\ref{div}) and the divergence theorem, any classical solution (i.e. $C^2$ solution) must be equal to the unique weak solution when interpreted as an $H^2_0$ function.
\par
Theorem 4 in Chapter 6.3 of \cite{evans} implies that there exists a constant $C=C_f$ (allowed to depend on $f$) such that for all $\psi\in L^2$, we have the norm estimate
$\|V_f\left[\psi\right]\|_{H^2}\leq C_f\|\psi\|_{L^2},$ and we need a result that tracks the dependence of $C_f$ on $f$ in a quantitative way. We first establish that when we only seek an $L^p\to L^p$-estimate, $p\in\{2,\infty\}$, rather than an $L^2 \to H^2$-estimate, the constant merely depends on the lower bound $K_{min}$ for $f$.
\begin{lem}\label{div-lem-lp}
Let $K_{min}>0$. Then there exists $C=C(d,\mathcal O, K_{min})$ such that for all $f\in C^2(\mathcal O)$ with $f\geq K_{min}>0$ and $\psi\in L^2$, we have
\begin{equation}\label{div-l2}
\|V_f\left[\psi\right]\|_{L^2}\leq C\|\psi\|_{L^2}
\end{equation}
and for all $\psi\in C^\eta(\mathcal O), \eta>0$,
\begin{equation}\label{div-linf}
\|V_f\left[\psi\right]\|_{\infty}\leq C\|\psi\|_{\infty}.
\end{equation}
\end{lem}
\begin{proof}
Assume first that $\psi \in C^\eta(\mathcal O)$ so that $V_f[\psi]\in C(\bar{\mathcal O})\cap C^2(\mathcal O)$ (see after (\ref{div})). Then we have the Feynman-Kac formula
\begin{equation}\label{div-fk}
V_f[\psi](x)=-\frac{1}{2}\mathbb{E}^x\left[\int_0^{\tau_\mathcal O}\psi(X^f_s)ds \right], \qquad x\in\mathcal O,
\end{equation}
where $(X^f_s:s \ge 0)$ is a diffusion Markov process started at $x \in \mathcal O$ with infinitesimal generator $L_f/2$ and expectation operator $\mathbb{E}^x$, and where $\tau_\mathcal O$ is the exit time of $X^f_s$ from $\mathcal O$, see, e.g., Theorem 1.2 in Section II of \cite{bass}. We also record that, by Theorem 4.3 in Section VII of \cite{bass} and inspection of its proof, there exists a constant $c_1$ only depending on the lower bound $K_{min}<f$ and on $d$, such that the transition densities of $(X^f_s: s \ge 0)$ exist and satisfy the estimate
\begin{equation}\label{div-heat}
p_f(t,x,y) \le c_1 t^{-d/2}, \qquad t>0,\quad x,y\in\mathbb{R}^d.
\end{equation}
Then, arguing as in the proof of Theorem 1.17 in \cite{chungzhao}, with (\ref{div-heat}) replacing the standard heat kernel estimate for Brownian motion, we obtain that $\sup_{x \in \mathcal O}\mathbb{E}^x \tau_\mathcal O\leq c,$ with $c=c(\mathcal O, d, c_1)$, and hence (\ref{div-linf}) follows from
\begin{equation}\label{vf1}
\|V_f[1]\|_\infty \le \sup_{x \in \mathcal O}\mathbb{E}^x \tau_\mathcal O \le c.
\end{equation}
Using what precedes one further shows that $V_f$ has a representation via a non-negative and symmetric integral kernel $G_f(\cdot,\cdot)$, such that
\begin{equation}\label{div-green}
V_f[\psi](x)=-\int_{\mathcal O}G_f(x,y)\psi(y)dy, ~ x\in\mathcal O,~~\forall~ \psi\in C^\eta(\mathcal O).
\end{equation}
Then using (\ref{vf1}), the Cauchy-Schwarz inequality and the positivity of $G$ we have for all $\psi\in C^\eta(\mathcal O)$,
\[\|V_f[\psi]\|_{L^2}^2 \le \int_{\mathcal O} \int_{\mathcal O} G_f(x,y)dy \int_{\mathcal O} G_f(x,y)\psi^2(y)dy dx \le \|V_f[1]\|_\infty^2 \|\psi\|_{L^2}^2,\]
whence (\ref{div-l2}) follows for $\psi\in C^\eta(\mathcal O)$, and extends to $\psi\in L^2$ by approximation since $V_f$ is a continuous operator on $L^2(\mathcal O)$ (as established above).
\end{proof}
Lemma \ref{div-lem-lp} will be used in the proof of the following stronger elliptic regularity estimate.
\begin{lem}\label{lem-div-h2}
Let $K_{min}>0$. Then there exists a universal constant $C>0$ such that for all $f\in C^{2}(\mathcal O)$ with $f\geq K_{min}$ and $\psi\in L^2(\mathcal O)$, the unique weak solution $w_{f,\psi}=V_f[\psi]$ to (\ref{div2}) satisfies
\begin{align}
\|V_f\left[\psi\right]\|_{H^2}&\leq C\left(1+\|f\|_{C^1}\right)\|\psi\|_{L^2}, \label{div-l2h2}\\
\|V_f\left[\psi\right]\|_{L^2}&\leq C\left(1+\|f\|_{C^1}\right)\|\psi\|_{(H^2_0)^*} \label{div-h-2},
\end{align}
where $C$ only depends on $K_{min}$ and $\mathcal O, d$.
\end{lem}
\begin{proof}
Let $f\in C^1$ and $\psi\in L^2$. By (\ref{h-isom}), there exists a constant $C>0$ depending only on $\mathcal O, d$ such that for all $u\in H^2_0$,
\begin{equation}\label{lapl-isom}
C^{-1}\|\Delta u\|_{L^2}\leq \|u\|_{H^2}\leq C\|\Delta u\|_{L^2}.
\end{equation}
Moreover we have by the definition of $L_f$ that
\begin{equation}\label{div-lapl}
\Delta u=f^{-1}(L_fu-\nabla f\cdot \nabla u).
\end{equation}
Writing $w=w_{f,\psi}$ and utilising (\ref{lapl-isom}) and (\ref{div-lapl}), we can estimate
\begin{equation}\label{h2start}
\begin{split}
\|w\|_{H^2}&\leq C\|\Delta w\|_{L^2}=C\left\|f^{-1}(\psi-\nabla w\cdot\nabla f)\right\|_{L^2}\\
&\leq CK_{min}^{-1}\left(\|\psi\|_{L^2}+\|f\|_{C_1}\|w\|_{H^1}.\right)
\end{split}
\end{equation}
By choosing the test function $-w\in H^1_0$ in the weak formulation (\ref{div-wk}), we have that
\[K_{min}\int_{\mathcal O}|Dw|^2\leq \int_{\mathcal O}\sum_{i=1}^{d}f(D_iw)^2 =\int_{\mathcal O}-\psi w \;\leq \frac 12 \int_{\mathcal O}(\psi^2+w^2) \;. \]
Combining this with (\ref{h2start}) and Lemma \ref{div-lem-lp}, we finally obtain that for constants $C',C'',C'''$ only depending on $K_{min}$ and $\mathcal O$, we have
\begin{equation*}
\begin{split}
\|w\|_{H^2}&\leq C'K_{min}^{-1}\left(\|\psi\|_{L^2}+\|f\|_{C^1}C''(\|\psi\|_{L^2}+\|w\|_{L^2} )\right)\\
&=C'''\left(1+\|f\|_{C^1}\right)\|\psi\|_{L^2},
\end{split}
\end{equation*}
which proves (\ref{div-l2h2}).
\par
Next, using the divergence theorem and (\ref{div-l2h2}), we obtain (\ref{div-h-2}) from
\begin{align*}
\|V_{f}[\psi]\|_{L^2}&=\sup_{\varphi\in C^\infty_c,\;\; \|\varphi\|_{L^2}\leq 1}\left|\int_{\mathcal O} V_{f}[\psi]\varphi \right|\\
&=\sup_{\varphi\in C_c^\infty (\mathcal O),\;\; \|\varphi\|_{L^2}\leq 1}\left|\int_{\mathcal O} V_{f}[\psi]L_fV_f[\varphi] \right|\\
&=\sup_{\varphi\in C_c^\infty (\mathcal O),\;\; \|\varphi\|_{L^2}\leq 1}\left|\int_{\mathcal O} \psi V_f[\varphi] \right|\\
& \leq C(1+\|f\|_{C^1})\sup_{\varphi \in H^2_0,\;\; \|\varphi\|_{H^2}\leq 1}\left|\int_{\mathcal O} \psi \varphi \right| =C(1+\|f\|_{C^1})\|\psi\|_{(H^2_0)^*}.
\end{align*}
\end{proof}
\subsubsection{Estimates for $G$}\label{sec-div-G}
Now we turn to the forward map $G$ representing the solutions of the PDE (\ref{div}). The following norm estimate for the $\mathcal C^2$-H\"older-Zygmund norm of $G(f)=u_f$ is needed.
\begin{lem}\label{lem-div-c}
Suppose that for some $K_{min}>0$, $\alpha>d/2+2$ and $g\in \mathcal C^\eta(\mathcal O), \eta>0$, $\tilde{\mathcal F}$ is as in (\ref{ftilde}) and $u_f$ denotes the unique solution of (\ref{div}). Then there exists $C=C(d, \mathcal O, K_{min}, \|g\|_\infty)$ such that for all $f\in\tilde{\mathcal F}$,
\begin{equation}\label{div-c}
\|u_f\|_{\mathcal C^2}\leq C\left(1+\|f\|_{\mathcal C^1}^2\right).
\end{equation}
\end{lem}
\begin{proof}
The proof is similar to that of Lemma \ref{lem-div-h2}. By (\ref{c-isom}), there exists a constant $C>0$ depending only on $\mathcal O, d$ such that for all functions $u\in \mathcal C^2_0(\mathcal O)$, we have
\begin{equation}\label{c2-isom}
C^{-1}\|\Delta u\|_{\mathcal C^0}\leq \|u\|_{\mathcal C^2}\leq C \|\Delta u\|_{\mathcal C^0}.
\end{equation}
Using this, the PDE (\ref{div}), the multiplicative inequality (\ref{c-mult}) and the interpolation inequality (\ref{c-int}), we can estimate as in (\ref{h2start})
\begin{equation*}
\begin{split}
\|u_{f}\|_{\mathcal C^2}&\lesssim \|f^{-1}(g-\nabla f\cdot \nabla u_f)\|_{\mathcal C^0} \lesssim \|f^{-1}\|_{\mathcal C^0}\left(\|g\|_{\mathcal C^0}+\|f\|_{\mathcal C^1} \|u_f\|_{\mathcal C^1}\right)\\
&\lesssim K_{min}^{-1}\left(\|g\|_{\mathcal C^0}+\|f\|_{\mathcal C^1}\|u_f\|_{\mathcal C^2}^{1/2}\|u_f\|_{\mathcal C^0}^{1/2}\right).
\end{split}
\end{equation*}
Dividing this inequality by $\|u_f\|_{\mathcal C^2}^{1/2}$ whenever $\|u_f\|_{\mathcal C^2}^{1/2}\geq 1$ and otherwise estimating it by $1$, we obtain that
\begin{equation*}
\|u_{f}\|_{\mathcal C^2}\lesssim 1+ K_{min}^{-2}\left(\|g\|_{\mathcal C^0}^2+\|f\|_{\mathcal C^1}^2\|u_f\|_{\mathcal C_0}\right)\lesssim 1+ K_{min}^{-2}\left(\|g\|_{\infty}^2+\|f\|_{\mathcal C^1}^2\|g\|_{\infty}\right)
\end{equation*}
where in last step we used $\|\cdot\|_{\mathcal C^0}\lesssim\|\cdot\|_{\infty}$ and Lemma \ref{div-lem-lp}.
\end{proof}
We also need that the forward map $G$ maps bounded sets in $H^\alpha$ onto bounded sets in $H^{\alpha+1}$.
\begin{lem}\label{lem-div-bd}
Suppose that $\alpha,\tilde{\mathcal F}$ are as in Lemma \ref{lem-div-c} and for some $g\in H^{\alpha-1}(\mathcal O)$, let $u_f=w_{f,g}, f \in \tilde {\mathcal F},$ be the unique solution of (\ref{div}). Then $u_f \in H^{\alpha+1}(\mathcal O)$ and there exists a constant $C=C(\alpha, d, \mathcal O, K_{min})>0$ such that
\begin{equation}\label{div-h-est}
\|u_{f}\|_{H^{\alpha+1}}\leq C\big(1+\|f\|_{H^\alpha}^{\alpha^2+\alpha}\big)\big( \|g\|_{H^{\alpha-1}}^{\alpha+1} \vee \|g\|_{H^{\alpha-1}}^{1/(\alpha+1)} \big).
\end{equation}
\end{lem}
\begin{proof}
First, suppose $f\in C^\infty\cap \tilde{\mathcal F}$.
By (\ref{h-isom}), the standard Laplacian $\Delta$ establishes an isomorphism between $H^{\alpha+1}_0$ and $H^{\alpha-1}$, and by Theorem 8.13 in \cite{gt}, $u_f \in H^{\alpha+1}_0$. Then (\ref{div-lapl}) and the multiplicative inequality (\ref{h-mult}) give
\begin{equation*}
\begin{split}
\|u_f\|_{H^{\alpha+1}}&\lesssim \|f^{-1}(g-\nabla f\cdot \nabla u_f)\|_{H^{\alpha-1}}\\
& \lesssim\|f^{-1}\|_{H^{\alpha-1}}(\|g\|_{H^{\alpha-1}}+\|f\|_{H^\alpha}\|u_f\|_{H^\alpha}).
\end{split}
\end{equation*}
Noting that the map $\Psi: (K_{min},\infty)\to\mathbb R,~ x\mapsto x^{-1}$ satisfies (\ref{phibddder}), (\ref{phi-sob}) implies that there exists $c>0$ such that for all $f\in\mathcal F$,
\[\|f^{-1}\|_{H^{\alpha-1}}\leq c(1+\|f\|_{H^{\alpha-1}}^{\alpha-1}).\]
Using this and (\ref{h-int}), we obtain
\begin{equation*}
\begin{split}
\|u_f\|_{H^{\alpha+1}}&\lesssim(1+ \|f\|_{H^{\alpha-1}}^{\alpha-1})(\|g\|_{H^{\alpha-1}}+\|f\|_{H^\alpha}\|u_f\|_{H^\alpha})\\
& \lesssim(1+ \|f\|_{H^{\alpha}}^\alpha)\big(\|g\|_{H^{\alpha-1}}+\|u_f\|_{H^{\alpha+1}}^{\frac \alpha{\alpha+1}}\|u_f\|_{L^2}^{\frac{1}{\alpha+1}}\big)
\end{split}
\end{equation*}
When $\|u_f\|_{H^{\alpha+1}}\leq 1$ we use (\ref{div-l2}) to deduce
\begin{equation*}
\|u_f\|_{H^{\alpha+1}} \lesssim (1+ \|f\|_{H^{\alpha}}^\alpha)\big(\|g\|_{H^{\alpha-1}}+\|g\|_{L^2}^{\frac{1}{\alpha+1}}\big),
\end{equation*}
and when $\|u_f\|_{H^{\alpha+1}}\geq 1$, then dividing both sides by $\|u_f\|_{H^{\alpha+1}}^{\frac \alpha{\alpha+1}}$ and using again (\ref{div-l2}) yields
\begin{equation*}
\|u_f\|^{1/(\alpha+1)}_{H^{\alpha+1}}\lesssim (1+ \|f\|_{H^{\alpha}}^\alpha)\big(\|g\|_{H^{\alpha-1}}+\|g\|_{L^2}^{\frac{1}{\alpha+1}}\big).
\end{equation*}
Combining the preceding bounds and using $\|\cdot\|_{L^2} \lesssim \|\cdot\|_{H^{\alpha-1}}$ implies (\ref{div-h-est}) for smooth $f \in \tilde{\mathcal F}$. Now for any $f\in\tilde{\mathcal F}$, take $f_n\in C^\infty(\mathcal O), f_n>K_{min}/2,$ such that $f_n\to f$ in $H^{\alpha}$ as $n \to \infty$, and hence by (\ref{div-h-est}) the sequence $u_{f_n}$ is bounded in $H^{\alpha+1}$. Then applying (\ref{obpertu}) to $u_{f_n}-u_{f_m}, m,n \in \mathbb N,$ and applying (\ref{div-h-est}) with $g=\nabla \cdot ((f_m- f_n) \nabla u_{f_m}),$ one shows that $u_{f_n}$ is a Cauchy sequence in $H^{\alpha+1}$ converging to $u_f$, and taking limits extends the inequality (\ref{div-h-est}) to the general case $f\in\mathcal F$.
\end{proof}
\subsubsection{Stability Estimates for $G^{-1}$}
The following estimate for the inverse map $u_f \mapsto f$ allows to obtain convergence rates for $\|\hat f-f_0\|_{L^2}$ via rates for $\|u_{\hat f} - u_{f_0}\|_{H^2}$, with choices $f_0 = f_1$ and $\hat f=f_2$. As $\hat f$ is random we explicitly track the dependence of the constants on $f_2$.
\begin{lem}\label{lem-div-stab}
Let $\alpha>d/2+2, g_{min}, K_{min}, B, \eta$ be given, positive constants and let $\tilde{\mathcal F}$ be given by (\ref{ftilde}). For $g\in C^\eta(\mathcal O)$ with $\inf_{x \in \mathcal O}g(x)\geq g_{min}$, denote by $u_f$ the unique solution of (\ref{div}). Then there exists $C=C(g_{min}, K_{min}, B, \mathcal O,d)<\infty$ such that for all $f_1,f_2\in \tilde{\mathcal F}$ with $\|f_1\|_{C^1} \vee \|u_{f_1}\|_{C^2} \le B$, we have
\begin{equation*}
\|f_1-f_2\|_{L^2}\leq C\|f_2\|_{C_1}\|u_{f_1}-u_{f_2}\|_{H^2}.
\end{equation*}
\end{lem}
\begin{proof}
For $f_1,f_2\in\tilde{\mathcal F}$ write $h=f_1-f_2$. By (\ref{div}), we have
\begin{equation}\label{div-diff}
\begin{split}
\nabla \cdot (h \nabla u_{f_1}) &= \nabla\cdot (f_1 \nabla u_{f_1}) - \nabla\cdot (f_2 \nabla u_{f_2}) - \nabla\cdot (f_2 \nabla (u_{f_1}-u_{f_2})) \\
&=\nabla \cdot (f_2 \nabla (u_{f_2}-u_{f_1})).
\end{split}
\end{equation}
We can upper bound the $\|\cdot\|_{L^2}$-norm of the right hand side by
\begin{align}
\|\nabla \cdot (f_2 \nabla (u_{f_2}-u_{f_1}))\|_{L^2}&\leq \|\nabla f_2\|_{\infty}\|u_{f_2}-u_{f_1}\|_{H^1} +\|f_2\|_{\infty}\|u_{f_2}-u_{f_1}\|_{H^2}\notag \\
&\leq 2\|f_2\|_{C^1}\|u_{f_2}-u_{f_1}\|_{H^2}.\label{diff-ub}
\end{align}
Next, we lower bound the $\|\cdot\|_{L^2}$-norm of the left side of (\ref{div-diff}). For regular enough $v$ we see from Green's identity (p.17 in \cite{gt}) that
\begin{align*}
\langle \Delta u_{f_1}, v^2 \rangle_{L^2} + \frac{1}{2} \langle \nabla u_{f_1}, \nabla (v^2) \rangle_{L^2} = \frac{1}{2}\langle \Delta u_{f_1}, v^2 \rangle_{L^2} + \frac{1}{2}\int_{\partial \mathcal O} \frac{\partial u_{f_1}}{\partial n} v^2.
\end{align*}
Moreover for $v=e^{-\lambda u_{f_1}}h$ with $\lambda>0$ to be chosen we have
\begin{equation*}
\frac{1}{2} \int_{\mathcal O} \nabla (v^2) \cdot \nabla u_{f_1} =- \int_{\mathcal O} \lambda \|\nabla u_{f_1}\|^2 v^2 + \int_\mathcal O v e^{-\lambda u_{f_1}}\nabla h \cdot \nabla u_{f_1},
\end{equation*}
so that by the Cauchy-Schwarz inequality
\begin{align} \label{keylb}
& \left|\int_{\mathcal O}\Big(\frac{1}{2}\Delta u_{f_1}+\lambda \|\nabla u_{f_1}\|^2\Big)v^2 + \int_{\partial \mathcal O}\frac 12 \frac{\partial u_{f_1}}{\partial n}v^2\right| \notag \\
& = \left|\langle (\Delta u_{f_1} + \lambda \|\nabla u_{f_1}\|^2), v^2 \rangle_{L^2} + \frac{1}{2} \langle \nabla u_{f_1}, \nabla (v^2)\rangle_{L^2}\right| \notag \\
&= \left|\langle h \Delta u_{f_1} + \nabla h \cdot \nabla u_{f_1}, h e^{-2\lambda u_{f_1}} \rangle_{L^2} \right| \le \mu \|\nabla \cdot (h \nabla u_{f_1})\|_{L^2} \|h\|_{L^2}
\end{align}
for $\mu=\exp(2\lambda \|u_{f_1}\|_{\infty})$. [The preceding argument is adapted from the proof of Theorem 4.1 in \cite{itokunisch}.] We next lower bound the multipliers of $v^2$ in the integrands in the first line of the last display. First we have
\begin{equation*}
0<g_{min} \le g= L_{f_1}u_{f_1} = f_1(x) \Delta u_{f_1} + \nabla f_1 \cdot \nabla u_{f_1},~\text{ on } \mathcal O,
\end{equation*}
so that either $\Delta u_{f_1}(x) \ge g_{min}/2\|f_1\|_\infty$ or $\|\nabla u_{f_1}(x)\|^2 \ge (g_{min} / 2\|f_1\|_{C^1})^2$ on $\mathcal O$. Since $\|\Delta u_{f_1}\|_\infty\le c(B)$ this implies for $\lambda=\lambda (g_{min}, B)$ large enough that
\begin{equation}\label{richterlb}
\frac{1}{2}\Delta u_{f_1}(x) + \lambda \|\nabla u_{f_1}(x)\|^2 \ge c_0>0,~~ x \in \mathcal O,
\end{equation}
for some $c_0=c_0(g_{min}, B)$. Next, for the integral over $\partial \mathcal O$, we use again $L_{f_1}u_{f_1}=g> 0$ and apply the Hopf boundary point Lemma 6.4.2 in \cite{evans}: We have $u_{f_1}(x_0)=0$ for any $x_0 \in \partial \mathcal O$ but $u_{f_1}(x)<0$ for all $x \in \mathcal O$: Indeed, by $g \ge g_{min}>0$ and the Feynman-Kac formula (\ref{div-fk}) (with $g=\psi$), it suffices to lower bound $\mathbb E^x \tau_\mathcal O$ which satisfies, by Markov's inequality $$\mathbb E^x \tau_\mathcal O \ge \mathbb P^x(\tau_\mathcal O >1) \ge \mathbb P^x\big(\sup_{0 <s\le 1}\|X_s-x\|<\|x-\partial \mathcal O\|\big)>0$$ in view Theorem V.2.5 in \cite{bass} with $\psi(s)=x$ identically for all $s$. Lemma 6.4.2 in \cite{evans} now gives $\partial u_{f_1}/\partial n \ge 0$ for all $x\in \partial \mathcal O.$ Combining this with (\ref{keylb}) and (\ref{richterlb}) we deduce
\begin{equation*}
\|\nabla \cdot (h \nabla u_{f_1})\|_{L^2}\|h\|_{L^2}\geq c'(g_{min}, K_{min}, B, \mathcal O,d)\|v\|^2_{L^2}\gtrsim \|h\|^2_{L^2},
\end{equation*}
which together with (\ref{diff-ub}) yields the desired estimate.
\end{proof}
\subsection{Schr\"odinger equation}\label{sec-schr-facts}
\subsubsection{Estimates for $V_f$ and $G$}\label{sec-schr-G}
In this section, for each $f\in C(\mathcal O)$ with $f\geq 0$, let $L_f$ denote the Schr\"odinger differential operator
\begin{equation*}
L_f:H^2_0(\mathcal O)\to L^2(\mathcal O),\qquad L_f[u]= \Delta u-2fu,
\end{equation*}
where $H^2_0$ is given by (\ref{h0}). As in the divergence form case, $L_f$ is a bijection with a linear, continuous inverse operator which we again denote by
\begin{equation*}
V_f:L^2(\mathcal O)\to H^2_0(\mathcal O),\qquad \psi\mapsto V_f\left[\psi\right].
\end{equation*}
In other words, for any $f\in C(\mathcal O)$ and $\psi\in L^2$ the \emph{inhomogeneous} equation
\begin{equation}\label{schr-2}
\begin{cases}
\Delta u-2fu =\psi \quad \textnormal{on} \quad \mathcal O,\\
u=0\quad \textnormal{on} \quad \mathcal \partial O
\end{cases}
\end{equation}
has a unique weak solution which we shall denote by $\omega_{f,\psi}:=V_f[\psi]\in H^2_0(\mathcal O)$, see Theorem 4 in Chapter 6.3 of \cite{evans} for this standard result for elliptic PDEs.
\par
As in the divergence form case, we first observe that for $p\in \{2,\infty\}$, the $L^p\to L^p$ operator norm of $V_f$ can be upper bounded uniformly in $f$.
\begin{lem}\label{schafk}
There exists a constant $C>0$ such that for all $f\in C(\mathcal O)$ with $f\geq 0$ and $\psi\in L^2(\mathcal O)$, $w_{f,\psi}=V_f[\psi]$ satisfies
\[\|V_f\left[\psi\right]\|_{L^2}\leq C\|\psi\|_{L^2} \]
and if $\psi\in C(\mathcal O)$, then also
\[\|V_f\left[\psi\right]\|_{\infty}\leq C\|\psi\|_{\infty}. \]
\end{lem}
\begin{proof}
We have the Feynman-Kac representation
$$w_{f,\psi}(x) = -\frac{1}{2}\mathbb E^x \Big[\int_0^{\tau_\mathcal O} \psi(X_s) e^{-\int_0^s f(X_r)dr}ds \Big], ~x \in \mathcal O,~ \psi \in C(\mathcal O),$$ where $(X_s: s \ge 0)$ is a standard $d$-dimensional Brownian motion started at $x$, with exit time $\tau_{\mathcal O}$ from $\mathcal O$, see p.84 and Theorem 3.22 of \cite{chungzhao} . [These results are applicable as $C(\mathcal O)\subseteq J$ with $J$ defined on p.62 of \cite{chungzhao}, and $C(\mathcal O)\subseteq \mathbb F(D,q)$ with $\mathbb F(D,q)$ defined on p.80 of \cite{chungzhao}.] The proof is now similar to that of Lemma \ref{div-lem-lp}, using $f \ge 0$ and that $\sup_{x\in\mathcal O}\mathbb{E}^x[\tau_{\mathcal O}]\leq K(vol(\mathcal O), d)<\infty$ by Theorem 1.17 in \cite{chungzhao}.
\end{proof}
Using the above lemma, we now show the following regularity estimate.
\begin{lem}\label{lem-schr-h2}
There exists a constant $C$ such that for all $f\in C^1(\mathcal O)$ with $f\geq 0$ and $\psi\in L^2(\mathcal O)$, we have
\begin{align}
\|V_f\left[\psi\right]\|_{H^2}&\leq C(1+\|f\|_{\infty})\|\psi\|_{L^2}, \label{schr-l2h2}\\
\|V_f\left[\psi\right]\|_{L^2}&\leq C(1+\|f\|_{\infty})\|\psi\|_{(H^2_0)^*}. \label{schr-h-2l2}
\end{align}
\end{lem}
\begin{proof}
By the norm equivalence (\ref{lapl-isom}) and (\ref{schr-2}), we have that
\begin{equation*}
\begin{split}
\|V_f\left[\psi\right]\|_{H^2}&\lesssim \left\|\Delta V_f\left[\psi\right]\right\|_{L^2}\leq \|L_fV_f\left[\psi\right]\|_{L^2}+\|fV_f\left[\psi\right]\|_{L^2}\\
&\leq \|\psi\|_{L^2}+\|f\|_{\infty}\|V_f\left[\psi\right]\|_{L^2}\lesssim (1+\|f\|_{\infty})\|\psi\|_{L^2},
\end{split}
\end{equation*}
which proves (\ref{div-l2h2}). The second estimate (\ref{schr-h-2l2}) now follows from the same duality argument as in the proof of (\ref{div-h-2}).
\end{proof}
Next, we prove some basic boundedness properties of the forward map $G:f\mapsto u_f$.
\begin{lem}\label{lem-schr-c2}
Suppose that for some $g\in C^\infty(\partial \mathcal O)$, $\alpha>d/2$ and $K_{min}\geq 0$, $\tilde{\mathcal F}$ is as in (\ref{ftilde}), and let $u_f$ be the unique solution of (\ref{schroedeq}).
\par
1. There exists $C>0$ (independent of $g$) such that for all $f\in\tilde{\mathcal F}$, we have
\[\|u_f\|_{\mathcal C^2(\mathcal O)}\leq C(1+\|f\|_{\infty})\|g\|_{\mathcal C^2(\mathcal O)}. \]
\par
2. There exists $C>0$ (possibly depending on $g$) such that for all $f\in\tilde{\mathcal F}$,
\[\|u_f\|_{ H^{\alpha+2}(\mathcal O)}\leq C(1+\|f\|_{H^\alpha}^{\alpha/2+1}). \]
\end{lem}
\begin{proof}
By (\ref{c-isom}), $(\Delta,\textnormal{tr}[\cdot])$ is a topological isomorphism between the spaces $\mathcal C^2(\mathcal O)$ and $\mathcal C^0(\mathcal O)\times \mathcal C^{2}(\partial O)$, whence we deduce that for all $u\in\mathcal C^2(\mathcal O)$, we have the norm estimate
\[\|u\|_{\mathcal C^2(\mathcal O)}\leq C\left(\|\Delta u\|_{\mathcal C^0(\mathcal O)}+\|\textnormal{tr}[u]\|_{\mathcal C^2(\partial \mathcal O)} \right). \]
Using this, the PDE (\ref{schroedeq}) and the triangle inequality, we have for $f\in \mathcal F$,
\begin{equation}\label{ufc2}
\begin{split}
\|u_f\|_{\mathcal C^2(\mathcal O)}&\lesssim \left\|L_fu_f \right\|_{\mathcal C^0(\mathcal O)}+\|fu_f\|_{\mathcal C^0(\mathcal O)}+\left\|\text{tr}[u_f]\right\|_{\mathcal C^2(\partial\mathcal O)}\\
&\leq \|f\|_{\infty}\|u_f\|_{\infty}+\|g\|_{\mathcal C^2(\partial \mathcal O)}.
\end{split}
\end{equation}
Next, we claim that there exists a constant $C>0$ such that for all $f,g$ as in the hypotheses, we have
\begin{equation}\label{schr-hom-linf}
\|u_f\|_{\infty}\leq C\|g\|_{\infty}.
\end{equation}
Indeed, this can be seen immediately from the fact that $f\geq 0$ and the Feynman-Kac representation (see \cite{chungzhao}, Theorem 4.7)
\begin{equation}\label{schr-fk}
u_f(x)=\frac{1}{2}\mathbb E^x\left[g(X_{\tau_{\mathcal O}})e^{-\int_0^{\tau_{\mathcal O}}f(X_s)ds}\right], \quad x\in\mathcal O,
\end{equation}
where $(X_s:s\geq 0), \tau_\mathcal O$ are as in the proof of Lemma \ref{schafk}. Hence, combining (\ref{schr-hom-linf}) with (\ref{ufc2}) yields the desired estimate
\[\|u_f\|_{\mathcal C^2(\mathcal O)}\lesssim \|f\|_{\infty}\|g\|_{L^\infty(\mathcal O)}+\|g\|_{\mathcal C^2(\partial \mathcal O)}\leq (1+\|f\|_{\infty})\|g\|_{\mathcal C^2(\partial \mathcal O)}. \]
\par
For the second part, we initially assume $f \in C^\infty(\mathcal O)$ so that $u_f \in C^\infty(\mathcal O)$ too (see Corollary 8.11 in \cite{gt}), and then use the topological isomorphism $(\Delta,\textnormal{tr})$ between $H^{\alpha+2}(\mathcal O)$ and $H^\alpha(\mathcal O)\times H^{\alpha+3/2}(\partial\mathcal O)$, which yields
\begin{equation*}
\begin{split}
\|u_f\|_{H^{\alpha+2}(\mathcal O)}&\lesssim \|\Delta u_f\|_{H^\alpha(\mathcal O)}+\|\textnormal{tr}[u_f]\|_{H^{\alpha+3/2}({\partial\mathcal O})}\lesssim \|fu_f\|_{H^\alpha(\mathcal O)}+\|g\|_{C^{\alpha+2}(\partial \mathcal O)}\\
&\lesssim 1+ \|f\|_{H^\alpha}\|u_f\|_{H^\alpha}\lesssim 1+\|u_f\|_{H^{\alpha+2}}^{\frac{\alpha}{\alpha+2}}\|u_f\|_{L^2}^{\frac{2}{\alpha+2}}\|f\|_{H^\alpha}.
\end{split}
\end{equation*}
Dividing this by $\|u_f\|_{H^{\alpha+2}}^{\frac{\alpha}{\alpha+2}}$ when $\|u_f\|_{H^{\alpha+2}}\ge 1$ and otherwise estimating it by $1$, and using (\ref{schr-hom-linf}), we have that
\[\|u_f\|_{H^{\alpha+2}}\lesssim 1+\|u_f\|_{L^2} \|f\|_{H^\alpha}^{\frac{\alpha+2}{2}} \lesssim 1+\|g\|_{\infty} \|f\|_{H^\alpha}^{\frac{\alpha+2}{2}}\lesssim1+\|f\|_{H^\alpha}^{\frac{\alpha+2}{2}}. \] The case of general $f \in \tilde {\mathcal F}$ now follows from taking smooth $f_n >K_{min}/2, f_n \to f $ in $H^\alpha$, showing that $u_{f_n}$ is Cauchy in $H^{\alpha+2}$ (by using (\ref{lapl-isom}), (\ref{h-int}), Lemma \ref{schafk}), and taking limits in the last inequality. Details are left to the reader.
\end{proof}
\subsubsection{Estimates for $G^{-1}$}
\begin{lem}\label{lem-schr-stab}
Suppose that for some $\alpha>d/2,K_{min}\geq 0$, $g_{min}>0$ and $g\in C^\infty(\partial \mathcal O)$ with $\inf_{x \in \partial \mathcal O}g(x)\geq g_{min}$, $\tilde{\mathcal F}$ is given by (\ref{ftilde}), and let $u_f$ denote the unique solution of (\ref{schroedeq}). Then there exist constants $c_1,c_2>0$ such that for all $f_1,f_2\in\tilde{\mathcal F}$, we have
\begin{equation*}
\begin{split}
\|f_1-f_2\|_{L^2}\leq c_1\big( &e^{c_2\|f_1\|_{\infty}}\left\|u_{f_1}-u_{f_2}\right\|_{H^2}\\
&\qquad +\|u_{f_2}\|_{C^2}\;e^{c_2\|f_1\vee f_2\|_{\infty}}\|u_{f_1}-u_{f_2}\|_{L^2}\big).
\end{split}
\end{equation*}
\end{lem}
\begin{proof}
Applying Jensen's inequality to the Feynman-Kac representation (\ref{schr-fk}), and since $\sup_x \mathbb{E}^x\tau_\mathcal O \le c<\infty$ (see the proof of Lemma \ref{schafk}) yields
\begin{equation}\label{schr-inf-est}
\inf_{x\in\mathcal O}u_f(x)\geq g_{min}\inf_{x\in\mathcal O}e^{-\|f\|_{\infty}\mathbb{E}^x\tau_{\mathcal O}}\geq g_{min}e^{-c\|f\|_{\infty}}>0.
\end{equation}
Moreover, (\ref{schroedeq}) yields that we have $f=\frac{\Delta u_f}{2u_f}$ on $\mathcal O$, for all $f\in\tilde{\mathcal F}$. Thus, for any $f_1,f_2\in\tilde{\mathcal F}$, we can estimate
\begin{equation}\label{schr-stab-1}
\begin{split}
\| f_1-&f_2\|_{L^2}=\frac 12\|\frac{\Delta u_{f_1}}{u_{f_1}}-\frac{\Delta u_{f_2}}{u_{f_2}} \|_{L^2}\\
&\lesssim \left\|\left(\Delta u_{f_1}- \Delta u_{f_2}\right)u_{f_1}^{-1} \right\|_{L^2}+\left\|\Delta u_{f_2}\left(u_{f_1}^{-1}-u_{f_2}^{-1} \right)\right\|_{L^2}\\
&\lesssim \big(\inf_{x\in\mathcal O}\big|u_{f_1}(x)\big|\big)^{-1}\left\|u_{f_1}-u_{f_2}\right\|_{H^2}+\big\|\Delta u_{f_2}\big\|_{C^2}\big\|u_{f_1}^{-1}-u_{f_2}^{-1}\big\|_{L^2}.
\end{split}
\end{equation}
Further, using the mean value theorem and (\ref{schr-inf-est}), we have that
\begin{equation*}
\left|u_{f_1}^{-1}-u_{f_2}^{-1}\right|\leq \max\{u_{f_1}^{-2},u_{f_2}^{-2}\}\left|u_{f_1}-u_{f_2}\right|\leq g_{min}^{-2}e^{2c\|f_1\vee f_2\|_{\infty}}\left|u_{f_1}-u_{f_2}\right|.
\end{equation*}
Combining this with (\ref{schr-stab-1}) and using (\ref{schr-inf-est}) once more, we obtain that
\[\left\| f_1-f_2\right\|_{L^2}\lesssim e^{c\|f_1\|_{\infty}}\left\|u_{f_1}-u_{f_2}\right\|_{H^2}+ e^{2c\|f_1\vee f_2\|_{\infty}}\left\|u_{f_1}-u_{f_2}\right\|_{L^2}, \]
which concludes the proof.
\end{proof}
\section{Some properties of regular link functions}\label{sec-reg}
We define $L^p$-norms, $0<p \le \infty$, in the usual way. By obvious modifications, the following lemma holds also for regular functions $\Phi:(a,b)\to \mathbb{R}$ with arbitrary $-\infty\leq a<b\leq \infty$ and suitable $F,J:\mathcal O\to (a,b)$, we restrict to the case $(a,b)=\mathbb{R}$ here.
\begin{lem}\label{lem-phi}
Suppose $\Phi:\mathbb{R} \to\mathbb{R}$ is a smooth and regular function in the sense of (\ref{phibddder}).
\par 1. There exists $C<\infty$ such that for all $p\in[1,\infty]$,
\begin{equation}\label{phi-lp}
\forall F\in L^p(\mathcal O), ~~\|\Phi\circ F\|_{L^p}\leq C(1+\|F\|_{L^p}).
\end{equation}
\par
2. For each integer $m\geq 0$, there exists $C<\infty$ such that
\begin{equation}\label{phi-cm}
\forall F\in C^m(\mathcal O),~~\|\Phi\circ F\|_{C^m}\leq C\left(1+\|F\|_{C^m}^m\right).
\end{equation}
\par
3. For each integer $m\geq d/2$, there exists $C<\infty$ such that for all $F\in H^m(\mathcal O)$, we have $\Phi\circ F\in H^m(\mathcal O)$ and
\begin{equation}\label{phi-sob}
\|\Phi\circ F\|_{H^m}\leq C(1+\|F\|_{H^m}^m).
\end{equation}
\par
4. There exists $C<\infty$ such that for $\kappa\in\{1,2\}$ and all $F, J\in C^\kappa(\mathcal O)$,
\begin{equation}\label{phi-neg-sob}
\left\|\Phi\circ F-\Phi\circ J\right\|_{(H^{\kappa})^*}\leq C\left\|F-J\right\|_{(H^{\kappa})^*}\left(1+ \|F\|_{C^\kappa}^\kappa\vee \|J\|_{C^\kappa}^\kappa\right).
\end{equation}
\end{lem}
The rest of this section is devoted to proving Lemma \ref{lem-phi}. To prove (\ref{phi-cm})-(\ref{phi-sob}), we need Fa\'a di Bruno's formula (a generalization of the chain rule), which classically holds for $C^m$ functions, and by the chain rule for Sobolev functions (see e.g. \cite{ziemer}, Thm. 2.1.11) also holds for $H^m$ functions.
\begin{lem}
Let $m\in\mathbb{N}$ and suppose that $F:\mathcal O\to \mathbb R$ and $\Phi:\mathbb R\to \mathbb R$ are of class $H^m(\mathcal O)$ and $C^m(\mathbb{R})$ respectively. Then for any $\alpha\in\{1,...,d\}^{m}$, the $m$-th order partial derivative of $f:=\Phi\circ F$ in direction $x_{\alpha_1}... x_{\alpha_m}$ is given by
\begin{equation}\label{faadibruno}
\frac{\partial^m f}{\partial x_{\alpha_1}...\partial x_{\alpha_m}}(x)=\sum_{\pi\in \Pi}\Phi^{(|\pi|)}(F(x))\prod_{B\in\pi}\frac{\partial ^{|B|}F}{\prod_{j\in B} \partial x_{\alpha_j}}(x),
\end{equation}
where $\pi$ runs through the set $\Pi$ of all partitions of $\{1,...,m\}$, and the $B\in\pi$ runs over all `blocks' $B$ of each partition $\pi$.
\end{lem}
\begin{proof}[Proof of (\ref{phi-lp})-(\ref{phi-cm})]
By (\ref{phibddder}), there exists a constant $c>0$ only depending on the values of $\Phi(0)$ and $\|\Phi'\|_{\infty}$ such that for all $x\in\mathbb{R}$, $|\Phi(x)|\leq c(1+|x|)$, which yields (\ref{phi-lp}). For (\ref{phi-cm}), let $\alpha\in\{1,...,d\}^{|\alpha|}$, $1\leq |\alpha|\leq m$ and let $\pi$ be a partition of $\{1,...,|\alpha|\}$. Then the corresponding summand on the right side of (\ref{faadibruno}) can be estimated by
\[ \left\| \prod_{B\in\pi}\frac{\partial ^{|B|}F}{\prod_{j\in B} \partial x_{\alpha_j}}\right\|_{\infty} \leq \|F\|_{C^m}^{|\pi|}\lesssim \left(1+\|F\|_{C^m}^{m}\right). \]
By summing the above display over all such $\alpha,\pi$ and using (\ref{phi-lp}) with $p=\infty$, we obtain (\ref{phi-cm}).
\end{proof}
To prove (\ref{phi-sob}), we also need the Gagliardo-Nirenberg interpolation inequality (see \cite{nirenberg59}, p.125) in the special case $r=q=2$.
\begin{lem}\label{lem-gns}
Suppose that $\mathcal O\subseteq \mathbb R^d$ is a bounded $C^\infty$ domain and that $i=1,...,m$, $a\in [i/m,1]$ and $p\in [1,\infty)$ satisfy
\begin{equation}\label{gns-cond}
\frac 1p=\frac 12 +\frac i d - \frac md a.
\end{equation}
Then for any $s>0$, there exist constants $C_1,C_2$ depending only on $m, d, i, a, \mathcal O$ and $s$ such that for all $F \in H^m$, we have that $D^iF\in L^p$, and
$$\|D^iF\|_{L^p}\leq C_1\|D^mF\|_{L^2}^a\|F\|_{L^2}^{1-a}+C_2\|F\|_{L^s}.$$
\end{lem}
\begin{proof}[Proof of (\ref{phi-sob})]
Let us write $f=\Phi\circ F$. By (\ref{phi-lp}), we have that $\|f\|_{L^2}\leq C(1+\|F\|_{L^2})$ whence we only need to estimate $\|D^mf\|_{L^2}$. For any $\alpha\in\{1,...,d\}^m$ we have by $(\ref{faadibruno})$ that
\begin{equation*}
\begin{split}
\left|\frac{\partial^m f}{\partial x_{\alpha_1}...\partial x_{\alpha_m}}(x)\right|^2\lesssim \sum_{\pi\in\Pi}\left| \prod_{B\in\pi}\frac{\partial ^{|B|}F}{\prod_{j\in B} \partial x_{\alpha_j}}(x)\right|^2
\end{split}
\end{equation*}
Similarly to the proof of (\ref{phi-cm}), it thus suffices to prove that for all $\alpha\in\{1,...,d\}^m$ and partition $\pi$ of $\{1,...,m\}$,
\begin{equation}\label{hm-red}
\left\|\prod_{B\in\pi}\frac{\partial ^{|B|}F}{\prod_{j\in B} \partial x_{\alpha_j}} \right\|_{L^2}\lesssim(1+\|F\|_{H^m}^m).
\end{equation}
Fix some $\pi$ for the rest of the proof. For $i=1,...,m$, define
$$\pi_i:=\left\{B\in\pi\;\middle| \;|B|=i\right\}, \qquad p_i:=\frac{2m}{i}.$$
Then we have $\sum_{i=1}^{m}i|\pi_i| = m$, and hence by H\"older's inequality
\begin{equation}\label{holder}
\begin{split}
\left\|\prod_{B\in\pi}\frac{\partial ^{|B|}F}{\prod_{j\in B} \partial x_{\alpha_j}} \right\|_{L^2}&\leq \left\|\prod_{i=1}^m|D^iF|^{|\pi_i|} \right\|_{L^2}\leq \prod_{i=1}^m\left\||D^iF|^{|\pi_i|}\right\|_{L^{p_i/|\pi_i|}}\\
&=\prod_{i=1}^m\left\|D^iF\right\|_{L^{p_i}}^{|\pi_i|}.
\end{split}
\end{equation}
Next, define
\begin{equation}\label{ai}
a_i:=\left(\frac id+\frac 12-\frac i{2m}\right)\frac dm\qquad \textnormal{for } i=1,...,m.
\end{equation}
To apply Lemma \ref{lem-gns}, we verify that for each $i=1,...,m$, $(i,a_i,p_i)$ satisfies the conditions of Lemma \ref{lem-gns}. By definition, (\ref{gns-cond}) is satisfied. Moreover, as $i\leq m$, it follows that
$$ma_i=i+\left(\frac d2-\frac{di}{2m}\right)\geq i,$$
whence we have $\frac{i}{m}\leq a_i$. Finally, we need to verify $a_i\leq 1$. For this, we note that for $i=1,...,m$, choosing $m=d/2$ in (\ref{ai}) yields $a_i=a_i(m)=1$. Moreover, for $m\geq d/2$, we have
\begin{equation}
\begin{split}
\frac{\partial a_i(m)}{\partial m}=\frac{2di-2mi-dm}{2m^3}\leq \frac{di-dm}{2m^3}\leq 0,
\end{split}
\end{equation}
so that $\alpha_i\leq 1$.
\par
Applying Lemma \ref{lem-gns} with $s=2$ to (\ref{holder}) and using that $\sum_{i=1}^m|\pi_i|\in [1,m]$ yields that
\begin{align*}
\left\|\prod_{B\in\pi}\frac{\partial ^{|B|}F}{\prod_{j\in B} \partial x_{\alpha_j}} \right\|_{L^2}&\lesssim
\prod_{i=1}^m \left(\|D^mF\|_{L^2}^{a_i}\|F\|_{L^2}^{1-a_i}+\|F\|_{L^2}\right)^{|\pi_i|}\\
&\lesssim \prod_{i=1}^m \|F\|_{H^m}^{|\pi_i|}\lesssim 1+\|F\|_{H^m}^m.
\end{align*}
\end{proof}
\begin{proof}[Proof of (\ref{phi-neg-sob})]
1. Let $\kappa\in \{1,2\}$ and fix $F,J\in C^\kappa(\mathcal O)$. Define the function
\[\omega:\mathcal O\to \mathbb R,\qquad \omega(x):=
\begin{cases}
\frac{\Phi(F(x))-\Phi(J(x))}{F(x)-J(x)} \quad \text{if} \;\; x\in\{F\neq J\}\;\\
\Phi'(F(x))\quad \text{if} \;\; x\in \{F= J\}.
\end{cases} \]
Then we have, using also (\ref{c-h-mult}), that
\begin{equation*}
\begin{split}
\|\Phi\circ F-\Phi\circ J\|_{(H^{\kappa})^*}&=\sup_{\varphi\in C^\infty(\mathcal O), \; \|\varphi\|_{H^\kappa}\leq 1}\left|\int_{\mathcal O} \varphi(\Phi\circ F-\Phi\circ J)\mathbbm 1_{\{F\neq J\}} \right|\\
&=\sup_{\varphi\in C^\infty(\mathcal O), \; \|\varphi\|_{H^\kappa}\leq 1}\left|\int_{\mathcal O} (F-J) \varphi \omega \right|\\
&\leq \|F-J\|_{(H^{\kappa})^*}\sup_{\varphi\in C^\infty(\mathcal O), \; \|\varphi\|_{H^\kappa}\leq 1}\left\|\varphi \omega\right\|_{H^k}\\
&\lesssim \|F-J\|_{(H^{\kappa})^*}\left\|\omega\right\|_{C^\kappa}.
\end{split}
\end{equation*}
\par
2. Thus it suffices to prove that $\left\|\omega\right\|_{C^\kappa}\leq C (1 + \|F\|_{C^\kappa}^\kappa\vee \|J\|_{C^\kappa}^\kappa)$ for some $C>0$ independent of $F$ and $J$. Writing $\omega=\psi\circ \phi$, where
\[\phi:\mathcal O\to \mathbb R^2, \qquad \phi(z)=\left(F(z),J(z)\right),\]
\[\psi:\mathbb R^2\to (0,\infty), \qquad \psi(x,y)=
\begin{cases}
\frac{\Phi(x)-\Phi(y)}{x-y} \quad \text{if} \;\; x\neq y\\
\Phi'(x)\quad \text{if} \;\; x=y,
\end{cases}\]
we see by the multivariate chain rule that it suffices to show that $\psi$ is $\kappa$-times continuously differentiable with bounded derivatives, and we achieve this by showing that the partial derivatives of $\psi$ of order $\kappa$ exist and are continuous throughout $\mathbb R^2$.
\par
3. We will repeatedly use the following basic fact: Let $h:\mathbb R\to \mathbb R$ be continuous and continuously differentiable on $\mathbb R\setminus \{0\}$. If $h'$ has a continuous extension $g$ to $\mathbb R$ with some value $g(0)=\xi$, then $h\in C^1(\mathbb R)$ with $h'(0)=\xi$.
\par
4. Clearly, $\psi$ is smooth on $\mathbb R^2\setminus\{x=y\}$. For $k\geq 0$ and $x,y\in \mathbb R$, we denote the remainder of the $k$-th order Taylor expansion by
$$R_{k,x}(y):=\Phi(y)-\sum_{j=0}^{k}\frac{\Phi^{(j)}(x)}{j!}(y-x)^j.$$
For $x\neq y$, we have $\psi(x,y)=\frac{R_{0,x}(y)}{y-x}$ and also, by induction
\begin{equation}\label{partder}
\partial_1^k\psi(x,y)=\frac{k!R_{k,x}(y)}{(y-x)^{k+1}}, \quad k\geq 0,
\end{equation}
where $\partial_1$ denotes the partial derivative with respect to $x$. By the mean value form of the remainder, we know that $R_{k,x}(y)=\frac{\Phi^{(k+1)}(\xi)}{(k+1)!}(y-x)^{k+1}$ for some $\xi$ between $x$ and $y$. Thus we can continuously extend $\partial_1^k\psi$ to $\{x=y\}$ by
$$\partial_1^k\psi(x,x)=\frac{\Phi^{(k+1)}(x)}{k+1}.$$
It follows that the partial derivatives with respect to $x$ of all orders exist and are continuous on $\mathbb R^2$. The same holds for the partial derivatives with respect to $y$, by symmetry, concluding the proof of the case $\kappa=1$. The case $\kappa=2$ follows by adapting the previous arguments for mixed partial derivative $\partial_1\partial_2$ and is left to the reader.
\end{proof}
\section{Proof of Theorem \ref{thm-gen}, Part 1} \label{sec-ex-pf}
Let $\lambda,\varepsilon>0$ be fixed throughout and let us write $\mathscr J=\mathscr J_{\lambda, \varepsilon}$. We denote by $\mathcal T_w=\mathcal T_{w,\alpha}$ the weak topology on $\mathcal H$ (recall $\mathcal H=H^\alpha(\mathcal O)$ if $\kappa<1/2$ and $\mathcal H=H^\alpha_c(\mathcal O)$ if $\kappa\geq 1/2$), i.e. the coarsest topology with respect to which all bounded linear functionals $L:\mathcal H\to \mathbb R$ are continuous. We also denote the subspace topology on subsets of $\mathcal H$ by $\mathcal T_w$. On any closed ball $\mathcal H(R):=\{F\in \mathcal H:\|F\|_{H^\alpha} \le R\}$, this topology is metrisable by some metric $d$, see e.g. Theorem 2.6.23 in \cite{megg}.
\paragraph{Step 1: Localisation} In Lemma \ref{lem-main}, by assumption on $\alpha$, we have that $\Psi_*(\lambda,R)/R^2\xrightarrow{R\to\infty}0$ and so there exists $\delta>0$ such that for all $R\geq \delta$, we have that $R^2\geq c_1\varepsilon\Psi_*(\lambda,R)$, where $c_1$ is the constant from (\ref{delta-req}). Thus, applying Theorem \ref{thm-ep}, we have that the events
\[A_j:=\left\{\mathscr J\text{ has a maximizer } \hat F\notin \mathcal V\cap \mathcal H(2^j) \right\}\]
satisfy $\P(A_j)\xrightarrow{j\to\infty} 0$, whence choosing $j\in\mathbb{N}$ large enough ensures that
$$\sup_{F\in\mathcal V\cap H^{\alpha}(2^j)}\mathscr J(F)=\sup_{F\in\mathcal V}\mathscr J(F)$$
holds with probability as close to one as desired.
\paragraph{Step 2: Local existence via direct method}
By the previous step, it suffices to show that for any $j\in\mathbb{N}$, $\mathscr J$ almost surely has a maximizer over $\mathcal V\cap \mathcal H(2^j)$. We fix some $j\in\mathbb{N}$. As $\mathcal V$ is weakly closed and $\mathcal H(2^j)$ is weakly sequentially compact by the Banach-Alaoglu Theorem, it follows that any sequence $F_n\in \mathcal V\cap \mathcal H(2^j)$ has a weakly convergent subsequence $F_n\to F$ with weak limit $F\in\mathcal V\cap \mathcal H(2^j)$. Moreover, we claim that $-\mathscr J:\mathcal V\cap \mathcal H(2^j)\to \mathbb{R}$ is lower semicontinuous with respect to $\mathcal T_w$. To see this, we decompose $-\mathscr J$ as
\[-\mathscr J(F)=-2\langle Y,\mathscr G(F)\rangle_\mathbb H+\|\mathscr G(F)\|_{\mathbb H}^2+\lambda^2\|F\|_{H^\alpha}^2=:I+II+III.\]
The term $I$ is, almost surely under $\mathbb P_{F_0}^\varepsilon$, continuous w.r.t. $\mathcal T_w$ by Lemma \ref{lem-Ycont}, $II$ is continuous w.r.t. $\mathcal T_w$ by Lemma \ref{lem-IIcont} and $III$ is lower semicontinuous by a standard fact from functional analysis. Thus the existence of minimisers follows from the direct method of the calculus of variations.
The next three lemmas are needed to prove lower semicontinuity of $-\mathscr J$.
\begin{lem}\label{lem-weakcon}
Let $\alpha>0$ and let $\left(F_n: n\in\mathbb N\right)\subseteq \mathcal H$, for $\mathcal H=H^\alpha$ or $H^\alpha_c$, be a sequence such that $F_n\to F$ for $\mathcal T_w$. Then also $F_n\to F$ in $L^2$.
\end{lem}
\begin{proof}
It suffices to show that for any subsequence $(F_{n_j}:j\in\mathbb N)$, there exists a further subsequence $(F_{n_{j'}}:j'\in\mathbb N)$ such that $F_{n_{j'}}\to F$ in $L^2$. By the uniform boundedness principle, there exists $R>0$ such that for all $n\in\mathbb N$, $\|F_n\|_{H^\alpha}\leq R$. By the Rellich-Kondrashov compactness theorem, the closed ball $\mathcal H(R)$ is pre-compact with respect to $L^2$ topology, hence for any subsequence $(F_{n_j})$ of $(F_n)$, there exists a further convergent subsequence $(F_{n_{j'}})$ with limit $\tilde F$ in $L^2$. In particular, we have $F_n\to F$ weakly in $L^2$ and $F_{n_{j'}}\to \tilde F$ in $L^2$, so that by the uniqueness of weak limits, we have $\tilde F=F$ as elements in $L^2$, and therefore $\tilde F=F$ a.e. in $\mathcal O$ and $F_n\to F$ in $L^2$.
\end{proof}
\begin{lem}\label{lem-IIcont}
Let $\alpha>0$, $\kappa, \gamma\in \mathbb{R}_+$ and $\mathcal V_0\subseteq \mathcal V$ be a bounded subset of $\mathcal H=H^\alpha$ or $H^\alpha_c$. If a map $\mathscr G: \mathcal V\to \mathbb H$ is $(\kappa,\gamma,\alpha)$-regular, then it is continuous as a mapping from $(\mathcal V_0,d)$ to $\mathbb H$.
\end{lem}
\begin{proof}
Take any $F_n,F\in \mathcal V_0$ such that $F_n\to F$ for $\mathcal T_w$ and note that $\|F_n\|_{H^\alpha} \le R$ for some $R>0$. By Lemma \ref{lem-weakcon} we have $\|F_n - F\|_{L^2}\to 0$ and by (\ref{entrcond}) and the continuous imbedding $L^2 \subseteq (H^\kappa)^*, \kappa \ge 0$, we obtain
\begin{equation}
\|\mathscr G(F_n)-\mathscr G(F)\|_{\mathbb H}\leq C\left(1+R^\gamma\right) \|F_n-F\|_{L^2} \xrightarrow{n\to\infty} 0.
\end{equation}
\end{proof}
We finally establish a continuity result for the Gaussian process $Y^{(\varepsilon)}$.
\begin{lem}\label{lem-Ycont}
Suppose that $Y^{(\varepsilon)}$ and $\mathscr G$ are as in Theorem \ref{thm-gen}. Then there exists a version of the Gaussian white noise process $\mathbb{W}$ in $\mathbb H$ such that for all $R>0$, the map (between metric spaces)
\[ \Psi: (\mathcal V\cap \mathcal H(R),d)\to \mathbb R, \qquad F\mapsto \langle Y^{(\varepsilon)}, \mathscr G(F)\rangle_{\mathbb H}\]
is almost surely uniformly continuous.
\end{lem}
\begin{proof}
For any $\delta>0$, define the modulus of continuity \[M_\delta :=\sup_{F,H\in \mathcal V\cap \mathcal H(R),\;\; d(F,H)\leq \delta}\left|\langle Y^{(\varepsilon)},\mathscr G(F)-\mathscr G(H)\rangle_\mathbb H \right|,\] a random variable.
Moreover, we define the set
\[A:=\left\{\omega\in\Omega\;\middle|\; M_\delta\xrightarrow{\delta\to 0}0 \right\},\] where $\Omega$ is a probability space supporting the law $\mathbb P$ of $\mathbb W$. It is sufficient to show that $\mathbb P(A)=1$, and noting that $M_\delta$ is decreasing in $\delta$, it hence suffices to prove $\mathbb E\left[M_\delta \right]\xrightarrow{\delta\to 0}0$. To see this, similarly to the proof of Lemma \ref{thm-ep}, we apply Dudley's theorem (see \cite{nicklgine}, Theorem 2.3.7) to the Gaussian process
\[\left(\mathcal \mathbb{W}(\psi):\;\psi\in \mathcal D_R\right), \qquad \mathcal D_R:=\left\{\mathscr G(F)\;\middle|\; F\in \mathcal V\cap \mathcal H(R)\right\}. \]
For any $\delta>0$, define
\[R_\delta:=\sup_{F,H\in \mathcal V\cap \mathcal H(R),\;\; d(F,H)\leq \delta}\|\mathscr G(F)-\mathscr G(H)\|_{\mathbb H}.\]
By Lemma \ref{lem-IIcont}, we know that $\mathscr G$ is continuous as a mapping from $(\mathcal V\cap \mathcal H(R), d)$ to $\mathbb H$. As $(\mathcal V\cap \mathcal H(R), d)$ is a compact metric space, $\mathscr G$ is in fact uniformly continuous, so we have that $R_\delta\xrightarrow{\delta\to 0} 0$. By the same argument as in the proof of Lemma \ref{lem-main} (but choosing here $m:=(1+R^\gamma)$) we can use (\ref{entrcond}) to obtain
\[H(\rho, \mathcal D_R, \|\cdot\|_{\mathbb H})\lesssim \left(\frac{Rm}{\rho}\right)^{\frac{d}{(\alpha+\kappa)}}, ~~~\rho>0, \]
whence by Dudley's theorem, the modulus of continuity is controlled by
\begin{equation*}
\begin{split}
\mathbb E\left[M_\delta\right]&\leq \mathbb E\left[\sup_{\psi,\varphi\in \mathcal D_R,\; \|\psi-\varphi\|_{\mathbb H}\leq R_\delta}\left|\langle\mathbb{W},\psi-\varphi\rangle_\mathbb H\right|\right] \lesssim \int_0^{R_\delta}\left(\frac{Rm}{\rho}\right)^{\frac{d}{2(\alpha+\kappa)} }d\rho,
\end{split}
\end{equation*}
which converges to zero as $\delta \to 0$ since $\alpha>d/2-\kappa$.
\end{proof}
\section*{Acknowledgments}
The authors are grateful to two anonymous referees and an associate editor for their remarks and suggestions. RN thanks Francois Monard and Gabriel P. Paternain for helpful discussions.
|
\section{Introduction}
Combustion instabilities have long been a hurdle in the development of modern rocket engines. These instabilities are characterized by the coupling between acoustics, hydrodynamics and heat release. In propulsion systems, the triggering of combustion instabilities can lead to catastrophic engine failures, and the underlying physical mechanisms are sensitive to many parameters including the geometry and fuel-to-oxidizer ratio. Therefore, a significant number of geometric configurations and operating conditions have to be evaluated and analyzed in rocket engine design. Though the advancement of modern computational technology has enabled routine modeling of laboratory-scale rockets ~\cite{harvazinski2015coupling} and other experimental configurations such as atmospheric combustors~\cite{matsuyama2016large}, direct numerical simulations (DNS)~\cite{domingo2005dns} and large eddy simulations (LES)~\cite{ihme2012large,huang2014computational,harvazinski2015coupling,lacaze2009large,hernandez2011laboratory,srinivasan2015flame} continue to be prohibitively expensive for the simulation of full-scale engines~\cite{urbano2016exploration,staffelbach2009large,wolf2012using}. This motivates research on alternative approaches such as reduced-fidelity modeling \cite{baukal2000computational,frezzotti2014low, sirignano2013two, popov2016transverse,you2005generalized} and reduced order modeling~\cite{lieu2006reduced,lucia2004reduced}.
Reduced-fidelity modeling approaches usually adapt certain levels of simplifications in physical and numerical models to achieve higher computational efficiency.
Sirignano and Popov \cite{sirignano2013two} developed a two-dimensional model for transverse-mode combustion instability in a cylindrical rocket motor and further extended it to rectangular configurations~\cite{popov2016transverse}. Though reduced-fidelity modeling approaches can be efficiently used for engine design with satisfying accuracy for critical quantities of interest, many challenges are encountered in accurately predicting crucial combustion dynamics that can trigger instability.
In addition to approaches that model combustion dynamics by solving transport equations, empirical approaches have been pursued. These approaches formulate the combustion response as a function of well-characterized flow variables such as pressure and velocity. This class of techniques include the so-called flame transfer (FTF)~\cite{durox2009experimental} and flame describing functions (FDF)~\cite{noiray2008unified}. In Ref. \cite{you2005generalized}, more factors such as variations in geometry are taken into account and the response function is successfully applied to a swirl-stabilized combustor. Though FTF/FDF approaches impose a simple model form for the flame dynamics that can be easily derived from both experimental and computational data, their applications are usually restricted to combustion system with low-amplitude perturbations with regards to the mean flow due to the fact that the derivation of the flame model is based on linear assumptions in frequency domain. Therefore, to model highly nonlinear flame dynamics which is typical in rocket engines, the validity of the FTF/FDF approach remains to be evaluated.
In contrast, reduced order modeling (ROM) approaches pursue mathematically formal reductions, offering the promise of generalization. Projection-based ROMs~\cite{chatterjee2000introduction,lucia2003projection} have recently been applied to combustion instability problems~\cite{xu2017reduced,xu2018multi,huang2018exploration,huang2016multi,huang2017multi}.
In projection-based ROMs, the full order model is projected onto a set of informative basis functions using proper orthogonal decomposition (POD)-Galerkin method~\cite{rowley2004model}, resulting in a low dimensional set of equations that can retain most of the modeling fidelity. Thus, the need for commonly used empirical modeling strategies such as combustion response functions can be eliminated. Preliminary explorations of projection-based ROMs on representative combustion model problems can be found in Refs. \cite{xu2017reduced,huang2018exploration,huang2016multi,huang2017multi}. It should be pointed out that POD-based ROMs have been more widely used in non-reacting flow problems such as aerodynamics~\cite{ravindran2000reduced, rowley2004model}, aeroelasticity~\cite{amsallem2008interpolation,lieu2007adaptation} and flow control~\cite{barbagallo2012closed,barbagallo2011input}.
However, applications of ROMs to rocket engines involve the following challenges:
\begin{itemize}
\item A single high fidelity simulation or Full Order Model (FOM) of a full engine may be prohibitively expensive.
\item Even if full engine FOM simulations are affordable, conventional ROM construction is based on FOM simulations for different injector configurations. If there is any change to the geometry, the ROM has to be re-generated or re-parameterized, which restricts the scope of ROM applications.
\item POD-ROMs may be incapable of providing predictions beyond the training interval and parameters.
\item
Since POD basis construction involves global minimization, the resulting basis will be dominated by high-energy regions. The dynamics in low-magnitude regions can be under-resolved and consequently affect stability and robustness.
\end{itemize}
The present work extends the multi-fidelity framework methodology developed by Huang et al.~\cite{huang2016multi,huang2017multi,huang2019multifidelity} and assesses ROM performance on the major aspects listed above based on a quasi-1D model combustor problem, which was originally developed by Smith et al.~\cite{smith2008computational} to model the experimental work by Yu et al. on the single-element combustor~\cite{yu2012spontaneous,yu2009experimental}. It should be mentioned that additional challenges exist in ROMs of more realistic problems, where complex physics such as three dimensional turbulent combustion, vortex shedding, and flame-vortex interactions are present. While progress is being made in addressing these issues~\cite{huang2018exploration,huang2019investigations,huang2018challenges}, the present work focuses on the issue of ROM development in problems where the full order solution is not available for the entire domain. The approach itself is generic for problems of different levels of complexity. The framework pursues a multi-domain approach where the computational domain is divided into two components: the upstream one covering the critical physics in the heat addition/acoustic-flame interaction region and the downstream part dominated by flow and acoustics dynamics. Reduced order models are used to model the critical dynamics in the upstream component. The ROM is developed and trained based on FOM solutions of the truncated domain corresponding to the upstream component with a perturbed characteristic downstream boundary condition. This training strategy is designed to incorporate rich information in the resulting ROM and proves to be effective in enhancing predictive capabilities. The truncated domain ROM is then integrated within a multi-fidelity solver~\footnote{For the current demonstration, we replace the reduced-fidelity part with the full order model for consistency of evaluations}. The multi-fidelity model is then compared to ROMs based on the FOM solutions of the entire computational domain, which will be referred as conventional ROMs in the current paper.
The rest of this paper is organized as follows: The quasi-1D combustor setup and formulations for the FOM and ROM are described in Sec. \ref{sec formulation}. In Sec. \ref{sec framework}, the present framework and ROM training with characteristic boundary are introduced. Numerical experiments are presented in Sec. \ref{sec result_domain}. Concluding remarks are given in Sec. \ref{sec conclusion}
\section{Formulation}\label{sec formulation}
\input{solver.tex}
\input{rom.tex}
\section{Framework Overview}\label{sec framework}
In this section, we explore characteristic-based ROM training on a truncated domain and contrast it with conventional ROM development.
A schematic is given in Fig. \ref{fig frameworkschematic}. The major steps include
\begin{enumerate}
\item Truncate the domain into two sub-domains, the first containing the physics-complex area that covers the fuel injection and the flame, and the second containing the variable-length chamber dominated by acoustics. The ROM will be trained for the first sub-domain.
\item Perform a FOM training simulation on the first sub-domain with a broadband perturbation added on the truncation interface, which is treated as a characteristic boundary.
\item Generate the POD bases for different variables using SVD of the solution from the training simulation.
\item Use the POD bases in a ROM solver for the first sub-domain and couple it with a FOM solver for the variable-length, acoustics-dominated chamber for predictions.
\end{enumerate}
It can be noted that in the framework ROM is only developed for the first sub-domain, whereas the second sub-domain is solved in FOM. There are two reasons for this choice:
\begin{enumerate}
\item The complex, computation-intensive area is fully covered in the first sub-domain. The accurate modeling of this domain requires high-resolution simulations. The FOM computation in the second sub-domain is expected to be much less challenging than the first domain and affordable with coarser resolution modeling approaches (e.g. coarse-mesh LES and URANS), which makes a ROM replacement unnecessary.
\item In design evaluations, the chamber length in the second sub-domain is variable, thus a new ROM training is required for each chamber length, which violates the goal of reducing the computing cost using ROM.
\end{enumerate}
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{frameworkschematic.png}
\caption {Schematic of framework}\label{fig frameworkschematic}
\end{center}
\end{figure}
\subsection{Characteristic ROM training on reduced-domain}
In traditional ROM formulations, POD bases are generated using FOM solutions of the target full domain. In the proposed approach, the domain is split into two sub-domains with the interface at $x=0.096$ m. The first sub-domain includes the injector, back-step and the leading part of the chamber where the flame is located, and the second includes the rest of the geometry, which is variable in design evaluations.
To obtain a basis that is representative of the physics in the first sub-domain, a FOM simulation is performed in a reduced domain, treating the truncation interface as a characteristic outlet boundary, which helps to eliminate the undesirable resonant acoustic modes corresponding to the reduced-domain geometry. As demonstrated by Huang et al.~\cite{huang2017multi}, the presence of such resonant acoustic modes can significantly affect the predictive capabilities of the ROMs. A similar training procedure as in Refs. ~\cite{huang2016multi,huang2017multi} is used. The training FOM simulation is also started from the steady state as in the self-excited simulations on the full domain (Sec. \ref{sec formulation}). The difference is that instead of a single frequency inlet perturbation, a broadband perturbation is imposed on the incoming characteristic wave at the truncation boundary to cover the range of responses created by the resonant frequencies corresponding to different chamber lengths. This perturbation is imposed over the entire duration of the FOM simulation.
To maintain consistency, the rest of the specified properties at the boundary are as follows:
\begin{eqnarray}
\Omega_{bc} = \{J, u, T, Y_{ox}\}^T,
\end{eqnarray}
where $c$ is the speed of sound, $J = -\frac{p}{\overline{\rho c}} + u$ is one-dimensional approximation of the characteristic invariant for the in-coming acoustic wave, $\overline{\rho c}$ is obtained from the steady state solution, and the other three primitive variables are extrapolated from the values of the interior cells.
\subsection{Multi-domain coupling}
Following the aforementioned simulations, the bases $\mathbf{V}$ for the conserved variables are obtained from the reduced domain solution snapshots, and used in POD-Galerkin projection (Eq. \eqref{eq ROM1}) to derive the ROM for the first sub-domain. The second sub-domain is computed using the full order equations \eqref{eq fom}. Due to the hybrid integration, chamber lengths can be changed without any further modifications to the ROM. The two sub-domains communicate at every time-step by exchanging interface conditions to be the value at the neighboring cell of the adjacent domain. Similar to the interior of the computing domain, Roe's upwind flux \cite{roe1986characteristic} is used at the interface. Using subscript \RNum{1} and \RNum{2} for the first and second sub-domain, and superscript $(1)$ and $(end)$ for the first and last cell in the corresponding sub-domain respectively, the flux at the interface is given by
\begin{eqnarray}
{\bf{\hat f}} = \frac{1}{2}\left( {\bf{f}}^{(end)}_{\text{\RNum{1}}} + {{\bf{f}}^{(1)}_{\text{\RNum{2}}}} \right) - \frac{1}{2}\left| {\frac{{\partial {\bf{f}}}}{{\partial {\bf{q}}}}\left( {{{\bf{q}}^*}} \right)} \right|\left( {{{\bf{q}}^{(1)}_{\text{\RNum{2}}}} - {{\bf{q}}^{(end)}_{\text{\RNum{1}}}}} \right),
\end{eqnarray}
where $\mathbf{q}^*$ is the Roe-averaged state calculated from $\mathbf{q}^{(end)}_{\text{\RNum{1}}}$ and $\mathbf{q}^{(1)}_{\text{\RNum{2}}}$. $\mathbf{q}^{(end)}_{\text{\RNum{1}}}$ is computed from the last row of the basis and the reduced variable, $\mathbf{q}^{(end)}_{\text{\RNum{1}}} = \mathbf{V}^{(end)}\mathbf{q}_r$.
\subsection{Control groups for comparison}
The methodology detailed above is compared to the following control groups using the conventional ROM approach:
\begin{description}
\item[Control group A] also uses a hybrid multi-domain solver, i.e. the first sub-domain is solved using the ROM and the second solved using the FOM. The difference from the proposed framework is in the training data generation and collection stage. In control group A, for each combination of chamber length and $\alpha$, a full-domain FOM simulation is performed instead of the proposed characteristic training on the truncated-domain. Then the solution is restricted to the first sub-domain and collected to generate the POD bases for the ROM of the sub-domain.
\item[Control group B] uses a traditional full-domain ROM. The same FOM training simulations as in control group A are used and the POD bases are directly generated on the full-domain solution.
\end{description}
To summarize, let $n_{L_c}$ and $n_{\alpha}$ be the number of chamber lengths and amplification factors studied, respectively. In the proposed framework, $n_{\alpha}$ FOMs simulated on the reduced domain are used in total~\footnote{although as shown in the next section, this can be reduced} using characteristic perturbations. Then the first sub-domain is simulated using the ROM, second sub-domain using the FOM. In control groups A\&B, $n_{L_c}\times n_{\alpha}$ self-excited FOM simulations are conducted. In group A, the first sub-domain is solved using a ROM, and the second is solved using the FOM. In group B, the whole domain is solved using a ROM.
For conciseness, the proposed framework will be referred to as ``reduced-domain training and multi-domain solver (RD-MD)", control group A as ``full-domain training and multi-domain solver (FD-MD)", and control group B as ``full-domain training and full-domain solver (FD-FD)". A schematic of the different methods is given in Fig. \ref{fig flowchart}.
\begin{figure}
\begin{center}
\includegraphics[width=1\textwidth]{flowchart.png}
\caption {Schematic of training (outlined in black) and prediction (outlined in blue) stages, curves indicating perturbation added at the boundary}\label{fig flowchart}
\end{center}
\end{figure}
\section{Results}\label{sec result_domain}
Numerical tests are conducted for chamber lengths $L_c$ ranging from 0.254 m to 0.762 m with an interval of $\Delta L_c=0.0635$ m and at $\alpha = 3.1,3.25$ and $3.4$. In the proposed training method, a FOM simulation is performed in reduced domain for each $\alpha$.
From \textit{a priori}~\cite{grenda1995application} analysis, the longitudinal frequency of the chamber spans approximately between 700 and 2000 Hz, and thus the broadband perturbation signal is set to
\begin{eqnarray}\begin{aligned}\label{eq train signal}
J'\left( t \right) = 0.01{J_0}\sum\limits_{i_f = 1}^{n_f} {\sin \left( {2\pi (f_0+(k-1)\Delta{f})t} \right)},
\end{aligned}\end{eqnarray}
which is a superimposition of $n_f=14$ frequencies, starting from $f_0 = 700\text{ Hz}$ with an incremental interval $\Delta f = 100\text{ Hz}$.
To train the control groups, full-domain self-excited FOM simulations are conducted for each combination of $L_c$ and $\alpha$.
In all three methods, snapshots of the training solution are collected every 100 time-steps over a period $t=0 - 0.05$ s, with the initial steady state set at $t = 0$. Once the basis $\mathbf{V}$ is obtained using the SVD of the corresponding snapshot, ROM simulations are conducted over $t=0 -0.1$ s.
\input{fom.tex}
\input{basis.tex}
\input{sanitytest.tex}
\input{solution.tex}
\input{off_design.tex}
\section{Conclusions}\label{sec conclusion}
In this work, a multi-fidelity ROM development framework is investigated on a quasi-1D version of the CVRC model rocket combustor. An $\alpha-\tau$ model that couples the pressure oscillation and heat release is used to represent the unstable behavior of the combustor. In the proposed framework, the domain is split into two sub-domains with the first sub-domain containing the main reaction dynamics and the second covering the adjustable chamber of the CVRC. Characteristic training is conducted in the first sub-domain, resulting in a ROM that can be directly integrated with a FOM solver of the second sub-domain with any length.
Numerical tests are conducted at different chamber lengths $L_c$ and amplification factors $\alpha$. The framework is compared against the FOM solution, and other ROM approaches where traditional training approaches are taken. A major advantage of the current framework is that it significantly reduces the number of FOM simulations required to train ROMs while the traditional method requires a separate FOM simulation for each individual target chamber lengths, which does not fit in the needs for more efficient rocket engine design. Moreover, the proposed method shows faster decay of the singular value spectrum at medium-to-high chamber lengths, which implies better basis quality and ROM reliability at a low number of modes. The advantage is especially distinct in the tests with $k=20$, where the proposed method shows significantly improved stability.
In predictive tests at conditions outside the training set, the framework showed significant improvement in both stability and accuracy over the traditional methods, especially when the instability is more pronounced at high $\alpha$, and when number of modes is low. Comparison of L2 error, growth rate and LCO peak-to-peak amplitude shows that the framework is able to predict the these quantities accurately at all combinations of $\alpha$ and $L_c$ at $k=100$.
In summary, the multi-fidelity framework proves to be promising approach for modeling rocket combustion instability. The modularity of the framework will be useful in efficient training and integration of components such as injector elements. The flexibility of the multi-domain approach also enables potential developments such as the application of different time-marching schemes, basis sizes and acceleration methods in different parts of the computational domain to achieve a better balance between efficiency and reliability in the future ROM development.
While the results are encouraging, it should be recognized that the flow and combustion models used in this study are highly simplified and further studies are required on more complex flow problems to evaluate the capabilities of the multi-fidelity framework.
\subsection{Modal decomposition}
\subsubsection{Singular values}
The singular values from a POD of snapshots of density, $\rho$ from different training methods are compared at $\alpha = 3.25, L_c = 0.254,0.508,0.762$ m. There are 472 cells in the first sub-domain, 897 in the full domain at $L_c = 0.254$ m, 1405 at 0.508 m and 1913 at 0.762 m. Since the number of snapshots is larger than the mesh size, the maximum number of modes and singular values from SVD is limited to the mesh size.
The result is visualized by the complementary part of the cumulative energy in Fig. \ref{fig spectrum}. For $k$ modes, the cumulative energy is defined as
\begin{eqnarray}
\eta_{k} = \frac{{\sum\limits_{i = 1}^{k} {{\sigma _i}} }}{{\sum\limits_{i = 1}^n {{\sigma _i}} }}.
\end{eqnarray}
It should be noted that there is only one curve for RD-MD as it uses the same training simulation and POD bases for all chamber lengths. It can be observed that for both FD-MD and FD-FD, the decay in the complementary part increases as the chamber length decreases. This is due to the fact that the dynamics have a higher frequency when the chamber length is shorter, which is easier to be captured by fewer modes in SVD. In all cases, the decay in FD-FD is slower than FD-MD, because it covers a larger domain with more spatial variations in geometry and physics, and more modes are therefore needed to contain the same portion of information stored in the training data. For medium-to-high $L_c$, the proposed framework gives the best decay since its training data contains more high-frequency contents. At small $L_c$, the decay in the leading modes for the two full-domain training methods become slightly better than RD-MD as the response frequency in their training data is comparable to the highest one included in the reduced-domain training, whereas there are also lower frequencies contained in the latter. Results for other variables and $\alpha$ follow a similar pattern.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{eta.jpg}
\caption{Complementary part of cumulative energy in the singular values for $\rho$ at $\alpha = 3.25$}\label{fig spectrum}
\end{figure}
\subsubsection{Mode shapes}
Comparison of just the singular values does not provide enough information on the compatibility of the basis and physical features, since each training method is different. To provide further insight, the first five most energetic modes for $\rho$ at $L_c = 0.508$ m, $\alpha = 3.25$ are shown in Fig. \ref{fig mode}. It should be noted that the basis for FD-FD is on the full domain and only truncated in the plot to the first sub-domain for comparison purposes. The first mode for all methods appears to be identical. For lower energy modes, the basis for the RD-MD is quite different from the other two methods. The basis for the FD-MD and FD-FD are also different for $k>4$, which suggests that performing POD on different domains can result in different mode shapes, even when the training data is collected from the same simulation. This is because that the SVD process for POD bases generation solves a global minimization problem that will yield a basis with minimal least square projection error for the input snapshots. Thus, the scope of the minimization problem is changed with the size of the domain. There is hence no guarantee that the same local mode shapes will be obtained, though the input snapshots are the same locally. Consistent observation has been made in Ref.~\cite{huang2016analysis}, where the mode shape in full domain and different sub-domains of a 3D simulation of the CVRC is compared (equivalent to the FD-MD and FD-FD comparison).
To better demonstrate the evolution of the modes, the FOM solution at $\alpha=3.25, L_c=0.508$ m is projected onto the POD bases. Due to orthogonality, the projection can be performed independently and the resulting reduced variables are given by
\begin{eqnarray}
q_r^{(i)}(t)=({\mathbf{v}}^{(i)})^T\mathbf{q},
\end{eqnarray}
where ${\mathbf{v}}^{(i)}$ is the $i$-th POD mode.
The coefficients for the five leading modes present in Fig.~\ref{fig mode} are shown in Fig. \ref{fig coefficient}. For better visualization, only a period of $t=0.3-0.4$ s is used. It can be seen that the first mode evolves at a single frequency corresponding to the large-scale instability characterized by the pressure oscillation described in the previous section. This frequency is exhibited by the latter modes, while there are also additional frequencies corresponding to the high-frequency dynamics contained in the modes.
It can be noticed from Fig.~\ref{fig mode} that the spatial modes that exhibit high-frequency temporal oscillations also have a higher spatial frequency. To make this trend clearer, the dominant spatial frequencies of the 50 leading modes of $\rho$ are plotted in Fig.~\ref{fig dominant_freq}. While showing a similar trend, the growth rates are different between the methods. The reasons are two-fold: 1) using the same single-frequency training, FD-MD shows a faster decay in the singular values, which indicates that the low-frequency high-energy dynamics is better captured in the leading modes, and thus the high-frequency modes are realized at smaller mode numbers than in FD-FD. 2) There are multiple frequencies in the reduced-domain training. These frequencies are still significantly higher than those of the corresponding POD mode numbers in FD-FD and FD-MD. The corresponding modes representing these frequencies cause the RD-MD to have the slowest increase in the dominant spatial frequencies. Also presented in Fig.~\ref{fig dominant_freq} are the dominant temporal frequencies exhibited in the evolution of the POD modes. For RD-MD, the frequency increases with mode number steadily before the 35-th mode. This demonstrate a good correlation between the temporal and spatial frequency captured in the modes. In contrast, many high-spatial-frequency modes from the two full-domain training methods actually evolve at a low temporal frequency. This inconsistency indicates a less efficient basis compared to RD-MD.
\begin{figure}
\centering
\subfloat[RD-MD]{
\includegraphics[width=0.33\textwidth]{mode_RDMD_rho.jpg}}
\subfloat[FD-MD]{
\includegraphics[width=0.33\textwidth]{mode_FDMD_rho.jpg}}
\subfloat[FD-FD]{
\includegraphics[width=0.33\textwidth]{mode_FDFD_rho.jpg}}
\caption{Leading spatial modes for $\rho$ at $\alpha=3.25, L_c=0.508$ m}\label{fig mode}
\end{figure}
\begin{figure}
\centering
\subfloat[RD-MD]{
\includegraphics[width=0.33\textwidth]{coefficient_RDMD_rho.jpg}}
\subfloat[FD-MD]{
\includegraphics[width=0.33\textwidth]{coefficient_FDMD_rho.jpg}}
\subfloat[FD-FD]{
\includegraphics[width=0.33\textwidth]{coefficient_FDFD_rho.jpg}}
\caption{Evolution of coefficients $\mathbf{q}_r^\rho$ at $\alpha=3.25, L_c=0.508$ m}\label{fig coefficient}
\end{figure}
\begin{figure}
\centering
\subfloat[Spatial]{
\includegraphics[width=.45\textwidth]{dominant_spatial.jpg}}
\subfloat[Temporal]{
\includegraphics[width=.45\textwidth]{dominant_temporal.jpg}}
\caption{Dominant spatial/temporal frequency of POD modes for $\rho$ at $\alpha=3.25, L_c=0.508$ m.
\\Solid line: first mode, shaded: band for the first five frequencies}\label{fig dominant_freq}
\end{figure}
\subsubsection{Projection error}
For a given set of $k$ POD modes, the projection error is defined as
\begin{eqnarray}
\epsilon = \frac{{{{\left\| {{\bf{Q}} - {{\bf{V}}_{1:k}}{\bf{V}}_{1:k}^T{\bf{Q}}} \right\|}_2}}}{{{{\left\| {\bf{Q}} \right\|}_2}}},
\end{eqnarray}
where $\mathbf{Q}$ is the FOM solution snapshot matrix, ${\bf{V}}_{1:k}$ is the first $k$ columns of the left singular vector from the SVD of training data. Again, at $\alpha = 3.25, L_c = 0.254,0.508,0.762$ m, the projection error of $\rho$ up to $k=100$ is shown in Fig. \ref{fig projection}. It can be observed that both RD-MD and FD-MD provide a monotonic decrease in the projection error. However for FD-FD, the error increases with $k$ at a few cases, which is not desirable for ROM development as it makes the choice of basis size more uncertain. The increase can be again attributed to the fact that the POD method solves a global minimization problem and the optimal bases for the full domain is not necessarily optimal for the reduced domain. The improved projection error property with the multi-domain method implies a higher accuracy and stability of ROM, which is confirmed in the following sections.
The above analysis are focused on $\rho$ and results for other variables follow the same trend.
\begin{figure}
\centering
\subfloat[$L_c=0.254$ m]{
\includegraphics[width=0.33\textwidth]{reconstruction_1.jpg}}
\subfloat[$L_c=0.508$ m]{
\includegraphics[width=0.33\textwidth]{reconstruction_2.jpg}}
\subfloat[$L_c=0.762$ m]{
\includegraphics[width=0.33\textwidth]{reconstruction_3.jpg}}
\caption{Projection error in $\rho$ for $\alpha = 3.25$}\label{fig projection}
\end{figure}
\subsection{FOM results}
Results from the full-domain FOM simulations are presented to characterize the instability behavior. The predicted response is visualized using pressure signals obtained 0.0127 meters upstream of the converging part of the nozzle, which is a typical location selected to probe combustion instabilities in the work conducted in Ref. \cite{frezzotti2015response}.
With different combinations of parameters, two general categories of responses are identifiable: one with positive growth rate, in which the pressure oscillation grows and settles into a limit cycle oscillation (LCO), and one with negative growth rate in which the instability starts to decay after the perturbation ends and the flow converges to a steady state. The definition of the growth rate, $gr$, is based on the peak-to-peak amplitude of the unsteady part of the pressure signal, whose growth is fitted to a exponential function plotted in Fig. \ref{fig growth rate}, given by
\begin{equation}\label{eq gr}
p(t)=p(0)e^{t\cdot gr}.
\end{equation}
To better distinguish the categories, two representative examples (at $\alpha=3.4,L_c=0.5715$ m for the growing category and $\alpha=3.1,L_c=0.254$ m for the decaying category, respectively) are presented in Fig.~\ref{fig p_grow} and \ref{fig p_decay}, where the difference in growth rate can be clearly observed via the monitored pressure signals and the spatio-temporal diagrams of the pressure evolution. Similar responses can be found in previous studies~\cite{xu2017reduced, wang2018non}. To provide more insight into the dynamics and the coupling of pressure and heat release described by Eq. \eqref{eq alpha-tau}, the pressure distribution and the unsteady heat release term are also plotted in the same figures. Each figure includes two snapshots at the beginning of the simulation, and two snapshots towards the end. It can be seen that in the case showing pressure amplitude growth (typically referred to as unstable case), the unsteady heat release has grown significantly with the pressure oscillation at the end of the simulation compared with the beginning due to the fact that the oscillations of pressure and heat release rate are in-phase, which is considered favorable to drive combustion instability. While for the other case with pressure amplitude decay (usually referred to as stable case), for some time instances, the pressure and heat release rate oscillations are out-of-phase, which is recognized as a combustion instability damping mechanism.
\begin{figure}
\centering
\subfloat{
\includegraphics[width=0.5\textwidth]{gr_def.jpg}}\\
\caption{Definition of growth rate}\label{fig growth rate}
\end{figure}
\begin{figure}
\centering
\subfloat[Pressure signal at the monitored location]{
\includegraphics[width=0.5\textwidth]{fom_225.jpg}}
\subfloat[Spatio-temporal diagram of the pressure evolution]{
\includegraphics[width=0.5\textwidth]{spatio_temporal_growth.jpg}}\\
\subfloat{
\includegraphics[width=0.45\textwidth]{p_growth5525.jpg}
\includegraphics[width=0.45\textwidth]{p_growth6030.jpg}}\\
\subfloat[Pressure and unsteady heat release profiles. Dotted line: steady state pressure, dashed line: pressure signal monitor location][Pressure and unsteady heat release profiles. \\Dotted line: steady state pressure, dashed line: pressure signal monitor location]{
\includegraphics[width=0.45\textwidth]{p_growth99620.jpg}
\includegraphics[width=0.45\textwidth]{p_growth99816.jpg}}
\caption{Example response with positive growth rate, $\alpha=3.4, L_c=0.5715$ m.}\label{fig p_grow}
\end{figure}
\begin{figure}
\centering
\subfloat[Pressure signal at the monitored location]{
\includegraphics[width=0.5\textwidth]{fom_10.jpg}}
\subfloat[Spatio-temporal diagram of the pressure evolution]{
\includegraphics[width=0.5\textwidth]{spatio_temporal_decay.jpg}}\\
\subfloat{
\includegraphics[width=0.45\textwidth]{p_decay5541.jpg}
\includegraphics[width=0.45\textwidth]{p_decay5883.jpg}}\\
\subfloat[Pressure and unsteady heat release profiles. Dotted line: steady state pressure, dashed line: pressure signal monitor location][Pressure and unsteady heat release profiles. \\Dotted line: steady state pressure, dashed line: pressure signal monitor location]{
\includegraphics[width=0.45\textwidth]{p_decay96978.jpg}
\includegraphics[width=0.45\textwidth]{p_decay99455.jpg}}
\caption{Example response with negative growth rate, $\alpha=3.1, L_c=0.254$ m.}\label{fig p_decay}
\end{figure}
\section{Acknowledgments}
The authors acknowledge support from the Air Force under the Center of Excellence grant FA9550-17-1-0195, titled ``Multi-Fidelity Modeling of Rocket Combustor Dynamics.''
\input{ref.bbl}
\end{document}
\section*{Nomenclature}
{\renewcommand\arraystretch{1.0}
\noindent\begin{longtable*}{@{}l @{\quad=\quad} l@{}}
{$\alpha$}&{amplification factor}\\
{$A$}&{cross sectional area}\\
{${C}_{f/o}$}&{stoichiometric fuel-to-oxidizer ratio}\\
{$\eta$}&{cumulative energy in singular values}\\
{$gr$}&{pressure oscillation growth rate}\\
{$J$}&{characteristic invariant for incoming acoustic wave}\\
{$k$}&{number of modes in proper orthogonal decomposition (POD) basis}\\
{$L_c$}&{chamber length}\\
{$l_s$}&{starting location of fuel injection}\\
{$l_f$}&{ending location of fuel injection}\\
{${\dot m }_f$}&{fuel mass flow rate}\\
{${\dot \omega }_{f}$}&{fuel consumption rate}\\
{${\dot \omega }_{ox}$}&{oxidizer consumption rate}\\
{$q'$}&{unsteady heat release rate}\\
{$\sigma$}&{singular value}\\
{$\tau$}&{time lag}\\
{$\mathbf{V}$}&{POD basis}\\
{$\mathbf{V}\bot$}&{complementary basis of $\mathbf{V}$}\\
{$\xi$}&{variable for fuel injection shape description}\\
{$Y_{ox}$}&{oxidizer mass fraction}\\
\end{longtable*}}
\subsection{Off-design condition performance}
To further assess the RD-MD framework, the ROM trained using $\alpha = 3.25, k = 100$ is evaluated at several off-design conditions. The evaluations include two cases at chamber lengths $L_c = 0.1905$ m and 1.016 m. These cases are characterized by dominant acoustic frequencies 2550 and 500 Hz, respectively, which are outside the range of the training frequencies given by Eq. \eqref{eq train signal}. The other conditions have amplification factors that deviate significantly from the training simulation, including $\alpha = 2.85,3.05,3.45,3.65$. The results for the L2 error ($\epsilon_2$), frequency ($f$), growth rate ($gr$), LCO peak-to-peak amplitude ($amp$) and their relative errors are summarized in Table \ref{table off-design}. It should be noted that due to the characteristic training method, the training case does not have a growing response as in the predictions, thus no growth rate or LCO amplitude is reported for the training set. Moreover, the results are plotted along with the designed conditions in Fig. \ref{fig off-design lc} and \ref{fig off-design alpha} for a better illustration of their relation.
It is observed that, at all the off-design conditions, the relative L2 errors are below $\num{1.2e-3}$. For the case with shorter $L_c$ and higher instability frequency (OD $L_{c1}$), and the four cases with deviated $\alpha$ (OD $\alpha_1$ to OD $\alpha_4$), the ROM performance is comparable to that in the designed conditions. However for the case with longer $L_c$ and lower instability frequency (OD $L_{c2}$), the error in LCO amplitude is $2.5\%$, which is more than 5 times higher than the other cases. The result indicates that for this problem, when operating within the frequency range of the training perturbation, the RD-MD method is not only independent of the chamber lengths, but also insensitive to changes in the unsteady heat release term. When beyond the training range, the instability frequency influences the predictive capabilities of the framework, and demonstrates the importance of the multi-frequency perturbation in the characteristic training.
\begin{table}
\begin{center}
\caption {Off-design condition results \\($f,gr,amp$ are listed in a ``FOM/ROM" style, all errors are relative)}\label{table off-design}
\begin{small}
\begin{tabular}{ |c| c c c c c c c c c|}
\hline
Case &$L_c$ (m) &$\alpha$ &$\epsilon_2$ &$f$ (Hz) & $\epsilon_{f}$ &$gr$ ($s^{-1}$) & $\epsilon_{gr}$ & $amp $ (Mpa) & $\epsilon_{amp}$\\
\hline
Training &N/A & 3.25 & N/A & 700 to 2100 & N/A & N/A & N/A & N/A & N/A\\
\hline
OD $L_{c1}$ &0.1905 & 3.25 &\num{9.6e-5} & 2516/2516 & 0 &-139.98/-139.79 &\num{1.4e-3} &0/0 &0 \\
OD $L_{c2}$ &1.016 & 3.25 &\num{2.7e-4} & 490/490 &0 &12.91/12.84 &\num{5.0e-3} &0.4576/0.4688 &\num{2.5e-2} \\
OD $\alpha_1$ &0.508 & 2.85 &\num{1.2e-4} & 1000/1000 &0 &28.87/28.87 &\num{2.3e-5} &0.0522/0.0521 &\num{1.4e-3} \\
OD $\alpha_2$ &0.508 & 3.05 &\num{3.3e-4} & 1000/1000 &0 &61.99/62.00 &\num{9.1e-5} &0.1545/0.1543 &\num{1.2e-3} \\
OD $\alpha_3$ &0.508 & 3.45 &\num{1.2e-3} & 980/980 &0 &122.38/122.41 &\num{2.5e-4} &0.6366/0.6366 &\num{8.5e-5} \\
OD $\alpha_4$ &0.508 & 3.65 &\num{1.6e-3} & 980/980 &0 &130.19/130.27 &\num{6.3e-4} &0.8368/0.8368 &\num{4.4e-5} \\
\hline
\end{tabular}
\end{small}
\end{center}
\end{table}
\begin{figure}
\centering
\subfloat{
\includegraphics[width=0.49\textwidth]{od_lc_gr.png}}
\subfloat{
\includegraphics[width=0.49\textwidth]{od_lc_lco.png}}
\caption{Results at off-design $L_c$}\label{fig off-design lc}
\end{figure}
\begin{figure}
\centering
\subfloat{
\includegraphics[width=0.49\textwidth]{od_alpha_gr.png}}
\subfloat{
\includegraphics[width=0.49\textwidth]{od_alpha_lco.png}}
\caption{Results at off-design $\alpha$}\label{fig off-design alpha}
\end{figure}
\subsection{Projection-based ROM}
To derive the projection-based ROM equations, we first rewrite the full order model given by Eq. \eqref{eq fom} for each variable as
\begin{equation}\label{eq FOM1}
\frac{{\partial {\mathbf{q}^i}}}{{\partial t}} = {\mathbf{r}^i}\left( {{\mathbf{q}}\left( {x,t} \right)} \right),
\end{equation}
where the variable index $i$ corresponds to the four conserved variables respectively. The residual vector for variable $i$ consists of the corresponding flux and source terms
\begin{equation}
\mathbf{r}^i = - \frac{{\partial \mathbf{f}^i}\left( {{\mathbf{q}}\left( {x,t} \right)} \right)}{\partial x} + \mathbf{s}^i.
\end{equation}
For each variable, defining an individual POD projection basis $\mathbf{V}^i \in \mathds{R}^{n_{grid}\times k}$ spanning a subspace $\mathcal{V} \subset \mathds{R}^{n_{grid}}$, and a complementary basis $\mathbf{V}^i_\bot$ spanning $\mathcal{V}_\bot$, such that $\mathcal{V} \oplus \mathcal{V}_\bot = \mathds{R}^{n_{grid}}$, the following decomposition can be derived:
\begin{eqnarray}\label{split_q}
\mathbf{q}^i = \mathbf{V}^i \mathbf{q}^i_r + \mathbf{V}^i_\bot \mathbf{q}^i_\bot = \tilde{\mathbf{q}}^i + \hat{\mathbf{q}}^i.
\end{eqnarray}
The two bases $\mathbf{V}^i$ and $\mathbf{V}^i_\bot$ are obtained by performing singular value decomposition (SVD) on the snapshots of the $i$-th conserved variable. Defining $k$ to be the designed size of the reduced order system for variable $i$, $\mathbf{V}^i$ is constructed column-wise from the first $k$ left-singular vectors from the SVD~\footnote{The remaining left-singular vectors are $\mathbf{V}^i_\bot$}. The conserved variables are then approximated as
\begin{equation}\label{eq ROM0}
{\mathbf{q}^i}(x,t) \approx {\bf{\tilde q}}^i = \mathbf{V}^i{\mathbf{q}^i_r}=\sum\limits_{n=1}^{k} {q^{i(n)}_r(t)\mathbf{v}^{i(n)}(x)},
\end{equation}
where $q^{i(n)}_r$ is the $n$-th element in the reduced order variable vector $\mathbf{q}^i_r \in \mathds{R}^k$, and $\mathbf{v}^{i(n)}$ is the $n$-th column in the POD basis $\mathbf{V}^i$. Since each left-singular vector represents a spatial mode of the FOM solution, the approximation can be viewed as a superimposition of the first $k$ leading modes, with $q^{i}_r$ being the temporal coefficient in the evolution of the $i$-th mode.
The Galerkin-projected ROM for $k$ coefficients is then
\begin{equation}\label{eq ROM1}
\frac{{d {\mathbf{q}^i_r}}}{{d t}} = {(\mathbf{V}^i)^T}{\mathbf{r}^i}\left({\mathbf{V}\mathbf{q}_r} \right).
\end{equation}
For clarity of presentation, $\mathbf{V}\mathbf{q}_r$ is used to denote the approximated conserved variables, including ${\mathbf{V}^1\mathbf{q}^1_r},\dots,{\mathbf{V}^4\mathbf{q}^4_r}$.
In practice, to achieve further reduction in computational cost and to improve robustness, additional constructs such as sparse sampling techniques and stabilization will be required. While we are developing such techniques~\cite{xu2017reduced,xu2018multi,huang2018exploration,ericpaper}, they will not be considered in the present work, as the focus is on multi-domain modeling.
\subsection{ROM evaluations}
Before integrating the ROM into the Multi-fidelity Framework, the predictive capability of the ROM obtained from the reduced-domain simulation is evaluated. In this test, the ROM trained with broadband characteristic perturbation is tested by feeding single frequency perturbations on the characteristic boundary. This is to ensure that the ROM is able to predict the essential dynamics when the downstream section is dominated by single frequency acoustics. Aside from the perturbation frequency, there is no difference between the conditions used in the test and the truncated domain training (Sec. \ref{sec framework}).
The tests are conducted over $t=0-0.05$ s at five frequencies between 700 and 2000 Hz. It should be noted that some of the frequencies are not included in the broadband training. The number of ROM modes and amplification factor are $k=100, \alpha = 3.25$, respectively. The results are compared with the reference FOMs using the same conditions, and the L2 error over the simulated period as well as at the last time step is plotted in Fig. \ref{fig sanitytest}. It can be seen that at all perturbation frequencies, the ROM is able to accurately predict the FOM behavior with an L2 error below \num{1.1e-11}. These evaluations confirm the fact that the ROM trained with broadband perturbation is capable of predicting dynamics over a wide range of frequencies.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{error_inlet.jpg}
\caption {Sanity test for L2 error}\label{fig sanitytest}
\end{center}
\end{figure}
\subsection{Multi-fidelity framework evaluations}
The solution from the three different techniques are compared against the FOM over the same parameters and time period. From Fig.~\ref{fig projection} it can be seen that the projection error for RD-MD at $k=20$ is comparable to that for FD-FD at $k=100$. The three methods are compared at the two basis sizes, $k=20,100$ in this section. Comparisons of the relative L2 error of the ROM solution are shown in Fig. \ref{fig L2}. It should be mentioned that in several cases, the ROMs in the control groups are numerically unstable. The error is set to 1 under these circumstances, which is higher than the error in all stable cases.
When the basis size is sufficiently large, i.e. at $k=100$, RD-MD performs similarly to FD-MD, except for one case at $\alpha = 3.25, L_c = 0.5715$ m, where the latter is unstable. At high $\alpha$, where the instability is stronger, the advantage of RD-MD becomes more significant. This can be observed from the medium $L_c$ cases at $\alpha = 3.4$, where the gap between RD-MD and FD-MD is apparent. Also, FD-FD performs worse as $\alpha$ increases, resulting in higher errors and more unstable cases. It should be noted that FD-FD sometimes outperforms the other two methods at low $L_c$, which is consistent with an improved decay in the singular values. However at $L_c = 0.254$ m, it becomes unstable.
When the basis size is reduced to $k=20$, all the predictions deteriorate considerably. The two full-domain training methods become unstable in most conditions. In contrast, RD-MD remains stable at all combinations of parameters, although the ROM solution error is approximately one order higher than that of the $k=100$ cases. The advantage in stability confirms the conclusion that the reduced-domain training results in a better set of basis functions. By virtue of using a smaller basis set, the computational cost, e.g. the projection computation in Eq.~\eqref{eq ROM1}, decreases linearly when an explicit time-marching scheme is used. In contrast to the traditional full-domain training methods, for which the stability is less predictable, the proposed framework provides a broader stability envelope, and consequentially a more flexible balance between ROM accuracy and efficiency.
\begin{figure}
\centering
\subfloat[$\alpha=3.1$]{
\includegraphics[width=0.32\textwidth]{err_310.jpg}}
\subfloat[$\alpha=3.25$]{
\includegraphics[width=0.32\textwidth]{err_325.jpg}}
\subfloat[$\alpha=3.4$]{
\includegraphics[width=0.32\textwidth]{err_340.jpg}}
\caption{Relative L2 error of ROM solutions, numerically unstable cases set to 1}\label{fig L2}
\end{figure}
Major quantities of interest in rocket combustor design include the dominant acoustic frequencies, the growth rates and LCO peak-to-peak amplitudes of the pressure oscillations. It should be noted that the dominant acoustic frequencies have been well-predicted by all three methods. Therefore only the comparisons of the other two quantities, growth rates and LCO amplitudes, are shown in the current study, as can be seen in Figs. \ref{fig gr} and \ref{fig lco} respectively for all the methods. The relative performance between different methods follows a similar relation as in the L2 error analysis. It can be seen that the proposed RD-MD framework is able to predict the relation between the growth rate and $L_c$ accurately with an error below $0.5\%$ at $k=100$ and below $5\%$ at $k=20$ in most cases, which illustrates its effectiveness.
\begin{figure}
\centering
\subfloat{
\includegraphics[width=0.32\textwidth]{gr_310.jpg}}
\subfloat{
\includegraphics[width=0.32\textwidth]{gr_325.jpg}}
\subfloat{
\includegraphics[width=0.32\textwidth]{gr_340.jpg}}
\hfill
\subfloat[$\alpha=3.1$]{
\includegraphics[width=0.32\textwidth]{err_gr_310.jpg}}
\subfloat[$\alpha=3.25$]{
\includegraphics[width=0.32\textwidth]{err_gr_325.jpg}}
\subfloat[$\alpha=3.4$]{
\includegraphics[width=0.32\textwidth]{err_gr_340.jpg}}
\caption{Growth rate of pressure oscillation 0.0127 m upstream of nozzle}\label{fig gr}
\end{figure}
\begin{figure}
\centering
\subfloat{
\includegraphics[width=0.32\textwidth]{lco_310.jpg}}
\subfloat{
\includegraphics[width=0.32\textwidth]{lco_325.jpg}}
\subfloat{
\includegraphics[width=0.32\textwidth]{lco_340.jpg}}
\hfill
\subfloat[$\alpha=3.1$]{
\includegraphics[width=0.32\textwidth]{err_lco_310.jpg}}
\subfloat[$\alpha=3.25$]{
\includegraphics[width=0.32\textwidth]{err_lco_325.jpg}}
\subfloat[$\alpha=3.4$]{
\includegraphics[width=0.32\textwidth]{err_lco_340.jpg}}
\caption{LCO peak-to-peak amplitude of pressure oscillation 0.0127 m upstream of nozzle}\label{fig lco}
\end{figure}
Finally, to provide a more direct comparison between the ROM solution from the framework and the FOM solution, spatial pressure profiles for $L_c=0.508$ m, $\alpha = 3.25, k=100$ are provided at several time instances in Fig.~\ref{fig xp}. These plots are focused on the back-step area for better visualization, and are selected by the end of the simulation, when the error is maximum. It is observed that the RD-MD solution correlates well with the FOM, and connects smoothly across the interface.
\begin{figure}
\centering
\subfloat[$t=0.098$ s]{
\includegraphics[width=0.3\textwidth]{interface_20_9800.jpg}}
\subfloat[$t=0.09825$ s]{
\includegraphics[width=0.3\textwidth]{interface_20_9825.jpg}}
\subfloat[$t=0.0985$ s]{
\includegraphics[width=0.3\textwidth]{interface_20_9850.jpg}}
\caption{Spatial profiles of pressure}\label{fig xp}
\end{figure}
\subsection{Full order model}
In this work, a quasi-1D version of the single element Continuously Variable Resonance Combustor (CVRC)~\cite{yu2012spontaneous,yu2009experimental} is used as the full order model. The geometry of the CVRC is sketched in Fig. \ref{fig geometry}. Numerical studies are performed with multiple chamber lengths, which exhibit various instability behaviors. The geometric parameters for other sections are given in Table \ref{table geometry}. To avoid invalidating the quasi-1D assumption, the back-step and the converging part of the nozzle are sinusoidally contoured in this study, in contrast to a discontinuous change in the cross-sectional area in the experimental setup. The same gas properties and operating conditions as in Ref.~\cite{xu2017reduced} are adopted and listed in Table \ref{CVRC}.
\begin{table}[h]
\begin{center}
\caption {Geometry parameters}\label{table geometry}
\begin{tabular}{ |c| c| c| c| c|}
\hline
Section &Injector &Back-step &Nozzle converging part &Nozzle diverging part\\
\hline
Length (m) &0.1397 &0.0064 &0.0127 &0.034\\
Radius (m) &0.0102 &0.0102 to 0.0225 &0.0225 to 0.0104 &0.0104 to 0.0195\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}
\begin{center}
\caption {CVRC operating conditions}\label{CVRC}
\begin{tabular}{ |c| c| }
\hline
Parameter & Value\\
\hline
Fuel mass flow rate, kg/s & 0.027\\
Fuel temperature, K & 300\\
Oxidizer mass composition & $57.6\%\: H_2O+42.4\%\:O_2$\\
Oxidizer mass flow rate, kg/s & 0.32\\
Oxidizer temperature, K & 1030\\
Fuel composition & $100\%\:CH_4$\\
Equivalence ratio & 0.8\\
\hline
\end{tabular}
\end{center}
\end{table}
The governing equations in this problem are the quasi-1D unsteady Euler equations with species transport, represented as
\begin{equation}\label{eq fom}
\frac{{\partial \mathbf{q}}}{{\partial t}} + \frac{{\partial \mathbf{f}}}{{\partial x}} = \mathbf{s}_A+\mathbf{s}_f+\mathbf{s}_q,
\end{equation}
where
\small
\begin{equation}
\mathbf{q} = \left( {\begin{array}{*{20}{c}}
\rho A \\
{\rho uA} \\
{\rho{E}A} \\
{\rho {Y_{ox}A}}
\end{array}} \right),
\mathbf{f} = \left( {\begin{array}{*{20}{c}}
{\rho uA} \\
{\left(\rho {u^2} + p\right)A} \\
{\left(\rho{E} + p\right)uA} \\
{\rho u{Y_{ox}}A}
\end{array}} \right),
\mathbf{s}_A = \left( {\begin{array}{*{20}{c}}
0 \\
{p\frac{{dA}}{{dx}}} \\
0 \\
0
\end{array}} \right),
\mathbf{s}_f = \left( {\begin{array}{*{20}{c}}
{{{\dot \omega }_f}} \\
{{{\dot \omega }_f}u} \\
{{{\dot \omega }_f}\Delta {h_0}} \\
{ - {{\dot \omega }_{ox}}}
\end{array}} \right),
\mathbf{s}_q = \left( {\begin{array}{*{20}{c}}
0 \\
0 \\
{q'} \\
0
\end{array}} \right).
\end{equation}
\normalsize
In the conservative variable vector $\mathbf{q}$, $\rho$ is the density, $u$ is the velocity, $E$ is the total internal energy, $Y_{ox}$ is the oxidizer mass fraction, and $A$ is the cross sectional area. The corresponding convective fluxes are represented by the vector $\mathbf{f}$, where $p$ is the static pressure.
In the three source terms, $\mathbf{s}_A$ is due to area variations and the latter two terms represent combustion. In the experiment~\cite{yu2009experimental}, the fuel is injected through an annular ring located at the back-step and reacts at a finite rate with the oxidizer injected. In this work, we follow the choice of Frezzotti et al.~\cite{frezzotti2015response, frezzotti2015parametric} and assume an infinitely-fast one-step combustion model. An important assumption behind the model is that when fuel is injected, it will react with the oxidizer instantaneously to form products and no intermediate species are produced, thus only one species transport equation is needed. This process is accounted for in $\mathbf{s}_f$. As suggested by Frezzotti et al.~\cite{frezzotti2015response, frezzotti2015parametric}, to reproduce the combustion region of a finite width and avoid discontinuities, the fuel injection process is modeled in a sinusoidal shape at the rate of ${\dot \omega }_f$, yielding
\begin{eqnarray}\begin{aligned}\label{eq flame}
{{\dot \omega }_f} = \frac{{{{\dot m}_f}}}{{\int {\left( {1 + \sin \xi } \right)} dx}}\left( {1 + \sin \xi } \right),\\
\xi = - \frac{\pi }{2} + 2\pi \frac{{x - {l_s}}}{{{l_f} - {l_s}}},{\text{with }}{l_s} < x < {l_f},
\end{aligned}\end{eqnarray}
where ${\dot m}_f$ is the total mass flow rate of fuel injection, and $l_s$ and $l_f$ are the starting and ending location of the flame, respectively.
Due to the infinitely-fast model, the consumption rate of the oxidizer is
\begin{eqnarray}
{\dot \omega }_{ox} = \frac{{\dot \omega }_f}{C_{f/o}},
\end{eqnarray}
where $C_{f/o}$ is the stoichiometric fuel-to-oxidizer ratio. The shape of the fuel injection and resulting flame given by Eq. \eqref{eq flame} is also shown in Fig. \ref{fig geometry} along with the CVRC geometry.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{alphatau.jpg}
\caption {Computational domain and flame shape}\label{fig geometry}
\end{center}
\end{figure}
The final source term $\mathbf{s}_q$ models the unsteady heat release and represents the response function introduced by Crocco et al.~\cite{crocco1958importance}, given in Eq. \eqref{eq alpha-tau}. The function takes into account the coupling between acoustics and combustion by expressing the unsteady part of the heat release as a function of pressure with an amplification parameter $\alpha$ and a time lag $\tau $.
\begin{equation}\label{eq alpha-tau}
q' = {\dot \omega _f}\Delta {h_0}\alpha \frac{{p\left( {x,t - \tau } \right) - \bar p\left( x \right)}}{{\bar p\left( x \right)}}.
\end{equation}
In the simulation, a steady state is first achieved with the unsteady source term $\mathbf{s}_q$ turned off. From the steady state, a low-amplitude perturbation is applied to the inlet boundary to trigger the instability at the beginning of predictive unsteady simulations (either FOM or ROM). After the perturbation, the unstable response will become self-excited and start to grow under certain parameters. The instability behavior is described in Sec. \ref{sec result_domain}.
Numerical solutions of Eq. \eqref{eq fom} will be used as a surrogate for the high-fidelity solution in the current work, and the corresponding simulation is referred as FOM.
|
\section{Introduction}
What should you bid in a second price ad auction for a display with a known click-through rate (CTR), for a given cost-per-click (CPC)?
The most probable (and possibly incorrect) answer is "CPC times CTR". However, the right answer is "What do you mean by cost-per-click?".
Indeed, if the CPC is the maximal amount the buyer is ready to pay for each click, the first answer is correct. But the story is different when CPC means the maximal average cost per click.
In this article, we challenge the current modeling approach for ad auctions. We argue that when some advertisers' business constraints apply, the expected outcome of the auction may depart from the traditional literature.
Advertising is a major source of revenue for Internet publishers, and as such, is financing a large part of the Internet. About 200 billion USD spent in 2017 on digital advertising \cite{Kafka2017}.
Banners for display advertising are usually bought through a high frequency one unit auction mechanism called RTB (Real-Time Bidding) by or on behalf of advertisers.
When a user reaches a publisher page, it triggers a call to an RTB platform. The RTB platform then calls potential buyers, who have a few milliseconds to answer with a bid request.
This results in an allocation of the display to a bidder, in exchange for a payment. The allocation and the payment are defined by the rules of the auction, which depends on the RTB platforms.
The winner of the auction can show a banner to the internet user.
Yet, for many end buyers, what really matters is not the display itself, but what will result from it: clicks, conversions, etc...
This is why empirical CPA measures such as average cost per click (CPC) or average cost per order (CPO) are of paramount importance.
The literature on display advertising auctions has been growing over the last decade, pushed by the emergence of Internet giants whose business models rely on digital advertising.
One track of research takes the seller's point of view and focuses on how to build "good" auctions, and relies on mechanism design theory
\cite{balseiro2015repeated,golrezaei2017boosted}, while the dual track brings the buyer perspective under scrutiny, and focuses on the design of bidding strategies
\cite{zhang2014multi,agarwal2014budget,diemert2017attribution,fernandez2016optimal,cai2017real}.
The reader can refer to \cite{roughgarden2016twenty} and \cite{krishna2009auction} for an introduction to auction theory.
Usually, in performance markets, the advertiser has a target Cost Per Action (CPA) and/or a budget in mind.
We will focus the analysis on the CPA constraint.
We will assume that the budget constraint is absent or not binding. This topic of budget constraints has already been addressed in the literature \cite{balseiro2015repeated,agarwal2014budget,fernandez2016optimal}. The CPA is measured in term of the average quantity of money spent per action, the action is a click, a visit, a conversion...
For instance, if
(a) we are facing a competition uniformly distributed between $0$ and $1$,
(b) the CTR of every display is $0.5$,
(c) the auction is second price and
(d) we are ready to pay 1 USD on average for a click,
then if we bid $0.5$, we will win half of the time. For every display won, we get on average half a click, and we pay $2\int_0^{0.5} t dt = 1/4$, thus our expected CPC is $1/(4*0.5)=0.5<1$. We should increase the bid to raise the empirical CPC.
One reason for using CPA constraints instead of budgets constraints in performance markets is that if the campaign performs well, there is no reason to shut it down at the middle of the month because of a strict budget limit. Those constraints, implemented by algorithms, can be modified by comparing the performances of different channels.
Auctions with ROI constraints are a hot topic \cite{Auerbach2008,Wilkens2016,Golrezaei2018}.
In particular \cite{Golrezaei2018} provide some empirical evidence that some buyers are ROI constrained and propose an optimal auction design.
Our work departs from them because we use a different notion of ROI: in the definition of CPA, we do not subtract the cost to the revenue in the numerator.
The reason to do so is that end buyers reason in « cost per click » or « cost per order ». The numerator is not expressed in a unit of money and does not include the buyer payment to the seller.
Also, our focus is more buyer than seller oriented.
In this work, we focus on the buyers perspective, but we also show on simple market instances that the CPA constraints may impact the seller design, or give rise to undesirable competitive behaviors.
This work brings several contributions to the table.
First, we introduce the buyer's CPA constrained optimization problem and exhibit a solution,
then we find the symmetric equilibrium in this setting and compare its properties with standard results from the literature. In particular, no reserve price should be used.
Last, since in practice, the CPA is computed over a time window of repeated auctions, we move the discussion to the dynamic settings for which we explain why one can expect adaptive bid multipliers to provide solutions close to the optimal for the buyer.
\section{Modeling Assumptions and Notations}
We start by introducing the CPA-constrained bidding problem.
A seller is selling one item (a display opportunity in our context) through an auction.
The item brings a value $v_i$ to bidder $i\in{1..n}$.
These values are independently distributed among bidders.
For example, in the context of the ad auction, $v$ can be thought of as the expected number of clicks (or sales) the bidder would get after winning the auction. Thus, this value is not expressed in money but in unit of action.
We will see later how these values, expressed in actions relates to bids, expressed in dollars.
Let $F_i$ be the cumulative distribution and $f_i$ the density distribution of $v_i$. For simplicity, we assume the support of these density distributions to be compact intervals.
When we take the viewpoint of bidder $i$,
we denote by $b^{-i}$ the greatest bid of the other bidders (the price to beat). We denote by $g^{i}$ and $G^{i}$ its density and cumulative distributions.
Unless otherwise stated, we also assume the auction to follow a second price rule, with $n>1$ bidders competing for the item.
Bidders with CPA constraints compare:
\begin{itemize}
\item the expected value they get in the auction
\item the expected payment they incur.
\end{itemize}
We will refer to the CPA with the letter $\T$ because it is our "targeted cost per something".
We will denote by $b^i$ the bid of a given bidder $i$ of interest.
The bidder wants to maximize his expected value, subject to an \emph{ex ante} constraint in expectancy representing the targeted CPA. The constraint is \emph{ex ante} because, in practice, the same bidding strategy is going to be applied repeatedly (or even simultaneously) on several similar auctions.
If for a random variable $X$, we denote by $[X]\in\{0,1\}$ the binary random variable that takes the value $1$ when $X\geq 0$, and $0$ otherwise, then the expected value earned by the buyer bidding ${{b^i}}$ is $E_{v,b^-}[b^-<{{b^i}}(v)] v$.
Here $E_{v,b^-}$ is there expectation operator on the distribution of $v$ and $b^-$.
Similarly the bidder expected cost is $\mathbb{E}_{v,b^-} [b^-<{{b^i}}(v)]b^- $.
We can now express the bidder optimization problem. The bidder is looking for a bid function ${{b^i}}:[0,1]\rightarrow\R^+$ solution of
\begin{eqnarray}
\label{eq:bidderProblem}
\max_{{{b^i}}}\quad \mathbb{E}_{v,b^{-i}} [b^{-i}<{{b^i}}(v)]v,
\end{eqnarray}
subject to $ \mathbb{E}_{v,b^{-i} }[b^{-i}<{{b^i}}(v)]b^{-i} \leq \T^i \mathbb{E}_{v,b^{-i}}[b^{-i}<{{b^i}}(v)] v$ (CPA).
Observe that if we remove the constraint, the buyer would buy all the opportunities no matter the cost.
We pinpoint that problem \eqref{eq:bidderProblem} is not equivalent to a budget constraint problem: we would have had $ \mathbb{E}_{v,b^{-i}} [b^{-i}<{{b^i}}(v)]b^{-i} \leq budget$
instead of the (CPA) constraint.
Nor is it equivalent to the maximization of the yield $\mathbb{E}_{v,b^-}(T^iv - b^{-i})[b^{-i}<b^i(v)]$, as illustrated in the introduction.
We model the interactions among buyers with a game. Observe that this game has constraints on the strategies profile.
For a given set of competing bidding strategies, a best reply of the bidder is a solution of (\ref{eq:bidderProblem}).
A (constrained) Nash equilibrium is a strategies profile with components that are best replies against the others .
\textit{In the following, the superscript $i$ is often omitted to help readability.}
\section{Bidding Behavior and Symmetric Equilibrium}
The main result of this section is the derivation of a symmetric Nash equilibrium for an CPA constrained second price auction. We then generalize this result to linear combinations of first price and second price auctions.
\subsection{Second Price}
Going back to the motivational example of the introduction, if a buyer bids $Tv$, then he would pay \emph{at most} $T$ per unit of action.
Since he can increase his bid until he pays $T$ per unit of action \emph{on average}, it is intuitive that he should bid higher than $Tv$ to maximize the criterion. The next theorem formalizes this idea.
\begin{theorem}[Best Reply]
\label{theorem:secondprice-best-reply}
For any bounded distribution of the price to beat $b^-$, there exists $\lambda^*\in \bar{\R}^+$ such that a solution of \eqref{eq:bidderProblem} writes:
\begin{equation}
\label{eq:best-reply}
{{b^i}}(v,\lambda^*) = (\nicefrac {1} {\lambda^*} +\T^i )v.
\end{equation}
\end{theorem}
\begin{lemma}[CPA monotony]
\label{lemma:roi-monotony-second-price}
If the bidder bids proportionally to the value i.e. if there exists $\alpha$ such that $b(v) = \alpha v$, then the CPA is non-decreasing in $\alpha$.
\end{lemma}
\begin{proof}
Fix $u\in[0,1]$, $\alpha>0$, then the expected cost knowing that the value is $v$ and the auction is won is $\int_0^{\alpha v} t g(t) {\rm d} t$, where $g$ is the probability density of the price to beat. This quantity is increasing in $\alpha$. Meanwhile the value is $v$.
We get the result by integration on $v$.
\end{proof}
If $v_1, v_2, \ldots v_n$ are $n$ independent draws from $F$, we denote by $v_i^{(n)}$ the average of the $i^{th}$ greater draw (the ith order statistic).
\begin{theorem}[Equilibrium Bid]
\label{theorem:equilibrium-bid-second-price}
The unique symmetric constrained Nash equilibrium strategy is
\begin{equation}
{{b^i}}^*(v) = \frac{ \T}{\gamma} v,
\end{equation}
where
\begin{equation}
\label{eq:gamma-def}
\gamma(F,n) = n \left(\frac{v_1^{(n-1)}}{v_1^{(n)}} - \frac{n-1}{n}\right).
\end{equation}
\end{theorem}
\paragraph{Discussion}
Theorem \ref{theorem:secondprice-best-reply} is the answer to the introductory question. Despite its simplicity, it shows something that is probably overlooked: it may happen in a second price auction that the bidder's optimal bid depends on the competition. Another useful insight is that the bid is linear w.r.t. the value, which implies that simple bid multipliers could be used optimally. Observe that the result holds also for non-symmetric settings.
The proof relies on the strong incentive compatibility of the second price auction in standard setting reinterpreted on the Lagrangian of the optimization problem.
The parameter $\lambda^*$ should be interpreted as the Lagrangian multiplier associated with the CPA constraint.
Informally speaking, the harder the constraint, the bigger $\lambda$, the smaller the bid.
When $\lambda$ is set to $+\infty$, one bids only for displays that do not consume the constraint, while when $\lambda$ goes to zero, the bid diverges.
Lemma \ref{lemma:roi-monotony-second-price} is useful understand what is happening. Observe that since the bid is linear w.r.t. the value and the objective increasing in the bid multiplier, the optimal bid multiplier is the maximal admissible one.
Theorem \ref{theorem:equilibrium-bid-second-price} is one of our main result. Observe that
the \textbf{competition factor} $\gamma$ is only a function of the number of opponents and $F$. Since it is smaller than one (see Appendix) the bid is proportional to $\T v$ with a factor greater than one.
Compared with the standard setting, the seller's expected revenue is multiplied by $\nicefrac{ \T}{\gamma} $.
The asymptotic value of $\gamma$, as well as other natural questions concerning the equilibrium will be studied in Section \ref{sec:properties-of-the-equilibrium}.
Note that Theorem \ref{theorem:equilibrium-bid-second-price} is a necessary and sufficient condition for a symmetric Nash equilibrium, therefore it is unique, but we cannot claim that we have identified the only Nash Equilibrium (see Section \ref{sec:properties-of-the-equilibrium}).
\subsection{Generalization}
\label{subseq:generalization}
We now generalize the argument to auctions which are convex combinations of first price and second price.
We motivate this extension by the existence in the industry of auctions which mix together first and second pricing rules, such as soft-floor.
If we denote by $S(b,b^-)$ the payment rule of the auction when the buyer bids $b$ and the best competing bid is $b^-$, then we can make the following observations:
(1) for any $s\geq 0$, $S(s b,sb^-) = sS(b,b^-)$,
(2) $S(b,b) = b$,
(3) $S(b,b^-)$ is the sum of a linear function of $b$ and a linear function of $b^-$.
Consider a standard auction with payment rule $S$ and symmetric buyers with i.i.d. values distributed according to F (no CPA constraints). Let ${{\hat{b}}}$ be a symmetric equilibrium strategy in such setting.
\begin{theorem}
\label{theorem:equilibirum_strategy-general}
The bid ${{b^i}}^*(v) = \frac{ \T}{\gamma} {{\hat{b}}}(v)$ is a symmetric equilibrium strategy, of the CPA constrained auction with payment rule $S$.
\end{theorem}
\paragraph{Example}:
Consider a first price auction with 2 buyer having uniformly distributed values in $[0,1]$.
It is known \cite{krishna2009auction} that an equilibrium bid for a standard auction is $v \rightarrow 0.5 v$.
By \eqref{eq:gamma-power-case}, $\gamma = 0.5$.
Thus an equilibrium bid for the CPA constrained auction is
$v \rightarrow T v$.
Indeed: in first price, we ensure the saturation of the CPA constraint by bidding $Tv$.
\section{Properties of the Equilibrium}
\label{sec:properties-of-the-equilibrium}
Observe that the expected payment of a buyer is equal to his expected value times $\T$. Since the total welfare is fixed, the expected payment of every buyer is the same no matter the auction.
We thus recover a standard result \cite{vickrey1962auctions} from auction theory, adapted to CPA constrained auctions.
\begin{theorem}[Revenue equivalence]
All the auctions described in Section \ref{subseq:generalization} bring the same expected revenue to the seller.
\end{theorem}
\begin{proof}
Using the proof of Theorem 3.3, we see that
constraint in \eqref{eq:bidderProblem} is binding at the symmetric equilibrium. Therefore the payment of a buyer to the seller is equal to $T\mathbb{E} v[v>v^{-}] =T\mathbb{E} \max_{i\in\{1..n\}}(v_{i})/n$.
\end{proof}
On the other hand, the following observation is quite unusual.
\begin{theorem}[Optimal Reserve Price]
\label{th:reserve-price}
With CPA constraints, the optimal reserve price in a symmetric setting is zero.
\end{theorem}
Explanation: a reserve price would decrease the social welfare, and thus decrease the expected payment because of the CPA constraint.
\begin{proof}
Same idea as before:
In the presence of a reserve price, the constraint in \eqref{eq:bidderProblem} tells us that
the buyer expected payment is smaller than
$T\mathbb{E} v[v>v^{-i}] [v>r]$, which is smaller than $T\mathbb{E} v[v>v^{-i}]$ for r>0.
Using the proof of Theorem 3.3 we see that the constraint should be binding at the equilibrium, and thus
this last quantity corresponds to the expected payment when $r=0$.
Thus the reserve price should be set to zero.
\end{proof}
\begin{figure}%
\centering
\includegraphics[width=7cm]{reserve-price.png} %
\caption{Potential outcomes (buyers' profits and payments) for different values of the bid multiplier in a symmetric setting. Each curve corresponds to a different value of the reserve price. The equilibrium points are the intersections of those curves with the line profit = $T$*payment. \emph{We took $T = 0.4$, uniform distribution of values, and two bidders for the numerical simulation.}} %
\label{fig:optinalReservePrice}%
\end{figure}
We display in Figure \ref{fig:optinalReservePrice} the results of a simulation that illustrates Theorem \ref{th:reserve-price}.
Observe that the proofs of the two last results rely on the symmetry of the setting (same CPA target $T$, same value distribution $F$).
We argue that this should not be seen as a limitation of our results, because (a) Revenue Equivalence
in the classical setting also requires symmetry, (b)
Theorem \ref{th:reserve-price} provides a striking counterexample to the common belief that the reserve price increases the revenue of the seller.
We believe Theorem \ref{th:reserve-price} to be extendable to nonsymmetric setting, as long as one can guarantee that the CPA constraint is binding.
This completely departs from the idea that reserve prices should be used to increase the seller's revenue. Observe that in practice in the case of display advertising, the buyer may take some time to react to a change of empirical CPA.
Consequently, measuring the seller's long term expected revenue uplift is a technical and business challenge.
Moreover, the reserve price, by decreasing the welfare, may on the long run trigger an increase of the target CPA, since buyers have to do an arbitrage between volume and CPA.
Last but not least, we do not claim that reserve prices should be set to zero for CPA constrained bidders in all situations.
A trivial counterexample to such claim is embodied by a setting with only one buyer and no competition.
\begin{theorem}[Convergence]
\label{convergence}
For any number of bidders $n$, $\gamma(F,n)\leq1$. Moreover
if the value $v$ is bounded, then
\begin{equation}
\lim_{n\rightarrow \infty} \gamma(F,n) = 1
\end{equation}
\end{theorem}
\begin{proof}
Let us denote by $Y^{(k)}$ the maximum of $k$ random variables drawn with the distribution $F$. One just need to observe that
\begin{equation}
\label{eq:new-gamma-def}
\gamma(F,n) = 1 -\frac{\mathbb{E}(v-Y^{n-1}|v>Y^{n-1})}{v_1^{(n)}}.
\end{equation}
\end{proof}
This can be interpreted in the following way: when there is a great number of bidders, the competition is such that the payment tends to become first price.
\begin{remark}
There may be a nonsymmetric equilibrium.
\end{remark}
The basic idea is that one bidder can bid higher than necessary to force another bidder to leave the auction because his (CPA) constraint cannot be satisfied.
\begin{proof}
We concentrate on exhibiting a counter-example.
Take $n=2$, $T=1$, $F$ is the uniform distribution over $[2,3]$, and denote by $(\alpha_i)_{i=1,2}$ the bid multipliers of the two buyers.
Set $\alpha_2 =6$. Then if $\alpha_1\leq 4$, buyer 1 does not win any auction, but the (CPA) constraint is satisfied.
On the other hand, if $\alpha_1 > 4$, then the cost vs value ratio is bigger than 4, and the (CPA) constraint is violated.
We can check that if $\alpha_1 =0$ the (CPA) constraint is satisfied for $2$.
Therefore $(\alpha_1,\alpha_2) =(0,6)$ is an asymmetric Nash equilibrium.
\end{proof}
In addition to this, observe that the buyers may be tempted to bid other strategies than the linear best reply.
For example consider this example:
Take $n=2$, an exponential distribution with parameter 1, $T=1$.
One buyer may increase its profit by bidding with an \emph{affine function}.
Compare $(b_1(v_1),b_2(v_2)) =(2v_1 +1,2v_2) $ with $(b_1(v_1),b_2(v_2)) =(3v_1 ,3v_2) $.
On a simulation with $10^8$ auctions, we get in the first case an empirical CPA of
$0.92$, and a revenue of $0.84$ for bidder 1 (resp. $1.0$ and $0.61$ for bidder 2), while in the second case, we get
an empirical CPA of
$1$, and a revenue of $0.75$ for bidder 1 (resp. $1.0$ and $0.75$ for bidder 2).
Those simple, informal examples indicate that the bidders may be tempted to bid aggressively to weaken the other bidders CPA.
\section{Dynamic Bidder Problem}
\label{sec:dynamic}
In practice, the buyer behaviors may differ from the static case:
(1) the buyer can adapt his bidding strategy to the past events,
(2) the linear constraint (CPA) does not reflect the buyer risk aversion,
(3) the benefit the buyer gets from a won auction is stochastic (for example, if he is only interested in clicks or conversions),
(4) from a business perspective, there is a trade-off between CPA and volumes in the buyer mind. This trade-off can be expressed in different ways.
We propose in this section a continuous time optimal control framework to express and study the buyer's problem in a dynamic fashion. The approach combines the advantages of a powerful and mathematically clean expressiveness with theoretical insights and numerical tools.
We refer the reader to \cite{pham2009continuous} for a reference on stochastic control, and to \cite{Falcone2013} for a presentation of the numerical methods involved. The User Guide \cite{bonnans:hal-01192610} provides a hands-on presentation of the topic.
We only model one individual buyer. Yet note that the optimal control formulation is the first building block for the study of the full market dynamics (with mean field games\cite{Carmona2013,gueant2011} or differential games\cite{Isaacs1999}).
Our main discovery in the stochastic case is that (a) the optimal bid is not linear in the value anymore, but (b) we can still reasonably propose a linear bid (at the cost of an approximation to be discussed thereafter).
\subsection{Deterministic Dynamic Formulation}
We use the deterministic case to introduce some notations,
and recover with optimal control tools some of the results we already derived with the static model. Indeed, when we neglect the stochastic aspect of the reward, the problem is very similar to the static one.
We consider a continuous time model. The buyer receives a continuous flow of requests.
The motivation to propose a continuous time model is that it makes the theoretical analysis easier \cite{pham2009continuous,Heymann2018}.
We denote by ${\mathcal{R}}$ the instantaneous revenue and by ${\mathcal{C}}$ the instantaneous cost:
\begin{equation}
{\mathcal{R}}(b) = \mathbb{E} v [b^{-}<b(v)],
\quad
\mbox{and}
\end{equation}
\begin{equation}
{\mathcal{C}}(b) = \mathbb{E} b^{-} [b^{-}<b(v)],
\end{equation}
If we denote by $\tau$ the time horizon, the bidder is now maximizing with respect to $b$
\begin{eqnarray}
\label{eq:MDP}
\int_0^\tau {\mathcal{R}}(b_t){\rm d} t,
\end{eqnarray}
subject to
$\dot{X}_t = \T {\mathcal{R}}(b_t) - {\mathcal{C}}(b_t)$,
$X_\tau\geq 0$ and
$X_0 = x_0$.
The state $X_t$ represents the (CPA) constraint is should be positive at $t=\tau$ for the (CPA) constraint to be satisfied.
We use the Pontryagin Maximum Principle to argue that a necessary condition for an optimal control is that there exists a multiplier $p(t)$ so that
$b(t)$ maximizes $H = p(\T{\mathcal{R}} - {\mathcal{C}}) +{\mathcal{R}}$.
Moreover $\dot{p} = -\frac{\partial H}{\partial x }=0$. We thus recover the main result of the static case: the bid is linear in $v$. Moreover
in the absence of stochastic components, the multiplier is constant.
Observe that in practice, one could use the solution of problem (\ref{eq:MDP}) to implement a Model Predictive Control (MPC), using $x_0$ as the current state and thus get an online bidding engine.
\subsection{Stochastic Formulation}
The number of displays is much higher than the number of clicks/sales, therefore we neglect the randomness of the competition/price over the randomness that comes from the \emph{actions}.
Because of the number of displays involved, we argue that by the law of large number,
the uncertainty on the action outcome can be apprehended a white noise.
We thus add an additional, stochastic term in the revenue:
$ \T\sigma_i{\rm d} W_t$, where $W_t$ is a Brownian motion.
We get that an optimal bid $b$ should maximize (see \cite{pham2009continuous} for a reference)
\begin{equation}
\mathbb{E}_{b^-,v} [b >b^-]((T v - b^-)p + v + \T^2 \sigma^2 M ),
\end{equation}
for some $p$ and $M$.
Since the realization of the action follows a binomial law, $\sigma \propto v(1-v)$.
Assuming $v<<1$, we can approximate it as $v$.
Therefore every infinitesimal terms of the Hamiltonian becomes
$
((T v - b^-)p + v + v M \T^2)[b>b^-]
$
reproducing the previous reasoning we get
\begin{equation}
b^* = v (\T+\frac{M\T^2+1}{p}).
\end{equation}
\paragraph{Conclusion:} Once again, the bid factor approach is justified!
Observe that this conclusion comes from an approximation, (therefore the bid is not strictly speaking optimal for the initial problem), but by doing so, we have reduced the control set from $\R^+\rightarrow\R^+$ to $\R^+$.
This reduction gives access to a well understood approach to solve this kind of problem: the continuous time equivalent of Dynamic Programming.
Observe that our problem definition is incomplete: we need to take into account the constraint on the final state.
We restate this constraint with a penalty $K(X)$.
Using the dynamic programming principle, the buyer problem is equivalent to the following Hamilton-Jacobi-Bellman equation:
\begin{equation}
V_t + \sup_{\alpha} \left(
(\T {\mathcal{R}}(\alpha) - {\mathcal{C}}(\alpha)) V_x+
\frac{\T^2}{2} {\mathcal{R}}(\alpha) V_{xx} +
{\mathcal{R}}(\alpha)
\right) =0
\end{equation}
\begin{equation}
\mbox{ and } \quad V_\tau = K,
\end{equation}
with
\begin{equation}
{\mathcal{R}}(\alpha) = {\mathcal{R}}(v\rightarrow \alpha v) ,
\quad
\mbox{and}
\end{equation}
\begin{equation}
{\mathcal{C}}(\alpha) = {\mathcal{C}}(v\rightarrow \alpha v)
\end{equation}
\subsection{Numerical Resolution}
Our aim here is to illustrate that HJB approaches provide powerful numerical and theoretical tools to model the buyer's preferences. A quantitative comparison of the performance of the stochastic method is out of the scope of this article.
We solve an instance of the problem with the numerical optimal stochastic control solver BocopHJB \cite{bonnans:hal-01192610}.
BocopHJB computes the Bellman function by solving the non-linear partial differential equation (HJB).
Observe that the state is one dimensional, which makes this kind of solver very efficient.
We take $G(b^-) = (b^-)^{a}$ on $[0,1]$
and $v$ uniformly distributed on $[0,1]$.
We get that ${\mathcal{R}}(\alpha) = \frac{\alpha^{a}}{a+2}$ if $\alpha <1$ and
${\mathcal{R}}(\alpha) = \frac{\alpha^{a}}{(a+2)(\alpha^{a+2})} + \frac{1}{2}(1- \frac{1}{\alpha^2})$ otherwise,
similarly, ${\mathcal{C}}(\alpha) = \frac{a\alpha^{a+1}}{(a+1)(a+2)}$ for $\alpha \leq 1$ and
${\mathcal{C}}(\alpha) = \frac{a}{(a+1)(a+2)\alpha}((a+2)\alpha -a -1)$ for $\alpha \geq 1$ (see Figure \ref{fig:R}).
We take $\T=0.8$ and a linear penalty for negative $X_\tau$. The result of a simulated trajectories is displayed in Figure \ref{fig:simulation2}(a).
We see that on this specific instance of the stochastic path, the control is decreased at the end of the time horizon to adapt to the sudden spike of CPA. This is the kind of
behavior we can expect from a stochastic controller. By contrast, a constant multiplier may result in the kind of trajectories displayed on Figure \ref{fig:simulation2} (b).
\begin{figure}%
\centering
\subfloat[a=1]{{\includegraphics[width=7cm]{a_1simulationChecker} }}%
\subfloat[a=5]{{\includegraphics[width=7cm]{a_5simulation_checker} }}%
\caption{${\mathcal{R}}$ and ${\mathcal{C}}$ and empirical $\T$ for two values of $a$ }%
\label{fig:R}%
\end{figure}
\begin{figure}%
\centering
\subfloat[stochastic controller]{{\includegraphics[width=7cm]{newTrajectory.png} }}%
\subfloat[deterministic controller]{{\includegraphics[width=7cm]{constantControl.png}} }%
\caption{Simulation result}%
\label{fig:simulation2}%
\end{figure}
\section{Conclusion}
We have formalized the existence of CPA constraints in the buyers bidding problem.
These constraints are of paramount importance in performance markets, yet they tend to be neglected in the current literature.
We have seen how standard results translate into this context, sometimes with surprising insights.
We also envisioned the dynamic, stochastic setting by combining a first-order approximation with the Hamilton-Jacobi-Bellman approach. This allowed us in particular to derive a numerical method.
This work raises questions that deserve more extended study.
In particular, can we generalize the first three theorems to other constraints, auction rules and settings?
Can the existence of the aggressive behaviors pointed out at the end of Section 5 pose a problem to the market?
One may also want to put under scrutiny the harder, but more realistic case of correlated values.
The dynamic case also deserves a specific study.
\begin{acks}
I would like to thank J\'er\'emie Mary for his helpful comments.
\end{acks}
\bibliographystyle{plainnat}
|
\section{Machine learning context}
\subsection{Vertex labelling problems}
Categorical machine learning tasks may be cast as vertex labelling problems
on large, edge-weighted graphs. Examples include
(1) mapping the pixels of an image to parts of an object depicted in the image,
(2) mapping web sites to functional categories of web services,
(3) discriminating between fraudulent and authentic credit card transactions,
(4) allocating mobile handsets to specific communication channels.
Suppose $V_o$ denotes the objects to be labelled, and $V_c$ the set of labels.
Information available to us is of three kinds:
\begin{enumerate}
\item
\textbf{Partial labelling: }
Some objects in $V_o$ already carry labels.
Express this by an edge $\{v, \phi(v)\}$ between $v \in V_o$ and its label $\phi(v) \in V_c$,
weighted to reflect strength of association.
\item
\textbf{Label equality: }
For some pairs $u, v \in V_o$, a weighted edge $\{u, v\}$ expresses evidence that $u$ and $v$
should carry the same (as yet undeclared) label.
\item
\textbf{Prohibitions: } There are some subsets of $V_o$ called \textbf{forbidden sets}.
Not all vertices in a forbidden set may be assigned the same label.
In data base contexts, for example, such prohibitions occur
when there is evidence that two objects with similar or identical names are actually distinct.
\end{enumerate}
\subsection{Conflicted-supervised learning}
The three kinds of information mentioned above typically come from disparate sources, and
may partially contradict each other. In
Table \ref{t:ml}, we call this \textbf{conflicted-supervised learning}.
The data scientist is asked to perform a \textit{best merge}
of these data sources to arrive at a vertex labelling.
In a semi-supervised learning graph problem, by contrast, few vertices
have been labelled, and the task is to extend this labelling to all vertices in a way
that respects the adjacency structure, as in \cite{zha}. Such label \textbf{propagation} in graphs
can be achieved through random walks, belief propagation, semi-definite programming, etc.;
see references in Darling \& Velednitsky \cite{dar}.
Combinatorial data fusion, as introduced here, is a rigorous approach to conflicted-supervised learning.
By deleting graph edges, break the graph
into components, each of which contains at most one of the label vertices in $V_c$.
If $v \in V_o$ is in a component with a unique $V_c$ vertex then $v$ receives that as its label. For a component without a $V_c$ vertex a new distinct label is created and all vertices within that component receive the new label. Note that the label received by a vertex may not agree
with the vertex's original label, and some label vertices may end up isolated.
The aim is to minimize the total weight of deleted edges of types 1. and 2.,
so to enforce all the prohibitions in 3.
Instead of label propagation, existence of forbidden sets leads to
label \textbf{dissociation}. The approach taken here builds on the machinery of graph cuts.
\begin{table}
\centering
\begin{tabular}{ |c|c| }
\hline
Problem Class & Label Amount \\
\hline
\textsc{Unsupervised Learning} & none \\
\textsc{Semi-supervised Learning} & partial \\
\textsc{Supervised Learning} & all \\
\textsc{conflicted-supervised Learning} & excess $+$ conflict \\
\hline
\end{tabular}
\caption{Combinatorial data fusion is a rigorous approach to conflicted-supervised learning. } \label{t:ml}
\end{table}
\subsection{Example: product categorization as combinatorial data fusion}
For the sake of concreteness, here is an over-simplified example, which we abstract in
Section \ref{s:anomaly}. You are the manager
of an online market place in which various vendors offer products for sale, whose union is a set $V_o$.
Vendors are expected to label each item in $V_o$ with one of a set $V_c$ of exclusive
categories, but some vendors fail to do so, or label incorrectly, perhaps because the categories
are difficult to understand. Set $V:=V_o \cup V_c$ as a vertex set, with two types of weighted edges:
\begin{enumerate}
\item
If vendor $k$ assigns product $v \in V_o$ to category $c :=\phi_k(v)\in V_c$, place weight
$w_k > 0$ on edge $\{v, c\}$. Less reliable vendors receive less weight.
\item
If transaction data indicate that $m$ buyers have bought both $v, v' \in V_o$,
assign weight $w_m' > 0$ to edge $\{v, v'\}$, where $m \mapsto w_m'$ is increasing.
The idea is that common buyers suggest two items belong to the same category.
\end{enumerate}
In exclusive categorization, the forbidden sets are $\{ \{c, c'\}, c \neq c' \in V_c\}$, and
combinatorial data fusion becomes a multiway cut problem, discussed in Section \ref{s:mmwcut}.
When categories are difficult to understand, a forbidden set might take the form $\{c, c', c''\}$,
meaning that the three categories are not identical, but possibly two out of three
may coincide. In summary, the goal is best merge of vendor labelling data with sales transaction
data. The output of the algorithm will be an assignment of each product to a category, possibly different
to the one assigned by the vendor.
\subsection{Forthcoming works}
The present work covers fundamental definitions and functorial properties, and
solution methods when the underlying graph is a tree, or is approximated by a tree.
Future works will cover linear programming relaxations,
and cases where forbidden sets are not given explicitly, but are returned by an
oracle which decides whether a system of linear inequalities has a feasible point.
\section{Problem statement}
\subsection{Circuits of an independence system}
Given a set $V$, and a collection $\mathcal{S}$ of subsets of $V$, we call
$(V, \mathcal{S})$ an \textbf{independence system} if
$U \in \mathcal{S}$ and $W \subset U$ implies $W \in \mathcal{S}$, and $\emptyset \in S$.
Elements of $\mathcal{S}$ are called \textbf{independent sets}.
A subset $W \subset V$ which is not in $\mathcal{S}$ is called a \textbf{dependent set}.
Minimal dependent sets are called \textbf{circuits}.
The collection of circuits of an independence system $(V, \mathcal{S})$ is denoted $\mathbf{C}(\mathcal{S})$.
The next assertion is part of Korte and Vygen \cite[Theorem 13.12]{kor}.
\begin{lemma} \label{l:circuits}
Let $\mathcal{F}$ be a collection of non-empty subsets of $V$, such that $F_1, F_2 \in \mathcal{F}$
and $F_1 \subseteq F_2$ implies $F_1 = F_2$.
Let $\mathcal{S}$ consist of those $U \subset V$ which contain no element of $\mathcal{F}$
as a subset. Then $(V, \mathcal{S})$ is an independence system, whose set of circuits
$\mathbf{C}(\mathcal{S}) = \mathcal{F}$.
\end{lemma}
In practical applications, the most parsimonious specification of an independence system may be its set of circuits.
Circuits are discussed further in Schrijver \cite[vol. B, Section 39.6]{sch}.
\subsection{Two combinatorial structures on the same vertex set} \label{s:2cs}
Let $V$ be a finite vertex set on which two separate combinatorial structures are defined:
\begin{enumerate}
\item
An edge-weighted graph $G = (V,E_0, w)$, with weights $w:E_0\rightarrow (0, \infty)$.
With no loss of generality, assume $G$ is connected\footnote{
If $G$ is not connected, then each graph component may be treated as a separate problem.}.
\item
An independence system $\mathcal{S}$, which includes all singletons\footnote{
This implies every forbidden set has cardinality at least two.
} and all edges in $E_0$ (i.e. $\{ \{v\} | v \in V\} \cup E_0 \subset S$).
The circuits $\mathbf{C}(\mathcal{S}) =: \mathcal{F}$
are called \textbf{forbidden sets}. Call $H:=(V, \mathcal{F})$ the \textbf{forbidden hypergraph},
and call $A:=(V, \mathcal{S})$ the \textbf{allowable hypergraph}.
\end{enumerate}
\begin{figure}
\caption{ \textbf{EXAMPLE: }\textit{A connected graph with sixteen edges and eleven vertices
was generated by a randomized method.
Independent uniform$(0,1)$ random variables were used to assign the edge weights.
The combinatorial data fusion problem, with
forbidden sets $\{v_1, v_2, v_3\}$ and $\{v_5, v_8\}$, was solved by exact search, giving
a subgraph whose edges are shown with solid lines; omitted edges
(total weight $2.648$) are shown as dashed lines.
Components of this subgraph were $\{v_1, v_7, v_8, v_{10}, v_{11}\}$, and $\{v_2, v_3, v_4, v_5, v_6, v_9\}$,
respectively.
The edge $\{v_1, v_3\}$ has maximum weight $0.944421$ out of all sixteen edges,
but is absent from the optimum subgraph.
This example demonstrates that Best-In Greedy Algorithm,
which would have included $\{v_1, v_3\}$ as its first choice, need not deliver an optimum.
}
} \label{f:greedynonopt}
\begin{center}
\scalebox{0.375}{\includegraphics{DataFusion-11vert-X123-58.png}}
\end{center}
\end{figure}
\subsection{Combinatorial data fusion problem}
What we call \textbf{data fusion} occurs between the forbidden hypergraph
$H=(V, \mathcal{F})$, and the weighted graph $G = (V,E_0, w)$. More specifically, the data fusion problem is looking for an optimal, with
respect to $w$, subgraph of $G$ which avoids all the structures that the forbidden hypergraph disallows. To make this precise, three
combinatorial formulations of the same problem are given below.
\subsubsection{Maximum subgraph whose components include no forbidden sets} \label{s:msg}
A subgraph formulation of the problem is as follows.
A subset $E_1 \subseteq E_0$ induces a partition of $V$ into disjoint graph components
$V_1 \cup \cdots \cup V_d$, for some $d \geq 1$.
Take $\mathcal{E}$ to be those subsets $E_1 \subset E_0$, called \textbf{feasible}
subgraphs, such that none of the
induced graph components of $(V, E_1)$ contains a forbidden set.
Then $\mathcal{E}$ is an independence system, but typically not a matroid, as Figure \ref{f:extraconstraintcounterex} shows.
We seek $E_1 \in \mathcal{E}$ to maximize the sum of weights:
\begin{equation} \label{e:mp4is}
\sum_{e\in E_1}w(e).
\end{equation}
This is an instance of what Korte and Vygen \cite{kor} call
the \texttt{maximization problem for independence systems}.
The subgraph formulation leads naturally to a
\textbf{Best-In Greedy Algorithm}: starting with the empty graph on $V$,
insert edges of $E_0$ in decreasing weight order, omitting any edge whose insertion
would create a component containing a forbidden set.
The failure of this algorithm to deliver an optimum solution
is shown in Figure \ref{f:greedynonopt}. For the example in Figure \ref{f:greedynonopt}
the optimal $E_1$ is
$E \setminus \{\{v_1,v_2\}, \{v_1,v_3\},\{v_1,v_5\},\{v_4,v_8\}\}$.
\subsubsection{Optimum coloring of the forbidden hypergraph} \label{s:fhg}
A coloring formulation of the problem is as follows.
A \textbf{hypergraph coloring} of $(V, \mathcal{F})$
means a mapping $\chi: V \to \mathbb{N}$ such that no hyperedge in $\mathcal{F}$ is monochromatic;
in other words, there is no integer $j$ or forbidden set $F$ for which $F \subseteq \chi^{-1}(j)$.
The optimization problem is to find a hypergraph coloring $\chi: V \to \mathbb{N}$
for which the total weight under $w:E_0\rightarrow \mathbb (0, \infty)$ of edges with differently
colored endpoints is minimum. The objective function is
\begin{equation} \label{e:hgcof}
\min_{\chi} \sum_{\substack{
\{u, v\} \in E_0 \\
\chi(u) \neq \chi(v)}
} w(\{u, v\}).
\end{equation}
As the sum of weights of all edges in $E_0$ is fixed, an equivalent objective function is
\begin{equation} \label{e:hgcof_max}
\max_{\chi} \sum_{\substack{
\{u, v\} \in E_0 \\
\chi(u) = \chi(v)}
} w(\{u, v\}).
\end{equation}
For the example in Figure \ref{f:greedynonopt} the optimum coloring is
\[
\chi(v_i) =
\begin{cases}
1 & \textrm{if } i \in \{1,7,8,10,11\}, \\
2 & \textrm{if } i \in \{2,3,4,5,6,9\}.
\end{cases}
\]
Comparing this optimum coloring to the optimum subgraph given in Section \ref{s:msg}
we see that the optimum coloring
assigns a distinct color to each connected component of the optimum subgraph and that the
edges with different colored endpoints are exactly the edges removed from $E$ to form the optimum subgraph.
The coloring formulation leads naturally to a
\textbf{Greedy Coloring Algorithm}, extending the Greedy Graph Coloring in
Korte and Vygen \cite[Section 16.2]{kor}. Order\footnote{
Possibly at random. Or possibly depending on total weight of incident edges,
or on vertex degree in the forbidden hypergraph.
} the vertices as $V:= \{v_1, v_2, \ldots, v_n\}$.
For each vertex $v_i$, let $\mathcal{F}_i \subset \mathcal{F}$ consist of those forbidden sets $F$
for which $v_i$ is the highest numbered vertex; in other words,
$F \ni v_i$ and
\[
F \cap \{v_{i+1}, \ldots, v_n\} = \emptyset.
\]
Start by assigning $\chi(v_1) = 1$.
For $1 < i \leq n$, let $C = \{1, \dots, k \}$ be the current set of colors used and let $C' \subset C$ be those colors $c$ for which there is no $F \in \mathcal{F}_i$ with all the vertices in $F\setminus \{v_i\}$ colored $c$. There are two cases: (1) If $C' = \emptyset$ then assign $\chi(v_i)$ the new color $k+1$; (2) Otherwise assign $\chi(v_i)$ a $k \in C'$ which maximizes the total weight of edges with endpoints of the same color.
This procedure guarantees that no forbidden set will be monochromatic.
This coloring is parsimonious rather than greedy, in the sense that introduction of new
colors is delayed until it is inevitable. When implemented on
the example in Figure \ref{f:greedynonopt}, using the given vertex order, the result is
\[
\chi(v_i) =
\begin{cases}
1 & \textrm{if } i \in \{1,2,4, 5, 6, 7,10,11\}, \\
2 & \textrm{if } i \in \{3,8,9\}.
\end{cases}
\]
\subsubsection{Maximum weight matching of the allowable hypergraph} \label{s:mwm}
A \textbf{matching} of a hypergraph is a collection of
pairwise disjoint hyperedges. A matching is called a \textbf{perfect
matching} if every vertex of the hypergraph is contained in a matching edge. A
matching formulation of the problem is as follows.
Use the edge-weighted graph $G = (V,E_0, w)$ to turn the allowable hypergraph
into a \textbf{weighted hypergraph} $A_w:=(V,\mathcal{S},w)$,
by taking the weight $w(U)$ of $U \in \mathcal{S}$
to be the total edge weight of the subgraph of $G$
induced on the vertex set $U \subset V$.
Let $\mathcal{M}$ denote the set of matchings of the allowable hypergraph.
We seek a matching $M \in \mathcal{M}$ which maximizes the sum of the weights
of the matching hyperedges:
\begin{equation} \label{e:hgmwm}
w(M) := \sum_{U \in M} w(U).
\end{equation}
Since $\{v\} \in \mathcal{S}$ for all $v \in V$ by assumption, and all
such singleton hyperedges have weight zero, any matching of the allowable
hypergraph can be extended to a perfect matching of equal weight. This allows us
to restrict our maximum weight matching problem to perfect matchings.
Let's consider the example in Figure \ref{f:greedynonopt} again.
The optimum perfect matching for this problem is
$\{\{v_1,v_7,v_8,v_{10},v_{11}\},\{v_2,v_3,v_4,v_5,v_6,v_9\}\}$.
Comparing this matching to the optimum coloring given in Section \ref{s:fhg}
we see that the optimum coloring
mono-colors each of the hyperedges of the optimum perfect matching with a distinct color.
The matching formulation leads naturally to a
\textbf{Greedy Matching Algorithm}. Start with the trivial weight
zero perfect matching $M_0$ consisting of all the singleton hyperedges.
Given a (perfect) matching $M_1$ of $A_w$ and $U_1,U_2
\in M_1$ such that $U_1 \cup U_2 \in \mathcal{S}$, the collection of hyperedges
$M_2 := (M_1 \setminus \{U_1,U_2\}) \cup \{U_1 \cup U_2\}$ is a (perfect)
matching of $A_w$ and $w(M_2) \geq w(M_1)$. Call the latter a merge operation, with a weight gain of
$w(M_2) - w(M_1)$.
At each step of Greedy Matching, perform a merge operation with maximum weight gain\footnote{
This is not the same as the Best-In Greedy Algorithm
of Section \ref{s:msg} because the weight of all edges between $U_1$ and $U_2$ is considered
when computing the weight gain.
}. Stop when no further merges are possible.
\subsection{Interchangeability of three formulations}
The \textbf{combinatorial data fusion problem} $(V, E_0, w, \mathcal{F})$
refers to any of the three optimization problems above.
Indeed:
\begin{lemma} \label{l:equiv}
The maximum subgraph (Section \ref{s:msg}), optimum coloring (Section \ref{s:fhg}) and maximum matching (Section \ref{s:mwm})
formulations of combinatorial data fusion are interchangeable in the following sense:
an optimum solution to any one of the formulations yields, in linear time, feasible solutions to the other two problems, with objective
function values equal to the value of the optimum solution to the initial formulation (using the maximization version of the objective function
for the optimum coloring formulation).
\end{lemma}
\textbf{Remark: } An optimum may be called an \textbf{optimum hypergraph coloring}
in the context (\ref{e:hgcof}), a \textbf{maximum subgraph} in the context (\ref{e:mp4is}),
or a \textbf{maximum hypergraph matching} in the context (\ref{e:hgmwm}).
\begin{proof}
Let $G=(V,E_0,w)$,$H=(V,F)$ be an instance of a combinatorial data fusion problem.
Case 1: Suppose that $E_1$ is the edge set of an optimum solution to the subgraph formulation of the problem.
The objective function value of $E_1$ is $\sum_{e \in E_1} w(e)$.
Let $\chi$ be a coloring of $H$ that monocolors the vertices of each connected component of $(V,E_1)$ a distinct color. As $E_1$ was a
feasible solution to the maximum subgraph formulation of the problem, no forbidden set of $H$ is contained within a single component of $(V,E_1)$. Thus $\chi$
is a valid coloring of $H$. Additionally, the objective function value for the maximization version of the optimum coloring formulation for $\chi$ is
the sum of the $E_0$ edges whose ends are in the same component of $(V,E_1)$, which by the optimality of $E_1$ is $\sum_{e \in E_1} w(e)$.
Let $M$ be the partition of $V$ consisting of the vertex sets of the connected components of $(V,E_1)$. As no connected component of $(V,E_1)$
contains a forbidden set, $M$ is a perfect matching of the weighted allowable hypergraph $A_w$ (see Section \ref{s:mwm}). By the
definition of $A_w$, the weight of each hyperedge $m$ in $M$ equals the sum of the weights of edges of $E_0$ with both
ends in $m$. By the optimality of $E_1$, the weight of each hyperedge $m$ in $M$ equals the sum of the weights of edges
of $E_1$ with both ends in $m$ and each edge $e \in E_1$ counts towards the weight of exactly one $m \in M$. Thus the objective function value of
$M$ is $\sum_{e \in E_1} w(e)$.
Case 2: Suppose that $\chi$ is an optimum coloring of the forbidden hypergraph. The value of the maximization version of the objective
function for $\chi$ is
\begin{equation*} \sum_{\substack{
\{u, v\} \in E_0 \\
\chi(u) = \chi(v)}
} w(\{u, v\}),
\end{equation*}
which we will denote by $Z$.
Let $E_1 = \{e \in E | \chi(u) = \chi(v)\}$. By construction the objective function value for $E_1$ in the maximum subgraph formulation is $Z$ as
required.
Since all edges of $(V,E_1)$ join vertices with equal colors, every connected component of $(V,E_1)$ is monochromatic. Since $\chi$ is a feasible
coloring of the forbidden hypergraph it follows that no forbidden set is contained in a connected component of $(V,E_1)$.
Let $M$ be the partition of $V$ consisting of the preimages of the colors of $\chi$. As
$\chi$ is a valid coloring of the forbidden hypergraph, $M$ is a perfect matching of the
weighted allowable hypergraph $A_w$. By the definition of $A_w$, the weight of each hyperedge $m$ in $M$ equals the sum
of the weights of edges of $E_0$ with both ends in $m$. Thus the objective function value of $M$ is $Z$.
Case 3: Suppose that $M$ is an optimum perfect matching of the weighted allowable hypergraph $A_w$. The objective function value for $M$ is
\begin{equation*}
\sum_{m \in M} w(m) =
\sum_{m \in M} \sum_{ \{u, v\} \in E_0 \cap m} w(\{u, v\}) =
\sum_{\substack{
\{u, v\} \in E_0 \\
M(u) = M(v)}
} w(\{u, v\}),
\end{equation*}
where $M(u)$ is the $m \in M$ for which $u \in m$. Denote this objective function value by $Z$.
Let $E_1 = \{ \{u,v\} \in E_0 | M(u)=M(v)\}$. By construction $\sum_{e \in E_1} w(e) = Z$. As all edges of $E_1$ join vertices within an $m \in M$,
if $C$ is a connected component of $(V,E_1)$ then there exists an $m \in M$ such that all vertices of $C$ are in $m$. As $M$ is a perfect matching of
$A_w$ it follows that no forbidden set can be contained in a connected component of $(V,E_1)$.
Form a hypergraph coloring $\chi$ which monocolors each $m \in M$ a distinct color. Since each $m \in M$ is a hyperedge of the allowable hypergraph,
no forbidden set is monocolored by $\chi$. The objective function value of $\chi$ is
\[
\sum_{\substack{
\{u, v\} \in E_0 \\
\chi(u) = \chi(v)}
} w(\{u, v\}) =
\sum_{\substack{
\{u, v\} \in E_0 \\
M(u) = M(v)}
} w(\{u, v\})
=Z
\]
Therefore the result holds in all three cases.
\end{proof}
\subsection{Special case: multi-multiway cut} \label{s:mmwcut}
Let $G = (V,E_0, w)$ be a weighted graph and let $S_1, S_2, \dots,
S_k$ be $k$ sets of vertices. Avidor \& Langberg \cite{avi} define the
\texttt{multi-multiway cut problem} (MMC) as finding a subset $E_2 \subset E_0$ of
minimum total weight such that the $S_i$ are completely disconnected in $(V, E_0 \setminus E_2)$
(i.e. if $u,v \in S_i$ for some $i$ then $u$ and $v$ are
disconnected in $(V, E_0 \setminus E_2)$ ). The \texttt{multiway cut problem} is the
sub-case where $k=1$, meaning that there is only one set of vertices that have to
be pairwise disconnected; see Chapter 8 of Williamson \& Shmoys
\cite{shm} and Chapters 4 and 19 of Vazirani \cite{vaz}.
Dahlhaus et al \cite{dah} proved this problem to be NP-Hard.
Recently Velednitsky \& Hochbaum \cite{vel} showed that
an algorithm which gives a 2-approximation actually gives an optimum
when the graph is 2-stable.
The more familiar \texttt{multicut problem},
discussed in section \ref{s:multicut}, is the sub-case where $|S_i| = 2$ for $1
\leq i \leq k$. This is NP-hard except when $k = 2$, and is
analyzed at length by Williamson \& Shmoys \cite{shm} and Vazirani \cite{vaz}.
Contrast the multi-multiway cut problem with the combinatorial data fusion problem
on the same $G$ where $S_1, S_2, \dots,S_k$ are the forbidden sets. For the latter,
removal of edges needs only to ensure that the $S_i$ are not connected, whereas
in the former each $S_i$ must be completely disconnected.
The multi-multiway cut problem is a special case of the combinatorial data fusion problem where the forbidden sets are all pairs of vertices contained in the $S_i$ (i.e. $\mathcal(F) = \{ (u,v) | u,v \in S_i, 1 \leq i \leq k \}$).
\subsubsection{Multi-multiway cut where the $S_i$ partition $V$}
Of interest is the subclass of the MMC data fusion problems for which the $S_i$
partition the vertex set. For this subclass there exists a surjective mapping
from the vertex set to $k$ types, and the forbidden sets
$\mathcal{F}$ can be reduced to simply the edges $\{v, v'\}$ whose endpoints
have the same type, i.e. $\phi(v) = \phi(v')$. Thus the
forbidden hypergraph is a graph with $k$ components, each of which is a complete
graph on the vertices of a given type. Here we seek to cut edges of minimum total weight so no component
contains more than one vertex of any type.
\textbf{Human resources example: } A collection of ships is awaiting departure at a seaport.
Each ship has a captain, but lacks other officers. The rules say that no ship may contain
more than one exemplar of each of the officer types: navigator, cook, surgeon, carpenter, and purser.
Officers of these types are available at the seaport. The conviviality score between
two officers (of different types) is the number of previous voyages they have taken together.
The task is to hire officers for the ships, so as to maximize the total of
conviviality scores among officers on the same ship. Here vertices are officers,
type is officer type, and the coloring $\chi$ assigns each officer to a ship.
\subsection{Application where forbidden set cardinality exceeds two}\label{s:multicut}
The combinatorial data fusion problem extends multicut in that
forbidden sets of cardinality larger than two are allowed.
Here is an example showing the need for this greater generality.
Bishop and Darling \cite{bis} consider sporadic observation of a moving target.
Take an {\em observation set}
$\mathcal{O}$, which consists of a collection of pairs
\begin{equation}\label{e:reports}
\mathcal{O}=\left\{(t_1, K_1), (t_2, K_2), \ldots, (t_{n}, K_{n})\right\}
\end{equation}
where $0 \leq t_1 < t_2 < \cdots < t_{n} \leq T$ are the {\em observation times}, and
each $K_i$ is a convex polygon ${\mathbf R}^d$ under the Manhattan metric.
In that case, membership in $K_i$ can be checked using a finite system of linear inequalities,
as can the non-emptiness of the set
\begin{equation} \label{e:convex}
\Lambda_B(\mathcal{O})= \left\{(\mathbf{y}_1,
\mathbf{y}_2, \ldots,\mathbf{y}_n) \in K_1
\times K_2 \times \cdots \times K_n:
\max_{1\leq j\leq n-1 }
\frac{\|\mathbf{y}_{j+1} -\mathbf{y}_j\|}{t_{j+1} - t_j} \leq B \right\}
\end{equation}
where $B$ is a positive speed bound. To say that $\Lambda_B(\mathcal{O})$ is
empty is equivalent to saying that a traveler with maximum speed $B$ could not
have visited zone $K_i$ at time $t_i$ for every $i$, in which case we call the
observation set \textbf{incompatible} with $B$; otherwise it is \textbf{compatible}.
Evidently the collection of subsets $X \subset \mathcal{O}$ such that $X$ is compatible with $B$
is an independence system, and its circuits can be taken as the forbidden sets $\mathcal{F}$ in
a combinatorial data fusion problem.
The enumeration of such circuits is the
problem addressed by Boros et al \cite{bor}.
For example, a forbidden set of size three
consists of three observations $(t_1, K_1), (t_2, K_2), (t_{3}, K_{3})$ which are
incompatible with $B$, while every subset of size two is compatible.
See Figure \ref{f:3incompatible} for a planar example under the Euclidean metric.
\begin{figure}
\caption{ \textbf{CIRCUIT OF SIZE THREE: }
\textit{
A traveler is observed in the left diamond $K_0$ at time $t=0$, the
lower diamond $K_1$ at time $t=1$, and the right diamond $K_2$ at time $t=2$. The
two shorter dashed arrows each indicate the unique $\ell_1$ or $\ell_2$
paths between respective diamonds that can be traveled in one time unit.
The long dashed arrow is the only possible path from $K_0$
to $K_2$ in two time units.
All the pairs $\{(i, K_i), (j, K_j)\}$, for $0 \leq i < j \leq 2$,
are compatible with this speed bound, but the triple
$\{(0, K_0), (1, K_1), (2, K_2) \}$ is incompatible.
}
}\label{f:3incompatible}
\begin{center}
\begin{tikzpicture}[scale=1.5]
\draw [color=blue, fill] (2.5,-1) -- (1.5,-2) -- (2.5,-3) -- (3.5,-2) -- (2.5,-1);
\node at (2.5,-2) {\footnotesize{$t = 1$}};
\draw [color = red, thick, dashed, ->] (1.5,0) -- (2.5,-1);
\draw [color=green, fill] (4.5,-1) -- (5.5,0) -- (6.5,-1) -- (5.5,-2) -- (4.5,-1);
\node at (5.5,-1) {\footnotesize{$t = 2$}};
\draw [color = blue, thick, dashed, ->] (3.5,-2) -- (4.5,-1);
\draw [color=pink, fill] (1.5,0) -- (2.5,1) -- (1.5,2) -- (0.5,1) -- (1.5,0);
\node at (1.5,1) {\footnotesize{$t = 0$}};
\draw [color = pink, thick, dashed, ->] (2.5,1) -- (4.5,-1);
\end{tikzpicture}
\end{center}
\end{figure}
At least two types of combinatorial data fusion problem appear in this context.
\subsubsection{Maximum set of compatible observations}
Suppose we are sure that all observations refer to the same traveler, but
some observations must be wrong, because $\mathcal{O}$ itself is not compatible with $B$.
Construct an edge-weighted star
$(V, E, w)$, with the $\{(t_i, K_i), 1 \leq i \leq n\}$ as leaf vertices, and an auxiliary
vertex $v_o$ of degree $n$. The weight $w_i$ on the edge $\{v_o, (t_i, K_i)\}$
measures trust in the reliabity of the observation, in some additive way.
The forbidden hypergraph $(V, \mathcal{F})$ consists of incompatible circuits
mentioned above. The combinatorial data fusion problem consists of finding
a set of compatible observations, namely those in the component of $v_o$,
of maximum total weight.
\subsubsection{Distinguishing travelers}
Tag each $(t_i, K_i)$ with a label $a_i \in V'$, where $V'$ is a set of
name strings. We believe $(t_i, K_i, a_i)$ is correct, but we do not know whether the name strings
$V'$ all refer to the same traveler, or to several different travelers. Create a connected
edge-weighted graph $(V', E, w)$, where the weight $w({u,v})$ on a pair of name strings $u, v$
represents the strength of external evidence that $u$ and $v$ refer to the same person,
such as similarity between the two strings. The forbidden sets $\mathcal{F}$
are the circuits in the independence system, whose elements are those $U \subset V'$
such that $\{(t_i, K_i), a_i \in U\}$ forms an observation set compatible with $B$.
In an optimum hypergraph coloring $\chi:V' \to \mathbb{N}$
for the combinatorial data fusion problem $(V', E, w, \mathcal{F})$,
name strings with the same color
refer to observations which could refer to the same traveler.
The number of distinct colors is the number of distinct travelers.
As we shall see in Section \ref{s:mincolnotopt},
this optimum need not correspond to a minimum number of distinct
travelers compatible with the observations. In such a context $V'$ could be small, while
$n$ could be large; hence determining the forbidden sets may be much more work than
solving the combinatorial data fusion problem itself.
\section{Category theory perspective on combinatorial data fusion }
\subsection{Category of partitioned sets}
Carlsson and M{\'e}moli \cite{car} propose that a clustering algorithm should be a functor
from a suitable category of finite metric spaces
into a suitable category of partitioned sets\footnote{
Carlsson and M{\'e}moli call this the category of outputs of standard clustering schemes.}.
Their categorical definitions are designed to achieve the following algorithmic property.
Suppose $\phi:X \to Y$ is an injective, non-expansive map between finite metric spaces.
Such a map could arise, for example, when a data set $X$ is augmented by additional data points.
Suppose points $x, x' \in X$ are assigned to the same cluster when a specific clustering algorithm
is applied to $X$; then $\phi(x), \phi(x') \in Y$ should be in the same cluster when the same algorithm is applied to $Y$.
Use the term \textbf{partitioned set} to denote a pair $(X, \mathcal{P})$,
where $X$ is a set, and $\mathcal{P}$ is a partition of $X$. View $\mathcal{P}$ as a collection
of subsets $\{X_\alpha \}_{\alpha \in A}$ of $X$ which we call the
\textbf{blocks} of the partition.
Define the \textbf{category of partitioned sets} as follows.
Objects are partitioned sets $(X, \mathcal{P})$ with $X$ finite.
The morphisms from $(X, \mathcal{P})$ to $(Y, \mathcal{Q})$ consist of set mappings $\phi:X \to Y$
with the property\footnote{
Equivalent to: $\mathcal{P}$ is a refinement of the pullback of $\mathcal{Q}$ under $\phi$.
}: if points $x, x' \in X$ are in the same block of $\mathcal{P}$,
then $\phi(x), \phi(x') \in Y$ are in the same block of $\mathcal{Q}$.
Carlsson and M{\'e}moli's program inspired us to ask whether
there exists some notion of morphism between a pair of combinatorial data fusion problems,
such that data fusion algorithms are functors into the category of partitioned sets.
\subsection{Morphisms of graphs and of independence systems}
Suppose $(V, E)$ and $(V', E')$ are graphs, while $(V, \mathcal{S})$ and
$(V', \mathcal{S}')$ are independence systems, also called abstract simplicial complexes.
\begin{definition}
An injective map $\phi:V \to V'$ is called a \textbf{graph morphism} if $\{\phi(u), \phi(v) \} \in E'$
whenever $\{u, v\} \in E$, and a \textbf{morphism of simplicial complexes} if $\phi(X) \in \mathcal{S}'$
whenever $X \in \mathcal{S}$. If map $\phi:V \to V'$ is injective and has both these properties,
call it a \textbf{weak morphism} of combinatorial data fusion problems:
\[
\phi: (V, E, \mathcal{S}) \to (V', E', \mathcal{S}').
\]
\end{definition}
Observe that the weight functions on $E$ and $E'$ play no part in this definition.
\subsection{Failure of functoriality: counterexamples}
Let $(V, E, w, \mathcal{F})$ and $(V', E', w', \mathcal{F}')$
be a pair of combinatorial data fusion problems,
where $\mathcal{F}$ are the circuits of $\mathcal{S}$, and
$\mathcal{F}'$ are the circuits of $\mathcal{S}'$.
Let $\phi:V \to V'$ be a weak morphism of these combinatorial data fusion problems.
Suppose a partition $V_0, V_1, \ldots, V_{d-1}$ of $V$ forms a maximum weight matching
of $(V, \mathcal{S})$ under $w$, as in Section \ref{s:mwm}. Is it necessarily a refinement of
$\phi^{-1}(V_0'), \phi^{-1}(V_1'), \ldots, \phi^{-1}(V_{f-1}')$,
for some maximum weight matching
$V_0', V_1', \ldots, V_{f-1}'$ of $(V', \mathcal{S}')$ under $w'$,
by analogy with Carlsson and M{\'e}moli's paradigm?
The counterexamples of Figures \ref{f:inducedcounterex} and \ref{f:extraconstraintcounterex}
show two important kinds of mis-match between two such combinatorial data fusion problems,
linked by a weak morphism.
\begin{enumerate}
\item
Suppose $(V, E)$ is a \textbf{vertex-induced subgraph} of $(V', E')$, which means that $E$ is precisely
the edges in $E'$ with both endpoints in $V$. Suppose $\mathcal{F}' = \mathcal{F}$,
so the only forbidden sets in $V'$ are the ones in $V$. Certainly the inclusion map
$\iota: V \to V'$ is a weak morphism.
Figure \ref{f:inducedcounterex} shows how, even with a single forbidden set $\{v_1, v_2, v_3\}$,
the unique maximum weight matching of $(V, \mathcal{S})$ is not a refinement of the
unique maximum weight matching of $(V', \mathcal{S}')$.
\begin{figure}
\caption{ \textbf{INDUCED SUBGRAPH COUNTEREXAMPLE: }
\textit{An edge-weighted graph with five edges on $\{v_1, \ldots, v_5\}$ is shown
on the left. The only forbidden set is $\{v_1, v_2, v_3\}$.
In the optimum solution, edge $\{v_2, v_3\}$ is discarded, and the
maximum weight matching of the allowable hypergraph is
$\mathcal{C}':= \{\{v_3, v_4\}, \{v_1, v_2, v_5\} \}$.
For the induced subgraph on vertices $\{v_1, v_2, v_3\}$, shown on the right,
the maximum weight matching partitions the vertices into
$\mathcal{C}:= \{ \{v_1\}, \{v_2, v_3\} \}$.
Notice that $v_2$ and $v_3$ are in the same block of $\mathcal{C}$,
but in different blocks of $\mathcal{C}'$. So $\mathcal{C}$ is not a refinement of $\mathcal{C}'$.}
} \label{f:inducedcounterex}
\begin{center}
\scalebox{0.4}{\includegraphics{CounterExample-InducedSubgraph.png}}
\end{center}
\end{figure}
\item
Take the graphs $(V, E, w)$ and $(V', E', w')$ as identical, but $\mathcal{S} \subsetneq \mathcal{S}'$.
From the forbidden sets perspective, $\mathcal{F} \supsetneq \mathcal{F}'$
In Figure \ref{f:extraconstraintcounterex}, $|\mathcal{F}|=2$ (right side), while
$|\mathcal{F}'|=1$ (left side). The unique maximum weight matching of $(V, \mathcal{S})$
is not a refinement of the
unique maximum weight matching of $(V', \mathcal{S}')$.
\begin{figure}
\caption{ \textbf{EXTRA CONSTRAINT COUNTEREXAMPLE: }
\textit{An edge-weighted graph with six edges on $\{v_1, \ldots, v_6\}$ is shown.
When $\{v_1, v_2, v_5\}$ is the only forbidden set,
one edge is discarded (shown dashed), and the maximum weight matching of $(V, \mathcal{S}')$ is
$\mathcal{C}_1:= \{\{v_1, v_2\}, \{v_3, v_4, v_5, v_6\} \}$.
If a second forbidden set $\{v_4, v_5\}$ is added,
then two edges are discarded, and the maximum weight matching of $(V, \mathcal{S})$ is
$\mathcal{C}_2:= \{ \{v_5, v_6\}, \{v_1, v_2, v_3, v_4\} \}$.
Notice that $\mathcal{C}_2$ is not a refinement of $\mathcal{C}_1$.}
\textbf{MATROID COUNTEREXAMPLE: }
\textit{The five solid edges on the left, and the four solid edges
on the right, are maximal elements of the independence system $\mathcal{E}$ defined
in Section \ref{s:msg}, when $\{v_1, v_2, v_5\}$ is the only forbidden set; since
$\mathcal{E}$ contains maximal elements of different sizes, it is not a matroid.
}
} \label{f:extraconstraintcounterex}
\begin{center}
\scalebox{0.4}{\includegraphics{CounterExample-6Vertices.png}}
\end{center}
\end{figure}
\end{enumerate}
\section{Number of colors in an optimum coloring}
\subsection{Goal}
Before studying algorithms, let us take a look at the space of optimum solutions.
In this section we shall study the number of colors that are required in an optimal coloring.
The number of colors (or equivalently the number of connected components in a subgraph,
or hyperedges in an optimal perfect matching) does not appear in the combinatorial data fusion objective function.
However, monochromatic $G$ edges do appear in the objective function and the more colors used in a coloring of $H$
the harder it is to maintain monochromatic edges in $G$. Additionally, some algorithmic techniques put a limit on the
number of allowable colors and/or have the number of colors appear in their run time.
\subsection{Lack of transversality}
It is natural to wonder about the interaction between the forbidden sets of a combinatorial data fusion problem,
and the color classes of an optimal solution. For example,
when every forbidden set has size at least $k$, and the number of colors in
an optimum hypergraph coloring does not exceed $k$, must
each color necessarily have a representative vertex in each forbidden set?
This is tautologically true if $k = 2$, but false in general for $k \geq 3$.
Figure \ref{f:nosdr} offers a counterexample.
Even when the forbidden sets are pairwise disjoint, it is possible that
a maximum weight matching of the allowable hypergraph partitions
the vertices into blocks, none of which contains a set of
distinct representatives\footnote{Such a set is called a transversal.} of all the forbidden sets.
\begin{figure}
\caption{ \textbf{TRANSVERSALITY AND 2-COLORING COUNTEREXAMPLE: }
\textit{An edge-weighted graph with 18 edges on $\{v_0, \ldots, v_9\}$ is shown.
The forbidden sets are $\{v_1, v_4, v_7\}$, $\{v_2, v_5, v_8\}$, and $\{v_3, v_6, v_9\}$,
shown in different colors.
The maximum weight matching of the allowable hypergraph partitions the vertices into blocks
$\{v_0, v_2, v_6, v_9\}$, $\{v_3, v_4, v_7\}$, $\{v_1, v_5, v_8\}$, none of which
intersects all three forbidden sets. The forbidden hypergraph is 2-colorable,
for example by mapping $v_1, v_2, v_3$ to color 0, and other vertices to color 1,
but the optimum coloring uses 3 colors.}
} \label{f:nosdr}
\begin{center}
\scalebox{0.45}{\includegraphics{CounterExample-Transversality.png}}
\end{center}
\end{figure}
\subsection{Solution in the case of a single forbidden set}
Suppose there is a single forbidden set $F \subset V$ for the weighted graph $(V,E_0,w)$
and that $|F| = k \geq 2$.
For each of the distinct non-empty subsets $K \subset F$ of size not exceeding $k/2$,
we specialize the combinatorial data fusion problem for $(V, E_0, w, \{F\})$
to the case where the solution has two graph components, one of which contains $K$,
while the other contains $F \setminus K$. In the latter case, solve the
Minimum $\{s,t\}$ Cut problem on an augmented graph with vertex set $V \cup \{s, t\}$ where a source
node $s$ is joined to each $v \in K$ by an infinite weight edge, and a sink node $t$
is joined to each $v' \in F \setminus K$ by an infinite weight edge.
Korte \& Vygen \cite[Ch. 8]{kor} describe the classical algorithms of Ford \& Fulkerson, and others,
for the equivalent Maximum Flow problem on the augmented graph. Having solved the problem to find a
Minimum $\{s,t\}$ Cut, we associate with $K$ the total weight of the edges in the Minimum Cut
between $s$ and $t$. Finally choose the $K$ for which this weight is least, and declare
the edges outside this Minimum Cut to be an optimum solution.
We claim that:
\begin{proposition} \label{p:onef}
When there is only a single forbidden set $F$, with cardinality $|F| = k$,
the combinatorial data fusion problem $(V, E_0, w, \{F\})$ can be solved
by picking the least costly of $2^{k-1} - 1$
Minimum $\{s,t\}$ Cut problems on $V \cup \{s, t\}$.
\end{proposition}
Proposition \ref{p:onef} is established by the argument above, together with the fact that
an optimum solution for $(V, E_0, w, \{F\})$ has exactly two components, which we now prove.
\begin{lemma} \label{l:oneforb}
When there is only one forbidden set, a maximum subgraph (Section \ref{s:msg}) for a
combinatorial data fusion problem $(V, E_0, w, \{F\})$ has exactly two components.
\end{lemma}
\textbf{Remark: } See Theorem \ref{t:numcolorsmax} for a stronger result.
\begin{proof}
Since $F \subset V$ is forbidden, a maximum subgraph cannot have one component.
We prove that a subgraph with more than two components cannot be optimal.
Suppose that $E_1 \subset E_0$ and that $F$ is not connected in $G_1 = (V,E_1)$. If $G_1$ has more that $2$ connected components then let $C_0, C_1, \dots, C_k$ be the connected components of $G_1$ ordered such that $C_{1}$ and $C_2$ both intersect $F$. As $G=(V,E_0)$ is connected there must exist an edge $e \in E_0 \setminus E_1$ with exactly one endpoint in $C_0$. Let $E'_1 = E_1 \cup \{e\}$ and $G'_1 = (V,E'_1)$. The graph $G'_1$ can be formed from $G_1$ be adding the edge $e$. The effect of adding $e$ to $G_1$ will be to merge $C_0$ with one of $C_1$, \dots, $C_k$. Thus the vertices of $C_1$ and $C_2$ remain in different components of $G'_1$ and we have that $F$ is not connected in $G'_1$. Therefore $E'_1$ is a feasible solution with a larger objective function value than $E_1$. It follows that every optimal solution has exactly two components.
\end{proof}
\subsection{Upper bound on optimum number of colors}
Here is a stronger form of Lemma \ref{l:oneforb}. Pike et. al. \cite{pik} proved
\begin{theorem} \label{t:numcolorsmax}
When a combinatorial data fusion problem has $|\mathcal{F}|=b$
forbidden sets, then there exists an optimum coloring of
the forbidden hypergraph $H=(V,\mathcal{F})$ which uses at most $t$ colors, where $2b \geq t^2 - t$.
\end{theorem}
\textbf{Examples: } Two colors suffice for $|\mathcal{F}|=2$; three colors suffice for $|\mathcal{F}| \in \{3, 4, 5\}$.
We shall elaborate on the proof from \cite{pik}.
\begin{proof}
Let $t$ be the minimum number of colors needed for an optimal solution and let $\chi$ be an optimal coloring on the
$t$ colors $\{1,2,\dots, t\}$. Let $\chi_i$ denote the set of vertices which receive color $i$.
Suppose that there exists a pair of distinct colors $i$ and $j$ such that the set $\chi_i \cup \chi_j$ does not contain
a forbidden set. Then the coloring $\chi'$ equal to $\chi$ with colors $i$ and $j$ merged into a single color would be a
valid coloring of $H$. By construction the objective function value of $\chi'$ must be at least that of $\chi$.
Thus $\chi'$ would be an optimal coloring of $H$ on fewer colors than $\chi$, a contradiction.
Therefore for every pair of distinct colors $\{i,j\}$ there exists a forbidden set $f_{i,j}$ contained in $\chi_i \cup \chi_j$. As $\chi$ is a valid
coloring of $H$, no $\chi_i$ contains a forbidden set. This implies that the $\binom{t}{2}$ forbidden sets $f_{i,j}$ are distinct
and thus $b \geq \binom{t}{2}$ as required.
\end{proof}
\begin{figure}
\caption{ \textbf{TWO FORBIDDEN SETS: }
\textit{This illustrates the $b=2$ version of Theorem \ref{t:numcolorsmax}.
Cut 1 disconnects forbidden set $F$, while cut 2 disconnects forbidden set $F'$. This gives a 3-coloring.
However we may merge the color class above cut 2 with the color class to the left of cut 1, giving cut 0, shown on the right,
which is equivalent to a 2-coloring. Cut 0 maintains feasibility, since $F$ and $F'$ both remain disconnected,
while not increasing the weight of cut edges in the graph.
}
} \label{f:2col42fs}
\begin{center}
\scalebox{0.5}{\includegraphics{2Colorsfor2ForbiddenSets.png}}
\end{center}
\end{figure}
The following Corollary, which uses the technique of Theorem \ref{p:onef},
extends the well known result that optimum multicut with two
source-sink pairs can be computed in polynomial time. We shall return to the case of
two forbidden sets in Section \ref{s:2forbsetsgreedy}.
\begin{corollary}\label{c:2forbset}
When a combinatorial data fusion problem has exactly two forbidden sets,
say $\mathcal{F}:=\{F,F'\}$,
an optimum coloring of $(V,\mathcal{F})$, with two colors, can be found using
$p$ Minimum $\{s,t\}$ Cut computations on graphs with vertex set $V \cup \{s, t\}$
and $|F| + |F'|$ additional edges,
where $p$ is the number of partitions $\mathcal{P}_2$ of $F \cup F'$ into two subsets,
both of which non-trivially intersect $F$ and $F'$.
\end{corollary}
\begin{proof}
Theorem \ref{t:numcolorsmax} shows that there is an optimum which is a 2-coloring.
Such an optimum partitions $V$ into $V_0 \cup V_1$, where $V_i \cap F \neq \emptyset$
and $V_i \cap F' \neq \emptyset$, for $i = 0, 1$.
To find such a partition, iterate through all partitions $\{A, B\} \in
\mathcal{P}_2$; for each we connect a source $s$ to all vertices in $A$, and a sink $t$
to all vertices in $B$, where the new edges have infinite weight,
and find a Minimum $\{s,t\}$ Cut in the augmented graph on vertex set $V \cup \{s, t\}$.
Having found a pair
$\{A, B\} \in \mathcal{P}_2$ whose cut has minimum weight, set $V_0\supset A$ to
be all vertices of $V$ in the same component as $s$, and set $V_1 \supset B$ to
be all vertices of $V$ in the same component as $t$, respectively.
\end{proof}
\subsection{Minimum number of colors need not give optimum coloring} \label{s:mincolnotopt}
The number of colors used in an optimum hypergraph coloring
of a combinatorial data fusion problem may exceed the chromatic number
of the forbidden hypergraph. Figure \ref{f:nosdr} shows one example where
the forbidden hypergraph is 2-colorable, but the optimum hypergraph coloring
uses three colors. Here is a simpler example of the same phenomenon.
Consider the path of length three shown in Figure \ref{f:path}, with forbidden sets
$\{v_1, v_2\}, \{v_1,v_3,v_4\}$, and $\{v_2,v_3,v_4\}$. Two of the edges have weight 1, while
the bottom edge has weight $b > 2$.
Different colors must assigned to $v_1$ and $v_2$, since $\{v_1, v_2\}$ is a forbidden set; without
loss of generality take $\chi(v_1)=0$ and $\chi(v_2)=1$. The optimum 2-coloring has
$\chi(v_4)=0$ and $\chi(v_3)=1$, shown on the lower left, with cost $b$.
The optimum 3-coloring has $\chi(v_4)=2 = \chi(v_3)$ shown on the lower right, with cost $2$.
Hence the best 2-coloring is $b/2$ times more expensive than the best 3-coloring,
and this factor can be made arbitrarily large.
\begin{figure}
\caption{\textit{Example where the best 2-coloring is arbitrarily more expensive than the best 3-coloring}}\label{f:path}
\begin{center}
\begin{tikzpicture}[scale=1.75]
\node (v1) at (-1,1){};
\node (v2) at (1,1){};
\node (v3) at (1,-1){};
\node (v4) at (-1,-1){};
\fill (v1) circle (0.01) node [above left] {$v_1$};
\fill (v2) circle (0.01) node [above right] {$v_2$};
\fill (v3) circle (0.01) node [below right] {$v_3$};
\fill (v4) circle (0.01) node [below left] {$v_4$};
\draw (v2) -- (v3)--(v4) -- (v1);
\node at (1.2,0) {\tiny{$1$}};
\node at (0,-0.8) {\tiny{$b$}};
\node at (-1.2,0) {\tiny{$1$}};
\end{tikzpicture}
\begin{tabular}{l r}
\begin{tikzpicture}[scale=1.75]
\node (v1) at (-1,1){};
\node (v2) at (1,1){};
\node (v3) at (1,-1){};
\node (v4) at (-1,-1){};
\fill [color = red] (v1) circle (0.01) node [above left] {$v_1$};
\fill [color = blue] (v2) circle (0.01) node [above right] {$v_2$};
\fill [color = blue] (v3) circle (0.01) node [below right] {$v_3$};
\fill [color = red](v4) circle (0.01) node [below left] {$v_4$};
\draw (v1) -- (v4);
\draw (v2) -- (v3);
\node at (1.2,0) {\tiny{$1$}};
\node at (-1.2,0) {\tiny{$1$}};
\end{tikzpicture} &
\begin{tikzpicture}[scale=1.75]
\node (v1) at (-1,1){};
\node (v2) at (1,1){};
\node (v3) at (1,-1){};
\node (v4) at (-1,-1){};
\fill [color = red] (v1) circle (0.01) node [above left] {$v_1$};
\fill [color = green] (v2) circle (0.01) node [above right] {$v_2$};
\fill [color = blue] (v3) circle (0.01) node [below right] {$v_3$};
\fill [color = blue](v4) circle (0.01) node [below left] {$v_4$};
\draw (v4) -- (v3);
\node at (0,-0.8) {\tiny{$b$}};
\end{tikzpicture}\\
1. \textit{Two color solution, cut cost $b>2$} & 2. \textit{Three color solution, cut cost 2}\\
\end{tabular}
\end{center}
\end{figure}
\textbf{Open Problem: } \textit{Under what conditions on a
combinatorial data fusion problem $(V, E, w, \mathcal{F})$
does an optimum hypergraph coloring of $(V,\mathcal{F})$ have exactly two colors?}
\textbf{Remark: }
The literature on the question of when a hypergraph is 2-colorable includes
Seymour \cite{sey} and McDiarmid \cite{mcd}. The latter proves, using the Lov\'{a}sz Local Lemma,
that a hypergraph, where every hyperedge contains at least $k$ points and meets at most $d$ other hyperedges,
is 2-colorable if $e(d+2) \leq 2^k$.
\section{The combinatorial data fusion problem for trees}
Let us move on to the study of solution methods. Begin with the case where $G$ is a tree.
\subsection{Set cover formulation}\label{s:ilpf}
When $(V, E, w)$ is an edge-weighted tree with positive weights,
the combinatorial data fusion problem $(V, E, w, \mathcal{F})$
may be formulated as a set cover problem, without
requiring the number $d$ of colors to be specified in advance.
Each forbidden set $F$ will induce a linear constraint as follows.
Let $E_F \subseteq E$ denote
the union, over all $v, v' \in F$, of the edges on the unique path from $v$ to $v'$ in the tree $(V, E)$.
The subtree\footnote{
We will revisit this subtree in Definition \ref{d:pathdiscrim}.
} $E_F$ may be constructed in time which is linear in $|E|$:
\begin{enumerate}
\item
Create an auxiliary vertex $v_o$, and auxiliary edges $E_o:=\{\{v_o, u\}, u \in F\}$.
\item
Let $E_F'$ denote the set of edges in the 2-core\footnote{
The maximal subgraph in which all vertices have degree at least $2$.
} of the graph
\[
(V \cup \{v_o\}, E \cup E_o).
\]
Observe that, for $u, u' \in F$, every edge on a path $u$ to $u'$ is part of a cycle
passing through $v_o$, $u$, and $u'$, and hence is in $E_F'$. However an edge not on any
such path will lead to a leaf vertex in $(V \cup \{v_o\}, E \cup E_o)$, and will therefore
be missing from the 2-core.
\item
The desired edge set $E_F$ is given by:
\(
E_F = E_F' \setminus E_o.
\)
\end{enumerate}
\textbf{Constraint Matrix: }Suppose $\{e_1, e_2, \ldots, e_m\}$ are the edges of the tree, and
$\{F_1, F_2, \ldots, F_b\}$ are the forbidden sets.
Encode all $b$ constraints in a $b \times m$ matrix $A:=(a_{i,j})$,
where $a_{i,j} \in \{0, 1\}$, and $a_{i,j} = 1$ precisely when edge $e_j$ belongs to $E_{F_i}$,
in other words when removal of $e_j$ disconnects the vertices in $F_i$.
The combinatorial data fusion problem $(V, E, w, \mathcal{F})$ now reduces to a
minimum weight \textbf{set cover} problem:
we seek a subset $K \subset E = \{e_1, e_2, \ldots, e_m\}$ of minimum weight
so that $K \cap F_i \neq \emptyset$ for every forbidden set $F_i$.
Here is the set cover formulation.
\begin{proposition} \label{p:hitset}
When the graph $(V,E)$ is a tree, an optimum set of deleted edges in the
combinatorial data fusion problem $(V, E, w, \mathcal{F})$
is given by $\{e_j \in E: x_j = 1\}$ in the set cover problem: minimize
\begin{equation} \label{e:hitwt}
\sum_{j = 1}^m w(e_j) x_j
\end{equation}
for variables $x_j \in \{0, 1\}$, over the constraints
\begin{equation} \label{e:hitconstr}
\sum_{j = 1}^m a_{i,j} x_j \geq 1, \quad i = 1, 2, \ldots,b.
\end{equation}
\end{proposition}
\begin{proof}
Take any forbidden set $F$. Under the constraints,
there is some edge $e \in E_F$ for which $x_e = 1$, and that edge is on
the unique path between some pair of vertices $v, v' \in F$. Hence $v$ and $v'$ are not in the same
component of the forest $(V, E_1)$, where $E_1:=\{e \in E: x_e = 0\}$. Therefore $F$ cannot be contained
in a component of $(V, E_1)$. Thus the constraints of the integer linear program precisely
identify the subsets $E_1 \subset E$ such that no component of $(V, E_1)$ contains any forbidden set.
\end{proof}
\subsection{Practical implementation}
\subsubsection{Multicut problem on trees}
Williamson and Shmoys \cite[Exercise 7.2]{shm} gives a 2-approximation algorithm based on Primal Dual Schema
for the case where every forbidden set has cardinality 2, i.e the multicut problem on trees.
\subsubsection{Set cover approximations}\label{s:setcovappr}
The Set Cover problem (\ref{e:hitwt}), (\ref{e:hitconstr}) is treated at length in
Williamson \& Shmoys \cite[Ch. 1, 7]{shm} and in Vazirani \cite{vaz}.
Available algorithms include the following.
\begin{itemize}
\item
The \textbf{Primal-Dual} algorithm gives a $c$-factor approximation, where $c$ is the maximum of
the column sums of the matrix $A$, or in other words the maximum number of forbidden sets which
can be disconnected by the removal of a single edge.
\item
In the \textbf{Greedy Set Cover} algorithm, $J_n \subset \{1,2, \ldots, m\}$ will denote the
indices of the edges deleted after $n$ steps, initialized at $J_0 = \emptyset$.
Stop when every forbidden set has been disconnected. After $n-1$ steps, the next iteration goes as follows:
\begin{enumerate}
\item
For each $j \notin J_{n-1}$, score edge $e_j$ by $w_j$ divided by the number of
connected forbidden sets which will be disconnected if $e_j$ is removed.
\item
Select $e_j$ with minimum score, using random tie-breaking, and set $J_n:=J_{n-1} \cup \{j\}$.
Reduce the collection of connected forbidden sets accordingly.
\end{enumerate}
Taking $c$ as the maximum of the column sums of the matrix $A$ as above,
\cite[Section 1.6]{shm} proves that $\max J_n$ gives an $H_c$-approximation, where
\begin{equation} \label{e:harmonic}
H_c:= \sum_{s=1}^c \frac{1}{s} \approx \log{c}.
\end{equation}
This is a better bound than that of the Primal-Dual algorithm.
\end{itemize}
\subsubsection{Exact solution through linear programming relaxation}
Call a matrix A \textbf{totally unimodular} if each subdeterminant of A is
0, 1, or -1. According to a theorem of Ghouila and Houri proved in \cite[Ch. 5]{kor},
a $b \times m$ integer matrix $A:=(a_{i,j})$ is totally unimodular if and only if, for every
$R \subset \{1, 2, ,\ldots, b\}$, there is a partition $R = R_1 \cup R_2$ such that
\begin{equation} \label{e:tunim}
\sum_{i \in R_1} a_{i,j} - \sum_{i \in R_2} a_{i,j} \in \{-1, 0, 1\}, \quad j = 1, 2, \ldots, m.
\end{equation}
Consider the linear programming relaxation of the integer linear program of Proposition \ref{p:hitset}.
Instead of $x_j \in \{0, 1\}$, take $x_j \geq 0$ and as before
\[
A \mathbf{x} \geq \mathbf{1}
\]
where $\mathbf{1}$ is the vector whose $b$ components are all 1.
\begin{corollary}
Suppose the constraint matrix $A$ satisfies (\ref{e:tunim}), and is therefore totally unimodular.
Then the linear programming relaxation of (\ref{e:hitwt}), (\ref{e:hitconstr}) has an optimum integer solution,
which is optimum for the combinatorial data fusion problem on the tree.
\end{corollary}
\begin{proof}
See \cite[Ch. 5]{kor} for material on total dual integrality when $A$ is totally unimodular.
\end{proof}
\section{Data fusion based on Gomory-Hu trees}
Let us consider solution methods when $G$ is not a tree.
\subsection{Definition of Gomory-Hu tree } \label{s:gomhu}
Recall the definition (Schrijver \cite[Ch. 15]{sch})
of a Gomory-Hu tree $(V, E_T, \omega_T)$ for the weighted graph
$G:=(V, E_0, w)$. This is a tree on the same vertex set as $G$,
with edge weight function $\omega_T$, such that for any
two vertices $v, v' \in V$, the cost $c_{v, v'}$ of a minimum cut
(i.e. the total weight of deleted edges)
between any distinct $v$ and $v'$ in $G$ equals the minimum of $\omega_T(e)$ over edges $e$
along the unique path from $v$ to $v'$ in $(V, E_T)$. Note that
\begin{enumerate}
\item
The tree edges $E_T$ need not be a subset of graph edges $E_0$.
\item
The construction is not canonical. For example, given four vertices $v_1, v_2, v_3, v_4$ in $V$,
there may be one Gomory-Hu tree in which removing the lowest weight edge on the path between $v_1$
and $v_3$ also disconnects $v_2$ from $v_4$, and another Gomory-Hu Tree for the same $G$ in which
it does not.
\end{enumerate}
See Vazirani \cite[Ch. 4.2]{vaz} for a description of how to construct the Gomory-Hu Tree,
and an explanation of its uses in multicut problems.
We tested the \texttt{JGraphT} implementation of \texttt{GusfieldGomoryHuCutTree} on random graphs with
comparable vertex degree distributions, from 90 to 98000 edges. The run time grew at a sub-quadratic rate
in the number of edges.
\subsection{Gomory-Hu tree for example of Figure \ref{f:greedynonopt}}
When a single edge is removed from a Gomory-Hu tree,
the weight of that edge corresponds to the cost (\ref{e:hgcof}) of the 2-coloring
which assigns color 0 to one tree component, and color 1 to the other tree component.
However when two edges $e, e'$ are removed from a Gomory-Hu tree, breaking it into three components,
the cost (\ref{e:hgcof}) of the corresponding 3-coloring could be less than $\omega_T(e) + \omega_T(e')$.
This is because some edges
of $G$ could be counted both towards $\omega_T(e)$ and towards $\omega_T(e')$.
This is illustrated in Figure \ref{f:gomoryhu}, where we remove two edges from a
Gomory-Hu Tree for the graph shown in Figure \ref{f:greedynonopt}.
Here the optimum coloring of the forbidden hypergraph is a
2-coloring (Figure \ref{f:greedynonopt}), whereas a 3-coloring of the
Gomory-Hu tree was needed to disconnect both forbidden sets. The caption explains how two of these colors
can be merged, giving the optimum solution.
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=1.75]
\node (v1) at (1,0){};
\node (v2) at (-1,0){};
\node (v3) at (0,-1){};
\node (v4) at (0,0){};
\node (v5) at (-1,-1){};
\node (v6) at (-2,0){};
\node (v7) at (0,2){};
\node (v8) at (0,1){};
\node (v9) at (0,-2){};
\node (v10) at (1,-1){};
\node (v11) at (2,0){};
\fill (v1) circle (0.01) node [above] {$v_1$};
\fill (v2) circle (0.01) node [below] {$v_2$};
\fill (v3) circle (0.01) node [right] {$v_3$};
\fill (v4) circle (0.01) node [below right] {$v_4$};
\fill (v5) circle (0.01) node [below left] {$v_5$};
\fill (v6) circle (0.01) node [left] {$v_6$};
\fill (v7) circle (0.01) node [above] {$v_7$};
\fill (v8) circle (0.01) node [right] {$v_8$};
\fill (v9) circle (0.01) node [below] {$v_9$};
\fill (v10) circle (0.01) node [below] {$v_{10}$};
\fill (v11) circle (0.01) node [right] {$v_{11}$};
\draw [thick] (v6) -- (v2) -- (v4)--(v1) -- (v11);
\draw [thick] (v7) -- (v8) -- (v4) -- (v3) -- (v9);
\draw [thick] (v1) -- (v10);
\draw [thick] (v4) --(v5);
\node at (-1.5,-0.1) {\tiny{$1.56$}};
\node at (-0.5,-0.1) {\tiny{$2.63$}};
\node at (0.5,-0.1) {\tiny{$2.47$}};
\node at (1.5,-0.1) {\tiny{$0.37$}};
\node at (0.2,-0.5) {\tiny{$3.06$}};
\node at (0.2,-1.5) {\tiny{$1.33$}};
\node at (0.2,0.5) {\tiny{$1.29$}};
\node at (0.2,1.5) {\tiny{$1.31$}};
\node at (-0.5,-0.7) {\tiny{$2.05$}};
\node at (1.2,-0.5) {\tiny{$0.73$}};
\end{tikzpicture}
\end{center}
\caption{
\textbf{Gomory-Hu tree} \textit{computed from the graph of Figure \ref{f:greedynonopt}, whose
forbidden sets are $\{v_1, v_2, v_3\}$ and $\{v_5, v_8\}$.
By inspection, the least costly way to disconnect both forbidden sets in this Gomory-Hu tree
is to remove two tree edges, $\{v_4,v_8\}$ and $\{v_1,v_4\}$.
The total weight of graph edges to be removed seems to be $1.29+2.47= 3.76$, but
the graph edge $\{v_1, v_7\}$ is removed twice, so the true cost is $3.76 - 0.56 = 3.2$.
This gives a 3-coloring of the forbidden hypergraph, of which $\{v_1, v_{10}, v_{11}\}$ is
one color class, and $\{v_7, v_8\}$ is another. These two color classes can be merged, because their
union contains no forbidden set. Thus the graph edge $\{v_1, v_7\}$ is restored, and
the total weight of removed edges drops to $3.20 - 0.56 = 2.64$, which coincides with
the optimal solution for the original graph of Figure \ref{f:greedynonopt}.
So the Gomory-Hu tree algorithm plus a merger yields the optimum solution in this case.
}
}\label{f:gomoryhu}
\end{figure}
\subsection{Single forbidden set: why Gomory-Hu tree succeeds}
The question arises as to whether a Gomory-Hu tree $(V, E_T, \omega_T)$ may offer a better way to
solve the combinatorial data fusion problem for
the weighted graph $(V, E_0, w)$, than Proposition \ref{p:onef},
when there is just one forbidden set $F \subset V$.
When $|F|= 2$, it provides an optimum solution,
by definition.
It would appear that when $|F| \geq 3$, the structure of the Gomory-Hu Tree
disallows some partitions $K, F \setminus K$ from being considered. Suppose $F:=\{v_0, v_1, v_2\}$,
and the optimal partition of $V$ places $v_1$ in one component and
$\{v_0, v_2\}$ in the other. In the specific Gomory-Hu tree which is selected,
the unique path from $v_0$ to $v_2$ may pass through $v_1$. Hence the only partitions
whose cost is visible in this Gomory-Hu tree are (a) $v_0$ in one component,
$\{v_1, v_2\}$ in the other, and (b) $\{v_0, v_1\}$ in one component,
$v_2$ in the other. Despite this objection, Proposition \ref{p:gomhuone}
shows that optimum combinatorial data fusion can be derived from any Gomory-Hu tree.
\begin{definition}\label{d:pathdiscrim}
Given a tree $T:=(V, E)$ and a subset $F \subset V$ with $|F| \geq 3$,
apply the following iterative algorithm:
whenever a leaf vertex is not an element of $F$, delete it and the incident edge.
The terminal state is a tree $T':=(V', E')$, which we call the \textbf{$F$-bounded subtree},
all of whose leaf vertices are elements of $F$. If all elements of $F$ are leaves
of the $F$-bounded subtree, we say that the tree $T$ \textbf{path-discriminates} the
vertex set $F$.
\end{definition}
The important feature of a tree $T$ which path-discriminates $F$ is that
there do not exist three vertices $v, v', v'' \in F$ such that the unique path in $T$ from
$v$ to $v'$ passes through $v''$.
\begin{proposition} \label{p:gomhuone}
Consider a combinatorial data fusion problem on $G:=(V, E_0, w)$ in which there is a
single forbidden set $F$ with $|F| \geq 3$.
Suppose $T:=(V, E_T, \omega_T)$ is any Gomory-Hu tree for $G$, which may or may not
path-discriminate the vertex set $F$, in the sense of Definition \ref{d:pathdiscrim}.
A 2-coloring $\chi \to \{0, 1\}$ obtained as follows is an optimum solution:
\begin{enumerate}
\item
Construct the $F$-bounded subtree $T':=(V', E_T')$ of $T$, whose leaf vertices
are a subset of $F$ by Definition \ref{d:pathdiscrim}.
\item
Identify an edge $e_o \in E_T' \subset E_T$ of minimum weight. By construction, removal of this edge
from $E_T$ breaks the Gomory-Hu tree into two vertex components
$V_0$ and $V_1$, neither of which contains $F$.
\item
Return the 2-coloring $\chi$ with value 0 on $V_0$, and value 1 on $V_1$.
\end{enumerate}
Then $\omega_T(e_o)$ is the total weight of edges of $G$ with differently colored
end points under $\chi$.
\end{proposition}
\begin{proof}
Removal of any $e_o \in E_T' \subset E_T$ partitions $V'$ into two components, both
of which contain at least one leaf vertex of $T'$, and hence at least one element of $F$.
Neither component contains $F$, so the proposed 2-coloring $\chi$ is not constant on $F$,
and is indeed a proper coloring.
By Lemma \ref{l:oneforb}, an optimum solution
of this combinatorial data fusion problem is a 2-coloring of $V$. Hence it suffices to prove that
there is no proper 2-coloring $\chi'$ of $(V, \{F\})$ for which the total weight of edges of $G$
with differently colored end points is less than $\omega_T(e_o)$.
Suppose $\chi'$ is a proper coloring of $(V,\{F\})$ that partitions $V$ into $V_0'$ and $V_1'$.
Let $h(u,v)$ denote the weight of a minimum cut between vertices $u, v \in V$ in the graph $G$,
and consider the value:
\[
h_*: = \min \{h(u, v), \quad u \in F \cap V_0', \quad v \in F \cap V_1'\}.
\]
There are distinct vertices $u_0 \in F \cap V_0'$ and $u_1 \in F \cap V_1'$ for which
$h(u_0, u_1)$ attains this minimum value $h_*$. The vertices $u_0$ and $u_1$ are in the $F$-bounded
subtree\footnote{
$u_0$ and $u_1$ do not need to be leaf vertices in $T'$; the path-discrimination
assumption is unnecessary.
} $T':=(V', E_T')$. Therefore, by the properties of a Gomery-Hu tree, $h(u_0,u_1)$ is the minimum weight of the
edges on the $(u_0,u_1)$-path in $T'$, denote this edge of $T'$ by $e'$.
It follows that
\[
h_* = h(u_0,u_1) = \omega_T(e') \geq \omega_T(e_o),
\]
where edge $e_o \in E_T' \subset E_T$ has minimum weight. However, $h_*$ is also a lower bound on
the total weight of edges of $G$ with differently colored end points under the coloring $\chi'$.
Hence $\chi'$ gives a deleted edge weight at least as large as that of $\chi$, as claimed.
\end{proof}
\subsection{Limitations on power of Gomory-Hu tree in general}
Suppose $\chi:V \to \{1,2,\ldots, t\}$ is an optimum forbidden hypergraph coloring for a combinatorial
data fusion problem $(V, E, w, \mathcal{F})$ where $|\mathcal{F}|=b \geq \binom{t}{2}$;
existence of an optimum $\chi$ with such a range follows from
Theorem \ref{t:numcolorsmax}. Take a Gomory-Hu tree
$T:=(V, E_T, \omega_T)$ for $G:=(V, E, w)$. We would like to compare the cost of discarded edges
in $T$ with those discarded in $G$, as in (\ref{e:hgcof}), i.e. to compare:
\[
C_G(\chi):=\sum_{\substack{
\{u, v\} \in E \\
\chi(u) \neq \chi(v)}
} w(\{u, v\}); \quad
C_T(\chi):=\sum_{\substack{
\{u, v\} \in E_T \\
\chi(u) \neq \chi(v)}
} \omega_T(\{u, v\}).
\]
Proposition \ref{p:gomhuone} ensures that if $b = 1$ then for each Gomery-Hu tree $T$ there exists
an optimal $2$ coloring such that $C_G(\chi) = C_T(\chi)$. However, as the next proposition
shows, the ratio between $C_G(\chi)$ and $C_T(\chi)$ is not bounded even for
the multicut problem ($b \geq1$, but all forbidden sets of size $2$).
\begin{proposition} \label{p:ugcght}
Assuming the Unique Games Conjecture (see \cite[Ch. 13 ]{shm}), there is no
constant $\alpha \geq 1$ such that, for every optimum coloring $\chi$ for
a multicut problem on a graph $G$, there is a Gomory-Hu tree
$T$ for $G$ such that
\begin{equation} \label{e:treecompare}
C_T(\chi) \leq \alpha C_G(\chi)
\end{equation}
unless \textbf{P = NP}.
\end{proposition}
\begin{proof}
Restrict to the multicut problem, meaning that $|F|=2$ for all $F \in \mathcal{F}$.
Let $\chi$ be an optimal coloring of $(V,\mathcal{F})$ with respect to $G$.
Select a Gomory-Hu tree $T:=(V, E_T, \omega_T)$ for $G$, and let
$\psi:V \to \{1,2,\ldots, b\}$ be an optimum coloring of $(V, \mathcal{F})$
with respect to $T$ (i.e. a solution to the combinatorial data fusion problem with $G$ replaced by $T$).
Williamson and Shmoys \cite[Exercise 7.2]{shm} gives a 2-approximation algorithm
for the multicut problem on trees. That means that we may construct
$\psi':V \to \{1,2,\ldots, b\}$ so that
\[
C_T(\psi') \leq 2 C_T(\psi) \leq 2 C_T(\chi) .
\]
The second inequality is valid because $\chi$ is also a coloring (possibly not an optimum)
with respect to $T$. If the inequality (\ref{e:treecompare}) were valid for some $\alpha \geq 1$,
then
\[
C_T(\psi') \leq 2 \alpha C_G(\chi),
\]
In other words, the coloring $\psi'$ gives a $2 \alpha$-approximation to the multicut problem.
According to \cite[Theorem 8.10]{shm}, assuming the Unique Games Conjecture, such an
approximation cannot exist unless \textbf{P = NP}.
\end{proof}
\section{Greedy data fusion algorithm based on Gomory-Hu tree}
Despite the impossibility of constant-factor approximation, demonstrated in Proposition \ref{p:ugcght},
we shall attempt to use the Gomory-Hu tree to approximate combinatorial data fusion.
Experiments reported in Section \ref{s:ghuxpts} show surprisingly good performance.
The following algorithm, which we have implemented in Java on top of the \texttt{JGraphT} library,
combines the idea of Proposition \ref{p:gomhuone} with the Greedy
Set Cover algorithm of Section \ref{s:setcovappr}.
\subsection{Algorithm}\label{s:greedygomhu}
A combinatorial data fusion problem $(V, E, w, \mathcal{F})$ is given.
\begin{enumerate}
\item
Compute a Gomory-Hu tree $T:=(V, E_T, \omega_T)$ for
$G:=(V, E, w)$.
\item
Compute, for each forbidden set $f \in \mathcal{F}$, the edge set $E_f \subset E_T$ of
the $f$-bounded subtree of $T$ (see Definition \ref{d:pathdiscrim}), for example by the method described in Section \ref{s:ilpf}.
Initialize Boolean arrays $(y_f, f \in \mathcal{F})$ at 1, and $(x_e, e \in E_T)$ at zero.
Subsequently $y_f=1$ will mean that forbidden set $f$ is connected, and $x_e = 1$
will mean that edge $e$ has been deleted from $T$.
\item
\textbf{Iterative Splitting: }
Continue until $y_f = 0$ for all $f \in \mathcal{F}$, i.e. all forbidden sets have been disconnected.
\begin{enumerate}
\item
For each undeleted tree edge, i.e. $e \in E_T$ with $x_e = 0$, compute
the number of forbidden sets which will be newly disconnected if $e$ is removed:
\begin{equation} \label{e:phi_e}
\phi(e):=\sum_{f \in \mathcal{F}} y_f 1_{ \{e \in E_f \} }.
\end{equation}
\item
Select (with random tie-breaking) an undeleted $e \in E_T$ for which
the ratio $\omega_T(e)/\phi(e)$ is minimum, as in Greedy Set Cover.
\item
Delete this $e$ by setting $x_e = 1$, and disconnect those $f$ for which $e \in E_f$
by setting $y_f = 0$.
\end{enumerate}
\item \textbf{Color Mergers: }
If $d$ is the number of steps needed to disconnect
every forbidden set, then $d$ edges of the Gomory-Hu tree
have been removed, leaving a forest with $d+1$ components.
This corresponds to a $(d+1)$-coloring $\chi: V \to \{0, 1, \ldots, d\}$ of $(V,\mathcal{F})$.
\begin{enumerate}
\item
Construct an edge-weighted graph $M=(V(M), E(M))$ whose nodes are the color classes:
\[
V(M):= \{C_0, C_1, \ldots, C_d\}; \quad C_j:=\chi^{-1}(j) \subset V.
\]
The weight on an edge $\{C_i, C_j\}$ is defined as the weight of graph edges in
$G:=(V, E, w)$ which would be restored if these two color classes were merged,
and where merging would not monocolor a forbidden set:
\[
\theta(\{C_i, C_j\}):=\sum_{u \in C_i, v \in C_j} w(\{u,v\}).
\]
Take the edge set $E(M)$ to be those pairs of color classes with non-zero weight.
\item
We wish to merge color classes without their merger including a forbidden set.
This is a combinatorial data fusion sub-problem $(V(M), E(M), \theta, \mathcal{C})$
where a collection of colors forms a forbidden set in $\mathcal{C}$ if their merger
would cause some forbidden set in $\mathcal{F}$ to become monochromatic.
\item
For a small number of colors, this sub-problem can be solved
exhaustively, by considering all possible color mergers. If $E(M)$ is too large for exhaustion,
perform color mergers in a greedy way, with those of highest $\theta$ weight first.
\item
If $m+1$ mergers are possible, the final number of tree components is $d-m$.
Apply a different color to each component, giving a $d-m$-coloring of the
original forbidden hypergraph.
\end{enumerate}
\end{enumerate}
We are not yet able to present a satisfactory analysis of the approximation
properties of the algorithm of Section \ref{s:greedygomhu}.
See Table \ref{comparetab} for experimental results.
Here is a weak bound;
\begin{proposition} \label{p:2approx}
In the case of a combinatorial data fusion problem with exactly $b$ forbidden sets,
the greedy algorithm of Section \ref{s:greedygomhu}
provides a coloring whose cost is no more than $b$ times the cost of an optimum coloring.
\end{proposition}
\textbf{Remark: } The approximation for multicut given in \cite[Theorem 8.9]{shm} suggests
that there may exist a $c \log{(b+1)}$-approximation algorithm,
where $c$ is a constant depending on the maximum or minimum cardinality of any forbidden set.
\begin{proof}
Let $C$ denote the cost of an optimum coloring. Let $e_F \in E_T$ be a tree edge which is of minimum cost
among those which disconnect $F$, for each forbidden set $F$. Then
\[
\omega_T(e_F) \leq C, \quad \forall F \in \mathcal{F}
\]
because in particular the optimum coloring disconnects $F$. Suppose the iterative splitting
part of the algorithm deletes tree edges $e_1, e_2, \ldots, e_m$, where necessarily
\[
\sum_{j=1}^m \phi(e_j) = b,
\]
($\phi(e)$ defined as in (\ref{e:phi_e}))
which is the number of forbidden sets to be split. For any $j \in \{1,2,\dots,m\}$, $\omega_T(e_j)$ does not exceed the sum
of those values of $\omega_T(e_F)$ for which forbidden set $F$ is split at step $j$, which is bounded
above by $C \phi(e_j)$.
The total weight of edges discarded by the iterative splitting of the greedy algorithm does not exceed
\[
\sum_{j=1}^m \omega_T(e_j) \leq C \sum_{j=1}^m \phi(e_j) = b C.
\]
The merge phase can only lessen the weight of discarded edges.
Thus the desired factor $b$ approximation is established.
\end{proof}
\subsection{Greedy data fusion: two forbidden sets}\label{s:2forbsetsgreedy}
We already analysed the case of exactly two forbidden sets $\{F, F'\}$ in Corollary \ref{c:2forbset}.
We will use this case to shed light on the greedy algorithm.
Recall that $E_F$ refers to the minimal set of edges of the Gomory-Hu tree which connect the vertices of $F$.
Thus removal of any one edge in $E_F$ disconnects $F$ in the tree.
There are two possibilities.
\begin{figure}
\caption{\textbf{Counterexamples to the optimality of the greedy algorithm when $|\mathcal{F}|=2$: }
\textit{Two examples where the forbidden sets are $F:=\{s, s'\}$ and $F':=\{t, t'\}$.
\textbf{Left: }
Graph itself is a tree. $E_F$ and $E_{F'}$ have an edge in common.The optimum is to remove the edge of
weight 6, which disconnects both $F$ and $F'$. However the greedy algorithm
removes the edges of weight 2 and 5, with total cost 7, which is worse.
\textbf{Upper Right: }Graph on five vertices with six edges.
\textbf{Lower Right: }Its Gomory-Hu tree, in which $E_F$ and $E_{F'}$ are disjoint.
The greedy algorithm on the Gomory-Hu tree removes the tree edges with weights 9 and 10.
Back in the original graph, this leads to a 2-coloring where edges of weight $8 + 9 = 17$
are cut. The optimum is to cut graph edges of weight $7 + 7 = 14$.
}
} \label{f:greedyfailson2fs}
\begin{center}
\scalebox{0.25}{\includegraphics{Counterexample1-2FS.png}}
\end{center}
\end{figure}
\begin{itemize}
\item
\textbf{$E_F$ and $E_{F'}$ are disjoint: } In this case there is no single edge of the
Gomory-Hu tree whose removal disconnects both $F$ and $F'$.
\begin{enumerate}
\item
Figures \ref{f:greedynonopt} and \ref{f:gomoryhu} illustrate one possible outcome of the algorithm.
The two edges removed from the Gomory-Hu tree are $\{v_4,v_8\}$ and $\{v_1,v_4\}$,
each of which disconnects one forbidden set. At the end of the iteration, there are three
components, two of which can be merged, to give the optimum solution illustrated in
Figure \ref{f:greedynonopt}.
\item
The graph on the right in Figure \ref{f:greedyfailson2fs} shows how the greedy algorithm fails.
The greedy algorithm fails to take account of the extra weight of edges (4 or 1 in this case)
which are restored when two of the colors are merged.
\end{enumerate}
\item
\textbf{$E_F$ and $E_{F'}$ have an edge in common: }
Suppose there is some edge of the
Gomory-Hu tree whose removal disconnects both $F$ and $F'$.
Let $e_0 \in E_F$, $e_1 \in E_{F'}$, and $e_2 \in E_F \cap E_{F'}$
each have least tree weight among those in their respective edge sets.
Possibly they are not distinct. Evidently
\(
\max { \{\omega_T(e_0), \omega_T(e_1) \} } \leq \omega_T(e_2).
\)
If
\[
\omega_T(e_2) < 2 \min{ \{ \omega_T(e_0), \omega_T(e_1) \} },
\]
then the criterion of step 3(b) above chooses $e_2$ as the first edge to delete.
There are four cases to consider:
\begin{enumerate}
\item
If $2 \min{ \{\omega_T(e_0), \omega_T(e_1) \} } >
\omega_T(e_2) = \max { \{\omega_T(e_0), \omega_T(e_1) \} }$,
then the deletion
of $e_2$ partitions the vertex set $V$ into two components, neither of which
contains $F$ or $F'$, and this is an optimum 2-coloring as provided by Corollary \ref{c:2forbset}.
\item
If $2 \min{ \{\omega_T(e_0), \omega_T(e_1) \} } > \omega_T(e_2) > \max { \{\omega_T(e_0), \omega_T(e_1) \} }$, then $e_2$ is the first edge
to be deleted by the greedy algorithm, but this need not be optimal (e.g. when $\omega_T(e_0) = \omega_T(e_1)$ and the cuts generated by
$e_0$ and $e_1$ have an edge in common).
\item
If $\omega_T(e_2) > 2 \min{ \{\omega_T(e_0), \omega_T(e_1) \} }$, then the greedy algorithm
will delete one of $e_0$ and $e_1$ followed by the other, possibly giving a non-optimum result, as
in the graph on the left in Figure \ref{f:greedyfailson2fs}.
\item If $\omega_T(e_2) = 2 \min{ \{\omega_T(e_0), \omega_T(e_1) \} }$ then the greedy algorithm will randomly choose to add either $e_2$ or the minimum of $e_0$ and $e_1$ followed by the other.
As with cases 2 and 3, this need not be optimal.
\end{enumerate}
\end{itemize}
\subsection{Empirical performance of the Gomory-Hu-based algorithm} \label{s:ghuxpts}
In our experiments, we made the following choices as to the number, size, and construction of forbidden sets.
\begin{enumerate}
\item The graph $G = (V, E)$ was the giant component of a random graph
sampled uniformly from those with degree sequence $(D_1, \ldots, D_n)$, where
$(D_1 - 1, \ldots, D_n - 1)$ was a Multinomial$(2m - n, (1/n, \ldots, 1/n))$ random vector.
Typically $|V| \approx n$, and $|E| \approx m$.
\item The number of forbidden sets was $b = \lceil\log(n)\rceil$.
\item We selected $a = \lceil 0.75 b\rceil$ ``bad'' vertices at random. The other vertices were ``good'' vertices.
\item Each forbidden set was constructed by selecting one bad vertex uniformly at random, and two good vertices uniformly at random.
The forbidden sets were statistically independent. Since $b > a$,
there were cases where the same bad vertex occurred in multiple forbidden sets.
\end{enumerate}
We tried the the following two algorithms on several examples:
\begin{enumerate}
\item {\bf A greedy algorithm based on Gomory Hu trees} as described in Section \ref{s:greedygomhu}.
\item {\bf An exhaustive search algorithm among all proper 2-colorings} along the lines of Proposition \ref{p:onef}.
Given a combinatorial data fusion problem $(V, E, w, \mathcal{F})$, compute the union $X$ of all the forbidden sets,
and compute all proper\footnote{
A 2-coloring is proper if and only if no $F\in\mathcal{F}$ is monochromatic.
} 2-colorings of $(X, \mathcal{F})$.
For each proper 2-coloring, solve the minimum cut problem in $(V \cup \{s, t\}, E\cup E_0 \cup E_1)$,
where $E_0$ consists of edges between a source $s$ and every vertex in $X$ colored 0, and
$E_1$ consists of edges between a sink $t$ and every vertex in $X$ colored 1.
Output a 2-coloring of X for which the weight of this minimum cut is least.
Note that this algorithm delivers an optimum if and only if there exists an optimum solution in which there are just two colors.
Indeed if there are no feasible 2-colorings then the algorithm will fail.
\end{enumerate}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& Graph Size & $|\mathcal{F}|$ & Cut Weight & Time (secs) & Cut Weight & Time (secs)\\
& $({\rm Nodes}, {\rm Edges})$ & & (2-color opt.) & (2-color opt.) & (Greedy) & (Greedy) \\
\hline
1 & $(60,90)$ & 5 & 2.01 & 0.2 & 2.01 & 0.02\\
\hline
2 & $(64, 192)$ & 5 & 14.93 & 1.2 & 14.93 & 0.02\\
\hline
3 & $(1024, 1536)$ & 7 & 4.15 & 562 & 4.51 & 0.4\\
\hline
4 & $(1K, 3K)$ & 7 & 11.72 & 1900 & 11.83 & 0.4\\
\hline
6 & $(32K, 98K)$ & 11 & \mbox{n/a} & \mbox{n/a} & 18.47 & 6948\\
\hline
\end{tabular}
\caption{{\bf 2-color optimum vs. Gomory-Hu Based Greedy Algorithm:} Code run in Java 1.8 on a single core,
on top of the \texttt{JGraphT} implementation of Gomory-Hu Tree.} \label{comparetab}
\end{center}
\end{table}
Observe in Table \ref{comparetab} that the cut weight obtained by both algorithms is close in all examples, and identical in examples 1 and 2.
In all examples the greedy algorithm delivers a solution with two colors, but in problems 3 and 4
it fails to find an optimum coloring.
The compute time for exhaustive 2-coloring reflects exponential growth in the number of proper 2-colorings
of the union of the forbidden sets, and a quadratic growth in the time to solve a single minimum cut problem on the whole of $(V, E, w)$.
We believe that the compute time for the greedy algorithm is dominated by the time to construct a Gomory-Hu tree.
\section{The multiway cut problem}
\subsection{Review} \label{s;muliwaycut}
Return to the multiway cut case, mentioned in Section \ref{s:mmwcut},
where the forbidden hypergraph $\mathcal{F}$ consists
of the complete graph on a proper subset $T \subset V$.
Assume $|T| \geq 3$, since cases where $|T| < 2$ are vacuous, and the case where $|T|= 2$
is the standard Minimum Cut problem on graphs.
No edge in $E_0$ can have both endpoints in $T$, but, since $G$ is connected,
every vertex in $T$ is incident to at least one edge in $E_0$.
A hypergraph coloring for $(V, \mathcal{F})$, in the sense of Section \ref{s:fhg}, means a
vertex partition $U_1, U_2, \ldots \subset V$ such that $|U_i \cap T| \leq 1$ for all $i$.
We seek such a partition for which the total weight of edges with endpoints
in different $(U_i)$ is minimum.
By Lemma \ref{l:equiv}, this is the same as a maximum
weight edge subset $E_1 \subset E_0$ such that no component of the graph $(V, E_1)$
contains more than one element of $T$; we seek to remove
edges of minimum total weight, such that
each of the vertices in $T$ must be in a different graph component after the cut.
\subsubsection{Application: vertex label anomaly correction} \label{s:anomaly}
Darling and Velednitsky \cite{dar} consider a large bipartite graph encoding transactional data. In an abstract sense, the left vertex set $L$
consists of clients, the right vertex set $R$ consist of servers, and each edge records the transaction
volume between a specific client - server pair.
Furthermore there is a set $C$ of categories, and each right vertex is connected by a weighted edge to exactly one of these categories.
The complete vertex set is $V = L \cup R \cup C $, and the edge set $E_0$ includes both the transaction records and the category labelling
of right vertices.
The ideal state would be that each component of this graph contains exactly one vertex of $C$, which leads to a partition of $L \cup R$
among the categories. If this is not the case, we seek a maximum weight subset $E_1 \subset E_0$ for which this condition holds.
This is a multiway cut problem: all vertices in $C$ must belong to different components.
\subsection{Matroid perspective}
Given $T \subset V$ with $2 \leq |T| < |V|$, and given a connected weighted graph
$G=(V, E_0, w)$ such that no edge in $E_0$ has both endpoints in $T$,
define $\mathcal{E}_T$ to be the collection of subsets $X \subseteq E_0$ such that:
\begin{enumerate}
\item
$(V, X)$ is a forest, and
\item
None of the tree components of the graph $(V, X)$
has more than one vertex in $T$.
\end{enumerate}
\begin{figure}
\caption{\textbf{Construction for matroid proof in Proposition \ref{p:mwcmatroid} }
\textit{The set $T=\{t_1, t_2, t_3\}$. The auxiliary vertex is $t_0$.
}
} \label{f:mwcutcyclematroid}
\begin{center}
\scalebox{0.25}{\includegraphics{Forest.png}}
\end{center}
\end{figure}
We call $\mathcal{E}_T$ the \textbf{multiway cut cycle matroid}, because of
Proposition \ref{p:mwcmatroid}.
\begin{proposition} \label{p:mwcmatroid}
$\mathcal{E}_T$ is a matroid.
\end{proposition}
Before proving Proposition \ref{p:mwcmatroid}, we introduce an auxiliary construction,
illustrated in Figure \ref{f:mwcutcyclematroid}.
Let $T=\{t_1, t_2, \ldots, t_k\}$. Introduce a new graph $G':=(V', E')$, on the vertex set
$V':= V \cup \{t_0\}$, where $t_0$ is a new vertex not in $V$,
and with edge set
\[
E':= E_0 \cup E_*; \quad E_*:=\{ \{t_0, t_i\}_{1 \leq i \leq k} \} .
\]
In other words, $E'$ is obtained by augmenting $E_0$ with a star centered at the new vertex $t_0$,
connected to each of the vertices in $T$; see Figure \ref{f:mwcutcyclematroid}.
\begin{lemma}
There is a bijection between $\mathcal{E}_T$ and the edge sets $Z \supseteq E_*$, for which
$(V', Z)$ is a forest contained in $(V', E')$.
\end{lemma}
\begin{proof}
Given $X \in \mathcal{E}_T$, define $X_i \subset X$
to be those edges in the component containing $t_i$, for $1 \leq i \leq k$.
In other words, for $1 \leq i \leq k$, $X_i \subset X$ is maximal such that
$(V(X_i), X_i)$ is connected and $V(X_i) \ni t_i$.
Let $X_0$ denote the edges in all components with no vertices in $T$. Hence
\[
X = X_0 \cup X_1 \cup \cdots \cup X_k
\]
is a partition of $X$. By definition of $\mathcal{E}_T$, none of the graphs
$(V(X_i), X_i)_{0 \leq i \leq k}$ contain a cycle. The graph $(V', E_* \cup X)$
contains a cycle-free component with edge set $E_* \cup (X \setminus X_0)$,
whose vertex set is
\[
\{t_0, t_1, t_2, \ldots, t_k\} \cup \bigcup_{i=1}^k V(X_i).
\]
In this component, the unique path between $t_i$ and $t_j$ for $1 \leq i < j \leq k$ passes through $t_0$.
The forest $(V, X_0)$ is unaffected by adding edges in $E_*$, and remains cycle-free. Hence
$(V', Z:=E_* \cup X)$ is a forest as claimed. The same argument can be reversed to give
the desired bijection.
\end{proof}
\subsection{Proof of Proposition \ref{p:mwcmatroid}}
\begin{proof}
\textbf{Step I: cycle matroid: }
For the sake of completeness, we recapitulate from \cite{kor}, Proposition 13.4,
the short proof that for any graph $(V, E)$,
the collection $\mathcal{T}$, consisting of subsets $X\subseteq E$ for which $(V, X)$ is a forest,
is a matroid. It is clear $\mathcal{T}$ is an independence system.
Suppose $X$ and $Y$ both belong to $\mathcal{T}$, but $X \cup \{e \} \notin \mathcal{T}$
for all $e \in Y$. For every edge $e$ in $Y$, both endpoints of $e$ are in the same component
of $(V, X)$, so each connected component of $(V, Y)$ is a subset of a connected component of $(V, X)$.
Hence $|C(X)| \leq |C(Y)|$, where $C(X)$ denotes the connected components of $(V, X)$.
It follows that
\[
|X| = |V| - |C(X)| \geq |V| - |C(Y)| = |Y|.
\]
So if $X, Y \in \mathcal{T}$ and $|X| < |Y|$, then $X \cup \{e \} \in \mathcal{T}$
for some $e \in Y$, proving that $\mathcal{T}$ is a matroid. It is called the \textbf{cycle matroid}
of $(V, E)$
\textbf{Step II: matroid contraction: }
To conclude, we show that $(E_0, \mathcal{E}_T)$ is the contraction of another matroid, and therefore a matroid.
First recall some terminology.
If $M = (E,I)$ is a matroid and $X \subset E$ then the \textbf{deletion matroid} $M \setminus X = ( E \setminus X, I')$
where $I' = \{ Z | Z \subset (E \setminus X), Z \in I\}$.
The \textbf{dual matroid} $M^* = (E, I^*)$, where $I^*$ consists of those sets $F \subset E$ for which
there is a base $B$ of $M$ such that $F \cap B = \emptyset$. For $X \subset E$, the
\textbf{matroid contraction} $M / X = (M^* \setminus X)^*$.
Note that if $X \subset E$, and $Z$ is a basis of $X$, then a subset $B$ of $E \setminus X$ is independent in
$M / X$ if and only if $B \cup Z$ is independent in $M=(E,I)$.
In this setting, take $M = (E',I)$ to be the cycle matroid of $(V', E_* \cup E_0)$ .
Since $E_*$ is independent in $M$ it is a base for itself.
Therefore the independent sets of the $M / E_*$ are exactly the forests of $G$ with
the $t_i$ in different components (A forest $F$ in $G$ with $t_0$ and
the edges of $E_*$ added is acyclic if and only if the $t_i$ are in different components in $F$.)
Hence the matroid $M / E_*$ is $(E_0, \mathcal{E}_T)$ as desired.
\end{proof}
\subsection{Greedy multiway cut forest algorithm}
As for any matroid, the \textbf{Best-In Greedy Algorithm} takes the following form.
\begin{enumerate}
\item
Sort $E_0 = \{e_1, e_2, \ldots, e_n\}$ such that $w(e_1) \geq w(e_2) \geq \cdots \geq w(e_n)$.
\item
Set $X = \emptyset$.
\item
For $i = 1$ to $n$, \textbf{do: } if $X \cup \{e_i\} \in \mathcal{E}_T$, then set $X:= X \cup \{e_i\}$.
\end{enumerate}
The context here is multiway cut with respect to $T \subset V$ with $2 \leq |T| < |V|$.
The connected weighted graph $G=(V, E_0, w)$ has no edge in $E_0$ with both endpoints in $T$.
In practice, Step 3. entails checking that the next edge $e_i$ neither causes a cycle, nor causes
two elements of $T$ to fall in the same component.
Proposition \ref{p:mwcmatroid} has the Corollary:
\begin{corollary}
Suppose the forbidden hypergraph consists
of the complete graph on a proper subset $T \subset V$ with $2 \leq |T| < |V|$, and
the connected weighted graph $G=(V, E_0, w)$ has no edge in $E_0$ with both endpoints in $T$.
The Best-In Greedy Algorithm finds an optimum solution to the problem of
finding a maximum weight forest $E_1 \subset E_0$ such that no component of $(V, E_1)$ contains
more than one element of $T$.
\end{corollary}
Call this $E_1$ an \textbf{optimum multiway cut forest}.
\begin{proof}
Since $\mathcal{E}_T$ is a matroid by Proposition \ref{p:mwcmatroid},
the Corollary follows from the Edmonds-Rado Theorem, \cite[Theorem 13.20]{kor}.
\end{proof}
\textbf{Remark: }
Costa \& Billionnet \cite{cos} have given a construction which solves
the multiway cut problem on trees in linear time.
The Best-In Greedy Algorithm falls short of linear time in that (1)
the sort alone is $O(|V| \log (|V|))$. and (2)
for the overall runtime to be linear in $|V|$,
the check for each proposed edge would need to be done in constant time.
\textbf{Acknowledgments: } The authors thank Imtiaz Manji and Adam Logan of the Tutte Institute of
Mathematics and Computing for insightful comments, and the organizers and participants
of STAMP 2017 for encouraging this work.
|
\section{Introduction}
Text to speech (TTS) is a very important task for user interaction, aiming to synthesize intelligible and natural audios which are indistinguishable from human recordings. Traditional TTS systems have two components: front-end and back-end. Front-end is responsible for text analysis and linguistic feature extraction, such as word segmentation, part of speech tagging, multi-word disambiguation and prosodic structure prediction; back-end is built for speech synthesis based on linguistic features from front-end, such as speech acoustic parameter modeling, prosody modeling and speech generation. In the past decades, concatenative and parametric speech synthesis systems were mainstream techniques. However, both of them have complex pipelines, and defining good linguistic features is often time-consuming and language specific, which requires a lot of resource and manpower. Besides, synthesized audios often have glitches or instability in prosody and pronunciation compared to human speech, and thus sound unnatural.
Recently, with the rapid development of neural networks, end-to-end generative text-to-speech models, such as Tacotron \cite{wang2017tacotron} and Tacotron2 \cite{shen2017natural}, are proposed to simplify traditional speech synthesis pipeline by replacing the production of these linguistic and acoustic features with a single neural network. Tacotron and Tacotron2 first generate mel spectrograms directly from texts, then synthesize the audio results by a vocoder such as Griffin Lim algorithm \cite{griffin1984signal} or WaveNet \cite{van2016wavenet}. With the end-to-end neural network, quality of synthesized audios is greatly improved and even comparable with human recordings on some datasets. The end-to-end neural TTS models contain two components, an encoder and a decoder. Given the input sequence (of words or phonemes), the encoder tries to map them into a semantic space and generates a sequence of encoder hidden states, and the decoder, taking these hidden states as context information with an attention mechanism, constructs the decoder hidden states then outputs the mel frames. For both encoder and decoder, recurrent neural networks (RNNs) are usually leveraged, such as LSTM \cite{hochreiter1997long} and GRU \cite{Cho2014Learning}.
However, RNNs can only consume the input and generate the output sequentially, since the previous hidden state and the current input are both required to build the current hidden state. The characteristic of sequential process limits the parallelization capability in both the training and inference process. For the same reason, for a certain frame, information from many steps ahead may has been biased after multiple recurrent processing. To deal with these two problems, Transformer \cite{vaswani2017attention} is proposed to replace the RNNs in NMT models.
Inspired by this idea, in this paper, we combine the advantages of Tacotron2 and Transformer to propose a novel end-to-end TTS model, in which the multi-head attention mechanism is introduced to replace the RNN structures in the encoder and decoder, as well as the vanilla attention network. The self-attention mechanism unties the sequential dependency on the last previous hidden state to improve the parallelization capability and relieve the long distance dependency problem. Compared with the vanilla attention between the encoder and decoder, the multi-head attention can build the context vector from different aspects using different attention heads. With the phoneme sequences as input, our novel Transformer TTS network generates mel spectrograms, and employs WaveNet as vocoder to synthesize audios. We conduct experiments with 25-hour professional speech dataset, and the audio quality is evaluated by human testers. Evaluation results show that our proposed model outperforms the original Tacotron2 with a gap of 0.048 in CMOS, and achieves a similar performance (4.39 in MOS) with human recording (4.44 in MOS). Besides, our Transformer TTS model can speed up the training process about 4.25 times compared with Tacotron2.
Audio samples can be accessed on \url{https://neuraltts.github.io/transformertts/}
\section{Background}
In this section, we first introduce the sequence-to-sequence model, followed by a brief description about Tacotron2 and Transformer, which are two preliminaries in our work.
\subsection{Sequence to Sequence Model}
A sequence-to-sequence model \cite{sutskever2014sequence,bahdanau2014neural} converts an input sequence $(x_1,x_2,...,x_T)$ into an output sequence $(y_1,y_2,...,y_{T'})$, and each predicted $y_t$ is conditioned on all previously predicted outputs $y_1,...,y_{t-1}$. In most cases, these two sequences are of different lengths ($T \neq T'$).
In NMT, this conversion translates the input sentence in one language into the output sentence in another language, based on a conditional probability $p(y_1,...,y_T'|x_1,...,x_T)$:
\begin{align}
h_t &= encoder(h_{t-1}, x_t) \\
s_t &= decoder(s_{t-1}, y_{t-1}, c_t)
\end{align}
where $c_t$ is the context vector calculated by an attention mechanism:
\begin{align}
c_t &= attention(s_{t-1}, \mathbf{h})
\end{align}
thus $p(y_1,...,y_T'|x_1,...,x_T)$ can be computed by
\begin{equation}
p(y_1,...,y_T'|x_1,...,x_T)=\prod_{t=1}^{T'} p(y_t|\mathbf{y_{<t}},\mathbf{x})
\end{equation}
and
\begin{equation}
p(y_t|\mathbf{y_{<t}},\mathbf{x}) = softmax(f(s_t))
\end{equation}
where $f(\cdot)$ is a fully connected layer. For translation tasks, this softmax function is among all dimensions of $f(s_t)$ and calculates the probability of each word in the vocabulary. However, in the TTS task, the softmax function is not required and the hidden states $\mathbf{s}$ calculated by decoder are consumed directly by a linear projection to obtain the desired spectrogram frames.
\subsection{Tacotron2}
Tacotron2 is a neural network architecture for speech synthesis directly from text, as shown in Fig. \ref{fig:tacotron2} . The embedding sequence of input is firstly processed with a 3-layer CNN to extract a longer-term context, and then fed into the encoder, which is a bi-directional LSTM. The previous mel spectrogram frame (the predicted one in inference, or the golden one in training time), is first processed with a 2-layer fully connected network (decoder pre-net), whose output is concatenated with the previous context vector, followed by a 2-layer LSTM. The output is used to calculate the new context vector at this time step, which is concatenated with the output of the 2-layer LSTM to predict the mel spectrogram and stop token with two different linear projections respectively. Finally the predicted mel spectrogram is fed into a 5-layer CNN with residual connections to refine the mel spectrogram.
\begin{figure}[tb]
\centering
\includegraphics[width=0.85\linewidth]{Tacotron2.jpg}
\caption{System architecture of Tacotron2.}
\label{fig:tacotron2}
\end{figure}
\subsection{Transformer for NMT}
\label{subsec:transformer}
Transformer \cite{vaswani2017attention}, shown in Fig. \ref{fig:transformer}, is a sequence to sequence network, based solely on attention mechanisms and dispensing with recurrences and convolutions entirely. In recent works, Transformer has shown extraordinary results, which outperforms many RNN-based models in NMT. It consists of two components: an encoder and a decoder, both are built by stacks of several identity blocks. Each encoder block contains two subnetworks: a multi-head attention and a feed forward network, while each decoder block contains an extra masked multi-head attention comparing to the encoder block. Both encoder and decoder blocks have residual connections and layer normalizations.
\begin{figure}[!tb]
\centering
\includegraphics[width=0.98\linewidth]{Transformer.jpg}
\caption{System architecture of Transformer.}
\label{fig:transformer}
\end{figure}
\section{Neural TTS with Transformer}
Compared to RNN-based models, using Transformer in neural TTS has two advantages. First it enables parallel training by removing recurrent connections, as frames of an input sequence for decoder can be provided in parallel. The second one is that self attention provides an opportunity for injecting global context of the whole sequence into each input frame, building long range dependencies directly. Transformer shortens the length of paths forward and backward signals have to traverse between any combination of positions in the input and output sequences down to 1. This helps a lot in a neural TTS model, such as the prosody of synthesized waves, which not only depends on several words in the neighborhood, but also sentence level semantics.
In this section we will introduce the architecture of our Transformer TTS model, and analyze the function of each part. The overall structure diagram is shown in Fig. \ref{fig:transtron}.
\subsection{Text-to-Phoneme Converter}
English pronunciation has certain regularities, for example, there are two kinds of syllables in English: open and closed. The letter "a" is often pronounced as /e\i/ when it's in an open syllable, while it is pronounced as /\ae/ or /a\textlengthmark/ in closed syllables. We can rely on the neural network to learn such a regularity in the training process. However, it is difficult to learn all the regularities when, which is often the case, the training data is not sufficient enough, and some exceptions have too few occurrences for neural networks to learn. So we make a rule system and implement it as a text-to-phoneme converter, which can cover the vast majority of cases.
\subsection{Scaled Positional Encoding}
\label{subsec:scl_tri_pe}
Transformer contains no recurrence and no convolution so that if we shuffle the input sequence of encoder or decoder, we will get the same output. To take the order of the sequence into consideration, information about the relative or absolute position of frames is injected by triangle positional embeddings, shown in Eq. \ref{eq:tri_pe}:
\begin{align}
PE(pos,2i)&=\sin(\frac{pos}{10000^{\frac{2i}{d_{model}}}}) \\
PE(pos,2i+1)&=\cos(\frac{pos}{10000^{\frac{2i}{d_{model}}}})
\label{eq:tri_pe}
\end{align}
where $pos$ is the time step index, $2i$ and $2i+1$ is the channel index and $d_{model}$ is the vector dimension of each frame. In NMT, the embeddings for both source and target language are from language spaces, so the scales of these embeddings are similar. This condition doesn't hold in the TTS scenarioe, since the source domain is of texts while the target domain is of mel spectrograms, hence using fixed positional embeddings may impose heavy constraints on both the encoder and decoder pre-nets (which will be described in Sec. \ref{subsec:enc_prenet} and \ref{subsec:dec_prenet}). We employ these triangle positional embeddings with a trainable weight, so that these embedding can adaptively fit the scales of both encoder and decoder pre-nets' output, as shown in Eq. \ref{eq:scl_tri_pe}:
\begin{equation}
x_i = prenet({phoneme}_i)+\alpha PE(i)
\label{eq:scl_tri_pe}
\end{equation}
where $\alpha$ is the trainable weight.
\subsection{Encoder Pre-net}
\label{subsec:enc_prenet}
In Tacotron2, a 3-layer CNN is applied to the input text embeddings, which can model the longer-term context in the input character sequence. In our Transformer TTS model, we input the phoneme sequence into the same network, which is called "encoder pre-net". Each phoneme has a trainable embedding of 512 dims, and the output of each convolution layer has 512 channels, followed by a batch normalization and ReLU activation, and a dropout layer as well. In addition, we add a linear projection after the final ReLU activation, since the output range of ReLU is $[0,+\infty)$, while each dimension of these triangle positional embeddings is in $[-1,1]$. Adding 0-centered positional information onto non-negative embeddings will result in a fluctuation not centered on the origin and harm model performance, which will be demonstrated in our experiment. Hence we add a linear projection for center consistency.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.97\linewidth]{Transtron.jpg}
\caption{System architecture of our model.}
\label{fig:transtron}
\end{figure}
\subsection{Decoder Pre-net}
\label{subsec:dec_prenet}
The mel spectrogram is first consumed by a neural network composed of two fully connected layers(each has 256 hidden units) with ReLU activation, named "decoder pre-net", and it plays an important role in the TTS system.
Phonemes has trainable embeddings thus their subspace is adaptive, while that of mel spectrograms is fixed. We infer that decoder pre-net is responsible for projecting mel spectrograms into the same subspace as phoneme embeddings, so that the similarity of a $\left \langle phoneme, mel\ frame \right \rangle$ pair can be measured, thus the attention mechanism can work. Besides, 2 fully connected layers without non-linear activation are also tried but no reasonable attention matrix aligning the hidden states of encoder and decoder can be generated. In our other experiment, hidden size is enlarged from 256 to 512, however that doesn't generate significant improvement but needs more steps to converge. Accordingly, we conjecture that mel spectrograms have a compact and low dimensional subspace that 256 hidden units are good enough to fit. This conjecture can also be evidenced in our experiment, which is shown in Sec. \ref{subsubsec:pe_method}, that the final positional embedding scale of decoder is smaller than that of encoder. An additional linear projection is also added like encoder pre-net not only for center consistency but also obtain the same dimension as the triangle positional embeddings.
\subsection{Encoder}
\label{subsec:encoder}
In Tacotron2, the encoder is a bi-directional RNN. We replace it with Transformer encoder which is described in Sec. \ref{subsec:transformer} . Comparing to original bi-directional RNN, multi-head attention splits one attention into several subspaces so that it can model the frame relationship in multiple different aspects, and it directly builds the long-time dependency between any two frames thus each of them considers global context of the whole sequence. This is crucial for synthesized audio prosody especially when the sentence is long, as generated samples sound more smooth and natural in our experiments. In addition, employing multi-head attention instead of original bi-directional RNN can enable parallel computing to improve training speed.
\subsection{Decoder}
In Tacotron2, the decoder is a 2-layer RNN with location-sensitive attention \cite{chorowski2015attention}. We replace it with Transformer decoder which is described in Sec. \ref{subsec:transformer}.
Employing Transformer decoder makes two main differences, adding self-attention, which can bring similar advantages described in Sec. \ref{subsec:encoder}, and using multi-head attention instead of the location-sensitive attention. The multi-head attention can integrate the encoder hidden states in multiple perspectives and generate better context vectors.
Taking attention matrix of previous decoder time steps into consideration, location-sensitive attention used in Tacotron2 can encourage the model to generate consistent attention results.
We try to modify the dot product based multi-head attention to be location sensitive, but that doubles the training time and easily run out of memory.
\subsection{Mel Linear, Stop Linear and Post-net}
Same as Tacotron2, we use two different linear projections to predict the mel spectrogram and the stop token respectively, and use a 5-layer CNN to produce a residual to refine the reconstruction of mel spectrogram. It's worth mentioning that, for the stop linear, there is only one positive sample in the end of each sequence which means "stop", while hundreds of negative samples for other frames. This imbalance may result in unstoppable inference. We impose a positive weight ($5 .0\sim 8.0$) on the tail positive stop token when calculating binary cross entropy loss, and this problem was efficiently solved.
\section{Experiment}
In this section, we conduct experiments to test our proposed Transformer TTS model with 25-hour professional speech pairs, and the audio quality is evaluated by human testers in MOS and CMOS.
\subsection{Training Setup}
We use 4 Nvidia Tesla P100 to train our model with an internal US English female dataset, which contains 25-hour professional speech (17584 $\left \langle text, wave \right \rangle$ pairs, with a few too long waves removed). 50ms silence at head and 100ms silence at tail are kept for each wave. Since the lengths of training samples vary greatly, fixed batch size will either run out of memory when long samples are added into a batch with a large size or waste the parallel computing power if the batch is small and into which short samples are divided. Therefore, we use the dynamic batch size where the maximum total number of mel spectrogram frames is fixed and one batch should contain as many samples as possible. Thus there are on average 16 samples in single batch per GPU. We try training on a single GPU, but the procedures are quiet instable or even failed, by which synthesized audios were like raving and incomprehensible. Even if training doesn't fail, synthesized waves are of bad quality and weird prosody, or even have some severe problems like missing phonemes. Thus we enable multi-GPU training to enlarge the batch size, which effectively solves those problems.
\subsection{Text-to-Phoneme Conversion and Pre-process}
Tacotron2 uses character sequences as input, while our model is trained on pre-normalized phoneme sequences. Word and syllable boundaries, punctuations are also included as special markers. The process pipeline to get training phoneme sequences contains sentence separation, text normalization, word segmentation and finally obtaining pronunciation. By text-to-phoneme conversion, mispronunciation problems are greatly reduced especially for those pronunciations that are rarely occurred in our training set.
\subsection{WaveNet Settings}
We train a WaveNet conditioned on mel spectrogram with the same internal US English female dataset, and use it as the vocoder for all models in this paper.
The sample rate of ground truth audios is 16000 and frame rate (frames per second) of ground truth mel spectrogram is 80.
Our autoregressive WaveNet contains 2 QRNN layers and 20 dilated layers, and the sizes of all residual channels and dilation channels are all 256. Each frame of QRNN's final output is copied 200 times to have the same spatial resolution as audio samples and be conditions of 20 dilated layers.
\subsection{Training Time Comparison}
Our model can be trained in parallel since there is no recurrent connection between frames. In our experiment, time consume in a single training step for our model is $\sim$0.4s, which is 4.25 times faster than that of Tacotron2 ($\sim$1.7s) with equal batch size (16 samples per batch). However, since the parameter quantity of our model is almost twice than Tacotron2, it still takes $\sim$3 days to converge comparing to $\sim$4.5 days of that for Tacotron2.
\subsection{Evaluation}
We randomly select 38 fixed examples with various lengths (no overlap with training set) from our internal dataset as the evaluation set. We evaluate mean option score (MOS) on these 38 sentences generated by different models (include recordings), in which case we can keep the text content consistent and exclude other interference factors hence only examine audio quality.
For higher result accuracy, we split the whole MOS test into several small tests, each containing one group from our best model, one group from a comparative model and one group of recordings. Those MOS tests are rigorous and reliable, as \textbf{each audio is listened to by at least 20 testers, who are all native English speakers } (comparing to Tacotron2's 8 testers in \citet{shen2017natural}), and each tester listens less than 30 audios.
We train a Tacotron2 model with our internal US English female dataset as the baseline (also use phonemes as input), and gain equal MOS with our model. Therefore we test the comparison mean option score (CMOS) between samples generated by Tacotron2 and our model for a finer contrast. In the comparison mean option score (CMOS) test, testers listen to two audios (generated by Tacotron2 and our model with the same text) each time and evaluates how the latter feels comparing to the former using a score in $\left[-3, 3\right]$ with intervals of 1. The order of the two audios changes randomly so testers don't know their sources. Our model wins by a gap of 0.048, and detailed results are shown in Table \ref{tab:mos_system}.
\begin{table}[t!]
\centering
\begin{tabular}{ccc}
\toprule
System & MOS & CMOS \\
\midrule
Tacotron2 & $4.39 \pm 0.05$ & $0$\\
Our Model & $4.39 \pm 0.05$ & \textbf{0.048}\\
\midrule
Ground Truth & $4.44 \pm 0.05$ & -\\
\bottomrule
\end{tabular}
\caption{MOS comparison among our model, our Tacotron2 and recordings.}
\label{tab:mos_system}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\linewidth]{mel_comp.jpg}
\caption{Mel spectrogram comparison. Our model (6-layer) does better in reconstructing details as marked in red rectangles, while Tacotron2 and our 3-layer model blur the texture especially in high frequency region. Best viewed in color.}
\label{fig:mel_comp}
\end{figure}
We also select mel spectrograms generated by our model and Tacotron2 respectively with the same text, and compare them together with ground truth, as shown in column 1,2 and 3 of Fig. \ref{fig:mel_comp}. As we can see, our model does better in reconstructing details as marked in red rectangles, while Tacotron2 left out the detailed texture in high frequency region.
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\linewidth]{scale.png}
\caption{PE scale of encoder and decoder.}
\label{fig:pe_scale}
\end{figure}
\subsection{Ablation Studies}
In this section, we study the detail modification of network architecture, and conduct several experiments to show our improvements.
\subsubsection{Re-centering Pre-net's Output}
As described in Sec. \ref{subsec:enc_prenet} and \ref{subsec:dec_prenet}, we re-project both the encoder and decoder pre-nets' outputs for consistent center with positional embeddings. In contrast, we add no linear projection in encoder pre-net and add a fully connected layer with ReLU activation in decoder pre-net. The results imply that center-consistent positional embedding performs slightly better, as shown in Table \ref{tab:mos_re_projection}.
\begin{table}[t!]
\centering
\begin{tabular}{ccc}
\toprule
Re-center & MOS \\
\midrule
No & $4.32 \pm 0.05$ \\
Yes & \textbf{4.36} $\pm 0.05$ \\
\midrule
Ground Truth & $4.43 \pm 0.05$ \\
\bottomrule
\end{tabular}
\caption{MOS comparison of whether re-centering pre-net's output.}
\label{tab:mos_re_projection}
\end{table}
\subsubsection{Different Positional Encoding Methods}
\label{subsubsec:pe_method}
We inject positional information into both encoder's and decoder's input sequences as Eq. \ref{eq:scl_tri_pe}. Fig. \ref{fig:pe_scale} shows that the final positional embedding scales of encoder and decoder are different, and Table \ref{tab:mos_scale_type} shows model with trainable scale performs slightly better. We think that the trainable scale relaxes the constraint on encoder and decoder pre-nets, making positional information more adaptive for different embedding spaces.
We also try adding absolute position embeddings (each position has a trainable embedding) to the sequence, which also works but has some severe problems such as missing phonemes when the sequences became long. That's because long sample is relatively rare in the training set, so the embeddings for large indexes can hardly be trained and thus the position information won't be accurate for rear frames in a long sample.
\subsubsection{Model with Different Hyper-Parameter}
Both the encoder and decoder of the original Transformer is composed of 6 layers, and each multi-head attention has 8 heads. We compare performance and training speed with different layer and head numbers, as shown in Table \ref{tab:mos_ln}, \ref{tab:mos_hn} and \ref{tab:speed_hp}. We find that reducing layers and heads both improve the training speed, but on the other hand, harm model performance in different degrees.
We notice that in both the 3-layer and 6-layer model, only alignments from certain heads of the beginning 2 layers' are interpretable diagonal lines, which shows the approximate correspondence between input and output sequence, while those of the following layers are disorganized. Even so, more layers can still lower the loss, refine the synthesized mel spectrogram and improve audio quality. The reason is that with residual connection between different layers, our model fits target transformation in a Taylor-expansion way: the starting terms account most as low ordering ones, while the subsequential ones can refine the function. Hence adding more layer makes the synthesized wave more natural, since it does better in processing spectrogram details (shown in column 4, Fig. \ref{fig:mel_comp}). Fewer heads can slightly reduce training time cost since there are less production per layer, but also harm the performance.
\begin{table}[t!]
\centering
\begin{tabular}{ccc}
\toprule
PE Type & MOS \\
\midrule
Original & $4.37 \pm 0.05$ \\
Scaled & \textbf{4.40} $\pm 0.05$ \\
\midrule
Ground Truth & $4.41 \pm 0.04$ \\
\bottomrule
\end{tabular}
\caption{MOS comparison of scaled and original PE.}
\label{tab:mos_scale_type}
\end{table}
\begin{table}[t!]
\centering
\begin{tabular}{ccc}
\toprule
Layer Number & MOS \\
\midrule
3-layer & $4.33 \pm 0.06$ \\
6-layer & \textbf{4.41} $\pm 0.05$ \\
\midrule
Ground Truth & $4.44 \pm 0.05$ \\
\bottomrule
\end{tabular}
\caption{Ablation studies in different layer numbers.}
\label{tab:mos_ln}
\end{table}
\begin{table}[t!]
\centering
\begin{tabular}{ccc}
\toprule
Head Number & MOS \\
\midrule
4-head & $4.39 \pm 0.05$ \\
8-head & \textbf{4.44} $\pm 0.05$ \\
\midrule
Ground Truth & $4.47 \pm 0.05$ \\
\bottomrule
\end{tabular}
\caption{Ablation studies in different head numbers.}
\label{tab:mos_hn}
\end{table}
\begin{table}[t!]
\centering
\begin{tabular}{c|cc}
\toprule
$ $ & 3-layer & 6-layer \\
\hline
4-head & - & $0.44$~~~ \\
8-head & $0.29$ & $0.50$~~~ \\
\bottomrule
\end{tabular}
\caption{Comparison of time consuming (in second) per training step of different layer and head numbers.}
\label{tab:speed_hp}
\end{table}
\section{Related Work}
Traditional speech synthesis methods can be categorized into two classes: concatenative systems and parametric systems. Concatenative TTS systems \cite{hunt1996unit,black1997automatically} split original waves into small units, and stitch them by some algorithms such as Viterbi \cite{viterbi1967error} followed by signal process methods \cite{charpentier1986diphone,verhelst1993overlap} to generate new waves. Parametric TTS systems \cite{tokuda2000speech,zen2009statistical,ze2013statistical,tokuda2013speech} convert speech waves into spectrograms, and acoustic parameters, such as fundamental frequency and duration, are used to synthesize new audio results.
Traditional speech synthesis methods require extensive domain expertise and may contain brittle design choices. Char2Wav \cite{sotelo2017char2wav} integrates the front-end and the back-end as one seq2seq \cite{sutskever2014sequence,bahdanau2014neural} model and learns the whole process in an end-to-end way, predicting acoustic parameters followed by a SampleRNN \cite{mehri2016samplernn} as the vocoder. However, acoustic parameters are still intermediate for audios, thus Char2Wav is not a really end-to-end TTS model, and their seq2seq and SampleRNN models need to be separately pre-trained,
while Tacotron, proposed by \citet{wang2017tacotron}, is an end-to-end generative text-to-speech model, which can be trained by $\left \langle text, spectrogram \right \rangle$ pairs directly from scratch, and synthesizes speech audios with generated spectrograms by Griffin Lim algorithm \cite{griffin1984signal}.
Based on Tacotron, Tacotron2 \cite{shen2017natural}, a unified and entirely neural model, generates mel spectrograms by a Tacotron-style neural network and then synthesizes speech audios by a modified WaveNet \cite{van2016wavenet}. WaveNet is an autoregressive generative model for waveform synthesis, composed of stacks of dilated convolutional layers and processes raw audios of very high temporal resolution (e.g., 24,000 sample rate), while suffering from very large time cost in inference. This problem is solved by Parallel WaveNet \cite{oord2017parallel}, based on the inverse autoregressive flow (IAF) \cite{kingma2016improved} and reaches $1000 \times$ real time. Recently, ClariNet \cite{ping2018clarinet}, a fully convolutional text-to-wave neural architecture, is proposed to enable the fast end-to-end training from scratch.
Moreover, VoiceLoop \cite{taigman2018voiceloop} is an alternative neural TTS method mimicking a person's voice based on samples captured in-the-wild, such as audios of public speeches, and even with an inaccurate automatic transcripts.
On the other hand, Transformer \cite{vaswani2017attention} is proposed for neural machine translation (NMT) and achieves state-of-the-art result. Previous NMT models are dominated by RNN-based \cite{bahdanau2014neural} or CNN-based (e.g. ConvS2S \cite{gehring2017convolutional}, ByteNet \cite{kalchbrenner2016neural}) neural networks. For RNN-based models, both training and inference are sequential for each sample, while CNN-based models enable parallel training. Both RNN and CNN based models are difficult to learn dependencies between distant positions since RNNs have to traverse a long path and CNN has to stack many convolutional layers to get a large receptive field, while Transformer solves this using self attention in both its encoder and decoder. The ability of self-attention is also proved in SAGAN \cite{zhang2018self}, where original GANs without self-attention fail to capture geometric or structural patterns that occur consistently in some classes (for example, dogs are often drawn without clearly defined separate feet). By adding self-attention, these failure cases are greatly reduced.
Besides, multi-head attention is proposed to obtain different relations in multi-subspaces. Recently, Transformer has been applied in automatic speech recognition (ASR) \cite{zhou2018comparison,zhou2018syllable}, proving its ability in acoustic modeling other than natural language process.
\section{Conclusion and Future Work}
We propose a neural TTS model based on Tacotron2 and Transformer, and make some modification to adapt Transformer to neural TTS task. Our model generates audio samples of which quality is very closed to human recording, and enables parallel training and learning long-distance dependency so that the training is sped up and the audio prosody is much more smooth. We find that batch size is crucial for training stability, and more layers can refine the detail of generated mel spectrograms especially for high frequency regions thus improve model performance.
Even thought Transformer has enabled parallel training, autoregressive model still suffers from two problems, which are slow inference and exploration bias. Slow inference is due to the dependency of previous frames when infer current frame, so that the inference is sequential, while exploration bias comes from the autoregressive error accumulation. We may solve them both at once by building a non-autoregressive model, which is also our current research in progress.
|
\section{Introduction}
We develop the following stochastic volatility (SV) model for the real-world (P-measure) evolution of a financial asset $S_t$, such as a broad-based stock index:
\bea \mbox{Under P}: \quad \left \{
\begin{array}{ll}
d S_t = (\alpha + \beta \, \sigma_t^2) S_t \, dt + \sigma_t S_t \, dB_t, \quad & S_t \in \RBB_+, \\
d \sigma_t = \sigma_t (\omega - \theta \, \sigma_t) \, dt + \xi \sigma_t \, dW_t, \quad & \sigma_t \in \RBB_+, \\
dB_t \, dW_t = \rho \, dt.
\end{array} \right. \label{eq:XGBM1}
\aeb
Here $(B_t,W_t)$ are correlated Brownian motions with correlation $\rho$. Also \newline $(\alpha,\beta,\omega,\theta,\xi,\rho)$ are constant parameters of the model, meant to be estimated from financial time series. We generally assume throughout that ($\omega,\theta,\xi^2)>0$,
although we sometimes admit $\omega = \theta = 0$ or even $\theta < 0$.
We also develop
a risk-neutral (Q-measure) version of the model, which is used for option valuation.
In that one, $\alpha$ becomes a cost-of-carry: $\alpha \ra (r-q)$, where $r$ is an interest rate and $q$ is a dividend yield. In addition,
$\beta \ra 0$, and
$\omega \ra \omega_Q$, (a possibly different parameter) while the other parameters
$(\theta,\xi,\rho)$ remain identical to their P-measure values.
An attractive feature of the model is that the stochastic volatility $\sigma_t$ is driven by a geometric Brownian motion (GBM)-type noise. GBM volatility seems to be favored by time series analysis over the ``square-root" variance noise of popular affine-type diffusion models, such as
the Heston '93 model.\footnote{See, for example \cite{cjm:2010} and \cite{tegpou:2018}.} Extended with the drifts shown, I call it the `Extended GBM' model, or XGBM for short.
\newpage
As we show, the model admits an exact solution for the transition density, vanilla option values,
and some other quantities of interest. In these solutions, we find two qualitatively
different regimes:
\begin{itemize}
\item Case 1: \quad $\Smallfrac{1}{2} \xi^2 \le \omega < \infty$,
\item Case 2: \quad $0 \le \omega < \Smallfrac{1}{2} \xi^2$.
\end{itemize}
In Case 1, the driving volatility process is mean-reverting and has a stationary density.
In Case 2, with probability 1, as $t$ grows large $\sigma_t \ra 0$, similar to the
SABR model. Indeed, the lognormal SABR model is a special case of (\ref{eq:XGBM1})
when the drifts are absent.
\Pbreak
However, a weakness of the SABR model is its lack of
a stationary density for the volatility. Thus (under Case 1), we have here what may represent the first example of an SV model combining: GBM-type volatility noise, a stationary volatility density, and exact solutions.
\Pbreak
Numerically, we find option values (and
thus implied volatilities) continuous vs. $\omega$, including at the
borderlines case: $\omega = \Smallfrac{1}{2} \xi^2$.
\Pbreak
In brief, the problem is solvable because it is reducible to the evolution problem for a 1D
diffusion operator admitting a spectral expansions in terms of confluent hypergeometric functions $M(a,b,x)$ and $U(a,b,x)$. The other special function that will appear frequently is
\[ F(a,b,c;z) = \sum_{n=0}^{\infty}\frac{(a)_n (b)_n}{(c)_n} \frac{z^n}{n!}, \]
the Gauss hypergeometric function. The $(a)_n$ etc are Pochhammer symbols. All these functions are
built-in in Mathematica, where everything is implemented.
\Pbreak
Structurally, the model is similar to the one-factor short-term
interest rate model, $d r_t = r_t (\omega - \theta \, r_t) \, dt + \xi r_t \, dW_t$,
originally due to Merton and which I fully solved in \cite{lewis:1998} (reprinted in \cite{lewis:2016}).
The solvability of Merton's model suggests, but does not prove, that associated 2D
SV models such as (\ref{eq:XGBM1}) are also solvable. Another suggestion that my XGBM SV model
might be solvable is the discussion found in \cite{hlab:2009} (Sec. 9.5).
As it turns out, moving from the interest rate model to the associated stochastic volatility model is
tricky. One reason is the ``reduced" PDE coefficients become complex-valued,
which complicates the associated spectral expansion. Another reason is that
the ``fundamental transform" needed for option valuation (using the language of \cite{lewis:2000a}) requires a regularization for Case 2. All that is explained below; key formulas are boxed.
\newpage
\section{The Joint Transition Probability Density}
With $(X_t,Y_t) \equiv (\log S_t, \sigma_t)$ the system (\ref{eq:XGBM1}) reads:
\bea \left \{
\begin{array}{ll}
d X_t = [\alpha + (\beta- \smallfrac{1}{2}) \, Y^2_t] \, dt + Y_t \, dB_t,
\quad & X_t \in \RBB, \\
d Y_t = (\omega Y_t - \theta \, Y^2_t) \, dt + \xi \, Y_t \, dW_t, \quad & Y_t \in \RBB_+, \\
dB_t \, dW_t = \rho \, dt.
\end{array} \right. \label{eq:XGBM2}
\aeb
We seek the bivariate transition probability density $p(t,x',y'|x,y)$ defined by $p(t,x',y'|x,y) \, dx' dy' = \Pr{X_t \in dx', Y_t \in dy' | X_0 = x, Y_0 = y}$. The initial condition is $p(0,x',y'|x,y) = \delta(x' - x) \,\delta(y' - y)$, using the Dirac delta. All parameters are real; assume $(\omega,\theta) \ge 0$. Using subscripts for derivatives, the corresponding Kolmogorov backward PDE problem is
\bea \left \{
\begin{array}{l}
p_t = \frac{1}{2} y^2 (p_{xx} + 2 \rho \xi p_{xy} + \xi^2 p_{yy}) + [\alpha + (\beta-\frac{1}{2}) y^2] p_x + (\omega y - \theta y^2) p_y, \\
p(0,x',y' | x, y) = \delta(x-x') \, \delta(y-y') .
\end{array} \right. \label{eq:KBE}
\aeb
The system (\ref{eq:XGBM2}) is a MAP (Markov Additive Process), where $X_t$ is the additive component and $Y_t$ is the Markov component. This is also seen in the $x$-independence in the coefficients of (\ref{eq:KBE}). The MAP property implies \newline $p(t,x',y'|x,y) = p(t,x' - x,y'|0,y)$ which in turn implies the existence of a Fourier representation:
\be p(t,x',y'|x,y) = \int_{-\infty}^{\infty} \e^{-\i z (x'-x)} \, \Phi(t,z, y, y') \, \frac{dz}{2 \pi}, \label{eq:prep} \eb
for a characteristic function $\Phi$ to be determined below.
In the Heston '93 model (also a MAP), $\Phi$ is found in terms of Bessel functions. Here the situation is not quite so simple: we'll find that $\Phi$ itself requires an integration using confluent hypergeometric functions $M(a,b,x)$ and $U(a,b,x)$. The additional integration means that numerical evaluations of the
exact formulas will take an order of magnitude more computer time than Heston. Pure numerics, such as a PDE approach, will be similar in the two models.
\newpage
\subsection{Reduction to an auxiliary problem for $G(\tau,y,y')$} \label{sec:reduction}
Because of the MAP property it suffices to solve (\ref{eq:KBE}) with $x' = 0$. Let $\Phi(t,y,y')= \int \e^{-\i z x} p(t,0,y'|x,y) dx$, suppressing the display of the $z$ dependence. Applying $\int \e^{-\i z x} (\cdots) dx$ to both sides of (\ref{eq:KBE}) and using parts integrations (the boundary terms vanish) yields:
\bea \left \{
\begin{array}{l}
\Phi_t = \frac{1}{2} \xi^2 y^2 \Phi_{yy} + b(y;z) \Phi_y - c(y;z) \Phi, \\
\Phi(0,y,y') = \delta(y-y'),
\end{array} \right. \label{eq:KBE2}
\aeb
\[ \mbox{where} \quad b(y;z) = \omega y - (\theta - \i z \rho \xi) y^2 \,\, \mbox{and}
\,\, c(y;z) = -\i z \alpha + [\Smallfrac{1}{2} z^2 - \i z (\beta - \Smallfrac{1}{2})] y^2.\]
\Pbreak
Letting $\Phi(t,y,y') = \e^{\i z \alpha t} G(\tau,y,y';z)$, where $\tau = \Smallfrac{1}{2} \xi^2 t$ yields
\bea \left \{
\begin{array}{l}
G_{\tau} = y^2 G_{yy} + (\tilde{\omega} y - \theta_z^- y^2) G_y - c^+_z(\beta) y^2 G, \\
G(0,y,y') = \delta(y-y'),
\end{array} \right. \label{eq:Greenfunc0}
\aeb
\[ \mbox{introducing} \quad \tilde{\omega} = \Smallfrac{2 \omega}{\xi^2}, \quad
\theta_z^- = \Smallfrac{2}{\xi^2} (\theta - \i z \rho \xi), \quad \mbox{and}
\quad c^+_z(\beta) = \Smallfrac{2}{\xi^2} [\Smallfrac{1}{2} z^2 - \i z (\beta - \Smallfrac{1}{2})].\]
The $\pm$ superscripts on $\theta_z$ and $c_z$ distinguish them from similar
expressions introduced later where the sign of $z$ is flipped from the convention here.
The mnemonic is that $\theta_z^-$ indicates that $z$ occurs as $-\i z$, so the sign is \emph{minus}.
That is, $\theta_z^{\pm} \equiv \Smallfrac{2}{\xi^2} (\theta \pm \i z \rho \xi)$. The
sign on $c^+_z(\beta)$ refers to the risk-neutral case where $\beta = 0$; i.e.,
$c^{\pm}_z(0) = (z^2 \pm \i z)/\xi^2$.
\Pbreak
As we see, starting from real $(\omega,\theta) \ge 0$, and regardless of the $\pm$
superscripts, the reduced problems have real
$\tilde{\omega} \ge 0$, while $(\theta_z,c_z) \in \CBB$. If the Fourier inversion is
performed along the real $z$-axis, and of course $\beta$ is real, we also have
$\Re \, \theta_z \ge 0$ and $\Re \, c_z \ge 0$. In terms of $G$, (\ref{eq:prep}) reads:
\be p(t,x',y'|x,y) = \int_{-\infty}^{\infty} \e^{-\i z (x'-x - \alpha t)}
\, G(\Smallfrac{1}{2} \xi^2 t, y, y';z) \, \frac{dz}{2 \pi}, \label{eq:soln} \eb
We call $G$ the (auxiliary model) Green function, although we may also refer to its Laplace
transform as a Green function.
\newpage
\newpage
\subsection{Spectral expansion for $G$ (overview)}
Suppressing the $z$-dependence, $G(\tau,y,y')$ has a Laplace transform with respect to $\tau$,
call it $\hat{G}(s,y,y')$, where $s$ is the transform variable. When we
invert the transform via a Bromwich contour in the complex $s$-plane, we'll
discover various singularities which are associated to the spectrum.
See Fig. \ref{fig:XGBMInversionContour} for a typical case.
The simplest situation has real parameters, $(\omega,\theta,c) \ge 0$.
Then, we'll find that $\hat{G}(s,y,y')$ has
\begin{enumerate}
\item A branch cut singularity at $s = s_c \equiv -\Smallfrac{1}{4}(\omega-1)^2$.
\item A set of poles (or sometimes an empty set) at $s_n, (n=0,1,\cdots,n_{max})$. These poles are
are a finite set of non-positive reals, lying in $s \in (s_c,0)$.
\end{enumerate}
The net result of the Laplace inversion is a spectral expansion of the form:
\be G(\tau,y,y') =
1_{pc} \sum_{n=0}^{n_{max}} \phi_{n}(y,y') \, \e^{s_n \tau}
+ \e^{s_c \tau} \,
\int_{0}^{\infty} \phi_{\nu}(y,y') \, \e^{-\nu^2 \tau} \, d \nu, \label{eq:Gqualitative}
\eb
The specifics are given in Sec. \ref{sec:Gresult} and developed in detail in Appendix A..
\Pbreak
The poles generate the discrete spectrum and the discrete sum term in (\ref{eq:Gqualitative}).
Here $1_{pc}$ denotes a ``pole condition", which is a criterion on the parameters for the
discrete contribution to appear at all. This term may be absent.
\Pbreak
The branch cut singularity marks the right edge of the
$s$-plane interval $(-\infty,s_c)$, which is the continuous spectrum, leading to the integral term in (\ref{eq:Gqualitative}). This term is always present.
\Pbreak
As we saw in Sec. \ref{sec:reduction}, for XGBM we need $\omega \ge 0$,
but $(\theta,c) \ra (\theta_z,c_z)$, complex functions of the Fourier parameter $z$.
In that case, the general form (\ref{eq:Gqualitative}) still holds, but the structure of the spectrum is altered.
A continuous spectrum remains along $(-\infty,s_c)$. However, the poles generally leave the
real $s$-axis, as illustrated in Fig. \ref{fig:XGBMInversionContour}.
The figure illustrates two poles, shown as the `crossed-circles'.
\Pbreak
Indeed, since (complex) $z$ varies continuously under a Fourier inversion, we have \emph{pole trajectories} $s_n(z)$. Examples are shown in Fig. \ref{fig:SpectrumTrajectories} for a case with (initially) three poles. The illustrated trajectories start in the real interval $(s_c,0)$ as in the real parameter case, then
leave the real $s$-axis, and finally return to touch the branch cut -- at which point they leave the
discrete spectrum. In turn, this leads to other complications which
are seen in Fig. \ref{fig:TransitionPt} and discussed later. Overall, the complex parameter case
is, well, complex!
\newpage
\subsection{Spectral expansion for $G$ (result)} \label{sec:Gresult}
As shown, the transition density we seek is found in terms of a Green function for an auxiliary 1D PDE with (two) complex coefficients. Simplifying the notation,
the reduced problem (\ref{eq:Greenfunc0}) reads:
\bea \left \{
\begin{array}{l}
G_{\tau} = y^2 G_{yy} + (\omega y - \theta y^2) G_y - c y^2 G, \\
G(0,y,y') = \delta(y-y').
\end{array} \right. \label{eq:Greenfunc}
\aeb
As before, real $\omega \ge 0$, while $(\theta,c) \in \CBB$ with $\Re \, c \ge 0$.
Solving (\ref{eq:Greenfunc}) is the key to the XGBM model. From that solution, all the other results in this article follow. I find:
\begin{theorem}[auxiliary Green function] \label{thm:Green}
Let $\lfloor x \rfloor$ denote the greatest integer in $x$, $(a)_n = a (a+1) \cdots (a+n-1)$
is the Pochhammer symbol, and the further notations
\[ \gamma = -\Smallfrac{1}{2} (\theta - \sqrt{\theta^2 + 4 c}), \quad R = \theta + 2 \gamma,
\quad \psi = \frac{\omega \gamma}{R}, \quad r_n = -(n+\psi),
\]
\[ s_n = (n+\psi)^2 + (1-\omega)(n+\psi), \,\, b_n = \omega + 2 r_n, \,\, a_{\nu} = \Smallfrac{1}{2}(1-\omega) + \psi + \i \nu, \,\, b_{\nu} = 1 + 2 \i \nu. \]
\newline
Assume real $\omega \ge 0$, while $(\theta,c) \in \CBB$ with
$\Re \, c \ge 0$. Then, the spectral representation solution to PDE problem $($\ref{eq:Greenfunc}$)$ is given by
\begin{empheq}[box=\fbox]{align}
&G(\tau,y,y'; \omega,\theta,c) = \nonumber \\
&\,\, R \, (R y')^{\omega-2} \e^{-\gamma y} \e^{-(\theta+\gamma)y'}
\times 1_{\{\frac{\omega-1}{2} > \Re \psi\}}
\sum_{n=0}^{\lfloor (\omega-1)/2 - \Re \psi \rfloor} \phi_{n}(y,y') \, \e^{s_n \tau} \nonumber \\
&+ (y')^{\omega-2} (y y')^{(1-\omega)/2} \, \e^{-\gamma y} \e^{-(\theta+\gamma)y'}
\e^{-(1-\omega)^2 \tau/4} \,
\int_{0}^{\infty} \phi_{\nu}(y,y') \, \e^{-\nu^2 \tau} \, d \nu, \label{eq:Gtotal}
\end{empheq}
where
\begin{empheq}{align*}
\phi_{n}(y,y') &= \frac{(\omega-1-2n-2\psi) (b_n)_n}{n! \, \Gamma(b_n) }
\, (R^2 y y')^{r_n} \, M(-n,b_n, R y) \, M(-n,b_n,R y')
\end{empheq}
and
\be \phi_{\nu}(y,y') = \frac{1}{2 \pi} \frac{\Gamma(a_{\nu}) \Gamma(a_{-\nu})}{\left| \Gamma(2 i \nu) \right|^2} \,
(R^2 y y')^{\i \nu} \, U \left( a_{\nu}, b_{\nu}, R y \right) \, U \left( a_{\nu}, b_{\nu}, R y' \right). \nonumber \eb
\label{prop:ae2}
\end{theorem}
Proof: In outline, we construct the solution by Laplace transform of (\ref{eq:Greenfunc}) with respect to $\tau$. This leads to an ODE reducible to Kummer's differential equation, with Laplace transform parameter
$s$. The ODE solution
$\hat{G}(s,y,y')$ is constructed; we perform the Laplace inversion by a Bromwich contour, which yields a first solution for $G$. Finally the spectral representation (\ref{eq:Gtotal}), which is a second solution
for $G$, is found by applying the Residue Theorem to the Bromwich inversion. Details are found in Appendix A.
\newpage
\subsection{Summary for the transition density}
To summarize at this point, the transition density for the P-model XGBM system (\ref{eq:XGBM2}) is given by
\begin{empheq}[box=\fbox]{align}
&p_{XGBM}(t,x',y'|x,y) = \int_{-\infty}^{\infty} \e^{-\i z (x'-x - \alpha t)} \,
G \left( \Smallfrac{1}{2} \xi^2 t, y, y'; \tilde{\omega},\theta_z^-,c_z^+ \right) \, \frac{dz}{2 \pi}, \label{eq:solnfinal} \\
\quad &\mbox{where} \,\, \tilde{\omega} = \Smallfrac{2 \omega}{\xi^2}, \quad
\theta_z^- = \Smallfrac{2}{\xi^2} (\theta - \i z \rho \xi), \quad c_z^+ = \Smallfrac{2}{\xi^2}
\left( \Smallfrac{1}{2} z^2 - \i z (\beta - \Smallfrac{1}{2}) \right), \nonumber \\
& \mbox{and} \,\, G(t,y,y'; \omega, \theta,c) \,\, \mbox{is given at} \,\, (\ref{eq:Gtotal}). \nonumber
\end{empheq}
\vspace{-30pt}
\Pbreak
For numerics, it is more efficient to use
\begin{empheq}{align*}
p_{XGBM}(t,x',y'|x,y) = \int_{0}^{\infty} \Re \left\{ \e^{-\i z (x'-x - \alpha t)} \,
G \left( \Smallfrac{1}{2} \xi^2 t, y, y'; \tilde{\omega},\theta_z^-,c_z^+ \right) \right\} \, \frac{dz}{ \pi}.
\end{empheq}
\subsection{Spectral expansion for $G$ (special cases)}
\subsubsection{The stationary limit.} Suppose $\omega > 1$, $\theta > 0$, and $c=0$ $\Rightarrow \gamma = \psi = 0$. In the large $\tau$ limit, (\ref{eq:Gtotal}) yields
\be \Psi(y') \equiv \lim_{\tau \rightarrow \infty} G(\tau,y,y') =
\frac{\theta \, (\theta y')^{\omega-2}}{\Gamma(\omega-1)} \, \e^{-\theta y'}, \quad (\omega > 1). \label{eq:stationayvol} \eb
This is the stationary density for the (standardized) stand-alone volatility process; i.e., when $\xi^2=2$. When $0 \le \omega \le 1$, then (\ref{eq:Gtotal}) still holds without a stationary limit. In that case, the $Y_t$ particle mass eventually accumulates arbitrarily close to $Y = 0$.
For general $\xi^2$, (\ref{eq:stationayvol}) holds with
\be (\omega,\theta) \ra (\tilde{\omega},\tilde{\theta})
\equiv \frac{2}{\xi^2}(\omega,\theta) \,\, \mbox{and} \,\,
\Ex{\sigma_{\infty}} = \int y \, \Psi(y) \, dy =
\frac{\tilde{\omega}-1}{\tilde{\theta}}, \,\, (\tilde{\omega}>1). \label{eq:sigmabar} \eb
\subsubsection{The norm-preserving case: $\bm{c=0}$.} Suppose $(\omega,\theta) > 0$ in (\ref{eq:Greenfunc}) are real and $c=0$. With no killing, the Green function should be norm-preserving on $\RBB_+$: $\int_0^{\infty} G(t,y,y') \, dy'=1$.
Let's check that for the sub-case $\omega > 1$.
\Pbreak
With $c=0$ and $\theta > 0$, then $\gamma = \psi = 0$. Then, using the stationary
density from (\ref{eq:stationayvol}), Theorem 1 reads:
\begin{empheq}{align}
&G(\tau,y,y') = 1_{\{ \omega > 1\}} \left[ \Psi(y')
+ \theta \, (\theta y')^{\omega-2} \e^{-\theta y'}
\sum_{n=1}^{ \lfloor (\omega-1)/2 \rfloor} \phi_{n}(y,y') \, \e^{s_n \tau} \right] \nonumber \\
&\quad + (y)^{(1-\omega)/2}
(y')^{-3/2 + \omega/2} \, \e^{-\theta y'} \e^{-(1-\omega)^2 \tau/4} \,
\int_{0}^{\infty} \phi_{\nu}(y,y') \, \e^{-\nu^2 \tau} \, d \nu, \label{eq:Gtotalspecial}
\end{empheq}
where
\begin{empheq}{align*}
\phi_{n}(y,y') &= \frac{(\omega-1-2n) (b_n)_n}{n! \, \Gamma(b_n) }
\, (\theta^2 y y')^{-n} \, M(-n,b_n, \theta y) \, M(-n,b_n,\theta y')
\end{empheq}
and
\be \phi_{\nu}(y,y') = \frac{1}{2 \pi} \frac{\Gamma(a_{\nu})
\Gamma(a_{-\nu})}{\left| \Gamma(2 i \nu) \right|^2} \,
(\theta^2 y y')^{\i \nu} \, U \left( a_{\nu}, b_{\nu}, \theta y \right) \,
U \left( a_{\nu}, b_{\nu}, \theta y' \right) \nonumber \eb \label{prop:aespecial}
now with
\[ s_n = n (n+1-\omega), \quad b_n = \omega - 2 n, \quad
a_{\nu} = \Smallfrac{1}{2}(1-\omega) + \i \nu, \quad b_{\nu}=1+ 2 \i \nu \]
We are also using the convention that $\sum_n^m (\cdots) = 0$ if $m < n$.
\Pbreak
Recall we suppose $\omega > 1$, so the stationary density exists. Since $\int \Psi(y') \, dy'=1$, the
remaining terms must have a zero $(y')$-integral. Indeed this is true because,
\[ (i) \quad \int (y')^{-3/2 + \omega/2 + \i \nu} \e^{-y'} U(a_{\nu},b_{\nu}, y') \, dy' = 0, \]
which follows from taking the parameters $\psi \ra 0$ and $\zeta=1$ in the integral (\ref{eq:SlaterApplied}) below, and
\[ (ii) \int (y')^{\omega - 2 - n } \e^{-y'} M(-n,\omega - 2 n, y')\, dy' = 0,
\quad \mbox{for} \quad n = 1,2, \cdots, \lfloor \omega-1 \rfloor. \]
Here (ii) follows from the relation
\be \int (y')^{\omega - 2 - n } \e^{-y'} M(-n,\omega - 2 n, y') \, dy'
=
\frac{(\omega - 2 \, n)}{(\omega-1-n) \, \Gamma(1-n)}, \quad (\omega > 1 + n), \eb
which we found from Mathematica, where $n$ is arbitrary real,
and then specializing to $n=1,2,\cdots, \lfloor \omega-1 \rfloor$. \qed
\pbold{Remarks for the sub-case: $0 \le \omega < 1$.} Norm preservation for this
sub-case is best shown with the fundamental transform $H$ developed later in Sec. \ref{sec:case2}.
With that, the property amounts to showing that $H(T,y;z=0)=1$. But, referring to
(\ref{eq:Hinf}), it is easy to see that $h(y;z=0)=1$ and then $H(T,y;z=0)=1$
is immediate from (\ref{eq:HCase2})-(\ref{eq:Hreg}).
\subsubsection{The SABR model limit: no drifts} Drop the drifts
in (\ref{eq:XGBM1}) and you have the (lognormal) SABR model. Then (\ref{eq:Gtotal})
yields previously
known SABR results. Details are found in Appendix B.
\newpage
\section{Risk-neutral evolution}
Given a P-measure diffusion, to avoid arbitrage opportunities, options must be valued under an equivalent Q-measure diffusion. The usual short-hand trick to finding it are Girsanov substitutions: $dB_t \ra dB^*_t - \lambda^e_t$ and $dW_t \ra dW^*_t - \lambda^v_t$
in (\ref{eq:XGBM1}). Here $(B^*_t,W^*_t)$ are Q-Brownian motions and $(\lambda^e_t,\lambda^v_t)$ are, respectively, market prices of equity and volatility risk.
We assume a world with a constant ``cost-of-carry" $b = r - q$, where
$r$ is a short-term interest rate, and $q$ is a constant dividend yield thrown off by the asset.
Then, given our
parametrization in (\ref{eq:XGBM1}), the absence of arbitrage requires
\[ \lambda^e_t = \frac{\alpha + \beta \sigma^2_t - (r-q)}{\sigma_t} \]
In principle, any functional choice $\lambda^v_t = \lambda^v(t,S_t,\sigma_t)$ that preserves the inaccessibility of
all the spatial boundaries will bring closure to the model and satisfy the ``no-arbitrage" principle: P-Q equivalence
as measures. In practice, one wants to preserve `closed formness'. Indeed, a convention in financial modelling (although not required by the general theory) is to
arrange both the P and Q evolutions to have similar parameterizations. This allows a successful
solution method for one to work for the other. In this spirit, our
simple choice here takes $\lambda^v$ a constant. With that,
\bea \mbox{under Q:} \quad \left \{
\begin{array}{l}
d S_t = (r-q) S_t \, dt + \sigma_t S_t \, dB^*_t, \\
d \sigma_t = \sigma_t (\omega_Q - \theta \, \sigma_t) \, dt + \xi \sigma_t \, dW^*_t, \\
dB^*_t \, dW^*_t = \rho \, dt,
\end{array} \right. \label{eq:XGBM3}
\aeb
where $\omega_Q = \omega - \lambda^v \xi$, and now (\ref{eq:XGBM1}) and (\ref{eq:XGBM3}) together complete the model.
To simplify the notations, let us agree that all the evolutions and Brownian motions in this section are `under Q'. Then we have, equivalently, the Q-model:
\bea \left \{
\begin{array}{ll}
d X_t = (r - q - \frac{1}{2} \, Y^2_t) \, dt + Y_t \, dB_t, \\
d Y_t = (\omega_Q Y_t - \theta \, Y^2_t) \, dt + \xi \, Y_t \, dW_t, \\
dB_t \, dW_t = \rho \, dt.
\end{array} \right. \label{eq:XGBM4}
\aeb
\newpage
\subsection{The issue of martingality} \label{sec:martingality}
The XGBM model nests the lognormal SABR model as a special case: see
(\ref{eq:SABR}) in Appendix B. Recall that, under the lognormal SABR
model, the stock price process can suffer `loss of martingality' due to the
correlation parameter $\rho$. Specifically, $S_t$ is: (i) a true
martingale when $-1 \le \rho \le 0$, and (ii) a strictly local martingale when
$0 < \rho \le 1$. Since every local martingale that is bounded from below is a
supermartingale: $\Esub{S_0,\sigma_0}{S_t} < S_0$ (when $\rho>0)$.
Indeed, when $\rho > 0$,
I calculate explicitly in \cite{lewis:2016} the martingale defect. For the full XGBM
model, correspondingly: when is the discounted stock price process
$\tilde{S}_t \equiv \e^{(q-r)t} S_t$ a true martingale? Equivalently, fixing a forward settlement date $T$, when is the forward price process $F_{t,T} = \e^{(r-q)(T-t)}S_t$ a true martingale? Suppressing $T$, the evolution for that is:
\bea \mbox{under Q:} \quad \left \{
\begin{array}{l}
d F_t = \sigma_t F_t \, dB^*_t, \\
d \sigma_t = \sigma_t (\omega_Q - \theta \, \sigma_t) \, dt + \xi \sigma_t \, dW^*_t, \\
dB^*_t \, dW^*_t = \rho \, dt,
\end{array} \right. \label{eq:XGBMF}
\aeb
A prescription for answering this question is given in \cite{lewis:2000a},\cite{lewis:2016}
for the general stochastic volatility evolution
$d\sigma_t = b(\sigma_t) \, dt + a(\sigma_t) \, dW_t$.
To determine if this process is martingale-preserving (for the forward),
introduce the \emph{auxiliary} volatility process $\{\hat{\sigma}_t\}$:
\[ d \hat{\sigma}_t = [b(\hat{\sigma}_t)+ \rho \, \hat{\sigma}_t a(\hat{\sigma}_t)] dt + a(\hat{\sigma}_t) dW_t. \]
Then, the forward $F_t$ suffers a loss of martingality if and only if the auxiliary process can
`explode'; i.e., reach $\hat{\sigma} = +\infty$ in finite time with strictly positive probability. From
(\ref{eq:XGBMF}), for the case under consideration,
\[ d \hat{\sigma}_t = (\omega_Q \, \hat{\sigma}_t - \delta \, \hat{\sigma}^2_t) dt +
\xi \hat{\sigma}_t dW_t, \quad
\mbox{where} \quad \delta = \theta - \rho \, \xi. \]
Then, by applying the \emph{Feller explosion criteria} (details in \cite{lewis:2000a}), one
can establish that $\hat{\sigma}_t$ can explode if and only if $\delta < 0$; i.e.,
$\rho > \theta/\xi$.
Even better, when $\rho > \theta/\xi$, one can find the \emph{martingale defect} by,
equivalently, finding the \emph{absorption-at-zero probability} $A(x,t)$ for the inverse process
$X_t = 1/\hat{\sigma}_t$. From \Ito,
\be dX_t = (\alpha X_t + \delta) dt + \xi X_t \, dW_t, \quad
\mbox{where} \quad \alpha = \xi^2 - \omega_Q. \label{eq:auxX} \eb
With $\tau_0$ the first time the $X$-process hits zero, $A(x,t)$ is defined by
\[ A(x,t) = \Pr{X(\tau_0) \le t| X_0 = x }. \]
Then, the martingale defect is given by
$\Esub{S_0,\sigma_0}{S_t/S_0} = 1 - A(1/\sigma_0,t)$, with an explicit formula found
in Appendix C.
In summary, the XGBM model, parametrized at (\ref{eq:XGBM3})
suffers a loss of martingality for the forward only when $\rho$ is found in the interval
$\theta/\xi < \rho \le 1$.
For (broad-based) equity applications, the condition is typically harmless:
one will usually estimate both $\rho < 0$ (negative option `skews') and $\theta > 0$ (volatility
has a stationary density). But for other applications (possibly with
positive option skews or negative $\theta$'s), it is an issue to keep in mind.
\newpage
\section{Option valuation -- Case 1: $\omega_Q \ge \xi^2/2$}
Let $q(t,x',y'|x,y)$ denote the XGBM transition density under the Q-evolution. From (\ref{eq:solnfinal}),
with $\alpha \ra (r-q)$, $\beta \ra 0$, and taking dummy $z \ra -z$, we have
\begin{empheq}{align}
&q(t,x',y'|x,y) = \int_{-\infty}^{\infty} \e^{\i z (x'-x - (r-q) t)} \,
G \left( \Smallfrac{1}{2} \xi^2 t, y, y'; \tilde{\omega},\theta_z^+,c_z^- \right) \, \frac{dz}{2 \pi}, \label{eq:solnfinalq} \\
\quad &\mbox{where} \,\, \tilde{\omega} = \Smallfrac{2 \omega_Q}{\xi^2}, \quad
\theta_z^+ = \Smallfrac{2}{\xi^2} (\theta + \i z \rho \xi), \quad c_z^- = \Smallfrac{1}{\xi^2}
\left( z^2 - \i z \right). \label{eq:qparms}
\end{empheq}
Here $G(t,y,y'; \omega, \theta,c)$ is again given at (\ref{eq:Gtotal}). Let $V(x,y,T)$ denote
the time-0 value of a Euro-style option with time-$T$ {\underline{bounded}} payoff function $w(x)$.
Then, since $V$ is the time-0 discounted expected value of the payoff under the
$Q$-evolution, we have generally
\begin{empheq}{align}
& V(x,y,T) = \frac{\e^{-r T}}{2 \pi} \nonumber \\
&\times \smallint_{-\infty}^{\infty} \smallint_{0}^{\infty} \smallint_{-\infty}^{\infty}
\e^{\i z (x'-x - (r-q) T)} w(x') \,
G \left( \Smallfrac{1}{2} \xi^2 T, y, y'; \tilde{\omega},\theta_z^+,c_z^- \right) \, dz \,dy' \, dx'. \label{eq:optionval}
\end{empheq}
While (\ref{eq:optionval}) offers a quasi-analytic solution, it has an embarrassing number of
integrations: 4(!), counting the integration for $G$ itself. We'll eliminate two.
\Pbreak
Proceeding formally, suppose it's legitimate to do the
$y'$ integral in (\ref{eq:optionval}) \emph{first}. Then, \emph{if the following integral
exists}, define a \emph{fundamental transform}
\be \mbox{?} \quad H(T,y;z) \equiv \int_{0}^{\infty}
G \left( \Smallfrac{1}{2} \xi^2 T, y, y'; \tilde{\omega},\theta_z,c_z \right) \, dy',
\quad z \in \mathcal{S}_1, \label{eq:H} \eb
where $\mathcal{S}_1$ is an analyticity strip for $H$ in the complex $z$-plane.
Note that we are now generalizing (\ref{eq:optionval}) by moving the $z$-integration off
the real $z$-axis.
For options, our preferred strip for Fourier inversions is: $0 < \Im z < 1$. It can be shown that
our preferred strip is contained within $\mathcal{S}_1$ for XGBM because the model is
(i)norm-preserving and (ii) martingale-preserving
(under the mild restriction $-1 \le \rho \le \min(1,\theta/\xi))$.\footnote{To
show that (i) and (ii) lead to analyticity in the preferred strip, see Th. 4.7
in \cite{lewis:2016}.}
We will show below that the question mark in front of (\ref{eq:H}) can
be removed for Case 1.
Next, in a familiar argument, define the (generalized) \emph{payoff transform}
\be \hat{w}(z) = \int_{-\infty}^{\infty} \e^{i z x} w(x) \, dx, \quad z \in \mathcal{S}_w, \label{eq:wtransform} \eb
with $z$ located in some other analyticity strip $\mathcal{S}_w$. We can always
find a nice payoff such that these two horizontal strips intersect: $\mathcal{S}_V = \mathcal{S}_1 \cap \mathcal{S}_w \not= \emptyset$.
Finally, by the arguments in my article ``A Simple Option Formula for General Jump-diffusion and other Exponential \Levy Processes" (reprinted in \cite{lewis:2016}), (\ref{eq:optionval}) becomes
\be V(x,y,T) = \frac{\e^{-r T}}{2 \pi} \int_{\i c -\infty}^{\i c + \infty}
\e^{-\i z (x + (r-q) T)}
H(T,y;z) \, \hat{w}(z) \ dz, \quad z \in \mathcal{S}_V. \label{eq:optionval3} \eb
In other words, $\i c$ is a point in the $z$-plane marking the intersection of a horizontal integration contour in $\mathcal{S}_V$ with the imaginary $z$-axis. Again, we can arrange things
so that we are working in the preferred strip: $0 < c < 1$. Reported numerics are found from versions of (\ref{eq:optionval3}).
For example, with $x = \log S_0$, $X = (r-q)T + \log S_0/K$ with strike price $K$, and
$c = \Smallfrac{1}{2}$, all reported put option values $P(x,y,T)$ are computed from:
\be P(x,y,T) = K \e^{-r T} \left( 1 - \int_{\i c}^{\i c + \infty}
\Re \left\{ \e^{-\i z X}
\frac{H(T,y;z)}{z^2 - \i z} \right\} \, \frac{dz}{\pi} \right). \label{eq:putval} \eb
\pbold{Removing the question mark in (\ref{eq:H}).} Recall that the SABR model is a special
case of the XGBM model, with the transition density given at (\ref{eq:pSABRalt}) in Appendix B.
It turn out that, for the SABR model, the integral in (\ref{eq:H}) does \emph{not} exist.
This problem was discussed in \cite{lewis:2016} (Sec. 8.13, pg. 407), but we will recap briefly
here. For SABR, the existence of (\ref{eq:H}) amounts to integrating (\ref{eq:pSABRalt}) from
Appendix B with respect to
$y'$, at fixed $(\nu,z)$. Extracting the few terms in (\ref{eq:pSABRalt})
with $y'$ dependence yields the integral
\[ \int_0^{\infty} (y')^{-3/2} e^{-\i z \rho y'/\xi} K_{\i \nu}(\chi_z y') \, dy', \]
which does not exist because the integrand is not integrable
as $y' \downarrow 0$.\footnote{\label{ft:bes}$K_{\i \nu}(y)
\sim c_1 y^{-\i \nu} + c_2 y^{\i \nu}$, as $y \ra 0$, where $c_{1,2}$ depend upon $\nu$ but not $z$.}
Now, for the SABR limit, there is no discrete spectral contribution to $G$
in (\ref{eq:Gtotal}). The lesson from SABR is that, in the general XGBM case, problems may lie with
integrating the continuous spectrum term. Indeed, extracting all the $y'$ dependencies
from the continuous spectrum term in (\ref{eq:Gtotal}), the corresponding XGBM integral is
\[ (*) \quad \int_0^{\infty} (y')^{\tilde{\omega}/2-3/2}
e^{-(\theta_z^+ + \gamma_z) y'} (y')^{\i \nu} U(a_{\nu},b_{\nu}, R_z y' ) \, dy'. \]
\[ \mbox{Recall} \quad \tilde{\omega} = \Smallfrac{2 \omega_Q}{\xi^2}, \quad
\theta_z^+ = \Smallfrac{2}{\xi^2} (\theta + \i z \rho \xi), \quad
c_z^- = \Smallfrac{1}{\xi^2}
\left( z^2 - \i z \right), \]
\[ \mbox{and so} \quad \gamma_z = -\Smallfrac{1}{2} \left( \theta_z^+ -
\sqrt{(\theta_z^+)^2 + 4 c_z^-} \right), \quad R_z = \theta_z^+ + 2 \gamma_z = \sqrt{(\theta_z^+)^2 + 4 c_z^-}.
\]
Now, first consider the integrand as $y' \downarrow 0$. As $x \ra 0$,
$U(a,b,x) \sim c_1 + c_2 x^{1-b}$, where $c_{1,2}$ are bounded constants independent of $x$.
Since $b_{\nu} = 1 + 2 \i \nu$,
\[ (y')^{\i \nu} U(a_{\nu},b_{\nu}, R_z y' ) \sim c_1 (y')^{\i \nu} +
c_2 (y')^{-\i \nu}, \]
much like the SABR case with Bessel functions discussed in footnote \ref{ft:bes}. The
exponential in harmless at small $y'$, so we see the crux of the
matter is the term $(y')^{\tilde{\omega}/2-3/2}$, which is integrable
iff\footnote{Integrability is clear when $\tilde{\omega}>1$. The
borderline case $\tilde{\omega}=1$ is not as immediate, but (\ref{eq:SlaterApplied}) below
shows that (*) also exists when $\tilde{\omega}=1$ as long as $\nu>0$.}
\begin{empheq}[box=\fbox]{align}
\tilde{\omega} \ge 1, \quad (\mbox{i.e.,} \,\,
\omega_Q \ge \Smallfrac{1}{2} \xi^2), \quad \mbox{(Case 1: integrability condition)}
\label{eq:integrability}
\end{empheq}
When the integrability condition holds, so does (\ref{eq:H}).
Recall we have seen this condition before at (\ref{eq:stationayvol}) in the P-model: there
needed for the existence of a stationary limit of the stand-alone volatility process. In
most financial applications (with calibrated parameters), the integrability condition will be satisfied.
Thus, to borrow a characterization from physics, Case 1 is the ``physical case". Case 2 is the
``unphysical case" (including SABR). Case 2 is, of course, still mathematically well-defined.
\Pbreak
What about $y' \ra \infty$? Since
\[ \theta_z^+ + \gamma_z = \Smallfrac{1}{2} \left( \theta_z^+ +
\sqrt{(\theta_z^+)^2 + 4 \, c_z^-} \right), \]
this term will have positive real part in the preferred $z$-plane strip. As the
remaining integrand terms in (*) have power law behavior as $y' \ra \infty$, the
net effect is exponential decay in $y'$ at large $y'$, so all is well in that limit.
\Pbreak
To summarize at this point, we have the following situation.
In the risk-neutral model, when the integrability condition (\ref{eq:integrability}) holds,
option values are given by (\ref{eq:optionval3}), with $H$ given by (\ref{eq:H}). If the integrability condition does \emph{not} hold
(Case 2), option values are \emph{still} given by (\ref{eq:optionval3}),
but $H(\cdot)$ is \emph{not}
given by (\ref{eq:H}). Instead, that integral needs to be regularized:
see Sec. \ref{sec:case2} for details. In this section, we now continue, assuming
(\ref{eq:integrability}) holds.
\pbold{Option valuation -- continued}.
To perform the integration in (*), use a formula from
Slater (\cite{slater:1960}, (3.2.51)-(3.2.52)):
\begin{empheq}{align*}
& \int_0^{\infty} \e^{-s t} t^{c-1} U(a,b,t) \, dt \nonumber \\
&= \frac{\Gamma(c) \Gamma(1+c-b)}{\Gamma(1+a+c-b)} \times \left \{
\begin{array}{ll}
F(c,1+c-b,1+a+c-b;1-s), \quad &|1-s| < 1, \\
s^{-c} F \left(a,c,1+a+c-b;1-\frac{1}{s} \right), \quad &\Re \, s > \Smallfrac{1}{2},
\end{array} \right.
\end{empheq}
\be \mbox{for} \,\, \Re c>0, \quad \Re b < \Re c + 1. \label{eq:SlaterUintegral} \eb
Here $F(a,b,c;z) = \sum_{n=0}^{\infty}\frac{(a)_n (b)_n}{(c)_n} \frac{z^n}{n!}$ is
the Gauss hypergeometric function. Note that the two conditions on $s$ above
can overlap, allowing both integral relations to
be true at the same time.
Applying (\ref{eq:SlaterUintegral}) to our problem (*) above yields:
\begin{empheq}{align*}
\int_0^{\infty} (y')^{\tilde{\omega}/2-3/2+\i \nu}
& e^{-(\theta_z^+ +\gamma_z) y'} U(a_{\nu},b_{\nu}, R_z y' ) \, dy' \nonumber \\
&=\frac{|\Gamma(c_{\nu})|^2}{R_z^{c_{\nu}}\Gamma(\psi)} \times \left \{
\begin{array}{ll}
F(c_{\nu},c^*_{\nu},\psi_z;1- \zeta_z), \quad &|1 - \zeta_z| < 1, \\
s^{-c_{\nu}} F \left(a_{\nu},c_{\nu},\psi_z;1-\frac{1}{\zeta_z} \right), \quad &\Re \, \zeta_z > \Smallfrac{1}{2},
\end{array} \right.
\end{empheq}
\be \mbox{where} \quad
a_{\nu} = \Smallfrac{1}{2}(1-\tilde{\omega}) + \psi_z + \i \nu, \quad
c_{\nu} =- \Smallfrac{1}{2}+\Smallfrac{1}{2} \tilde{\omega}+\i \nu, \quad
\psi_z = \frac{\tilde{\omega} \gamma_z}{R_z}, \label{eq:SlaterApplied} \eb
\be \mbox{and} \quad \zeta_z = \frac{1}{2}
\left( 1 + \frac{\theta_z^+}{\sqrt{(\theta_z^+)^2 + 4 \, c_z^-}} \right). \label{eq:sdef} \eb
Note $c^*_{\nu} = - \Smallfrac{1}{2}+ \Smallfrac{1}{2} \tilde{\omega}-\i \nu$ denotes the
complex conjugate of $c_{\nu}$ (since $\tilde{\omega}$ is real). The integration
conditions $\Re c>0$ and $\Re b < \Re c + 1$ from Slater reduce to: $\tilde{\omega}>1$.
(Recall in fact $\tilde{\omega}=1$ is OK for Case 1 as long as $\nu \not= 0$).
The $\zeta_z$-value restrictions can be checked along any putative
$z$-plane integration contour; I have found the $|1 - \zeta_z|<1$ formulas suffice for all
the test examples of Sec. \ref{sec:numerics}.
\pbold{The discrete term integration.}
From (\ref{eq:Gtotal}) and (\ref{eq:H}), the discrete term in the spectral expansion requires
\[ (**) \quad \int_0^{\infty} (y')^{\tilde{\omega}-2+r_n}
e^{-(\theta_z^+ +\gamma_z) y'} M(-n,b_n, R_z y' ) \, dy', \]
where $r_n = -(n+\psi)$ and $\psi$ is given at (\ref{eq:SlaterApplied}).
The integral (**) can be performed using another formula from Slater, (3.2.16):
\be
\int_0^{\infty} \e^{-s t} t^{c-1} M(-n,b,kt) \, dt =
\Gamma(c) \, s^{-c} F \left(-n,c,b;\frac{k}{s} \right),
\label{eq:SlaterMintegral} \eb
\[ \mbox{for} \,\, \Re c>0, \,\, \Re s > 0, \,\, \mbox{and} \,\, n=0,1,2,\cdots \]
Note that, since $n$ is an integer for our application, $F(-n,c,b,z)$ is a terminating polynomial of $O(z^n)$. Thus,
\begin{empheq}{align*}
\int_0^{\infty} (y')^{\tilde{\omega}-2+r_n}
e^{-(\theta_z^+ +\gamma_z) y'} & M(-n,b_n, R_z y' ) \, dy' \\
&= \Gamma(c_n) \, (R_z \zeta_z)^{-c_n}
F \left(-n,c_n,b_n;\frac{1}{\zeta_z} \right),
\end{empheq}
\[ \mbox{where} \quad c_n = \tilde{\omega} -1 + r_n, \quad b_n= \tilde{\omega} + 2 \, r_n,
\,\, \mbox{and $\zeta_z$ is defined at (\ref{eq:sdef})}. \]
\pbold{Remarks.} Note that the last Slater integral (\ref{eq:SlaterMintegral}) requires
$\Re \, c > 0$, which translates to $\Re \, c_n > 0$. It can be seen that this condition is
automatically satisfied here as follows. First, $\Re \, c_n > 0$ is equivalent to
$\tilde{\omega} - 1 > (n + \Re \, \psi)$. But, if $s_n$ contributes to the discrete spectrum, then
$\tilde{\omega} - 1 > 2(n + \Re \, \psi)$ (see Appendix A or the sum cutoff
in (\ref{eq:Htotal}). When this last inequality is
satisfied, the l.h.s. will be positive. And certainly if $A > B$ with $A$
positive but otherwise arbitrary, then $A > B/2$.
\newpage
\subsection{Frequently used notations under risk-neutrality} \label{sec:notations}
Here we collect in one place notation we have introduced and will
frequently refer to subsequently.
\begin{empheq}[box=\widefbox]{align*}
&\tilde{\omega} = \Smallfrac{2 \omega}{\xi^2}, \quad
\theta_z^+ = \Smallfrac{2}{\xi^2} (\theta + \i z \rho \xi), \quad
c_z^- = \Smallfrac{1}{\xi^2}
\left( z^2 - \i z \right), \\
&R_z = \sqrt{(\theta_z^+)^2 + 4 \, c_z^-}, \quad
\gamma_z = -\Smallfrac{1}{2} \theta_z^+ +
\Smallfrac{1}{2} R_z,
\quad
\zeta_z = \Smallfrac{1}{2}
\left( 1 + \frac{\theta_z^+}{R_z} \right), \quad
\psi_z = \frac{\tilde{\omega} \, \tilde{\gamma_z}}{R_z}, \\
&a_{\nu} = \Smallfrac{1}{2}(1-\tilde{\omega}) + \psi_z + \i \nu, \quad
b_{\nu} = 1 + 2 \i \nu, \quad
c_{\nu} =- \Smallfrac{1}{2}+\Smallfrac{1}{2} \tilde{\omega}+\i \nu, \\
&r_n = -(n + \psi_z), \quad s_n(z) = (n+\psi_z)^2 + (1-\tilde{\omega})(n+\psi_z),\\
&b_n= \tilde{\omega} + 2 \, r_n, \quad
c_n = \tilde{\omega} -1 + r_n.
\end{empheq}
Note that here and throughout, we frequently suppress $z$-dependencies to ease notations,
as we have done with $(a_{\nu},r_n,b_n,c_n)$.
\subsection{Summary: fundamental transform for Case 1}
Putting it all together, for $\omega \ge \xi^2/2$, we have the spectral representation:
\begin{empheq}[box=\widefbox]{align}
&\,\, \, H(T,y;z) = \int_0^{\infty} G(\tau,y,y'; \tilde{\omega},\theta_z^+,c_z^-) \, dy' \label{eq:Htotal} \\
&\,\,= \e^{-\gamma_z y} \left\{
1_{\{\frac{\tilde{\omega}-1}{2} > \Re \psi_z \}} \hspace{-18pt}
\sum_{n=0}^{ \lfloor (\tilde{\omega}-1)/2 - \Re \psi_z \rfloor} \phi_{n}( R_z y) \, \e^{s_n(z) \tau}
+
\int_{0}^{\infty} \phi_{\nu}( R_z y) \,
\e^{-(\nu^2 + \frac{1}{4}(1-\tilde{\omega})^2) \tau} \, d \nu \right\} \nonumber
\end{empheq}
\vspace{-40pt}
\begin{empheq}{align*}
& \mbox{where} \,\, \tau = \Smallfrac{1}{2} \xi^2 T, \\
& \phi_{n}(y) = \frac{(\tilde{\omega}-1-2n-2\psi) (b_n)_n}{n!}
\frac{\Gamma(c_n)}{\Gamma(b_n)}
\,\, \zeta_z^{-c_n} F \left(-n,c_n,b_n;\frac{1}{\zeta_z} \right)
y^{r_n} M(-n,b_n, y), \\
& \phi_{\nu}(y) = \frac{1}{2 \pi}
\frac{\Gamma(a_{\nu}) \Gamma(a_{-\nu})|\Gamma(c_{\nu})|^2}{\left| \Gamma(2 i \nu) \right|^2 \Gamma(\psi_z)} \, y^{(1-\tilde{\omega})/2 + \i \nu}
\, U \left( a_{\nu}, b_{\nu}, y \right) \\
&\quad\quad\quad\quad \times \left \{
\begin{array}{ll}
F(c_{\nu},c^*_{\nu},\psi_z;1-\zeta_z), \quad &|1-\zeta_z| < 1, \\
\zeta^{-c_{\nu}} F \left(a_{\nu},c_{\nu},\psi_z;1-\frac{1}{\zeta_z} \right), \quad &\Re \, \zeta_z > \Smallfrac{1}{2},
\end{array} \right. \\
\end{empheq}
with other notations in Sec. \ref{sec:notations}. Again, some $z$-dependencies are suppressed.
\newpage
\section{Option valuation -- Case 2: $0 \le \omega_Q < \xi^2/2$} \label{sec:case2}
To handle this case, we'll adapt the method I used for the SABR model in \cite{lewis:2016}.
Consider $h(y)$, a putative time-independent solution to the PDE
in the top line of (\ref{eq:Greenfunc}). Suppose $h(y) \sim 1 + O(y^{1- \tilde{\omega}})$
as $y \ra 0$ and $h(y)$ decays at large $y$. Note an initial condition plays no role in this time-independent
problem. (With $\theta \ra \theta_z$ and $c \ra c_z$,
we write this solution as $h(y;z)$). With that putative solution, the
fundamental transform $H(T,y;z)$ can be constructed as
\be H(T,y;z) = h(y;z) + H_{reg}(\tau,y;z), \quad (\tau = \Smallfrac{1}{2} \xi^2 T).
\label{eq:HCase2} \eb
Here a \emph{regularized} fundamental transform $H_{reg}(\tau,y;z)$ is defined by
\be H_{reg}(\tau,y;z) \equiv \int_{0}^{\infty} (1 - h(y';z)) \,
G \left( \tau, y, y'; \tilde{\omega},\theta_z^+,c_z^- \right) \, dy',
\quad z \in \mathcal{S}_1. \label{eq:Hreg} \eb
Let's review why this works. First, both terms on the r.h.s. of (\ref{eq:HCase2})
manifestly satisfy the PDE associated to the fundamental transform -- the top line of (\ref{eq:Greenfunc}) again. Second, as $T \ra 0$, because $G(\cdots)$ satisfies a Dirac mass initial
condition, we have $H_{reg}(0,y;z) = 1 - h(y;z)$. Thus, $H(0,y;z) = 1$, which is the correct initial condition for the fundamental transform. Third, the integral in (\ref{eq:Hreg})
exists because $1 - h(y;z) \sim O(y^{1- \tilde{\omega}})$ as $y \ra 0$, so the problematic integrand in (*)
is tamed by this new behavior. Finally, $H(T,y;z)$ correctly decays at large $y$ because both terms on the r.h.s. of (\ref{eq:HCase2}) do.
\Pbreak
Using the notations in Sec. \ref{sec:notations} and the confluent hypergeometric $U$, I find
\begin{empheq}[box=\fbox]{align}
h(y;z) = \e^{-\gamma_z y}
\frac{U(\psi_z,\tilde{\omega}, R_z y)}{U(\psi_z,\tilde{\omega}, 0)},
\label{eq:Hinf}
\end{empheq}
which has the advertised $y$-behaviors.
Now recall
\be U(a,b,z) = \frac{\Gamma(1-b)}{\Gamma(a-b+1)} M(a,b,z)
+ z^{1-b} \frac{\Gamma(b-1)}{\Gamma(a)} M(a-b+1,2-b,z). \label{eq:Udef} \eb
Since $M$ has a Taylor expansion about $z=0$, this implicitly defines coefficients $(A_n,B_n)$ such that
\[ h(y;z) = \sum_{n=0}^{\infty} A_n(z) \, (R_z y)^n + (R_z y)^{1-\tilde{\omega}}
\sum_{n=0}^{\infty} B_n(z) \, (R_z y)^n. \]
I find (again using the Gauss hypergeometric function $F(a,b,c;z)$) that
\begin{empheq}{align*}
A_n(z) &= f_n(\psi_z, \tilde{\omega}, \zeta_z -1), \\
B_n(z) &= \frac{\Gamma(\tilde{\omega}-1) \Gamma(\psi_z - \tilde{\omega} +1)}
{\Gamma(1-\tilde{\omega}) \Gamma(\psi_z)} f_n(\psi_z-\tilde{\omega}+1, 2-\tilde{\omega}, \zeta_z -1), \quad \mbox{where} \\
f_n(a,b,c) &= c^n \, \frac{F(a,-n,b;-\frac{1}{c})}{\Gamma(1+n)}.
\end{empheq}
Note that, since $f_0=1$, the first sum of $h$ is $1 + O(y)$ and the second sum is
$O(y^{1-\tilde{\omega}})$; since $\tilde{\omega}<1$, the net effect is $h(y) \sim 1 + O(y^{1- \tilde{\omega}})$, as promised.
The term-by-term integrals that are now needed in (\ref{eq:Hreg}) can be done
by the previously given formulas from Slater. The net result is that the previously introduced $\phi_{\nu}(y)$ (for Case 1) generalizes to $\phi_{\nu,n}(y)$ where
\begin{empheq}{align*}
& \phi_{\nu,n}(y) = \frac{1}{2 \pi}
\frac{\Gamma(a_{\nu}) \Gamma(a_{-\nu})|\Gamma(c_{\nu,n})|^2}{\left| \Gamma(2 i \nu) \right|^2 \Gamma(\psi_z)} \, y^{(1-\tilde{\omega})/2 + \i \nu}
\, U \left( a_{\nu}, b_{\nu}, y \right) \\
&\quad\quad\quad\quad \times \left \{
\begin{array}{ll}
F(c_{\nu,n},c^*_{\nu,n},\psi_z+n;1-\zeta_z), \quad &|1-\zeta_z| < 1, \\
(\zeta_z)^{-c_{\nu,n}} F \left(a_{\nu},c_{\nu,n},\psi_z+n;1-\frac{1}{\zeta_z} \right), \quad &\Re \, \zeta_z > \Smallfrac{1}{2},
\end{array} \right. \\
\end{empheq}
In addition, we need to introduce the new function $\psi_{\nu,n}(y)$ where
\begin{empheq}{align*}
& \psi_{\nu,n}(y) = \frac{1}{2 \pi}
\frac{\Gamma(a_{\nu}) \Gamma(a_{-\nu})|\Gamma(d_{\nu,n})|^2}{\left| \Gamma(2 i \nu) \right|^2 \Gamma(1-\tilde{\omega}+n+\psi_z)} \, y^{(1-\tilde{\omega})/2 + \i \nu}
\, U \left( a_{\nu}, b_{\nu}, y \right) \\
&\quad\quad\quad\quad \times \left \{
\begin{array}{ll}
F(d_{\nu,n},d^*_{\nu,n},1-\tilde{\omega}+\psi_z+n;1-\zeta_z), \quad &|1-\zeta_z| < 1, \\
(\zeta_z)^{-d_{\nu,n}} F \left(a_{\nu},d_{\nu,n},1-\tilde{\omega}+\psi_z+n;1-\frac{1}{\zeta_z} \right), \quad &\Re \, \zeta_z > \Smallfrac{1}{2},
\end{array} \right. \\
\end{empheq}
These functions use the new notations:
\[ c_{\nu,n} = n - \Smallfrac{1}{2}+\Smallfrac{1}{2} \tilde{\omega}+\i \nu, \quad (n \ge 1) \]
\[ d_{\nu,n} = n + \Smallfrac{1}{2} - \Smallfrac{1}{2} \tilde{\omega}+\i \nu, \quad (n \ge 0). \]
From these, form
\be \kappa(y;z) \equiv \sum_{n=1}^{\infty} \, A_n(z) \, \phi_{\nu,n}(y) + \sum_{n=0}^{\infty}
\, B_n(z) \, \psi_{\nu,n}(y) \label{eq:kappadef} \eb
and obtain finally, for Case 2,
\begin{empheq}[box=\fbox]{align}
H(T,y;z) = h(y;z) - \e^{-\gamma_z y} \int_0^{\infty}
\kappa_{\nu}(R_z y; z) \, \e^{-(\nu^2 + \frac{1}{4}(1 - \tilde{\omega}^2)) \tau} \, d \nu.
\label{eq:HCase2final}
\end{empheq}
Don't forget that $\tau = \Smallfrac{1}{2} \xi^2 T$ -- both here and in (\ref{eq:Htotal}).
Equation (\ref{eq:HCase2final}) is the Case 2 replacement for (\ref{eq:Htotal}); option values
are computed from this new $H$ using the previous (\ref{eq:optionval3}) and (\ref{eq:putval}).
\newpage
\section{Small and large time asymptotics}
By `time', we refer to the time-to-option-expiration $T$.
\subsection{Small $T$}
For small $T$, option prices at strike $K$ tend to their parity
values, and the implied volatility smile, $\sigma_{imp}(T,K,S_0,\sigma_0)$, tends
to a non-trivial limit.
Indeed, it is well-known that the limiting asymptotic smile in general SV models
does not depend upon the diffusion drifts; it is solely determined
by the variance-covariance system. In our case, that system is (\ref{eq:SABR}), which is the (lognormal) SABR model.
The asymptotic smile formula for that one is well-known. The version I like is\footnote{I am copying in (\ref{eq:smallTIV}) from eq. (12.33) in \cite{lewis:2016} (Ch. 12). It is easily shown to be equivalent
to other SABR model sources. For example, to show it is equivalent to (8.32) in \cite{paulot:2010}, identify my $z$ with $-(2 \nu q/\alpha)$ in Paulot and apply some
routine algebra.}
\be \sigma^{imp}_0 \equiv \lim_{T \ra 0} \sigma_{imp}(T,K,S_0,\sigma_0) =
\frac{ z \, \sigma_0}{F(z)},
\quad \mbox{where} \quad z = \left(\frac{2 \xi}{\sigma_0} \right) \log \frac{S_0}{K}, \label{eq:smallTIV} \eb
\[ \mbox{and} \quad
F(z) = 2 \log \left\{ \frac{z/2-\rho + \sqrt{1 - \rho z + z^2/4}}{1- \rho} \right\}. \]
\subsection{Large $T$}
\label{sec:largeT}
The large $T$ option behavior is determined from the large $T$ behavior
of the fundamental transform $H(T,\sigma;z)$.
\pbold{Case 1: $\tilde{\omega} \ge 1$}. In this case, following the method in \cite{lewis:2000a} (Ch. 6),
one looks for the behavior (as $T$ grows large) $H(T,\sigma;z) \sim \e^{-\lambda_0(z) T} u_0(\sigma;z)$ in the spectral expansion (\ref{eq:Htotal}). When there is a discrete
component in that expansion, then $\lambda_0$ is the principal eigenvalue.
When there is only a continuous spectrum, but still real $\lambda_0 > 0$, a better nomenclature
for $\lambda_0$ (borrowing from physics) might be be `mass gap'. In any event, we define $\lambda_0$ by
\[ \lambda_0(z) \equiv \lim_{T \ra \infty} -\frac{1}{T} \log H(T,\sigma;z). \]
Tentatively, assume the parameters are such that there is indeed a discrete spectrum,
and identify $\lambda_0(z)$ from the leading term of that. Using the notation
associated to (\ref{eq:Htotal}), one has
\[ \lambda_0(z) = -\Smallfrac{1}{2} \xi^2 s_0(z) = -\Smallfrac{1}{2} \xi^2
\left[ \psi^2(z) + (1 -\tilde{\omega}) \psi(z) \right],\]
\[ \mbox{where} \quad \psi(z) = \Smallfrac{1}{2} \tilde{\omega}
\left\{1 - \frac{(\theta + \i z \rho \xi)}{\sqrt{(\theta + \i z \rho \xi)^2 +
(z^2 - \i z) \xi^2} } \right\}. \]
Then, the prescription in the cited reference is to look for a saddle point
$z^* = \i y^*$ in the complex $z$-plane along the purely imaginary axis (so $y^*$ is real).
The saddle point is a solution to $\lambda_0'(z)=0$, where the prime indicates
the $z$-derivative. Having found that, the asymptotic $T \ra \infty$
option implied volatility $V^{imp}_{\infty} = (\sigma^{imp}_{\infty})^2$ is given by the very simple relation
\begin{empheq}[box =\fbox]{align}
V^{imp}_{\infty} = 8 \lambda_0(z^*). \label{eq:VimpInf}
\end{empheq}
With $\psi^* \equiv \psi(\i y^*)$, it's easy to find
\be y^* = \frac{\theta}{2 \, \theta - \rho \xi}
\quad \Rightarrow \quad \psi^* = \frac{\tilde{\omega}}{2} \,
\left\{ 1 - \zeta \right\}.
\label{eq:deltastar} \eb
\[ \mbox{defining} \quad \zeta \equiv
\frac{1}{(1+\frac{\xi^2}{4 \theta (\theta - \rho \xi)})^{1/2}}, \quad
(\theta > \rho \,\xi). \]
Notice that we work in the ``martingale-preserving" regime (recall Sec. \ref{sec:martingality}),
where $-1 \le \rho \le (\theta/\xi)$. Since we assume throughout that $\theta \ge 0$,
in that regime $\zeta \in [0,1)$ and $y^* \in [0,1]$, which is the preferred (and
guaranteed) analyticity strip for $H$.
Thus, we have found that \emph{when $\tilde{\omega}$ is such that a
discrete spectrum exists}, then
\[ \lambda_0^* \equiv \lambda_0(z^*) =
\Smallfrac{1}{2} \xi^2
\left[ (\tilde{\omega} -1) \psi^* - (\psi^*)^2 \right],\]
\[ \mbox{and} \quad \sigma^{imp}_{\infty} = 2 \, \xi \sqrt{(\tilde{\omega}-1) \psi^* - (\psi^*)^2}, \]
where $\psi^*$ is given by (\ref{eq:deltastar}). But, what is the criterion for
a discrete spectrum to exist? From (\ref{eq:Htotal}), and since $\psi^*$ is real,
a discrete spectrum exists when
\[ \tilde{\omega} \ge 1 + 2 \psi^* = 1 + \tilde{\omega} (1 - \zeta)
\quad \Rightarrow \quad \tilde{\omega} \ge \frac{1}{\zeta}. \]
Now, since we are discussing Case 1, which has $\tilde{\omega} \ge 1$, the remaining
sub-case to be addressed here is $1 \le \tilde{\omega} < (1/\zeta)$. Since there
is no discrete spectrum, only the continuous spectrum
term contributes. One reads off from (\ref{eq:Htotal}) that
\[ H(T,\sigma;z) \sim \e^{-(1 - \tilde{\omega})^2 \xi^2 T/8} u_0(\sigma;z); \]
\[ \mbox{i.e., the mass gap} \quad \lambda_0^* =
\Smallfrac{1}{8} \xi^2 (\tilde{\omega}-1)^2, \quad
( 1 \le \tilde{\omega} < \frac{1}{\zeta} ), \]
with some $u_0$ whose specific form doesn't matter for our purpose.
Let's introduce the notation $\tilde{\omega}_c \equiv 1/\zeta$, which
marks the critical value of $\tilde{\omega}$ such that a discrete spectrum emerges.
To summarize, we've found under Case 1 (and the martingale-preserving regime) that
the principal eigenvalue/mass gap is:
\begin{empheq}[box=\fbox]{align}
& \lambda^*(\tilde{\omega}) = \Smallfrac{1}{2} \xi^2
\left \{
\begin{array}{ll}
(\tilde{\omega} -1) \psi^*(\tilde{\omega}) - (\psi^*(\tilde{\omega}))^2, \quad &
\tilde{\omega}_c \le \tilde{\omega} < \infty, \\
\Smallfrac{1}{4} (\tilde{\omega}-1)^2, \quad & 1 \le \tilde{\omega} < \tilde{\omega}_c,
\end{array} \right. \label{eq:lamstar} \\
& \mbox{where} \,\, \tilde{\omega}_c = \frac{1}{\zeta}, \quad
\psi^*(\tilde{\omega}) = \frac{\tilde{\omega}}{2} \, \left\{ 1 - \zeta \right\},
\quad \mbox{and} \,\,\, \zeta =
\frac{1}{(1+\frac{\xi^2}{4 \theta (\theta - \rho \xi)})^{1/2}}. \nonumber
\end{empheq}
\pbold{Smoothness.} Notice that $\psi^*(\tilde{\omega}_c) = (1-\zeta)/(2 \zeta)=
\Smallfrac{1}{2}(\tilde{\omega}_c -1).$ Thus
\[ (\tilde{\omega}_c -1) \psi^*(\tilde{\omega}_c) - (\psi^*(\tilde{\omega}_c))^2
= \Smallfrac{1}{4} (\tilde{\omega}_c -1)^2, \]
which shows that $\lambda^*(\tilde{\omega})$ of (\ref{eq:lamstar})
is continuous at $\omega = \tilde{\omega}_c$. By differentiating, it is
easy to see that $d \lambda^*(\tilde{\omega})/d \tilde{\omega}$ is
also continuous at $\omega = \tilde{\omega}_c$. In other
words $\lambda^*(\tilde{\omega})$ is \emph{smooth}, in the sense
of being continuously differentiable, throughout the Case 1 regime:
$1 \le \tilde{\omega} < \infty$.
\pbold{Examples.}
With $\omega = \xi = 1$ and $\rho = 0$, then $\tilde{\omega}=2$,
$\psi^* = 0.007722$, and $\sigma^{imp}_{\infty} = 0.1751$. Or,
with $\omega = \xi = 1$ and $\rho = -0.75$, then
$\psi^* = 0.006515$, and $\sigma^{imp}_{\infty} = 0.1609$. This is how the
$T=\infty$ entries in Table \ref{tab:NumericalExample1} were found.
\pbold{Case 2.}
For simplicity, suppose no cost-of-carry parameters, so
$r=q=0$. As $T$ gets large under Case 2 the fundamental transform $H$ does not decay,
but instead has a non-trivial, but relatively simply limit:
$H(T,\sigma_0; z) \sim H_{\infty}(z,\sigma_0)$. From (\ref{eq:putval}):
\be P_{\infty}(S_0,\sigma_0;K) = K \left[1 -
\frac{1}{2 \pi} \int_\Gamma \e^{-\i z X} \frac{H_{\infty}(z,\sigma_0)}{z^2 - \i z} \, dz \right]. \label{eq:PutFromHinf} \eb
Here the contour $\Gamma$ is defined by $0 < \Im \, z < 1$, and $X = \log (S_0/K)$.
Note that $H_{\infty}(z,y)$ is not a new function in our development; indeed, $H_{\infty}(z,y) = h(y;z)$, where the r.h.s. was found previously at (\ref{eq:Hinf}). Hence,
\be H_{\infty}(z,\sigma_0) = \e^{-\gamma_z \sigma_0}
\times \left \{
\begin{array}{ll}
\frac{U(\psi_z,\tilde{\omega},R_z \sigma_0)}{U(\psi_z,\tilde{\omega},0)},
\quad & 0 < \tilde{\omega} \le 1, \\
1, \quad & \tilde{\omega}=0,
\end{array} \right.
\label{eq:Hinf2} \eb
and recall Sec. \ref{sec:notations} for other notations.\footnote{The
case $\tilde{\omega}=0$ most easily follows by seeking a
stationary solution $h(y;z)$ to the top line of (\ref{eq:Greenfunc})
with $\omega = 0$. One easily find $h(y;z) = \e^{-\gamma_z y}$. Alternatively,
it can be checked by proving $\lim_{b \ra 0} U(a,b,z)/U(a,b,0) =1$.}
Numerical examples using (\ref{eq:PutFromHinf}-\ref{eq:Hinf2}) are found in the $T=\infty$ entry in Table \ref{tab:NumericalExample2}. A similar discussion for the SABR model
is found in \cite{lewis:2016} (Sec. 8.12.1).
\newpage
\section{Numerical examples for option prices} \label{sec:numerics}
Formulas are implemented in Mathematica; prices use (\ref{eq:putval}).
All required special functions are built in.
We have a double integration. Additional complications that must be handled are now discussed.
\pbold{Implementation notes: Case 1.}
For the `physical' Case 1 ($\omega \ge \xi^2/2$), we have a double integration using confluent
hypergeometric functions with complex arguments. Results are not immediate.
For example, with my choices for cutoffs and Precision parameters, option prices are found in 0.5-2 minutes on my desktop
machine.
The main complication in this case is that, prior to the Fourier integration over $z$, one must identify
the transition points (if any: see Appendix A) where the discrete and continuous parts of $H$ are discontinuous.
In Mathematica, I wrote a module {\texttt{GetAllTransitionPts[..]}}, which returns a list of
such points. As I integrate in the $z$-plane over the line $z = \i/2 + x$ for $x \in [0,x_{max}]$,
the list only includes transition points in that (finite) interval. Now, in Mathematica, when
integrating using {\texttt{NIntegrate[..]}} and typical defaults, the integration is adaptive and you
don't know in advance what points will be sampled. In the function returning the integrand,
I first determine if each sample point $x$
is near any transition point $x^*$, nearness meaning $x \in (x^*-\epsilon,x^*+\epsilon)$, where
$\epsilon > 0$ is small. If not near, I just return
the `normal' integrand. If near, I return the linearly interpolated $H$-value, using the endpoint
values at $x_1 = x^* - \epsilon$ and $x_2 = x^* + \epsilon$. This keeps $H$ continuous, while
avoiding troublesome evaluations of the discontinuous components too close to their
transition points.
\pbold{Implementation notes: Case 2.}
For the unphysical Case II ($\omega < \xi^2/2$), the good news is that there are no discontinuity points. The bad news is that the integrand
requires the infinite sum at (\ref{eq:kappadef}). I truncate that sum, and
make some use of Mathematica's {\texttt{Parallelize}}. This case
is really computationally tedious: results can take a half-hour or more.
\pbold{Examples.}
As a first example, see Tables \ref{tab:NumericalExample1} and \ref{tab:NumericalExample2}
for Cases 1 and 2 respectively. Table \ref{tab:NumericalExample3} shows
additional at-the-money implied volatilities at larger $T$, the easiest regime for spectral
solutions. There
is good convergence to the exact analytic results at $T = \infty$,
found in Sec. \ref{sec:largeT}. Table \ref{tab:NumericalExample3} results are plotted Fig. \ref{fig:XGBMLargeTIV}.
Various results were spot-checked for consistency with Monte Carlos. As a stronger check, Table \ref{tab:PDEcompares} shows some
comparisons and good agreement with option values from a PDE solver.\footnote{\label{ft:PDE}The PDE solver was developed by Y. Papadopoulos, who
provisionally extended the method in \cite{2018:paplew} to the XGBM model (private communication).}
\newpage
\begin{table}[h]
\begin{center}
\begin{tabular}{crlrlrl}
\toprule & \multicolumn{2}{c}{$(\omega=1,\rho=0)$}
& \multicolumn{2}{c}{$(\omega=1,\rho=-0.75)$} &
\multicolumn{2}{c}{$(\omega=2,\rho=-0.75)$} \\
\cmidrule(r){2-3} \cmidrule(r){4-5} \cmidrule(r){6-7}
& Option & Implied & Option & Implied & Option & Implied \\
T & Price & Vol(\%) & Price & Vol(\%) & Price & Vol(\%) \\
\midrule
0.25 & 4.126 &20.69 & 4.033 & 20.23 & 4.543 & 22.79 \\
2 & 11.57 &20.57 & 10.78 & 19.17 & 18.15 & 32.46 \\
20 & 30.97 &17.82 & 28.68 & 16.44 & 61.26 & 38.65 \\
100 & 61.94 &17.54 & 58.03 & 16.14 & 95.06 & 39.30 \\
500 & 94.98 &17.51 & 92.81 & 16.10 & 99.999 & 39.42 \\
$\infty$& 100 & 17.51 & 100 & 16.09 & 100 & 39.46 \\
\bottomrule
\end{tabular}
\end{center}
\caption{{\bf{At-the-money option values for Case 1: $\mathbf{\omega \ge \xi^2/2}$.}}
Other risk-neutral model parameters: $S_0=K=100$, $\sigma_0 = 0.20$, $\xi=1$,
$\theta=4$, $r=q=0$. Under the stationary density, $\Ex{\sigma}=(0.125,0.375)$ for $\omega=(1,2)$ respectively. The $T=\infty$ Implied Vol entries are found from the square-root of (\ref{eq:VimpInf}).}
\label{tab:NumericalExample1}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{crlrlrl}
\toprule & \multicolumn{2}{c}{$(\omega=\Smallfrac{1}{4},\theta=4)$} &
\multicolumn{2}{c}{$(\omega=\Smallfrac{1}{2},\theta=4)$} &
\multicolumn{2}{c}{$(\omega=0, \theta=0)$} \\
\cmidrule(r){2-3} \cmidrule(r){4-5} \cmidrule(r){6-7}
& Option & Implied
& Option & Implied & Option & Implied \\
T & Price & Vol(\%) & Price & Vol(\%) & Price & Vol(\%) \\
\midrule
0.25 & 3.705 & 18.58 & 3.809 & 19.11 & 3.962 & 19.87 \\
2 & 7.287 & 12.93 & 8.268 & 14.68 & 10.35 & 18.39 \\
20 & 9.577 & 5.38 & 13.56 & 7.64 & 15.49 & 8.74 \\
100 & 9.756 & 2.45 & 18.50 & 4.68 & 15.554 & 3.92 \\
500 & 9.758 & 1.10 & 25.82 & 2.95 & 15.554 & 1.75 \\
$\infty$ & 9.758 & 0 & 100 & 0 & 15.554 & 0 \\
\bottomrule
\end{tabular}
\end{center}
\caption{{\bf{At-the-money option values for Case 2: $\mathbf{\omega < \xi^2/2}$.}}
Other risk-neutral model parameters: $S_0=K=100$, $\sigma_0 = 0.20$, $\xi=1$, $\rho=-0.75$, $r=q=0$.
The middle column is technically a Case 1 edge case, but it was computed two ways
with the same result: (i) using Case 1 code with $\omega = 0.5$ and (ii) using
Case 2 code with $\omega = 0.49999$. For (ii), the sums in (\ref{eq:kappadef})
were truncated at $n=10$ terms. The $T=\infty$ price entries are computed from (\ref{eq:PutFromHinf}).}
\label{tab:NumericalExample2}
\end{table}
\newpage
\begin{table}[h]
\begin{center}
\begin{tabular}{rcccccccccccc}
\multicolumn{13}{c}{Implied Volatility (decimal)} \\
\toprule
& \multicolumn{4}{c}{Case 2} & \multicolumn{8}{c}{Case 1} \\
\cmidrule(r){2-5} \cmidrule(r){6-13}
T &$\omega=0.1$ & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9 & 1.0 &1.1 &1.2 \\
\cmidrule(r){2-5} \cmidrule(r){6-13}
$10$& 0.1672 & 0.2044 & 0.2518 & 0.3105 & 0.381 & 0.465& 0.560 & 0.667& 0.783 &0.908 &1.041 &1.180 \\
$50$ & 0.0813 & 0.1070 & 0.1470 & 0.2080 & 0.294 & 0.405& 0.537 & 0.683& 0.840 &1.004&1.172 &1.344 \\
$250$& 0.0364 &0.0481 & 0.0678 & 0.1061 & 0.184 & 0.311& 0.469 & 0.642 &0.822 &1.007&1.192&1.376\\
$\infty$ &0 &0 &0 &0 &0 &0.2 &0.4 &0.6 &0.8 &1.0 &1.196 &1.386 \\
\bottomrule
\end{tabular}
\end{center}
\caption{{\bf{Large $T$ behavior of the Implied Volatility.}}
Other risk-neutral model parameters: $S_0=K=100$, $\sigma_0 = 0.20$,
$\theta=1/\sqrt{12}$, $\xi=1$, $\rho=0$, $r=q=0$. The $T=\infty$ entries
are computed from the exact relations in Sec. \ref{sec:largeT}. The
data are plotted in Fig. \ref{fig:XGBMLargeTIV}.}
\label{tab:NumericalExample3}
\end{table}
\newpage
\begin{table}[h]
\begin{center}
\begin{tabular}{crlrl}
\multicolumn{5}{c}{Implied Volatility at Various Strikes} \\
\toprule & \multicolumn{2}{c}{Case A}
& \multicolumn{2}{c}{Case B} \\
\cmidrule(r){2-3} \cmidrule(r){4-5}
Strike & PDE & Exact & PDE & Exact \\
\cmidrule(r){1-1} \cmidrule(r){2-3} \cmidrule(r){4-5}
3400 & 0.6384 & 0.6384 & & \\
3600 & 0.5854 & 0.5854 & 0.2813 & 0.2812 \\
3800 & 0.5322 & 0.5322 & 0.2765 & 0.2765 \\
4000 & 0.4787 & 0.4787 & 0.2719 & 0.2719 \\
4200 & 0.4252 & 0.4252 & 0.2675 & 0.2676 \\
4400 & 0.3742 & 0.3742 & 0.2634 & 0.2635 \\
4600 & 0.3344 & 0.3344 & 0.2594 & 0.2595 \\
4800 & 0.3196 & 0.3196 & 0.2557 & 0.2558 \\
5000 & 0.3263 & 0.3263 & 0.2522 & 0.2524 \\
5200 & 0.3420 & 0.3420 & 0.2490 & 0.2492 \\
5400 & 0.3604 & 0.3604 & 0.2462 & 0.2463 \\
5600 & 0.3792 & 0.3792 & 0.2437 & 0.2437 \\
\bottomrule
\end{tabular}
\end{center}
\caption{{\bf{Numerical check: PDE vs. (quasi) exact solutions.}} Table entries show
the implied volatility (decimal) for two cases. For Case A: $T=\frac{14}{365}, \,\, r = 0.0357$. For Case B: $T=\frac{714}{365}, \,\,
r = 0.0401$.
Common parameters: $S_0=4468.17$, $q=0$, $\sigma_0 = 0.365113$,
$\omega=24.8424$, $\theta=63.2858$, $\rho=-0.517545$, $\xi=5.33384$. As $2 \omega/\xi^2 > 1$,
this is a Case 1 comparison. Parameters and PDE values come from an XGBM model calibration against a (July 5, 2002) DAX index option chain using software developed by Y. Papadopoulos (footnote \ref{ft:PDE}). Exact values are implied vols using prices computed from (\ref{eq:putval}).}
\label{tab:PDEcompares}
\end{table}
\section{Other applications: $P$-model parameter estimation}
Maximum likelihood estimation (MLE) is the preferred approach -- feasible using the transition density at (\ref{eq:solnfinal}) with a volatility proxy for the $y$'s.
Key are dimensionless ratios such as
$\sigma_i^2 (\theta/\xi)^2 (T_i-T_{i-1})$ and $\sigma_i (\omega \theta/\xi^2) (T_i-T_{i-1})$,
where $\sigma_i = \sigma(T_i)$.
If the data observations are such that these ratios
are not particularly small, then a fast and efficient (say C/C++) implementation
is likely needed. However, when the $\mbox{ratios} \ll 1$,
as may be the case with daily or weekly observations, then small-time asymptotics
for the transition density should be effective and the exact formulas can simply serve
as checks. Results along these lines will be reported in another publication.
\newpage
|
\section{Introduction}
Hypothesis testing is ubiquitous in modern statistical applications, which
permeates many different fields such as biology, medicine,
phycology, economics, and engineering etc.
As a critical component of the hypothesis testing procedure (Lehmann and Romano, 2005),
$p$-value is defined as the probability of observing the
random data as or more extreme than the observed
given the null hypothesis being true.
In general, the statistical significance level
or the type I error rate is set at 5\%, so that a $p$-value below
5\% is considered significant leading to rejection of the null hypothesis,
and that above 5\% insignificant resulting in failure to reject the null.
Although $p$-value is the most commonly used summary measure for evidence
or strength in the data regarding the null hypothesis, it
has been the center of controversies and debates for decades.
To clarify ambiguities surrounding $p$-value,
the American Statistical Association (2016) gave statements on $p$-value
and, in particular, the second point states that ``$P$-values do not measure
the probability that the studied
hypothesis is true, or the probability that the data
were produced by random chance alone.'' It is often argued
that $p$-value only gives information on how incompatible
the data are with the null hypothesis, but it does not provide
any information on how likely the data would occur under
the alternative hypothesis.
Extensive investigations have been conducted on the inadequacy of the $p$-value.
Rosenthal and Rubin (1983) studied how $p$-value can be adjusted to
allow for greater power when an order of importance exists on the hypothesis tests.
Royall (1986) investigated the effect of sample size on $p$-value. Schervish (1996)
described computation of the $p$-value for one-sided point null hypotheses, and also
discussed the intermediate interval hypothesis.
Hung et al. (1997) studied the behavior of
$p$-value under the alternative hypothesis,
which depends on both the true value of
the tested parameter and sample size.
Rubin (1998) proposed an alternative randomization-based
$p$-value for double-blind trials with non-compliance.
Sackrowitz and Samuel-Cahn (1999) promoted more
widespread use of the expected $p$-value in practice.
Donahue (1999) suggested that the distribution of the $p$-value
under the alternative hypothesis provide more information for rejection of implausible
alternative hypotheses. As there is a widespread notion that medical research
is interpreted mainly based on $p$-value,
Ioannidis (2005) claimed that most of the published findings are false.
Hubbard and Lindsay (2008) showed that $p$-value
tends to exaggerate the evidence against the null hypothesis.
Simmons et al. (2011) demonstrated that $p$-value is subject to
manipulation to achieve the threshold of 0.05 and cautioned against its use.
Nuzzo (2014) gave an editorial on why $p$-value alone
cannot serve as adequate statistical evidence for inference.
Criticisms on $p$-value and null hypothesis significance
testing have become even more contentious in recent years.
If the key words ``misuse of $p$-value'' or ``ban $p$-value'' are
used in Google search, millions of queries can be found
to attack and bash $p$-value.
More seriously, several journals, e.g.,
{\em Basic and Applied Social Psychology} and {\em Political Analysis},
have made claims to ban the use of $p$-value in their publications
(Trafimow and Marks, 2015; Gill, 2018).
The controversy over $p$-value has recently been reignited,
which is more centered around the proposals to adjust,
abandon or provide alternatives to $p$-value.
Fidler et al. (2004) and Ranstam (2012) recommended use of
the confidence interval as an alternative to $p$-value,
and Cumming (2014) called for abandoning
$p$-value in favor of reporting the confidence interval.
Colquhoun (2014) investigated the issue of misinterpretation of
$p$-value as a culprit for the high false discovery rate.
Concato and Hartigan (2016) suggested that $p$-value should not be the
primary focus of attention or the sole basis for evaluation of scientific results.
McShane et al. (2017) recommended that the role of $p$-value
as a threshold for screening scientific findings should be demoted, and that
$p$-value should not take priority over other statistical measures.
In the aspect of reproducibility concerns of scientific research,
Johnson (2013) traced one major cause of nonreproducibility as
the routine use of the null hypothesis testing procedure.
Leek et al. (2017) proposed abandonment of $p$-value thresholding and
transparent reporting of false positive risk as remedies to the replicability
issue in science. Benjamin et al. (2018) recommended shifting the
significance threshold from 0.05 to 0.005, while Trafimow et al.
(2018) argued that such a shift is futile and unacceptable.
Bayesian approaches are often advocated as a solution to the
crisis resulting from abusing the $p$-value.
Goodman (1999) strongly supported use of the Bayes factor in contrast to
$p$-value as a measure of evidence for medical evidence-based
research. Rubin (1984) proposed the predictive $p$-value as the tail-area
probability of the posterior predictive distribution, and Meng (1994)
further studied its properties.
In the applications to psychology, Wagenmakers (2007)
revealed the issues with $p$-value and recommended
use of the Bayesian information criterion instead.
In an effort to support the wider use of Bayesian statistics,
Lee (2010) demonstrated that Bayesian approaches
provide a superior alternative to the frequentist methods using $p$-values.
Alongside its ban on $p$-value, the journal of
{\em Basic and Applied Social Psychology} gave endorsement of
Bayesian approaches (Trafimow and Marks, 2015).
Briggs (2017) proposed that $p$-value should be proscribed and be substituted with the Bayesian posterior probability, while Savalei and Dunn (2015) expressed skepticism on the utility of abandoning $p$-value and resorting to alternative hypothesis testing paradigms, such as the Bayesian approach, in solving the reproducibility issue.
On the other hand, extensive research has been conducted in an attempt to reconcile
or account for the differences between frequentist and Bayesian
hypothesis testing approaches (Berger, 2003; and Bayarri and Berger, 2004).
For hypothesis testing, Berger and Sellke (1987),
Berger and Delampady (1987), and Casella and Berger (1987)
investigated the relationships between $p$-value and
the Bayesian measure of evidence against
the null hypothesis. In particular, they provided an in-depth study of one-sided hypothesis
testing and point null cases, and also discussed the posterior probability of the null hypothesis
with respect to various prior distributions including the mixture prior distribution with a
point mass at the null and the other more broad distribution over the alternative (Lindley, 1957). Sellke, Bayarri, and Berger (2001) proposed to calibrate $p$-value
for testing precise null hypotheses.
Although $p$-value is often regarded as an inadequate and insufficient representation
of statistical evidence, it did not stall the scientific advancement
in the past years.
Jager and Leek (2014) surveyed high-profile medical journals
and estimated the rate of false discoveries in the medical
literature using reported $p$-values as the data, which
led to a conclusion that the medical literature remains a reliable
record of scientific progress.
Murtaugh (2014) defended the use of $p$-value based on the ground
that it is closely linked to the confidence interval and to the difference in
Akaike's information criterion.
Despite the fact that Bayesian alternatives are often recommended
as superior solutions to the various notorious drawbacks of $p$-value,
in many common cases, $p$-value in fact has a simple and clear Bayesian interpretation.
We present the relationship between the frequentist $p$-value and
Bayesian posterior probability in several commonly encountered settings
in clinical trials,
and show that in both one-sided and two-sided hypothesis tests,
asymptotic equivalence, sometimes exact equivalence, can be established.
Although in terms of definition, $p$-value is not the probability
that the null hypothesis is true, contrary to the conventional notion,
it does have a close correspondence to the Bayesian posterior probability of the null hypothesis being true. Based on the theoretical results of Dudley and Haughton (2002), we present several
cases where $p$-value and the posterior probability of the null
are equivalent for one-sided tests. Further, we extend such equivalence
results to two-sided hypothesis testing problems,
where most of the controversies and discrepancies lie.
In particular, we introduce
the notion of two-sided posterior probability which matches the $p$-value
from a two-sided hypothesis test.
After all, we conclude that $p$-value is not all that bad.
The rest of the paper is organized as follows. In Section 2,
we present a motivating example that shows the similarity in
operating characteristics of a frequentist hypothesis test and a
Bayesian counterpart using the posterior probability. In Section 3,
we show that $p$-value and the posterior probability have an
equivalence relationship for the case of binary outcomes. In Section 4, we present such equivalence properties for univariate normal data with known and unknown variances respectively,
and in Section 5, we develop similar results for hypothesis tests involving
multivariate data.
Finally, Section 6 concludes with some remarks.
\section{Motivating Example}
The use of binary endpoint is common in clinical trial design. Frequentist design typically utilizes an exact binomial test or $Z$-test based on normal approximation, and Bayesian design often bases the decision on the posterior probabilities.
As a motivating example, we consider a two-arm clinical trial comparing the response rate of an experimental drug $p_E$ versus that of the standard drug $p_S$.
We are interested in testing a one-sided hypothesis,
\begin{equation}\label{binaryh}
H_0\mbox{:} \ p_E \le p_S \quad {\rm versus} \quad H_1\mbox{:} \ p_E > p_S.
\end{equation}
When there is sufficient evidence to support $H_1$, we would reject $H_0$ and claim
that the experimental treatment is superior.
Under the frequentist approach, we
construct a $Z$-test statistic,
\begin{equation}\label{binaryztest}
Z = \frac{\hat p_E-\hat p_S}{[{\{ {{\hat p_E}(1-{\hat p_E})} + {{\hat p_S}(1-{\hat p_S})} \}/n}]^{1/2}},
\end{equation}
where $n$ is the sample size per arm, $\hat p_E=y_E/n$ and $\hat p_S=y_S/n$ are the sample proportions, $y_E$ and $y_S$ are the numbers of responders in the respective arms.
We reject the null hypothesis if $Z > z_{\alpha}$, where $z_{\alpha}$ is the $100(1-\alpha)$th
percentile of the standard normal distribution.
Under the Bayesian framework,
we assume beta prior distributions for $p_E$ and $p_S$, i.e.,
$p_E\sim{\rm Beta}(a_E,b_E)$ and $p_S \sim {\rm Beta}(a_S,b_S)$.
The binomial likelihood function for group $g$
can be formulated as
$${P}({y_g}|{p_g}) = {\displaystyle {n \choose y_g}}
{p_g^{y_g}}{(1 - p_g)^{n - y_g}},
\quad g=E, S.$$
The posterior distribution of $p_g$
is given by
\begin{eqnarray}
{p_g}|{y_g}&\sim& {\rm Beta}(a_g+y_g,b_g+n-y_g),\nonumber
\end{eqnarray}
for which the density function is denoted by $f({p_g}|{y_g})$.
Let $\eta$ be a prespecified cutoff probability boundary. We declare
treatment superiority if the posterior probability of $p_E$ greater than $p_S$ exceeds threshold $\eta$.
Based on the posterior probability, we can construct a Bayesian decision rule
so that the experimental treatment is declared as superior if
\begin{equation}\label{binarybtest}
\Pr(H_1|y_E,y_S) = \Pr(p_E>p_S|y_E,y_S)> \eta,
\end{equation}
where
\[
\Pr({p_E} > {p_S}|{y_E},{y_S})
=\int_0^1{\int_{{p_S}}^1{f({p_E}|{y_E})f({p_S}|y_S)}}d{p_E}d{p_S}.
\]
Otherwise, we fail to declare treatment superiority, i.e., fail to reject the null hypothesis.
To maintain the frequentist type I error rate at $\alpha$, we need to set $\eta = 1-\alpha$. The exact probabilities of committing type I and type II errors under the frequentist design are respectively given by
$$
\alpha={\sum\limits_{{y_E}=0}^{{n}}{\sum\limits_{{y_S}=0}^{{n}}}}{P}({y_E}|{p_E=p_S})
{P}({y_S}|{p_S})I(Z>z_\alpha),$$
and
$$\beta={\sum\limits_{{y_E}=0}^{{n}}{\sum\limits_{{y_S}=0}^{{n}}}}{P}({y_E}|{p_E=p_S+\delta})
{{P}}({y_S}|{p_S})I(Z<z_\alpha),
$$
where $\delta$ is the desired treatment difference and $I(\cdot)$ is the indicator function.
The exact error rates under the Bayesian test can
be derived similarly by replacing
$Z>z_\alpha$ with $\Pr(p_E>p_S|y_E,y_S) > 1-\alpha$ inside the indicator function.
As a numerical study, we consider a two-arm randomized trial with
a type I error rate of 10\% and 5\% and target power of 80\% and 90\% when $(p_S,p_E)=(0.2,0.3)$ and $(p_S,p_E)=(0.2,0.35)$, respectively.
Under equal randomization, to achieve the desired power, the required sample size per arm is
$$
n=\frac{(z_{\alpha}+z_{\beta})^2}{\delta^2}\{p_E(1- p_E)+p_S(1- p_S)\},
$$
where we take $\delta=0.1$ and 0.15.
Under the Bayesian design, we assume non-informative prior
distributions, $p_S \sim {\rm Beta}(0.2,0.8)$ and $p_E\sim{\rm Beta}(0.2,0.8)$. For comparison, we compute the type I error rate and power for both the
Bayesian test with $\eta=1-\alpha$ and the frequentist
$Z$-test with a critical value $z_{1-\alpha}$. As shown in
Figure \ref{comparet1}, both designs
produce similar operating characteristics: the type I error rate can
be maintained at the nominal level, and the power attains the target level of 80\% or 90\% at
the specified values of $(p_S,p_E)$. It is worth noting that
because the endpoints are binary and the trial outcomes are discrete,
exact calibration of the empirical type I error rate to the nominal level is not possible, particularly when the sample size is small.
When we adopt a larger sample size by setting the type I error rate to be 5\% and the target power to be 90\%, the empirical type I error rate is closer to the nominal level as shown in the blue lines.
\section{Hypothesis Test for Binary Data}
\subsection{Two-Sample Hypothesis Test}
We first study the relationship between $p$-value and the posterior probability in a two-arm randomized clinical trial with dichotomous outcomes.
We consider the one-sided hypothesis test in (\ref{binaryh}), and under the frequentist $Z$-test for two proportions given by (\ref{binaryztest}),
the $p$-value is
$$
p\mbox{-value}_1 = 1-\Phi(Z),
$$
where $\Phi(\cdot)$ denotes the cumulative distribution function (CDF) of the standard normal distribution. At the significance level of $\alpha$, we reject the null hypothesis if $p$-value is smaller than $\alpha$.
In the Bayesian paradigm, we base our decision on the posterior probability, as given in (\ref{binarybtest}). We reject the null hypothesis if the posterior probability of $p_E\le p_S$ is smaller than $\alpha$,
$${\rm PoP}_1= \Pr(p_E \le p_S|y_E,y_S)<\alpha.$$
As a numerical study, we set $n=20$, 50, 100 and 500, and randomly draw integers between 0 and $n$ to be the values for $y_E$ and $y_S$, and for each replication we compute the posterior probability of the null hypothesis $\Pr(H_0|y_E,y_S)$ and the $p$-value. As shown in Figure \ref{os}, all the paired values lie very close to the straight line of $y=x$, indicating the equivalence between the $p$-value and posterior probability of the null.
Figure \ref{three} shows the differences between $p$-values and posterior probabilities $\Pr({p_E} \le {p_S}|{y_E},{y_S})$ under sample sizes of 20, 50, 100 and 500, respectively. As sample size increases, the differences diminish toward 0, corroborating the asymptotic equivalence between $p$-value and the posterior probability.
For two-sided hypothesis tests, we are interested in
examining whether there is any difference in the treatment effect between the experimental drug and the standard drug,
\begin{equation*}\label{Hypothesis}
H_0\mbox{:} \ p_E = p_S \quad {\rm versus} \quad H_1\mbox{:} \ p_E \neq p_S.
\end{equation*}
The $p$-value under the two-sided hypothesis test is
\begin{eqnarray*}
p \mbox{-value}_2 &=& 2-2\Phi(|Z|) = 2[1-{\rm max}\{\Phi(Z),\Phi(-Z) \}].
\end{eqnarray*}
It is worth emphasizing that under the frequentist paradigm, the two-sided test can be viewed as a combination of two one-sided tests along the opposite directions. Therefore, to construct an equivalent counterpart under the Bayesian paradigm, we may regard the problem as two opposite one-sided Bayesian test and compute the posterior probabilities of the two opposite hypotheses; this approach to Bayesian hypothesis testing is different from the one commonly adopted in the literature, where a prior probability mass is imposed on the point null, e.g., see Berger and Sellke (1987), Berger and Delampady (1987), and Berger (2003).
If we define the two-sided posterior probability (${\rm PoP}_2$) as
$${\rm PoP_2} = 2[1-{\rm max}\{\Pr({p_E} > {p_S}|{y_E},{y_S}), \Pr({p_E} < {p_S}|{y_E},{y_S})\}],$$
then its relationship with $p$-value is similar to that of one-sided hypothesis testing as shown in Figure \ref{ts}.
The equivalence of the $p$-value and the posterior probability in the case of binary outcomes can be established by applying the Bayesian central limit theorem. Under large sample size, the posterior distribution of $p_E$ and $p_S$ can be approximated as
$$
p_g|y_g \sim {\rm N}({\hat p_g},{\hat p_g}(1-{\hat p_g}) / n), \quad g=E,S.
$$
As $y_E$ and $y_S$ are independent, the posterior distribution of $p_E-p_S$ can be derived as
$$p_E-p_S | y_E,y_S \sim {\rm N}({\hat p_E} - {\hat p_S} , \{ {{\hat p_E}(1-{\hat p_E})} + {{\hat p_S}(1-{\hat p_S})} \}/n ).$$
Therefore, the posterior probability of $p_E \le p_S$ is
$$
{\rm PoP}_1= \Pr(p_E \le p_S|y_E,y_S) \approx \Phi\bigg(-\frac{{\hat p_E} - {\hat p_S}}{[\{ {{\hat p_E}(1-{\hat p_E})} + {{\hat p_S}(1-{\hat p_S})} \}/n]^{1/2}}\bigg) = \Phi(-Z),
$$
which is equivalent to $p \mbox{-value}_1 =1-\Phi(Z)= \Phi(-Z)$.
The equivalence relationship for a two-sided test can be derived along similar lines.
More generally, Dudley and Haughton (2002) proved that under mild regularity conditions, the posterior probability of a half space converges to the standard normal CDF transformation of the likelihood ratio test statistic. In a one-sided hypothesis test, the posterior probability of the half space is $\Pr(H_1|D) = 1-{\rm PoP}_1$, whereas the standard normal CDF transformation of the likelihood ratio test statistic equals to one minus $p \mbox{-value}_1$, an therefore ${\rm PoP}_1$ and $p \mbox{-value}_1$ are asymptotically equivalent.
\subsection{One-Sample Hypothesis Test}
In a single-arm clinical trial with dichotomous outcomes, we are interested in
examining whether the response rate of the experimental drug $p_E$ exceeds a prespecified threshold $p_0$, by formulating a one-sided hypothesis test,
\begin{equation*}\label{Hypothesis}
H_0\mbox{:} \ p_E \le p_0 \quad {\rm versus} \quad H_1\mbox{:} \ p_E > p_0.
\end{equation*}
In the frequentist paradigm, the $p$-value can be computed based on the exact binomial test. In the Bayesian paradigm, we assume a beta prior distribution for $p_E$, e.g., $p_E\sim{\rm Beta}(a_E,b_E)$.
The posterior distribution of $p_E$
is given by $
{p_E}|{y_E}\sim {\rm Beta}(a_E+y_E,b_E+n-y_E)$,
for which the density function
is denoted by $f({p_E}|{y_E})$.
Based on the posterior probability, we can construct a Bayesian decision rule
so that the experimental treatment is declared as promising if
$$\Pr(H_1|y_E) = \Pr(p_E>p_0|y_E)> \eta,$$
where
\[
\Pr({p_E} > {p_0}|{y_E})
=\int_{p_0}^1 f(p_E|y_E)d{p_E}.
\]
Otherwise, we fail to declare treatment efficacy.
As a result, the one-sided posterior probability is defined as
$${\rm PoP}_1=\Pr(H_0|y_E) = \Pr(p_E \le p_0|y_E).$$
For two-sided hypothesis tests, we are interested in
examining whether the response rate of the experimental drug is different from $p_0$,
\begin{equation*}\label{Hypothesis}
H_0\mbox{:} \ p_E = p_0 \quad {\rm versus} \quad H_1\mbox{:} \ p_E \neq p_0.
\end{equation*}
The $p$-value can be computed based on the exact binomial test.
If we define the two-sided posterior probability,
$${\rm PoP_2} = 2[1-{\rm max}\{\Pr({p_E} > {p_0}|{y_E}), \Pr({p_E} < {p_0}|{y_E})\}],$$
then its relationship with $p$-value is similar to that of one-sided hypothesis testing as shown in Figure \ref{ts}.
In a numerical study, we set $n=20$, 50, 100 and 500, $p_0=0.2$, and randomly draw integers between 0 and $n$ to be the values of $y_E$. We assume a noninformative prior for $p_E$, i.e., $a_E = 1$ and $b_E = 1$. Figure \ref{os} shows the relationship between the posterior probability of the null hypothesis $\Pr(H_0|y_E)$ and the $p$-value, which clearly indicates that all the points lie very close to the straight line of $y=x$.
\section{Hypothesis Test for Normal Data}
\subsection{Hypothesis Test with Known Variance}
In a two-arm randomized clinical trial with normal endpoints, we are interested in comparing the means of the outcomes between the experimental and standard arms.
Let $n$ denote the sample size for each arm, and let $D = \{(y_{E1},y_{S1}),\ldots,(y_{En},y_{Sn}) \}$ denote the paired data under the experimental and standard treatments. Assume $y_{Ei} \sim {\rm N}(\mu_E, \sigma^2)$ and $y_{Si} \sim {\rm N}(\mu_S, \sigma^2)$ with unknown means $\mu_E$ and $\mu_S$ but a known variance $\sigma^2 = 1$.
Let $\bar y_E = \sum_{i=1}^{n} y_{Ei}/n$ and $\bar y_S = \sum_{i=1}^{n} y_{Si}/n$ denote the sample means, and let $\theta = \mu_E-\mu_S$ and $\hat \theta = \bar y_E-\bar y_S$ denote the true and the observed treatment difference, respectively.
Considering the one-sided hypothesis test,
\begin{equation*}\label{Hypothesis}
H_0\mbox{:} \ \theta \le 0 \quad {\rm versus} \quad H_1\mbox{:} \ \theta > 0,
\end{equation*}
the frequentist $Z$-test statistic is formulated as
${\hat \theta}/{\sqrt{2/n}}$,
which follows the standard normal distribution under the null hypothesis. Therefore, the $p$-value under the one-sided hypothesis test is given by
\begin{align*}
p \mbox{-value}_1 &= \Pr(Z\ge \hat \theta \sqrt{n/2}|H_0)
= 1 - \Phi( \hat \theta \sqrt{n/2}),
\end{align*}
where $Z$ denotes the standard normal random variable.
In the Bayesian paradigm, if we assume an improper flat prior distribution, $p(\theta) \propto 1$, the posterior distribution of $\theta$ is
$$\theta|D \sim {\rm N}(\hat \theta, 2/n).$$
Therefore, the posterior probability of $\theta$ smaller or equal to 0 is
$$
{\rm PoP}_1 = \Pr(\theta \le 0 | D) = 1- \Phi( \hat \theta \sqrt{n/2}).
$$
Under such an improper prior distribution of $\theta$, we can establish an exact equivalence relationship between $p$-value and $\Pr(\theta \le 0 | D)$.
Under the two-sided hypothesis test, $H_0:\theta=0$ versus $H_1:\theta \neq 0$, the $p$-value is given by
\begin{align*}
p \mbox{-value}_2 &= 2[1- {\rm max}\{\Pr(Z\ge z|H_0),\Pr(Z \le z|H_0)\} ]\\
&= 2 - 2 {\rm max} \{\Phi( \hat \theta \sqrt{n/2}),\Phi(- \hat \theta \sqrt{n/2})\}.
\end{align*}
Correspondingly, the two-sided posterior probability is defined as
\begin{align*}
{\rm PoP}_2 &= 2[1 - {\rm max}\{\Pr(\theta < 0 | D), \Pr(\theta > 0 | D)\}]\\
&= 2 - 2 {\rm max} \{\Phi( \hat \theta \sqrt{n/2}),\Phi(- \hat \theta \sqrt{n/2})\},
\end{align*}
which is exactly the same as the (two-sided) $p$-value.
\subsection{Hypothesis Test with Unknown Variance}
In a more general setting, we consider the case where $\mu_E$, $\mu_S$ and $\sigma$ are all unknown parameters. We define $x_i = y_{Ei}-y_{Si}$, which follows the normal distribution ${\rm N}(\theta, 2\sigma^2)$. For notational simplicity, let $\nu = 2\sigma^2$ and we are interested in modeling the joint posterior distribution of $\theta$ and $\nu$.
In the frequentist paradigm, Student's $t$-test statistic is
$$
T = \frac{\hat \theta }{\sqrt{\sum_{i=1}^{n}(x_i-\hat \theta)^2 / \{(n-1)n\}}}.
$$
Therefore, the $p$-value under the one-sided hypothesis test is
\begin{align*}
p \mbox{-value}_1 &= 1 - F_{t_{n-1}}(T),
\end{align*}
where $F_{t_{n-1}}(\cdot)$ denotes the CDF of Student's $t$ distribution with $n-1$ degrees of freedom.
In the Bayesian paradigm, if we assume Jeffreys' prior for $\theta$ and $\nu$,
$
p(\theta,\nu) \propto \nu^{-{3}/{2}}
$,
the corresponding posterior distribution is
$$
p(\theta,\nu|D) \propto \nu^{-(n+3)/2}\exp\bigg\{ -\frac{\sum_{i=1}^{n}(x_i-\hat \theta)^2 + n(\hat \theta - \theta)^2 }{2\nu} \bigg\},
$$
which matches the normal-inverse-chi-square distribution,
$$(\theta,\nu)|D \sim {\rm N}\mbox{--Inv} \ {\chi^{2}} \bigg(\hat \theta, n, n, \sum_{i=1}^{n}(x_i-\hat \theta)^2/n \bigg).$$
Based on the posterior distribution, the one-sided posterior probability of the null hypothesis is ${\rm PoP}_1 = \Pr({\theta} \le {0}|{D})$.
As an alternative to Jeffreys' prior distribution, we also consider a normal-inverse-gamma prior distribution for $\theta$ and $\nu$,
$
(\theta,\nu) \sim {\rm N} \mbox{--IG}(\theta_0,\nu_0,\alpha,\beta)
$,
which belongs to the conjugate family of prior distributions for the normal likelihood function.
As a result, the corresponding posterior distribution is also a normal-inverse-gamma prior distribution,
$$
(\theta,\nu)|D \sim {\rm N}\mbox{--IG}\bigg(\frac{\theta_0\nu_0 + n\hat\theta}{\nu_0 + n},\nu_0+n,\alpha+\frac{n}{2},\beta+ \frac{1}{2}\sum_{i=1}^{n}(x_i-\hat \theta)^2 + \frac{n\nu_0}{\nu_0+n}\frac{(\hat \theta - \theta_0)^2}{2} \bigg).
$$
For a two-sided hypothesis test, the $p$-value is
\begin{align*}
p \mbox{-value}_2 &= 2-2F_{t_{n-1}}(|T|)\\
&= 2[1-{\rm max}\{ F_{t_{n-1}}(T),F_{t_{n-1}}(-T) \}].
\end{align*}
Similarly, we define the two-sided posterior probability as
$${\rm PoP}_2 = 2[1-{\rm max}\{\Pr({\theta} > {0}|{D}), \Pr({\theta} < {0}|{D})\}].$$
In a numerical study, we simulate a large number of trials, and for each replication we compute the posterior probability $\Pr(\theta \le 0|D)$ and $p$-value.
To ensure that the simulated $p$-values can cover the entire range of $(0,1)$, we generate values of $\theta$ from ${\rm N}(0,0.05)$ and $\nu$ from ${\rm N}(1,0.05)$ truncated at zero.
To construct a vague normal-inverse-gamma prior distribution, we take $\theta_0 = 0$, $\nu_0 = 100$, and $\alpha=\beta=0.01$. Under Jeffreys' prior and the vague normal-inverse-gamma prior distributions, the equivalence relationships between $p$-values and the posterior probabilities $\Pr(\theta \le 0|D)$ are shown in Figure \ref{normal_jeff}, with sample size of 20, 50 and 100, respectively.
In addition, we generate values of $x_i$ from a Gamma$(2,0.5)$ distribution, a Beta$(0.5,0.5)$ distribution, as well as a mixture of normal distributions of N$(-1,1)$ and N$(1,1)$ with equal weights. To ensure that the simulated $p$-values can cover the entire range of $(0,1)$, the simulated values of $x_i$ are further deducted by the mean value of the corresponding distribution plus a uniform random variable. Under Jeffreys' prior, the equivalence relationships between $p$-values and the posterior probabilities $\Pr(\theta \le 0|D)$ are shown in Figure \ref{normal_gamma}.
To study the effect of informative prior and sample size on the relationship between $p$-values and the posterior probabilities, we construct an informative prior distribution on $\theta$ by setting $\theta_0 = \theta + 0.01$, $\nu_0 = 0.01$, and $\alpha=\beta=0.01$. Under such an informative prior distribution, the relationships between $p$-values and the posterior probabilities $\Pr(\theta \le 0|D)$ under increasing sample sizes are shown in Figure \ref{informative}. As sample size increases, the equivalence relationship is gradually established.
Moreover, we consider the case where the sample size is fixed but the prior variance increases, i.e., we take $\theta_0 = \theta + 0.01$ and let $\nu_0$ change from 0.001 to 1. As shown in Figure \ref{informative}, as the prior distribution becomes less informative, the equivalence relationship becomes more evident.
\section{Hypothesis Test for Multivariate Normal Data}
In
hypothesis testing on the mean vector of a multivariate normal random variable, we consider $\mbox{\bf X} \sim {\rm N}_p(\boldsymbol{\mu},\boldsymbol{\Sigma})$, where $p$ is the dimension of the multivariate normal distribution. For the ease of exposition, the covariance matrix $\boldsymbol{\Sigma}$ is assumed to be known.
Let $D = \{\mbox{\bf X}_1,\ldots,\mbox{\bf X}_n\}$ denote the observed multivariate vectors, let $\bar \mbox{\bf X} = \sum_{i=1}^{n}\mbox{\bf X}_i/n$ denote the sample mean vector, and thus $\bar \mbox{\bf X} \sim {\rm N}_p(\boldsymbol{\mu},\boldsymbol{\Sigma}/n)$.
Consider the one-sided hypothesis test,
\begin{equation*}\label{Hypothesis}
H_0\mbox{:} \ \mbox{\bf c}_k^{\top}\boldsymbol{\mu} \le 0 \ {\rm for} \ {\rm some} \ k = 1,\ldots,K \quad {\rm versus} \quad H_1\mbox{:} \ \mbox{\bf c}_k^{\top}\boldsymbol{\mu} > 0 \ {\rm for} \ {\rm all} \ k = 1,\ldots,K,
\end{equation*}
where $\mbox{\bf c}_1,\ldots,\mbox{\bf c}_K$ are $K$ prespecified $p$-dimensional vectors.
The likelihood ratio test statistics (Sasabuchi, 1980) are
\begin{equation}\label{sasa}
Z_k = \frac{\mbox{\bf c}_k^{\top} {\bar {\mbox{\bf X}}}}{\sqrt{\mbox{\bf c}_k^{\top} { \boldsymbol{\Sigma}}\mbox{\bf c}_k/n}}, \ \ k = 1,\ldots,K,
\end{equation}
and the corresponding $p$-values are
$$
p \mbox{-value}(k)_1 = 1 - \Phi(Z_k).
$$
The null hypothesis is rejected if all of the $K$ $p$-values are smaller than $\alpha$.
In the Bayesian paradigm, we assume a conjugate multivariate normal prior distribution for $\boldsymbol{\mu}$, $
\boldsymbol{\mu} \sim {\rm N}_p(\boldsymbol{\mu}_0,\boldsymbol{\Sigma}_0)$.
The corresponding posterior distribution is $
\boldsymbol{\mu}|D \sim {\rm N}_p(\boldsymbol{\mu}_n,\boldsymbol{\Sigma}_n)$,
where
\begin{align*}
\boldsymbol{\mu}_n & = \boldsymbol{\Sigma}_0\left(\boldsymbol{\Sigma}_0+ \frac{\boldsymbol{\Sigma}}{n}\right)^{-1}{\bar \mbox{\bf X}} + \frac{1}{n} \boldsymbol{\Sigma} \left(\boldsymbol{\Sigma}_0+ \frac{\boldsymbol{\Sigma}}{n}\right)^{-1}\boldsymbol{\mu}_0, \\
\boldsymbol{\Sigma}_n & = \frac{1}{n}\boldsymbol{\Sigma} \left(\boldsymbol{\Sigma}_0+ \frac{\boldsymbol{\Sigma}}{n}\right)^{-1}\boldsymbol{\Sigma}.
\end{align*}
The one-sided posterior probability corresponding to $\mbox{\bf c}_k$ is
$$
{\rm PoP}(k)_1 = \Pr(\mbox{\bf c}_k^{\top}\boldsymbol{\mu} \le 0 | D).
$$
For two-sided hypothesis testing (Liu and Berger, 1995), we are interested in
\begin{eqnarray*}\label{Hypothesis}
&&H_0\mbox{:} \ \mbox{\bf c}_k^{\top}\boldsymbol{\mu} \le 0 \ {\rm for} \ {\rm some} \ k = 1,\ldots,K, {\rm and} \\
&& \ \ \ \ \ \ \mbox{\bf c}_k^{\top}\boldsymbol{\mu} \ge 0 \ {\rm for} \ {\rm some} \ k = 1,\ldots,K \\
&&{\rm versus} \\
&&H_1\mbox{:} \ \mbox{\bf c}_k^{\top}\boldsymbol{\mu} > 0 \ {\rm for} \ {\rm all} \ k = 1,\ldots,K, \ {\rm or} \\
&& \ \ \ \ \ \ \mbox{\bf c}_k^{\top}\boldsymbol{\mu} < 0 \ {\rm for} \ {\rm all} \ k = 1,\ldots,K.
\end{eqnarray*}
Based on (\ref{sasa}), the $p$-values are given by
$$
p \mbox{-value}(k)_2 = 2 - 2\Phi(|Z_k|) = 2[1-{\rm max}\{ \Phi(Z_k),\Phi(-Z_k) \}].
$$
The null hypothesis is rejected if all of the $K$ $p$-values are smaller than $\alpha$.
Similar to the univariate case, we define the two-sided posterior probability,
$${\rm PoP}(k)_2 = 2[1-{\rm max}\{\Pr(\mbox{\bf c}_k^{\top}\boldsymbol{\mu} > 0 | D), \Pr(\mbox{\bf c}_k^{\top}\boldsymbol{\mu} < 0 | D)\}].$$
In a numerical study, we compute the posterior probabilities of $\mbox{\bf c}_k^{\top}\boldsymbol{\mu} \le 0$ for $k = 1,\ldots,K$, and compare them with the corresponding $p$-values. We take $K=2$ and $\mbox{\bf c}_k$ to be a unit vector with 1 on the $k$th element and 0 otherwise, and assume a vague normal prior distribution for $\boldsymbol{\mu}$, i.e., $\boldsymbol{\mu}_0 = {\bf 0}$ and $\boldsymbol{\Sigma}_0 = 1000 \mbox{\bf I}_p$, where $\mbox{\bf I}_p$ is a $p$-dimensional identity matrix. The relationship between the posterior probabilities and $p$-values is shown in Figure \ref{multi}, which is very similar to that in the univariate setting, which again demonstrates their equivalence.
\section{Discussion}
Berger and Sellke (1987) studied the point null for two-sided hypothesis tests, and noted discrepancies between the frequentist test and the Bayesian test based on the posterior probability.
The major difference between their work and the equivalence relationship between the posterior probability and $p$-value established here lies in the assumption of the prior distribution. Berger and Sellke (1987) assumed a point mass prior distribution at the point null hypothesis, which violates the regularity condition of continuity in Dudley and Haughton (2002), leading to the discrepancy between the posterior probability and $p$-value.
The equivalence relationship between the posterior probability and $p$-value for one-sided tests can be established from the theoretical results of Dudley and Haughton (2002), where the posterior probability of a half space is proven to converge to the standard normal CDF transformation of the likelihood ratio test statistic. A future direction of research is on more complex composite hypotheses tests involving multivariate normal outcomes.
Berger (1989) and Liu and Berger (1995) constructed a uniformly more powerful test than the likelihood ratio test for multivariate one-sided tests involving linear inequalities. Follman (1996) proposed a simple alternative to the likelihood ratio test. It would be of interest to study the relationship of these tests with the Bayesian counterparts based on posterior probabilities.
%
\newpage
\section*{References}
\begin{description}
\item
Bayarri, M. J. and Berger, J. O. (2004). The interplay of Bayesian and frequentist analysis. {\em Statistical Science} {\bf 19}, 58--80.
\item
Berger, J. O. (2003). Could Fisher, Jeffreys and Neyman have agreed on testing? (with discussion) {\em Statistical Science
} {\bf 18}, 1--32.
\item
Berger, J. O. and Delampady M. (1987). Testing precise hypotheses. {\em Statistical Science} {\bf 2}, 317--335.
\item
Berger, J. O. and Sellke, T. (1987). Testing a point null hypothesis: the irreconcilability of P values and evidence. {\em Journal of the American Statistical Association} {\bf 82}, 112--122.
\item
Benjamin, D. J., Berger, J. O., Johannesson, M., Nosek, B. A., Wagenmakers, E., et al. (2017). Redefine statistical significance. {\em Nature Human Behaviour} {\bf 2}, 6--10.
\item
Briggs, W. M. (2017). The substitute for p-values. {\em Journal of the American Statistical Association} {\bf 112}, 897--898.
\item
Berger, R. L. (1989). Uniformly more powerful tests for hypotheses concerning linear inequalities and normal means. {\em Journal of the American Statistical Association} {\bf 84}, 192--199.
\item
Casella, G. and Berger, R. L. (1987). Reconciling Bayesian and frequentist evidence in the one-sided testing problem. (with discussion) {\em Journal of the American Statistical Association} {\bf 82}, 106--111.
\item
Concato, J. and Hartigan, J. A. (2016). P values: from suggestion to superstition. {\em Journal of Investigative Medicine} {\bf 64}, 1166--1171.
\item
Colquhoun, D. (2014). An investigation of the false discovery rate and the misinterpretation of p-values. {\em Royal Society of Open Science} {\bf 1}, 140--216.
\item
Cumming, G. (2014). The new statistics: why and how. {\em Psychological Science} {\bf 25}, 7--29.
\item
Donahue, R. M. J. (1999). A note on information seldom reported via the P value.
{\em The American Statistician} {\bf 53}, 303--306.
\item
Dudley, R. M. and Haughton, D. (2002). Asymptotic normality with small relative errors of
posterior probabilities of half-spaces.
\emph{ The Annals of Statistics} {\bf 30}, 1311--1344.
\item
Fidler, F., Thomason, N., Cumming, G., Finch, S., Leeman, J. (2004). Editors can lead researchers to confidence intervals, but can't make them think: Statistical reform lessons from medicine. {\em Psychological Science} {\bf 15}, 119--126.
\item
Follmann, D. (1996). A simple multivariate test for one-sided alternatives.
{\em Journal of the American Statistical Association} {\bf 91}, 854--861.
\item
Gill, J. (2018). Comments from the New Editor. {\em Political Analysis} {\bf 26}, 1--2.
\item
Goodman, S. N. (1999). Toward evidence-based medical statistics. 1: the p value fallacy. {\em Annals of Internal Medicine} {\bf Volume 130}, 995--1004.
\item
Hubbard, R. and Lindsay, R. M. (2008). Why P values are not a useful measure of evidence in statistical significance testing. {\em Theory $\&$ Psychology} {\bf 18}, 69--88.
\item
Hung, H. J., O'Neill, R. T., Bauer, P., Kohne, K. (1997). The behavior of the p-value when the alternative hypothesis is true. {\em Biometrics} {\bf 53}, 11--22.
\item
Ioannidis, J. P. (2005). Why most published research findings are false. {\em PLoS Medicine} {\bf 2}, 124.
\item
Jager, L. R. and Leek, J. T. (2014).
An estimate of the science-wise false discovery rate
and application to the top medical literature. {\em Biostatistics} {\bf 15}, 1--12.
\item
Johnson, V. E. (2013).
Revised standards for statistical evidence.
{\em Proceedings of the National Academy of Sciences} {\bf 110}, 19313--19317.
\item
Lee, J. J. (2010). Demystify statistical significance--time to move on from the p-value to Bayesian analysis. {\em Journal of the National Cancer Institute} {\bf 103}, 16--20.
\item
Leek, J., McShane, B. B., Gelman, A., Colquhoun, D., Nuijten, M. B., Goodman, S. N. (2017). Five ways to fix statistics. {\em Nature} {\bf 551}, 557--559.
\item
Lehmann, E. L. and Romano, J. P. (2005). Testing Statistical Hypotheses. New
York: Springer.
\item
Lindley, D. V. (1957). A statistical paradox. {\em Biometrika} {\bf 44}, 187--192.
\item
Liu, H. and Berger, R. L. (1995). Uniformly more powerful, one-sided tests for hypotheses about linear inequalities.
{\em The Annals of Statistics} {\bf 23}, 55--72.
\item
McShane, B. B., Gal, D., Gelman, A., Robert, C., Tackett, J. L. (2018). Abandon statistical significance. arXiv:1709.07588
\item
Meng, X. L. (1994). Posterior predictive p-values.
{\em The Annals of Statistics} {\bf 22}, 1142--1160.
\item
Murtaugh, P. A. (2014). In defense of P values.
{\em Ecology} {\bf 95}, 611--617.
\item
Nuzzo, R. (2014).
Statistical errors: P values, the `gold standard'
of statistical validity, are not as reliable as many scientists assume.
{\em Nature} {\bf 506}, 150--152.
\item
Ranstam, J. (2012). Why the P-value culture is bad and confidence intervals a better alternative. {\em Osteoarthritis Cartilage} {\bf 20}, 805-808.
\item
Rosenthal, R. and Rubin, D. B. (1983). Ensemble-adjusted p values. {\em Psychological Bulletin} {\bf 94}, 540--541.
\item
Royall, R. M. (1986). The effect of sample size on the meaning of significance tests. {\em The American Statistician} {\bf 40}, 313--315.
\item
Rubin, D. B. (1984). Bayesianly justifiable and relevant frequency calculations for the applies statistician.
{\em The Annals of Statistics} {\bf 12}, 1151--1172.
\item
Rubin, D. B. (1998).
More powerful randomization-based p-values in double-blind trials with non-compliance.
{\em Statistics in Medicine} {\bf 17}, 371--385.
\item
Sackrowitz, H. and Samuel-Cahn, E. (1999). P values as random variable-expected P values. {\em The American Statistician} {\bf 53}, 326--331.
\item
Sasabuchi, S. (1980). A test of a multivariate normal mean with
composite hypotheses determined by linear inequalities. {\em Biometrika}
{\bf 67}, 429--439.
\item
Savalei, V. and Dunn, E. (2015). Is the call
to abandon p-values the red herring of
the replicability crisis? {\em Frontiers in Psychology} {\bf 6}, 245.
\item
Schervish, M. J. (1996). P values: what they are and what they are not. {\em The American Statistician} {\bf 50}, 203--206.
\item
Sellke, T., Bayarri, M. J., and Berger, J. O. (2001). Calibration of p-values for testing precise null hypotheses. {\em The American Statistician} {\bf 55}, 62--71.
\item
Simmons, J. P., Nelson, L. D., Simonsohn, U. (2011). False-positive psychology: undisclosed flexibility in data collection and analysis allows presenting anything as significant. {\em Psychological Science} {\bf 22}, 1359--1366.
\item
Trafimow, D., Amrhein, V., Areshenkoff, C. N.,
Barrera-Causil, C. J., Beh, E. J., et al. (2018). Manipulating the
alpha level cannot cure significance testing. {\em Frontiers in Psychology} {\bf 9}, 699.
\item
Trafimow, D. and Marks, M. (2015). Editorial. {\em Basic and Applied Social Psychology} {\bf 37}, 1--2.
\item
Wagenmakers, E. J. (2007). A practical solution to the pervasive problems of p values. {\em Psychonomic Bulletin $\&$ Review} {\bf 14}, 779--804.
\item
Wasserstein, R. L. and Lazar, N. A. (2016). The ASA's statement on p-values: context, process, and purpose. {\em The American Statistician} {\bf 70}, 129--133.
\end{description}
\newpage
\begin{figure}[htb]
\begin{center}
\includegraphics[height=8cm,width=8cm]{t1.eps}
\includegraphics[height=8cm,width=8cm]{t12.eps}\\
\includegraphics[height=8cm,width=8cm]{pwr1.eps}
\includegraphics[height=8cm,width=8cm]{pwr.eps}
\end{center}
\caption{Comparison of the type I error rate and power under the frequentist $Z$-test and Bayesian test based on the posterior probability for detecting treatment difference $\delta = 0.15$ (left) and $\delta=0.1$ (right).
}
\label{comparet1}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_single_onesided20.eps}
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_onesided20.eps}\\
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_single_onesided50.eps}
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_onesided50.eps}\\
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_single_onesided100.eps}
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_onesided100.eps}\\
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_single_onesided500.eps}
\includegraphics[height=5cm,width=5cm]{pv_vs_pop_onesided500.eps}\\
\end{center}
\caption{The relationship between $p$-value and the posterior probability over 1000 replications under one-sided one-sample and two-sample hypothesis tests with binary outcomes under sample sizes of 20, 50, 100 and 500 per arm, respectively.
}
\label{os}
\end{figure}
\newpage
\begin{figure}[htb]
\begin{center}
\includegraphics[height=8cm,width=8cm]{sum_n20.eps}
\includegraphics[height=8cm,width=8cm]{sum_n50.eps}\\
\includegraphics[height=8cm,width=8cm]{sum_n100.eps}
\includegraphics[height=8cm,width=8cm]{sum_n500.eps}
\end{center}
\caption{The differences between $p$-values and posterior probabilities over 1000 replications in one-sided two-sample hypothesis tests with binary outcomes under sample sizes of 20, 50, 100 and 500, respectively.
}
\label{three}
\end{figure}
\newpage
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6.5cm,width=6.5cm]{pv_vs_pop_single.eps}
\includegraphics[height=6.5cm,width=6.5cm]{pv_vs_pop.eps}
\\
\includegraphics[height=6.5cm,width=6.5cm]{pv_vs_pop_single2.eps}
\includegraphics[height=6.5cm,width=6.5cm]{pv_vs_pop2.eps}
\\
\includegraphics[height=6.5cm,width=6.5cm]{pv_vs_pop_single3.eps}
\includegraphics[height=6.5cm,width=6.5cm]{pv_vs_pop3.eps}
\end{center}
\caption{The relationship between $p$-value and the posterior probability over 1000 replications under two-sided one-sample and two-sample hypothesis tests with binary outcomes under sample size of 500 per arm.}
\label{ts}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop1.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop2.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop3.eps}\\
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop4.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop5.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop6.eps}\\
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop7.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop8.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop9.eps}\\
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop10.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop11.eps}
\includegraphics[height=5cm,width=5cm]{mu_pv_vs_pop12.eps}
\end{center}
\caption{The relationship between $p$-value and the posterior probability over 1000 replications under one-sided and two-sided hypothesis tests with normal outcomes assuming Jeffreys' prior and vague normal-inverse-gamma prior under sample size of 20, 50 and 100, respectively.
}
\label{normal_jeff}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=5cm,width=5cm]{mu_gamma1.eps}
\includegraphics[height=5cm,width=5cm]{mu_gamma2.eps}
\includegraphics[height=5cm,width=5cm]{mu_gamma3.eps}\\
\includegraphics[height=5cm,width=5cm]{mu_gamma4.eps}
\includegraphics[height=5cm,width=5cm]{mu_gamma5.eps}
\includegraphics[height=5cm,width=5cm]{mu_gamma6.eps}
\end{center}
\caption{The relationship between $p$-value and the posterior probability over 1000 replications under one-sided hypothesis tests with outcomes generated from Gamma, Beta and mixture normal distributions, assuming Jeffreys' prior for the normal distribution under sample size of 20 and 50, respectively.
}
\label{normal_gamma}
\end{figure}
\newpage
\begin{figure}[htb]
\begin{center}
\includegraphics[height=6.5cm,width=6.5cm]{pop_normal_informative_1000.eps}
\includegraphics[height=6.5cm,width=6.5cm]{pop_normal_informative_prior1000.eps}
\\
\includegraphics[height=6.5cm,width=6.5cm]{pop_normal_informative_10000.eps}
\includegraphics[height=6.5cm,width=6.5cm]{pop_normal_informative_prior100.eps}
\\
\includegraphics[height=6.5cm,width=6.5cm]{pop_normal_informative_100000.eps}
\includegraphics[height=6.5cm,width=6.5cm]{pop_normal_informative_prior1.eps}
\end{center}
\caption{The relationship between $p$-value and the posterior probability $\Pr(\mu_E \le \mu_S|D)$ over 1000 replications under one-sided hypothesis tests with normal outcomes; left panel: assuming a fixed informative normal-inverse-gamma prior under increasing sample sizes of 1000, 10000 and 100000 (from top to bottom), right panel: assuming
a fixed sample size of 1000 with an increasing prior variance of 0.001, 0.01 and 1 (from top to bottom).}
\label{informative}
\end{figure}
\begin{figure}[htb]
\begin{center}
\includegraphics[height=8cm,width=8cm]{multivariate_normal.eps}
\includegraphics[height=8cm,width=8cm]{multivariate_normal3.eps}\\
\includegraphics[height=8cm,width=8cm]{multivariate_normal2.eps}
\includegraphics[height=8cm,width=8cm]{multivariate_normal4.eps}
\end{center}
\caption{The relationship between $p$-value and the posterior probability over 1000 replications under one-sided and two-sided hypothesis tests with multivariate normal outcomes under sample size of 100.
}
\label{multi}
\end{figure}
\end{document}
|
\section{Introduction}
Suspended thin films are widely used in photonics and sensing applications. Among them, silicon nitride films deposited with low pressure chemical vapor deposition benefit from excellent optical properties, high refractive index and ultralow loss in the visible and near-infrared range, as well outstanding mechanical properties, high frequency/quality factor mechanical resonances and high burst pressure. Due to these features as well as their low mass, thin SiN films with thickness ranging from tens to a few hundreds of nanometers have been applied in a number of cavity optomechanics investigations~\cite{Thompson2008,Jayich2008,Wilson2009,Camerer2011,Karuza2012,Purdy2013,Sawadsky2015,Xu2016,Nielsen2017}. Increasing the otherwise relatively low reflectivity of such membranes without increasing their mass is beneficial for cavity optomechanics with single~\cite{Kemiktarak2012apl,Kemiktarak2012njp,Bui2012,Stambaugh2015,Bernard2016,Norte2016,Reinhardt2016,Chen2017,Moura2018} or multiple resonators~\cite{Bhattacharya2008,Hartmann2008,Xuereb2012,Xuereb2014,Li2016,Nair2016,Piergentili2018}, as well as for sensing~\cite{Bruckner2010,Guo2017,Naesby2018} or lasing~\cite{Huang2008,Kemiktarak2014,Yang2015} applications. Such an increase in reflectivity for films with subwavelength thickness can be realized by patterning the film with a suitable subwavelength periodic structure, such as a high contrast grating~\cite{ChangHasnain2012} or a photonic crystal structure~\cite{Zhou2014}. The interference between the incoming light and guided modes in the structure allows for tailoring the optical properties of the film. In particular, the destructive interference occuring for specific wavelengths/polarizations can give rise to the appearance of either broad- or narrowband Fano-like resonances in the reflectivity spectrum~\cite{Miroshnichenko2010}. In a broader context, resonant waveguide gratings and photonic crystal structures are relevant for a variety of optical applications, including waveguide coupling, optical filtering, polarizers, spectrometry, biosensing, solar cells, photodetection, lasing, etc.~\cite{Quaranta2018}.
Enhanced reflectivity of suspended silicon nitride membranes has been observed both for two-dimensional photonic crystals~\cite{Bui2012,Bernard2016,Norte2016,Chen2017,Moura2018,Gartner2018} and for one-dimensional subwavelength gratings (SG)~\cite{Kemiktarak2012apl,Kemiktarak2012njp,Stambaugh2015} patterning. We focus here on the realization of polarization-dependent, one-dimensional SGs similar to those of Refs.~\cite{Kemiktarak2012apl,Kemiktarak2012njp,Wang2015,Stambaugh2015}, albeit by following the approach of~\cite{Chen2017}~and \textit{directly} patterning a $50\times 50$ $\mu$m$^2$ subwavelength grating on a commercial, 200 nm-thick silicon nitride membrane. A Fano resonance is observed in the transmission spectrum of transverse magnetic (TM) polarized light impinging at normal incidence, showing an increase in reflectivity at 937 nm from 10\% for an unpatterned membrane to 78\% for a patterned one. By using localized cuts of the structure with a Focused Ion Beam~\cite{Ierardi2014} we also show that the grating transverse grating profile of the suspended film can be measured. We compare the observed transmission spectrum with the results of Rigorous Coupled Wave Analysis numerical simulations taking into account the grating profile and discuss collimation broadening and finite grating size effects.
Last, by mounting the membrane chip on a ring piezoelectric transducer and applying a compressive force to the corners of the substrate we show as a proof of principle that it is possible to shift the transmission spectrum by 0.23 nm towards lower wavelengths. Optimizing and increasing the piezoelectric compressive force further would make such patterned membranes interesting for realizing tunable optical filters~\cite{Wang2015}, waveplates~\cite{Mutlu2012} or strongly focusing lenses~\cite{Fattal2010,Lu2010,Klemm2013}. Increasing the patterned area~\cite{Chen2017,Moura2018} and the reflectivity would also make them attractive for vertical-cavity surface-emitting lasers~\cite{Huang2008,Zhou2008} or optical sensors~\cite{Guo2017,Naesby2018} for biophysics~\cite{Schuler2009,Dong2017} and biomedical~\cite{Leinders2015,Sinha2017} applications. Last, combined with piezoelectric actuation of their mechanical modes~\cite{Wu2018,Naserbakht2019}, such enhanced and electrically tunable reflectivity membranes would be particularly interesting for improving and tailoring the optomechanical response of arrays of nanomembranes~\cite{Nair2017,Piergentili2018,Gartner2018} and for investigating collective and strong coupling optomechanics~\cite{Xuereb2012,Xuereb2013,Xuereb2014,Xuereb2015,Cernotik2019}.
\section{Fabrication}
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{fig1.jpg}
\caption{(a-f) SG fabrication process. (a) Suspended SiN membrane on Si, (b) Coating with PMMA and CP, (c) EBL, (d) Development, (e) Etching, (f) PMMA removal. (g) Schematic cross section of a SG with tilted walls. (h) Topview SEM image of the $50\times 50$ $\mu$m$^2$ SG patterned area.}
\label{fig:fab}
\end{figure}
The stochiometric silicon nitride membranes used in this work are commercial~\cite{Norcada}, high stress ($\sim$GPa), 0.5~mm-square and 200~nm-thick films on a 5~mm-square, 500~$\mu$m-thick silicon frame. The steps for patterning the SG are depicted in Fig.~\ref{fig:fab}a and are as follows: after cleaning with an O$_2$ plasma the sample is spin-coated with a 9\% 950k molecular weight PMMA resist and a conductive polymer layer (SX-AR-PC 5000/90.2) to avoid charging effects during Electron Beam Lithography (EBL). A $50\times50$ $\mu$m$^2$ grating mask with a target period and finger width is then written by EBL at 30 kV. After writing the conductive polymer layer is removed by immersion in deionized water and the PMMA resist is developed in a solution of 3:7 H$_2$O:IPA (Iso-propyl alcohol). The membrane is then etched in a STS Pegasus ICP DRIE system using reactive ion etching with C$_4$F$_8$ (flow rate 59 sccm) and SF$_6$ (flow rate 36 sccm) for 170 seconds at 800 W. The PMMA layer is removed in acetone followed by rinsing in IPA and N$_2$ blow drying. The sample is finally cleaned by means of the O$_2$ plasma again.
The etching parameters are observed to be quite critical for the success of the patterning, and etching only a fraction of the 200 nm films is determinant in avoiding rupture/damage during the process. A top-view SEM picture of the resulting grating is shown in Fig.~\ref{fig:fab}h. From such SEM images the period and grating finger width are estimated to be $a=(810\pm 15)$ nm and $w=(450\pm 20)$ nm, respectively, where the uncertainties come from the imaging system calibration. The latter number is potentially deceiving, however, as the vertical profile of the grating is not expected to be rectangular and the observed mean finger width depends on the saturation level of the SEM image.
\begin{figure}[h!]
\centering\includegraphics[width=\columnwidth]{fig2.jpg}
\caption{(a) Schematic of the FIB cutting setup. (b) SEM topview image of the SG after cutting. (c) SEM cross section view of the SG profile at position 1.}
\label{fig:fib}
\end{figure}
To get more insight into the vertical profile of the grating structure a Focused Ion Beam (FIB) was used to cut through the SG at different locations of the patterned area, as shown in Figs.~\ref{fig:fib}a and b. Before the FIB cut the membrane was made more robust by coating on both the etched and non etched sides with $\sim 50$~nm and $\sim 200$~nm of gold, respectively, and a thick ($\simeq 2\mu$m) carbon layer was deposited at the various cutting positions. The gold layers prevent the structure from fracturing and breaking when exposed to the ion beam. Cuts in several locations of the same grating can then be performed, as illustrated in Fig.~\ref{fig:fib}b. A high energy Ga$^+$ ion beam is focused at the edge of the cutting positions defined by the carbon layers, as shown in Figs.~\ref{fig:fib}a and b. The grating cross section can then be imaged at a 45$^{\circ}$ angle with a SEM microscope (Fig.~\ref{fig:fib}c). As expected, the grating fingers are observed to have smooth edged walls, whose width substantially varies from top to bottom. Analyzing these images yield for this particular sample a period $a= (806\pm 10)$~nm, mean/top/bottom/mean finger widths of $w_m= (512\pm10)$~nm, $w_t= (395\pm14)$~nm and $w_b= (628\pm15)$~nm, respectively. The grating height is $h= (109\pm 8)$~nm and the underlying Si$_3$N$_4$ layer thickness $h'= (87\pm 7)$~nm. The transverse profile obtained from these images can be used as input to the numerical simulations to predict transmission spectra, as will be discussed below.
\section{Optical characterization}
\begin{figure}[h!]
\centering\includegraphics[width=0.8\columnwidth]{fig3.pdf}
\caption{Normal incidence normalized transmission of the patterned membrane for TE (full symbols) and TM (open symbols) polarized light incident on (blue and magenta) and outside (red) the patterned area. The solid/dashed lines are the results of the RCWA simulations for the grating parameters given in the text. For TM polarized light incident on the SG the dashed line shows the result of RCWA simulations of a plane wave impinging at normal incidence on an infinite grating, while the solid line shows the result of simulations of a Gaussian beam (superposition of plane waves incident at different incidence angles) on the grating (see text for details).}
\label{fig:trans}
\end{figure}
The transmission of a patterned membrane similar to the one described in the previous section was measured by focusing onto the sample linearly polarized light issued from a tunable external cavity diode laser with a $f=60$ mm focal length achromat doublet (spotsize $\sim$30 $\mu$m). Figure~\ref{fig:trans} shows the normalized transmission measured for TE and TM polarized light impinging at normal incidence either on or outside the patterned area.
Outside the patterned area the transmissivity of the film is independent of the polarization of the light and its level is determined by the film thickness and refractive index. From ellipsometry measurements and broadband transmission spectroscopy~\cite{Nair2017} the thickness and refractive index of unprocessed samples from the same fabrication batch were determined to be $t\simeq 200$ nm and $n\simeq 1.982$ in the plotted range (920-960 nm). Consistently with these values the measured transmissiondecreases from 12 to 8\% in that range. When light impinges on the patterned area an overall small drop in transmission is observed for TE polarized light, on account of the reduced effective thickness of the patterned film. A much more stronger wavelength-dependent drop is observed for TM polarized light around 937 nm, as a result of the interference between the incoming light and guided modes in the structure in this wavelength range.
To investigate the observed spectrum we performed numerical simulations based on a RCWA approach~\cite{MIST} and using as input parameters the refractive index $n=1.982$ determined previously, the grating period $a=815$ nm and mean finger width $w_m=530$ nm estimated from topview SEM images and the approximate grating profile shown in Fig.~\ref{fig:fab}g with the thicknesses $h=109$ nm and $h'=87$ nm determined from the FIB cut measurements. In this approach a plane wave is incident on an infinite grating, whose unit cell is defined by the previous parameters (20 layers were used to discretize the structure), and the transmitted or reflected fields can be computed. The dashed blue line in Fig.~\ref{fig:trans} shows the result of such a simulation for a plane wave impinging on the grating at normal incidence ($\theta=0$), which predicts a narrower Fano resonance with zero transmission occuring around 934.5 nm. The beam incident on the grating is however a focused Gaussian beam with a waist $w_0\sim 15$ $\mu$m. This beam can be considered as a superposition of plane waves impinging on the grating with various incident angles $\theta$~\cite{Crozier2006,Bernard2016,Moura2018}. The transmittivity of the collimated beam is then computed by performing a weighted average of the RCWA-simulated transmittivities of plane waves incident at different angles, with a Gaussian distribution $e^{-2\theta^2/\theta_D^2}$, where $\theta_D=\lambda/\pi w_0$ is the Gaussian beam divergence angle ($\theta_D\simeq 1.1^{\circ}$ in our case). The resulting spectrum (solid blue line in Fig.~\ref{fig:trans}) shows a shifted and broadened Fano resonance displaying non-zero minimal transmission, as compared to the infinite plane wave case, and is closer to the measured spectrum, although the latter is somehow still more shifted and broadened. Slightly altering the grating parameters in the simulations around the measured values within the systematic uncertainties does not qualitatively change this picture; while collimation broadening can be expected to play a non-negligible role, other effects are likely to contribute to the broadening and shifting of the spectrum. Effects due to the finite size of the grating could also play a role. These could in principle be numerically simulated using full three-dimensional finite element analysis; however, such simulations are computationally heavy and were not performed here. Instead, we note that an estimate of the broadening of a waveguide resonance due to finite size effects can be obtained following analysis of guided-mode resonance filters and couplers~\cite{Brazas1995,Saarinen1995,Boye2000} and is approximately given by $C\lambda a/l$, where $l$ is the grating dimension and $C$ a constant of order unity. In our case ($l=50$ $\mu$m), this broadening would be of the order of 15 nm, but may be expected to be overestimated given the relatively high contrast of the grating structure~\cite{Quaranta2018,Ko2018}. It is however reasonable to think that finite size effects contribute to some extent to the broadening of the observed spectrum. Let us point out that both collimation broadening and finite size effects could be reduced by increasing the size of the patterned area in the future~\cite{Chen2017,Moura2018}. Another possible broadening effect would be inhomogeneities in the grating structure; these could in principle be investigated by applying the method demonstrated in the previous section and performing cuts in several locations of the grating.
\section{Piezoelectric tuning}
\begin{figure}[h]
\centering\includegraphics[width=0.33\columnwidth]{fig4a.jpg}\\
\centering\includegraphics[width=0.65\columnwidth]{fig4b.pdf}\\
\centering\includegraphics[width=0.65\columnwidth]{fig4c.pdf}
\caption{Top: Topview schematic (to scale) of the patterned membrane chip mounted on the ring piezoactuator (blue: SiN membrane, dark grey: Si chip, light grey: piezoactuator). The red dot indicate the position of the SG on the membrane. Middle: Normalized transmission spectra for TM polarized light at normal incidence on the SG, with 0 V (blue) and with 180 V (orange) applied to the piezoactuator. The dots are the average of 3 spectra and the statistical errors are smaller than the dot size. The solid lines are the results of fits with a Fano profile. Bottom: zoom of these spectra around the Fano resonance.}
\label{fig:actuation}
\end{figure}
In subsequent experiments three corners of the patterned membrane were glued on a ring piezoelectric actuator (Noliac NAC2123) with 6 mm inner diameter and a specified load-free inner radius contraction of $\Delta r\sim -9$ $\mu$m for an applied 180 V voltage. The grating being offset from the membrane center by about $d\sim160$ $\mu$m its period is slightly reduced by the inward contraction of the SiN film and a global shift of the Fano resonance can then be expected. The same transmission measurements as in the previous section were performed with and without voltage applied to the actuator. The results are shown in Fig.~\ref{fig:actuation} for TM polarized light at normal incidence. The spectrum is clearly shifted towards lower wavelengths when a 180 V voltage is applied. The solid lines show the result of fits with a Fano profile of the form
\begin{equation}
T(\lambda)=A+B\frac{\left(\lambda-\lambda_0+q\gamma\right)^2}{(\lambda-\lambda_0)^2+\gamma^2}
\end{equation}
and match the experimental spectra fairly well. A shift of the resonance frequency $\lambda_0$ of $-0.23\pm 0.03$ nm can be inferred from the fits to the Fano profiles. This is consistent with what can be expected from the resulting reduction in grating period taking into account its off-centered position. The RCWA simulations show that the relative shift in resonance wavelength scales like the relative change in the grating period, i.e. $\Delta\lambda_0/\lambda_0\sim\Delta a/a$. Assuming a linear contraction between the edges of the chip and its center, one has that $\Delta a/a\sim \sqrt{2}(\Delta r/L)(d/L)$, where $2L=5$ mm is the chip lateral dimension. One then gets an expected resonant wavelength shift $\Delta \lambda_0\sim-0.3$ nm, in reasonable agreement with the measured shift.
Let us note that the shift is relatively modest owing to the piezoelectric transducer and geometry used for this device, which was not optimized for this purpose, but rather for optomechanical applications in which preservation of the high quality factor of the mechanical resonances requires minimizing clamping losses as much as possible~\cite{Wu2018,Naserbakht2019}. The 6 mm-inner diameter of the piezoelectric ring actuator was thus chosen only slightly smaller than the 7 mm chip diagonal dimension. For photonics applications, such as low loss tunable optical filters and waveplates, much larger shifts could in principle be obtained by increasing the compression force and optimizing the position of the grating with respect to the piezoelectric transducer. For instance, using larger membrane windows and/or different actuation geometry, as in e.g.~\cite{Wang2015}, could readily yield grating period changes $\Delta a$ of several nanometers. Second, the optical quality factor of the Fano resonance used here is not particularly high ($\sim 90$). Making use of subwavelength gratings with much higher optical quality factor resonances ($Q>14 000$ was reported in~\cite{Zhou2008}) would readily improve the switching contrast of such a filter. For instance, assuming a reasonable ten times larger period change of 3 nm and an optical Q of $\sim 1000$ would give $\Delta\lambda_0/\lambda_0\sim 3\times 10^{-3}\gg 1/Q$ and a switching contrast of about 10 dB. To verify this, the results of RCWA simulations of the transmission of an infinite SG with $a=729$, $w_m=705$ nm, $h=50$ nm and $h'=150$ nm with and without a compression corresponding to such a 3 nm grating period change are shown in Fig.~\ref{fig:sim_act}.
\begin{figure}[h]
\centering\includegraphics[width=0.65\columnwidth]{fig5.pdf}
\caption{RCWA simulated spectra of a high-Q ($\sim1000$) SG (see text for parameters) without (blue) and with (orange) a compression corresponding to a 3 nm grating period change.}
\label{fig:sim_act}
\end{figure}
Finally, let us mention that the use of a high optical Q subwavelength grating, even with modest electrical tunability, as the flexible end-mirror of an ultrashort cavity would be interesting for investigating new regimes of cavity optomechanics with strongly-wavelength dependent reflectors~\cite{Naesby2018,Cernotik2019}.
\section{Conclusion}
The possibility to directly pattern one-dimensional subwavelength gratings on commercial, high-stress, 200~nm-thick suspended silicon nitride films was demonstrated. An enhancement of the reflectivity from 10\% to 78\% was observed for 937 nm TM polarized light focused on a $50\times 50$ $\mu$m$^2$ grating. The broadening and non-zero minimal transmission of the Fano resonance observed in the transmission spectrum were discussed based on RCWA simulations and using transverse profile images of the grating obtained by FIB cutting. Increasing the patterned area and further optimizing of the writing/etching process using the FIB cutting method are expected to reduce collimation broadening and finite grating size effects, and thereby increase the reflectivity. Fine tunability of the optical spectrum by piezoelectric contraction of the suspended film was furthermore evidenced. Such enhanced and electrically tunable reflectivity membranes would be interesting for investigations of collective phenomena in optomechanical arrays of nanomembranes as well as for a number of photonics and optical sensing applications.
\section*{Acknowledgments}
We are grateful to Pia Bomholt Jensen for assistance with FIB cutting and Folmer Lyckegaard for assistance with gold coating, and acknowledge support from the Velux Foundations, the Danish Council for Independent Research (Sapere Aude initiative) and the Carlsberg Foundation.
\bibliographystyle{iopart-num.bst}
|
\section{Introduction}
Schizophrenia is a chronic psychiatric disorder that affects more than 21 million people worldwide~\cite{WHOSchizophrenia}. Up to 80\% of the risk factors appear to be genetic, although it has proven difficult to identify the specific genes that are involved in the disease~\cite{Bertolino2009II}. The disease usually commences in early adulthood, and symptoms range from hallucinations and avolition to cognitive deficits (such as impaired working memory)~\cite{WHOSchizophrenia,Dawson2014}. It is believed that the cognitive deficits arise from compromised functional integration between neural subsystems~\cite{Bullmore1997,Peled1999, Bassett2008,Dawson2014}. There can be significant differences in the properties of time series from imaging measurements of healthy versus schizophrenic individuals, although different studies have found seemingly contradictory results when comparing functional magnetic resonance imaging (fMRI) time series from two distinct brain regions in a schizophrenia patient and a healthy control. The majority of studies have concluded that schizophrenia patients have less-similar time series than healthy controls across different brain regions
\cite{Fornito2012}. Zalesky \emph{et al.}~\cite{Zalesky2012} suggested that such reduced similarity may arise from an altered coupling between brain regions and local decoherence within brain regions in schizophrenia patients. However, some studies have observed that schizophrenia patients have more-similar series than healthy controls across brain regions. For a detailed discussion of these seemingly contradictory findings, see~\cite{Fornito2015}. In some cases, methodological steps in fMRI analyses seem to yield increases in these similarities, but abnormal neurodevelopment or drug treatment may play a role in increasing them in other cases~\cite{Fornito2015}.
One approach for studying the human brain is to construct a (possibly time-dependent) neuronal network based on experimental data and then analyze the network's structure and dynamics to gain insights into its properties~\cite{Bullmore2009,Bullmore2011,Sporns2014,Papo2014,Papo2014II,Betzel2016,bassett2017,bassett2018}.
One can form a so-called \emph{functional network}~\cite{Bullmore2011,Bullmore2009,Sporns2015,Petersen2015,Stolz2017}, in which each node represents a brain region and one weights the edges between them based on some measure of the similarity between the nodes' fMRI time series. In Fig.~\ref{fig:FunctionalNetwork}, we show a pipeline of how to construct a functional network from fMRI time-series data. Additionally, when interpreting functional networks in fMRI studies, it is very important to consider the cautionary notes in \cite{eklund2016}.
Studies of functional networks of schizophrenia patients have revealed that such networks differ significantly from the functional networks of healthy controls~\cite{Lynall2010,Rubinov2013,Alexander-Bloch2012,Bassett2008,Liu2008,Fornito2012,Singh2016}. For example, schizophrenia patients can have rather different community structure
than controls~\cite{Alexander-Bloch2012,flanagan2018}.
For example, Alexander-Bloch \emph{et al.}~\cite{Alexander-Bloch2012} observed that a small subset of brain regions lead to significant differences in the community assignments in schizophrenia patients, whereas the communities for healthy subjects appear to be consistent with each other. Moreover, the maximum modularity of functional networks appears to be smaller for schizophrenia patients than for healthy controls~\cite{Alexander-Bloch2012, Alexander-Bloch2010}. Two recent papers, Flanagan \emph{et al.}~\cite{flanagan2018} and Towlson \emph{et al.}~\cite{towlson2019}, compared the network structures of schizophrenia patients and healthy controls under the effects of different drugs and a placebo.
\begin{figure}[h!]
\centering\includegraphics[width=\textwidth]{fMRIFunctional8-eps-converted-to.pdf}
\caption{Pipeline to construct functional networks from imaging data (e.g., fMRI data).
}
\label{fig:FunctionalNetwork}
\end{figure}
An increasingly popular approach for the analysis of functional networks is to use ideas from computational algebraic topology, as these approaches allow one to characterize topological invariants (such as connectedness and loops) in high-dimensional data
~\cite{Edelsbrunner2002,Edelsbrunner2008, Ghrist2008,Carlsson2009,Edelsbrunner2010,sizemore2018,batt2020}. In contrast to standard methods of network analysis \cite{Newman2018}, employing computational topology allows one to explicitly go beyond pairwise connections; this is helpful for gaining global understanding of low-dimensional structures in networks. Although one can also use frameworks such as hypergraphs~\cite{Bollobas1998} to study higher-order network structures (see, e.g., ~\cite{Bassett2014}), such a formalism does not by itself give direct information about the shape or scale of mesoscale features in networks. By contrast, \emph{persistent homology} (PH), the most prominent approach in topological data analysis, allows one to explore the persistence of features (such as connectedness and loops) in data sets \cite{otter2017,Patania2017}. Persistent homology has led to interesting insights in a variety of fields (for {examples, see~\cite{kramar2013,taylor2015,bgk2015,topaz2014,Bendich2014,Feng2020,Byrne2019})}; and it has been used increasingly in neuronal networks, leading to {several promising insights~\cite{Curto2008,Dabaghian2012,Petri2013,Lee2011,Giusti2015,Spreemann2015,Curto2016,Giusti2016,Reimann2017,Dabaghian2016,Stolz2017,Lee2019,Bardin2018,chung2019,babichev2018,geniesse2019,Ibanez2019}. }
In the present paper, we construct functional networks using fMRI data from schizophrenia patients, healthy controls, and siblings of schizophrenia patients. We create a nested sequence of networks in which we add edges, one by one, to the networks in order from the largest edge weights to the smallest. (In the unlikely case of two edges having the exact same weight, we add both edges simultaneously in one step.) We then construct a
weight-rank clique filtration (WRCF) \cite{Petri2013} by determining cliques and tracking their changes in each step of the network sequence.
We then compute PH and Betti numbers~\cite{Edelsbrunner2010,Croom} of the WRCF and examine the results by applying tools from statistics and machine learning, respectively, on the persistence landscapes and persistence images that result from our computation of PH. We compare our findings from these two approaches. We focus on loops (with four or more edges) \footnote{We use the term `loop' to refer to at least four edges in a network that are connected in a way that forms a cycle. Conventionally, loops (other than self-loops) in undirected graphs must
have at least 3 edges, and loops in directed graphs must have at least 2
edges. In our paper, we adapt this terminology to represent the topological
features that we detect in our simplicial complexes.} in the networks in our nested sequence, rather than on connected components, because one can also study the latter using more conventional approaches (such as by computing the spectrum of the combinatorial graph Laplacian matrix \cite{Bollobas1998,Newman2018}). Although it is interesting to also consider higher-dimensional topological features in these functional networks, the computational cost of PH is very high~\cite{otter2017} (especially in higher dimensions), and we therefore focus on the analysis of loops.
Our paper proceeds as follows. We introduce the data set and the mathematical methods in an intuitive way in Section~\ref{Sec:MatMethods}, present our findings in Section~\ref{Sec:Results}, and discuss our comparisons in the context of current biological research in Section~\ref{Sec:Discussion}. We give some additional details about a few results and report our results of our computation of Betti curves in \ref{S1}.
\section{Methods}\label{Sec:MatMethods}
\subsection{Data set: {fMRI} data of schizophrenia patients, siblings of schizophrenia patients, and healthy controls}
We use a data set that consists of time series from blood oxygen level-dependent (BOLD) functional magnetic resonance imaging (fMRI) data
that was collected from 281 subjects (encompassing 54 schizophrenia patients, 50 healthy siblings of schizophrenia patients, and 177 healthy controls) with 120 time steps (where the length of $1$ time step corresponds to $\Delta t = 2\, \text{s}$). The brain regions were determined according to the Montreal Neurological Institute template~\cite{Talairach1988}. Prior to obtaining the time series, the fMRI data were corrected for head motion, and they were normalized and smoothed with a Gaussian filter. The voxel-wise signal intensities were normalized to the whole-brain global mean. The data set was acquired by Bertolino, Blasi, and their collaborators as part of a larger fMRI data set over a period of approximately 10 years. Subsets of the data set have been studied previously \cite{Bertolino2010,Sambataro2009,Rampino2014}, although these previous studies of the data did not include the data for siblings.
The experimentalists obtained fMRI images while subjects were performing a block paradigm of a so-called `$n$-back task'.
During an $n$-back task, subjects are presented with a sequence of {stimuli (such as numbers)}. In each step $m$ of the sequence, subjects are first shown a number and then asked to recall the number from sequence step $m-n$. For example, during a $2$-back task, subjects are shown a sequence
$\{\dots, x_{i-1}, x_i, x_{i+1}, x_{i+2}, \dots\}$ and are asked to recall number $x_{i-1}$ while being shown number $x_{i+1}$, recall number $x_i$ while being shown number $x_{i+2}$, and so on. For the present data set, the {block paradigm} consisted of alternating blocks of {four} $0$-back tasks and {four} $2$-back tasks.
We preprocess the data to remove noise that arises due to (1) contributions from brain white matter~\cite{Weissenbacher2009} and cerebrospinal fluid~\cite{Weissenbacher2009,Dagli1999} (in these areas, one does not expect a response that is related to neuronal processes), (2) spontaneous global signal fluctuations~\cite{Weissenbacher2009,Birn2006,Fox2007}, and (3) signal mismatches between images from the head motion of subjects~\cite{Friston1996}. For each subject and time step, we calculate the mean signal for white-matter brain regions, the mean signal for regions that consist of cerebrospinal fluid, and the mean of the global signal. In addition to these mean values, we also include the squares and cubes of the global signal means, as well as head-motion parameters (3 translation and 3 rotation parameters), as rows in our $11 \times 120$ subject-specific design matrices. We then perform linear regression for each time series using {\sc Matlab}'s command for the Moore--Penrose pseudoinverse \footnote{Because some of the matrices are ill-conditioned, the resulting networks differ across different runs of the preprocessing. However, in our observations, the matrices differ by only up to $0.2\%$ of the entries between two separate realizations of preprocessing.} {\sc pinv()}; we exclude brain regions without grey matter from our calculations. We then use the residuals from the regression as our time series for the $120$ brain regions that we list in Tables \ref{S1_Table: BrainRegionsI}--\ref{S1_Table: BrainRegionsV}.
Such preprocessing steps are common when working with fMRI data, but they are not uncontroversial. In particular, the effects of global signal regression can alter correlations between time series. See, for example, \cite{Murphy2009,Fox2009} and~\cite{Fornito2015} in the context of schizophrenia.
\subsection{Functional connectivity}\label{construct}
We construct functional networks from the fMRI time series for each subject by using the 120 distinct brain regions (see Tables \ref{S1_Table: BrainRegionsI}--\ref{S1_Table: BrainRegionsV}) as the nodes of the networks and calculating Pearson correlations \footnote{There are numerous ways to measure functional connectivity \cite{Smith2011,Zhou2009,Bullmore2011}. For a discussion in the context of schizophrenia research, see \cite{Fornito2012}.} (without a time lag) between the nodes' time series as a measure of pairwise functional connectivity. The values of the pairwise functional connectivity give the edge weights between the brain regions in the functional networks.
In our key computations in the present paper, we consider four contiguous time regimes of 30 time points each; this yields four functional networks per subject. (The lone exception to this approach is Section~\ref{S1_3}, in which we use each subject's full time series, which consists of 120 time points, to construct a single functional network for each subject.) Although the four time regimes {each overlap temporally with times during which subjects performed one $0$-back and one $2$-back task}, our separation into time regimes is motivated by an interest in potential developments in the dynamics over time, rather than in relating the fMRI response to the task. We represent each functional network using an adjacency matrix $A = A(\text{subject}, \text{time regime})$, whose entry $A_{ij}$ is given by the edge weight between node $i$ and node $j$. Due to the high computational cost of PH~\cite{otter2017}, we reduce the number of edges in the networks that we analyze to enable computations. We apply a statistical threshold, described in \cite{Bassett2011}, to the weighted adjacency matrices without modifying the remaining edge weights. To obtain the thresholded adjacency matrices, we estimate p-values for the correlations using the {\sc Matlab} function {\sc corrcoef} and retain only those entries whose p-value is less than $0.05$. Using this type of thresholding, we retain at most 44\% of the edges in a network and retain {a mean} of 20--30\% of the edges {in each subject group}. We then separate each adjacency matrix into a positive and a negative part, $A = A^+ + A^-$, and study only the positive $A^+ $ part of the adjacency matrix \footnote{We choose to only include positive edge weights to avoid the need to interpret negative correlations between time series.}. In discarding negative edge weights, we retain a mean of slightly more than 50\% of the entries in our thresholded adjacency matrices. In Fig.~\ref{fig:TimeSeries}, we show a diagram of the steps that we perform to construct our functional networks. Although we consider the four time regimes separately, we treat all subjects and all time regimes together as one data set.
\begin{figure}[h!]
\centering\includegraphics[width=\textwidth]{TimeSeriesDiagramThesis.pdf}
\caption{Steps that we perform on the preprocessed time series of each brain region
to construct a functional network for each subject during each of four time regimes. We study the positive parts of the resulting networks using persistent homology.
}\label{fig:TimeSeries}
\end{figure}
\subsection{Persistent homology}
\emph{Persistent homology} (PH) is a technique from topological data analysis, which aims to understand the `shape' of data \cite{otter2017}. PH is based on the topological concept of \emph{homology}, which is used to study the shape of objects in a way that disregards changes from stretching and bending.
We motivate our use of PH for brain networks by considering different types of cheese and how they differ in their homology.
Calculating homology allows one to differentiate between the shape of a stereotypical Swiss cheese (of the Emmental sort) with holes and the shape of a mozzarella cheese by providing information about the presence or absence of holes in the cheeses. (See Fig.~\ref{fig:Cheeses1} for examples of the aforementioned cheeses.) One can thereby consider the space that surrounds the holes; these are the so-called \emph{loops}. However, homology does not give information about the geometry of the cheeses; for example, it does not `see' that the Swiss cheese is a cube or that the mozzarella cheese is a sphere (unless it happens to be hollow), as it only detects differences in the number of holes.
We now give a brief intuitive introduction to a few concepts behind homology and PH for network data. For more mathematical introductions, see~\cite{Ghrist2008,Edelsbrunner2008,Edelsbrunner2010,otter2017,Stolz2017,Croom}.
\begin{figure}[h!]
\centering
\subcaptionbox{Swiss (Emmental) Cheese}{\includegraphics[width=0.49\textwidth]{Emmentaler.jpg}}%
\hspace{0.01\textwidth}
\subcaptionbox{Mozzarella Cheese}{\includegraphics[width=0.312\textwidth]{Mozzarella.jpg}}%
\caption{An example of two topologically-different objects. Homology detects the topological differences by counting the number of holes in the cheeses.}\label{fig:Cheeses1}
\end{figure}
\subsubsection{Simplicial complexes}
To study the characteristics of topological spaces~\cite{Kosniowski1980}, such as the Swiss cheese or the mozzarella cheese, we consider small pieces (`morsels'), on which we can perform computations more easily. When reassembled, the morsels carry the same overall topological information as the original space. We begin building these morsels (i.e., `spaces', to be more formal) using a discrete set of points, which we call `nodes'. We then add `edges' to connect pairs of nodes; `triangles', which consist of three nodes, three edges, and a face; `tetrahedra'; and so on. Formally, these elements are called $k$-simplices, where $k$ indicates the dimension of the simplex. A point is a $0$-simplex, an edge is a $1$-simplex, a triangle is a $2$-simplex, and a tetrahedron is a $3$-simplex.
We can combine different simplices to capture different aspects of a topological space. For example, to capture the holes in the Emmental cheese, we glue together a collection of triangles and edges around the holes; we enclose the same number of holes as in the original cheese. Note that we can only
capture the holes that are enclosed inside the cheese (using the triangles), as one can deform the visible
holes on the surface into a smooth surface of the cheese. For demonstrative purposes, we therefore assume that the Emmental cheese in Fig.~\ref{fig:Cheeses1} is a cross section of a larger cheese that encloses the holes that are visible in the image.
One can combine simplices to obtain a \emph{simplicial complex} $\Sigma$, and we take the \emph{dimension} of $\Sigma$ to be the dimension of its highest-dimensional simplex. We show examples of simplicial complexes in Fig.~\ref{fig:Cheeses2}, where we again note that we are assuming that the Emmental cheese is only a cross section of a larger hunk of cheese.
\begin{figure}[h!]
\centering
\subcaptionbox{Swiss (Emmental) Cheese}{\includegraphics[width=0.49\textwidth]{Emmentaler2.pdf}}%
\hspace{0.01\textwidth}
\subcaptionbox{Mozzarella Cheese}{\includegraphics[width=0.312\textwidth]{Mozzarella2.pdf}}%
\caption{Simplicial complexes approximate topological spaces and capture their properties. (In the cross section of the Emmental cheese, all holes are topologically the same in the original object. However, we only visualize parts of the simplicial complex in which one of the holes is tiled with triangles but the others are merely surrounded by edges.)
}\label{fig:Cheeses2}
\end{figure}
\subsubsection{Homology and Betti numbers}
Homology assigns a family of vector spaces (called \emph{homology groups} in more general settings) to a simplicial complex. For a given dimension, the vector spaces capture the topological features in that dimension. For example, for dimension $0$, homology gives a vector space whose elements are connected components; for dimension $1$, homology gives a vector space that has loops as its elements. The dimensions of these vector spaces are called \emph{Betti numbers}, where $\beta_D$ denotes the Betti number for dimension $D$. The first three Betti numbers ($\beta_0$, $\beta_1$, and $\beta_2$) count, respectively, the number of connected components, the number of $1$-dimensional holes (i.e., loops), and the number of $2$-dimensional holes (as found in the Emmental cheese) in a simplicial complex.
\subsubsection{Weight rank clique filtration (WRCF)}
Similarly to being able to distinguish between two types of cheese, we are interested in whether we can use homology
(and specifically persistent homology) to distinguish between functional networks of schizophrenia patients, siblings of schizophrenia patients, and healthy controls.
In a network, we take a loop to consist of a sequence of four or more nodes and edges that begins and ends at the same node.
If two loops surround the same hole and can be deformed into one another in the space without tearing open either of the loops, then one counts the loops only once, and we construe them to be different \emph{representatives} (also called \emph{generators}) of a loop.
To obtain simplicial complexes from a weighted network, we construct a so-called \emph{filtration}. A filtration is a sequence of embedded simplicial complexes that starts with the empty complex:
\begin{equation*}
\emptyset = \Sigma_0 \subseteq \Sigma_1 \subseteq \Sigma_2 \subseteq \dots \subseteq \Sigma_\text{max} = \Sigma\,.
\end{equation*}
One can obtain a filtration from data in various ways \cite{feng2019}. When given data in the form of a weighted network, the easiest method is to filter by weights \cite{Lee2012}. In the first filtration step, one includes all nodes and the edge(s) with the largest weight in the simplicial complex. In the second step of the filtration, one adds the edge(s) with the second-largest weight to the simplicial complex from step one, and so on. In this way, one obtains a sequence of embedded simplicial complexes that satisfies the properties of a filtration. To construct a \emph{weight rank clique filtration} (WRCF)~\cite{Petri2013}, one performs one additional step: Whenever three edges in a simplicial complex of a filtration form a triangle, one fills in the associated face and one interprets the triangle as a $2$-simplex. Similarly, when four nodes are all connected pairwise by edges, the nodes form a (filled) tetrahedron (i.e., a $3$-simplex). We use the WRCF to analyze our weighted networks. The WRCF has been applied to weighted neuronal networks in several previous studies, including \cite{Petri2013, Petri2014, Giusti2015,Stolz2017}.
One can use homology to study topological features, such as loops, in every step of a filtration and determine the extent to which a feature persists with respect to the filtration \cite{otter2017}. We say that a topological feature $h$ in a given dimension is \emph{born} at filtration step $m$ if the homology group of $\Sigma_m$ is the first homology group of a simplicial complex in the filtration to include that feature. Similarly, we say that a topological feature \emph{dies} at filtration step $n$ if it is present in the homology group of $\Sigma_{n-1}$ but not in the homology group of $\Sigma_n$. The lifetime of a feature in a filtration is defined as the \emph{persistence} $p$. That is,
\begin{equation}
p = n - m \,.
\end{equation}
If a feature persists until the last filtration step, we say that it has \emph{infinite persistence}. Persistence was first used as a measure to rank topological features based on their lifetime in a filtration in \cite{Edelsbrunner2002}.
Ideally, one performs a WRCF on a fully connected functional network. However, because of the high computational cost, this is often impossible in practice. We avoid this issue by thresholding our weighted networks before analyzing them.
\subsection{Representations of persistent homology}\label{SubSec:PHRepresentation}
There are multiple ways to represent the output of persistent homology (PH) calculations and to visualize the persistence of topological features and their location within a filtration. The most common representations are barcodes and persistence diagrams. In recent years, a desire to leverage the output of PH computations for machine-learning and data-mining tasks has resulted in the development of alternative representations to both barcode and persistence diagrams \cite{otter2017}. Two of these alternative representations are persistence landscapes~\cite{Bubenik2015I,Bubenik2015} and persistence images~\cite{Adams2015}. In the following subsubsections, we describe barcodes, persistence diagrams, persistence landscapes, and persistence images.
\subsubsection{Barcodes} \label{bar}
A common representation of the output of PH calculations is a \emph{barcode}~\cite{Carlsson2005,Ghrist2008}. See Fig.~\ref{fig:HouseNeuro} for an example. A $D$-dimensional barcode is a plot of {a collection of} filtration parameter intervals $\{[\text{birth},\text{death})_l \}_{l = 1}^t$ that indicate the births and deaths of topological features of dimension $D$. The horizontal axis represents the filtration steps, and each $D$-dimensional topological feature in a filtration is represented by a bar that starts at the filtration step at which the feature is born and ends at the filtration step at which it dies. In a $0$-dimensional barcode, each bar corresponds to a connected component, and the length of a bar indicates how long a particular component is disconnected from other components in a simplicial complex. Similarly, in a $1$-dimensional barcode, each bar corresponds to a loop in a simplicial complex.
\begin{figure}[h!]
\centering\includegraphics[width=\textwidth]{test27.png}
\centering\includegraphics[width=\textwidth]{WRCFNeuroExample3-eps-converted-to.pdf}
\caption{Example of a weight rank clique filtration (WRCF) of a neuronal network and the corresponding (a) barcodes and (b) persistence diagrams (PDs) in dimensions $0$ and $1$. The neuronal network consists of different brain regions (indicated by circles), which we interpret as the nodes (indicated by dots) of a network, and weighted edges between the nodes. To construct the filtration, we add the nodes in step $0$, followed by the edge with the largest weight in step $1$, the edge with the second-largest weight in step $2$, and so on. As soon as three nodes are all connected pairwise by edges, we cover the resulting region with a triangle. When four nodes are all connected pairwise, we fill in a tetrahedron. In a $0$-dimensional barcode, we represent each connected component by a bar that starts when the component is born and ends when it dies (e.g., when two components combine with each other). In a $1$-dimensional barcode, each bar represents a loop, which consists of $4$ or more edges and starts and ends at the same node.
In a PD, one represents topological features by points, rather than by bars. The distance of a point to the diagonal (the purple line) indicates the persistence of the corresponding feature in the filtration.
}\label{fig:HouseNeuro}
\end{figure}
\subsubsection{Persistence diagrams} \label{perst}
As an alternative to a barcode, one can use a \emph{persistence diagram} (PD)~\cite{Cohen-Steiner2005}, which is a planar representation of a barcode that conveys the same information. One maps each $[\text{birth}, \text{death})$ interval in a barcode to birth--death coordinates, where the horizontal coordinate of a point represents the birth time of a feature in the associated filtration and its vertical coordinate represents the death time of that feature. Alternatively, one can use a birth--persistence coordinate system, which is particularly useful when examining persistence images (which we will discuss shortly). Points that are farther away from the diagonal identity line represent more-persistent topological features in a filtration. We show an example of a PD in Fig.~\ref{fig:HouseNeuro}. As with barcodes, one can treat PDs as mathematical objects, and one can endow the space of PDs with a distance.
\subsubsection{Persistence landscapes}
A \emph{persistence landscape} (PL)~\cite{Bubenik2015I,Bubenik2015}
is a sequence of piecewise-linear functions that one can use to visualize and analyze the information in a barcode or PD. Instead of using a bar and its length to represent a feature and its persistence, one now interprets each topological feature as a peak, whose height is determined by the feature's persistence and whose location corresponds to the feature's location in the filtration. In contrast to a barcode or a PD, a PL has three dimensions. As in a barcode, the horizontal axis represents the filtration step. The other two dimensions of a PL are the persistence of a feature and the different layers of the PL.
To create a PL from a barcode, one first defines {a peak function} for each bar. For a given $[\text{birth}, \text{death})$ interval in a barcode, one constructs the function
\begin{equation}
f_{[\text{birth},\text{death})} (x) = \begin{cases} 0 \,, & \text{if } x \notin (\text{birth},\text{death}) \\
x - \text{birth} \,, &\text{if } x \in \left(\text{birth},\frac{\text{birth}\, + \,\text{death}}{2}\right] \\
-x + \text{death} \,, &\text{if } x \in \left(\frac{\text{birth}\, + \, \text{death}}{2},\text{death} \right) \,.
\end{cases}
\end{equation}
One then collapses the collection of peak functions onto the horizontal axis of the barcode.
For a barcode that consists of the collection $\{[\text{birth},\text{death})_l \}_{l = 1}^t$ of intervals, the $q$th layer (with $q \geq 0$) of the PL (i.e., the \emph{$q$}th \emph{PL}) is the following set of functions:
\begin{align}
& \lambda_q: \ensuremath{\mathbb{R}} \rightarrow \ensuremath{\mathbb{R}}\,, \\
& \lambda_q(x) = q \text{th-largest value of } \left\{f_{[\text{birth},\text{death})_l}(x) \right\}_{l = 1}^t \,. \nonumber
\end{align}
If the $q$th-largest value does not exist, $\lambda_q(x) = 0$. The $0$th layer of a PL consists of the maximum function values among the collection of functions that one evaluates across a filtration.
Similarly, the $1$st layer of a PL consists of the second-largest values of the collection of functions that one evaluates across a filtration. One defines other layers in an analogous way. The \emph{persistence landscape} $\lambda$ of a barcode $\{[\text{birth},\text{death})_l \}_{l = 1}^t$ is defined as the sequence $\{ \lambda_q \}$ of the functions $\lambda_q$. We illustrate the pipeline from a barcode to a PL in Fig.~\ref{Fig: Landscapes}.
\begin{figure}[h!]
\centering\includegraphics[width=\textwidth]{WRCFExampleBarcodeFull3-eps-converted-to.pdf}
\caption{Schematic illustration of the steps for converting a barcode into a persistence landscape (PL). We use an example based on a weight rank clique filtration (WRCF) in dimension $1$. (a) Example barcode. (b) One defines peak functions on the bars of a barcode. (c) One collapses the images of the peak functions onto the horizontal axis. (d) The PL consists of the collection of layers $q$ (with $q$ = $0$, $q = 1$, and $q = 2$ in this example), which indicate the $q$th-largest values of the collection of peak-function values. To visualize the third dimension, we show the different layers using different colors.
(This figure is a modified version of a figure in~\cite{Stolz2017}.)
}\label{Fig: Landscapes}
\end{figure}
An advantage of PLs is that one can construct a mean PL for a set of landscapes. A mean landscape no longer corresponds to a barcode or a PD. However, one can define pairwise distances between two or more mean landscapes and use them to quantify the difference between two sets of barcodes. We use the $L_2$ distance. One can also use a variety of statistical tools on PLs~\cite{Bubenik2015I}. Such calculations have been used for applications like conformational changes in protein binding sites~\cite{Kovacev2015}, the origin of seizures in electroencephalographic (EEG) data from epileptic patients~\cite{Wang2015}, phase separation in binary metal alloys~\cite{Dlotko2016II}, brain geometry in neurodegenerative diseases~\cite{Garg2017}, audio signals in music~\cite{Liu2016}, and motor learning in humans~\cite{Stolz2017}.
\subsubsection{Persistence images}
Another representation of topological features in PH calculations are \emph{persistence images} (PIs), which are based on PDs and take the form of real-valued vectors \footnote{We use the term `vectorization' for the production of such a vector from a PD. One PD produces one PI, which yields one vector after it is reshaped.} that one can use as an input to a variety of machine-learning approaches. The transformation from a PD to a PI is stable with respect to the $1$-Wasserstein distance and maintains a clear and interpretable connection to the original PD \cite{Adams2015}. For example, PIs have been used to classify different types of neurons~\cite{Kanari2019,Kanari2018}.
We show a schematic of the mapping from a PD to a PI in Fig.~\ref{fig:schematic}, which depicts the various stages that are involved in the transformation. Recall that the output of a PH computation is a set of points (or intervals) corresponding to the birth and death times of each topological feature for a specified homological dimension. In Fig.~\ref{fig:schematic}(b), we show a sample PD \footnote{This example is the output of running a WRCF for dimension $1$ on the functional network of the first sibling from the first time regime in our data set.} for the sample point cloud from Fig.~\ref{fig:schematic}(a). In Fig.~\ref{fig:schematic}(c), we show the PD in the birth--persistence coordinate system. In Fig.~\ref{fig:schematic}(d), we show an overlay of the surface that is generated by centering a two-dimensional (2D) Gaussian on each point in the rotated PD in Fig.~\ref{fig:schematic}(c). Finally, in Fig.~\ref{fig:schematic}(e), we show an example PI that is produced by computing the volume under the surface in Fig.~\ref{fig:schematic}(d) over a uniformly-spaced grid. (We set the resolution so that we have a $20\times20$ grid of elements.) One can then reshape this final PI into a vector by stacking the columns (or, equivalently, the rows), as is often done in image processing. As described in \cite{Adams2015}, the generation of a PI involves the choice of (1) a 2D probability density function to center at each point in the birth--persistence PD, (2) a resolution, and (3) a weighting function. The role of the weighting function is, when necessary, to
suppress points in a PD that lie very close to the diagonal and are often construed as `noisy' features. For all of the PIs that we examine in the present paper, we use the default settings for the code: 2D Gaussian probability density functions, a linear weighting function, and a $50\times50$ grid of elements. We choose additional parameters that are associated with these choices (e.g., the variance of the Gaussians) according to the defaults in \cite{PIcode}. When analyzing and comparing multiple PIs, there is an additional pair of values that one must choose based on the data; these are the maximum birth and persistence values. We will see in Section~\ref{subsec: PIs} that this choice can influence our results.
\begin{figure}
\begin{centering}
\includegraphics[width=\textwidth]{PersistenceImagePipeline.pdf}
\caption{Schematic that illustrates the primary steps for converting a persistence diagram (PD) to a persistence image (PI). (a) Sample point cloud in $\mathbb{R}^{3}$.
(b) PD in birth--death coordinates (i.e., the standard choice), with the diagonal identity line shown. (c) PD in birth--persistence coordinates. (d) The persistence surface generated by centering 2D Gaussian distributions at each point in panel (c). (e) One generates a PI by summing the volume under the 2D Gaussian distributions over the area of a pixel (i.e., the area of a square) in a uniformly-spaced grid overlay.
}
\label{fig:schematic}
\end{centering}
\end{figure}
\subsubsection{Software employed}
For our PH calculations, we implement {\sc Matlab} code {that we construct} using {\sc javaPlex} \cite{javaPlex}, a software package for PH \footnote{For an overview of available PH software and additional references, see \cite{otter2017}.}. For a given filtration of a simplicial complex, \textit{{\sc javaPlex}} can output $[\text{birth}, \text{death})$ barcode intervals, representatives for each topological feature, and PDs. It outputs PDs in standard birth--death coordinates, from which one computes birth--persistence coordinates as $(\text{birth},\, \text{death}-\text{birth})$. For the WRCF, we also use a maximal clique-finding algorithm (that is based on the Bron--Kerbosch algorithm \cite{bron1973}) from the Mathworks library \cite{Wildmann2011}. For the analysis and interpretation of our barcodes, we use the {\sc Persistence Landscapes Toolbox}~\cite{Bubenik2015}. We create PIs using the code at~\cite{PIcode} with the default parameters.
\subsection{Clustering methods from data mining and network analysis}
Given the output of PH calculations, it can be insightful to use clustering methods to compare the PHs of different networks. There are myriad ways to proceed. In the present paper, we use a few different approaches. First, we apply the $k$-means clustering algorithm and community detection to examine whether we can separate the three subject groups based on the topological features of their functional networks. Second, we apply a linear sparse support vector machine (SSVM) to identify pixels in PIs to discriminate between the subject groups and examine which brain regions are generators of loops that help discriminate between groups. We describe these techniques in the following subsubsections.
\subsubsection{Employing $k$-means clustering for subject-group separation}
The method of $k$-means clustering aims to produce a partition of a metric space into $k$ clusters of points \cite{siamcluster}. Suppose that there are $\mu$ data points in a metric space. One selects $k$ of the $\mu$ points as `centers' and assigns all other points of a data set into clusters based on their closest center point. The `score' of such a clustering is the sum of the distances from each point to its nearest center. The desired output of $k$-means clustering is an assignment of points to clusters with the minimum clustering score. However, an exhaustive search for a global minimum is often prohibitively expensive.
A typical approach to search for a global minimum is to choose a large selection of $k$ initial centers uniformly at random, iteratively improve each selection of centers until the clustering score stabilizes, and then return the identified final clustering with the lowest score for each initialization. One iteratively updates the centers by setting the new center to be the mean of the points that are assigned to the center in the current iteration. One can apply $k$-means clustering either to a distance matrix (which one can calculate for either PDs or PLs) or to a set of input vectors (such as those that one obtains from a PI).
\subsubsection{Community detection for persistence-landscape classification}\label{Sec: Community Detection}
Community detection is a method from network analysis that attempts to partition a network into sets (called `communities') of nodes that are more densely connected to themselves than to other sets of networks~\cite{Newman2018,Porter2009,fortunato2016}. One can detect communities in either weighted or unweighted networks. In a weighted network, one finds larger total edge weights within communities than between them.
One can also use community detection to partition data (e.g., for classification) by studying a given distance matrix of data objects (such as mean PLs). We interpret the $n$ PLs as $n$ nodes of a network and convert the pairwise distances between them into edge weights, where a large edge weight signifies closeness in the distance matrix and a small edge weight signifies a long distance between two landscapes. We convert the distance $d(i,j)$ between landscapes $i$ and $j$ into an edge weight $A_{ij}$ between nodes $i$ and $j$ with the following formula:
\begin{equation}
A_{ij} = 1.01 - \frac{d(i,j)}{\max_{i,j \in\{1, \ldots, n\}}\{ d(i,j)\}}\,.
\end{equation}
This yields an adjacency matrix $A$ with elements $A_{ij}$. Naturally, there are many choices for converting from pairwise distances to pairwise weights, and one has to be careful about how that influences community structure and other network computations.
There are numerous methods that one can use for community detection in networks \cite{fortunato2016}. One approach for decomposing a network into communities (i.e., for performing a `hard partitioning') is to seek a partition that maximizes an objective function $Q$. The objective function that we use is modularity
\begin{equation}\label{eq:QF}
Q = \sum_{i,j}[A_{ij}-\gamma P_{ij}] \delta(g_i,g_j)\,,
\end{equation}
where $P$ (with elements $P_{ij}$) is a null-model matrix (which specifies the expected edge weight between nodes $i$ and $j$), the resolution parameter $\gamma$ is a factor that determines how much weight one gives to the null model, and $\delta(g_i,g_j) = 1$ if nodes $i$ and $j$ are in the same community (i.e., if $g_i = g_j$) and $\delta(g_i,g_j) = 0$ otherwise \cite{Porter2009,fortunato2016}.
For our computations, we use the {\sc GenLouvain} package~\cite{MuchaComm,Mucha2010}, which maximizes $Q$ using a variant of the Louvain algorithm \cite{blondel2008} to algorithmically detect communities in our mean PLs. We vary the weighting factor $\gamma$ (which is often called a `resolution parameter') to compare results for different values of $\gamma$.
\subsubsection{Linear sparse support vector machines for discriminatory feature selection}\label{ssvm}
The $1$-norm, regularized, linear support vector machine (i.e., SSVM) classifies data by generating a separating hyperplane between data points that depends on very few input-space features \cite{Bradley1998,Zhu2003,Zhang2010b}.
A hyperplane is a flat surface that cuts an ambient space into two parts. One can use an SSVM to identify discriminatory features between different groups of data points. One implements linear SSVM feature selection on data points in the form of vectors, so we can use it on our PIs to select `distinguishing pixels' during classification. In a PI, a \emph{distinguishing pixel} is a bounded region in the birth--persistence coordinate system. For
clarity, we use the term `distinguishing pixel' to signify a region
that is selected by SSVM and a `feature' to refer to a topological feature from a PH computation.
During the analysis of our results (see Section~\ref{subsec: PIs}),
we aim to match distinguishing pixels to their corresponding features.
We apply a `one-against-all' (OAA) SSVM to dimension-1 PIs from each subject to identify pixels in PIs that can discriminate between the subject groups.
In a one-against-all SSVM, there is one binary SSVM for each class to separate members of that class from the members of all other classes. In our case, this amounts to defining three hyperplanes: one that separates patients from controls and siblings, one that separates siblings from patients and controls, and one that separates controls from patients and siblings. We use a 5-fold cross-validated SSVM. We specify an optimal separating hyperplane by a normal vector, and we use the term `SSVM weights' for the values of the components of the normal vector. We select distinguishing pixels for each classifier by retaining the vector components (which are pixels in this application) with nonzero SSVM weights, ordering the nonzero SSVM weights by decreasing magnitude, and discarding SSVM weights when the ratio of successive SSVM weights drops below a user-specified tolerance. For details, see~\cite{chepushtanova2014}.
Given a set of distinguishing pixels, we can see for each subject whether the associated functional network have any loops that are born and persist in the corresponding bounded PI region. If there are loops in this region, we can identify a set of brain regions that are representative of that loop in the network. We are thereby able to leverage PIs to obtain (biologically) interpretable information about the involvement of different brain regions in the task (as measured with fMRI) for different subject types.
\section{Results}\label{Sec:Results}
We now present the results of our PH computations to examine loops in functional brain networks. We focus exclusively on topological features in dimension $1$. Additionally, we perform our computations on all four time regimes as part of one data set, rather than separating the data for each time regime. We run our PH computations on four functional networks per subject. From the PH output, we create either PLs or PIs. We then perform our computations either on (i) the full data set of PLs or PIs of 281 subjects and four time regimes (which gives 1124 landscapes or images, respectively, for the data set) or on (ii) the 12 subject-group means of the landscapes or images (from three subject groups with four time regimes each). We indicate which case we are examining in the relevant subsections.
For both PLs and PIs, we find that there seem to be differences in the topological features of the functional networks between subject groups, although we only observe these for PLs when examining means across groups. To illustrate limitations of the methods, we also discuss results in which we were unable to find differences between subject groups.
\subsection{Results of $k$-means clustering on PLs}
By applying $k$-means clustering to the mean PLs, we are able to separate siblings of schizophrenia patients from
controls and patients. For these calculations, recall that we use all four time regimes in each of the 12 mean landscapes.
We construct mean PLs from the 1D barcodes (i.e., the barcodes that represent loops in the networks) for each time regime and each subject group. We obtain 12 mean landscapes and exclude infinitely-persisting bars, because all of our landscapes include persistent infinite features and these tend to dominate the first several layers of the landscapes. Other researchers have excluded layers of landscapes (e.g., the first twenty) to filter out `topological noise' \cite{patrangenaru}. Although we threshold our weighted networks prior to analyzing them, this does not necessarily imply that we lose significant information by disregarding the persistent infinite features. Additionally, such features do not necessarily correspond to the most-persistent features in barcodes, as even features that are born in the last filtration steps are infinitely persisting if they do not die during the filtration. In our case, the presence of persistent infinite features prevented us from discriminating between landscapes based on their pairwise landscape distances. When we considered {the number of infinitely persisting} features without the other features, we did not observe any noticeable differences between the three subject groups.
We calculate a pairwise $L_2$ distance matrix of the mean landscapes, and we then perform $k$-means clustering on the distance matrix (which has $12\times 12$ entries). For $k = 3$, we obtain the expected division of the mean landscapes into patients, controls, and siblings. Although the fact that one can separate the three cohorts based on fMRI data is not a new finding --- see, for example,~\cite{Lynall2010,Rubinov2013,Alexander-Bloch2012,Bassett2008,Liu2008,Fornito2012,Singh2016} for patients versus controls and ~\cite{Sepede2010} for patients versus siblings --- the novelty of our calculation is that $k$-means clustering successfully distinguishes between the three different cohorts based on topological information (in the form of loops) in the functional networks.
We also perform $k$-means clustering for $k = 2$. Surprisingly, we find that the patients and controls are grouped in one cluster for all time regimes, whereas the siblings are in a separate cluster for all time regimes. We show the mean landscapes and clusters in Fig.~\ref{Fig:k-means Average Landscapes}.
\begin{figure}[h!]
\centering\includegraphics[width=\textwidth]{Schizophrenia2_corr2-eps-converted-to.pdf}
\caption{Mean PLs for each of the four time regimes and subject groups. Using $k$-means clustering with $k = 2$ on the set of 12 PLs (which consists of all subject-group means and time regimes as one data set) assigns patients and controls to one group. We show the mean PLs and their $k$-means-clustering grouping for the four time regimes separately.
}\label{Fig:k-means Average Landscapes}
\end{figure}
For $k \geq 4$, we do not observe a clear subject-group separation. To compare our results with ones from other clustering methods, we also apply average linkage clustering to the distance matrix and perform community detection on networks that we construct from those distance matrices (see Section~\ref{Sec: Community Detection}). We obtain the same qualitative result for these two methods as we did for $k$-means clustering. For community detection, we observe a clear separation for resolution-parameter values of $\gamma = 0.82, 0.83, \dots, 1.14$ into two communities (the siblings versus the patients and controls). These results appear to support our prior observation that the sibling cohort is particularly distinct from the other two cohorts, as compared to any other pairwise comparison among the three cohorts, with respect to their loop topology in the functional networks.
We also perform a permutation test on the mean PLs for each time regime {to determine the significance of the landscape distances, as suggested in~\cite{Bubenik2015}}. In this permutation test, we regroup the individual landscapes into three groups uniformly at random, create a new mean landscape for each newly assigned group, and calculate the pairwise $L_2$ distances between them. We then count how many of the $L_2$ distances of the new groups are larger than the ones that we observe when using the mean landscapes of the three subject groups. We use 10000 permutations to obtain our results, which we summarize in Table~\ref{Tab:pValuesLandscapes}.
\begin{table}[!ht]
\caption{Using a permutation test, we calculate p-values for the mean landscape distances between the three subject groups in each time regime.}\label{Tab:pValuesLandscapes}
\centering
\begin{tabular}{cccc}
\br
p-values for &Controls versus Patients & Controls versus Siblings & Patients versus Siblings \\ \mr
time regime 1 & 0.302 & 0.200 & 0.051 \\
time regime 2 & 0.460 & 0.009 & 0.052 \\
time regime 3 & 0.477 & 0.102 & 0.270\\
time regime 4 & 0.736 & 0.110 & 0.229\\ \br
\end{tabular}
\end{table}
Interestingly, for time regimes $1$ and $2$, we find {almost} significant distances (i.e., the p-values are slightly larger than $0.05$) between the patient and sibling mean landscapes, whereas the p-values for time regime $3$ and $4$ suggest that the distance is not significant (even though the p-values are {small in comparison to the p-values that we observe for controls versus patients}). The distance between the mean landscapes of the controls and the siblings appears to be significant for time regime $2$, but this does not appear to be the case for the other time regimes, although the p-values are again much smaller than for the distances between the mean landscapes of the patients and controls. For the controls and the patients, there are many other divisions into two groups that lead to more extreme distances between the mean landscapes than what one obtains by simply assigning them to a control group and a patient group.
To see if we can further support our result from $k$-means clustering for $k = 2$, we artificially group the controls and patients into one group to create a mean landscape and again perform a permutation test to verify whether the distance between the mean landscapes for the two groups is significant. In Table~\ref{Tab:pValuesLandscapes2}, we show the p-values that we obtain with 10000 permutations.
\begin{table}[!ht]
\caption{Using a permutation test, we calculate p-values for the controls-and-patients mean landscape versus the siblings mean landscape.}\label{Tab:pValuesLandscapes2}
\centering
\begin{tabular}{cccc}
\br
Time regime 1 &Time regime 2 & Time regime 3 & Time regime 4 \\ \mr
0.112 & 0.008 & 0.092 & 0.110 \\ \br
\end{tabular}
\end{table}
For time regime 2, we obtain a significant distance, but the p-values for time regimes 1, 3, and 4 are about $0.1$. Given the artificial grouping of the two subject groups, we interpret these values as small, although they are not statistically significant.
\subsection{Results of community detection using a distance matrix from individual PLs}
We construct PLs from each of the 1D barcodes, which we calculate by examining each subject in each of the four time regimes, and we calculate the $L_2$ distance matrix for the resulting 1124 PLs. We again use the distance matrix to construct a network between the PLs, and we detect communities in this network by maximizing modularity. For $\gamma = 0.92, 0.93, \dots, 1$, we obtain a separation into two communities. The partition that is closest to what we observe with $2$-means clustering for the mean landscape distance occurs for the resolution-parameter value $\gamma = 0.93$. We summarize our results in Table~\ref{Tab:CommunityClusters}.
\begin{table}[!ht]
\caption{Number of subjects from each subject group that are assigned to communities $1$ and $2$ by community detection using modularity maximization.}\label{Tab:CommunityClusters}
\centering
\begin{tabular}{ccc}
\br
Subject group & Number of subjects in community $1$ & Number of subjects in community $2$ \\ \mr
Patients & 122 & 94 \\
Controls & 418 & 290 \\
Siblings & 93 & 107\\ \br
\end{tabular}
\end{table}
We also apply $k$-means clustering and average linkage clustering to the distance matrix from the individual PLs (results not shown).
Of all of the classification methods that we perform on these distance matrices, community detection appears to perform best at `separating' the subject groups, although we do not observe a very clear separation.
\subsection{Results from our analysis of PIs}\label{subsec: PIs}
We find that PIs can identify discriminatory topological features across the three subject groups that we considered.
We generate PIs for each of the subjects for each of the four time regimes for the 1D PDs. We set the resolution, probability density function, and weighting function to the defaults in the PI code at \cite{PIcode}.
There is an additional pair of values --- the maximum birth and maximum persistence values --- that one must choose from the data that one is analyzing using PIs. These values determine the discretization of the pixel boundaries in the PIs once one sets the resolution. Possibilities include taking the maximum birth and persistence values across all PDs or normalizing each PD relative to its individual maximum. In the original paper on PIs~\cite{Adams2015} took maximum values across all PDs under consideration, although no theoretical rationale was provided for this choice. We were unable to obtain clear results using either of the above two conventions. For example, in the left image of Fig.~\ref{fig:PIstatsSame}, we see the mean vectorized PI for each subject group when we generate the PIs using the maximum birth time and maximum persistence across all subjects. (We create the mean vectorized PI for each subject group by taking the mean of each vector entry. {We take the mean across all four time regimes to ensure that we have enough data to make a meaningful comparison.}) Observe that the means look very similar, aside from a slight variation in their amplitudes. For each group, the mean PI is the mean of the vectorized PIs for all of the individuals in the group. The top row of Fig.~\ref{fig:PIstatsImages} has the mean PIs in image form when one selects the maximum birth and persistence across all subjects.
\begin{figure}[h!]
\centering
\subcaptionbox{}{\includegraphics[width=0.49\textwidth]{same_max_means2.pdf}}%
\hspace{0.01\textwidth}
\subcaptionbox{}{\includegraphics[width=0.49\textwidth]{separate_max_averages_all_time_regimes2.pdf}}%
\caption{(a) The mean vectorized PI for each subject group that we generate using the maximum values of birth and persistence across all subjects to create the PIs. We then take the means over the PIs of each group. The horizontal axis corresponds to individual pixels in the PIs, and the vertical axis indicates their intensity values. (b) Mean vectorized PI for each subject group that we generate using maximum values of birth and persistence that we determine by calculating the maximum birth and persistence for each of the three groups separately and using this group-specific information to create the PIs for each subject within its group. We then take the means over the PIs of each group.
}\label{fig:PIstatsSame}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[width=.6\textwidth]{averageImages.png}
\caption{The mean PI for each subject group. We generate panels (a)--(c) using the maximum values of birth and persistence across all subjects to compute all PIs before creating the depicted means over the PIs in each subject group. We generate panels (d)--(f) using the maximum values of birth and persistence using the maximum birth and persistence separately for each subject group to compute PIs within each group before creating the depicted means over the PIs in each group. The color axis is the same across rows. From left to right in each row, we show the mean PIs for siblings, controls, and patients.}
\label{fig:PIstatsImages}
\end{centering}
\end{figure}
Alternatively, if we use a priori knowledge of subject-group membership and fix the maximum birth values separately for each subject group (based on the collection of PDs that we compute separately for each subject group), we can discriminate between the three subject groups. This provides a first interesting observation from the PIs: there appears to be nontrivial information in the maximum birth time that corresponds --- or almost corresponds, in exceptional cases in which multiple edges have exactly the same weight --- to the number of pairs of regions in the brain with positive functional connectivity. (Recall that we do not add edges that correspond to negative Pearson correlations.) In Fig.~\ref{fig:PIstatsSame}(b), we show the mean vectorized PI for each subject group, where we set the maximum birth and maximum persistence values separately for each subject group (instead of setting the maximum birth and persistence values to be the same for all subjects). Observe that the sibling and control means both have two humps, whereas the patients have one hump that is clearly discernible.
Similarly, in Fig.~\ref{fig:PIstatsImages}, we observe two patches along the prominent diagonal with high intensity for the means of the siblings and the controls. However, in the bottom row, we only observe one clear (and elongated) hot spot for the mean of the patients. Therefore, there are multiple, smaller regions in which loops often occur in the filtrations of the functional networks of siblings and controls, whereas there is seemingly a single, larger region of loops in the filtrations of the networks of the patients.
It is also worth commenting on the locations of the local maxima for each subject type. Relative to the maximum values across each class, groupings of loops occur at different locations. From the values of the vectors, we see that the controls and patients have more similar maximum magnitudes than do the patients and their siblings. Therefore, we conclude that we are able to accurately separate the populations using PIs. Surprisingly, despite the pronounced difference in SSVM performance based on the PIs when we use different maximum values for each class, the distributions of the maximum birth times and maximum persistences for each subject type are not statistically-significantly different from each other. In Fig.~\ref{fig:dists}, we show Gaussian fits to the set of maximum birth times and maximum persistences for each subject type. Observe the strong similarity across all classes and the especially close similarity between the control and patient distributions. Because the maximum values are linked closely to the preprocessing of the data, it is important to conduct further research into how to account for these observations. One can also normalize PIs in other ways, such as by normalizing each PI individually by its own maximum value. However, we find that such individual normalization of PIs obscures information (specifically, the maximum birth time of each subject group) that appears to be relevant. The following results are based on the PIs that we generated using the maximum values that we determined by class membership.
\begin{figure}[h!]
\centering
\subcaptionbox{}{\includegraphics[width=0.49\textwidth]{max_birth_dist2-eps-converted-to.pdf}}%
\hspace{0.01\textwidth}
\subcaptionbox{}{\includegraphics[width=0.49\textwidth]{max_persistence_dist2-eps-converted-to.pdf}}%
\caption{The distribution of (a) the maximum birth times across all samples for each subject type and (b) the maximum persistences across all samples for each subject type.
}\label{fig:dists}
\end{figure}
As we discussed in Section~\ref{ssvm}, it is possible to apply a linear SSVM to the set of PIs to identify distinguishing pixels to help us interpret our classification results. Using a one-against-all SSVM with 5-fold cross validation, we obtain a 100\% classification accuracy. In Fig.~\ref{fig:selectedfeatures}, we show the distinguishing pixels from each of the three binary classifiers. By taking the union of these pixels, we obtain $41$ distinct pixels from the total of $625$ pixels in the PIs. We again emphasize that each of these pixels corresponds to a bounded region in the birth--persistence plane.
\begin{figure}
\begin{centering}
\includegraphics[width=\textwidth]{pixels_chosen.png}
\caption{The set of 41 distinguishing pixels (of 625 in total) that we determine via SSVM to be critical for obtaining 100\% classification accuracy on the testing set of PIs.
}
\label{fig:selectedfeatures}
\end{centering}
\end{figure}
Interpreting these distinguishing pixels requires discussing their relationships with particular regions of the brain. We make these connections as follows. For each subject, it is possible to determine whether a topological feature (in our case, a loop) in a filtration of a network exists in the bounded region of the birth--persistence plane that corresponds to a particular distinguishing pixel. If a loop does exist, one can identify a set of brain regions (see Section~\ref{Sec:Discussion}) that comprise the loop (i.e., representatives of this loop). We are particularly interested in brain regions that are consistently involved in the generation of particular loops across subjects. We identify the set of nodes, which we call \emph{top node(s)} \footnote{One can interpret our calculation of top nodes in a similar spirit as calculations of node centralities \cite{Newman2018}.}, that are involved in the generation of loop(s) for each distinguishing pixel in each of the four time regimes for each subject. We then create histograms of the union of the nodes that we select in this fashion to examine the relative importances of top nodes across each subject type.
In Fig.~\ref{fig:MAX_nodes}, we give the relative importances of different brain regions for each pixel. In Fig.~\ref{fig:MAX_nodes}(a), we show the top nodes for each subject type based the proportion of the subject types for which that top node is involved in the generation of a loop in the distinguishing-pixel region. In Fig.~\ref{fig:MAX_nodes}(b), we show the proportion of the subjects for which the top node(s) is (are) present. The vertical gaps in each plot signify that there are no nodes that are consistently involved in loops for that distinguishing pixel. We make several observations from Fig.~\ref{fig:MAX_nodes}(a). First, there are only five distinguishing pixels for which we find top nodes for the patients. Therefore, we are unable to predict which brain regions are involved in loops in the functional networks during the given task for schizophrenia patients. By contrast, there are many distinguishing pixels for which we find top nodes for the siblings. The control group lies between the other two in terms of its number of distinguishing pixels with top nodes, but there are still few top nodes in comparison to the number of distinguishing pixels that have top nodes. In Tables \ref{table:features}--\ref{table:features4} (see also Figs.~\ref{fig:MAX_nodes_Sib}--\ref{fig:MAX_nodes_Pat} of~\ref{S1}), we indicate which brain regions (as well as their locations) we identify as top nodes. We include only the distinguishing pixels in which top nodes exist within a cohort.
An equivalent way to identify a top node is to calculate the percentage of a given subject class that has a topological feature in the corresponding pixel region (see the bar graph in Fig.~\ref{fig:MAX_nodes}) and determine if a specific node is in the group of representatives for all of the subjects that have a topological feature in the pixel region. We identify a node as a top node if it occurs in the list of representatives for a topological feature for every subject of the class with a topological feature in the pixel region. Therefore, when considering Tables \ref{table:features}--\ref{table:features4}, it is possible for the same brain regions to be listed for more than one distinguishing pixel index. This is also reflected in Fig.~\ref{fig:MAX_nodes} by the occurrence of multiple markers along the same horizontal line.
\begin{figure}[ht!]
\centering
\subcaptionbox{}{\includegraphics[width=0.49\textwidth]{max_node_figures_a.pdf}}%
\hspace{0.01\textwidth}
\subcaptionbox{}{\includegraphics[width=0.49\textwidth]{max_node_figures_b.pdf}}%
\caption{(a) The index (indices) of the top node(s) that are associated with each distinguishing pixel that we determine via SSVM. (b) A stacked bar graph of the proportion of each subject type that have the corresponding node(s). Pale green indicates siblings, greenish blue indicates controls, and black indicates patients.
}\label{fig:MAX_nodes}
\end{figure}
\begin{table}
\centering
\caption{{\bf Top nodes that are involved in loop representatives within the bounds of distinguishing pixel birth--persistence regions (part I).}
We include only distinguishing pixels for which there is (are) top node(s) within a cohort. `Left' and `Right' refer to the hemispheres of the brain. We use the following abbreviations: superior frontal gyrus medial segment (MSFG), superior temporal gyrus (STG), opercular part of the inferior frontal gyrus (OpIFG), transverse temporal gyrus (TTG), frontal operculum (FO), gyrus rectus (GRe), middle frontal gyrus (MFG), orbital part of the inferior frontal gyrus (OrIFG), precuneus (PCu), cuneus (CC), anterior insula (AIns), superior parietal lobule (SPL), lingual gyrus (LiG), cerebellum exterior (CE), parahippocampal gyrus (PHG), medial frontal cortex (MFC), medial orbital gyrus (MOrG), and posterior cingulate gyrus (PCgG).
}
\label{table:features}
\begin{tabular}{*7c}
\br
Pixel Indices & \multicolumn{2}{c}{Siblings} & \multicolumn{2}{c}{Controls} & \multicolumn{2}{c}{Patients}\\
\mr
{} & Node& Location & Node & Location & Node & Location\\ \mr
1 & 70 & Left MSFG & -- & -- & -- & -- \\
& 114 & Left STG &-- & -- & -- & -- \\
\mr
2 & 70 & Left MSFG & 77 & Left OpIFG & 40 & Left FO \\
& 114 & Left STG &120 & Left TTG & -- & -- \\
\mr
3 & 45 & Right GRe &-- & -- & -- & -- \\
& 119 & Right TTG & -- & -- & -- & -- \\
\mr
5 & 45 & Right GRe & 60 & Left MFG &80 & Left OrIFG \\
& 119 & Right TTG & -- & -- & -- & -- \\
\mr
6 & 84 & Left PCu & 60 &Left MFG & -- & -- \\
\mr
8 & 84 & Left PCu & 60 &Left MFG &35 & Right CC \\
\mr
9 & 120 & Left TTG & 60 &Left MFG & -- & -- \\
\mr
11 & 120 & Left TTG & 60 &Left MFG & 26 & Left AIns \\
\mr
12 & 45 &Right GRe & 60 &Left MFG & -- &-- \\
\br
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{{\bf Top nodes that are involved in loop representatives within the bounds of distinguishing pixel birth--persistence regions (part II).}
We include only distinguishing pixels for which there is (are) top node(s) within a cohort. `Left' and `Right' refer to the hemispheres of the brain. We use the following abbreviations: superior frontal gyrus medial segment (MSFG), superior temporal gyrus (STG), opercular part of the inferior frontal gyrus (OpIFG), transverse temporal gyrus (TTG), frontal operculum (FO), gyrus rectus (GRe), middle frontal gyrus (MFG), orbital part of the inferior frontal gyrus (OrIFG), precuneus (PCu), cuneus (CC), anterior insula (AIns), superior parietal lobule (SPL), lingual gyrus (LiG), cerebellum exterior (CE), parahippocampal gyrus (PHG), medial frontal cortex (MFC), medial orbital gyrus (MOrG), and posterior cingulate gyrus (PCgG).
}
\label{table:features2}
\begin{tabular}{*7c}
\br
Pixel Indices & \multicolumn{2}{c}{Siblings} & \multicolumn{2}{c}{Controls} & \multicolumn{2}{c}{Patients}\\
\mr
{} & Node& Location & Node & Location & Node & Location\\ \mr
14 & 84 & Left PCu & 112& Left SPL & -- & -- \\
& 112 & Left SPL & -- & -- & -- & -- \\
\mr
15 & 51 & Right LiG & -- & -- & -- & -- \\
\mr
17 & -- & -- &112 &Left SPL &26 & Left AIns \\
\mr
18 &51 & Right LiG & 69 &Right MSFG & -- &-- \\
\mr
19 & 3 & Right Amyg. & -- & -- & -- & -- \\
\mr
21 & 3 & Right Amyg. & 8 &Left CE & -- & -- \\
& -- & -- & 112&Left SPL &-- &-- \\ \mr
22 & 80 & Left OrIFG & -- & -- & -- & -- \\
& 112& Left SPL & -- & -- & -- & -- \\
\mr
24 & 80 & Left OrIFG & 59 &Right MFG & -- & -- \\
& 112 &Left SPL & -- & -- & -- & -- \\ \mr
25 & 112 &Left SPL & -- & -- & -- & -- \\
\br
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{{\bf Top nodes that are involved in loop representatives within the bounds of distinguishing pixel birth--persistence regions (part III).}
We include only distinguishing pixels for which there is (are) top node(s) within a cohort. `Left' and `Right' refer to the hemispheres of the brain. We use the following abbreviations: superior frontal gyrus medial segment (MSFG), superior temporal gyrus (STG), opercular part of the inferior frontal gyrus (OpIFG), transverse temporal gyrus (TTG), frontal operculum (FO), gyrus rectus (GRe), middle frontal gyrus (MFG), orbital part of the inferior frontal gyrus (OrIFG), precuneus (PCu), cuneus (CC), anterior insula (AIns), superior parietal lobule (SPL), lingual gyrus (LiG), cerebellum exterior (CE), parahippocampal gyrus (PHG), medial frontal cortex (MFC), medial orbital gyrus (MOrG), and posterior cingulate gyrus (PCgG).
}
\label{table:features3}
\begin{tabular}{*7c}
\br
Pixel Indices & \multicolumn{2}{c}{Siblings} & \multicolumn{2}{c}{Controls} & \multicolumn{2}{c}{Patients}\\
\midrule
{} & Node& Location & Node & Location & Node & Location\\ \mr
27 & 112 & Left SPL &60 & Left MFG & -- & -- \\
\mr
28 & 119 & Right TTG & -- & -- & -- & -- \\
\mr
30 &119 & Right TTG &59 & Right MFG & -- & -- \\ \mr
31 & 119 & Right TTG &59 & Right MFG &-- & -- \\
\mr
32 & 120 & Left TTG & -- & -- & -- & -- \\
\mr
33 & -- & -- & 59 & Right MFG & -- & -- \\
\mr
34 & 120& Left TTG & 8 &Left CE & -- & -- \\
& -- & -- & 85 &Right PHG & -- & -- \\
\mr
36 & -- & -- & 8 & Left CE & -- & -- \\
& -- & -- & 85 & Right PHG & -- & -- \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}
\centering
\caption{{\bf Top nodes that are involved in loop representatives within the bounds of distinguishing pixel birth--persistence regions (part IV).}
We include only distinguishing pixels for which there is (are) top node(s) within a cohort. `Left' and `Right' refer to the hemispheres of the brain. We use the following abbreviations: superior frontal gyrus medial segment (MSFG), superior temporal gyrus (STG), opercular part of the inferior frontal gyrus (OpIFG), transverse temporal gyrus (TTG), frontal operculum (FO), gyrus rectus (GRe), middle frontal gyrus (MFG), orbital part of the inferior frontal gyrus (OrIFG), precuneus (PCu), cuneus (CC), anterior insula (AIns), superior parietal lobule (SPL), lingual gyrus (LiG), cerebellum exterior (CE), parahippocampal gyrus (PHG), medial frontal cortex (MFC), medial orbital gyrus (MOrG), and posterior cingulate gyrus (PCgG).
}
\label{table:features4}
\begin{tabular}{*7c}
\br
Pixel Indices & \multicolumn{2}{c}{Siblings} & \multicolumn{2}{c}{Controls} & \multicolumn{2}{c}{Patients}\\
\mr
{} & Node & Location & Node & Location & Node & Location\\ \mr
37 & 57 & Right MFC & 59 & Right MFG & -- & -- \\ & 58 & Left MFC &-- & -- & -- & -- \\
& 119& Right TTG & -- & -- & -- & -- \\
\mr
39 & 111& Right SPL & 26 &Left AIns & -- & -- \\
& -- & -- & 52 &Left LiG & -- & -- \\
\mr
40 & 46 &Left GRe & -- & -- & -- & -- \\
& 57 & Right MFC & -- & -- & -- &-- \\
& 58 & Left MFC & -- & -- & -- & -- \\
& 64 & Left MOrG & -- & -- & -- & -- \\ \mr
41 & 82 & Left PCgG & -- & -- & -- & -- \\
\br
\end{tabular}
\end{table}
\section{Discussion}\label{Sec:Discussion}
We applied methods from persistent homology to analyze loops in functional brain networks of schizophrenia patients, siblings of schizophrenia patients, and healthy controls. We constructed both persistence landscapes and persistence images of these networks, and we compared them to each other using several clustering techniques.
We observed topological differences in the functional brain networks of schizophrenia patients, siblings of schizophrenia patients, and healthy controls with respect to the loops in their networks. We also found that PLs and PIs have different
practical advantages and disadvantages when applied to the same data set; these insights may be useful for interpreting the results of PH computations in networks in diverse applications.
Computing PLs gave interesting results when the comparing mean PLs of the cohorts but not when comparing individual landscapes of the subjects. Using mean PLs, we were able to separate the sibling cohort from the other two subject groups in each of the four time regimes. This is supported by the p-values that we obtained for the distances between the mean landscapes of the sibling cohort versus the controls and patient cohorts, although not all of our p-values are statistically significant. The shape of the mean PLs seems to suggest that loops that occur in the functional brain networks of siblings are more persistent on average than those in the functional networks of controls or patients. This may imply either that (1) loops in the networks of siblings tend to be longer or that (2) the third edge between three nodes has a small edge weight and thus that three brain regions with a large pairwise Pearson correlation between one region and two of the other regions do not necessarily imply that there is a large correlation between the other two brain regions; such a third edge facilitates the creation of a loop structure in the filtration. (Recall that we need at least four nodes for our loops.) To examine this issue further, it may be useful to analyze cross-links in the functional networks, as in \cite{Bassett2014II}. For the above computations and their interpretation, we need to take into account that we did not include infinitely-persisting loops (which persist until the end of a filtration). We also include only positive edge weights in our networks, so we only analyzed loops that arise from brain regions with positive pairwise Pearson correlations.
Although we were able to obtain interesting insights about the data using mean PLs, we did not find interpretable results from comparing individual landscapes, and only using the mean landscapes reduces the amount of information the we can obtain from this approach. By contrast, using individual PIs and SSVMs allowed us to separate the entire set of subjects (with $100\%$ accuracy) in each of the four time regimes. In previous work, Anderson and Cohen~\cite{Anderson2013} obtained 65\% accuracy for schizophrenia classification by applying machine-learning techniques to functional brain networks. It is important to note, however, that our results are based on using a priori knowledge of group membership (specifically, by including the maximum birth times of loops within subject groups). These birth times seem to include nontrivial information, which is important to pursue further in future studies. Moreover, such a priori knowledge is tied closely to the choice of statistical thresholding when preprocessing fMRI data. Consequently, developing a statistical model that can classify a novel subject based on a PI representation also requires further explorations of how to choose such a threshold.
Computing PIs also allowed us to identify brain regions with consistent involvement in loops in the functional networks within subject cohorts. Of the three cohorts, we found that siblings have the highest level of consistent brain-region involvement in the performance of the mental task in this study across the four time regimes. That is, brain regions that are involved in loops for siblings in one of the time regimes are more likely to also be involved in loops in other time regimes than is the case for patients or controls. It is particularly noteworthy that the number of brain regions that are consistently involved in the separation of the three cohorts is larger in the siblings of schizophrenia patients than in the healthy controls. We view heterogeneous involvement of brain regions in loops as a notion of neurological `flexibility'. Various works have studied concepts of brain flexibility using community structure~\cite{Bassett2011,Braun2016}. In those studies, flexibility was defined differently --- based on how often a brain region changes its allegiance to a community of nodes over time, so it does not use loops directly --- than in the above characterization, but it is noteworthy that Braun \emph{et al.}~\cite{Braun2016} observed that relatives of schizophrenia patients have larger flexibility than healthy controls. In our work, we found that a specific group of brain regions leads to the separation of the three subject groups when using PIs and observed for the schizophrenia patients that the regions that lead to a separation consistently in each of the four time regimes are fewer in number than for the siblings and the controls. Braun \emph{et al.}\cite{Braun2016} reported that there is larger node flexibility in network organization of schizophrenia patients than in healthy controls. Additionally, Siebenh\"uhner \emph{et al.}~\cite{Siebenhuhner2013} observed a greater variability in temporal networks that were constructed from magnetoencephalography (MEG) data of schizophrenia patients than those that were constructed from the data of healthy controls.
We observed four time regimes, which each consist of fMRI signals that were recorded during one block of a $0$-back task and a $2$-back task. For time regime 2, we obtained very small and statistically significant p-values in our mean PL computations when comparing siblings to controls and when we were comparing siblings to a group that consists of all patients and all controls. We did not observe this for any of the other time regimes. We conclude that time regime 2 appears to capture significant changes in the persistence and/or appearance of loops in the networks of siblings during the working-memory task. It will be useful to conduct further {laboratory} experiments to draw biological conclusions.
Schizophrenia has a high genetic determinism, so siblings of schizophrenia patients have a significant genetic risk of developing the disease themselves~\cite{Bertolino2009}, and it has been demonstrated that they have abnormalities in their structural neuronal networks~\cite{Collin2014}.
Although our results that functional brain networks that are constructed from fMRI measurements of siblings differ both from those of patients and those of healthy controls do not agree completely with the current scientific literature, other studies have also reported that the features of fMRIs of siblings of schizophrenia patients differ both from schizophrenia patients and from controls. For example, Callicott \emph{et al.}~\cite{Callicott2003} observed in an fMRI study that there was no difference in task performance between healthy siblings of schizophrenia patients and healthy controls, yet they detected a physiological similarity between the sibling cohort and the schizophrenia patients in the associated fMRI data. Similarly, Sepede \emph{et al.}~\cite{Sepede2010} observed using fMRI data from a different data set that healthy siblings of schizophrenia patients exhibit differences in brain function from schizophrenia patients, although they did not differ significantly in task performance. Additionally, it was demonstrated recently that schizophrenia patients undergo a cortical normalization process over the course of the disease~\cite{Guo2016}. However, one needs further phenotypic information to assess whether it is possible to directly connect the results of such a study to our observations.
Because our results are somewhat inconsistent with prior observations, it is also possible that our data set contains experimental noise that is beyond our control. Using standard network-analysis techniques on the functional brain networks, we did not observe any differences between the three subject groups. Nevertheless, we believe that our comparison of PLs to PIs and the different types of results from these techniques provide a valuable example of an approach that uses topological data analysis for functional brain networks. To give another cautionary note, one needs to take into account that there are difficulties when interpreting the information about node participation in loops from computations of PH, as the software that is used for such computations (including, specifically, {\sc javaPlex}, which is what we used) only finds representatives of the loops. These representatives are not determined in an optimal way, and they need not be `geometrically nice'~\cite{Adams2014}. For example, in these calculations, one often encounters double loops or even triple loops as generators for one loop in a functional network. Selecting a basis of homology generators that behaves in a biologically representative way corresponds mathematically to solving a problem known as the `optimal homology-basis problem', which is difficult (and is NP-hard in the worst case)~\cite{Erickson2000}. Despite these difficulties, our list of discriminating nodes provides a useful starting point for further investigations into neuronal abnormalities in functional networks of schizophrenia patients.
Another important issue is that we preprocessed the data for our study. This is very common when working with fMRI data, but such steps are not uncontroversial, and studies on functional connectivity in schizophrenia patients have found contradictory results depending on whether one performs global signal correction~\cite{Fox2009,Fornito2015}. It is also relevant to keep in mind that the choice of functional connectivity measure can influence results~\cite{Smith2011}. We used the Pearson correlation because of its simplicity and the fact that it is a widely used measure of functional connectivity~\cite{Wang2010,Bassett2012}. Many other choices are also available.
Finally, we chose to threshold our networks and removed edges whose weights were below a certain amount. However, it has been observed previously \cite{Bassett2012} that edges with small weights can be important when comparing functional brain networks of schizophrenia patients to those of healthy controls. However, such missing edges --- depending on their location and distribution --- can result in loops in a network, our analysis indirectly includes some of this information. In future work, it seems interesting to consider only the parts of the functions networks that were below the threshold value that we employed in our present study and analyze them with PH to compare classification results.
\section*{Acknowledgements}
\label{sec:Acknowledgements}
We thank Alessandro Bertolino, Fabio Sambataro, and the Bari psychiatric neuroscience group for permission to study their data. All rights of the data lie with their research groups, and we are unable to release the data. We also thank Pawel D\l otko for his help with the {\sc Persistence Landscapes Toolbox} and for providing us with new versions of the code during our project. We are also grateful to Danielle Bassett, Peter Bubenik, Carina Curto, Parker Edwards, and several referees for helpful comments. We thank Florian Lipsmeier and Franziska Mech from Roche for useful discussions. We also acknowledge Advanced Research Computing (ARC) at University of Oxford for resources that we used in carrying out this work. BJS thanks the EPSRC and MRC (grant number EP/G037280/1) and F. Hoffmann-La Roche AG for funding her doctoral studies. HAH acknowledges funding from an EPSRC Fellowship (EP/K041096/1) and a Royal Society University Research Fellowship. BJS and HAH are members of the Centre for Topological Data Analysis, which is funded by an EPSRC grant (EP/R018472/1).
\providecommand{\newblock}{}
|
\section{Introduction}
Alkali diatomic molecules have been widely studied by theoreticians and experimentalists for many years due to their simple and rich internal structures. Investigations of molecular structure and dynamics of short-range potentials and photodissociation, predissociation, autoionization, and energy transfer processes using traditional short-range molecular spectroscopy are important for ultracold physics. In recent years cold and ultracold (bi)alkali diatomic molecules have been at the forefront of quantum chemistry and many-body physics. These molecules have been used to probe new states of quantum matter, to improve precision measurements of fundamental constants, and in the development of quantum information storage and molecule lasers~\cite{Stwalley78,Stwalley99,Jones06,Carr09,Bahns00,McGuyer15,Balakrishnan16,Krems08}. While the radiative properties of ultracold molecules are no longer restricted by collisional de-excitation, preparing dense samples of ultracold diatomic molecules in a specific quantum state still remains challenging. Diatomic alkali molecules exhibit an exotic behaviour of the $6\,^1\Sigma_g^+$ symmetry with double well (inner and outer) structures due to the sodium positive and negative ion pair potential energy interaction. The existence of an outer-well ($8a_o<R<40a_o$) of this potential curve may play an important role for the production of cold ground state molecules. This work deals with the experimental and theoretical studies of the $6\,^1\Sigma_g^+$(3s+5s) state in sodium diatomic molecules. \\
The potential curves of electronic states up to the Na$_2$ (3s+5s) dissociation limit for the singlet and triplet electronic states and radiative lifetimes of a band of vibrational levels for various electronic states have been studied using various methods~\cite{Demtroder76,Ducas76,Baumgartner84,Radzewick83,Magnier93,tsai94-2,Sanli15,Anunciado16}. Recent experimental and theoretical studies of the radiative lifetimes of the $2\,^1\Sigma_u^+$ state of sodium dimers have been reported using a molecular beam apparatus and pulsed dye lasers with a pump-delayed probe technique by Anunciado~\textit{et. al.}~\cite{Anunciado16}. Authors demonstrated for the first time the importance of the inclusion of the bound-free transitions into the lifetime calculations. Despite considerable interest in the electronic states of sodium diatomic molecules, the radiative lifetimes of the $6\,^1\Sigma_g^+$ inner-well state have not been measured by experimentalists. The lifetimes of the outer-well of this state have been reported in multi-step laser excitation method~\cite{Laue03}. In this paper we report for the first time lifetime measurements of the inner-well of the $6\,^1\Sigma_g^+ (7,31)$ state using time-correlated photon counting technique and theoretical calculations in various ro-vibrational levels of the $6\,^1\Sigma_g^+$ state using bound-bound and bound-free transitions. We also demonstrate the effect of inclusion of the bound-free transitions in the calculated radiative lifetime of the $6\,^1\Sigma_g^+$ (3s+5s) ro-vibrational levels in the range of $v=0-200$ with $J=1$ and $J=31$. We report the transition dipole moment functions coupling the $6\,^1\Sigma_g^+$ state to seven singlet, ungerade states and also the branching ratios for radiative transitions from the $6\,^1\Sigma_g^+$ ro-vibrational levels into those seven states. We discuss the results of both experimental and theoretical lifetimes of the $6\,^1\Sigma_g^+$ (3s+5s) state.\\
The remainder of this paper is organized as follows: section~\ref{sec:Experiment} contains a description of the experimental apparatus including the excitation scheme and potential energy curves of Na$_2$, section~\ref{sec:Measurements} presents new measurements including the techniques used, section~\ref{sec:Calculations} describes the theoretical calculations, followed by results and discussions in section~\ref{sec:Results}, and conclusion in section~\ref{sec:Conclusion}.
\section{Experiment}\label{sec:Experiment}
The time-resolved double-resonance experimental setup is shown in Fig.~\ref{fig:Fig1}. The sodium metal is contained in a 4-arm crossed heat-pipe oven~\cite{Vidal69,Vidal71,Vidal72,Vidal96}. The desired amount of buffer gas was introduced into the heat-pipe when the oven was cold and the the background pressure was about 15 mTorr. The gas valve was closed to keep a fixed amount of argon gas inside the oven. The oven, which has a length of 1 m, is then heated by eight non-magnetic ceramic heaters to produce sodium vapor while cold water is run through copper coils wrapped near the windows of the arms. The heat-pipe oven and the ceramic heaters were wrapped with a ceramic fiber blanket to maintain a homogenous temperature inside the oven. A stainless steel mesh lining the arms of the oven walls acts as a wick to return the liquid metal back to the central region of the oven in order to prevent condensation onto the windows. During the experiment argon pressure was continuously monitored by an absolute capacitance manometer (MKS 627E), temperature-controlled to 45~$^\circ$C. To create sodium molecules the heat-pipe is heated to 290~$^\circ$C and the temperature of the central region of the oven was monitored with a thermocouple probe. At the operating temperature of 290~$^\circ$C, about 32.6\% of the molecules are in the vibrational ground state $v''=0$, 21.8\% in $v''=1$, 14.6\% in $v''=2$, 9.85\% in $v''=3$,..., 1.45\% in $v''=8$ and less than 1 in levels $v''=9$ and higher. The Nesmeyanov vapor pressure formula is used to estimate the sodium molecular and atomic vapor densities at the operating temperature of the heat-pipe oven~\cite{Nesmeyanov63}, yielding a number density at 290~$^\circ$C for the sodium dimers of $1.45$x$10^{12}$ cm$^{-3}$ and for the sodium atoms of $7.35$x$10^{14}$ cm$^{-3}$. \\
\begin{figure*}[th]
\includegraphics[scale=0.85]{Fig1.eps}
\caption{(Color online) A schematic view of the experimental setup. The tunable dye lasers are pumped by the pulse Nd:YAG laser. The slit widths and swing-away mirror of the spectrometer are computer-controlled. The CCD is a charge-couple device, cooled to -40~$^\circ$C, with 14x10$^{-6}$ m pixel size. MCS is a multi-channel scaler for time-correlated photon counting, PMT refers to a photomultiplier tube, pd a photodiode, and a periscope, consisting of two mirrors, is used to invert the horizontal image to vertical in order to optimize the fluorescence collection at the entrance slit of the Horiba spectrometer. }\label{fig:Fig1}
\end{figure*}
The partial energy level diagram of Na$_2$ and experimental excitation scheme are shown in Fig.~\ref{fig:Fig2}. The excitation lasers are tunable home-built pulse dye lasers whose cavity designs are of the grazing incidence Littman-Metcalf~\cite{Littman78} type. Both dye lasers operate in a single transverse mode and are pumped by the second harmonics (532 nm) of a Nd:YAG laser with a pulse repetition rate of 20 Hz. They produce about 0.5 mW average power with bandwidths less than 6 GHz ($\leq$ 0.2 cm$^{-1}$). Dye laser 1, L1, drives the $X\,^1\Sigma_g^+ (0,29)$ $\rightarrow$ $A\,^1\Sigma_u^+ (8,30)$ transition at 15518.5 cm$^{-1}$ (644 nm) while the second dye laser, L2, drives the $A\,^1\Sigma_u^+ (8,30)$ $\rightarrow$ $6\,^1\Sigma_g^+ (7,31)$ transition at 17798.9 cm$^{-1}$ (561.8 nm). Laser wavelengths are measured using a Coherent wavemeter with a precision of 0.01 cm$^{-1}$. The lasers, both collimated to a 1-cm beam diameter, collinearly counterpropagate in the interaction region of the heat-pipe oven. Pulse duration of the lasers, about 6 ns, was measured using an ultrafast phototube with 270 ps risetime and 100 ps fall time. The L2 pulse arrives at the interaction region of the oven 4 ns after the arrival time of the L1 pulse. Due to the 2 ns temporal overlap time compared to the radiative lifetime of the excited level~\cite{Baumgartner84}, about 23\% of the excited molecules will have radiatively decayed during the pulse. This overlap time sufficiently populates the $6\,^1\Sigma_g^+ (7,31)$ state. Molecular fluorescence emission from the $6\,^1\Sigma_g^+ (7,31)$ state is collected at the right angles to the propagation directions of the lasers using a periscope and an imaging spectrometer-detector system. The imaging spectrometer, Horiba iHR320, acts as a narrowband filter and has three built in diffraction gratings (600, 1200 and 2400 groves/mm), 1200 g/mm is predominantly used throughout the experiment. The spectrometer has two exit ports (each port can be selected with a motorized swing away mirror): the front port has a back illuminated thermoelectrically-cooled (-50~$^\circ$C) charge-couple device (CCD) chip with 14x10$^{-6}$ m squared pixel size while the side port has a slit and a photomultiplier tube (PMT). The widths of the slits and swing away mirror are remotely controlled. The spectrometer is calibrated with a mercury light source. The FWHM resolution of the spectrometer, with the 1200 g/mm holographic grating (blazed at 500 nm) and the side slit at 90x10$^{-6}$ m, is 19.33 cm$^{-1}$ at 17798.9 cm$^{-1}$ and 14.19 cm$^{-1}$ at 15518.5 cm$^{-1}$. For lifetime measurements, the data was collected from the side slit (width at 90x10$^{-6}$ m) using a PMT-multichannel scaler (MCS). A second photomultiplier tube (Hamamatsu R928), with interference filter, was mounted on the second window of the heat-pipe and used to monitor the Na atomic fluorescence from the 4p state to the ground state. \\
\begin{figure}[th]
\includegraphics[scale=0.4]{Fig2.eps}
\caption{(Color online) Experimental excitation to the $6\,^1\Sigma_g^+ (3s+5s)$ state and detection scheme on a partial energy level diagram of Na$_2$.}\label{fig:Fig2}
\end{figure}
\section{Measurements}\label{sec:Measurements}
With the setup shown in Fig.~\ref{fig:Fig1}, a resolved molecular fluorescence spectrum was acquired from the $6\,^1\Sigma_g^+ (7,31)$ state as shown in Fig.~\ref{fig:Fig3}. The spectrum exists only when the double-resonance excitation to the $6\,^1\Sigma_g^+ (7,31)$ state is accomplished. This detection scheme is almost background-free because the frequencies of this spectrum are far away from the molecular fluorescence from the $A\,^1\Sigma_u^+$ state, excited by Laser 1. In addition, it was not possible to observe any molecular fluorescence from the $6\,^1\Sigma_g^+ (7,31)$ state using Laser 1 or Laser 2 alone, and thus no two-photon absorption was observed from the $6\,^1\Sigma_g^+ (7,31)$ state. As a result, the spectrum suggests the two-step double-resonance for the $X\,^1\Sigma_g^+ (0,29)$ $\rightarrow$ $A\,^1\Sigma_u^+ (8,30)$ $\rightarrow$ $6\,^1\Sigma_g^+ (7,31)$ transition. The spectral peaks were identified by comparing them with the output of the LEVEL 8.0 Fortran program. This program uses experimental potential energy curves and calculates quantities such as Franck-Condon factors, Einstein coefficients, transition dipole moment matrix elements and energies of the ro-vibrational levels. High Franck-Condon factors suggest $P-$ and $R-$branches from the $6\,^1\Sigma_g^+ (v=7,J=31)$ to the $A\,^1\Sigma_u^+ (v=3-8, J=30,32)$ state should be readily observed and identified, and they are (see Fig.~\ref{fig:Fig3}). We report the lifetime measurements from the $v=6$ doublets. The separation between the (6,30) and (6,32) doublet is 13.40 cm$^{-1}$ according to the output of the LEVEL fortran program. The photon counting technique is used to perform low-light-level measurement with high sensitivity and accuracy. \\
\begin{figure}[th]
\includegraphics[scale=0.35]{Fig3.eps}
\caption{(Color online) Double-resonance spectrum from the $6\,^1\Sigma_g^+$ ($7,31$) state. The spectrum, only obtained when both lasers are on resonance, was taken using a 1200 g/mm spectrometer-CCD with FWHM resolution of 3 cm$^{-1}$. The spectral peaks mainly consists of $R(30), P(32)$ doublets for the $A\,^1\Sigma_u^+ (v',J') \leftarrow 6\,^1\Sigma_g^+ (7,31)$ fluorescence emission and they are identified by comparing to the output of the LEVEL 8.0 program. The inset shows the Franck-Condon factors, extracted from the program, as a function of vibrational quantum number.}\label{fig:Fig3}
\end{figure}
To do time-correlated photon counting we used a multi-channel scaler (MCS) which has a fast amplifier, a built-in discriminator and 5-ns bin width (the temporal resolution). This technique is used to record low level light signals for the radiative lifetime measurements. The time accuracy is not limited by the width of the PMT detector pulse and thus there is no loss due to gating as in Boxcar devices or gated CCDs. The MCS works by measuring the time between a trigger pulse and an associated PMT pulse which results from a single photon. We measure about one photon per five or more triggers. When the MCS receives a trigger pulse a record of up to about 32,000 time bins, in our case about 1,000 time bins of 5-ns, starts. During a record, when a signal pulse is received, one count is added to the time bin corresponding to when the pulse is received in relation to the trigger. The MCS was triggered by laser pulses from the YAG laser using a photodiode and a scan of 20,000 records was accumulated and displayed on the screen on real-time. The resulting histogram yielded the fluorescence decay as a function of time. If the number of molecules leaving the $6\,^1\Sigma_g^+ (v,J)$ state during the time interval $dt$ is $N=-k_rNdt-k_{nr}Ndt$, the solution gives the emitted radiation with intensity proportional to $N(t)=N_oe^{(k_r+k_{nr})t}$, where $N_o$ is the initial population, $k_r$ is the radiative decay, and $k_{nr}$ is the non-radiative decay. Then, the lifetime is defined as $N(t)e^{-t/\tau_e}$. The typical result, a histogram with an exponential drop of counts as a function of time from the MCS, is shown in Fig.~\ref{fig:Fig4}. The recorded data were transferred to a computer and the effective lifetime, $\tau_e$, was obtained from the convolution of a Gaussian with an exponential fit to the data using Origin 2017. During the experiment, each scan was accumulated for 40 minutes to produce a full decay curve and more than ten independent measurements were done at one pressure point on various days. The average of the effective lifetime was calculated for each pressure setting. In order to determine the collision-free lifetime, this procedure was repeated at argon pressures ranging from 200 mTorr to 460 mTorr (at the operating temperature) and thus the effective lifetime as a function of pressure was obtained and plotted. A linear dependence of such a plot, known as Stern-Volmer, is expected if the probability of the multiple collisions is negligible. The Stern-Volmer relationship can be written as
\begin{equation}
\frac{1}{\tau_{e}} = \frac{1}{\tau_r} + \frac{1}{\tau_{nr}},
\end{equation}
where $\tau_{e}$ is the effective lifetime, $\tau_r$ is the radiative lifetime, and $\tau_{nr}^{-1}=N_p\bar{v}\sigma$. Here, $N_p$ is the number density of the buffer gas, $\bar{v}$ is the average velocities of the colliding Na$_2^*$-Ar atoms over the
Maxwell-Boltzmann distribution of relative velocities at the heat-pipe oven temperature, and $\sigma$ is the collisional cross section between the excited dimer and argon atoms. The effect of argon gas collisions is eliminated with the Stern-Volmer extrapolation. The radiative lifetime was extracted from the zero pressure extrapolation of the Stern-Volmer plot.
Figure~\ref{fig:Fig5} shows the inverse lifetimes (ns) as a function of argon pressure at 290~$^\circ$C in the range of 200 mTorr - 450 mTorr. The radiative lifetime at the collision-free limit (intercept of the Stern-Volmer plot) was extracted from a linear fit to the data. The radiative lifetime of the $6\,^1\Sigma_g^+ (7,31)$ state is found to be 39.56 ($\pm$ 2.23) ns with a correlation coefficient of 0.99 and residual sum square of 7.86x10$^{-9}$. \\
The effect of high pressure on the lifetime measurements was observed for pressures from about 500 mTorr to 1700 mTorr. A slight curving trend at high pressures may indicate the possibility of buffer gas quenching effect~\cite{Romalis09} which can be studied in the future. For the purpose of determining the radiative lifetime at the collision-free limit a linear fit to the data was done at low pressures. Additionally, we repeated the time-correlated photon counting experiments at different operating temperatures ranging from 290~$^\circ$C to 340~$^\circ$C and at different laser powers. The data showed no laser power dependency on lifetime measurements. However, as expected, lifetime decreased as we increased the temperature to 340~$^\circ$C. To ensure the reproducibility of the data we have repeated the experiments on different days. We have validated our measurements and setup by repeating the measurements for the known lifetimes of the Na$_2$ $A\,^1\Sigma_u^+ (7,31)$ state under the same experimental conditions. The extracted radiative lifetime from the Stern-Volmer plot is found to be 14.02 (93) ns, comparable with the literature values of 13.11 (47) ns~\cite{Ducas76} and 12.51 (62) ns~\cite{Baumgartner84}. \\
\begin{figure}[ht]
{{\includegraphics[scale=0.30]{Fig4.eps} }
\caption{(Color online) A typical time-correlated photon counting scan (40 minutes to complete one decay curve) illustrating a histogram consisting of a range of `time bins' from the MCS. The blue line is a convolution fit of a Gaussian with an exponential curve function to the data with a correlation coefficient of 0.99. The effective lifetime of the molecular excited state was extracted from this convolution fit. }\label{fig:Fig4}
\end{figure}
In addition, we observed atomic fluorescence from the Na $4p\,^{2}P_{J}$ state to the $3s\,^{2}S_{1/2}$ ground state using a 330 nm interference filter. This is due to the molecular dissociation into the Na$_2$ (3s+4p) asymptote. The signal exists only when molecules were formed in the heat-pipe. To ensure that the atomic fluorescence is not a result of an accidental dissociation into the Na$_2$ (3s+5s) asymptote and then branching into the $4p$ state we monitored fluorescence for the $5s\,^{2}P_{J} \rightarrow 4p\,^{2}P_{J}$ transition at 615 nm and found no evidence.
\begin{figure}[ht]
\includegraphics[scale=0.36]{Fig5.eps}
\caption{(Color online) Plot showing the variation in the inverse lifetime as a function of argon pressure at 290~$^\circ$C. Each data point is an average of 10 to 20 independent lifetime measurements, with the error bar equal one standard deviation. Red filled square data point is an outlier and was excluded in the fitting procedure. Red line shows a weighted linear fit to the data with a correlation coefficient of R$^2$=0.98 and residual sum square of 7.86x10$^{-9}$. }\label{fig:Fig5}
\end{figure}
\section{Theoretical Calculations}\label{sec:Calculations}
All lifetime calculations presented here were performed using the LEVEL 8.2 Fortran program, developed by Le Roy~\cite{LeRoy8.2}, for the bound-bound transitions and the BCONT Fortran program for the bound-free transitions. The LEVEL code solves the radial Schr{\"o}dinger equation of bound and quasibound levels with appropriate input for the transition dipole moment function~\cite{Magnier17} and relevant potential energy curves. The dipole selection rules allow spontaneous emission from the $6\,^1\Sigma_g^+ (v,J)$ (Tsai~\emph{et al.}~\cite{tsai94-2} and Laue~\emph{at al.}~\cite{Laue03}) to seven electronic potentials: $1(A)\,^1\Sigma_u^+$ from Tiemann~\emph{et al.}~\cite{Tiemann96}, $1(B)\,^1\Pi_u^+$ from Comacho~\emph{et al.}~\cite{Camacho05} and Tiemann~\emph{et al.}~\cite{tiemann87}, $2\,^1\Sigma_u^+$ from Pashov \emph{et al.}~\cite{Pashov00-1}, $2\,^1\Pi_u^+$ from Grochola \emph{et al.}~\cite{Grochola05}, $3\,^1\Sigma_u^+$~\cite{Sanli18}, $3\,^1\Pi_u^+$ from Grochola \emph{et al.}~\cite{Grochola06}, and $4\,^1\Sigma_u^+$ from Grochola \emph{et al.}~\cite{Grochola04}. The potentials and transition dipole moment functions for the seven dipole allowed transitions from the $6\,^1\Sigma_g^+ (v,J)$ state are shown in Fig.~\ref{fig:Fig1S} and Fig.~\ref{fig:Fig2S}.\\
\begin{figure}[th]
\includegraphics[scale=0.4]{Fig6.eps}
\caption{(Color online) Na$_2$ $6\,^1\Sigma_g^+$ potential energy function and the singlet, ungerade electronic states used in calculations into which excited molecules in the $6\,^1\Sigma_g^+$ state can decay.}\label{fig:Fig1S}
\end{figure}
\begin{figure}[th]
\begin{center}
\includegraphics[scale=0.4]{Fig7.eps}
\caption{(Color online) Transition dipole moment (TDM) functions that couple the $6\,^1\Sigma_g^+$ state to each of the other potentials used in the calculations.}\label{fig:Fig2S}
\end{center}
\par
\end{figure}
The LEVEL 8.2 program computes the Einstein $A$ coefficient for spontaneous emission, $A_{if}$, from an initial level $|i\rangle$ = $|\alpha,v,J\rangle$ to a final level $|f\rangle$ = $|\alpha{'},v',J'\rangle$. This coefficient can be expressed as~\cite{Herzberg50,Bernath16}
\begin{equation}\label{eq:A}
A_{if}=\frac{16\pi^3}{3}\frac{k_{if}^3}{\epsilon_oh}\frac{S_{JJ'}}{(2J+1)}|\langle\psi_{v',J'}^{\alpha{'}}|M(r)|\psi_{v,J}^\alpha\rangle|^2,
\end{equation}
where $\alpha$ denotes the electronic state, $v$ and $J$ are vibrational and rotational quantum numbers respectively, $A_{if}$ has units $s^{-1}$, $M(r)$ is the transition dipole moment function, $k_{if}$ the emission wavenumber, $S_{JJ'}$ the H{\"o}nl-London rotational intensity factor, $\psi_{v,J}^\alpha$ is the normalized radial nuclear wave functions belonging to initial and final state of the transition.\\
In cases where the sum of the Franck-Condon factors is less than one, we incorporated bound-free transitions using BCONT code which computes the intensity as a function of wavelength. All bound-free calculations were completed using a modified version of the BCONT 2.2~\cite{LeRoy2.2} program. The modifications have been described previously in Ref.~\cite{Anunciado16} and even greater detail in Ref.~\cite{Brett10} and therefore will not be repeated here. The bound-free calculations are more labor intensive than the bound-bound transitions, and therefore the ro-vibrational levels $6\,^1\Sigma_g^+ (v= 0, 7, 8, 20, 40, 75, 100, 125, 155, 174, 200, J = 1, 31)$ were selected because they include energy levels with wavefunctions that reside in the inner well below the double well region, in the double minimum region, and in the region just above where the inner and outer well merge. Figure~\ref{fig:Fig3S} shows some of the calculated wave functions used by the LEVEL and BCONT programs.\\
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.35]{Fig8.eps}
\caption{(Color online) $6\,^1\Sigma_g^+$ potential with the wavefunctions of the $v=0, 8, 20, 40, 75, 100, 125, 155, 174$, and $200$, all with $J=31$, plotted at the energy of their respective ro-vibrational level. Note that the wavefunctions existing in the region where two separated wells exist ($v=40, 75, 100, 125, 155$, and $174$) do oscillate in both wells, however all but $v=174$ must be scaled up several orders of magnitude for those oscillations to become visible. }\label{fig:Fig3S}
\end{center}
\end{figure}
The revised BCONT code calculates the bound-free intensity $dI$ per molecule emitted into a small energy band $dE$ at photon energy $E=hck$ given by $dI = E \cdot \Gamma_{if}^{\alpha{''}}dE$, where the photon emission rate is given by
\begin{equation}\label{eq:Gamma}
\Gamma_{if}^{\alpha{'}}dE=\frac{16\pi^3}{3}\frac{k_{if}^3}{\epsilon_oh}\frac{S_{JJ'}}{(2J+1)}|<\psi_{v',J'}^{\alpha{'}}|M(r)|\psi_{v,J}^\alpha>|^2dE
\end{equation}
The main difference between this photon emission rate and the bound-bound version in Eq.~(\ref{eq:A}) is the replacement of the final state bound wave function $\psi_{v,J}^\alpha$ with the continuum wave function, $\psi_{EJ}^\alpha$. The lifetime of a single ro-vibrational level of the $6\,^1\Sigma_g^+$ is determined using the relation
\begin{equation}\label{eq:tau}
\frac{1}{\tau_i}=\sum_{\alpha{'}}\bigg(\sum_{f}A_{if}+\int_{0}^{E(\alpha{'})}\Gamma_{if}^{\alpha{'}}dE\bigg),
\end{equation}
where the $A_{if}$ coefficients have been obtained from the LEVEL calculations, the $\Gamma_{if}^{\alpha{'}}dE$ from BCONT, and the overall sum includes all dipole allowed transitions. The bound-bound sum extends over all levels in the given electronic state $\alpha^\prime$, and the integral extends from zero frequency to the appropriate threshold $E(\alpha{'})=E(v,J)-E(\alpha^\prime$ asymptotic energy$)$. The sum of Einstein coefficients as a function of $6\,^1\Sigma_g^+$ vibrational level for transitions to the $A\,^1\Sigma_u^+$ is plotted in Fig.~\ref{fig:Fig4S}. \\
\begin{figure}[th]
\includegraphics[scale=0.3]{Fig9.eps}
\caption{Each point on this plot represents the sum of all Einstein coefficients originating from a single ro-vibrational level, $6\,^1\Sigma_g^+ (v,J)$, to all bound levels of final state $\alpha^\prime$. In this plot, the final state is $\alpha^\prime = 1(A)\,^1\Sigma_g^+$ and the vibrational levels $v=0-200$ with $J=31$ are shown. The upper branch is produced by $6\,^1\Sigma_g^+ (v,J)$ levels with predominantly inner well wavefunctions, while the lower branch corresponds to those with predominantly outer well wavefunctions. Around approximately $v=180$, the wavefunctions accrue substantial amplitude in both the inner and outer region as the two wells merge.}\label{fig:Fig4S}
\end{figure}
If the $A_{if}$ and $\Gamma_{if}^{\alpha^\prime}dE$ for a single final electronic state, $\alpha^\prime$, are summed and divided by the full sum over all $\alpha^\prime$ in Eq.~\ref{eq:tau}, one obtains the branching ratio, $r_b^\alpha$, into final state $\alpha^\prime$, i.e.,
\begin{equation}
r_b^a=\frac{(\sum_{f}A_{if}+\int_{0}^{E(\alpha{'})}\Gamma_{if}^{\alpha{'}}dE)}{\sum_{{\alpha{'}}}(\sum_{f}A_{if}+\int_{0}^{E(\alpha{'})}\Gamma_{if}^{\alpha{'}}dE)}.
\end{equation}
Branching ratios for $J$=1 and $J$=31, for the calculations that include both bound-bound and bound-free transitions, are plotted in Fig.~\ref{fig:Fig6S}. Note the rather small branching ratios for the $1(B)\,^1\Pi_u$, $2\,^1\Pi_u$, $3\,^1\Pi_u$, and $4\,^1\Sigma_u^+$ states indicate that they have minimal impact on the resulting lifetimes, since the excited molecule has a lower probability of decaying through those states. Conversely, the $1(A)\,^1\Sigma_u^+$, $2\,^1\Sigma_u^+$, and $3\,^1\Sigma_u^+$ states offer the most probable decay paths for the excited molecule.\\
\begin{figure*}[ht]
\subfloat[]{{\includegraphics[scale=0.35]{Fig10a.eps} }
\subfloat[]{{\includegraphics[scale=0.35]{Fig10b.eps} }
\caption{(Color online) Branching ratios for $6\,^1\Sigma_g^+$ ro-vibrational levels with $J$=1 and $J$=31.}\label{fig:Fig6S}
\end{figure*}
Figure~\ref{fig:Fig7S} shows our calculated radiative lifetimes of the $6\,^1\Sigma_g^+$ ro-vibrational levels in the range of $v=0-200$ with $J=1$ and $J=31$. This figure clearly demonstrates the importance of the bound-free transitions in the calculations.
\begin{figure*}[ht]
\centering
\subfloat[]{
{\includegraphics[scale=0.3]{Fig11a.eps} }
\subfloat[]{{\includegraphics[scale=0.3]{Fig11b.eps} }
\caption{(Color online) Calculated lifetimes of $6\,^1\Sigma_g^+(v,J=1,31)$ ro-vibrational levels in the range $v=0-200$. The upper branch is produced by ro-vibrational states with wavefunctions that have predominantly outer well amplitude, while the lower branch wavefunctions are predominantly inner well. The open circles were computed using only bound-bound transitions, $\tau^{-1}=\sum_{\alpha{'}}\sum_{f}A_{if}^{\alpha{'}}$, while the orange X represents lifetimes calculated as in Eq.~(\ref{eq:tau}). The inclusion of bound-free transitions causes a much larger reduction of the calculated lifetimes for the ro-vibrational levels with appreciable outer well wavefunction amplitude since there is generally a less favorable overlap of the $6\,^1\Sigma_g^+$ outer well.}\label{fig:Fig7S}
\end{figure*}
A strong dependence of the calculated radiative lifetime on the radial coordinate, $R$, is observed, causing two distinct branches to occur in the portion of the $6\,^1\Sigma_g^+$ potential in which the two wells remain separated. Figure~\ref{fig:Fig9S} shows the $6\,^1\Sigma_g^+$ radiative lifetimes obtained from the full (Eq.~\ref{eq:tau}) calculation for $J=1$ and $J=31$, plotted as a function of their vibrational level. The large difference of approximately a factor of three between the $J=1$ and $J=31$ lifetimes for $v=40$ and for $v=100$ occurs due to a ``switch'' of the wavefunction, which is predominantly outer well for the long lifetime value, $v=40,J=31$ and $v=100,J=1$, and is predominantly inner well for the shorter lifetime value, $v=40,J=1$ and $v=100,J=31$. The results of these calculated radiative lifetimes are presented in Table~\ref{tab:table1}. \\
\begin{figure}[ht]
\centering
\includegraphics[scale=0.4]{Fig12.eps}
\caption{(Color online) Comparison of radiative lifetimes of J=1 and J=31 calculated for selected vibrational levels $v=0,8,20,40,75,100,125,155,174,$ and $200$, using Eq.~(\ref{eq:tau}).}\label{fig:Fig9S}
\end{figure}
\begin{table}[th!]
\caption{\label{tab:table1}Calculated radiative lifetimes (in nanosecond) of selected ro-vibrational levels of the Na$_2$ $6\,^1\Sigma_g^+$ state.}
\begin{ruledtabular}
\begin{tabular}{ccc}
$v$ & $J=1$ & $J=31$ \\
\hline
0 & 46.4 & 45.3\\
7 & 42.5 & 42.8\\
8 & 42.5 & 42.9\\
20 & 48.2 & 48.8\\
40 & 49.9 & 153.2\\
75 & 142.9 & 142.4\\
100 & 139.4 & 46.9\\
125 & 114.8 & 115.1\\
155 & 101.9 & 102.6\\
174 & 54.3 & 44.9\\
200 & 91.7 & 92.3\\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Results and Discussions}\label{sec:Results}
A time-resolved double resonance spectroscopy technique has been applied to the sodium diatomic molecules in a heat-pipe and the radiative lifetime of the $6\,^1\Sigma_g^+ (7,31)$ state was measured using a time-correlated photon counting technique. Also, lifetime calculations in various ro-vibrational levels of the $6\,^1\Sigma_g^+$ state using bound-bound and bound-free transitions were performed.\\
The radiative lifetimes are measured and the effect of argon gas collisions is eliminated with the Stern-Volmer extrapolation, explained in section~\ref{sec:Measurements}. The measured radiative lifetime (extracted from the zero-pressure of the extrapolation) of the $6\,^1\Sigma_g^+$ ($7,31$) state is found to be 39.56~($\pm$2.23) ns and our calculations showed this value to be 42.8 ns. Experimental and calculated radiative lifetimes reported in this work have been compared and the results were in reasonable agreement. To validate our measurements and the techniques, we also measured the radiative lifetime of the $A\,^1\Sigma_u^+ (8,30)$ state by extrapolating the Stern-Volmer plot to zero pressure. The result is in good agreement with the literature values within the experimental uncertainties. \\
The observation of the non-linear behaviour on the lifetimes at high pressures in the range of 500 mTorr - 1800 mTorr may be due to the quenching effect. Further experimental and theoretical studies of the high pressure effect on the lifetime will be investigated in the future work. For the purpose of determining the radiative lifetime at the collision-free limit a linear fit to the data was done at pressures below 500 mTorr. In addition, the observed dissociation signal from the Na 4p state does not originate from any resonance transition to the $6\,^1\Sigma_g^+$ ($7,31$) state. The measured spectrum exists only when both lasers are turned on (see Fig. 3). Since the $6\,^1\Sigma_g^+$ ($7,31$) state is about 2500 cm$^{-1}$ below the (3s+4p) asymptote collisional energy ($kT\approx$ 600~cm$^{-1}$) is not sufficient to transfer $6\,^1\Sigma_g^+$ ($7,31$) molecules into this asymptote. In addition, dissociation signal appears only when molecules are formed in the heat-pipe ($\sim 280~^\circ$C), indicating that no atomic transitions exist with either laser. As a result, the observed dissociation signal has no effect on the measured lifetime. A possible channel for the dissociating molecule into the (3s+4p) asymptote may be through a one-color two-photon transition. The information about the origin of observed dissociation can be studied in detail in a separate work and will not be detailed here.\\
Table~\ref{tab:table2} is the tabulated outline of the measurements and theoretical calculations. A reasonable agreement between the experimental and theoretical results are achieved when bound-bound and bound-free transitions are considered.
\begin{table}[H]
\caption{\label{tab:table2}Comparison of measured and calculated radiative lifetimes of selected ro-vibrational levels of the Na$_2$ $6\,^1\Sigma_g^+$. Theoretical calculations are compared according to the bound-bound, denoted as bb, and bound-free, denoted as bf, transitions.}
\begin{ruledtabular}
\begin{tabular}{ccccc}
Method & Level & $\tau~$(ns) & Reference\\
\hline
Experiment & $6\,^1\Sigma_g^+ (7,31)$ & $39.56~(\pm 2.23)$ & This work\\
Theory~(bb+bf) & $6\,^1\Sigma_g^+ (7,31)$ & $42.8$ & This work\\
Theory~(bb+bf) & $6\,^1\Sigma_g^+ (7,1)$ & $42.5$ & This work\\
Theory~(bb)& $6\,^1\Sigma_g^+ (7,1)$ & $126$ & Ref.~15\\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Conclusion}\label{sec:Conclusion}
For the first time, experimental lifetime measurement of the $6\,^1\Sigma_g^+ (7,31)$ inner well and theoretical calculations in various ro-vibrational levels of the $6\,^1\Sigma_g^+$ state using bound-bound and bound-free transitions were performed. In addition, the effect of inclusion of the bound-free transitions in the calculated radiative lifetime of the $6\,^1\Sigma_g^+$ (3s+5s) ro-vibrational levels in the range of $v=0-200$ with $J=1$ and $J=31$ were reported. The results reveal the importance of the bound-free transitions and rotational quantum number (e.g. a large difference of about a factor of three between the $J$ = 1 and $J$ = 31 for the $v$ = 40 and $v$ = 100) on the lifetime calculations. Also, the transition dipole moment functions coupling the $6\,^1\Sigma_g^+$ state to seven singlet, ungerade states and also the branching ratios for radiative transitions from the $6\,^1\Sigma_g^+$ ro-vibrational levels into those seven states were demonstrated. The measured and calculated radiative lifetimes are found to be 39.56~($\pm$ 2.23) ns and 42.8 ns, respectively. The results are in reasonable agreement.
\section{Acknowledgements}
Financial support from the National Science Foundation (Grant No. NSF-PHY-1607601) is gratefully acknowledged. Also, we acknowledge Professor Sylvia Magnier of Universite Lille, France for providing transition dipole moment functions used in our calculations. Authors from Miami University thank Professor Mark Havey of Old Dominion University for lending the MCS and Professor Marjatta Lyyra of Temple University for the heat-pipe.
|
\section{Introduction et définitions}
En 1965, R. Thompson découvrit les premiers exemples de groupes $T\subset V$ de présentation finie, simples et infinis. Le groupe $T$ [resp. $V$] se représente comme groupe d'homéomorphismes [resp. échanges d'intervalles] affines par morceaux du cercle (voir \cite{CFP}, \cite{Ste}). En 1987, K. Brown (\cite{Bro}) a défini une famille $T_{r,m} \subset V_{r,m}$ englobant $T$ et $V$ et les groupes $V_{r,m}$ sont isomorphes aux groupes $G_{r,m}$ de Higman (\cite{Hig}).
Plus précisément, soit $r$ un entier strictement positif, on note $\mathbb S_r$ le cercle ${\mathbb R }/{r\mathbb Z}$ de longueur $r$. Le cercle de longueur $1$ est $\mathbb S_1$, nous le noterons plus classiquement $\mathbb S^1$.
\begin{definition}
Un homéomorphisme $f$ du cercle $\mathbb S_r$ est
{\em affine par morceaux} s'il existe une subdivision finie
$0<a_1<a_2<\dots<a_p=r$ de l'intervalle $[0,r]$ et un relevé ${\tilde
f}$ de $f$ à $\mathbb R$ tels que $\displaystyle {\tilde f}_{\vert [a_i,a_{i+1}] } (x)
= \lambda_i x + \beta_i, \quad \lambda_i, \beta_i\in \mathbb R.$
Les points $a_i$ sont appelés \emph{points de coupure} de~$f$ et les nombres $\lambda_i$, {\em pentes de~$f$}.
Le groupe des homéomorphismes affines par morceaux de $\mathbb S_r$ préservant l'orientation est
noté $PL^+ (\mathbb S_r)$.
\end{definition}
\begin{definition}Soient $r$ et $m\geq 2$ deux entiers strictement positifs. On définit le \textit {groupe de Brown-Thompson} $T_{ r,m}$ comme l'ensemble des éléments $f$ de $PL^+ (\mathbb S_r)$ tels que :
\begin{itemize}
\item les pentes de $f$ appartiennent à $\langle m \rangle =\{m^s, s\in \mathbb Z\}$.
\item les points de coupure de $f$ appartiennent à $\mathbb Z [\frac{1}{m}]=\{N. m ^s \vert N, s \in \mathbb Z\}$,
\item les images par $f$ des points de coupure de $f$ appartiennent à $\mathbb Z [\frac{1}{m}]$.
\end{itemize}
Le \textit {groupe de Thompson $T$} est $T_{1,2}$.
\end{definition}
De nombreux auteurs se sont intéressés aux invariants et à la question d'isomorphicité pour ces groupes de type Thompson (\cite{BiSt}, \cite{Bri}, \cite{Bri2}, \cite{BrGu}, \cite{Bro}, \cite{Hig}, \cite{Lio}, \cite{Mat}, \cite{Ste} \ $\cdots$). Dans cet article, nous nous concentrons sur les obstructions à l'isomorphicité entre groupes $T_{r, m}$ issues des éléments d'ordre fini et de leurs classes de conjugaison. Le calcul du nombre de ces classes fût effectué pour $G_{r,m}$ par Higman (\cite{Hig}, section 6), pour $T$ par Matucci(\cite{Mat}) puis ultérieurement par Geoghegan-Varisco(\cite{GeVa}) et Fossas(\cite{Fos}).
Comme dans \cite{Mat} et \cite{GeVa}, nous utilisons la représentation comme groupe d'homéomorphismes affines par morceaux du cercle et disposons ainsi d'un invariant dynamique suplémentaire : le nombre de rotation de Poincaré. Nous indiquons sa définition et ses premières propriétés (voir \cite{Her} ou \cite{KH}).
\begin{definition} Soit $f$ un homéomorphisme du cercle $\mathbb S_r$, on définit le \textit{nombre de rotation sur $\mathbb S_r$} de $f$ par : $\displaystyle \rho(f) = \lim\limits_{n\rightarrow \infty} ({\tilde
f^n (0)/ rn } ) \ (mod \ 1) \ \in \mathbb S^1.$
\end{definition}
Ce nombre ne dépend pas du choix du relevé $\tilde{f}$ et satisfait les propriétés classiques :
\begin{property} \
\begin{itemize}
\item $\rho (R_{\alpha}) =\frac{\alpha}{r}$ où $R_{\alpha}(x)= x+\alpha \ (mod \ r)$,
\item $\rho(f^n) = n \, \rho(f)$ pour tout $n\in \mathbb Z$,
\item si $f$ est d'ordre fini $q\in \mathbb N^{>1}$ alors $\rho(f) =\frac{p}{q}$ avec $p<q$ et $p\wedge q=1$,
\item soit $h:\mathbb S_r \rightarrow \mathbb S_{r'}$ un homéomorphisme préservant l'orientation, $\rho(h\circ f\circ h^{-1}) =\rho (f)$.
\end{itemize}
\end{property}
Commençons par cette observation : tout élément d'ordre $q$ est conjugué dans $PL^+ (\mathbb S^1)$ à une rotation d'angle $\frac{p}{q}$ avec $p\wedge q = 1$, une conjugante est construite par moyennisation (voir par exemple \cite{KH}, Proposition 11.2.2). Comme deux rotations d'angles différents ne sont jamais $C^0$-conjuguées, le nombre de classes de conjugaison d'éléments d'ordre $q$ dans $PL^+ (\mathbb S^1)$ est exactement le nombre d'entiers $p<q$ premiers avec $q$ c'est à dire $\varphi(q)$ (la fonction phi d'Euler).
Le Théorème 7.1.5 de \cite{Mat} (voir aussi \cite{GeVa} et \cite{Fos}) exprime qu'il est encore vrai pour le groupe de Thompson $T$ : \textit{"dans $T$, tout rationnel de $\mathbb S^1$ est réalisé comme nombre de rotation d'une unique classe de conjugaison d'éléments d'ordre fini"}.
Ici, nous établissons que cette propriété n'est plus satisfaite par les autres groupes $T_{r, m}$ :
\begin{theorem}\label{thm:1}
Soient $r\geq 1$, $m\geq 2$ et $q\geq 2$ des entiers.
\begin{enumerate}[A.]
\item Si pgcd $(m-1, q)$ ne divise pas $r$ alors il y a $0$ classes de conjugaison d'éléments d'ordre $q$ dans $T_{r,m}$.
\item Si pgcd $(m-1, q)$ divise $r$ alors il y a pgcd $(m-1, q)$ classes de conjugaison d'éléments d'ordre $q$ et de nombre de rotation $\frac{p}{q}$ dans $T_{r,m}$, pour tout entier $p$ premier avec $q$.
\item Si pgcd $(m-1, q)$ divise $r$ alors il y a $\varphi(q)$.pgcd$(m-1, q)$ classes de conjugaison d'éléments d'ordre $q$ dans $T_{r,m}$.
\end{enumerate} \end{theorem}
Nous en déduisons le
\begin{corollary}\label{coro :1} \
Tout rationnel de $\mathbb S^1$ est réalisé comme nombre de rotation d'une unique classe de conjugaison d'éléments d'ordre fini dans $T_{r,m}$ si et seulement si $m=2$.
Le groupe de Thompson $T$ n'est isomorphe à aucun des groupes $T_{r,m}$ avec $m\not=2$ et tout morphisme de $T$ dans $T_{r,m}$, avec $m\not=2$ et $r\not= 0$ $mod \ (m-1)$, est trivial.
\end{corollary}
Ce corollaire contraste avec le résultat d'ubiquité de $F$ montré par Brin (\cite{Bri2}).
Notre approche diffère de celles de \cite{Mat}, \cite{GeVa} et \cite{Fos} au sens où elle est essentiellement basée sur un critère dû à Bieri et Strebel \cite{BiSt}.
\section{Préliminaires}\label{Pre}
\subsection{Critère de Bieri-Strebel pour les groupes de Brown-Thompson.}
Nous reprennons ici le critère général de Bieri-Strebel déterminant à quelles conditions deux intervalles réels sont échangés par une application affine par morceaux avec points de coupure et pentes prescrites (voir Théorème A 4.1 de \cite{BiSt}).
\begin{definition} \
\begin{itemize}
\item Un \textit{$m$-intervalle} est un intervalle réel dont les extrémités sont dans $\mathbb Z \left[\frac{1}{m}\right]$.
\item Un homéomorphisme $f: I\rightarrow I'$ est dit $PL_m$ s'il est affine par morceaux avec pentes dans $\langle m \rangle$ et points de coupure dans $\mathbb Z \left[\frac{1}{m}\right]$.
\item Deux intervalles $I$ et $I'$ sont dits \textit{$PL_m$-équivalents} s'il existe un homéomorphisme $PL_m$ entre-eux.
\end{itemize}
\end{definition}
\begin{proposition} (\cite{BiSt}, \cite{Lio}) \label{BS} Deux $m$-intervalles $I$ et $I'$ sont $PL_m$-équivalents si et seulement si $\vert I \vert - \vert I' \vert\in ( m-1). \mathbb Z \left[\frac{1}{m}\right]$, où $\vert I \vert$ représente la longueur de l'intervalle $I$. \end{proposition}
\begin{proof}
Soient $I=[a,c]$ et $I'=[a',c']$ avec $a,a',c,c'\in \mathbb Z \left[\frac{1}{m}\right]$. Supposons qu'il existe $f$ un homéomorphisme $PL_m$ entre $I$ et $I'$. Notons $a=b_0<b_1 ...b_{n-1}<b_n =c$, $ b_i\in \mathbb{Z} \left[\frac{1}{m}\right]$, les points de coupure de $f$ et $\lambda_i=m^{k_i}$, $k_i\in \mathbb{Z}$, la pente de $f$ sur $[b_{i-1},b_{i}]$. Nous allons montrer que $\vert I \vert - \vert I' \vert= (c-a)-(c'-a')\in ( m-1). \mathbb Z \left[\frac{1}{m}\right]$.
\smallskip
Comme $ c-a = \sum_i (b_i -b_{i-1})$ et $c'-a' = \sum_i \lambda_i (b_i -b_{i-1})$, on a
$ (c-a) - (c'-a') = \sum_i (1-\lambda_i)(b_i -b_{i-1})$ et
$ (1-\lambda_i)= -(m-1)\sum_{p=0}^{k_i-1} m^{p}=(m-1)M_i$ avec $\displaystyle M_i \in\mathbb{Z}$ et finalement
$(c-a) - (c'-a') = (m-1) \sum_i M_i (b_i -b_{i-1}) \in (m-1).\mathbb{Z} \left[\frac{1}{m}\right]$.
\medskip
\noindent R\'eciproquement, supposons $ \vert I \vert - \vert I' \vert\in ( m-1).\mathbb{Z} \left[\frac{1}{m}\right]$ \ \ $(*)$.
Quitte à composer à la source et au but par des rotations d'angles convenables dans $\mathbb{Z} \left[\frac{1}{m}\right]$, on peut supposer que $I=[0,b]$ et
$I'=[0,b']$ avec $b,b' \in \mathbb{Z} \left[\frac{1}{m}\right]$ positifs.
La condition $(*)$ se traduit par le fait qu'il existe $a \in \mathbb{Z} \left[\frac{1}{m}\right]$
tel que $ b' = b + (m-1) a$ et il s'agit de construire un $PL_m$-homéomorphisme $f: [0,b] \to [0,b + (m -1)a]$ pour tous $a,b\in \mathbb Z \left[\frac{1}{m}\right]$ avec $b\geq 0$ et $b + (m-1) a\geq 0$. L'inverse d'un homéomorphisme $PL_m$ entre $m$-intervalles étant $PL_m$, on peut aussi supposer, quitte à échanger $b$ et $b'$, que $a\geq 0$.
\smallskip
{\bf Cas 1 : $a<b$} ($0 \leq b-a\leq b$). L'application $f_0 : [0,b] \rightarrow [0, b+ (m-1)a]$ définie par $$f_0(x) = \left\{ \begin{array}{ccc} & x & {\text{ si }} x\in [0, b-a] \cr
& m(x -(b-a)) +(b-a) & {\text{ si }} x\in [ b-a, b]
\end{array} \right.$$ est l'homéomorphisme $PL_m$ cherché.
\smallskip
{\bf Cas 2 : $a\geq b$.} Choisissons $p\in \mathbb N$ tel que $0\leq m^{-p} a <b$. D'après le cas 1, il existe $f_0: [0,b] \rightarrow [0, b+ (m-1) m^{-p} a]$ ayant les propriétés requises. On définit alors $f_1 : [0, b+ (m-1) m^{-p} a]\rightarrow [0, b+ (m-1) a]$ par $$f_1(x) = \left\{ \begin{array}{cc} x &{\text{ si }} x\in [0, b] \cr
m^p(x -b) + b &{\text{ si }} \ \ \ \ \ \ \ x\in [ b, b+ (m-1) m^{-p} a].
\end{array} \right. $$ L'application cherchée est $f= f_1\circ f_0$. \end{proof}
\medskip
\begin{consequence}[Isomorphisme de Bieri-Strebel] \label{cons:1}
Soient $m>1$ un entier, si $r$ et $r'$ sont deux entiers positifs congrus modulo $m-1$ alors les groupes $T_{r,m}$ et $T_{r',m}$ sont isomorphes. Par suite tout $T_{r',m}$ est isomorphe à l'un des $m-1$ groupes $T_{r,m}$, $r\in {1, ..., m-1}$.
\end{consequence}
\begin{remark}\label{rema :1}
Tous les intervalles dyadiques sont $PL_2$-équivalents et par suite tous les groupes $T_{r,2}$ sont isomorphes à $T$.
\end{remark}
\subsection{Nombres de rotation des éléments d'ordre fini.}
\begin{proposition} (\cite{Lio}) \label{prop :2} Soient $m\geq 2$, $r\geq 1$ et $q\geq 1$ des entiers.
\begin{enumerate}
\item Si le groupe $T_{r,m}$ contient un élément d'ordre $q$ alors pour tout $p\in \mathbb{N}^*$, le groupe $ T_{r,m}$ contient un élément d'ordre fini de nombre de rotation $\frac{p}{q}$.
\item Le groupe $T_{r,m}$ contient un élément d'ordre $q$ si et seulement si pgcd$(m-1,q)$ divise $r$.
\end{enumerate}
\end{proposition}
\begin{proof} Nous supposons $q\geq 2$, pour $q=1$ le résultat est trivial.
\textbf{Premier item.} Supposons qu'existe $f\in T_{r,m}$ d'ordre $q$. On a $\rho(f)=\frac{n}{q}$ où les entiers $n$ et $q$ sont premiers entre-eux. Par Bezout, il existe $u$ et $v$ entiers tels que $1 =un + vq$.
Soit $p\in \mathbb{N^*}$, on définit un élément d'ordre fini de $T_{r,m}$ par $g=f^{up}$. On a $ \rho(g)= \rho(f ^{up}) = up \rho(f) = \frac{upn}{q}=\frac{p(1-vq)}{q} = \frac{p}{q}-pv= \frac{p}{q} (mod 1)$.
\medskip
\textbf{Deuxième item.}
\textbf{Condition nécessaire}. Supposons que $r$ soit un multiple de $pgcd(q, m-1)$. D'après Bezout, $r=uq + v(m-1)$, donc $r= uq $ modulo $(m-1)$. L'isomorphisme de
Bieri-Strebel implique que les groupes $T_{uq, m }$ et $T_{r, m}$
sont isomorphes. De plus, le groupe $T_{uq, m }$ contient la rotation $ x\mapsto x+u$ d'ordre $q$ et de nombre de rotation $\frac{1}{q}$.
\medskip
\textbf{Condition suffisante}. Par hypothèse et d'après le premier item, il existe $f\in T_{r,m}$ d'ordre $q$ et $\rho{(f)} =\frac{1} {q}$. Fixons $\tilde f$ un relevé de $f$ à $\mathbb{R}$ et identifions $f$ à $\tilde f (mod \ r)$.
La $f$-orbite de $0$ est ordonnée comme suit $0 <f(0) <....< f^{q-1}(0)<r$.
Les $q$ intervalles $I_i:=[f^{i-1}(0),f^{i}(0)]$, $i=1,\cdots,q$, sont tous $PL_m$-équivalents, donc d'après le critère de Bieri-Strebel, $\vert I_i \vert =\vert I_1 \vert \ mod \ (m-1).\mathbb Z \left[\frac{1}{m}\right]$ et $\vert I_1 \vert=f(0)$. Par conséquent $\displaystyle r= \vert I_1 \vert +\cdots+ \vert I_q \vert = qf(0) \ mod \ (m-1).\mathbb Z \left[\frac{1}{m}\right].$
On en déduit qu'il existe des entiers $u,v,s$ tels que $r - qf(0) =(m-1)\frac {v} { m^s}$ \
et \ $f(0) = \frac {u} { m^s}$. Ainsi,
$ m^s r -qu = (m-1)v$, autrement dit $ m^s r = qu + (m-1)v$. Ceci implique
que $ m^s r$ est un multiple du $pgcd(q, m-1)$. Les entiers $(m-1)$ et $m^s$ étant premiers entre-eux, on conclut que $ r$ est un multiple de $pgcd(q, m-1)$. \end{proof}
\subsection{Critère de Conjugaison $PL_m$}
\begin{proposition}\label{prop :3}
Soient $f_1$ et $f_2$ deux éléments de $T_{r,m}$ d'ordre fini $q$ et de nombre de rotation $\frac{1}{q}$, on note $f_i(0) =\frac{N_i}{m^{s_i}}$, $i=1,2$. Les propriétés suivantes sont équivalentes :
\begin{enumerate}
\item $f_1$ et $f_1$ sont $PL_m$-conjugués (dans $T_{r,m}$),
\item $N_2-N_1$ est un multiple de $m-1$,
\item $f_1(0)- f_2(0) \in (m-1) . \mathbb Z[\frac{1}{m}]$ (autrement dit, les intervalles $[0,f_1(0)]$ et $[0,f_2(0)]$ sont $PL_m$-équivalents).
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{lemma}
Soit $a=\frac{N_a}{m^{s_a}}\in \mathbb Z[\frac{1}{m}]$, tout homéomorphisme $f\in T_{r,m}$ d'ordre $q$, de nombre de rotation $\frac{1}{q}$ et vérifiant $f(0)=a$ est $PL_m$-conjugué à la rotation $R_{N_a}$ de $\mathbb S_{qN_a}$.
\end{lemma}
\begin{proof} L'intervalle $[0,r[$ s'écrit $\displaystyle \bigcup_{i=1}^{q} I_i$, où $I_i =[f^{i-1}(0), f^{i}(0)]$.
On considère l'application affine $H_1 : I_1=[0,a] \rightarrow [0,N_a]$, $x \mapsto m^{s_a} .x$ et on définit par récurrence $H_i : I_i \rightarrow [i-1,i]$ par $H_{i+1} = R_1 \circ H_i \circ f^{-1}.$
\smallskip
On vérifie facilement que l'application $H : \mathbb S_r \rightarrow \mathbb S_{qN_a}$ définie par $H_{\vert I_i} = H_i$ est un $PL_m$-homéomorphisme qui conjugue $f$ à $R_{N_a}$. \end{proof}
\noindent $(1)\implies (2)$. D'après le lemme précèdent, $f_i$ est $PL_m$-conjuguée à la rotation $R_{N_i}$ de $\mathbb S_{qN_i}$. Il nous reste à étudier à quelles conditions deux telles rotations sont $PL_m$-conjuguées.
Soit $h :\mathbb S_{qN_1} \rightarrow \mathbb S_{qN_2}$ une $PL_m$-conjugaison entre $R_{N_1}$ et $R_{N_2}$, quitte à composer au but $h$ par la rotation $R_{-h(0)}$ de $\mathbb S_{qN_2}$, on peut supposer que $h(0)=0$.
Les intervalles $[0,N_1]$ et $[0,N_2]$ étant $PL_m$-équivalents, l'entier $N_2-N_1 \in (m-1).\mathbb Z [\frac{1}{m}]$ et par suite $N_2-N_1$ est un multiple de $m-1$.
\noindent $(2)\implies (1)$. Si $N_2-N_1$ est un multiple de $(m-1)$ alors $[0,N_1]$ et $[0,N_2]$ sont $PL_m$-équivalents et on peut reprendre la preuve du lemme précédent avec pour $H_1$ l'homéomorphisme de Bieri-Strebel qui échange ces 2 intervalles.
\noindent $(2) \Longleftrightarrow (3)$ résulte du calcul suivant :
$f_1(0)-f_2(0)=\frac{N_1}{m^{s_1}}-\frac{N_2}{m^{s_2}}= N_1(m^{-s_1} -1) -N_2(m^{-s_2} -1) + (N_1-N_2)$
\hskip 4.7 truecm $\ =(N_1-N_2) \ mod \ (m-1).\mathbb Z [\frac{1}{m}]$. \end{proof}
\section{Classes de conjugaison d'éléments d'ordre fini dans les groupes de Brown-Thompson.}
\begin{lemma}\label{lemm :2}
Soient $p$ et $q>1$ deux entiers premiers entre-eux et $u>0$, $v$ entiers tels que $up+vq=1$.
Deux éléments $f_1$ et $f_2$ de $T_{r,m}$ d'ordre fini $q$ et de nombre de rotation $\frac{p}{q}$ sont $PL_m$-conjugués si et seulement si $f_1^u$ et $f_2^u$ (de nombre de rotation $\frac{1}{q}$) sont $PL_m$-conjugués.
\end{lemma}
La preuve de ce lemme résulte du fait que la $PL_m$-conjugaison se transmet aux puissances et des généralités suivantes :
On a $\rho ( f^{u})= \frac{up}{q}= \frac{1-vq}{q}= \frac{1}{q}$ et $(f^u)^p = f^{1-vq} = f$ dès que $f$ est d'ordre fini $q$.
\begin{remark} \label{rema :2} Une conséquence de ce lemme est qu'étant donnés $p$ et $q$ deux entiers premiers entre-eux, le nombre de classes de conjugaison d'éléments d'ordre fini $q$ et de nombre de rotation $\frac{p}{q}$ ne dépend pas de $p$.
\end{remark}
\subsection{Classes de conjugaisons dans $T_{r,2}$.} \
Soit $q\in \mathbb N^*$, d'après l'invariance par conjugaison topologique du nombre de rotation et la remarque précédente, il suffit de déterminer le nombre de classes de conjugaison d'éléments d'ordre $q$ et de nombre de rotation $\frac{1}{q}$.
La Proposition \ref{prop :2} indique que $T_{r,2}$ contient un élément $f_1$ d'ordre $q$ et de nombre de rotation $\frac{1}{q}$. Tous les intervalles dyadiques étant $PL_2$-équivalents par la Remarque \ref{rema :1}, l'item (3) de la Proposition \ref{prop :3} est vérifié pour tout autre $f_2\in T_{r,2}$ d'ordre $q$ et de nombre de rotation $\frac{1}{q}$. On en déduit qu'il y a exactement une classe de conjugaison d'éléments d'ordre $q$ et de nombre de rotation $\frac{1}{q}$ ; le résultat de Matucci en découle.
\subsection{Preuve du Théorème \ref{thm:1}
\subsubsection{Preuve de l'item A} Il résulte directement de l'item (2) de la Proposition \ref{prop :2}.
\subsubsection{Preuve de l'item B} Par la Remarque \ref{rema :2}, il suffit d'établir le résultat pour $p=1$.
\begin{proposition}\label{prop :4} Soient $q\in \mathbb N^{>1}$ et $a=\frac{N_a}{m^{s_a}} \in \mathbb Z [\frac{1}{m}] \cap ]0,r[$. Les propriétés suivantes sont équivalentes
\begin{enumerate}
\item Il existe $f\in T_{r, m}$ d'ordre $q$, vérifiant $f(0)=a$ et de nombre de rotation $\frac{1}{q}$.
\item $r-qa \in (m-1) \mathbb Z [\frac{1}{m}]$.
\item $r-qN_a$ est un multiple de $m-1$.
\end{enumerate}
\end{proposition}
\begin{proof} \
\noindent $(1) \implies (2)$. Comme dans la preuve de la Proposition \ref{prop :2}, les $q$ intervalles $I_i:=[f^{i-1}(0),f^{i}(0)[$, $i=1,\cdots,q$, sont tous $PL_m$-équivalents et forment une partition de $[0,r[$. Par conséquent $ r= q\vert I_1 \vert =qf(0) \ mod \ (m-1) \mathbb Z [\frac{1}{m}]$.
\medskip
\noindent $(2) \implies (1)$. Supposons que $r-qa \in (m-1) \mathbb Z [\frac{1}{m}]$.
\noindent \textbf{Cas 1 :} $r-qa \geq 0$. Considère l'homéomorphisme $f_a\in PL_+(\mathbb S_r)$ défini par :
\smallskip
$\left\{ \begin{array}{ccccc}
& f_a(x) = &x + a &\text{ si } &x \in [0, (q-2)a] \cr
& f_a(x) = &h_{BS}(x) &\text{ si } &\ \ \ \ \ \ \ x \in [(q-2)a, (q-1)a] \cr
& f_a(x) = &h_{BS}^{-1}(x) -(q-2)a &\text{ si } &x \in [(q-1)a, r], \cr
\end{array}\right. $
\smallskip
\noindent où $h_{BS} : [(q-2)a, (q-1)a] \rightarrow [(q-1)a, r]$ est l'application $PL_m$ donnée par la Proposition \ref{BS}, son existence est garantie par $\vert [(q-1)a, r] \vert - \vert [(q-2)a, (q-1)a]\vert = r-qa \in (m-1) \mathbb Z [\frac{1}{m}]$. On vérifie facilement que $f_a(0)=a$ et $f_a$ est d'ordre $q$.
\medskip
\noindent \textbf{Cas 2 :} $r-qa < 0$, choisissons $p \in \mathbb N$ de sorte que $ r - q(m^{-p} a) \geq 0$.
On considère $H_{BS} \in T_{r,m}$ tel que $H_{BS} ([0,a])=[0,m^{-p} a] $, son existence est assurée par le critère de Bieri-Strebel, puisque $m^{-p} a-a = (m^{-p} -1) a \in (m-1) \mathbb Z [\frac{1}{m}]$.
Finalement, l'application $PL_m$ définie par $\displaystyle f_a =H_{BS}^{-1} \circ f_ {m^{-p} a} \circ H_{BS}$ est d'ordre $q$ et satisfait $f_a(0)=a$.
\medskip
\noindent $(2) \Longleftrightarrow (3)$ est conséquence du calcul suivant :
$\displaystyle r-qa=r-q\frac{N_a}{m^{s_a}}=\frac{r m^{s_a}-qN_a}{m^{s_a}} \in (m-1)\mathbb Z [\frac{1}{m}] \Longleftrightarrow $
$r m^{s_a}-qN_a = r (m^{s_a}-1) + (r-qN_a)$ est un multiple de $m-1$ $\Longleftrightarrow $
$r-qN_a$ est un multiple de $m-1$. \end{proof}
Nous pouvons maintenant calculer le nombre de classes de conjugaison d'éléments d'ordre $q$ de nombre de rotation $\frac{1}{q}$ dans $T_{r,m}$. D'après les Propositions \ref{prop :3} et \ref{prop :4}, cette quantité est égale au nombre de classes modulo $(m-1)$ d'entiers $N$ tels que $r-qN $ est un multiple de $m-1$, nous affirmons que c'est $d=pgcd(m-1,q)$.
En effet, sous la condition $d$ divise $r$, le critère d'isomorphisme de Bieri-Strebel nous permet de supposer que $r=qu$. Posons $P= u-N$, on a $r-qN = q(u-N)=qP$, le problème se ramène à déterminer le nombre de classes modulo $(m-1)$ d'entiers $P$ tels que $qP$ est un multiple de $m-1$.
Puisque $m-1 = m_0.d$ et $q=q_0.d$ avec $m_0\wedge q_0=1$, l'entier $qP $ est un multiple de $m-1$ si et seulement si $q_0P $ est un multiple de $m_0$ et donc si et seulement si $P$ est un multiple de $m_0$. Par conséquent, il y a exactement $d$ tels entiers entre $0$ et $m-2$.
\subsubsection{Preuve de l'item C} D'après l'invariance par conjugaison topologique du nombre de rotation et la Remarque \ref{rema :2}, le nombre de classes de conjugaison d'éléments d'ordre $q$ dans $T_{r,m}$ est $\varphi(q)$.pgcd$(m-1, q)$, si pgcd $(m-1, q)$ divise $r$ et $0$ si non.
\section{Problèmes d'isomorphisme et de plongement entre groupes de Brown-Thompson. Preuve du Corollaire \ref{coro :1}}
\begin{proposition}\label{prop :5} Soit $m\in \mathbb N^{>1}$.
\begin{enumerate}
\item Parmi les groupes $T_{r,m}$, pour $0< r \leq m-1$, seul le groupe $T_{m-1,m}$ contient des éléments d'ordre quelconque. Ainsi, il n'existe pas de morphisme injectif $T_{m-1,m}\rightarrow T_{r,m}$, pour $0< r <m-1$.
\item Si $m_1-1$ possède un diviseur qui ne divise pas $m_2-1$ alors il n'existe pas de morphisme injectif $T_{1, m_2} \rightarrow T_{1, m_1}$.
\end{enumerate} \end{proposition}
\smallskip
\begin{proof} \
(1) Puisque pour tout $q\in \mathbb N^{>1}$, pgcd$(m-1,q)$ est un diviseur de $m-1$, le groupe $T_{m-1,m}$ contient des éléments de tout ordre. Réciproquement, si $0<r < m-1$, pgcd$(m-1,m-1)=m-1$ ne divise pas $r$ et il n'existe pas d'élément d'ordre $m-1$ dans $T_{r,m}$.
\smallskip
(2) Soit $d_1$ un diviseur de $m_1-1$ ne divisant pas $m_2-1$. D'une part, pgcd$(m_1-1,d_1)=d_1$ ne divise pas $r=1$ et $T_{1, m_1}$ ne contient pas d'élément d'ordre $d_1$.
D'autre part, pgcd$(m_2-1,d_2)=1$ divise $r=1$ et $T_{1, m_2}$ contient des éléments d'ordre $d_1$. Par conséquent, il n'existe pas de morphisme injectif $T_{1, m_2} \rightarrow T_{1, m_1}$.\end{proof}
{\bf Preuve du Corollaire \ref{coro :1}.}
Comme $T$ est simple, tout morphisme de $T$ dans $T_{r,m}$ est injectif ou trivial. D'après l'item (1) de la Proposition \ref{prop :5}, si $T$ s'injecte dans $T_{r,m}$, pour $0< r \leq m-1$, alors $r=m-1$. De plus, $T_{m-1,m}$ contient $m-1$ classes de conjugaison d'éléments d'ordre $m-1$ et de nombre de rotation $\frac{1}{m-1}$ alors que $T$ n'en contient qu'une. Ces groupes ne sont isomorphes que lorsque $m-1=1$, correspondant au seul cas où tout rationnel se réalise comme nombre de rotation d'une unique classe de conjugaison d'éléments d'ordre fini.
\bibliographystyle{alpha}
|
\section{Introduction and main result}
\subsection{Objective and background}
In this paper we want to consider a linear two-body operator which determines the critical temperature of a superconductive or superfluid system.
This linear operator was studied recently in connection with the influence of a constant magnetic field on the critical temperature \cite{FHaiLa}.
The analysis of this operator was significantly complicated by the unboundedness of the magnetic vector potential as well as the non-commutativity of
the components of the magnetic momentum. For this reason we want to present here the method of \cite{FHaiLa} in the simplified situation where the external field consists of an electric potential.
We have the following situation in mind. Two particles interact via a two body potential $-2V(x-y)$ and both particles are placed in an external electric potential $h^2 W(hx)$, where $h>0$ is a small parameter. Thus, the external field is weak of order $h^2$ and varies on the scale of order $1/h$, whereas both the strength and the scale of the interaction are of order one determined by $V$. The energy is given by the linearized BCS (Bardeen--Cooper--Schrieffer) functional at positive temperature $T = 1/\beta$.
Therefore we are interested in the infimum of the spectrum of the two-body operator
\begin{equation}
\label{eq:twobodyk}
\frac{p_x^2 +h^2 W(hx) +p_y^2 +h^2 W(hy) - 2\mu}{\tanh \left( \frac\beta 2 \left( p_x^2 +h^2 W(hx) - \mu \right)\right) + \tanh \left( \frac\beta 2 \left(p_y^2 +h^2 W(hy) - \mu \right) \right)} - V(x-y)
\end{equation}
acting in
$$
L_{\rm symm}^2(\mathbb{R}^3\times\mathbb{R}^3) = \left\{\alpha\in L^2(\mathbb{R}^3\times\mathbb{R}^3):\ \alpha(x,y)=\alpha(y,x) \ \text{for all}\ x,y\in\mathbb{R}^3 \right\}.
$$
Here $p_x=-{\rm i}\nabla_x$ and $p_y=-{\rm i}\nabla_y$. The interaction potential $-2V(x-y)$ between the two particles is assumed to be spherically symmetric, i.e., to depend only on the distance $|x-y|$. (We will also assume that the interaction potential is non-positive and the minus sign, as opposed to the more usual plus sign, will simplify some formulas.) Moreover, $\mu\in\mathbb{R}$ is the chemical potential. We are interested in the dependence of the operator on two parameters, namely, the inverse temperature $\beta>0$ and the scale ratio $h>0$. More precisely, we are interested in identifying regimes of temperatures $T=\beta^{-1}$ such that the infimum of the spectrum of the above operator is positive or negative for all sufficiently small $h>0$.
As we explained in detail in \cite{FHaiLa} and will repeat below, the motivation for this question comes from the BCS theory of superconductivity and the operator \eqref{eq:twobodyk} arises through the linearization of the Bogolubov--de Gennes equation around the normal state. Therefore, the question whether the infimum of the spectrum of the operator \eqref{eq:twobodyk} is positive or negative corresponds to the local stability or instability of the normal state. In that sense it is not hard to imagine that the BCS critical temperature corresponds to the value of $T$ for which the infimum of the spectrum of this operator is exactly zero.
To describe our main result we introduce the effective one-body operator
\begin{equation}
\label{eq:onebody}
\frac{(-{\rm i}\nabla_r)^2 - \mu}{\tanh \left( \frac\beta 2 \left( (-{\rm i}\nabla_r)^2 - \mu \right)\right)} - V(r)
\end{equation}
acting in
$$
L^2_{\rm symm}(\mathbb{R}^3) = \{ \alpha \in L^2(\mathbb{R}^3):\ \alpha(-r)=\alpha(r) \ \text{for all}\ r\in\mathbb{R}^3\} \,.
$$
Later on, we will see that the variable $r\in\mathbb{R}^3$ arises as the relative coordinate $r=x-y$ of the two particles at $x$ and $y$. We will \emph{assume} that the operator $|(-{\rm i}\nabla_r)^2-\mu| -V(r)$ has a negative eigenvalue. Then it is easy to see (see, e.g., \cite{HHSS}) that there is a unique $\beta_c\in (0,+\infty)$ such that the operator \eqref{eq:onebody} is non-negative for $\beta\leq\beta_c$ and has a negative eigenvalue for $\beta>\beta_c$. Let $T_c = \beta_c^{-1}$. Then our main result is, roughly speaking, that the infimum of the spectrum of the two-particle operator \eqref{eq:twobodyk} is negative for $T \leq T_c+c_0 h^2 + o(h^2)$ and positive for $T\geq T_c +c_0 h^2 - o(h^2)$. Here $c_0$ is a positive constant which we compute explicitly in terms of the zero-energy ground state of \eqref{eq:onebody} at $\beta=\beta_c$. (In fact, $c_0 = -T_c D_c$ with $D_c$ from \eqref{eq:dc}.) Thus, the external electric field $h^2 W(hx)$ changes the critical temperature by an amount $c_0 h^2 + o(h^2)$. Informally (that is, ignoring issues like the possible non-uniqueness of a critical temperature), this says that
$$
T_c(h)= T_c + c_0 h^2 + o(h^2) \,.
$$
The mathematical challenge of this problem is that low energy states of the two-particle operator \eqref{eq:twobodyk} exhibit a two-scale structure. As function of the relative coordinate $r=x-y$ and the center of mass coordinate $X=(x+y)/2$ they vary on a scale of order one with respect to $r$ and on a (much larger) scale of order $1/h$ with respect to $X$. The variation on the former scale is responsible for the leading order term $T_c$ for the critical temperature, whereas the variation on the latter scale is responsible for the subleading correction $c_0 h^2$. This subleading correction is determined by an effective linear Ginzburg--Landau functional which emerges on the macroscopic scale $1/h$ determined by the external potential. We hereby recover a similar result for the critical temperature as in the full non-linear BCS theory in \cite{FHSS2}. This is of course not unexpected since we deal with the second derivative around the normal state of the BCS functional.
The work \cite{FHSS2} relied on \cite{FHSS} where the Ginzburg--Landau functional was derived from the BCS functional close to the critical temperature by means of a rather intricate proof. In view of this, the goal of the present paper is twofold. First, we explain the strategy from \cite{FHaiLa} in a simpler setting, and second, we derive the linearized Ginzburg--Landau equation in a simpler way as in the full non-linear case \cite{FHSS}. One difference compared to the work \cite{FHSS,FHSS2} is the fact that we do not restrict ourselves to a finite box and therefore omit the periodicity assumptions. Further, we work in relative and center-of-mass coordinates which is natural in terms of the before mentioned two-scale structure.
As in \cite{FHaiLa} we will not work directly with the two-particle operator \eqref{eq:twobodyk}, but rather with its Birman--Schwinger version.
Before we describe the precise set-up of our analysis, we would like to stress that in this paper we work with the BCS functional and its linearization around the normal state. This should not be confused with what is often called the BCS Hamiltonian or the BCS model and which was investigated, for instance, by Haag, Thirring and Wehrl from the point of view of algebraic quantum field theory. The BCS Hamiltonian is a many-body Hamiltonian which corresponds to a regularization of a $\delta$ interaction. The BCS functional arises as an effective non-linear functional by restricting the BCS Hamiltonian to quasi-free states and dropping the direct and exchange terms. We do allow, however, for more general interaction potentials. It remains an open problem to understand from a mathematically rigorous point of view the relation between the BCS functional and many-body quantum mechanics. Nevertheless, our analysis leads to quantitative estimates which agree with physics.
\subsection{Model and main result}
Our model has the following ingredients.
\begin{assumption}\label{ass2}
(1) External electric potential $h^2 W(hx)$ such that $W \in W^{1,\infty}(\mathbb{R}^3)$.\\
(2) Inverse temperature $\beta = T^{-1}>0$\\
(3) Chemical potential $\mu\in\mathbb{R}$\\
(4) Non-negative, spherically symmetric interaction potential $V$ such that $V\in L^\infty(\mathbb{R}^3)$ and $|r|V\in L^\infty(\mathbb{R}^3)$
\end{assumption}
We recall that the Sobolev space $W^{1,\infty}(\mathbb{R}^3)$ consists of all bounded, Lipschitz continuous functions with a finite global Lipschitz constant.
The non-negativity assumption on $V$ is for technical convenience.
To simplify notation and since the precise meaning is always clear from the context, we use the same symbol $V$ also for the corresponding multiplication operators on $L^2_{\rm symm}(\mathbb{R}^3)$ (i.e., $(V\alpha)(r) = V (r)\alpha(r)$) and on $L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ (i.e., $(V\alpha)(x, y) = V (x-y)\alpha(x, y)$).
The corresponding single-particle Hamiltonian, acting in $L^2(\mathbb{R}^3)$, is defined by
\begin{equation}
\label{eq:hw}
\mathord{\mathfrak h}_W = p^2 +h^2 W(hx) - \mu \,.
\end{equation}
with the notation $p=-{\rm i}\nabla$. The locations of the two particles are represented by coordinates $x,y\in\mathbb{R}^3$. If we want to emphasize the variables on which the operators act, we write
$$
\mathord{\mathfrak h}_{W,x} = p_x^2 + h^2 W(hx) -\mu \,,
\qquad
\mathord{\mathfrak h}_{W,y} = p_y^2 + h^2 W(hy) -\mu \,.
$$
As in \cite{FHaiLa} we introduce a function $\Xi_\beta:\mathbb{R}^2\to\mathbb{R}$ by
$$
\Xi_\beta(E,E') := \frac{\tanh\frac{\beta E}{2}+\tanh\frac{\beta E'}{2}}{E+E'}
$$
if $E+E'\neq 0$ and $\Xi_\beta(E,-E) = (\beta/2)/\cosh^2(\beta E/2)$. Since the operators $\mathord{\mathfrak h}_{W,x}$ and $\mathord{\mathfrak h}_{W,y}$ commute, we can define the operator
$$
L_{T,W} = \Xi_\beta(\mathord{\mathfrak h}_{W,x},\mathord{\mathfrak h}_{W,y}) \,.
$$
We will always consider this operator in the Hilbert space $L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$. Note that, with this notation, the operator in \eqref{eq:twobodyk} can be written as $L_{T,W}^{-1}-V$.
Next, in order to formulate our assumption on the critical temperature, we introduce the function $\chi_\beta:\mathbb{R}\to\mathbb{R}$ by
$$
\chi_\beta(E) := \frac{\tanh\frac{\beta E}2}{E}
$$
and set $\chi_\infty(E):=|E|^{-1}$. We consider the compact operator
$$
V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2}
$$
in $L^2_{\rm symm}(\mathbb{R}^3)$, where
$$
p_r = -{\rm i}\nabla_r
$$
denotes the momentum operator. (The operator $\chi_\beta(p_r^2-\mu)$ is denoted by $K_T^{-1}$ in \cite{HHSS} and several works thereafter.)
\begin{assumption}\label{ass0}
$\sup\spec V^{1/2}\chi_\infty(p_r^2-\mu)V^{1/2}>1$.
\end{assumption}
Since $\beta\mapsto\chi_\beta(E)$ is strictly increasing for each fixed $E\in\mathbb{R}$, Assumption \ref{ass0} implies that there is a unique $\beta_c\in (0,\infty)$ such that
\begin{align*}
\sup \spec V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} & \leq 1 \qquad\text{if}\ \beta\leq\beta_c \,,\\
\sup \spec V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} & > 1 \qquad\text{if}\ \beta>\beta_c \,.
\end{align*}
We set $T_c=\beta_c^{-1}$. Note that the operator $V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2}$ has eigenvalue $1$.
\begin{assumption}\label{ass1}
The eigenvalue $1$ of the operator $V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2}$ is simple.
\end{assumption}
We denote by $\phi_*$ a normalized eigenfunction of $V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2}$ corresponding to the eigenvalue $1$ which, by assumption, is unique up to a phase. Since $p_r^2$ and $V$ are real operators, so is $V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2}$ and we can assume that $\phi_*$ is real-valued.
The spherical symmetry of $V$ from Assumption \ref{ass2} and the non-degeneracy from Assumption \ref{ass1} imply that $\phi_*$ is spherically symmetric.
From a physics point of view, Assumption \ref{ass1} restricts us to potentials giving rise to s-wave superconductivity. It is known that this assumption is fulfilled for a large class of potentials, including those which have a non-negative Fourier transform \cite{HaSe}. For partial results in the case where Assumption \ref{ass1} is violated, we refer to \cite{FrLe}.
As the final preliminary before stating our main result, we will introduce some constants. They are defined in terms of the auxiliary functions
\begin{align}\label{eq:auxiliary}
g_0(z) & = \frac{\tanh(z/2)}{z} \,,\notag \\
g_1(z) & = \frac{e^{2z}-2ze^z-1}{z^2(e^z+1)^2} = \frac1{2z^2} \frac{\sinh z-z}{\cosh^2(z/2)}\,, \notag \\
g_2(z) & = \frac{2e^z (e^z-1)}{z(e^z+1)^3} = \frac1{2z} \frac{\tanh(z/2)}{\cosh^2(z/2)}\,,
\end{align}
as well as the function
\begin{equation}
\label{eq:t}
t(p) := \|\chi_{\beta_c}((-{\rm i}\nabla_r)^2-\mu) V^{1/2}\phi_*\|^{-1}\ 2 (2\pi)^{-3/2} \int_{\mathbb{R}^3} dx\, V(x)^{1/2} \phi_*(x) e^{-{\rm i} p\cdot x} \,.
\end{equation}
(The prefactor in front of the integral is irrelevant for us and only introduced for consistency with the definition in \cite{FHSS2}.) We now set
\begin{align}\label{eq:glcoeff}
\Lambda_0 & := \frac{\beta_c^2}{16} \int_{\mathbb{R}^3} \frac{dp}{(2\pi)^3}\, |t(p)|^2 \left(g_1(\beta_c(p^2-\mu)) + \frac23 \beta_c p^2 g_2(\beta_c(p^2-\mu)) \right) \,,\\
\Lambda_1 & := \frac{\beta_c^2}{4} \int_{\mathbb{R}^3} \frac{dp}{(2\pi)^3}|t(p)|^2\, g_1(\beta_c(p^2-\mu)) \, \,, \\
\Lambda_2 & := \frac{\beta_c}8 \int_{\mathbb{R}^3} \frac{dp}{(2\pi)^3}\, |t(p)|^2 \cosh^{-2}(\beta_c(p^2-\mu)/2) \,.
\end{align}
The constants $\Lambda_0$ and $\Lambda_2$ are positive (for a proof for $\Lambda_0$ see \cite{FHSS}). Note that the quotient $\Lambda_0/\Lambda_2$, which will appear in our main result, has the dimension of an inverse temperature.
We set
\begin{equation}
\label{eq:dc}
D_c := \frac{\Lambda_0}{\Lambda_2}\, \inf\spec \left( p_X^2 + \frac {\Lambda_1}{\Lambda_0} W(X) \right),
\end{equation}
where the operator on the right side is considered as an operator in $L^2(\mathbb{R}^3)$ and where $p_X=-{\rm i}\nabla_X$.
The following is our main theorem.
\begin{theorem}\label{main}
Under assumptions \ref{ass2}, \ref{ass0} and \ref{ass1} the following holds.
\begin{enumerate}
\item[(1)] Let $0<T_1<T_c$. Then there are constants $h_0>0$ and $C>0$ such that for all $0< h \leq h_0$ and all $T_1\leq T< T_c (1 - h^2 D_c) -C h^3$ one has
$$
\inf_{\Phi} \langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle <0 \,.
$$
\item[(2)] There are constants $h_0>0$ and $C>0$ such that for all $0< h \leq h_0$ and all $T> T_c(1- h^2 D_c) + C h^{5/2}$ one has
$$
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle >0 \,,
$$
unless $\Phi=0$.
\end{enumerate}
\end{theorem}
\begin{remark}
Let us restate this theorem in a heuristic form. Informally, we think of the critical temperature $T_c(h)$ as the value of the parameter $T$ such that
$$
\sup \, {\rm spec} V^{1/2} L_{T,W} V^{1/2} = 1 \,.
$$
This is not a precise definition because in contrast to the one-body operator $V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} $ it is not clear whether the two-body operator $V^{1/2} L_{T,W} V^{1/2}$, or at least the infimum of its spectrum, is monotone in $T$ and therefore the uniqueness of the value of $T$ such that $\sup \, {\rm spec} V^{1/2} L_{T,W} V^{1/2} = 1$ is not guaranteed. Ignoring this issue, as well as some technicalities connected with $T_1$ in part (1) which we discuss below, we see that our main theorem says that
$$
T_c(h)= T_c(1 - D_c h^2) + o(h^2) \,.
$$
Note that concerning the potential non-uniqueness of the critical temperature the theorem implies that, if it occurs at all, it occurs only in a temperature interval of size $o(h^2)$.
\end{remark}
\begin{remark}
Observe that $D_c$ can have either sign, depending on $W$. Thus, an external electric field $h^2 W(hx)$ can both raise and lower the critical temperature by an amount of order $h^2$. This is in contrary to the influence of magnetic fields where the critical temperature always goes down.
\end{remark}
\begin{remark}
Let us compare our results here with those in \cite{FHSS2} where we also computed the shift of the critical temperature. The results of \cite{FHSS2} concern a definition of the critical temperature in the non-linear BCS functional, whereas here we base our definition of critical temperature on a quadratic approximation to the BCS functional around the normal state. Both notions lead to the same result to order $h^2$. A minor difference is that the setting in \cite{FHSS2} is a finite sample whereas here we work on the whole space. Technically, the methods of proof in the two approaches are quite different.
\end{remark}
\begin{remark}
The assumption in part (1) that the temperature is bounded away from zero is probably only technical. Note, however, that our result is valid for arbitrarily small $T_1>0$, as long as it is uniform in $h$. The reason for this restriction is that our expansions diverge as the temperature goes to zero. Remarkably, there is no such restriction in part (2) of the theorem.
\end{remark}
\begin{remark}
Let us emphasize that our definition of the critical temperature $T_c$ coincides with that in \cite{HHSS} (and therefore with that in \cite{FHSS,FHSS2}) and that our Assumptions \ref{ass0} and \ref{ass1} coincides with \cite[Assumption 2]{FHSS}. This is a consequence of the Birman--Schwinger principle, which also implies that, if $\alpha_*$ denotes a normalized, real-valued eigenfunction of the operator \eqref{eq:onebody}, then
$$
V^{1/2}\alpha_* = \pm \|\chi_{\beta_c}(({\rm i}\nabla_r)^2-\mu) V^{1/2}\phi_*\|^{-1} \phi_* \,.
$$
(To get the normalization constant, we apply $\chi_{\beta_c}((-{\rm i}\nabla_r)^2-\mu) V^{1/2}$ to both sides and use the equation for $\alpha_*$ and its normalization.)
\end{remark}
\begin{remark}
In the physics literature the two-body interaction $V$ is usually replaced by a local contact interaction. With this modification the linear two-body operator \eqref{eq:twobodyk} was studied earlier in the literature in particular in the school by Gorkov and co-authors. In the presence of a constant magnetic fields this operator was used by Werthamer et al. \cite{HeWe, WeHeHo} in their study of the upper critical field. This approach was later extended in different directions, see e.g., \cite{SchSch, La, La2}. In particular, \cite{EL1} relaxed the local approximation and was an initial motivation for our work~\cite{FHaiLa}.
\end{remark}
\subsection{Connection to BCS theory}
In this subsection we repeat our argument from from \cite{FHaiLa} and describe how the two-body operators \eqref{eq:twobodyk} and $L_{T,W}$ arise in a problem in superconductivity. Our purpose here is to give a motivation and our presentation in this subsection will be informal. For background and references on the mathematical study of BCS theory we refer to our earlier works \cite{HHSS,FHNS,HaSe,FHSS,FHSS2,FrLe,FHaiLa} and, in particular, to the review \cite{HSreview}.
We consider a superconducting sample occupying all of $\mathbb{R}^3$ at inverse temperature $\beta>0$ and chemical potential $\mu\in\mathbb{R}$. The particles interact through a two-body potential $-2V(x-y)$ and are placed in an external electric field with potential $h^2 W(hx)$. In BCS theory the state of a system is described by two operators $\gamma$ and $\alpha$ in $L^2(\mathbb{R}^3)$, representing the one-body density matrix and the Cooper pair wave function, respectively. The operator $\gamma$ is assumed to be Hermitian and the operator $\alpha$ is assumed to satisfy $\alpha^* =\overline{\alpha}$, where for a general operator $A$ we write $\overline A = \mathcal C A \mathcal C$ with $\mathcal C$ denoting complex conjugation. Moreover, it is assumed that
$$
0\leq \begin{pmatrix}
\gamma & \alpha \\ \overline\alpha & 1-\overline\gamma
\end{pmatrix} \leq 1 \,.
$$
In an equilibrium state the operators $\gamma$ and $\alpha$ satisfy the (non-linear) Bogolubov--de Gennes equation
\begin{align*}
& \begin{pmatrix}
\gamma & \alpha \\ \overline\alpha & 1-\overline\gamma
\end{pmatrix}
= \left( 1+ \exp\left( \beta H_{\Delta_{V,\alpha}}\right) \right)^{-1} \,, \\
& \qquad\text{where}\qquad
\Delta_{V,\alpha}(x,y) = -2V(x-y)\alpha(x,y)
\qquad\text{and}\qquad
H_{\Delta} =
\begin{pmatrix}
\mathord{\mathfrak h}_W & \Delta \\ \overline\Delta & -\overline\mathord{\mathfrak h}_W
\end{pmatrix} \,.
\end{align*}
Here $\Delta$ is considered as an integral operator in $L^2(\mathbb{R}^3)$ with integral kernel $\Delta(x,y)$. Moreover, $\mathord{\mathfrak h}_W $ is the one-particle operator introduced in \eqref{eq:hw}.
Note that one solution of the equation is $\gamma= (1+\exp(\beta\mathord{\mathfrak h}_W))^{-1}$ and $\alpha=0$. This is the \emph{normal state}. We are interested in the local stability of this solution and therefore will linearize the equation around it.
It is somewhat more convenient to write the equation in the equivalent form
$$
\begin{pmatrix}
\gamma & \alpha \\ \overline\alpha & 1-\overline\gamma
\end{pmatrix}
= \frac12 - \frac12 \tanh\left( \frac{\beta}2 H_{\Delta_{V,\alpha}}\right) \,.
$$
Then, in view of the partial fraction expansion (also known as Mittag--Leffler series)
$$
\tanh z = \sum_{n\in\mathbb{Z}} \frac{1}{z -{\rm i} (n+1/2)\pi}
$$
(where we write $\sum_{n\in\mathbb{Z}}$ short for $\lim_{N\to\infty} \sum_{n=-N}^N$ for conditionally convergent sums like this one; convergence becomes manifest by combining the $+n$ and $-n$ terms),
$$
\tanh\left( \frac{\beta}2 H_{\Delta}\right) = - \frac2\beta \sum_{n\in\mathbb{Z}} \frac{1}{{\rm i}\omega_n - H_\Delta}
$$
with the \emph{Matsubara frequencies}
\begin{equation}
\label{eq:matsubara}
\omega_n = \pi(2n+1)T \,, \qquad n\in\mathbb{Z} \,.
\end{equation}
Using this formula we can expand the operator $\tanh(\beta H_\Delta/2)$ in powers of $\Delta$. Since
\begin{align*}
\frac{1}{{\rm i}\omega_n - H_\Delta} & = \frac{1}{{\rm i}\omega_n - H_0} + \frac{1}{{\rm i}\omega_n - H_0} \begin{pmatrix} 0 & \Delta \\ \overline\Delta & 0 \end{pmatrix} \frac{1}{{\rm i}\omega_n - H_0} + \ldots \\
& = \begin{pmatrix} ({\rm i}\omega_n - \mathord{\mathfrak h}_W)^{-1} & 0 \\ 0 & ({\rm i}\omega_n + \overline\mathord{\mathfrak h}_W)^{-1} \end{pmatrix} \\
& \qquad + \begin{pmatrix} 0 & ({\rm i}\omega_n - \mathord{\mathfrak h}_W)^{-1}\Delta ({\rm i}\omega_n + \overline\mathord{\mathfrak h}_W)^{-1} \\ ({\rm i}\omega_n + \overline\mathord{\mathfrak h}_W)^{-1}\overline \Delta ({\rm i}\omega_n - \mathord{\mathfrak h}_W)^{-1} & 0\end{pmatrix} + \ldots \,,
\end{align*}
the Bogolubov--de Gennes equation for the Cooper pair wave function becomes
$$
\alpha = \frac1\beta \sum_{n\in\mathbb{Z}} ({\rm i}\omega_n - \mathord{\mathfrak h}_W)^{-1}\Delta_{V,\alpha} ({\rm i}\omega_n + \overline\mathord{\mathfrak h}_W)^{-1} + \ldots \,,
$$
where $\ldots$ stands for terms that are higher order in $\Delta_{V,\alpha}$. The key observation now is that
\begin{equation}
\label{eq:ltsum}
\frac1\beta \sum_{n\in\mathbb{Z}} ({\rm i}\omega_n - \mathord{\mathfrak h}_W)^{-1}\Delta_{V,\alpha} ({\rm i}\omega_n + \overline\mathord{\mathfrak h}_W)^{-1} = L_{T,W} V\alpha \,.
\end{equation}
(Here $V\alpha$ on the right side is considered as a two-particle wave function, defined by $(V\alpha)(x,y)=V(x-y)\alpha(x,y)$.) This identity follows by writing
\begin{equation}
\label{eq:zetaidentity}
- \frac2\beta \sum_{n\in\mathbb{Z}} ({\rm i}\omega_n - E)^{-1} ({\rm i}\omega_n + E')^{-1} =
- \frac{2}{\beta} \sum_{n\in\mathbb{Z}} \frac{1}{E+E'} \left( \frac{1}{{\rm i}\omega_n - E} - \frac{1}{{\rm i}\omega_n + E'} \right)
\end{equation}
and using the partial fraction expansion of $\tanh$ to recognize the right side as $\Xi_\beta(E,E')$.
Thus, the linearized Bogolubov--de Gennes equation becomes
$$
\alpha = L_{T,W} V\alpha \,.
$$
There are two ways to make the operator appearing in this equation self-adjoint. The first one is to apply the operator $L_{T,W}^{-1}$ to both sides and to subtract $V\alpha$. In this way we obtain the operator \eqref{eq:twobodyk}. The other way is to multiply both sides of the equation by $V^{1/2}$, to subtract $V^{1/2} L_{T,W} V\alpha$ and to call $\Phi=V^{1/2}\alpha$. In this way we arrive at the operator $1-V^{1/2} L_{T,W} V^{1/2}$ which appears in our main result, Theorem \ref{main}.
The upshot of this discussion is that positivity of the operator \eqref{eq:twobodyk} (or, equivalently, of $1-V^{1/2} L_{T,W} V^{1/2})$ corresponds to local stability of the normal state and the existence of negative spectrum of \eqref{eq:twobodyk} corresponds to local instability. If we define two critical local temperatures $\overline{T_c^{\rm loc}(h)}$ as the smallest temperature above which the normal state is always stable and $\underline{T_c^{\rm loc}(h)}$ as the largest temperature below which the normal state is never stable, then our theorems says that both $\overline{T_c^{\rm loc}(h)}$ and $\underline{T_c^{\rm loc}(h)}$ are equal to $T_c(1 - D_c h^2) + O(h) $ as $h\to 0$.
\subsection*{Acknowledgements}
We thank Edwin Langmann who initiated and co-authored our previous work \cite{FHaiLa} which forms the basis of the present paper.
We further thank Robert Seiringer and Jan Philip Solovej for our long lasting collaboration on BCS theory.
Further, partial support by the U.S. National Science Foundation through grant DMS-1363432 (R.L.F.) is acknowledged.
\section{A representation formula for the operator $L_{T,W}$}
In this section we derive a useful representation formula for the operator $L_{T,W}$ as a sum over contributions from the individual Matsubara frequencies $\omega_n$ from \eqref{eq:matsubara}. Moreover, we express the formula in terms of center of mass and relative coordinates,
$$
r = x-y \,,
\qquad
X= (x+y)/2 \,.
$$
We recall that the corresponding momenta are denoted by $p_r=-{\rm i}\nabla_r$ and $p_X = -{\rm i}\nabla_X$.
Our starting point is \eqref{eq:ltsum}, which can be written in the form
\begin{align}\label{eq:repr}
\left( L_{T,W} \Delta \right)(x,y)
= -\frac2\beta \sum_{n\in\mathbb{Z}} \left( \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_W} \Delta \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_W}\right) (x,y) \,.
\end{align}
(Here we used the fact that $\mathord{\mathfrak h}_W=\overline{\mathord{\mathfrak h}_W}$.) This formula means that as an operator on $L^2(\mathbb{R}^3 \times \mathbb{R}^3)$ we have
$$
L_{T,W} =- \frac2\beta \sum_{n\in\mathbb{Z}} \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{W,x}} \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{W,y}} \,.
$$
The strategy now will be to expand the operators $1/({\rm i}\omega_n\mp\mathord{\mathfrak h}_W)$ with respect to $W$. Clearly the leading term is
$$
L_{T,0} = - \frac2\beta \sum_{n\in\mathbb{Z}} \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{0,x}} \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,y}}
$$
and the subleading correction is $h^2$ times
\begin{align*}
N_{T,W}:= - \frac2\beta \sum_{n\in\mathbb{Z}} & \left( - \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{0,x}} \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,y}} W(hy)\frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,y}} \right. \\
& \quad \left. + \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,x}} W(hx)\frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,x}} \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{0,y}} \right).
\end{align*}
The following lemma justifies this formal expansion.
\begin{lemma}\label{lemmadiffw0}
As an operator on $L^2(\mathbb{R}^3 \times \mathbb{R}^3)$ we have
$$
\| L_{T,W} - L_{T,0} \| \lesssim \beta^3 h^2
$$
and
$$
\| L_{T,W} - L_{T,0} -h^2 N_{T,W} \| \lesssim \beta^5 h^4
$$
\end{lemma}
\begin{proof}
Using the resolvent identity we write
\begin{align*}
\frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{W,x}} \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{W,y}}
& = \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{0,x}} \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,y}} \\
& \quad - \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{0,x}} \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{W,y}} h^2 W(hy)\frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,y}} \\
& \quad + \frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{W,x}} h^2 W(hx)\frac 1{{\rm i}\omega_n+ \mathord{\mathfrak h}_{0,x}} \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_{W,y}} \,.
\end{align*}
The first term on the right side, when summed with respect to $n$, corresponds to the operator $L_{T,0}$. In the remaining terms we use $W\in L^\infty(\mathbb{R}^3)$ and bound each resolvent in norm by $|\omega_n|^{-1}$. The resulting bound is summable with respect to $n$. This proves the first bound. For the proof of the second bound we expand the resolvents once more.
\end{proof}
In the remainder of this section we will do two things, namely bring the operator $L_{T,0}$ in a more explicit form and extract the leading term from the operator $N_{T,W}$. While in Lemma \ref{lemmadiffw0} we considered $L_{T,W}$ as an operator on $L^2(\mathbb{R}^3 \times \mathbb{R}^3)$, we will from now on restrict it to the subspace $L^2_{\rm symm}(\mathbb{R}^3 \times \mathbb{R}^3)$.
In order to investigate the operator $L_{T,0}$ we denote by $g^{z}$ the integral kernel of $1/(z- \mathord{\mathfrak h}_0)$, that is,
$$
\frac 1{z- \mathord{\mathfrak h}_0} (x,x') = g^z(x-x') \,.
$$
Using center-of-mass and relative coordinates we can rewrite \eqref{eq:repr} as
\begin{align*}
\left(L_{T,0} \Delta\right)(X+\frac r2,X-\frac r2)
& = -\frac2\beta \sum_{n\in\mathbb{Z}} \iint_{\mathbb{R}^3\times\mathbb{R}^3} dY ds \, \Delta(Y+\frac s2,Y-\frac s2) \\
& \qquad\qquad\qquad \times g^{{\rm i}\omega_n}(X-Y+\frac{r-s}{2}) g^{-{\rm i}\omega_n}(X-Y - \frac{r-s}{2}) \\
& = \iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ ds \, k_{T}(Z,r-s) \Delta(X-Z+\frac s2,X-Z-\frac s2)
\end{align*}
with
$$
k_{T}(Z,\rho) := -\frac2\beta \sum_{n\in\mathbb{Z}} g^{{\rm i}\omega_n}(Z+\frac{\rho}{2}) g^{-{\rm i}\omega_n}(Z - \frac{\rho}{2}) \,.
$$
Next, we use the fact that $\psi(X - Z) = (e^{-{\rm i} Z\cdot p_X} \psi)(X)$ to write
\begin{align}
\label{eq:reprproof}
\left(L_{T,0} \Delta\right)(X+\frac r2,X-\frac r2)
= \iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ ds \, k_{T}(Z,r-s) \left( e^{-{\rm i} Z\cdot p_X} \Delta\right)(X+\frac s2,X-\frac s2) \,.
\end{align}
We claim that in this formula we can replace $e^{-{\rm i} Z\cdot p_X}$ by $\cos(Z\cdot p_X)$. To do so, we change variables $Z\mapsto-Z$, $r\mapsto-r$ and $s\mapsto-s$ and use $\Delta(x,y)=\Delta(y,x)$ and $k_{T}(-Z,-r+s)=k_{T}(Z,r-s)$ in order to obtain the same formula as in \eqref{eq:reprproof}, but with $e^{-{\rm i} Z\cdot p_X}$ replaced by $e^{+{\rm i} Z\cdot p_X}$. Adding the two formulas we finally find
\begin{equation}\label{defpX}
\left( L_{T,0}\Delta\right)(X+\frac r2,X-\frac r2) = \iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ\,ds\, k_{T}(Z,r-s) \left(\cos(Z\cdot p_X)\Delta\right)(X+\frac s2,X-\frac s2).
\end{equation}
Next, we derive a convenient representation of $k_{T}(Z,\rho)$. Setting $\ell=p+q$ and $k=(p-q)/2$ and recalling \eqref{eq:ltsum} and \eqref{eq:zetaidentity}, we calculate
\begin{align}
\label{eq:semest1}
k_{T}(Z,\rho) & = - \frac2\beta \sum_{n\in\mathbb{Z}} \iint_{\mathbb{R}^3\times\mathbb{R}^3} \frac{dp}{(2\pi)^3}\,\frac{dq}{(2\pi)^3} \frac{e^{{\rm i} p\cdot(Z+\frac\rho2)}}{{\rm i}\omega_n-p^2+\mu} \frac{e^{{\rm i} q\cdot(Z-\frac\rho2)}}{{\rm i}\omega_n+q^2-\mu} \notag \\
& = \iint_{\mathbb{R}^3\times\mathbb{R}^3} \frac{dp}{(2\pi)^3}\,\frac{dq}{(2\pi)^3} L(p,q) e^{{\rm i} p\cdot(Z+\frac\rho2)+{\rm i} q\cdot(Z-\frac\rho2)} \notag \\
& = \iint_{\mathbb{R}^3\times\mathbb{R}^3} \frac{d\ell}{(2\pi)^3}\,\frac{dk}{(2\pi)^3} L(k+\frac\ell2,k-\frac\ell2) e^{{\rm i}\ell\cdot Z +{\rm i} k\cdot\rho}
\end{align}
with
\begin{equation}
\label{eq:semestl}
L(p,q) : = \frac{\tanh\frac{\beta(p^2-\mu)}2+\tanh\frac{\beta(q^2-\mu)}2}{p^2-\mu+q^2-\mu} \,.
\end{equation}
Let us explain the intuition for the following. Since the external field is varying on the scale $1/h$, which is much larger than the typical distance of between the particles, each momentum $p_X$ will pick up an additional factor of $h$. Therefore, we expect the leading term in \eqref{defpX} to be given by the corresponding operator with $\cos(Z\cdot p_X)$ replaced by $1$. We will justify this approximation in the following lemma. The next order, namely $-(1/2) (Z\cdot p_X)^2$, which will ultimately give rise to the Laplacian in Ginzburg--Landau theory, will be discussed in the following section.
In order to compute the right side of \eqref{defpX} with $\cos(Z\cdot p_X)$ replaced by $1$, we first compute, using \eqref{eq:semest1},
\begin{align}\label{eq:deltafcn}
\int_{\mathbb{R}^3} dZ \, k_{T}(Z,\rho) = \int_{\mathbb{R}^3} \frac{dk}{(2\pi)^3} L(k,k) e^{{\rm i} k\cdot \rho} \,.
\end{align}
This implies that
$$
\iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ\,ds\, k_{T}(Z,r-s) \Delta(X+\frac s2,X-\frac s2) = \left( \chi_\beta(p_r^2-\mu) \Delta \right)(X+r/2,X-r/2) \,,
$$
that is,
\begin{equation}\label{Lt0exp}
L_{T,0} = \chi_\beta(p_r^2 - \mu) - \int_{\mathbb{R}^3} dZ\, k_{T}(Z) \left(1- \cos(Z\cdot p_X)\right) \,,
\end{equation}
where $k_T(Z)$ denotes the operator in $L^2_{\rm symm}(\mathbb{R}^3)$ with integral kernel $k_T(Z,r-s)$.
We now quantify the replacement of $\cos(Z\cdot p_X)$ by $1$.
\begin{lemma}\label{Ltleading}
$$
\left\| \left( L_{T,0} - \chi_\beta(p_r^2 - \mu) \right)\Delta \right\| \lesssim \beta^3 \left\| p_X^2 \Delta \right\|
$$
\end{lemma}
\begin{proof}
We have to bound the integral on the right side of \eqref{Lt0exp}. For this we consider a single term in the definition of $k_T(Z,\rho)$. For fixed $r\in\mathbb{R}^3$ we estimate using Minkowski's inequality
\begin{align*}
& \left( \int_{\mathbb{R}^3} dX \left| \iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ\,ds\, g^{{\rm i}\omega_n}(Z+(r-s)/2) g^{-{\rm i}\omega_n}(Z-(r-s)/2) \right.\right. \\
& \qquad \qquad \times \left.\left. \phantom{\int_{\mathbb{R}^6}} \!\!\!\!\!\!\!\!\!\! \left((1-\cos(Z\cdot p_X))\Delta\right)(X+s/2,X-s/2) \right|^2 \right)^{1/2} \\
& \leq \iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ\,ds \left| g^{{\rm i}\omega_n}(Z+(r-s)/2) g^{-{\rm i}\omega_n}(Z-(r-s)/2) \right| \\
& \qquad\qquad \times \left( \int_{\mathbb{R}^3} dX \left| \left((1-\cos(Z\cdot p_X))\Delta\right)(X+s/2,X-s/2) \right|^2 \right)^{1/2}.
\end{align*}
Now we bound for fixed $Z,s\in\mathbb{R}^3$
\begin{align*}
& \left( \int_{\mathbb{R}^3} dX \left| \left((1-\cos(Z\cdot p_X))\Delta\right)(X+s/2,X-s/2) \right|^2 \right)^{1/2} \\
& \leq \left\| \frac{1-\cos(Z\cdot p_X)}{(Z\cdot p_X)^2} \right\| \left( \int_{\mathbb{R}^3} dX \left| \left((Z\cdot p_X)^2\Delta\right)(X+s/2,X-s/2) \right|^2 \right)^{1/2} \\
& \lesssim |Z|^2 t(s)
\end{align*}
where
$$
t(s) := \left( \int_{\mathbb{R}^3} dX \left| \left(p_X^2\Delta\right)(X+s/2,X-s/2) \right|^2 \right)^{1/2} \,.
$$
Thus, the quantity we are interested in is bounded by a constant times
$$
\iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ\,ds \left| g^{{\rm i}\omega_n}(Z+(r-s)/2) g^{-{\rm i}\omega_n}(Z-(r-s)/2) \right| |Z|^2 t(s) \,.
$$
Using
$$
|Z|^2 \leq \frac12 \left( \left| Z + \frac{r-s}{2} \right|^2 + \left| Z - \frac{r-s}{2}\right|^2 \right)
$$
we can bound the above quantity by
$$
\frac12 \left( \left( \left( |\cdot|^2 g^{{\rm i}\omega_n}\right)\ast g^{-{\rm i}\omega_n} * t \right)(r) + \left( g^{{\rm i}\omega_n} \ast \left( |\cdot|^2 g^{-{\rm i}\omega_n}\right) * t \right)(r) \right).
$$
The $L^2$ norm of this term with respect to $r$ is bounded according to Young's convolution inequality by
$$
\frac12 \left( \left\| |\cdot|^2 g^{{\rm i}\omega_n} \right\|_1 \left\| g^{-{\rm i}\omega_n} \right\|_1 + \left\| g^{{\rm i}\omega_n} \right\|_1 \left\| |\cdot|^2 g^{-{\rm i}\omega_n} \right\|_1 \right) \|t\|_2 \,.
$$
By \cite[Lemma 9]{FHaiLa} this expression is summable with respect to $n$ and therefore the left side in the lemma is bounded by a constant times $\|t\|_2 = \left\| p_X^2 \Delta \right\|$, as claimed.
\end{proof}
This concludes our discussion of the leading term $L_{T,0}$. We now aim at extracting the leading term from the operator $N_{T,W}$ and we concentrate on a term of the form
$$
\iint_{\mathbb{R}^3\times\mathbb{R}^3} dx' dy'\, \left( \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_0} W(h\cdot) \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_0} \right) (x,x') \frac 1{-{\rm i}\omega_n- \mathord{\mathfrak h}_0} (y,y') \Delta(x',y') \,.
$$
We introduce again center of mass and relative coordinates $X=(x+y)/2$, $r=x-y$, $Y=(x'+y')/2$ and $s=x'-y'$. In order to obtain concise expressions we introduce the abbreviation $$ \zeta_X^r = X + r/2, \quad \zeta_Y^{-s} = Y-s/2,$$
where the second term should just show the consistency of the symbol. With these definitions we obtain
\begin{align}
&\int_{\mathbb{R}^6} dx' dy' \left(\frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_0} W(h\cdot) \frac 1{{\rm i}\omega_n - \mathord{\mathfrak h}_0} \right) (x,x') \frac 1{-{\rm i}\omega_n- \mathord{\mathfrak h}_0} (y,y') \Delta(x',y') \nonumber \\
&=\int_{\mathbb{R}^9} dY ds dz' \, g^{{\rm i}\omega_n}(\zeta_X^r-z') W(h z') g^{{\rm i}\omega_n} (z' - \zeta_Y^{s} ) g^{-{\rm i}\omega_n}(\zeta_{X-Y}^{s-r} )\Delta (\zeta_Y^{s} ,\zeta_Y^{-s} ) \nonumber\\
& = \int_{\mathbb{R}^9} dY ds dz \, g^{{\rm i}\omega_n}(\frac r2-z) W(h X + hz) g^{{\rm i}\omega_n}(z + \zeta_{X-Y}^{-s} ) g^{-{\rm i}\omega_n} (\zeta_{X-Y}^{s-r} )\Delta (\zeta_Y^{s} ,\zeta_Y^{-s} ) \nonumber\\
& = \int_{\mathbb{R}^9} dZ ds dz \, g^{{\rm i}\omega_n}(\frac r2-z) W(h X + hz) g^{{\rm i}\omega_n}(z + \zeta_Z^{-s} ) g^{-{\rm i}\omega_n}(\zeta_Z^{s-r} )\Delta (\zeta_{X-Z}^{s} ,\zeta_{X-Z}^{-s})\nonumber \\ \label{Werr}
&= \int_{\mathbb{R}^9} dZ ds dz \, g^{{\rm i}\omega_n}(\frac r2-z) W(h X + hz) g^{{\rm i}\omega_n} (z + \zeta_Z^{-s}) g^{-{\rm i}\omega_n}(\zeta_{Z}^{s-r}) \left( e^{-{\rm i} Z\cdot p_X} \Delta\right)(\zeta_{X}^{s},\zeta_{X}^{-s}) \,,
\end{align}
where in the last step we used again
\begin{align*}
\alpha (\zeta_{X-Z}^{s} ,\zeta_{X-Z}^{-s}) & = \alpha(X-Z +s/2, X-Z - s/2) \\
&= \left( e^{-{\rm i} Z\cdot p_X}\alpha\right)(X + s/2, X-s/2) \\
&= \left( e^{-{\rm i} Z\cdot p_X} \alpha\right)(\zeta_{X}^{s},\zeta_{X}^{-s}) \,.
\end{align*}
We claim that to leading order we can replace $W(hX+hz)$ in this integral by $W(hX)$. Therefore we define
\begin{align}
\left(\tilde N_{T,W} \Delta\right)(\zeta_X^r, \zeta_X^{-r}) := W(hX) \iint_{\mathbb{R}^3\times\mathbb{R}^3} dZ ds \, \ell_{T}(Z,r-s) \left(e^{-{\rm i} Z\cdot p_X} \Delta\right)(\zeta_X^s,\zeta_X^{-s})
\end{align}
and
\begin{align}\label{equ17}
\ell_{T}(Z,\rho) := \frac2\beta \sum_{n\in\mathbb{Z}} \left ( \left({g^{{\rm i}\omega_n}} \ast {g^{{\rm i}\omega_n}}\right) (\zeta_{Z}^{\rho}) g^{-{\rm i}\omega_n}(\zeta_{Z}^{-\rho}) +
{g^{{\rm i}\omega_n}} (\zeta_{Z}^{\rho}) \left( {g^{-{\rm i}\omega_n}} \ast {g^{-{\rm i}\omega_n}}\right) (\zeta_{Z}^{-\rho}) \right).
\end{align}
\begin{lemma}\label{lemmaequ17}
$$
\left\| \left(N_{T,W} - \tilde N_{T,W}\right) \Delta \right\| \lesssim h \left( \left\| \Delta\right\| + \left\| |r| \Delta\right\|\right) \,.
$$
\end{lemma}
\begin{proof}
In \eqref{Werr}, we write
$$
W(hX + hz) = W(hX) + h \int_0^1 z\cdot \nabla W(hX + t hz) \, dt
$$
and then we have to estimate the norm of the error term coming from the $t$-integral. In order to calculate the $L^2(\mathbb{R}^3 \times \mathbb{R}^3)$-norm of the corresponding expression in the $(X,r)$-variables we first fix $r \in \mathbb{R}^3$ and consider the following term, which has a prefactor of $h$ in front,
\begin{multline}
\left( \int_{\mathbb{R}^3} dX \left | \int_{\mathbb{R}^9} dZ ds dz \, g^{{\rm i}\omega_n}(\frac r2-z) \int_0^1 z\cdot \nabla W(hX + t hz) dt \right.\right. \times \\ \left. \left. \times g^{{\rm i}\omega_n} (z + \zeta_Z^{-s}) g^{-{\rm i}\omega_n}(\zeta_{Z}^{s-r}) \left( e^{-{\rm i} Z\cdot p_X} \Delta\right)(\zeta_{X}^{s},\zeta_{X}^{-s}) \right|^2 \right)^{1/2} \,.
\end{multline}
Using Minkowski's inequality we can bound this by
\begin{align*}
&\int_{\mathbb{R}^9} dZ ds dz \, |g^{{\rm i}\omega_n}(\frac r2-z) | | g^{{\rm i}\omega_n} (z + \zeta_Z^{-s}) || g^{-{\rm i}\omega_n}(\zeta_{Z}^{s-r})| \\
& \qquad\qquad \times \left( \int_{\mathbb{R}^3} dX \left| \int_0^1 z\cdot \nabla W(hX + t hz) dt \left( e^{-{\rm i} Z\cdot p_X} \Delta\right)(\zeta_{X}^{s},\zeta_{X}^{-s}) \right|^2 \right)^{1/2} \\
& \qquad \leq \int_{\mathbb{R}^9} dZ ds dz \, |g^{{\rm i}\omega_n}(\frac r2-z) | | g^{{\rm i}\omega_n} (z + \zeta_Z^{-s}) || g^{-{\rm i}\omega_n}(\zeta_{Z}^{s-r})| \\
& \qquad\qquad \times |z| \| \nabla W\|_\infty \left( \int_{\mathbb{R}^3} dX \left | \left(e^{-{\rm i} Z\cdot p_X} \Delta\right)(X+s/2,X-s/2) \right|^2 \right)^{1/2} \\
& \qquad = \int_{\mathbb{R}^9} dZ ds dz \, |g^{{\rm i}\omega_n}(\frac r2-z) | | g^{{\rm i}\omega_n} (z + \zeta_Z^{-s}) || g^{-{\rm i}\omega_n}(\zeta_{Z}^{s-r})| |z| \|\nabla W\|_\infty m(s) \,
\end{align*}
where
$$
m(s) : = \left( \int_{\mathbb{R}^3} dX\, | \Delta(X+s/2,X-s/2)|^2 \right)^{1/2}
$$
and where we used the unitarity of $e^{-{\rm i} Z\cdot p_X}$ in the last equality.
The inequality
$$
|z| \leq \frac12 |z-r/2| + \frac12 |z+Z-s/2| + \frac12|Z-(r-s)/2| + \frac12 |s|
$$
leads to four terms, which we bound separately. The term with $|z-r/2|$ can be bounded by
\begin{align*}
& \| \nabla W\|_\infty \int_{\mathbb{R}^9} dZ ds dz \, |g^{{\rm i}\omega_n}( r/2-z) | |r/2 - z| | g^{{\rm i}\omega_n} (z + Z- s/2) || g^{-{\rm i}\omega_n}(Z + (s-r)/2)| m(s) \\
& \qquad = \| \nabla W\|_\infty \left( \left| \left|\cdot\right| g^{{\rm i}\omega_n} \right| \ast |g^{{\rm i}\omega_n}| \ast |g^{-{\rm i}\omega_n}| \ast m\right) (r) \,.
\end{align*}
According to Young's inequality, the $L^2$ norm of this term is bounded by $\|\nabla W\|_\infty$ times
$$
\||\cdot| g^{{\rm i}\omega_n}\|_1 \|g^{{\rm i}\omega_n}\|_1 \|g^{-{\rm i}\omega_n}\|_1 \|m\|_2 = \||\cdot| g^{{\rm i}\omega_n}\|_1 \|g^{{\rm i}\omega_n}\|_1 \|g^{-{\rm i}\omega_n}\|_1 \|\Delta\|_2 \,.
$$
According to \cite[Lemma 9]{FHaiLa} this expression is summable with respect to $n$ and therefore the contribution of this term to $\left(N_{T,W} - \tilde N_{T,W}\right) \Delta$ is bounded by a constant times $h \|\Delta\|_2$.
The argument for the terms involving $|z+Z-s/2|$ and $|Z-(r-s)/2|$ is similar.
The term with $|s|$ can be bounded by
\begin{align*}
& \| \nabla W\|_\infty \int_{\mathbb{R}^9} dZ ds dz \, |g^{{\rm i}\omega_n}( r/2-z) | |s| | g^{{\rm i}\omega_n} (z + Z- s/2) || g^{-{\rm i}\omega_n}(Z + (s-r)/2)| m(s) \\
& \qquad = \| \nabla W\|_\infty \left(|g^{{\rm i}\omega_n}|\ast |g^{{\rm i}\omega_n}| \ast |g^{-{\rm i}\omega_n}| \ast \left( |\cdot| m\right) \right) (r),
\end{align*}
According to Young's inequality, the $L^2$ norm of this term is bounded by $\|\nabla W\|_\infty$ times
$$
\| g^{{\rm i}\omega_n}\|_1 \|g^{{\rm i}\omega_n}\|_1 \|g^{-{\rm i}\omega_n}\|_1 \left\| \left|\cdot\right| m\right\|_2 = \|g^{{\rm i}\omega_n}\|_1 \|g^{{\rm i}\omega_n}\|_1 \|g^{-{\rm i}\omega_n}\|_1 \left\|\left|\cdot\right|\Delta\right\|_2 \,.
$$
Again by \cite[Lemma 9]{FHaiLa} this expression is summable with respect to $n$ and therefore the contribution of this term to $\left(N_{T,W} - \tilde N_{T,W}\right) \Delta$ is bounded by a constant times $h \left\|\left|\cdot\right|\Delta\right\|_2$. This proves the lemma.
\end{proof}
\section{Representation of $L_{T,W}$ on the states $\Delta = \psi(X) \tau(r)$}
We will argue below that we are able to restrict to a specific class of states, which are of the form
$\Delta(X+r/2,X-r/2) = \psi(X) \tau(r)$. Due to the symmetry of $\Delta$, $\tau$ has to be an even function, but
in fact we will later see that $\tau$ can be assumed as radial, and for the proof of our main theorem
$\tau$ will be proportional to $V^{1/2}(r) \phi_*(r)$, where $\phi_*(r)$ is the zero eigenstate of
$1 - V^{1/2} \chi_{\beta_c}(p_r^2 - \mu) V^{1/2}$.
The following corollary is an immediate consequence of the bounds in the previous section.
\begin{corollary}\label{reprproduct}
If $\Delta(X+r/2,X-r/2) = \psi(X)\tau(r)$ with $\tau$ even, then
\begin{align}
\label{eq:repr2a}
\langle \Delta, L_{T,W} \Delta \rangle & = \langle\psi,\psi\rangle
\langle \tau,\chi_\beta(p_r^2-\mu)\tau\rangle \nonumber \\
& - \int_{\mathbb{R}^3} dZ\, \langle\psi,(1 - \cos(Z\cdot p_X))\psi\rangle \iint_{\mathbb{R}^3\times\mathbb{R}^3} dr ds \, \overline{\tau(r)} k_{T}(Z,r-s) \tau(s)\nonumber \\
& + h^2 \int_{\mathbb{R}^3} dZ\, \langle\psi,W(hX) e^{-{\rm i} Z\cdot p_X} \psi\rangle \iint_{\mathbb{R}^3\times\mathbb{R}^3} dr ds \, \overline{ \tau(r)} \ell_{T}(Z,r-s) \tau(s)\nonumber \\
& + O(h^3) \|\psi\|^2 \|\tau\| \||\cdot| \tau\| \,.
\end{align}
\end{corollary}
We remark that with slightly more work we could replace the error term $\|\tau\| \||\cdot| \tau\|$ by $\||\cdot|^{1/2} \tau\|^2$.
The second term on the right side of \eqref{eq:repr2a} is given to leading order by the same expression with $1-\cos(Z\cdot p_X)$ replaced by $(Z\cdot p_X)^2/2$. Under the assumption that $\tau$ is a radial function, we therefore obtain $\langle \psi, p_X^2 \psi \rangle$ times a constant depending on $\tau$.
The third term on the right side of \eqref{eq:repr2a} is given to leading order by the same expression with $e^{-{\rm i} Z\cdot p_X}$ replaced by $1$. We therefore obtain $h^2 \langle\psi, W(h\cdot) \psi \rangle$ times a constant depending on $\tau$.
This tells us that the center-of-mass fluctuations are governed by a one-body operator of the form $c_1 p_X^2 + c_2 h^2 W(hX)$, which is unitarily equivalent to the operator
$$
h^2 \left( c_1 p_X^2 + c_2 W(X)\right).
$$
The precise value of the constants $c_1,c_2$ depends on the specific choice of $\tau$.
As we will show below, the errors made in these two approximations can be controlled by $\|p_X^2 \psi\|^2$ and $h^2\|p_X \psi\| \|\psi\|$. In order to get an intuition why the error terms are indeed of higher order in $h$ we recall the heuristic picture of our chosen scaling. The external field $W$ varies on the scale $1/h$. Therefore we expect the optimal function $\psi$ to match this behavior and vary as well on the macroscopic scale. More precisely, we expect that
$\psi$ will be of the form $\psi(X) = h^{3/2} \tilde \psi(hX) $ with a function $\tilde\psi$ which is bounded in $H^2$ uniformly for small $h$. Therefore the error bounds $\|p_X^2 \psi\|^2$ and $h^2\|p_X \psi\| \|\psi\|$ are $o(h^2)$.
Next, we formulate this intuitive picture as a precise mathematical statement.
\begin{theorem} \label{semest}
There is a constant $C$ such that for $\Delta$ of the form
$$
\Delta(X+r/2,X-r/2)=\psi(X)\tau(r)
$$
with $\tau$ radial, one has
\begin{align}
\label{eq:semest}
& \left| \langle \Delta, L_{T,W} \Delta \rangle - A^{(0)}_T[\tau] \|\psi\|^2 - A^{(1)}_T[\tau] \langle \psi, p_X^2 \psi\rangle - h^2 A_T^{(2)}[\tau] \langle \psi, W(h \cdot) \psi\rangle\right| \notag \\
& \qquad \leq C \left( \|\tau\|^2 \|p_X^2 \psi\|^2 + h^2 \left\| \tau\right\|^2 \| p_X \psi\| \|\psi\| + h^3 \|\psi\|^2\||\cdot|\tau\| \|\tau\| \right)
\end{align}
with
\begin{align*}
A^{(0)}_T[\tau] & = \beta \int_{\mathbb{R}^3} dp\, |\hat \tau(p)|^2 \ g_0(\beta(p^2-\mu)) \,, \\
A^{(1)}_T[\tau] & = - \frac{\beta^2}{4} \int_{\mathbb{R}^3} dp\, |\hat \tau(p)|^2 \left(g_1(\beta(p^2-\mu)) + \frac23 \beta p^2 g_2(\beta(p^2-\mu)) \right)\, ,\\
A^{(2)}_T[\tau] & = \frac{\beta^2}{4} \int_{\mathbb{R}^3} dp\, |\hat \tau(p)|^2\, g_1(\beta(p^2-\mu))
\end{align*}
in terms of the functions $g_0$, $g_1$ and $g_2$ from \eqref{eq:auxiliary}.
\end{theorem}
\begin{proof}
This theorem is essentially a consequence of \eqref{eq:repr2a}. We first notice that
$$
\langle \tau,\chi_\beta(p_r^2-\mu)\tau\rangle = A_T^{(0)}[\tau] \,.
$$
Moreover, using arguments as in the previous subsection one can verify that
$$
\left| \int_{\mathbb{R}^3} dZ\, \langle\psi,(1 - \cos(Z\cdot p_X) - (Z\cdot p_X)^2/2)\psi\rangle F_\tau(Z) \right| \lesssim \|\tau\|^2 \|p_X^2\psi\|^2
$$
where we have introduced
$$
F_\tau(Z) := \iint_{\mathbb{R}^3\times\mathbb{R}^3} dr\, ds\, \overline{\tau(r)} k_{T}(Z,r-s) \tau(s)\,.
$$
Since $\tau$ is radial, so is $F_\tau$ and therefore
$$
\frac12 \int_{\mathbb{R}^3} dZ\, \langle\psi,(Z\cdot p_X)^2\psi\rangle F_\tau(Z)
= \frac{1}{6} \int_{\mathbb{R}^3} dZ\, Z^2 F_\tau(Z) \langle\psi,p_X^2\psi\rangle \,.
$$
Now using \eqref{eq:semest1},
$$
\int_{\mathbb{R}^3} dZ\, Z^2 F_\tau(Z) = -\int_{\mathbb{R}^3} dk \, \nabla_{\ell}^2|_{\ell=0} L(k+\frac\ell2,k-\frac\ell2) |\hat \tau(k)|^2 \,,
$$
and a tedious, but straightforward computation yields
$$
\nabla_{\ell}^2|_{\ell=0} L(k+\frac\ell2,k-\frac\ell2) = - \frac{3\beta^2}2 \left(g_1(\beta(k^2-\mu)) + \frac23 \beta k^2 g_2(\beta(k^2-\mu)) \right),
$$
which shows that
$$
- \frac12 \int_{\mathbb{R}^3} dZ\, \langle\psi,(Z\cdot p_X)^2\psi\rangle F_\tau(Z)
= A^{(1)}_T[\tau] \langle\psi,p_X^2\psi\rangle \,.
$$
Finally, by estimating $1- e^{{\rm i} Z\cdot p_X}$ we obtain
$$
\left| \int_{\mathbb{R}^3} dZ\, \langle\psi,W(hX) \left( e^{-{\rm i} Z\cdot p_X} - 1\right) \psi\rangle G_\tau(Z) \right| \lesssim \|\tau\|^2 \| p_X\psi\| \|\psi\|
$$
with
$$
G_\tau(Z) := \iint_{\mathbb{R}^3\times\mathbb{R}^3} dr ds \, \overline{ \tau(r)} \ell_{T}(Z,r-s) \tau(s) \,.
$$
Rewriting \eqref{equ17} in Fourier space and summing over the Matsubara frequencies gives
$$
\int_{\mathbb{R}^3} dZ \,\ell_{T}(Z,\rho) = \int_{\mathbb{R}^3} \frac{dk}{(2\pi)^3} \frac{\beta^2}{4} g_1(\beta(k^2-\mu)) e^{ik\cdot\rho}
$$
and therefore
$$
\int_{\mathbb{R}^3} dZ\, G_\tau(Z) = A^{(2)}_T[\tau] \,.
$$
This concludes the proof of the theorem.
\end{proof}
\section{Lower bound on the critical temperature}
We now provide the \emph{Proof of part (1) of Theorem \ref{main}}, which will be a rather straightforward consequence of Theorem \ref{semest}. We will work under Assumptions \ref{ass2} and \ref{ass0}. Assumption \ref{ass1} is not needed in this part of Theorem \ref{main}.
We fix a parameter $T_1$ with $0<T_1<T_c$ and restrict ourselves to temperatures $T\geq T_1$. We consider functions $\Phi$ in $L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ of the form
$$
\Phi(x,y)=\phi(x-y)h^{3/2} \psi(h (x+y)/2)\,,
$$
where the functions $\phi\in L^2_{\rm symm}(\mathbb{R}^3)$ and $\psi\in L^2(\mathbb{R}^3)$ are still to be determined. At the moment we only require that $\|\psi\|=1$ and $\|p_X^2\psi\|<\infty$.
We first assume, in addition, that $T_1\geq T_c-Mh^2$ for some constant $M$ independent of $h$. In this case we choose $\phi$ radial and then, applying the expansion from Theorem~\ref{semest} with $\tau(r)=V(r)^{1/2}\phi(r)$, we find that
\begin{align}\label{eq:lowerboundexp}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle = & \|\phi\|^2 - \langle\tau(r)\psi(X), L_{T,W} \tau(r)\psi(X) \rangle \notag \\
\leq & \|\phi\|^2 - A_T^{(0)}[\tau] - h^2 A_T^{(1)}[\tau] \langle\psi,p_X^2\psi\rangle - h^2 A_T^{(2)}[\tau] \langle\psi,W \psi\rangle \notag \\
& + C h^3 \,.
\end{align}
The constant $C$ here depends only on upper bounds on $\|p_X^2\psi\|$, $\|\tau\|$ and $\||\cdot|\tau\|$ (as well as on $M$). The leading order term on the right side is
\begin{equation}
\label{eq:lowerboundmain}
\|\phi\|^2 - A_T^{(0)}[\tau] = \left\langle \phi, \left( 1- V^{1/2}\chi_\beta(p_r^2-\mu) V^{1/2}\right)\phi\right\rangle.
\end{equation}
We choose
$$
\phi :=(2\pi)^{-3/2} \|\chi_{\beta_c}(p^2-\mu) V^{1/2} \phi_*\| \ \phi_* \,,
$$
which makes \eqref{eq:lowerboundmain} equal to zero at $T=T_c$. With this choice of $\phi$ we therefore obtain
\begin{align}
\label{eq:upperproof1}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle
\leq & A^{(0)}_{T_c}[\tau] - A^{(0)}_T[\tau] - h^2 \left( A^{(1)}_T[\tau] \langle\psi,p_X^2 \psi\rangle + A_T^{(2)}[\tau] \langle\psi,W \psi\rangle\right) \nonumber \\& + C h^3 \,.
\end{align}
In order to proceed, we note the fact that $\tau = V^{1/2}\phi = (2\pi)^{-3/2} V \alpha_*$, and therefore, in terms of the function $t$ from \eqref{eq:t},
\begin{equation}
\label{eq:upperchoicetau}
\hat\tau = (1/2)(2\pi)^{-3/2} t \,.
\end{equation}
It follows from this identity that
$$
\frac{d}{dT}|_{T=T_c} A^{(0)}_{T}[\tau] = -T_c^{-1} \Lambda_2 \,,
$$
and some simple analysis of the function $g_0$ shows that
$$
A^{(0)}_{T_c}[\tau] - A^{(0)}_T[\tau] \leq -\Lambda_2 \frac{T_c-T}{T_c} + C(T_c-T)^2
$$
for all $T_1\leq T\leq T_c$. Using \eqref{eq:upperchoicetau} once again we also find that
$$
A^{(1)}_{T_c}[\tau] = - \Lambda_0
\qquad\text{and}\qquad
A^{(2)}_{T_c}[\tau] = - \Lambda_1 \,,
$$
which in turn can be used to prove that
$$
A^{(1)}_T[\tau] \geq -\Lambda_0 - C(T_c-T)
\qquad\text{and}\qquad
\left| A^{(2)}_T[\tau] +\Lambda_1 \right| \leq C(T_c-T)
$$
for all $T_1\leq T\leq T_c$.
Inserting these expansions into \eqref{eq:upperproof1} we obtain
\begin{align*}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle
\leq & - \Lambda_2 \frac{T_c-T}{T_c} + h^2 \left\langle\psi,\left( \Lambda_0 p_X^2 + \Lambda_1 W \right) \psi\right\rangle + C h^3
\end{align*}
for all $T_1\leq T\leq T_c$. Note that here we used the assumption $T\geq T_c - M h^2$, so that the error terms are independent of $T-T_c$.
In order to conclude the proof we assume first, for the sake of simplicity, that $\inf {\rm spec} \left(\Lambda_0 p^2_X + \Lambda_1 W(X)\right)$ is an eigenvalue. In this case we simply choose $\psi$ to be a corresponding normalized eigenfunction. With this choice we obtain, recalling the definition of $D_c$ from \eqref{eq:dc},
\begin{align*}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle
\leq & -\Lambda_2 \frac{T_c-T}{T_c} + h^2 \Lambda_2 D_c
+ C h^3 \,.
\end{align*}
The right side is negative if $T< T_c(1- D_c h^2 + (C/\Lambda_2) h^3)$, as claimed.
In case $\inf {\rm spec} \left(\Lambda_0 p^2_X + \Lambda_1 W(X)\right)$ is not an eigenvalue, we choose a sequence of functions $\psi_h$ with $\|\psi_h\|=1$,
$$
\left\langle \psi_h, \left(\Lambda_0 p_X^2 + \Lambda_1 W(X)\right)\psi_h \right\rangle \leq \Lambda_2 \left( D_c + h \right)
\qquad\text{and}\qquad
\| p_X^2 \psi_h \| \leq C
$$
for some $C$ independent of $h$. Such a sequence is obtained by choosing elements in the spectral subspace of $\Lambda_0 p^2_X + \Lambda_1 W(X)$ corresponding to the intervals $\left[\Lambda_2 D_c,\Lambda_2\left(D_c+h\right)\right]$. Since $\Lambda_0 p^2_X + \Lambda_1 W(X)$ has the same operator domain as $p_X^2$ we conclude that
$$
\| p_X^2 \psi_h \| \lesssim \left\| \left(\Lambda_0 p^2_X + \Lambda_1 W(X) + C' \right) \psi_h\right\| \leq \Lambda_2\left( D_c + h \right) + C' \,,
$$
which proves the last requirement.
We can now repeat the proof with $\psi$ replaced by $\psi_h$. Since all constants were uniform in $\psi$ as long as $\|\psi\|=1$ and $\|p_X^2\psi\|\leq C$, we arrive at the same conclusion as before. This proves the assertion in case $T\geq T_c- Mh^2$ for some fixed $M$ independent of $h$.
Thus, in order to complete the proof of part (1) in the theorem, we show that there is an $M>0$ such that if $T<T_c - M h^2$, then there are $\phi$ and $\psi$ such that the $\Phi$ defined as above satisfies $\langle\Phi, (1-V^{1/2} L_{T,W}V^{1/2})\Phi\rangle<0$.
We proceed similarly as before, but use Corollary \ref{reprproduct} instead of Theorem \ref{semest}. By similar, but simpler estimates as in the proof of Theorem \ref{semest} we obtain
\begin{align}\label{eq:lowerboundexp1}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle \leq \|\phi\|^2 - A_T^{(0)}[\tau] + C h^2 \,.
\end{align}
The constant $C$ here depends only on upper bounds on $\|p_X\psi\|$, $\|\tau\|$ and $\||\cdot|\tau\|$ (as well as on $T_1$). Thus the leading term on the right side is again \eqref{eq:lowerboundmain}.
To bound this term, we denote by $\lambda_T$ the largest eigenvalue of $V^{1/2} \chi_\beta(p_r^2-\mu)V^{1/2}$ in $L^2_{\rm symm}(\mathbb{R}^3)$. By definition of $T_c$ we have $\lambda_{T_c}=1$. Since $\beta\mapsto\chi_\beta(E)$ is monotone for any $E$ with positive derivative, we infer by analytic perturbation theory that there is a $c>0$ such that
$$
\lambda_T \geq \lambda_{T_c} + c(T_c-T) = 1 + c(T_c-T)
\qquad\text{for all}\ 0\leq T\leq T_c \,.
$$
Let $\phi_T$ be a normalized eigenfunction of $V^{1/2} \chi_\beta(p_r^2-\mu)V^{1/2}$ corresponding to $\lambda_T$. With $\phi=\phi_T$ and an arbitrary normalized function $\psi$ with $\|p_X^2\psi\|<\infty$ we obtain,
by inserting \eqref{eq:lowerboundmain} into \eqref{eq:lowerboundexp1} and using the above bound,
\begin{align*}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle
\leq 1- \lambda_{T} + C h^2 \leq - c (T_c -T) + C h^2 \, .
\end{align*}
The right side is negative for $T< T_c - (C/c) h^2$, as claimed. This completes the proof of part (1) of Theorem \ref{main}.
\qed
\section{The approximate form of almost minimizers}
In this and the following section we work under Assumptions \ref{ass2}, \ref{ass0} and \ref{ass1}.
\subsection{The decomposition lemma}
The remainder of this paper is devoted to proving an upper bound on the critical temperature. As a preliminary step we prove in this section a decomposition lemma, which says that, if $|T_c-T|\leq C_1 h^2$ and if $\Phi$ satisfies $\langle\Phi,(1-V^{1/2}L_{T,W}V^{1/2})\Phi\rangle\leq C_2h^2$ for some fixed constants $C_1$ and $C_2$ independent of $h$, then $\Phi$ has, up to a controllable error, the same form as the trial function that we used in the proof of the lower bound on the critical temperature.
\begin{theorem}\label{decomp}
For given constants $C_1,C_2>0$ there are constants $h_0>0$ and $C>0$ such that the following holds. If $T>0$ satisfies $|T-T_c|\leq C_1h^2$, if $\Phi\in L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ satisfies $\|\Phi\|=1$ and
$$
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle \leq C_2 h^2 \,,
$$
and if $\epsilon$ satisfies $\epsilon\in[h^2,h^2_0]$, then there are $\psi_\leq\in L^2(\mathbb{R}^3)$ and $\sigma\in L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ such that
$$
\Phi(X+r/2,X-r/2) = \phi_*(r) \psi_{\leq}(X)+\sigma \,,
$$
where
\begin{equation}
\label{eq:decomppsibounds}
\| (p_X^2)^{k/2}\psi_\leq \|^2 \leq C \epsilon^{k-1} h^2 \qquad \text{if}\ k\geq 1 \,,
\end{equation}
\begin{equation}
\label{eq:decompsigmabound}
\|\sigma\|^2 \leq C \epsilon^{-1} h^2
\end{equation}
and
\begin{equation}
\label{eq:decompsilower}
1\geq \|\psi_\leq\|^2 \geq 1- C\epsilon^{-1}h^2 \,.
\end{equation}
Moreover, $\psi_\leq\in\ran\1(p_X^2\leq\epsilon)$ and there is a $\psi_>\in L^2(\mathbb{R}^3)\cap\ran\1(p_X^2>\epsilon)$ such that
$$
\sigma_0(X+r/2,X-r/2):= \frac{ \phi_*(r) \cos(p_X\cdot r/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(r')|^2 \cos^2(p_X \cdot r'/2)\, dr' }} \psi_>(X)
$$
satisfies
\begin{equation}
\label{eq:diffsigmasigma0}
\|\sigma - \sigma_0\|^2 \leq C h^2 \,.
\end{equation}
\end{theorem}
Thus, $\Phi$ is of the form $\psi_\leq(X)\phi_*(r)$ up to a small error. The parameter $\epsilon$ provides a momentum cut-off similarly as in \cite{FHSS,FHSS2} and ensures that we have control on the expectation of $(p_X^2)^2$ in $\psi_\leq$.
\subsection{Upper bound on $L_{T,W}$}
Our goal in this subsection is to obtain an operator lower bound on $1-V^{1/2}L_{T,W}V^{1/2}$. In \cite{FHSS,FHSS2} such a bound was proved by means of a relative entropy inequality \cite[Lemma 3]{FHSS}, which controlled a two-particle operator by the sum of two one-particle operators, and by \cite[Lemma 5]{FHSS} which showed that the energy of the system is dominated by the kinetic energy of the center of mass motion. This was sufficient to recover the corresponding a-priori estimates.
In \cite{FHaiLa} this operator bound was performed in the presence of a constant magnetic field. Following the spirit of \cite{FHSS,FHSS2} we had to come up with new ideas in order to overcome the problems of non-commutativity of the components of the magnetic momentum operator. In the present much simpler situation we can choose a mixture of the two methods \cite{FHSS,FHSS2} and~\cite{FHaiLa}
We define the unitary operator
\begin{equation}
\label{eq:u}
U := e^{-{\rm i} p_X\cdot r/2}
\end{equation}
in $L^2(\mathbb{R}^3\times\mathbb{R}^3)$ where, as usual, $r=x-y$ and $X=(x+y)/2$.
\begin{proposition}\label{lowerboundlt}
There is a constant $C>0$ such that for all $T>0$,
\begin{align*}
V^{1/2} L_{T,W} V^{1/2} & \leq \frac12 \left( U V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} U^* + U^* V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} U \right) \\
& \qquad + C \beta^3 h^2 .
\end{align*}
\end{proposition}
\begin{proof}
Since for any real numbers $E$ and $E'$ one has
$$
\Xi_\beta(E,E') \leq \frac12 \left( \frac{\tanh\frac{\beta E}{2}}{E} + \frac{\tanh\frac{\beta E'}{2}}{E'} \right) = \frac12 \left( \chi_\beta(E) + \chi_\beta(E') \right),
$$
we have
$$
L_{T,0} =\Xi_\beta(\mathord{\mathfrak h}_{0,x},\mathord{\mathfrak h}_{0,y}) \leq \frac12 \left(\chi_\beta(\mathord{\mathfrak h}_{0,x}) + \chi_\beta(\mathord{\mathfrak h}_{0,y}) \right).
$$
In the variables $r=x-y$, $X=(x+y)/2$ we have $p_x = p_r + p_X/2$ and $p_y = p_r - p_X/2$ and therefore
$$
\mathord{\mathfrak h}_{0,x} = (p_r + p_X/2)^2 - \mu = U \left( p_r^2 -\mu\right) U^* \,,
\quad
\mathord{\mathfrak h}_{0,y} = (p_r - p_X/2)^2 - \mu = U^* \left( p_r^2 -\mu\right) U \,,
$$
so the previous bound can be written as
$$
L_{T,0} \leq \frac12\left( U \chi_\beta(p_r^2-\mu) U^* + U^* \chi_\beta(p_r^2-\mu) U \right).
$$
On the other hand, by Lemma \ref{lemmadiffw0} we have
$$
L_{T,W} \leq L_{T,0} + C \beta^3 h^2 \,.
$$
Since $V$ commutes with $U$ we obtain the claimed bound.
\end{proof}
\subsection{A priori bound on the critical temperature and an operator inequality}\label{sec:opineq}
As a first consequence of Proposition \ref{lowerboundlt} we obtain a rough a-priori upper bound on the critical temperature.
\begin{corollary}\label{aprioriuppertemp}
There are constants $h_0>0$ and $C>0$ such that for all $0< h \leq h_0$ and $T> T_c+Ch^2$ one has
$$
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle >0 \,,
$$
unless $\Phi=0$.
\end{corollary}
\begin{proof}
According to Proposition \ref{lowerboundlt} for all $T\geq T_c$,
\begin{align}\label{eq:lowerbound}
1- V^{1/2} L_{T,W} V^{1/2} & \geq 1 - \frac12 \left( U V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} U^* + U^* V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} U \right) \notag \\
& \qquad - C h^2 \,.
\end{align}
We next recall that the family of operators $V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2}$ is non-decreasing with respect to $\beta$ and has an eigenvalue $1$ at $\beta=\beta_c$. Moreover, since the function $\chi_\beta(E)$ is strictly increasing with respect to $\beta$ for every $E\in\mathbb{R}$, we learn from analytic perturbation theory that there are $c>0$ and $T_2>T_c$ such that for all $T_c\leq T\leq T_2$,
$$
V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} \leq 1 - c(T-T_c)\,.
$$
Again by monotonicity this implies that for all $T\geq T_c$
$$
V^{1/2} \chi_\beta(p_r^2-\mu) V^{1/2} \leq 1 - c\min\{T-T_c,T_2-T_c\}\,.
$$
Inserting this into the lower bound above we conclude that
$$
1- V^{1/2} L_{T,W} V^{1/2} \geq c\min\{T-T_c,T_2-T_c\} - C h^2 \,.
$$
The right side is positive if $T> T_c + (C/c)h^2$ and $h^2 \leq (c/C)(T_2-T_c)$, which proves the corollary.
\end{proof}
As a consequence of this corollary and the lower bound on the critical temperature, from now on we may and will restrict ourselves to temperatures $T$ such that $|T-T_c|$ is bounded by a constant times $h^2$.
Our next goal is to deduce from Proposition \ref{lowerboundlt} a lower bound on the operator $1- V^{1/2} L_{T,W} V^{1/2}$. We recall that by definition of $\beta_c$ the largest eigenvalue of the operator $V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2}$ equals one. Moreover, by Assumption \ref{ass1}, this eigenvalue is simple and $\phi_*$ denotes a corresponding real-valued, normalized eigenfunction. We denote by
$$
P:=|\phi_*\rangle\langle\phi_*|
$$
the corresponding projection and write $P^\bot = 1-P$. Since $V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2}$ is a compact operator, there is a $\kappa>0$ such that
\begin{equation}
\label{eq:gap}
V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2} \leq 1-\kappa P^\bot \,.
\end{equation}
Finally, we introduce the operator
\begin{equation}
\label{eq:defq}
Q:= \frac12\left( U P U^* + U^* P U \right).
\end{equation}
We can now state our operator inequality for $1- V^{1/2} L_{T,W} V^{1/2}$.
\begin{proposition}\label{opineq}
Given $C_1>0$ and $h_0>0$ with $C_1 h_0^2< T_c$, there is a constant $C>0$ such that for all $|T-T_c|\leq C_1h^2$ and $0< h \leq h_0$ one has
\begin{align}\label{eq:opineq}
1- V^{1/2} L_{T,W} V^{1/2} \geq \kappa\left( 1- Q\right) - C h^2 \,.
\end{align}
\end{proposition}
\begin{proof}
Our starting point is again inequality \eqref{eq:lowerbound}, which is valid for all $|T-T_c|\leq C_1 h_0^2$. Since the derivative of $\chi_\beta(E)$ with respect to $T$ is bounded uniformly in $E$ for $T$ away from $0$, we infer that there is a $C'>0$ such that for all $|T-T_c|\leq C_1 h^2_0$ and all $E\in\mathbb{R}$,
\begin{equation}
\label{eq:chiineq}
\left|\chi_\beta(E) - \chi_{\beta_c}(E)\right| \leq C'|T-T_c| \,.
\end{equation}
This, together with the gap inequality \eqref{eq:gap}, implies that for $|T-T_c|\leq C_1 h^2 \leq C_1 h^2_0$,
\begin{align*}
1- V^{1/2} L_{T,W} V^{1/2} & \geq
1- \frac12 \left( U V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2} U^* + U^* V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2} U \right) \\
& \qquad\qquad -C''|T-T_c| - C h^2 \\
& \geq \frac\kappa2 \left( U P^\bot U^* + U^* P^\bot U\right) - (C_1 C'' + C) h^2 \\
& = \kappa\left( 1- Q\right) - (C_1 C'' + C) h^2 \,,
\end{align*}
as claimed.
\end{proof}
Next, we observe that for functions $\Phi\in L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ one can write
\begin{align*}
& (Q\Phi)(X+r/2,X-r/2) \\
& \quad = \phi_*(r) \cos(p_X\cdot r/2) \int_{\mathbb{R}^3} ds\, \overline{\phi_*(s)} \cos(p_X\cdot s/2) \Phi(X+s/2,X-s/2) \\
& \quad =: |A_{p_X}\rangle \langle A_{p_X} | \Phi \rangle
\end{align*}
with
$$
A_p(r):= \phi_*(r) \cos(p\cdot r/2) \,.
$$
(More precisely, the expression $|A_{p_X}\rangle \langle A_{p_X} |$ can be written as a direct integral over the center of mass momenta $p_X$. In the case of magnetic fields \cite{FHaiLa} this did not work because the components of the magnetic momentum did not commute.)
Now we use the fact that in each fiber $Q$ can be estimated from above by its largest eigenvalue, hence we immediately conclude that
\begin{equation}
1 - Q \geq 1 - \langle A_{p_X}|A_{p_X}\rangle = 1 - R
\end{equation}
with
\begin{equation}
\label{eq:r}
R := \int_{\mathbb{R}^3}dr\, |\phi_*(r)|^2 \cos^2(r\cdot p_X/2)
\end{equation}
acting in $L^2(\mathbb{R}^3)$. Since $\cos(r\cdot p_X/2)^2 \leq 1$ and since $\phi_*$ is normalized, we have $R\leq 1$ and therefore $1-R\geq 0$. We now prove a more precise lower bound.
\begin{lemma}\label{rbound}
There are constants $E_0>0$ and $c>0$ such that
$$
1- R \geq c \ \frac{p_X^2}{E_0+p_X^2} \,.
$$
\end{lemma}
\begin{proof}
All operators involved are diagonal in Fourier space, so for the proof we can consider $p_X$ to be a vector in $\mathbb{R}^3$. Using the normalization of $\phi_*$ we are thus lead to considering the function
$$
1-R(p_X) = \int_{\mathbb{R}^3} dr \, |\phi_*(r)|^2 \left( 1-\cos^2(p_X\cdot r/2)\right) = \int_{\mathbb{R}^3} dr \, |\phi_*(r)|^2 \sin^2(p_X\cdot r/2)\,.
$$
First, we have
$$
\lim_{p_X\to 0} \frac{1-R(p_X)}{p_X^2} = \frac1{12} \int_{\mathbb{R}^3} dr\, |\phi_*(r)|^2 r^2 =:c \,.
$$
(The right side is finite, as shown in \cite{FHSS}.) Therefore, there is a $\delta>0$ such that $1-R(p_X)\geq (c/2) p_X^2$ for $|p_X|\leq\delta$.
Second, by the Riemann--Lebesgue lemma, we have
$$
\lim_{|p_X|\to\infty} \left(1-R(p_X)\right) = \frac12 \,,
$$
and therefore there is an $M>0$ such that $1-R(p_X)\geq 1/4$ for $|p_X|\geq M$.
Since for any $p_X\neq 0$ the function $r\mapsto \sin^2(p_X\cdot r/2)$ vanishes only on a set of measure zero, we have $1-R(p_X)> 0$ for all $p_X\neq 0$. Since $p_X\mapsto R(p_X)$ is continuous, there is a $c'>0$ such that $1-R(p_X)\geq c'$ for all $\delta\leq |p_X|\leq M$. This proves that
$$
1 - R(p_X) \geq \min\{ (c/2) p_X^2, c',1/4\} \,,
$$
which immediately implies the lemma.
\end{proof}
\subsection{Proof of the decomposition lemma}
As a consequence of Proposition~\ref{opineq} we now deduce a first decomposition result for almost maximizers $\Phi$ of $1-V^{1/2} L_{T,W} V^{1/2}$.
Let us now define the projection
$$
P_Q := \frac{| A_{p_X} \rangle \langle A_{p_X} |}{\langle A_{p_X}| A_{p_X}\rangle},
$$
where the last expression is again a direct integral over the momenta $p_X$.
To see how this operator acts define for a given $\Phi\in L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$,
\begin{equation}\label{eq:defpsi}
\psi(X) := \frac{\langle A_{p_X}|}{\|A_{p_X}\| } \Phi = \int_{\mathbb{R}^3} ds \frac{ \phi_*(s) \cos(p_X\cdot s/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(s')|^2 \cos^2(p_X \cdot s'/2) ds'}} \Phi(X+s/2,X-s/2) \,.
\end{equation}
Then
$$P_Q \Phi (X+ r/2,X-r/2) = \frac{ \phi_*(r) \cos(p_X\cdot r/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(r')|^2 \cos^2(p_X \cdot r'/2) dr' }} \psi(X),$$
and we define $\xi\in L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ by
\begin{equation}
\label{eq:defxi}
\Phi = P_Q \Phi + \xi \,.
\end{equation}
With these definitions we can formulate a first version of the decomposition lemma.
\begin{lemma}\label{decomp1}
Given $C_1,C_2>0$ there are $h_0>0$, $E_0>0$ and $C>0$ with the following properties. If $|T-T_c|\leq C_1 h^2 \leq C_1 h^2_0$ and if $\Phi\in L^2_{\rm symm}(\mathbb{R}^3\times\mathbb{R}^3)$ with $\|\Phi\|=1$ satisfies
\begin{equation}
\label{eq:almostmin}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle \leq C_2 h^2 \,,
\end{equation}
then, with $\psi$ and $\xi$ defined in \eqref{eq:defpsi} and \eqref{eq:defxi},
$$
\left\langle\psi,\frac{p_X^2}{E_0+p_X^2}\psi\right\rangle + \|\xi\|^2 \leq C h^2 \,.
$$
and
$$
1\geq \|\psi\|^2 \geq 1- Ch^2 \,.
$$
\end{lemma}
\begin{proof}
By Proposition \ref{opineq} and assumption \eqref{eq:almostmin} we obtain
\begin{equation}
\label{eq:apriorisize}
\langle\Phi,(1-Q)\Phi\rangle \leq \kappa^{-1} (C+C_2)h^2 \,.
\end{equation}
By construction, for every fixed value $p_X$ of the Fourier transform with respect to $X$, $P_Q \Phi $ and $\xi$ are orthogonal as functions of $r$. Therefore
$$
\langle \Phi, (1- Q) \Phi\rangle = \langle P_Q \Phi, (1- Q) P_Q \Phi\rangle + \| \xi \|^2 \,.
$$
On the other hand, it is easy to see that
$$
\langle P_Q \Phi, (1- Q) P_Q \Phi\rangle = \langle \psi,(1-R) \psi \rangle \,.
$$
Therefore the lower bound on $1-R$ from Lemma \ref{rbound} implies the first assertion in the lemma.
In order to prove the second assertion, we note that
$$
\|\psi\|^2 = \|P_Q\Phi\|^2 = \|\Phi\|^2 - \|\xi\|^2 = 1-\|\xi\|^2
$$
and use the bound on $\|\xi\|^2$ from the first assertion.
\end{proof}
\begin{proof}[Proof of Theorem \ref{decomp}]
Let $\psi$ be as in Lemma \ref{decomp1}. For $\epsilon\in[h^2,h^2_0]$ we set
$$
\psi_\leq := \1(p_X^2 \leq \epsilon)\psi \,,
\qquad
\psi_> := \1(p_X^2 > \epsilon)\psi \,.
$$
Recall from Lemma \ref{decomp1} that $\langle\psi,p_X^2(E_0+p_X^2)^{-1}\psi\rangle\leq C h^2$. This implies that for $k\geq 1$,
\begin{align*}
\left\|\left(p_X^2\right)^{k/2} \psi_\leq\right\|^2 & \leq \epsilon^{k-1} \|p_X\psi_\leq\|^2 \leq (E_0+\epsilon) \epsilon^{k-1} \left\langle\psi,\frac{p_X^2}{E_0+p_X^2}\psi\right\rangle \notag \\
& \leq C (E_0+\epsilon) \epsilon^{k-1} h^2
\end{align*}
and
\begin{equation}
\label{eq:apriorioutside}
\|\psi_>\|^2\leq \frac{E_0+\epsilon}{\epsilon}\left\langle\psi,\frac{p_X^2}{E_0+p_X^2}\psi\right\rangle \leq C \frac{E_0+\epsilon}{\epsilon} h^2 \,.
\end{equation}
We now define
$$
\sigma_0(X+r/2,X-r/2):= \frac{ \phi_*(r) \cos(p_X\cdot r/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(r')|^2 \cos^2(p_X \cdot r'/2)\, dr' }} \psi_>(X) \,,
$$
$$
\sigma_1(X+r/2,X-r/2):= - \phi_*(r) \left( 1 - \frac{\cos(p_X\cdot r/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(r')|^2 \cos^2(p_X \cdot r'/2)\, dr' }} \right) \psi_\leq(X)
$$
and
$$
\sigma:= \sigma_0 + \sigma_1 + \xi \,,
$$
so that, by Lemma \ref{decomp1},
$$
\Phi= \phi_*(r) \psi_{\leq}(X) + \sigma\,.
$$
According to Lemma \ref{decomp1} and \eqref{eq:apriorioutside}, we have
$$
\|\psi_\leq\|^2 = \|\psi\|^2 - \|\psi_>\|^2 \geq 1- Ch^2 - C \epsilon^{-1}h^2\geq 1- C'\epsilon^{-1} h^2 \,.
$$
and, again according to \eqref{eq:apriorioutside}, we have
$$
\|\sigma_0\|^2 = \|\psi_>\|^2 \leq C \frac{E_0+\epsilon}{\epsilon} h^2 \,.
$$
Moreover,
$$
\|\sigma_1\|^2 = \langle \psi_\leq, S \psi_\leq \rangle
$$
with the operator
\begin{align*}
S & := \int_{\mathbb{R}^3} dr\, |\phi_*(r)|^2 \left( 1 - \frac{\cos(p_X\cdot r/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(r')|^2 \cos^2(p_X \cdot r'/2)\, dr' }} \right)^2 \\
& = 2 \int_{\mathbb{R}^3} dr\, |\phi_*(r)|^2 \left( 1 - \frac{\cos(p_X\cdot r/2) }{ \sqrt{ \int_{\mathbb{R}^3} |\phi_*(r')|^2 \cos^2(p_X \cdot r'/2)\, dr' }} \right)
\end{align*}
acting in $L^2(\mathbb{R}^3)$. In the same way as in the proof of Lemma \ref{rbound} we can show that
$$
S \leq C \frac{p_X^2}{E_0+p_X^2} \,,
$$
and therefore
$$
\|\sigma_1\|^2 \lesssim \langle \psi_\leq, \frac{p_X^2}{E_0+p_X^2} \psi_\leq \rangle \leq \langle \psi, \frac{p_X^2}{E_0+p_X^2} \psi \rangle \lesssim h^2 \,.
$$
We conclude that
$$
\| \sigma - \sigma_0 \| = \|\sigma_1 + \xi \| \leq \|\sigma_1 \| + \|\xi\| \lesssim h \,.
$$
This concludes the proof of the theorem.
\end{proof}
\section{Upper bound on the critical temperature}
In this section we prove part (2) of Theorem \ref{main}. In view of Corollary \ref{aprioriuppertemp} and the lower bound on the critical temperature it suffices to consider $T$ satisfying $|T-T_c|\leq C_1 h^2$. Moreover, it clearly suffices to consider functions $\Phi$ with $\|\Phi\|=1$ satisfying
$$
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle \leq C_2 h^2
$$
(for if there are no such $\Phi$, then the theorem is trivially true). According to Theorem~\ref{decomp}, for any parameter $\epsilon\in [h^2,h^2_0]$, $\Phi$ can be decomposed as
$$
\Phi = \phi_*(r) \psi_\leq(X) + \sigma \,.
$$
Thus,
\begin{align*}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle = I_1 + I_2 + I_3
\end{align*}
with
\begin{align*}
I_1 &:= \left\langle \phi_*(r) \psi_\leq(X), \left(1- V^{1/2} L_{T,W} V^{1/2}\right) \phi_*(r) \psi_\leq(X) \right\rangle, \\
I_2 &:= \left\langle \sigma, \left(1- V^{1/2} L_{T,W} V^{1/2}\right) \sigma \right\rangle, \\
I_3 &:= 2\re \left\langle\sigma, \left(1- V^{1/2} L_{T,W} V^{1/2}\right)\phi_*(r) \psi_\leq(X) \right\rangle.
\end{align*}
The term $I_1$ is the main term and can be treated exactly as in the proof of the lower bound on the critical temperature. We obtain
$$
I_1 \geq -\Lambda_2 \frac{T_c-T}{T_c}\|\psi_\leq\|^2 + \left\langle \psi_\leq, \left( \Lambda_0 p_X^2 + \Lambda_1 h^2 W(hX) \right) \psi_\leq \right\rangle
- C \epsilon h^2 \,.
$$
The fact that the error $h^3$ is replaced by $\epsilon h^2$ comes from the bound $\|p_X^2 \psi_\leq\|^2 \lesssim \epsilon h^2$ from \eqref{eq:decomppsibounds}.
Let us therefore bound the error terms $I_2$ and $I_3$. Using the operator inequality from Proposition \ref{opineq}, dropping the non-negative term $\kappa (1-Q)$ and using the bound \eqref{eq:decompsigmabound} on $\sigma$, we obtain
$$
I_2 \gtrsim - h^2 \|\sigma\|^2 \gtrsim - \epsilon^{-1} h^4 \,.
$$
In order to bound $I_3$ we use the first bound in Lemma \ref{lemmadiffw0} and the bounds \eqref{eq:decompsigmabound} and \eqref{eq:decompsilower} on $\sigma$ and $\psi_\leq$ to obtain
\begin{align*}
I_3 & \geq 2\re \left\langle\sigma, \left(1- V^{1/2} L_{T,0} V^{1/2}\right)\phi_*(r) \psi_\leq(X) \right\rangle - C h^2 \|\sigma\| \|\phi_*(r)\psi_\leq(X) \| \\
& \geq 2\re \left\langle\sigma, \left(1- V^{1/2} L_{T,0} V^{1/2}\right)\phi_*(r) \psi_\leq(X) \right\rangle - C' \epsilon^{-1/2} h^3 \,.
\end{align*}
To bound the first term on the right side we decompose $\sigma=\sigma_0 +(\sigma-\sigma_0)$. We claim that
$$
\left\langle\sigma_0, \left(1- V^{1/2} L_{T,0} V^{1/2}\right)\phi_*(r) \psi_\leq(X) \right\rangle =0 \,.
$$
Indeed, to see this, we note that for fixed $r$, the Fourier transforms of $\sigma_0(X+r/2,X-r/2)$ and $V(r)^{1/2} \sigma_0(X+r/2,X-r/2)$ with respect to the variable $X$ are supported in $\{ p_X^2>\epsilon\}$ and likewise the Fourier transforms of $\phi_*(r)\psi_\leq(X)$ and $V(r)^{1/2} \phi_*(r)\psi_\leq(X)$ with respect to the variable $X$ are supported in $\{ p_X^2\leq \epsilon\}$. Thus $\left\langle\sigma_0, \phi_*(r) \psi_\leq(X) \right\rangle =0$, and the full claim follows by observing that the operator $L_{T,0}$ acts diagonally in Fourier space with respect to the $X$ variables, see \eqref{defpX}.
Thus, it remains to bound the term with $\sigma-\sigma_0$. We decompose $L_{T,0}=\chi_{\beta}(p_r^2-\mu) + (L_{T,0} - \chi_{\beta}(p_r^2-\mu))$ and, using the fact that $(1 - V^{1/2} \chi_{\beta_c}(p_r^2-\mu) V^{1/2}) \phi_*=0$, we find
\begin{align*}
& \left\langle \sigma -\sigma_0, \left(1- V^{1/2} L_{T,0} V^{1/2}\right)\phi_*(r) \psi_\leq(X) \right\rangle \\
& \qquad = \left\langle \sigma -\sigma_0, V^{1/2} \left( \chi_{\beta_c}(p_r^2-\mu) - \chi_{\beta}(p_r^2-\mu) \right) V^{1/2} \phi_*(r) \psi_\leq(X) \right\rangle \\
& \qquad \quad - \left\langle \sigma -\sigma_0, V^{1/2} \left(L_{T,0} - \chi_\beta(p_r^2-\mu) \right) V^{1/2} \phi_*(r) \psi_\leq(X) \right\rangle \,.
\end{align*}
Using inequality \eqref{eq:chiineq}, as well as the bounds \eqref{eq:diffsigmasigma0} and \eqref{eq:decompsilower} on $\sigma-\sigma_0$ and $\psi_\leq$, we find
\begin{align*}
\left\langle \sigma -\sigma_0, V^{1/2} \left( \chi_{\beta_c}(p_r^2-\mu) - \chi_{\beta}(p_r^2-\mu) \right) V^{1/2} \phi_*(r) \psi_\leq(X) \right\rangle
& \gtrsim - h^2 \|\sigma-\sigma_0\| \|\psi_\leq \| \\
& \gtrsim - h^3 \,.
\end{align*}
The remaining term we bound similarly using Lemma \ref{Ltleading}, as well as the bounds \eqref{eq:diffsigmasigma0} and \eqref{eq:decompsilower} on $\sigma-\sigma_0$ and $\psi_\leq$,
\begin{align*}
- \left\langle \sigma -\sigma_0, V^{1/2} \left(L_{T,0} - \chi_\beta(p_r^2-\mu) \right) V^{1/2} \phi_*(r) \psi_\leq(X) \right\rangle
& \gtrsim - \|\sigma-\sigma_0\| \|p_X^2\psi_\leq\| \\
& \gtrsim - \epsilon^{1/2} h^2 \,.
\end{align*}
To summarize, we have shown that
\begin{align*}
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle
& \geq
-\Lambda_2 \frac{T_c-T}{T_c}\|\psi_\leq\|^2 + \left\langle \psi_\leq, \left( \Lambda_0 p_X^2 + \Lambda_1 h^2 W(hX) \right) \psi_\leq \right\rangle \\
& \qquad - C h^2 \left( \epsilon + \epsilon^{-1} h^2 + \epsilon^{-1/2} h + h + \epsilon^{1/2} \right).
\end{align*}
In order to minimize the error we choose $\epsilon = h$. With this choice we obtain, recalling also the lower bound on $\|\psi_\leq\|$ from \eqref{eq:decompsilower},
$$
\langle\Phi, (1- V^{1/2} L_{T,W} V^{1/2})\Phi\rangle
\geq \left\langle \psi_\leq, \left( \Lambda_0 p_X^2 + \Lambda_1 h^2 W(hX)
-\Lambda_2 \frac{T_c-T}{T_c} - C' h^{5/2} \right) \psi_\leq \right\rangle
$$
By definition of $D_c$ plus a rescaling we can bound the right side from below by
$$
\left( h^2 \Lambda_2 D_c - \Lambda_2\frac{T_c-T}{T_c} - C' h^{5/2} \right) \|\psi_\leq\|^2 \,.
$$
Recalling that $\|\psi_\geq\|\neq 0$, we conclude that this is $>0$ provided $T> T_c(1- D_c h^2 + (C'/\Lambda_2) h^{5/2})$. This concludes the proof of the upper bound on the critical temperature.
\bibliographystyle{amsalpha}
|
\section{Introduction and model}
\label{sec:intro}
\par We consider a one-dimensional simple random walk surviving among a random field of static soft traps~: each time it meets a trap the walk is killed with probability $1-e^{-\beta}$ or survives with probability $e^{-\beta}$, where $\beta$ is a positive and fixed parameter, see Section~\ref{sec:rwrt} for a precise definition. The increments between consecutive traps, or gaps, are assumed to be mutually independent and independent from the walk, with a power-law decaying tail. This is what we refer to as {\it power-law renewal traps} in the title, see Section~\ref{sec:traps} for a precise definition. To be more precise, we deal with the {\it quenched} version of the model, meaning that the positions of the traps are frozen and the survival probability is computed only with respect to the law of the random walk.
Our main result, Theorem~\ref{thm0}, states a convergence in law for the properly rescaled logarithm of the quenched survival probability, seen as a random variable with respect to the field of traps, as time goes to infinity. The limiting law writes as a variational formula involving (i) a Poisson point process that emerges as the universal scaling limit of the properly rescaled gaps and (ii) a function of the parameter $\beta$ that we call {\it asymptotic cost of crossing per trap} and that may, in principle, depend on the details of the gap distribution, see the definition of $\lambda(\beta)$ in Proposition~\ref{pr:lambda.beta}. Even if we offer no path statement for the walk conditioned on surviving, our proof strongly suggests a {\it confinement strategy} according to which the walk remains in a large gap with an appropriate scale. Path localization could be considered as future work.
\par As we will see in Section~\ref{sec:polymer}, our model may also been seen as a $(1+1)$-directed polymer among many repulsive interfaces, in which case $\beta$ corresponds to the strength of repulsion. We will take advantage of this connection by using and revisiting some estimates obtained by Caravenna and Pétrélis~\cite{CP09b,CP09} in the case of periodic traps. We point out that the logarithm of the survival probability is the finite-volume free energy of the corresponding polymer model.\\
\par {\it Outline.} Section~\ref{sec:intro} contains the mathematical definition of the model and a discussion on the relation with other models, such as the directed polymer among multiple interfaces. Section~\ref{sec:result} contains the statement of Theorem~\ref{thm0}, which is our main result. At the beginning of the section we introduce several mathematical objects which are necessary for the understanding of the theorem. Comments and related open questions are listed at the end of Section~\ref{sec:result}. Sections~\ref{sec:KT} to~\ref{sec:pot} constitute the proof of the theorem. Key tools are gathered in Section~\ref{sec:KT}. The rest of the proof is split into a lower bound part (Section~\ref{sec:LB}), an upper bound part (Section~\ref{sec:UB}) and a conclusion (Section~\ref{sec:pot}). The more technical proofs are deferred to an appendix.
\subsection{A random walk in ${\ensuremath{\mathbb Z}} $ among soft traps}
\label{sec:rwrt}
We consider $S=(S_n)_{n\ge 0}$ a simple random walk on ${\ensuremath{\mathbb Z}} $ in presence of traps. We recall that the increments $(S_n - S_{n-1})_{n\ge 1}$ are independent and identically distributed (i.i.d.) random variables which are uniformly distributed on $\{-1,1\}$, and we shall write $\bP_x$ for the law of the walk started at $S_0 =x$, for $x\in{\ensuremath{\mathbb Z}} $, with the notational simplification $\bP_0 =\bP$. The positions of the traps are integers and they will be denoted by $\tau = \{\tau_n\}_{n\ge 0}$ with $\tau_0=0$. The increments $(\tau_{n+1} - \tau_n)_{n\ge0}$ may be referred to as {\it gaps}.
\par Let $\beta>0$ be a parameter of the model. Informally, each time the walk meets a trap, it is killed with probability $1-e^{-\beta}$ or survives with probability $e^{-\beta}$ independently from the past. The traps are called {\it soft}, by opposition to {\it hard} traps, because the walk has a positive chance to survive when sitting on a trap. For a precise mathematical definition let us first introduce $(\theta_n)_{n\ge 1}$, the clock process recording the times when the random walk visits $\tau$, that is
\begin{equation}
\theta_0 = 0,\quad \theta_{n+1} = \inf\{k > \theta_n \colon S_k \in \tau\},\quad n\ge 0.
\end{equation}
We enlarge the probability space so as to include a $\mathbb{N}$-valued geometric random variable ${\ensuremath{\mathcal N}} $ with success parameter $1-e^{-\beta}$. This plays the role of the clock that kills the walk after its ${\ensuremath{\mathcal N}} $-th meeting with the set of traps. We now define $\sigma$ the death time of the walk by
\begin{equation}
\sigma = \theta_{{\ensuremath{\mathcal N}} }.
\end{equation}
Note that our probability law now depends on the parameter $\beta$. We shall write $\bP^\beta_x$ when we want to stress this dependence or omit the superscript when no confusion is possible. Again we may write $\bP^\beta$ instead of $\bP_0^\beta$. We also point out that $\sigma$ depends on $\tau$ through $\theta$ even if it is not explicit in the notations.\\
\par The {\it hitting times} of the walk are defined by
\begin{equation}
H_x = \inf\{n\geq 1\colon S_n = x \}, \qquad x\in{\ensuremath{\mathbb Z}} ,
\end{equation}
and
\begin{equation}
H_{\mathbb{Z}^-} = \inf\{n\geq 1\colon S_n \leq 0 \}.
\end{equation}
\par In this paper we study the limiting behaviour of the probability that the walk survives up to time $n\in {\ensuremath{\mathbb N}} $, as $n$ gets large. For convenience, we consider walks that do not visit $\mathbb{Z}^-=-\mathbb{N}_0$. This extra condition allows to consider traps that are indexed by ${\ensuremath{\mathbb N}} _0$ instead of ${\ensuremath{\mathbb Z}} $ and does not hide anything deep nor change the main idea of the paper. Thus, our {\it survival probability} writes
\begin{equation}
\label{eq:defZn}
Z_n = \bP^{\beta}(\sigma \wedge H_{\mathbb{Z}^-} > n), \qquad n\in{\ensuremath{\mathbb N}} .
\end{equation}
We stress again that $Z_n$ is a function of the environment of traps $(\tau_n)_{n\geq 0}$.
\subsection{(De)pinning of a $(1+1)$-directed polymer by multiple interfaces}
\label{sec:polymer}
\par By integrating on ${\ensuremath{\mathcal N}} $ in~\eqref{eq:defZn}, we obtain
\begin{equation}
Z_n = \bE\Big[\exp\Big(-\beta \sum_{k=1}^n {\sf 1}_{\{S_k \in \tau\}}\Big) {\sf 1}_{\{H_{\mathbb{Z}^-}>n\}} \Big].
\end{equation}
This expression links the survival probability to a certain polymer model from statistical mechanics. More precisely, the expression above is the partition function of a $(1+1)$-directed polymer above an impenetrable wall and among many repulsive interfaces. Here, $(k,S_k)_{0\le k\le n}$ plays the role of a polymer with $n$ monomers and the parameter $n$, which is initially a time parameter, becomes the size of the polymer, see Figure~\ref{polymer}. Whenever the polymer touches one of the interfaces, located at the levels $\tau = \{\tau_n\}_{n\ge 0}$, it is penalized by a factor $e^{-\beta}$. Finally, the event $\{H_{\mathbb{Z}^-} > n\}$ reflects the presence of a hard wall at level $0$.
\begin{figure}[!h]
\centering
\begin{tikzpicture}[scale=1]
\draw[->] (-0,0) -- (10,0);
\draw [->] (0,-0.5) -- (0,5.5);
\draw (0,0) node[left] {$0$};
\draw (0,5.5) node[left] {$(\tau_i)_{i\geq 0}$};
\draw (10,0) node[right] {$\mathbb{N}$};
\draw (0,0.4) node[left] {$\tau_1$};
\draw (0,3.6) node[left] {$\tau_i$};
\draw (0,4.5) node[left] {$\tau_{i+1}$};
\draw (2,4.05) node[left] {$T_{i+1}$};
\draw [<->] (2,3.6) -- (2,4.5);
\draw [dashed] (0,1) -- (10,1);
\draw [dashed] (0,0.4) -- (10,0.4);
\draw [dashed] (0,0.8) -- (10,0.8);
\draw [dashed] (0,2.4) -- (10,2.4);
\draw [dashed](0,3.1) -- (10,3.1);
\draw [dashed](0,3.6) -- (10,3.6);
\draw [dashed](0,4.5) -- (10,4.5);
\draw[dashed] (0,5) -- (10,5);
\draw [dashed](0,4.8) -- (10,4.8);
\fill[pattern=north east lines] (0,0) -- (9.9,0) -- (9.9,-0.5)-- (0,-0.5) ;
\draw [very thick] (0,0) -- (0.1,0.1)--(0.2,0.2)--(0.3,0.1)--(0.4,0.2)--(0.5,0.3)--(0.6,0.4)--(0.7,0.5)--(0.8,0.4)--(0.9,0.5)--(1.0,0.6)--(1.1,0.7)--(1.2,0.6)--(1.8,1.2)--(1.9,1.1)--(2.4,1.6)--(2.6,1.4)--(2.7,1.3)--(2.9,1.5)--(3.5,2.1)--(3.8,1.8)--(4.1,2.1)--(4.3,1.9)--(4.6,2.2)--(5,1.8)--(5.4,2.2)--(7.1,3.9)--(7.2,3.8)--(7.5,4.1)--(7.8,3.8)--(7.9,3.9)--(8,3.8)--(8.6,4.4)--(8.8,4.2)--(8.9,4.3)--(9.5,3.7)--(9.8,4);
\end{tikzpicture}
\caption{Example of a polymer among repulsive interfaces. The dashed lines correspond to the interfaces, the thick one to the polymer and the shaded area to the hard wall.}
\label{polymer}
\end{figure}
\par There is a rich literature about the pinning phenomenon in polymer models, for which we refer to~\cite{dH09, Gia07, Gia11}. In general, the parameter $\beta$ can have any sign (thus $\beta<0$ corresponds to the attractive case with our notations) and in the simplest case of a unique interface at level zero without wall, one observes a localization transition at $\beta=0$. The case $\beta<0$ is known as the \textit{repulsive} or \textit{delocalized} phase as the polymer typically touches the interface only finitely many times. The case $\beta >0$ is called instead the \textit{attractive} or \textit{localized} phase as the polymer typically touches the interface a positive fraction of time. Caravenna and Pétrélis studied the case of multiple interfaces periodically located at levels $t_n{\ensuremath{\mathbb Z}} $, where the period $t_n\ge 1$ depends on the size of the polymer $n$, both in the attractive~\cite{CP09} and repulsive case~\cite{CP09b}. In contrast, the positions of our interfaces are random, but do not vary in $n$. However, the size of the {\it relevant gaps} does depend on $n$, which explains why we use (and extend) some estimates from~\cite{CP09b}. Note also the difference between our results, since Caravenna and Pétrélis obtained results at the level of paths, which means information on the (de)localization of the path of the random walk under the polymer measure
\begin{equation}
\frac{\dd \bP_n^{\beta}}{\dd \bP} \propto \exp\Big(-\beta \sum_{k=1}^n {\sf 1}_{\{S_k \in \tau\}}\Big),
\end{equation}
(where $\propto$ means ``proportional to'') while this is not the purpose of the present paper, see Comment $8$ in Section~\ref{subsec:c}.
\subsection{Assumption on the traps}
\label{sec:traps}
We now put a probability measure ${\ensuremath{\mathbb P}} $ on the environment of traps.
We denote by $T_k = \tau_k - \tau_{k-1}$, for $k\in{\ensuremath{\mathbb N}} $, the increments, that is the size of the intervals between two consecutive traps, which we call {\it gaps}. We assume that, under ${\ensuremath{\mathbb P}} $, $\tau$ is a discrete renewal process, that is the $(T_k)$'s are i.i.d. ${\ensuremath{\mathbb N}} $-valued random variables. We further assume that $\tau_0=0$ and that the increments have a power-tail distribution:
\begin{equation}
\label{eq:tail.ass}
{\ensuremath{\mathbb P}} (T_1 = n) \sim c_\tau\ n^{-(1+\gamma)},\qquad \gamma>0,\quad n\to\infty,
\end{equation}
where $c_\tau$ is a positive constant.
We recall the following standard limit theorem, see e.g.\ Petrov~\cite[Theorem 14, p. 91]{Petrov}.
\begin{proposition}
\label{thm:limit.stable}
If $\gamma\in(0,1]$, the sequence $(\tau_n/ n^{1/\gamma})_{n\ge 1}$ converges in law to a (totally skewed to the right) $\gamma$-stable random variable with scale parameter $c_\tau$. If $\gamma>1$ then $(\tau_n/n)_{n\ge 1}$ converges almost-surely to ${\ensuremath{\mathbb E}} (T_1)$.
\end{proposition}
\subsection{Parabolic Anderson model with a correlated potential}
Our model is also connected to a discrete analogue of the parabolic Anderson model (PAM) with potential $V(x) = -\beta {\sf 1}_{\{x\in \tau\}} - \infty {\sf 1}_{\{x \le 0\}}$, that is the heat equation with random potential $V$,
\begin{equation}
\partial_t u=\Delta u+ V u.
\end{equation}
There is a rich literature and intense activity around the PAM. We refer to König~\cite{Ko16} for a recent survey on this topic. Note that the potential is usually chosen as a sequence of random variables that are independent in the space parameter. In contrast, our potential exhibits long-range spatial correlations, that is one of the research direction suggested in~\cite[Section 7.2]{Ko16}. For a review of potentials in discrete and continuous space, we refer to ~\cite[Section 1.5]{Ko16}.
\par Let us end this section with a reference to the classical monograph by Sznitman~\cite{Sz98} on random motions in random media. Chapter 6 is of particular interest to us as it highlights the link with directed polymers in the presence of columnar defects and introduces the concept of pinning effect of quenched path measures, see again Comment $8$ in Section~\ref{subsec:c}.
\section{Results}
\label{sec:result}
In this section we first introduce the various tools needed to state our main theorem. We first prove the existence of the asymptotic exponential cost for the walker to cross a new trap without being killed (Section \ref{subsec:acct}). Then, in Section \ref{subsec:coe}, we consider the environment as a point process and investigate some properties of its asymptotic limit that is a Poisson point process with explicit intensity measure. We finally state our result in Section \ref{subsec:sor} and conclude with a few comments in Section \ref{subsec:c}.
\subsection{Asymptotic cost of crossing traps}
\label{subsec:acct}
We first recall without proof the following standard proposition:
\begin{proposition}
\label{pr:rw_exit.dist}
For all $t\in{\ensuremath{\mathbb N}} $,
\begin{equation}
\bP_1(H_t < H_0) = 1/t.
\end{equation}
\end{proposition}
Let us define, for $\ell\ge 1$, the random variable (with respect to $\tau$)
\begin{equation}
\label{def:gl.ell.gb}
\lambda(\ell,\beta) = -\frac{1}{\ell} \log \bP^{\beta}(H_{\tau_\ell} < H_{0} \wedge \sigma).
\end{equation}
Note that in the above definition $H_{0}$ could be safely replaced by $H_{\mathbb{Z}^-}$.
The next proposition gives the existence of an {\it asymptotic cost of crossing per trap}, which will play a crucial role in our theorem.
\begin{proposition}
\label{pr:lambda.beta}
For all $\beta >0$ there exists a positive constant $\lambda(\beta) = \lambda(\beta,{\ensuremath{\mathbb P}} )$ such that, ${\ensuremath{\mathbb P}} $-a.s.\ and in $L_1({\ensuremath{\mathbb P}} )$,
\begin{equation}
\lim_{\ell\to\infty} \lambda(\ell,\beta) = \lambda(\beta),
\end{equation}
with
\begin{equation}
\label{eq:encadr.lambda}
0 \le \lambda(\beta) - \beta \le {\ensuremath{\mathbb E}} (\log T_1)+\log 2.
\end{equation}
\end{proposition}
Note that $\log T_1$ is integrable because of~\eqref{eq:tail.ass}.
\begin{proof}[Proof of Proposition~\ref{pr:lambda.beta}]
Let us define a collection of random variables indexed by:
\begin{equation}
{\ensuremath{\mathcal Z}} (i,j) = - \log \bP_{\tau_i}(H_{\tau_j} < H_{\tau_i} \wedge \sigma), \qquad 0\le i < j.
\end{equation}
Thus, we are interested in the limit of $({\ensuremath{\mathcal Z}} (0,\ell)/\ell)_{\ell\ge 1}$. Let $1\le i<j<k$. To go from $\tau_i$ to $\tau_k$ without being killed, one strategy is to go first from $\tau_i$ to $\tau_j$ and survive, then from $\tau_j$ to $\tau_k$ and survive without coming back to $\tau_j$, which, by the Markov property, leads to the inequality
\begin{equation}
\label{eq:subadd}
{\ensuremath{\mathcal Z}} (i,k) \le {\ensuremath{\mathcal Z}} (i,j) + {\ensuremath{\mathcal Z}} (j,k).
\end{equation}
By stationarity of the sequence $(T_k)_{k\ge 1}$, ${\ensuremath{\mathcal Z}} (i,j)$ has the same law as ${\ensuremath{\mathcal Z}} (0,j-i)$. Moreover, by Proposition \ref{pr:rw_exit.dist},
\begin{equation}
{\ensuremath{\mathcal Z}} (i-1,i) \le \beta + \log T_i + \log 2,\qquad i\in{\ensuremath{\mathbb N}} ,
\end{equation}
and one gets by iterating~\eqref{eq:subadd}
\begin{equation}
\beta\ell \le {\ensuremath{\mathcal Z}} (0,\ell) \le (\beta+\log 2)\ell + \sum_{1\le i\le \ell} \log T_i
\end{equation}
(the lower bound follows simply from the fact that surviving to a newly visited trap costs at least $e^{-\beta}$). We may now conclude with Kingman's sub-additive ergodic theorem (see Theorem 7.4.1 in \cite{Durrett}) and the law of large numbers.
\end{proof}
\begin{proposition}\label{lem:continuity_lambda}
The function $\beta\to\lambda(\beta)$ is continuous on $\mathbb{R}_+^*$.
\end{proposition}
\begin{proof}[Proof of Proposition~\ref{lem:continuity_lambda}]
We prove that the function is concave on $(0,+\infty)$: as it is finite, it implies continuity. We observe by integrating over ${\ensuremath{\mathcal N}} $ that
\begin{equation}
\lambda(\ell, \beta)=-\frac{1}{\ell}\log\bE\Big(\exp\Big(-\beta \sum_{k=1}^{H_{\tau_\ell}}1_{\{S_k\in \tau\}} \Big) {\sf 1}_{\{H_{\tau_\ell} < H_{0}\}}\Big).
\end{equation}
A basic interchange theorem allows us to write
\begin{equation}
\partial_{\beta}^2 \lambda(\ell,\beta)= -\frac{1}{\ell}\var_{\tilde \bP_\beta}\Big(\sum_{k=1}^{H_{\tau_\ell}}1_{\{S_k\in \tau\}}\Big) \le 0,
\end{equation}
where $\tilde \bP_\beta$ is absolutely continuous with respect to $\bP$, with Radon-Nikodym derivative:
\begin{equation}
\frac{\dd \tilde \bP_\beta}{\dd \bP} = \exp\Big( - \beta \sum_{k=1}^{H_{\tau_\ell}}1_{\{S_k\in \tau\}} \Big) {\sf 1}_{\{H_{\tau_\ell} < H_{0}\}}.
\end{equation}
We deduce thereby that $\beta \mapsto \lambda(\ell,\beta)$ is concave. Therefore $\lambda$ is concave as the almost-sure limit of concave functions.
\end{proof}
\subsection{Convergence of the environment}
\label{subsec:coe}
In this section we recall a few elements of point processes, see~\cite{Res} for more background on this theory.
\par Define the quadrant $E:=[0,+\infty)\times (0,+\infty)$ and consider $\mathcal{E}$ the Borel $\sigma-$algebra on $E$. We say a measure $\mu$ on $(E,\mathcal{E})$ to be a point measure if $\mu$ can be written
\begin{equation}
\label{eq:mp}
\mu:=\sum_{i=1}^{+\infty}\delta_{z_i}
\end{equation}
where, for any $a$ in $E$, $\delta_a$ denotes the Dirac measure in $a$ and $(z_i)_{i\geq 1}$ is a family of points in $E$. If $\mu$ can be written as in \eqref{eq:mp} we say that \textit{$z_i \in \mu$} ($i\geq 1$) even if we should say that $z_i$ is in the support of $\mu$.
\par We call $M_p(E)$ the set of all Radon point measures on $E$, that are the point measures $\mu$ such that $\mu(K)<+\infty$ for all compact sets $K\subset E$. We endow $M_p(E)$ with the $\sigma-$algebra $\mathcal{M}_p(E)$ defined as the smallest $\sigma-$algebra that makes applications $\mu \mapsto \mu(F)$ measurable for all $F \in \mathcal{E}$.
Let $C^+_K(E)$ be the set of continuous non-negative functions on $E$ with compact support. A sequence $(\mu_n)_{n\geq 1}$ in $M_p(E)$ is said to converge vaguely to $\mu$, which we note $\mu_n \stackrel{v}{\to} \mu$, if for any $f$ in $C^+_K(E)$
\begin{equation}
\int f\ \dd\mu_n \to \int f\ \dd\mu,\qquad n\to +\infty.
\end{equation}
This provides a topology on $M_p(E)$ that turns out to be metrisable, separable, and complete.
In this context, a sequence of probability measures $(P_n)_{n\geq 1}$ on $\mathcal{M}_p(E)$ converges weakly to $P$, which we note $P_n \stackrel{w}{\to} P$, if for every $\Theta$ vaguely continuous and bounded on $M_p(E)$,
\begin{equation}
\int \Theta\ \dd P_n \to \int \Theta\ \dd P,\qquad n\to +\infty.
\end{equation}
\par We now come back to our context. For $n\geq 1$ we define
\begin{equation}
\begin{aligned}
&(X^n_i,Y^n_i) := \Big(\frac{i-1}{n}, \frac{T_i}{n^{1/\gamma}}\Big) \qquad \textrm{for all } i\geq 1,\\
&\textrm{and }\qquad \Pi_n =\sum_{i=1}^{+\infty}\delta_{(X_i^n,Y_i^n)}.
\end{aligned}
\end{equation}
We observe that $\Pi_n$ is a random variable that takes values in $M_p(E)$.
Recall the definition of $c_\tau$ in \eqref{eq:tail.ass} and define $\Pi$ to be a Poisson point process on $E$ with intensity
\begin{equation}
\label{eq:def_p}
p:=\dd x \otimes \frac{c_{\tau}\gamma}{y^{\gamma+1}}\ \dd y,
\end{equation}
that is, for all finite families of disjoint events $(A_i)_{1\leq i \leq n} \in \mathcal{E}^n$, $(\Pi(A_i))_{1\leq i \leq n}$ are independent Poisson random variables with respective parameters $(p(A_i))_{1\leq i \leq n}$.
\begin{proposition}
\label{prop:convergenceP}
It holds that
$
\Pi_n \stackrel{w}{\rightarrow} \Pi.
$
\end{proposition}
\begin{proof}
The proof follows from \cite[Proposition~3.21 p.~154]{Res} once noticed that for all $y>0$, the sequence $(n {\ensuremath{\mathbb P}} (T_1>y n^{1/\gamma}))_{n\geq 1}$ converges to $c_\tau / y^\gamma$ when $n$ goes to infinity.
\end{proof}
We define for any $\lambda>0$,
\begin{equation}
\label{eq:psi0}
\begin{array}{ccccc}
\psi^\lambda & : & E & \to & \mathbb{R}^+ \\
& & (x,y) & \mapsto & \lambda x + \frac{\pi^2}{2 y^2}, \\
\end{array}
\end{equation}
and for any $\mu$ in $M_p(E)$,
\begin{equation}
\label{eq:psi}
\Psi^\lambda (\mu):=\inf_{(x,y)\in \mu} \psi^\lambda(x,y).
\end{equation}
We can now define the variable that appears as the limit law in Theorem \ref{thm0},
\begin{equation}
F := \Psi^{\lambda(\beta)}(\Pi) = \inf_{(x,y) \in \Pi} \Big\{\lambda(\beta)x + \frac{\pi^2}{2y^2}\Big\},
\end{equation}
where we remind that $(x,y) \in \Pi$ means that $(x,y)$ is in the support of $\Pi$.
We shall write $F^\beta$ instead of $F$ when we want to stress the dependence on this parameter.
\par The rest of the section is devoted to various results relative to the environment $\Pi$.
\begin{proposition}
\label{prop:unique}
For all $\lambda>0$, $\Psi^{\lambda}(\Pi)$ is almost surely well defined and positive. Moreover the infimum in the definition of $\Psi^{\lambda}$ is almost surely achieved at a unique point $(x^*,y^*)$ in $E$.
\end{proposition}
\begin{proof}[Proof of Proposition~\ref{prop:unique}]
\par As there is almost surely at least one point in $\Pi$, the infimum is well defined. Moreover, there are almost surely finitely many points in $[0,1]\times [1,+\infty)$, which implies that $\bar Y$, the maximum of the second coordinates among these points, is almost surely finite. Thus, $\psi^\lambda(x,y)\geq \min\{{\pi^2}/{(2\bar Y^2)}, \lambda\}$ for any $(x,y)\in \Pi$, which proves that $\Psi^{\lambda}(\Pi)$ is almost surely positive.
\par For $\lambda >0$ and $u>0$, we introduce the set
\begin{equation}
\label{eq:defA}
A^\lambda_{u}:=\{(x,y)\in E \textrm{ such that }\psi^{\lambda}(x,y)\leq u\},
\end{equation}
and note that $\Pi (A^\lambda_{u})<+\infty$ almost surely.
\par Let us denote by $C$ the event that the minimum in the definition of $\Psi$ is achieved. We observe that for any $u>0$,
\begin{equation}
P(\Psi(\Pi)<u)=P(\Psi(\Pi)<u, \Pi(A_u^\lambda)<+\infty)\leq P(\Psi(\Pi)<u, \Pi \in C).
\end{equation}
As $\lim_{u\to +\infty}P(\Psi(\Pi)<u)=1$ and $\lim_{u \to +\infty}P(\Psi(\Pi)<u, \Pi \in C)=P(\Pi \in C)$, we obtain that $P(\Pi \in C)=1$.
\par It remains to prove that the infimum is almost surely achieved at only one point.
For $u>0$, let $D_{u}$ be the event that two points of $\Pi \cap A_u^\lambda$ have the same image by $\psi^{\lambda}$ and $D$ be the event that two points of $\Pi$ have the same image by $\psi^{\lambda}$.
For $n\in\mathbb{N}$ and conditional on $\{\Pi(A_u^\lambda)=n\}$, the restriction of $\Pi$ to $A_u^\lambda$ has the same law as $\sum_{i=1}^n \delta_{X_i}$, where the $(X_i)_{1\le i\leq n}$'s are i.i.d. with continuous law $p$ restricted to $A_u^\lambda$ and renormalised to a probability measure. This implies that $P(\Pi\in D_u | \Pi(A_u^\lambda))=0$ and thus $P(\Pi\in D_u)=0$. We obtain
\begin{equation}
P(\Psi(\Pi)<u) = P(\Psi(\Pi)<u, \Pi \in D_u^c).
\end{equation}
The first term converges to $1$ while the second one converges to $P(\Pi \in D)$ when $u$ goes to infinity. This proves $P(\Pi \in D)=1$ and concludes the proof.
\end{proof}
\par Unfortunately, it is difficult to work directly with $\Psi^\lambda$ as it appears not to be a vaguely continuous function on $M_p(E)$. For this reason, we introduce the function $\Psi^{\lambda}_K$, for any compact set $K\subset E$, defined by
\begin{equation*}
\Psi^\lambda_K (\mu):=\inf_{(x,y)\in \mu \cap K} \psi^\lambda(x,y),\qquad \mu\in M_p(E).
\end{equation*}
The advantage of restricting the infimum to a compact set lies in the following lemma.
\begin{lemma}\label{prop:continuity}
For any compact set $K\subset E$ and $\lambda>0$, the function $\Psi^\lambda_K$ is vaguely continuous on $M_p(E)$.
\end{lemma}
\begin{proof}
Consider a sequence $(\mu_n)_{n\geq 1}$ in $M_p(E)$ that converges vaguely to $\mu$. Due to Proposition~$3.14$ in \cite{Res}, $\mu \in M_p(E)$. Suppose $\mu(K)=k$, that is $\mu(\cdot \cap K)$ writes $\sum_{i=1}^k \delta _{(x_i,y_i)}$ where $(x_i,y_i)_{1 \leq i \leq k}$ is a family in $K$. By Proposition~$3.13$ in \cite{Res} there exists for all $n$ larger than some $n(K)$ a family $(x^n_i,y^n_i)_{1 \leq i \leq k}$ such that $\mu_n(\cdot \cap K)=\sum_{i=1}^k \delta _{(x^n_i,y^n_i)}$. Moreover, for all $1\leq i \leq k$, the sequence
$(x^n_i,y^n_i)_{n\geq 1}$ converges to $(x_i,y_i)$ as $n$ goes to infinity. This implies that $\Psi^{\lambda}_K(\mu_n)$ converges to $\Psi^{\lambda}_K(\mu)$.
\end{proof}
\par We conclude this section with the following technical lemma:
\begin{lemma}
\label{lem:Fgep}
The family $(F^{\beta-\epsilon})_{\gep \geq 0}$ (seen as functions of the random measure $\Pi$) converges non-decreasingly to $F^{\beta}$ when $\epsilon\to 0$, almost surely.
\end{lemma}
\begin{proof}
Using Proposition \ref{prop:unique} we can define almost surely a random point $(X^*(\beta),Y^*(\beta))$ such that
\begin{equation}
\psi^{\lambda(\beta)}(X^*(\beta),Y^*(\beta))=\Psi^{\lambda(\beta)}(\Pi).
\end{equation}
We first prove that $(X^*(\beta-\epsilon),Y^*(\beta-\epsilon))= (X^*(\beta),Y^*(\beta))$ almost surely if $\epsilon>0$ is chosen small enough. Let $\gep_0\in(0,\beta)$.
Consider some $x\geq x_0:=\frac{2\Psi^{\lambda(\beta)}(\Pi)}{\lambda(\beta-\epsilon_0)}$ and any $y>0$. As $\lambda(\cdot)$ and $\Psi^{\lambda(\cdot)}$ are non-decreasing functions of $\beta$ it holds that
\begin{equation}
\psi^{\lambda(\beta-\epsilon)}(x,y) \geq \lambda(\beta-\epsilon) x \geq 2\Psi^{\lambda(\beta-\epsilon)}(\Pi)
\end{equation}
for all $\epsilon<\epsilon_0$, and we conclude that $X^*(\beta-\epsilon)<x_0$ for any $\epsilon<\epsilon_0$.
\par Consider now some $y\leq y_0:= \frac{\pi}{2\sqrt{\Psi^{\lambda(\beta)}(\Pi)}}$ and any $x>0$.
We thus obtain
\begin{equation}
\psi^{\lambda(\beta-\epsilon)}(x,y) \geq \frac{\pi^2}{2 y^2} \geq 2\Psi^{\lambda(\beta-\epsilon)}(\Pi),
\end{equation}
and we conclude that $Y^*(\beta-\epsilon)>y_0$ for any $\epsilon<\epsilon_0$.
We deduce thereof that almost surely
\begin{equation}
F^{\beta-\epsilon}= \inf_{(x,y)\in \Pi \cap \{x<x_0 , y>y_0\}}\psi^{\lambda(\beta-\epsilon)}(x,y).
\end{equation}
Finally, observe that $\Pi (x<x_0 , y>y_0) <+\infty$ almost surely, so that there are only finitely many candidates for $(X^*(\beta-\epsilon),Y^*(\beta-\epsilon))$. Pick $(X,Y)\in \Pi \cap \{x<x_0 , y>y_0\}$ that is not $ (X^*(\beta),Y^*(\beta))$ (if there is no such point there is nothing more to prove as $(X^*(\beta),Y^*(\beta))$ is then the only candidate). The function $\epsilon \mapsto \lambda(\beta-\epsilon) X + \frac{\pi^2}{2Y^2}$ is (i) strictly larger than $\lambda(\beta) X^*(\beta) + \frac{\pi^2}{2(Y^*(\beta))^2}$ at $\epsilon=0$ due to Proposition \ref{prop:unique} and (ii) continuous due to Proposition \ref{lem:continuity_lambda}. Therefore,
\begin{equation}
\lambda(\beta-\gep) X + \frac{\pi^2}{2Y^2}>\lambda(\beta) X^*(\beta) + \frac{\pi^2}{2(Y^*(\beta))^2}
\end{equation}
for $\epsilon>0$ small enough. As $\Pi( \{x<x_0 , y>y_0\}) <+\infty$ almost surely, we can choose $\epsilon$ small enough so that the last inequality holds for all points in $\Pi \cap \{x<x_0 , y>y_0\}$. Therefore, we may from now on consider $\epsilon>0$ small enough so that $(X^*(\beta-\epsilon),Y^*(\beta-\epsilon))= (X^*(\beta),Y^*(\beta))$.
\par From what precedes, we get
\begin{equation}
|\Psi^{\lambda(\beta-\epsilon)}(\Pi)-\Psi^{\lambda(\beta)}(\Pi)|\leq |\lambda(\beta-\epsilon)-\lambda(\beta)| X^*(\beta).
\end{equation}
Again, as $\lambda$ is continuous, $\Psi^{\lambda(\beta-\epsilon)}(\Pi)$ converges almost surely to $\Psi^{\lambda(\beta)}(\Pi)$ when $\epsilon$ goes to $0$. Finally,
\begin{equation}
F^{\beta-\epsilon} \underset{\epsilon \to 0}{\to} F^{\beta} \qquad a.s.
\end{equation}
Moreover, as $\lambda$ is non-decreasing with $\beta$, the convergence is monotone:
\begin{equation}
F^{\beta-\epsilon} \underset{\epsilon \to 0}{\nearrow} F^{\beta} \qquad a.s.
\end{equation}
\end{proof}
\subsection{Statement of the result}
\label{subsec:sor}
From now on we set
\begin{equation}
\label{eq:norm}
N = N(n) = n^{\frac{\gamma}{\gamma+2}},\qquad F_n = - \frac{1}{N} \log Z_n,\qquad n\ge1.
\end{equation}
\par Let us first explain at a heuristic level the choice of this normalization. The argument is of the type one uses to find volume exponents in some polymer models and is sometimes referred to as a {\it Flory argument}. We assume that at a large time $n$ the walk has visited at most $N$ traps and has remained confined in the largest visible gap, and we find the value of $N$ with the best energy-entropy balance. By basic extreme-value theory, the size of that gap is of order $N^{1/\gamma}$, and by a standard small-ball estimate (see Proposition~\ref{pr:small_ball} below for a precise statement) the entropic cost of being confined in that gap during time $n$ is of order $nN^{-2/\gamma}$. Also, the cost of crossing $N$ traps should be roughly of order $N$, see Proposition~\ref{pr:lambda.beta} below for a rigorous proposition. Finally, by equating these two costs, one finds the optimal choice $N= n^{\frac{\gamma}{\gamma +2}}$. As a consequence, the walk has travelled a distance of order $n^{\frac{1\vee \gamma}{\gamma+2}}$ from the origin (see Proposition \ref{thm:limit.stable}). This {\it confinement strategy} will be justified during the proof of our theorem, which we now state:
\begin{theorem}\label{thm0}
The sequence of random variables $(F_n)_{n\ge 1}$ converges in ${\ensuremath{\mathbb P}} $-distribution to the random variable
\begin{equation}
\label{eq:th}
F := \Psi^{\lambda(\beta)}(\Pi) = \inf_{(x,y) \in \Pi} \Big\{\lambda(\beta)x + \frac{\pi^2}{2y^2}\Big\},
\end{equation}
where $\Pi$ is a Poisson point process on ${\ensuremath{\mathbb R}} ^+\times{\ensuremath{\mathbb R}} ^+_*$ with intensity $p =\dd x \otimes c_{\tau} \gamma \ y^{-(1+\gamma)}\dd y$.
\end{theorem}
\par The rest of the article is devoted to the proof of this result. The general strategy is the following. In Section \ref{sec:KT} we introduce the notion of \textit{good environments}, that are environments for which we can provide suitable bounds on the survival probability $Z_n$, see~\eqref{eq:ge}. In Sections~\ref{sec:LB} and \ref{sec:UB} we successively prove lower and upper bounds, assuming only that the environment is good. We complete the proof in Section \ref{sec:pot}: first we show that we can adjust parameters so that environments are asymptotically good, with a probability arbitrarily close to one; then we use the two bounds obtained in the previous sections to relate the quantity of interest to an explicit functional $\Psi$ of the properly renormalized environment, see \eqref{eq:psi}. Finally we let $n$ tend to infinity to prove that $F$ has the same law as $\Psi$ applied to the limiting environment, that is a Poisson point process (see Proposition \ref{prop:convergenceP}).
\subsection{Comments}
\label{subsec:c}
We collect here a number of comments and remarks about our result.\\
{\noindent \bf 1.} We do not consider the annealed survival probability, which decreases at most polynomially fast since
\begin{equation}
{\ensuremath{\mathbb E}} \bP(\sigma \wedge H_{\mathbb{Z}^-}>n) \ge \frac 12 \bP_1(H_0 \geq n) {\ensuremath{\mathbb P}} (\tau_1 > n) \sim \frac 12 c_\tau\ n^{-\gamma-\frac 12},
\end{equation}
as $n\to \infty$, and thus has a completely different behaviour than the quenched probability.\\
{\noindent \bf 2.} Note that we cannot hope for better than weak convergence. Indeed, if $F_n$ would converge to $F$ almost-surely, then $F$ would be measurable with respect to the tail $\sigma$-algebra of the family $(T_i)_{i\ge 1}$. As the increments are independent, the latter is trivial due to the $0-1$ law, and $F$ would be deterministic.\\
{\noindent \bf 3.} In the case $\gamma\le 1$, the variational formula in \eqref{eq:th} admits an alternative representation in terms of a subordinator, which reads
\begin{equation}
F = \inf_{t\ge 0} \Big\{\lambda(\beta)t + \frac{\pi^2}{2(\Delta{\ensuremath{\mathcal S}} _t)^2}\Big\},
\end{equation}
where $({\ensuremath{\mathcal S}} _t)_{t\ge 0}$ is a $\gamma$-stable subordinator and $\Delta{\ensuremath{\mathcal S}} _t = {\ensuremath{\mathcal S}} _t - {\ensuremath{\mathcal S}} _t^- = {\ensuremath{\mathcal S}} _t - \lim_{u\to t^-} {\ensuremath{\mathcal S}} _u$.\\
{\noindent \bf 4.} We can compute explicitly the tail distribution function of the limiting law $F = \Psi^{\lambda(\beta)}(\Pi)$ in Theorem \ref{thm0}. Recall \eqref{eq:def_p}, \eqref{eq:psi0} and \eqref{eq:defA}. For any $u \ge 0$ (we write $\lambda$ instead of $\lambda(\beta)$ to lighten notations),
\begin{equation}
\begin{aligned}
{\ensuremath{\mathbb P}} (F \ge u)&={\ensuremath{\mathbb P}} (\Pi(A^\lambda_{u})=0)=\exp(- p(A^\lambda_{u})).
\end{aligned}
\end{equation}
Since $A^\lambda_u = \Big\{(x,y)\colon 0 \le x < u/\lambda,\ y\ge \frac{\pi}{\sqrt{ 2(u-\lambda x)}} \Big\}$, we get by a straightforward computation that
\begin{equation}
\label{eq:Fcont}
{\ensuremath{\mathbb P}} (F \ge u) = \exp\Big( - \frac{c_\tau}{\lambda(\beta)\pi^\gamma (\gamma+2)} (2u)^{\frac{\gamma}{2}+1} \Big),\qquad u\ge 0.
\end{equation}
{\noindent \bf 5.} The case $\gamma=0$ is left open. In this case, a gap distribution of the form \eqref{eq:tail.ass} is no longer appropriate and one should instead assume that ${\ensuremath{\mathbb P}} (T_1=n) \sim L(n)/n$, where $L$ is a non-negative slowly varying function such that $\sum L(n)/n$ is finite. Complications may arise at two levels~: (i) the normalization of $\max_{1\le i\le n} T_i$, that we use to guess the value of $N$, and (ii) the integrability of $\log T_1$, that we use in Proposition~\ref{pr:lambda.beta}. For instance, if $L(n) = (\log n)^{-2}$ then ${\ensuremath{\mathbb E}} (\log T_1) = \infty$ and $\max_{1\le i\le n} T_i$ has a completely different type of renormalization since, as one can readily show, $(1/\sqrt{n}) \log \max_{1\le i\le n} T_i$ converges to a non-trivial probability law with cumulative distribution function $x \mapsto \exp(-x^{-2}){\sf 1}_{\{x>0\}}$, as $n\to\infty$.\\
{\noindent \bf 6.} We state without proof an alternative expression for $\lambda(\beta)$ based on ergodic theory considerations. To this end, let $\tilde\tau$ be an independent copy of $\tau$, as defined in Section~\ref{sec:traps}. Suppose that the random walk is now free to visit $\mathbb{Z}^-$ but is killed by the set $-\tilde\tau$ (note the minus sign), with the same probability $1-\exp(-\beta)$, and denote by $\tilde\sigma$ the corresponding killing time. Then,
\begin{equation}
\label{eq:alt_lambda.beta}
\lambda(\beta) = - {\ensuremath{\mathbb E}} \tilde{\ensuremath{\mathbb E}} \log \bP^\beta(H_{\tau_1} < \tilde\sigma).
\end{equation}
Assuming this last equality, we could readily prove using the dominated convergence theorem that $\lambda$ is also continuous at $0$.\\
{\noindent \bf 7.} Equation~\eqref{eq:encadr.lambda} does not give much information about the behaviour of $\lambda(\beta)$ at $0$, that remains an open question. We expect however that $\beta = o(\lambda(\beta))$ as $\beta\to 0$ and we now explain why. To this end, recall~\eqref{eq:alt_lambda.beta} and the related notations above. By integrating over ${\ensuremath{\mathcal N}} $ and differentiating in $\beta$ we obtain
\begin{equation}
\lim_{\beta \to 0} \lambda'(\beta) = \tilde{\ensuremath{\mathbb E}} \bbE\bE\Big(\sum_{k=1}^{H_{\tau_1}} {\sf 1}_{\{S_k \in-\tilde\tau\}} \Big),
\end{equation}
that we expect to be infinite. Indeed, by first restricting the walk to make its first step to the left and then using the symmetry of the random walk,
\begin{equation}
\tilde{\ensuremath{\mathbb E}} \bbE\bE\Big(\sum_{k=1}^{H_{\tau_1}} {\sf 1}_{\{S_k \in-\tilde\tau\}} \Big)
\ge \frac12 \tilde{\ensuremath{\mathbb E}} \bE_{-1}\Big(\sum_{k=1}^{H_0} {\sf 1}_{\{S_k \in-\tilde\tau\}} \Big)
= \frac12 \tilde{\ensuremath{\mathbb E}} \bE_1\Big(\sum_{k=1}^{H_0} {\sf 1}_{\{S_k \in \tilde\tau\}} \Big).
\end{equation}
We now interchange integrals and use the renewal theorem to obtain, at least for $\gamma \neq 1$,
\begin{equation}
\lim_{\beta \to 0} \lambda'(\beta) \ge \frac12 \bE_1\Big(\sum_{k=1}^{H_0} \tilde{\ensuremath{\mathbb P}} (S_k \in \tilde\tau) \Big)
\ge \frac{C}{2} \bE_1\Big(\sum_{k=1}^{H_0} (1+S_k)^{(\gamma-1)\wedge 0} \Big).
\end{equation}
Since, by Ray-Knight's theorem, the mean number of visits to $x\in{\ensuremath{\mathbb N}} _0$ between time $1$ and $H_0$ equals $1$ under $\bP_1$, we get
\begin{equation}
\lim_{\beta \to 0} \lambda'(\beta) \ge C \sum_{x\ge 0} (1+x)^{(\gamma-1)\wedge 0} = \infty.
\end{equation}
{\noindent \bf 8.} Note that we offer no path statement. In other words, we do not prove anything about the behaviour of the walk conditioned to survive for a long time $n$. However, as it is often the case with this type of model, our result and the method of proof suggest a path strategy, which in our context corresponds to a confinement (or localization) strategy. To be more precise, we roughly expect that as $n$ is large, the walk reaches the trap labelled $X^*(\beta)N$ and then remains in the corresponding gap, of size $Y^*(\beta)N^{1/\gamma}$, where $(X^*(\beta),Y^*(\beta))$ is distributed as the unique minimizer of the random variational problem in Theorem~\ref{thm0}. In other words, the path (or polymer) gets stuck in a slab which is {\it large enough} and {\it not too far from the origin}. Surprisingly, the repulsive interaction between the path and the traps leads to a pinning effect on the quenched path measures, as explained by Sznitman~\cite[Chapter 6]{Sz98}. Proving such a result should demand substantial additional work, as one would most likely need sharper bounds on the survival probability (partition function) and fine controls on the ratio of survival probabilities restricted to suitable events. Nevertheless, this can be considered as an interesting direction of research. Let us mention that Ding and Xu~\cite{DingXu} recently obtained a path confinement result in the case of quenched hard Bernoulli obstacles for $d\ge 2$.\\
{\noindent \bf 9.} Let us stress that the scaling $t^{\gamma/(\gamma+2)}$ that appears in our Theorem \ref{thm0} is different from the scaling of the PAM in a bounded i.i.d. potential. In this case \cite[Example $5.10$]{Ko16} states that the correct scaling is $t$ up to a logarithmic correction. Hence we are in a case where the correlations of the potential have a drastic effect on the asymptotic behaviour of the survival probability.
\section{Key tools}
\label{sec:KT}
In this section we introduce several tools which we will use to prove our result. For convenience, various notations are gathered together in Section~\ref{subsec:not}, to which the read may refer. In Section~\ref{subsec:conf} we remind the reader of the so-called small ball estimate and establish a rough upper bound on the probability that the walker stays confined in a fixed gap until time $n$, see Proposition~\ref{prop:roughUB}. Section~\ref{subsec:di} contains Proposition \ref{prop:FKG}, which states that a walk conditioned to hit a given point $x$ before coming back to $0$ does it faster when it is also conditioned on survival until its first visit to $x$. In Section~\ref{subsec:ttl} we state the two technical Lemmas~\ref{lem:control_ratio} and~\ref{lem:control_pinbad} that we will use in Section~\ref{sec:UB} while proving the upper bound on $Z_n$. Finally we introduce the key notion of \textit{good environment} in Section~\ref{subsec:ge}. Informally, \textit{good environments} are those for which we are able to efficiently bound $Z_n$. We thus give a list of events, see \eqref{eq:ge}, that are essentially the technical conditions we will need in Proposition~\ref{prop:ub} and~\ref{prop:lb}.
\subsection{Notations} \label{subsec:not} Let us introduce notations that will be necessary in what comes next.\\
{\it \noindent Killing clock.} We recall the definition of the $\mathbb{N}$-valued geometric random variable ${\ensuremath{\mathcal N}} $ with success parameter $1-e^{-\beta}$ that plays the role of the killing clock.\\
{\it \noindent Records.} As we already hinted in the heuristics, only the {\it largest} gaps matter. To be more precise, a particular attention is given to record values of the sequence $(T_\ell)$. Therefore, we let
\begin{equation}
\label{def:ik}
i(0)=0,\qquad i(k) = \inf\{i>i(k-1)\colon T_{i+1} > T_{i(k-1)+1}\},\qquad k\ge 1,
\end{equation}
be the sequence of record indexes, while
\begin{equation}
\label{eq:def.t.tau.star}
\tau^*_k=\tau_{i(k)} \qquad \textrm{ and } \qquad T^*_k = T_{i(k)+1},\qquad k\ge 0.
\end{equation}
We also define
\begin{equation}
R(a,b) = \{k \ge 1 \colon a\le i(k) \le b\},\qquad {\ensuremath{\mathcal R}} (a,b) = i(R(a,b)),\qquad a,b\in{\ensuremath{\mathbb N}} ,\ a<b,
\end{equation}
and
\begin{equation}
\label{eq:def.R.cR}
R_\gep(n) = R(\gep N, \gep^{-1}N),\qquad {\ensuremath{\mathcal R}} _\gep(n) = {\ensuremath{\mathcal R}} (\gep N, \gep^{-1}N), \qquad n\in {\ensuremath{\mathbb N}} , \qquad \gep>0.
\end{equation}
Finally we write
\begin{equation}
{\ensuremath{\mathcal R}} ={\ensuremath{\mathcal R}} (1,+\infty),
\end{equation}
for the set of all records.\\
{\it \noindent Auxiliary random walk.} We remind that the clock process $(\theta_n)_{n\ge 1}$ is defined by
\begin{equation}
\theta_0 = 0,\quad \theta_{n+1} = \inf\{k > \theta_n \colon S_k \in \tau\},\quad n\ge 0.
\end{equation}
The process that chronologically keeps track of the traps visited by the walk will be denoted by $X = (X_n)_{n\ge 0}$ and is uniquely determined by $\tau_{X_n} = S_{\theta_n}$.
It is not difficult to see that $X$ is a Markov chain on ${\ensuremath{\mathbb N}} _0$, the hitting times of which are denoted by
\begin{equation}
\label{def:zeta}
\zeta_x = \inf\{ n\ge 1 \colon X_n = x\},\qquad x\in{\ensuremath{\mathbb N}} _0,
\end{equation}
and
\begin{equation}
\label{def:zeta.star}
\zeta^*_k = \inf\{ n\ge 1 \colon X_n = i(k)\},\qquad x\in{\ensuremath{\mathbb N}} _0.
\end{equation}
\medskip
{\it \noindent Transition matrices and their moment-generating functions.} Let us define
\begin{equation}
\label{eq:defqij}
q_{ij}(n) = \bP_{\tau_i}(S_k \notin \tau,\ 1\le k < n,\ S_n = \tau_j),\qquad i,j\in {\ensuremath{\mathbb N}} _0, \quad n\ge 1,
\end{equation}
and the associated family of matrices $\{Q(\phi)\}_{\phi\ge 0}$ by
\begin{equation}
\label{eq:defQij}
Q_{ij}(\phi) = \sum_{n\ge 1} e^{\phi n} q_{ij}(n) = \bE_{\tau_i}\left(e^{\phi \theta_{1}}{\sf 1}_{\{S_{\theta_1}=\tau_j\}}\right),\qquad i,j\in {\ensuremath{\mathbb N}} _0, \ \phi \geq 0.
\end{equation}
Note that the matrix $\{Q_{ij}(0)\}_{i,j\ge 0}$ is nothing but the transition matrix of the Markov chain $X$ defined above. These quantities will appear in Lemma~\ref{lem:control_ratio} below and are zero as soon as $|j-i|>1$. Finally, we will also use the following notations for the gap associated to an non-oriented edge $\{i,j\}$ with $|j-i|\le 1$:
\begin{equation}
\label{eq:deftij}
t_{ij} =
\left\{
\begin{array}{lll}
t_{i+1} & \text{if} & j=i+1, \\
t_i & \text{if} & j = i- 1, \\
t_{i+1} \vee t_i & \text{if} & i=j,
\end{array}
\right.
\end{equation}
where $(t_i)$ is a sequence of integers.
\medskip
\subsection{Confinement estimates} \label{subsec:conf}
One of the key standard estimates in our arguments are the so-called {\it small-ball estimates}, that control the probability that a simple random walk stays confined in an interval:
\begin{proposition} \label{pr:small_ball}
There exist $t_0,c_1,c_2,c_3,c_4>0$ such that for all $t>t_0$, the following inequalities hold for all $n\ge 1$ such that $n\in 2{\ensuremath{\mathbb N}} $ or $n-t \in 2{\ensuremath{\mathbb N}} $:
\begin{equation}
\label{eq:sb}
\frac{c_1}{t \wedge n^{1/2}} e^{-g(t)n} \le \bP(H_t \wedge H_0 \wedge H_{-t} > n) \le \frac{c_2}{t \wedge n^{1/2}} e^{-g(t)n}
\end{equation}
\begin{equation}
\frac{c_3}{t^3 \wedge n^{3/2}} e^{-g(t)n} \le \bP(H_t \wedge H_0 \wedge H_{-t} = n) \le \frac{c_4}{t^3 \wedge n^{3/2}} e^{-g(t)n},
\end{equation}
where
\begin{equation}
\label{eq:g}
g(t) = -\log \cos \big(\frac{\pi}{t}\big) = \frac{\pi^2}{2t^2} + O\big(\frac{1}{t^4}\big),\qquad t\to +\infty.
\end{equation}
\end{proposition}
This proposition is taken from Lemma 2.1 in Caravenna and Pétrélis~\cite{CP09b}. A crucial point here is the uniformity of the constants, which gives the uniformity of the constant $C$ in Proposition~\ref{prop:roughUB}.
\par Caravenna and Pétrélis~\cite{CP09b} treated the case of equally spaced traps, which we refer to as the {\it homogeneous} case, in the sense that increments of $\tau$ are all equal. We summarize their results here.
\begin{proposition}[Homogeneous case, see Eq. (2.1)-(2.3) in \cite{CP09b}]
\label{pr:homo}
Let $t\in {\ensuremath{\mathbb N}} $ and $\tau = t{\ensuremath{\mathbb Z}} $. There exists a constant $\phi(\beta, t)$ such that
\begin{equation}
\label{eq:315}
\phi(\beta, t) = - \lim_{n\to\infty} \frac{1}{n} \log \bP(\sigma > n),
\end{equation}
with
\begin{equation}
\label{eq:phi}
\phi(\beta, t) = \frac{\pi^2}{2t^2} \Big(1 - \frac{4}{e^{\beta} - 1}\frac{1}{t} + o\Big(\frac{1}{t} \Big) \Big).
\end{equation}
Moreover, it is the only solution of the equation:
\begin{equation}
\bE(\exp(\phi \inf\{n\ge 1 \colon S_n \in \tau\}))) = \exp(\beta),\qquad \beta\ge 0.
\end{equation}
\end{proposition}
Note that the first order term in the expansion of $\phi$ does not depend on $\beta$.
It turns out that we are able to extend this proposition, at the price of additional technical work, to deal with the {\it weakly-inhomogeneous} case, that is when increments of $\tau$ follow a periodic pattern. We obtain the following:
\begin{proposition}[Weakly-inhomogeneous case]
\label{pr:periodic}
Let $p\ge 2$, $t_1, \ldots, t_p$ positive integers and $\tau$ be the periodic set $\{\tau_i \colon 0\le i < p\} + \tau_p {\ensuremath{\mathbb Z}} $, where $\tau_0 = 0$ and $\tau_i = \sum_{1\le j \le i} t_j$ for all $0<i<p$.
There exists a constant $\phi = \phi(\beta ; t_1, \ldots, t_p)$ such that
\begin{equation}
\phi(\beta ; t_1, \ldots, t_p) = - \lim_{n\to\infty} \frac{1}{n} \log \bP(\sigma > n).
\end{equation}
Moreover,
\begin{equation}
\bP(\sigma > n) \le Cn^2p \exp(-\phi(\beta ; t_1, \ldots, t_p) n),\quad n\geq 1,
\end{equation}
and
\begin{equation} \label{eq:comp.phi}
\phi(\beta, t_{\max}) \le \phi(\beta ; t_1, \ldots, t_p) < g(t_{\max}), \qquad t_{\max} = \max_{1\le i \le p} t_i.
\end{equation}
\end{proposition}
The proof is deferred to Appendix~\ref{sec:periodic}. Remark that both inequalities in \eqref{eq:comp.phi} are intuitive: the first one asserts that it is easier to survive in a homogeneous environment with gap $t_{max}$ than in the original environment. The second one states that one of the strategy to survive is to stay confined in the largest gap. With this estimate in hand, we get our first building block, that is an upper bound on the probability to survive in-between two traps, for a general environment $\tau$.
\begin{proposition}\label{prop:roughUB}
There exists a constant $C>0$ such that for all $0\le k < r < \ell$, one has
\begin{equation}
\bP_{\tau_r}(\sigma \wedge H_{\tau_k} \wedge H_{\tau_\ell} > n) \le Cn^2(\ell-k) \exp(-\phi(\beta;\max\{t_i \colon k<i\le \ell\})n),
\end{equation}
where $\phi(\beta;\cdot)$ is defined in Proposition~\ref{pr:homo}.
\end{proposition}
\begin{proof}[Proof of Proposition~\ref{prop:roughUB}]
The proof relies on periodization. Since the random walk does not leave the interval $(\tau_k,\tau_\ell)$ on the event considered, we may as well replace the renewal $\tau$ by a periodized version, and by translation invariance, consider that the random walk starts at zero. Finally, everything is as if the walk evolves in a new environment $\tilde\tau$, with periodic increments, defined by
\begin{equation}
\tilde{\tau}=\{\tau_i \colon k\le i \le \ell \} -\tau_r+ (\tau_\ell-\tau_k) {\ensuremath{\mathbb Z}} ,
\end{equation}
and we now have to bound from above
$
\bP^{\tilde\tau}(\sigma \wedge H_{\tilde\tau_{k-r}} \wedge H_{\tilde\tau_{\ell-r}} > n),
$
where we put a superscript on $\bP$ to stress that the walk evolves among $\tilde\tau$. This probability is certainly smaller than $\bP_0^{\tilde\tau}(\sigma> n)$, and we may now conclude thanks to Proposition~\ref{pr:periodic}.
\end{proof}
\subsection{A decoupling inequality} \label{subsec:di} %
The next building block is a control on the probability that the walk reaches a given point before a certain fixed time, conditional on survival and not coming back to $0$.
In essence, the following proposition tells us that the walk reaches this point {\it stochastically faster} in the presence of traps:
\begin{proposition}\label{prop:FKG}
For all $\beta>0$, $x\in {\ensuremath{\mathbb N}} $ and $n\in \mathbb{N}$,
\begin{equation}
\label{eq:propFKG}
\bP^{\beta}(H_x \le n\ |\ \sigma\wedge H_0 > H_x) \ge \bP(H_x \le n\ |\ H_0 > H_x).
\end{equation}
\end{proposition}
Let us stress that this proposition is general, as it does not depend on the position of the traps.
\begin{proof}[Proof of Proposition~\ref{prop:FKG}]
Let $x\in{\ensuremath{\mathbb N}} $. We first remark that the stopped process $(S_{k\wedge H_x})_{k\ge 0}$ is still a Markov chain under $\bP^{\beta}(\cdot | \sigma\wedge H_0 > H_x)$, with $0$ as initial state, $x$ as absorbing state, and transition probabilities given by
\begin{equation}
\begin{aligned}
\bar Q_\beta(a,b) :=
\begin{cases}
\frac{e^{-\beta{\sf 1}_{\{b\in\tau\}}}\bP_b(\sigma\wedge\tilde H_0 > \tilde H_x)}{e^{-\beta{\sf 1}_{\{a+1\in\tau\}}}\bP_{a+1}(\sigma\wedge\tilde H_0 > \tilde H_x) + e^{-\beta{\sf 1}_{\{a-1\in\tau\}}}\bP_{a-1}(\sigma\wedge\tilde H_0 > \tilde H_x)} & \text{if } |b-a| = 1\\
0 & \text{otherwise,}
\end{cases}
\end{aligned}
\end{equation}
where $1\le a < x$ and $\tilde H_z := \inf\{n\ge 0\colon S_n = z\}$. By applying the strong Markov property at $H_{a+1}$, observe that
\begin{equation}
\frac{\bar Q_\beta(a,a-1)}{\bar Q_\beta(a,a+1)} = \frac{e^{-\beta{\sf 1}_{\{a-1\in\tau\}}}\bP_{a-1}(\sigma\wedge\tilde H_0 > \tilde H_x)}{e^{-\beta{\sf 1}_{\{a+1\in\tau\}}}\bP_{a+1}(\sigma\wedge\tilde H_0 > \tilde H_x)}
= e^{-\beta{\sf 1}_{\{a-1\in\tau\}}}\bP_{a-1}(\sigma\wedge\tilde H_0 > H_{a+1}),
\end{equation}
and note that this ratio is non-increasing in $\beta$, for all $1\le a < x$. We may deduce by a standard coupling argument that $H_x$ is stochastically smaller under $\bP^{\beta}(\cdot | \sigma\wedge H_0 > H_x)$ than under $\bP(\cdot | H_0 > H_x)$, which corresponds to the case $\beta=0$. This concludes the proof.
\end{proof}
\subsection{Two technical lemmas} \label{subsec:ttl}
\par Recall the notations in~\eqref{eq:defQij} and~\eqref{eq:deftij}.
\begin{lemma}
\label{lem:control_ratio}
Define the function $f \colon z\in (0,\pi) \mapsto z/\sin(z)$. Let $n\in{\ensuremath{\mathbb N}} $. For $\gep>0$ small enough, there exist $\alpha=\alpha(\gep)>0$, $C>0$ and $T_0(\gep)\in{\ensuremath{\mathbb N}} $ such that for $T> T_0(\gep) \vee \max_{0\le x \le n} T_x$ and $\phi = \frac{\pi^2}{2T^2}$,
\begin{equation}
\label{eq:ratio}
\frac{Q_{x,y}(\phi)}{Q_{x,y}(0)}
\leq
\begin{cases}
\exp(\gep) & \text{if } x\neq y \text{ and } T_{xy} \le \alpha T, \text{ or } x=y\\
2 f\left(\pi \frac{\max_{0\leq x\leq n}T_{x}}{T}(1+\frac{C}{T^2})\right) & \text{else.}
\end{cases}
\end{equation}
\end{lemma}
The ratio in \eqref{eq:ratio} is the positive Laplace transform of the hitting time of $\tau$ for the walk conditioned to go from $\tau_x$ to $\tau_y$. As we will consider large values of $T$ and thus small values of $\phi$, Lemma \ref{lem:control_ratio} can be understood as a control of this transform near $0$.
\begin{proof}[Proof of Lemma~\ref{lem:control_ratio}]
\par We consider $\gep>0$ small enough (it will appear in the following how small it has to be). Let us start with the case $x\neq y$. From the explicit expressions of the Laplace transforms, see e.g. (A.5) in Caravenna and Pétrélis~\cite{CP09}, we get
\begin{equation}
\label{eq:defQxy}
Q_{x,y}(\phi) = \frac{\tan \Delta}{2\sin(T_{xy}\Delta)}, \quad \text{where } \Delta=\Delta(\phi) = \arctan(\sqrt{e^{2\phi} - 1}),
\end{equation}
and we note that
\begin{equation}
\label{eq:defQxy0}
Q_{x,y}(0) = \frac{1}{2T_{xy}}.
\end{equation}
\par Let us notice that \eqref{eq:defQxy} is well-defined if $T_{xy} \Delta < \pi$, which occurs as soon as $T$ is large enough. Indeed, by expanding $\Delta$ we see that there exists a constant $C>0$ and $T_1\in{\ensuremath{\mathbb N}} $ such that for $T\ge T_1$,
\begin{equation}
\label{eq:UBgD}
\Delta \le \frac{\pi}{T}\Big(1+\frac{C}{T^2}\Big).
\end{equation}
If we assume moreover that $T>T_{xy}$, we obtain, as $T$ and $T_{xy}$ are integers,
\begin{equation}
\label{eq:UBTxygD}
\frac{T_{xy}\Delta}{\pi} \le \frac{T-1}{T}\Big(1+\frac{C}{T^2}\Big) = 1- \frac{1+o(1)}{T}<\frac{1}{1+\gep},
\end{equation}
provided $T$ is larger than some $T_1(\gep) \in {\ensuremath{\mathbb N}} $. For the rest of the proof, we assume that $T> T_1 \vee T_1(\gep) \vee \max\{T_x\colon 0\le x\le n\}$.
\par By combining \eqref{eq:defQxy} and \eqref{eq:defQxy0}, we now obtain
\begin{equation}
\frac{Q_{x,y}(\phi)}{Q_{x,y}(0)} = \frac{T_{xy}\tan (\Delta)}{\sin(T_{xy}\Delta)}.
\end{equation}
By using \eqref{eq:UBgD} and expanding $\tan$ to first order, there exists $T_2(\gep)$ such that for $T\ge T_2(\gep)$,
\begin{equation}
\tan(\Delta) \le (1+\gep) \Delta.
\end{equation}
By adding this latter condition on $T$, we get, since $f$ is increasing,
\begin{equation}
\begin{aligned}
\label{eq:UBratioQ's}
\frac{Q_{x,y}(\phi)}{Q_{x,y}(0)} &\le (1+\gep) \frac{T_{xy}\Delta}{\sin(T_{xy}\Delta)} \\
&= (1+\gep)f(T_{xy}\Delta)\\
&\le (1+\gep)f\left(\pi \frac{T_{xy}}{T}\Big(1+\frac{C}{T^2}\Big)\right).
\end{aligned}
\end{equation}
As $\gep<1$, that concludes the proof of the second inequality in our statement.
To get the first inequality when $x\neq y$, notice first that, as $f(z) \to 1$ when $z \to 0$, there exists $z_\gep$ such that $(1+\gep)f(z)\le \exp(\gep)$ for $z\le z_\gep$. Therefore, it is enough to define
\begin{equation}
\alpha(\gep) = \frac{z_\gep}{\pi(1+\gep)},
\end{equation}
assume that $T> T_3(\gep):=(C/\gep)^{1/2}$ and use \eqref{eq:UBratioQ's} to get what we need.\\
We are left with the case $x = y$. Again, with the help of (A.5) in Caravenna and Pétrélis~\cite{CP09},
\begin{equation}
Q_{xx}(\phi) = 1 - \frac12 \frac{\tan(\Delta)}{\tan(T_{x-1,x}\Delta)} - \frac12 \frac{\tan(\Delta)}{\tan(T_{x,x+1}\Delta)},
\end{equation}
where $\Delta$ is defined as in \eqref{eq:defQxy}. We thereby retrieve the standard formula:
\begin{equation}
Q_{xx}(0) = 1 - \frac{1}{2T_{x-1,x}} - \frac{1}{2T_{x,x+1}}.
\end{equation}
Note that it is enough to treat the case $T_{x,x+1} = T_{x-1,x}$ since
\begin{equation}
Q_{xx}(\phi) = \frac12\Big(1 - \frac{\tan(\Delta)}{\tan(T_{x-1,x}\Delta)}\Big) + \frac12\Big(1 - \frac{\tan(\Delta)}{\tan(T_{x,x+1}\Delta)}\Big).
\end{equation}
We may now consider the ratio
\begin{equation}
\label{eq:Qratio}
\frac{Q_{xx}(\phi)}{Q_{xx}(0)}
=
\frac{1 - \frac{\tan(\Delta)}{\tan(T_{x,x+1}\Delta)}}{1- \frac{1}{T_{x,x+1}}}.
\end{equation}
By choosing $T\ge T_2(\gep)$ and expanding $\tan$ to first order, we obtain
\begin{equation}
1 - \frac{\tan(\Delta)}{\tan(T_{x,x+1}\Delta)} \le
\begin{cases}
1 - \frac{\Delta}{\tan(T_{x,x+1}\Delta)} & \text{if } \quad T_{x,x+1}\Delta \le \frac{\pi}{2},\\
1 - (1+\gep) \frac{\Delta}{\tan(T_{x,x+1}\Delta)} & \text{if } \quad \frac{\pi}{2} <T_{x,x+1}\Delta < \pi.
\end{cases}
\end{equation}
We remind that our conditions on $T$ guarantee that $T_{x,x+1}\Delta < \pi$. The reason why we split cases above is that $\tan$ changes sign at the value $\pi$. We further make a dichotomy : (i) $T$ is large and $T_{x+1}$ is at least $\gep T$ and (ii) $T$ is large and $T_{x+1}$ less than $\gep T$. Let us start with (i). If actually $T_{x,x+1}\Delta < \pi/2$, we may simply bound the denominator in \eqref{eq:Qratio} by $1$. Otherwise, we note that $z\mapsto - z/\tan(z)$ is increasing on $(\pi/2,\pi)$, so we may write, as $T_{x,x+1}\Delta < \pi/(1+\gep)$ by \eqref{eq:UBTxygD},
\begin{equation}
1 - (1+\gep)\frac{\Delta}{\tan(T_{x,x+1}\Delta)} \le 1 - \frac{1}{T_{x,x+1}} \frac{\pi}{\tan(\pi/(1+\gep))}.
\end{equation}
Thus, if we define
\begin{equation}
T_4(\gep) = \frac{3}{\gep^2}\Big(\frac{\pi}{|\tan(\pi/(1+\gep))|} \vee 1 \Big)
\end{equation}
and assume that $T>T_4(\gep)$, we obtain
\begin{equation}
\frac{Q_{xx}(\phi)}{Q_{xx}(0)} \le \frac{1-\frac{1}{T_{x,x+1}}\frac{\pi}{\tan(\pi/(1+\gep))}}{1-\frac{1}{T_{x,x+1}}} \le \frac{1+\gep/3}{1-\gep/3} = 1+\frac 23\gep+o(\gep), \quad \gep \to 0,
\end{equation}
which is smaller than $\exp(\gep)$ when $\gep$ is small enough.
We now deal with (ii) and to this end we assume $T\ge T_2(\gep)$ and $T_{x,x+1} \le \gep T$, in which case we expand $\tan(T_{x,x+1}\Delta)$. {By using \eqref{eq:UBgD} and assuming that $T>T_3(\gep)$ we get $T_{x,x+1} \Delta \le \gep(1+\gep)\pi$}. Thus, there exists a constant $C=C(\gep)>0$ such that
\begin{equation}
\tan(T_{x,x+1}\Delta) \le T_{x,x+1}\Delta + C(T_{x,x+1}\Delta)^3, \quad \text{so } 1 - \frac{\Delta}{\tan(T_{x,x+1}\Delta)} \le 1 - \frac{1}{T_{x,x+1}}(1-CT_{x,x+1}^2\Delta^2).
\end{equation}
Finally, since $T_{x,x+1}\ge 2$ necessarily,
\begin{equation}
\frac{Q_{xx}(\phi)}{Q_{xx}(0)} \leq 1 + 2CT_{x,x+1}\Delta^2 \le 1 +2C\gep(1+\gep)\Delta.
\end{equation}
Now we pick $T\ge T_5(\gep)$ such that $\Delta \le [2C(1+\gep)]^{-1}$ and we get the claim.
\par We conclude the proof by setting $T_0(\gep) = \max(T_1, T_1(\gep),T_2(\gep),T_3(\gep), T_4(\gep),T_5(\gep))$.
\end{proof}
Recall the notations in~\eqref{eq:def.t.tau.star} and~\eqref{eq:deftij}. Given $\alpha>0$ and $k\in{\ensuremath{\mathbb N}} $, we define a set of {\it bad} edges as
\begin{equation}
\label{eq:defB}
{\ensuremath{\mathcal B}} _{k,\alpha} = \{1\le x, y\le i(k) \colon x\neq y,\ T_{x,y} > \alpha T^*_{k}\}.
\end{equation}
and its cardinal
\begin{equation}
L_{k,\alpha} = |{\ensuremath{\mathcal B}} _{k,\alpha}|.
\end{equation}
These bad edges correspond to the second case in Lemma \ref{lem:control_ratio}.
Recall also~\eqref{def:zeta} and~\eqref{def:zeta.star}. The following lemma controls the visits to the bad edges:
\begin{lemma}
\label{lem:control_pinbad}
There exists a function $h$ such that, for any $A>0$, $k\geq 0$, and $\alpha>0$, if $T^*_k > h(A,L_{k,\alpha},\alpha)$ then
\begin{equation}
\bE^{\beta}\Big( A^{\sharp\{i\le \zeta^*_k \colon \{X_{i-1},X_i\} \in {\ensuremath{\mathcal B}} _{k,\alpha}\}} {\sf 1}_{\{\zeta^*_k< \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big) \leq 2A^{L_{k,\alpha}} \tau^*_k \bP^{\beta}(\zeta^*_k< \zeta_0\wedge {\ensuremath{\mathcal N}} ).
\end{equation}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{lem:control_pinbad}]
\par We start with the case $L_{k,\alpha}=1$ and denote by $(s,s+1)$ the bad edge. By using the geometric nature of ${\ensuremath{\mathcal N}} $ and applying the Markov property at $\zeta_{s+1}$, we get
\begin{equation}
\begin{aligned}
\label{eq:cpinbad1}
\bE^{\beta}\Big( & A^{\sharp\{i\le \zeta_k^* \colon \{X_{i-1},X_i\} \in {\ensuremath{\mathcal B}} _{k,\alpha}\}}{\sf 1}_{\{\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)=\bE^{\beta}\Big( A^{\sharp\{i\le \zeta_{s+1} \colon \{X_{i-1},X_i\} = \{s,s+1\}\}}{\sf 1}_{\{\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)\\
&\le \bP^\beta(\zeta_{s+1} < \zeta_0 \wedge {\ensuremath{\mathcal N}} )A\
\bE_{\tau_{s+1}}^{\beta}\Big( A^{\sharp\{i\le \zeta_{s+1} \colon \{X_{i-1},X_i\} = \{s,s+1\}\}} {\sf 1}_{\{\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big),
\end{aligned}
\end{equation}
and we now focus on the last factor in the line above. By considering the consecutive visits of $X$ to $s+1$, we may write
\begin{equation}
\bE_{\tau_{s+1}}^{\beta}\Big( A^{\sharp\{i\le \zeta_{s+1} \colon \{X_{i-1},X_i\} = \{s,s+1\}\}} {\sf 1}_{\{\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)
= \bE(v^G) \bP_{\tau_{s+1}}(\zeta_k^* < {\ensuremath{\mathcal N}} | \zeta_k^* < \zeta_{s+1}),
\end{equation}
where $G$ is a $\mathbb{N}_0$-valued geometric random variable with parameter $\bP_{\tau_{s+1}}(\zeta_k^* < \zeta_{s+1})$ and
\begin{equation}
v = \bE_{\tau_{s+1}}\Big( A^{\sharp\{i\le \zeta_{s+1} \colon \{X_{i-1},X_i\} = \{s,s+1\}\}} {\sf 1}_{\{ \zeta_{s+1} < {\ensuremath{\mathcal N}} \}} | \zeta_{s+1} < \zeta_k^* \Big).
\end{equation}
We are going to show now that $v \le 1$ when $T^*_k \geq h_0(A,\alpha)$ where
\begin{equation}
\label{eq:defH}
h_0(A,\alpha)=\frac{A^2}{2\alpha e^{\beta}(e^{\beta}-1)}.
\end{equation}
To this end, note that
\begin{equation}
v \le \frac12 e^{-\beta} + \frac12 \Big(1 - \frac{1}{T_{s,s+1}} \Big)e^{-\beta} + \frac{1}{2T_{s,s+1}}A^2 e^{-2\beta}.
\end{equation}
Indeed, the first term is given by walks which make their first step to the right. The second term comes from those who make their first step to the left but come back to $\tau_{s+1}$ before hitting $\tau_s$, whereas the third term comes from the walks who hit $\tau_s$ before coming back to $\tau_{s+1}$.
Then, as $T_{s,s+1}\geq \alpha T^*_k$,
\begin{equation}
v \leq e^{-\beta}+\frac{A^2e^{-2\beta}}{2\alpha T^*_k},
\end{equation}
which, by \eqref{eq:defH}, proves that $v\leq 1$.
To complete the proof in this case, we write
\begin{equation}
\begin{aligned}
\text{r.h.s\eqref{eq:cpinbad1}} &\le A\ \bP^\beta(\zeta_{s+1} < \zeta_0 \wedge {\ensuremath{\mathcal N}} )\bP_{\tau_{s+1}}(\zeta_k^* < {\ensuremath{\mathcal N}} | \zeta_k^* < \zeta_{s+1})\\
&\le A\ \bP^\beta(\zeta_{s+1} < \zeta_0 \wedge {\ensuremath{\mathcal N}} )
\frac{\bP^\beta_{\tau_{s+1}}(\zeta_k^* < {\ensuremath{\mathcal N}} \wedge \zeta_{s+1})}{\bP_{\tau_{s+1}}(\zeta_k^* < \zeta_{s+1})}\\
&\le 2A(\tau_k^* - \tau_{s+1}) \bP^\beta(\zeta_{s+1} < \zeta_0 \wedge {\ensuremath{\mathcal N}} )\bP^\beta_{\tau_{s+1}}(\zeta_k^* < {\ensuremath{\mathcal N}} \wedge \zeta_{s+1})\\
&\le 2A \ \tau_k^*\ {\bP^\beta(\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} )}.
\end{aligned}
\end{equation}
\par Let us now conclude the proof in the general case $L_{k,\alpha}\ge 1$. Our strategy is to decouple the contribution of each bad set by Holdër's inequality and reduce the problem to the case $L_{k,\alpha}=1$ with $A$ replaced by $A^L$. Indeed, if we note
\begin{equation}
{\ensuremath{\mathcal B}} _{k,\alpha}= \{(s_\ell, s_{\ell+1}) \colon 1\le \ell \le L,\ 1\le s_\ell < i(k) \},
\end{equation}
and suppose that $T_k^*\geq h(A,L_{k,\alpha},\alpha)$, where
\begin{equation}
\label{eq:def.gALga}
h(A,L,\alpha):=h_0(A^L,\alpha)= \frac{A^{2L}}{2\alpha e^{\beta}(e^{\beta}-1)},
\end{equation}
we get
\begin{equation}
\begin{aligned}
&\bE^{\beta}\Big( A^{\sharp\{i\le \zeta_k^* \colon \{X_{i-1},X_i\} \in {\ensuremath{\mathcal B}} _{k,\alpha}\}}{\sf 1}_{\{\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)\\
&\le
\prod_{i=1}^{L_{k,\alpha}}
\bE^{\beta}\Big( (A^{L_{k,\alpha}})^{\sharp\{i\le \zeta_k^* \colon \{X_{i-1},X_i\} = \{s_\ell, s_{\ell+1}\}\}}{\sf 1}_{\{\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)^{1/L_{k,\alpha}}\\
&\le 2A^{L_{k,\alpha}} \tau_k^* {\bP^\beta(\zeta_k^* < \zeta_0 \wedge {\ensuremath{\mathcal N}} )}.
\end{aligned}
\end{equation}
This concludes the proof.
\end{proof}
\subsection{Good environments}
\label{subsec:ge}
{We define here a notion of \textit{good environments}, that are environments where it is possible to give a good control on the survival probability. We will show in Section~\ref{sec:proba.good.env} that these environments are typical, meaning that by tuning some parameters and considering $n$ large enough, their probability of occurence can be made arbitrarily close to one.}
\subsubsection{Additional notations} Beforehand, we remind of the functions $f$ and $h$ introduced in Lemma \ref{lem:control_ratio} and Lemma \ref{lem:control_pinbad}. We define
\begin{equation}
\label{eq:def_fk}
f_{k}:=2 f\left(\pi \frac{T^*_{k-1}}{T^*_k}\Big[1+\frac{C}{(T^*_k)^2}\Big]\right),
\end{equation}
that appears in the events $A_n^{(6)}$ and $A_n^{(7)}$ below. The constant $C$ above is the same as the one in \eqref{eq:ratio} in Lemma \ref{lem:control_ratio}.
From \eqref{eq:g} and \eqref{eq:phi}, there exists (for a fixed $\beta$) a constant $C_1>0$ so that
\begin{equation}
\label{eq:encadr_g}
1/(C_1 t^2) \leq g(t) \wedge \phi(\beta,t) \leq g(t) \vee \phi(\beta,t) \le C_1 / t^2, \qquad t\ge 1.
\end{equation}
This constant appears in the event $A_n^{(9)}$. Finally, we define the exponent
\begin{equation}
\label{def:kappa}
\kappa =
\begin{cases}
\frac{\gamma}{4} & \text{if } \gamma \le 1\\
\frac{1}{2\gamma} - \frac14 & \text{if } 1 <\gamma < 2\\
\frac{1}{2\gamma} & \text{if } \gamma \ge 2,
\end{cases}
\end{equation}
which appears in the event $A_n^{(1)}$.
\subsubsection{Definition} Let $\delta,\gep_0,\gep,\eta>0$. The set of good environments, denoted by $\Omega_n(\delta,\gep_0,\gep,\eta)$, is defined as the intersection of the events defined below (we stress that $\alpha(\gep)$ and $T_0(\gep)$ appearing here are the same as in Lemma \ref{lem:control_ratio}):
\begin{equation}
\Omega_n(\delta,\gep_0,\gep,\eta)=\bigcap_{i=1}^{11} A_n^{(i)}(\delta,\gep_0,\gep,\eta),
\end{equation}
with
\begin{equation}
\label{eq:ge}
\begin{aligned}
A^{(1)}_n&=
\begin{cases}
{\{\tau_{N^{1+\kappa}}^2 < n^{1-\frac{\gamma\wedge (2-\gamma)}{4(\gamma+2)}}\}} & \text{if } \gamma < 2\\
{\{\tau_{N^{1+\kappa}}^2 < n^{1+\frac{2\gamma - 1}{2(\gamma+2)}}\}} & \text{if } \gamma \ge 2,\\
\end{cases}\\
A_n^{(2)}(\gep_0) &:= {\{ T_k \le \gep_0^{\frac{1}{2\gamma}} N^{\frac{1}{\gamma}},\quad \forall k\le \gep_0N\}}\\
A_n^{(3)}(\gep_0) &:= \{\tau_{N/\gep_0} < n\}\\
A_n^{(4)}(\delta,\gep) &:=\{\exists \ell \in \{N,\ldots, 2N\}\colon T_\ell \ge T_0 \vee \delta N^{\frac{1}{\gamma}}\}\\
A_n^{(5)}(\gep_0,\gep) &:=\{\forall k\in R_{\gep_0}(n),\ T^*_k > T_0 \vee \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}\}\\
A_n^{(6)}(\gep_0,\gep) &:=\{\forall k\in R_{\gep_0}(n),\ f_k^{L_k} \le \exp(n^{\frac{\gamma}{2(\gamma+2)}})\}\\
A_n^{(7)}(\gep_0,\gep) &:=\{\forall k\in R_{\gep_0}(n),\ T^*_k > h(f_k, L_{k,\alpha}, \alpha)\}\\
A_n^{(8)}(\gep_0) &:= \{|R(1,N/\gep_0)| \le [\log(N/\gep_0)]^2\}\\
A_n^{(9)}(\delta) &:=\{ |\lambda(2N,\beta) - \lambda(\beta)| \le \tfrac{C_1}{2\delta^{2}}\}\\
A_n^{(10)}(\gep_0,\gep,\eta) &:= \{ |\lambda(\ell-1,b) - \lambda(b)| \le \tfrac{\gep_0\eta}{2},\ \forall \ell\ge \gep_0N,\ b\in \{\beta, \beta-\gep\}\}\\
A^{(11)}_n(\gep_0)&:=\{ \Pi_N( \{y>1/\gep_0\}) = 0 \}.
\end{aligned}
\end{equation}
We might omit some or all parameters of $\Omega_n$ when no confusion is possible to alight notations.
Event $A_n^{(1)}$ is used to provide a lower bound on $Z_n$, see Proposition \ref{prop:lb}. Events from $A_n^{(2)}$ to $A_n^{(9)}$
are used to establish the upper bound, see Proposition \ref{prop:ub}. Finally, the events $A_n^{(10)}$ and $A_n^{(11)}$ are used in the conclusion of the proof, see \eqref{eq:ubGtilde} and \eqref{eq:useAn10} in Section \ref{subsec:conclusion}.
\section{Lower bound}\label{sec:LB}
In this section we prove a lower bound on $Z_n$ that is an upper bound on $F_n$.\\
Let us set for $\beta \ge 0$, $n\in{\ensuremath{\mathbb N}} $ and $\ell > 1$,
\begin{equation}
\label{eq:defG}
G^\beta_n(\ell) = -\frac{1}{N}\log \bP^\beta(\sigma \wedge H_0> H_{\tau_{\ell-1}}) + g(T_\ell)\frac{n}{N} = \lambda(\ell-1,\beta)\frac{\ell-1}{N}+ g(T_\ell)\frac{n}{N},
\end{equation}
where $\lambda(\ell,\beta)$ has been defined in~\eqref{def:gl.ell.gb}. Recall the definition of $\kappa$ in~\eqref{def:kappa}.
Our goal in this section is to prove
\begin{proposition}
\label{prop:lb}
On $\Omega_n$,
\begin{equation}
F_n \le \min_{1< \ell \le N^{1+\kappa}} G_n(\ell) + o(1)
\end{equation}
where the $o(1)$ holds when $n$ goes to $+\infty$.
\end{proposition}
Actually, only $A_n^{(1)}$ is necessary in Lemma \ref{lem:lem.proplb} and this event does not depend on any parameter. That is why we omit the parameters in $\Omega_n$ above.
In order to prove this proposition we need the following lemma, that states that the exponential cost of reaching level $N^{1+\kappa}$ (and so any level $\ell \leq N^{1+\kappa}$) before time $n$ is negligible in front of $N$.
\begin{lemma}
\label{lem:lem.proplb}
There exists a function $k(N)$ such that $k(N)=o(N)$ as $N\to\infty$ and, on $\Omega_n$, for all $1< \ell <N^{1+\kappa}$,
\begin{equation}
\bP(H_{\tau_{\ell}}\le n | H_{\tau_{\ell}} < H_0 ) \ge \exp(-k(N)).
\end{equation}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:lem.proplb}]
Observe that
\begin{equation}
\begin{aligned}
\bP(H_{\tau_{\ell}}\le n | H_{\tau_{\ell}} < H_0 ) &\geq \bP(S_n \geq \tau_{\ell}\ ;\ S_k >0,\ 0<k\leq n )\\
&\geq \bP(S_n \geq \tau_{\ell})\bP( S_k >0,\ 0<k\leq n )\\
&\geq \bP(S_n \geq \tau_{N^{1+\kappa}})\bP( S_k >0,\ 0<k\leq n ).
\end{aligned}
\end{equation}
To go from the first to the second line we use the FKG inequality, since both events are non-decreasing coordinate-wise in the family of i.i.d. increments $(S_i-S_{i-1})_{1\leq i \leq n}$. As there exists some constant $C>0$ such that $\bP( S_k >0,\ 0<k\leq n )\geq C/\sqrt{n}$ we only focus on $\bP(S_n \geq \tau_{N^{1+\kappa}})$.
Assume first that $\gamma<2$. Then, we notice that, on $A^{(1)}_n$, $\tau_{N^{1+\kappa}}= o(\sqrt{n})$ when $n$ goes to infinity so that
\begin{equation}
\bP(S_n \geq \tau_{N^{1+\kappa}})=\bP(S_n \geq 0)-\bP(0\leq S_n < \tau_{N^{1+\kappa}})=1/2+o(1).
\end{equation}
Assume now that $\gamma\ge 2$. In this case, $\tau_{N^{1+\kappa}}$ is not anymore negligible in front of $\sqrt{n}$. However, on $A_n^{(1)}$, a standard large deviation estimate for the simple random walk yields
\begin{equation}
\bP(S_n \geq \tau_{N^{1+\kappa}})\ge \exp\Big(-Cn^{\frac{2\gamma-1}{2(\gamma+2)}}\Big).
\end{equation}
We may conclude by setting $k(N) = CN^{1-\frac{1}{2\gamma}}$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:lb}]
We provide a lower bound on $Z_n$ by computing the cost of various strategies. Here is an informal description of the tested strategies: for $1 < \ell \le N^{1+\kappa}$, the walk reaches $\tau_{\ell-1}$ before time $n$ and before its death, which has a probability of order $e^{-\lambda \ell}$. Then, it remains confined in the gap $(\tau_{\ell-1}, \tau_{\ell})$ until time $n$, with a probability of order $e^{- g(T_\ell) n}$. We finally optimise on $1 < \ell \le N^{1+\kappa}$.
\par We may thus write for all $\ell > 1$,
\begin{equation}
Z_n \ge Z_n^{[\ell]} := \bP(H_{\tau_\ell}\wedge \sigma \wedge H_{0} > n \ge H_{\tau_{\ell-1}}),
\end{equation}
and then optimize { over $1 < \ell \le N^{1+\kappa}$}.
By decomposing on the value of $H_{\tau_{\ell-1}}$, we have on $\Omega_n$ and for $n$ large enough,
\begin{equation}
\label{eq:lb}
\begin{aligned}
Z_n^{[\ell]} \ &{ \geq } \sum_{0\le k \le n} \bP(\sigma \wedge H_{0} > H_{\tau_{\ell-1}} = k) \bP_{\tau_{\ell-1}}(H_{\tau_\ell} \wedge H_{\tau_{\ell-1}} > n-k)\\
& \ge \sum_{0\le k \le n} \bP(\sigma \wedge H_{0} > H_{\tau_{\ell-1}} = k) \bP_{\tau_{\ell-1}}(H_{\tau_\ell} \wedge H_{\tau_{\ell-1}} > n)\\
& \ge \bP(\sigma \wedge H_{0}> H_{\tau_{\ell-1}}, H_{\tau_{\ell-1}}\le n) \times \frac{c_1}{2\sqrt{n}} e^{- g(T_\ell) n} \\
& \ge \bP(\sigma \wedge H_{0}> H_{\tau_{\ell-1}})\bP(H_{\tau_{\ell-1}}\le n | H_{0}>H_{\tau_{l-1}}) \times \frac{c_1}{2\sqrt{n}} e^{- g(T_\ell) n}\\
& \ge \frac{c_1}{2\sqrt{n}} \bP(\sigma \wedge H_{0} > H_{\tau_{\ell-1}})e^{- g(T_\ell) n - k(N)}.
\end{aligned}
\end{equation}
Note that we have used Proposition~\ref{pr:small_ball} to go from the second to the third line and Proposition~\ref{prop:FKG} to go from the third to the fourth line. Finally, to go from the fourth to the fifth line we use Lemma~\ref{lem:lem.proplb} and the fact that the environment is in $\Omega_n$.
Therefore,
\begin{equation}
F_n \le -\frac1N \log \Big( \frac{c_1}{2\sqrt{n}} \Big) + \frac{k(N)}{N} + \inf_{ 1< \ell \le N^{1+\kappa}} G_n^\beta(\ell),
\end{equation}
where $G_n^\beta$ is defined in \eqref{eq:defG}. Since
\begin{equation}
-\frac1N \log \Big( \frac{c_1}{2\sqrt{n}} \Big) + \frac{k(N)}{N} = o(1),
\end{equation}
as $N\to \infty$, this concludes the proof.
\end{proof}
\section{Upper bound}\label{sec:UB}
In this section we prove an upper bound on $Z_n$ or, equivalently, a lower bound on $F_n$. Recall the definitions in \eqref{eq:def.R.cR}.
\begin{proposition}
\label{prop:ub}
Let $\epsilon,\delta>0$. There exists $\gep_0>0$ such that, on $\Omega_n(\delta,\gep_0,\gep)$,
\begin{equation}
\label{eq:goal}
F_n \geq \min_{\ell\in \mathcal{R}_{\gep_0}(n)} G_n^{\beta-\epsilon}(\ell)+o(1).
\end{equation}
where the $o(1)$ holds as $n$ goes to infinity and $G_n^\beta$ is defined in \eqref{eq:defG}.
\end{proposition}
Before starting the proof of this proposition we need additional notations. Recall~\eqref{def:ik}. We define the hitting time of the $k$-th record
\begin{equation}
H^*_k = H_{\tau_{i(k)}}, \qquad k\ge 0.
\end{equation}
We also define
\begin{equation}
\tilde H_0 = 0,\qquad \tilde H_i = \inf\{n> \theta_{i-1} \colon S_n \in \tau\} - \tilde H_{i-1}, \qquad i\ge 1.
\end{equation}
For all $n\geq 1$ and $k\geq 1$, we define
\begin{equation}
Z_n^{(k)} = Z_n(H^*_k \le n < H^*_{k+1})
\end{equation}
where $Z_n(A)= \bP^{\beta}(\sigma \wedge H_{\mathbb{Z}^-} > n, A)$ for any event $A$. The quantity above corresponds to the survival probability restricted to the event that the walker has reached the $k$-th record but not the $(k+1)$-th one. These events are clearly incompatible for different values of $k$.
\begin{proof}[Proof of Proposition \ref{prop:ub}] Let $\epsilon,\delta>0$.
We choose $\gep_0$ small enough so that
\begin{equation}
\label{eq:gep0}
\begin{aligned}
\frac{\beta}{\gep_0}& >2(C_1 \delta^{-2} + \lambda(\beta)),\\
\gep_0^{-1/\gamma}& >4C_1(C_1 \delta^{-2} + \lambda(\beta)),\\
\end{aligned}
\end{equation}
(these technical conditions will become clear soon).
We have to prove that \eqref{eq:goal} is satisfied on $\Omega_n$. We thus consider until the end of this proof an environment that is in $\Omega_n$.
\par \textbf{Step $1$.} We first prove that we can actually consider only the records in $R_{\gep_0}(n)$, that are the ones lying in the window $\{\gep_0 N , \cdots, N/\gep_0\}$, see \eqref{eq:step1} for a precise formulation. As the environment is in $A_n^{(1)} \cap A_n^{(4)}(\delta,\gep)$, using \eqref{eq:lb} and \eqref{eq:encadr_g}, we obtain for $n$ large enough, a rough lower bound on $Z_n$:
\begin{equation}
\begin{aligned}
Z_n \ &{\geq } \max_{ N \leq \ell \leq 2N} Z_n^{[\ell]} \\
& \ge \max_{N \leq \ell \leq 2N} \frac{c_1}{2\sqrt{n}} \bP(\sigma \wedge H_0 > H_{\tau_{\ell-1}})e^{- g(T_\ell) n - k(N)}\\
& \ge \frac{c_1}{2\sqrt{n}} \exp\Big(- \Big\{ C_1\delta^{-2}N + k(N)+ 2N \lambda(2N,\beta)\Big\}\Big).
\end{aligned}
\end{equation}
As the environment is in $A_n^{(9)}(\delta)$, we finally get
\begin{equation}
\label{1}
Z_n \ \geq \frac{c_1}{2\sqrt{n}} \exp\Big\{-N \Big(2C_1 \delta^{-2} + 2\lambda(\beta) + o(1)\Big)\Big\},
\end{equation}
where the $o(1)$ holds as $n\to\infty$.
Observe that
\begin{equation}
\label{2}
\sum_{k\in R(N/\gep_0,+\infty)} Z_n^{(k)}\leq e^{-\beta N/\gep_0},
\end{equation}
while due to Proposition \ref{prop:roughUB} and the fact that the environment is in $A_n^{(2)}(\gep_0)$, we have for $n$ large enough,
\begin{equation}
\begin{aligned}
\label{3}
\sum_{k\in R(0,\gep_0 N)} Z_n^{(k)}\leq Z_n(H_{\tau_{\gep_0 N}}>n)&\le Cn^2 (\gep_0 N) \exp(-\phi(\beta, \max\{T_i,\ i\le \gep_0 N\})n)\\
& \le Cn^2 (\gep_0 N) \exp(-\frac{n}{C_1(\gep_0^{1/2\gamma} N^{1/\gamma})^2})\\
& \le \exp \Big({-\frac{ \gep_0^{-1/\gamma}N}{2C_1}}\Big).
\end{aligned}
\end{equation}
Combining \eqref{1}, \eqref{2} and \eqref{3} and due to the choice of $\gep_0$ in \eqref{eq:gep0}, we easily get that for $n$ large enough
\begin{equation}
\label{eq:step1}
Z_n \leq 2 \sum_{k \in R_{\gep_0}(n)} Z_n^{(k)}.
\end{equation}
\textbf{Step $2$.} The main work is thus to derive an upper bound on $Z_n^{(k)}$ for $k \in R_{\gep_0}(n)$ (see \eqref{eq:step2}).
\par Using the Markov property at time $H^*_k$ we observe that for $k \in R_{\gep_0}(n)$
\begin{equation}
\begin{aligned}
\label{eq:auxZnk}
Z_n^{(k)}&=\bP^\beta \left( \sigma \wedge H_{\mathbb{Z}^-}> n, H^*_k \le n < H^*_{k+1} \right)\\
&=\sum_{m=0}^n \bP^\beta \left( \sigma \wedge H_{\mathbb{Z}^-} > m, H^*_k =m \right)\bP_{\tau_{i(k)}}^\beta \left( \sigma \wedge H^*_{k+1} \wedge H_{\mathbb{Z}^-}>n-m \right).
\end{aligned}
\end{equation}
For all $m\geq 0$, by Proposition \ref{prop:roughUB}, we have on $A^{(3)}_n(\gep_0)$, for $n$ large enough,
\begin{equation}
\begin{aligned}
\label{eq:auxZnk2}
\bP_{\tau_{i(k)}}^\beta \left( \sigma \wedge H^*_{k+1} \wedge H_{\mathbb{Z}^-} >n-m \right)&\leq C (n-m)^2 i(k+1)e^{-\phi(\beta;\max\{T_i\colon 0\leq i< i(k+1)\})(n-m)}\\
&\leq C\ n^3\ e^{-\phi(\beta, T^*_k)(n-m)}.
\end{aligned}
\end{equation}
\par It remains to bound $\bP^\beta \left( \sigma \wedge H_{\mathbb{Z}^-} > m, H^*_k =m \right)$ for $0\le m \le n$.
Recall~\eqref{def:zeta} and~\eqref{def:zeta.star}.
By using Tchebychev's inequality in the first line and then conditioning on $X$, we obtain for
$\phi:=\frac{\pi^2}{2(T^*_k)^2}$,
\begin{equation}
\begin{aligned}
\label{eq:auxZnk3}
\bP^\beta \left( \sigma \wedge H_{\mathbb{Z}^-} > m, H^*_k =m \right)&=\bP^\beta(\zeta^*_{k} < \zeta_0 \wedge {\ensuremath{\mathcal N}} ,\ H^*_{k} = m) \\
&\leq e^{-\phi m} \bE^\beta(e^{\phi H^*_{k}} {\sf 1}_{\{\zeta^*_{k} < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}})\\
&\leq e^{-\phi m} \bE^\beta( \bE(e^{\phi H^*_{k}}| X) {\sf 1}_{\{\zeta^*_{k} < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}})\\
&\leq e^{-\phi m} \bE^\beta\Big( \prod_{1\le i \le \zeta^*_{k}} \bE(e^{\phi \tilde H_i}| X_{i-1}, X_i) {\sf 1}_{\{\zeta^*_{k} < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big).\\
\end{aligned}
\end{equation}
Next, by integrating on ${\ensuremath{\mathcal N}} $ we obtain
\begin{equation}
\label{eq:UB0}
\bE^\beta\Big( \prod_{1\le i \le \zeta^*_{k}} \bE(e^{\phi \tilde H_i}| X_{i-1}, X_i) {\sf 1}_{\{\zeta^*_{k}< \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)=\bE\Big( \prod_{1\le i \le \zeta^*_{k}} e^{-\beta} \frac{Q_{X_{i-1},X_i}(\phi)}{Q_{X_{i-1},X_i}(0)} {\sf 1}_{\{\zeta^*_{k} < \zeta_0\}} \Big)
\end{equation}
with notations similar to \eqref{eq:defQij}.
\par On $A_n^{(5)}(\gep_0,\gep)$ the assumptions of Lemma~\ref{lem:control_ratio} are valid (with $T_k^*$ playing the role of $T$), which provides $\alpha>0$. Recall the definition of ${\ensuremath{\mathcal B}} _{k,\alpha}$ in \eqref{eq:defB}. We obtain
\begin{equation}
\begin{aligned}
\label{eq:interle}
&\bE\Big( \prod_{1\le i \le \zeta^*_{k}} e^{-\beta} \frac{Q_{X_{i-1},X_i}(\phi)}{Q_{X_{i-1},X_i}(0)} {\sf 1}_{\{\zeta^*_{k} < \zeta_0\}} \Big)\\
&\qquad\leq \bE^{\beta-\gep}\Big( 2 f\left(\pi \frac{T^*_{k-1}}{T^*_k}(1+\frac{C}{(T^*_k)^2})\right)^{\sharp\{i\le \zeta_k^* \colon \{X_{i-1},X_i\} \in {\ensuremath{\mathcal B}} _{k,\alpha}\}} {\sf 1}_{\{\zeta^*_k < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big).
\end{aligned}
\end{equation}
Recall the definition of $f_k$ in~\eqref{eq:def_fk}. On $A^{(7)}_n(\gep_0,\gep)$ the assumptions of Lemma \ref{lem:control_pinbad} (with $f_{k}$ playing the role of $A$) are satisfied and from \eqref{eq:interle} we obtain that on $A^{(3)}_n (\gep_0)$
\begin{equation}
\begin{aligned}
\label{eq:auxZnk4}
\bE\Big( \prod_{1\le i \le \zeta^*_k} e^{-\beta} \frac{Q_{X_{i-1},X_i}(\phi)}{Q_{X_{i-1},X_i}(0)} {\sf 1}_{\{\zeta^*_k< \zeta_0\}} \Big)
& \leq \bE^{\beta-\gep}\Big( f_{k}^{\sharp\{i\le \zeta_k^* \colon \{X_{i-1},X_i\} \in {\ensuremath{\mathcal B}} _{k,\alpha}\}} {\sf 1}_{\{\zeta^*_k < \zeta_0 \wedge {\ensuremath{\mathcal N}} \}} \Big)\\
& \leq 2 f_{k}^{L_{k,\alpha}} n\ \bP^{\beta-\epsilon}(\zeta^*_k < \zeta_0\wedge {\ensuremath{\mathcal N}} ).
\end{aligned}
\end{equation}
\par Finally, combining \eqref{eq:auxZnk}, \eqref{eq:auxZnk2}, \eqref{eq:auxZnk3} and \eqref{eq:auxZnk4} we obtain for $k\in R_{\gep_0}(n)$
\begin{equation}
\label{eq:inter}
\begin{aligned}
Z_n^{(k)} &\leq 2C\ n^4\ f_{k}^{L_{k,\alpha}} \ e^{-\phi(\beta, T^*_k)n}\ \bP^{\beta-\epsilon}(\zeta^*_k < \zeta_0\wedge {\ensuremath{\mathcal N}} ) \sum_{m=0}^n e^{-\Big(\frac{\pi^2}{2(T^*_k)^2} - \phi(\beta,T^*_k) \Big)m}\\
& \leq 2C\ n^4 f_{k}^{L_{k,\alpha}} e^{-G_n^{\beta-\epsilon}(i(k))N} e^{-\left(\phi(\beta ; T^*_k )-g(T^*_k)\right)n} \sum_{m=0}^n e^{-\Big(\frac{\pi^2}{2(T^*_k)^2} - \phi(\beta,T^*_k) \Big)m}.
\end{aligned}
\end{equation}
On $A_n^{(5)}(\gep_0,\gep)$ we control both the errors $\phi(\beta ; T^*_k )-g(T^*_k)$ and $\phi(\beta,T^*_k)-\frac{\pi^2}{2(T^*_k)^2}$.
Indeed due to \eqref{eq:g} and \eqref{eq:phi} there exists some constant $C(\beta)$ such that for $t$ large enough (depending on $\beta$),
\begin{equation}
\label{eq:erreur}
|\phi(\beta ; t)-\tfrac{\pi^2}{2t^2}| \vee |\phi(\beta ; t)-g(t)|<\frac{C(\beta)}{t^3},
\end{equation}
and we obtain that on $A_n^{(5)}(\gep_0,\gep)$ and for $n$ large enough
\begin{equation}
|\phi(\beta ; T^*_k)-\tfrac{\pi^2}{2(T^*_k)^2}| \vee |\phi(\beta ; T^*_k )-g(T^*_k)|\leq { \frac{C(\beta)}{\gep_0^{9/2\gamma}N^{3/\gamma}}.}
\end{equation}
From \eqref{eq:inter} we thus obtain for $n$ large enough and $k\in R_{\gep_0}(n)$
\begin{equation}
Z_n^{(k)} \leq 2C\ n^5 f_{k}^{L_{k,\alpha}} \exp\Big\{-N \min_{\ell \in \mathcal{R}_{\gep_0}(n)} G_n^{\beta-\epsilon} (\ell)+\ C(\beta)\gep_0^{-9/2\gamma}n^{\frac{\gamma-1}{\gamma+2}} \Big\}.
\end{equation}
We also remind that on $A^{(6)}_n(\gep_0,\gep)$, $f_{k}^{L_{k,\alpha}} \leq e^{n^{\frac{\gamma}{2(\gamma+2)}}}$ so that finally for $k \in R_{\epsilon_0}(n)$, and for $n$ large enough,
\begin{equation}
\label{eq:step2}
Z_n^{(k)} \leq 2C\ n^5\exp\Big(-N \min_{\ell \in \mathcal{R}_{\gep_0}(n)} G_n^{\beta-\epsilon} (\ell) +o(N)\Big).
\end{equation}
\textbf{Step $3$.} It remains to sum this last equation for $k \in R_{\gep_0}(n) $. As the environment is in $A_n^{(8)}(\gep_0)$, we finally obtain that for $n$ large enough
\begin{equation}
Z_n \leq \gep_0^{-1} 2C\ n^6\exp\Big(-N \min_{\ell \in \mathcal{R}_{\gep_0}(n)} G_n^{\beta-\epsilon}(\ell) +o(N)\Big)
\end{equation}
and
\begin{equation}
F_n \geq \min_{\ell \in \mathcal{R}_{\gep_0}(n)} G_n^{\beta-\epsilon}(\ell) +o(1),
\end{equation}
where the $o(1)$ holds as $n\to\infty$.
\end{proof}
\section{Proof of Theorem \ref{thm0}} \label{sec:pot}
This section is divided in two parts. In the first one we establish that, for $n$ large enough, environments are good with probability arbitrary close from $1$. The second one is devoted to the proof our main result Theorem \ref{thm0}. Due to the control on the environment we can actually restrict our analysis to the event of good environments so that results of Proposition \ref{prop:lb} and \ref{prop:ub} are in force and provide a precise control on $Z_n$.
\subsection{Environments are good with high probability}
The aim of this part is to prove the following proposition that assures a control of the probability that an environment is good:
\begin{proposition}\label{lem:ge}
For all $\theta>0$ there exists $\delta$ and $\gep_1(\delta)$ small enough such that for all $\gep_0<\gep_1(\delta)$, for all $\gep,\eta>0$
\begin{equation}
\liminf_{n\to +\infty} {\ensuremath{\mathbb P}} \left(\Omega_n(\delta,\gep_0,\gep,\eta) \right) \ge 1-\theta.
\end{equation}
\end{proposition}
We first establish various results on the \textit{records} of the sequence $(T_i)_{i\geq 1}$.
\subsubsection{Preliminaries on records}
\label{sec:records}
We say that $n\in{\ensuremath{\mathbb N}} $ is a record if $n=1$ or $n\ge 2$ and $T_n > \max\{T_1, \ldots, T_{n-1}\}$.
Let us define
\begin{equation}
I_n =
\begin{cases}
1 & \text{if $n$ is a record}\\
0 & \text{otherwise},
\end{cases}
\end{equation}
the indicator of a record, and
\begin{equation}
\mathfrak{R}_n = \sum_{k=1}^n I_k = |R(1,n)|
\end{equation}
the number of records between $1$ and $n$. It was noticed (see Rényi~\cite{Re62}) that when $T_1$ is a continuous random variable, the $I_n$'s are independent Bernoulli random variables with mean $\bE(I_n) = 1/n$. However, we deal here with the discrete case, where this simple structure breaks down because of the possibility of ties. Actually, this case seems to have attracted relatively little attention (see however~\cite{GLS} and references therein). In this section, we provide some results in the discrete case (moments and deviations from the mean for the number of records) that will be useful later, though we do not aim for optimality. We start with:
\begin{proposition}
\label{pr:records2}
For all $p\in{\ensuremath{\mathbb N}} $ and $1\le n_1 < n_2 < \ldots n_p$,
\begin{equation}
\bE(I_{n_1}\ldots I_{n_p}) \le (1/n_1)\ldots (1/n_p).
\end{equation}
\end{proposition}
As a consequence we obtain:
\begin{proposition}
\label{lem:records_dev}
For $b>1$ there exists a positive constant $c=c(b)$ (which we may choose smaller but arbitrarily close to $\sup_{\lambda>0}\{b\lambda + 1 - e^{\lambda}\}=1+b(\ln b -1) >0$) such that for $n$ large enough,
\begin{equation}
\label{eq:records_dev2}
\bP(\mathfrak{R}_n \ge b\log n) \le n^{-c}.
\end{equation}
\end{proposition}
\begin{proof}[Proof of Proposition \ref{pr:records2}]
We prove it by iteration on $p\ge 1$ and thus start with the case $p=1$. Let $n\ge 2$ (the statement is trivial for $n=1)$. Let $J_n$ be the indicator of the event that there is a strict maximum among the $n$ first variables, meaning that
\begin{equation}
\label{eq:Jn}
J_n = {\sf 1}_{\{\exists 1\le i \le n \colon \forall 1\le j \le n,\ j\neq i,\ T_i > T_j\}}.
\end{equation}
By exchangeability of the $n$ first variables we get that
\begin{equation}
\label{eq:Jn2}
\bE(I_n) = \frac1n \bE(J_n) \le \frac1n.
\end{equation}
Suppose now that the statement is true for $p\ge1$ and let $2\le n_1 < \ldots < n_{p+1}$ (if $n_1= 1$, we can safely remove it). Again, by exchangeability of $(T_1, \ldots, T_{n_1})$ we have
\begin{equation}
\bE(I_{n_1}\ldots I_{n_{p+1}}) = \frac{1}{n_1} \bE({J_{n_1}}I_{n_2}\ldots I_{n_{p+1}}) \le \frac{1}{n_1} \bE(I_{n_2}\ldots I_{n_{p+1}}),
\end{equation}
and the result follows by using the induction hypothesis.
\end{proof}
\begin{proof}[Proof of Proposition \ref{lem:records_dev}]
Let $\lambda>0$ to be specified later. By Chernoff's bound,
\begin{equation}
\bP(\mathfrak{R}_n \ge b\log n) \le e^{-b\lambda \log n}\bE[e^{\lambda R_n}].
\end{equation}
Since $\mathfrak{R}_n = \sum_{k=1}^n I_k$ and the $I_k$'s are $\{0,1\}$valued random variables, we get
\begin{equation}
e^{\lambda \mathfrak{R}_n} = \prod_{k=1}^n (1+ [e^\lambda-1]I_k) = 1 + \sumtwo{J\subseteq \{1,\ldots, n\}}{J\neq\eset} [e^{\lambda}-1]^{|J|} \prod_{j\in J} I_j.
\end{equation}
By taking the expectation and using Proposition~\ref{pr:records2}, we get
\begin{equation}
\bE[e^{\lambda \mathfrak{R}_n}] \le \prod_{k=1}^n \Big(1+ [e^\lambda-1]\frac1k\Big) \le \exp([e^\lambda-1][1+o(1)]\log n ),
\end{equation}
where the $o(1)$ holds as $n\to\infty$. Finally, we obtain
\begin{equation}
\bP(\mathfrak{R}_n \ge b\log n) \le \exp(\{[e^\lambda-1][1+o(1)]-b\lambda\}\log n ),
\end{equation}
which concludes the proof.
\end{proof}
\subsubsection{Proof of Proposition~\ref{lem:ge}} \label{sec:proba.good.env}
We will use notations from the previous subsection during the proof, and sometimes write $\mathfrak{R}(n)$ instead of $\mathfrak{R}_n$ for the sake of readability. We consider the events $A_n^{(i)}$ for $1\le i \le 11$ and conclude by combining the results obtained in each cases.\\
Along the proof we will use that, by~\eqref{eq:tail.ass}, there exists $c_1,c_2>0$ such that for all $m\in {\ensuremath{\mathbb N}} $,
\begin{equation}
\label{eq:UB_tail_T}
1 - c_1m^{-\gamma} \leq {\ensuremath{\mathbb P}} (T_1 \le m) \le 1 - c_2m^{-\gamma} \le \exp(-c_2m^{-\gamma}).
\end{equation}
{\noindent \it {Case $i=1$}}. Assume first that $\gamma \le1$, so that $\kappa = \frac{\gamma}{4}$.
Since
\begin{equation}
N^{(1+\frac{\gamma}{4})\frac{2}{\gamma}} = o(n^{1-\frac{\gamma}{4(\gamma+2)}}),
\end{equation}
one has $\lim_{n\to\infty}{\ensuremath{\mathbb P}} (A_n^{(1)}) = 0$ from Proposition~\ref{thm:limit.stable}. Assume now that $\gamma>1$. Then,
\begin{equation}
N^{2(1+\kappa)} =
\begin{cases}
o(n^{1-\frac{2-\gamma}{4(2+\gamma)}}) & \text{if } \gamma\in(1,2)\\
o(n^{1+\frac{2\gamma-1}{2(\gamma+2)}}) & \text{if } \gamma \ge2,
\end{cases}
\end{equation}
and again $\lim_{n\to\infty}{\ensuremath{\mathbb P}} (A_n^{(1)}) = 0$ from Proposition~\ref{thm:limit.stable}.\\
{\noindent \it {Case $i=2$}}. Note that, by \eqref{eq:UB_tail_T},
\begin{equation}
\begin{aligned}
{\ensuremath{\mathbb P}} \left(A_n^{(2)}(\gep_0)\right) = {\ensuremath{\mathbb P}} \left(T_1 \leq \gep_0^{\frac{1}{2\gamma}}N^{1/ \gamma}\right)^{\gep_0 N}\geq \left(1-\frac{c_1}{\gep_0^{\frac{1}{2}}N}\right)^{\gep_0 N}=e^{-c_1\gep_0^{1/2}}+o(1)\qquad n\to +\infty.
\end{aligned}
\end{equation}
We obtain $\lim_{\gep_0\to 0} \liminf_{n \to +\infty} {\ensuremath{\mathbb P}} (A_n^{(2)}(\gep_0))=1$.\\
{\noindent \it {Case $i=3$}}. Here we note that $N^{1/\gamma} = o(n)$ when $\gamma\le1$ and $N = o(n)$ when $\gamma>1$, and we conclude using Proposition \ref{thm:limit.stable} that for all $\gep_0>0$
\begin{equation}
\lim_{n\to\infty}{\ensuremath{\mathbb P}} (A_n^{(3)}(\gep_0)) = 0.
\end{equation}
{\noindent \it {Case $i=4$}.} By independence of the $T_\ell$'s, one has
\begin{equation}
{\ensuremath{\mathbb P}} (A_n^{(4)}(\delta,\gep)^c) \le {\ensuremath{\mathbb P}} (T_1 < T_0(\gep) \vee \delta N^{\frac{1}{\gamma}})^N.
\end{equation}
Therefore,
\begin{equation}
{\ensuremath{\mathbb P}} (A_n^{(4)}(\delta,\gep)^c) \le \exp( - c(NT_0(\gep)^{-\gamma} \wedge \delta^{-\gamma})),
\end{equation}
from which we deduce that for all $\gep>0$
\begin{equation}
\label{eq:cas4}
\lim_{\delta\to 0} \liminf_{n\to\infty} {\ensuremath{\mathbb P}} (A_n^{(4)}(\delta,\gep)) = 1.
\end{equation}
{\noindent \it {Case $i=5$}.} Recall the definitions of $R_{\gep_0}(n)$ and $T^*_k$ in~\eqref{eq:def.t.tau.star} and~\eqref{eq:def.R.cR}. We get
\begin{equation}
{\ensuremath{\mathbb P}} (A_n^{(5)}(\gep_0,\gep)^c) \le {\ensuremath{\mathbb P}} (\forall \ell \le \gep_0N,\ T_\ell \leq T_0(\gep) \vee \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}) \le {\ensuremath{\mathbb P}} (T_1 \leq T_0(\gep) \vee \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}})^{\gep_0 N}.
\end{equation}
By using again~\eqref{eq:UB_tail_T}, we obtain
\begin{equation}
{\ensuremath{\mathbb P}} (A_n^{(5)}(\gep_0,\gep)^c) \le \exp(-c_2 (\gep_0T_0(\gep)^{-\gamma}N \wedge \gep_0^{-1/2})),
\end{equation}
from which we get for all $\gep>0$
\begin{equation}
\label{eq:cas5}
\lim_{\gep_0\to 0} \liminf_{n\to\infty} {\ensuremath{\mathbb P}} (A_n^{(5)}(\gep_0,\gep)) = 1.
\end{equation}
{\noindent \it {Case $6\le i\le 8$}}. Let us first prove that for $\gep_0 >0$,
\begin{equation}
\label{eq:cas8}
\lim_{n\to\infty} {\ensuremath{\mathbb P}} (A_n^{(8)}(\gep_0)) =1.
\end{equation}
Indeed, since $|R(1,N/\gep_0)| = \mathfrak{R}(N/\gep_0)$, it is enough to prove that
\begin{equation}
{\ensuremath{\mathbb P}} (\mathfrak{R}_n \le (\log n)^2) \to 1,\qquad n\to\infty,
\end{equation}
which can be easily deduced from Proposition~\ref{lem:records_dev}. We now deal with the event corresponding to $i=7$.
Fix $\gep_0,\gep>0$, and note that
\begin{equation}
\label{eq:cas7aux1}
\begin{aligned}
{\ensuremath{\mathbb P}} (A_n^{(7)}(\gep_0,\gep)^c) \le {\ensuremath{\mathbb P}} ( & A_n^{(8)}(\gep_0)^c) + {\ensuremath{\mathbb P}} (A_n^{(5)}(\gep_0,\gep)^c)\\
& +{\ensuremath{\mathbb P}} \Big(\exists k\leq [\log(N/\gep_0)]^2 \colon \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}} \le T^*_k \le h(f_k,L_k,\alpha(\gep)),\ i(k)\le N/\gep_0 \Big).
\end{aligned}
\end{equation}
By a union bound we are left to prove that
\begin{equation}
\label{eq:but}
p_k(n,\gep_0) = o([\log n]^{-2}),
\end{equation}
where
\begin{equation}
\label{eq:628}
p_k(n,\gep_0,\gep) := {\ensuremath{\mathbb P}} (\gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}} \le T^*_k \le h(f_k,L_k,\alpha(\gep)),\ i(k)\le N/\gep_0).
\end{equation}
From~\eqref{eq:def.gALga} in the proof of Lemma~\ref{lem:control_pinbad}, we have for $n$ large enough,
\begin{equation}
\label{eq:pk1pk2}
p_k(n,\gep_0,\gep) \le {\ensuremath{\mathbb P}} (f_k^{L_k} \ge C(\gep_0,\gep) N^{\frac{1}{2\gamma}} ,\ T_k^* \ge \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}},\ i(k)\le N/\gep_0).
\end{equation}
Let us first show that
\begin{equation}
\label{eq:LBfk}
{\ensuremath{\mathbb P}} \Big(f_k > A,\ T_k^* \ge \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}\Big) \le
C \Big[A \wedge \gep_0^{\frac{3}{\gamma}}N^{\frac{2}{\gamma}}\Big]^{-1}.
\end{equation}
To this end, our first ingredient is the following upper bound, which holds for $u\in(0,1)$ and $k\in{\ensuremath{\mathbb N}} $:
\begin{equation}
\label{eq:ratio_rec_UB}
{\ensuremath{\mathbb P}} \Big(\frac{T^*_{k}}{T^*_{k+1}}\ge1-u\Big) \le C u.
\end{equation}
We show this inequality at the end of the proof, in order not to disrupt the main line of ideas.
We will also use the following elementary bound:
\begin{equation}
\label{eq:ineqf}
f(u)=\frac{u}{\sin(u)}\leq \frac{C}{\pi-u}, \qquad u\in (0,\pi).
\end{equation}
Coming back to \eqref{eq:LBfk}, we have by using~\eqref{eq:ineqf},
\begin{equation}
\begin{aligned}
{\ensuremath{\mathbb P}} \Big(f_k > A,\ T_k^* \ge \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}\Big) &\le
{\ensuremath{\mathbb P}} \Big( \frac{T_{k-1}^*}{T_{k}^*} \Big[1+\frac{C}{(T_{k}^*)^2}\Big] \ge 1 - \frac{C}{A},\ T_k^* \ge \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}\Big)\\
& \le {\ensuremath{\mathbb P}} \Big( \frac{T_{k-1}^*}{T_{k}^*}\ge 1 - \frac{C}{A} - \frac{C}{\gep_0^{\frac{3}{\gamma}}N^{\frac{2}{\gamma}}}\Big),
\end{aligned}
\end{equation}
and we obtain~\eqref{eq:LBfk} thanks to~\eqref{eq:ratio_rec_UB}. In view of \eqref{eq:but}, we choose $A=A_n=[\log n]^3$. For $k\geq 1$, from \eqref{eq:pk1pk2},
\begin{equation}
\label{eq:634}
p_k(n,\gep_0,\gep)\leq {\ensuremath{\mathbb P}} \Big(f_k \ge A_n ,\ T_k^* \ge \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}\Big)
+{\ensuremath{\mathbb P}} \Big(L_k \ge C \frac{\log n}{\log \log n},\ T_k^* \ge \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}},\ i(k)\le N/\gep_0\Big).
\end{equation}
Using \eqref{eq:LBfk}, the first term in the sum above is $o[(\log n)^{-2}]$. We now deal with the second one.
On the corresponding event, we have
\begin{equation}
\label{eq:event.bin}
\mathrm{card}\Big\{1\le j \le N/\gep_0 \colon T_j \ge \alpha \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}} \Big\} \ge C \frac{\log n} {\log \log n}.
\end{equation}
Furthermore, the random variable in the l.h.s. of the inequality follows a binomial law with parameters $N/\gep_0$ and ${\ensuremath{\mathbb P}} (T_1 \ge \alpha \gep_0^{\frac{3}{2\gamma}}N^{\frac{1}{\gamma}}) \le \alpha^{-\gamma}\gep_0^{-3/2}N^{-1}$ (up to a harmless constant). By using a standard binomial-to-Poisson approximation (cf.\ end of the proof)
\begin{equation}
\label{eq:approx.bin}
{\ensuremath{\mathbb P}} (\mathrm{Bin}(\ell,q)\ge m) \le \Big( \frac{q\ell}{m} \Big)^{m} e^{m-q\ell},\qquad \ell\in{\ensuremath{\mathbb N}} ,\ q\in(0,1),\ m\in{\ensuremath{\mathbb N}} \colon q\ell < m,
\end{equation}
with $q=\gep_0^{-3/2}N^{-1}$, $\ell = N/\gep_0$ and $m = C \log n/\log \log n$,
we get that
\begin{equation}
\label{eq:LB.Lk}
{\ensuremath{\mathbb P}} (\text{\eqref{eq:event.bin}}) = o(n^{-C/2}),
\end{equation}
which is enough to conclude. From what precedes we finally obtain that for all $\gep_0,\gep>0$,
\begin{equation}
\lim_{n\to\infty} {\ensuremath{\mathbb P}} (A_n^{(7)}(\gep_0,\gep)^c) = 0.
\end{equation}
The event corresponding to $i=6$ can be readily treated with the same idea, since the $N^{\frac{1}{2\gamma}}$ in \eqref{eq:pk1pk2} is less than $\exp(n^{\frac{\gamma}{2(\gamma+2)}})$. Finally,
\begin{equation}
\label{eq:cas67}
\lim_{n\to\infty} {\ensuremath{\mathbb P}} (A_n^{(6)}(\gep_0,\gep)^c \cup A_n^{(7)}(\gep_0,\gep)^c) = 0.
\end{equation}
{\noindent \it {Case $9\le i\le 10$}}. From the almost-sure convergence in Proposition~\ref{pr:lambda.beta}, we readily get that for any choice of $\delta,\gep_0,\gep,\eta$,
\begin{equation}
\label{eq:cas910}
\lim_{n\to \infty} {\ensuremath{\mathbb P}} (A_n^{(9)}(\delta) \cap A_n^{(10)}(\gep_0,\gep,\eta)) = 1.
\end{equation}
{\noindent \it {Case $i=11$}}. Note that by \eqref{eq:UB_tail_T},
\begin{equation}
\begin{aligned}
{\ensuremath{\mathbb P}} \left(A_n^{(11)}(\gep_0)\right) = {\ensuremath{\mathbb P}} \left(T_1 \leq \frac{N^{1/ \gamma}}{\gep_0}\right)^N \geq e^{-c_1 \gep_0^\gamma}+o(1) \qquad n \to +\infty .
\end{aligned}
\end{equation}
We obtain $\lim_{\gep_0\to 0} \liminf_{n \to +\infty} {\ensuremath{\mathbb P}} (A_n^{(11)}(\gep_0))=1$.
{\noindent \it Proof of \eqref{eq:ratio_rec_UB}.}
By writing $1+v = (1-u)^{-1}$ for convenience, one has
\begin{equation}
{\ensuremath{\mathbb P}} \Big(T^*_{k+1} \le (1+v){T^*_k} \Big|\ T^*_k = \ell \Big)
=
\begin{cases}
0 & \text{if } v < 1/\ell\\
{\ensuremath{\mathbb P}} (T_1 \le \lfloor(1+v)\ell\rfloor\ |\ T_1 > \ell) & \text{otherwise}.
\end{cases}
\end{equation}
From our assumption on the tail of $T_1$ and with the help of a standard comparison between series and integrals, the probability in the second case is bounded from above by
\begin{equation}
C\ell^{\gamma} \sum_{n=\ell+1}^{\lfloor(1+v)\ell\rfloor} n^{-(1+\gamma)} \le C\ell^{\gamma} \int_{\ell}^{\lfloor(1+v)\ell\rfloor-1} t^{-(1+\gamma)}\dd t.
\end{equation}
In turn, the integral above is controlled by
\begin{equation}
\int_{\ell}^{(1+v)\ell} t^{-(1+\gamma)}\dd t = C[1-(1+v)^{-\gamma}]\ell^{-\gamma} \le C u \ell^{-\gamma},
\end{equation}
which completes the proof of \eqref{eq:ratio_rec_UB}.\\
{\noindent \it Proof of \eqref{eq:approx.bin}.} By Chernoff's bound, one has for $x>0$,
\begin{equation}
{\ensuremath{\mathbb P}} (\mathrm{Bin}(\ell,q)\ge m) \le e^{-x m} {\ensuremath{\mathbb E}} (e^{x \mathrm{Ber}(q)})^\ell \le \exp(-x m+ q(e^x -1)\ell).
\end{equation}
Since $m> q\ell$, we may choose $x = \log(m/(q\ell))$ to minimize the r.h.s.\ in the line above and get the result.
\subsection{Conclusion : proof of Theorem \ref{thm0}} \label{subsec:conclusion}
To prove Theorem \ref{thm0}, we establish that for all $u\in {\ensuremath{\mathbb R}} $ which is a continuity point of the distribution function of $F$,
\begin{equation}
\label{main_eq}
\lim_{n\to\infty} {\ensuremath{\mathbb P}} (F_n \le u) = {\ensuremath{\mathbb P}} (F \le u).
\end{equation}
By \eqref{eq:Fcont} all real numbers are continuity points of $F$. Moreover since $F_n$ is positive, we only have to prove \eqref{main_eq} for $u>0$.
\par We start with the \textbf{upper bound} in \eqref{main_eq}
\begin{equation}
\label{main_eqUB}
\limsup_{n\to\infty} {\ensuremath{\mathbb P}} (F_n \le u) \leq {\ensuremath{\mathbb P}} (F \le u).
\end{equation}
Fix $\theta>0$ that we will let go to $0$ only at the very end of the proof. Fix also $\gep,\eta>0$. From Proposition~\ref{lem:ge}, there exists $\delta>0$ and $\gep_1(\delta)$ so that for $\gep_0<\gep_1(\delta)$
\begin{equation}
\label{eq:choice}
\liminf_{n}{\ensuremath{\mathbb P}} (\Omega_n(\delta,\gep_0,\gep,\eta))>1-\theta.
\end{equation}
Fix $\gep_0<\gep_1(\delta)$ small enough so that the conclusion of Proposition \ref{prop:ub} is satisfied. Thus we obtain that for $n$ large enough
\begin{equation}
\label{eq:grossier}
{\ensuremath{\mathbb P}} (F_n \leq u) \leq {\ensuremath{\mathbb P}} (\min_{\ell\in \mathcal{R}_{\gep_0}(n) } G_n^{\beta-\epsilon}(\ell)\leq u + \eta, \Omega_n) +{\ensuremath{\mathbb P}} (\Omega_n^c).
\end{equation}
From Proposition~\ref{lem:ge} and the choices of $\delta$ and $\gep_0$, ${\ensuremath{\mathbb P}} (\Omega_n^c)< \theta $ for $n$ large enough. Thus we just have to focus on the first term in the last equation.
We introduce for $n,\ell\geq 1$ and $\beta>0$, the random variable
\begin{equation}
\tilde{G}^{\beta}_n(\ell)=\frac{\lambda(\beta) (\ell-1)}{N}+\frac{\pi^2 n}{2 T_\ell^2 N}= \psi^{\lambda(\beta)}\left(\frac{\ell}{N}, \frac{T_{\ell}}{N^{1/\gamma}}\right).
\end{equation}
We replace $G$ by $\tilde{G}$ in \eqref{eq:grossier} and control the probability that both processes are not close
\begin{equation}
\label{eq:mainub}
\begin{aligned}
{\ensuremath{\mathbb P}} (\min_{\ell\in \mathcal{R}_{\gep_0}(n) } &G_n^{\beta-\epsilon}(\ell) \leq u + \eta, \Omega_n) \\
&\leq {\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} } \tilde{G}_n^{\beta-\epsilon}(\ell)< u +2 \eta, \Omega_n)+{\ensuremath{\mathbb P}} (\max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} |\tilde{G}_n^{\beta-\epsilon}(\ell)-G_n^{\beta-\epsilon}(\ell)| \geq \eta, \Omega_n).
\end{aligned}
\end{equation}
The first term in the sum gives the main contribution. Let us first prove that the second one is zero for $n$ large enough.
For $\ell\geq 1$ we define
\begin{equation}
\Delta_1(n,\ell):=\frac{\ell-1}{N}\left|\lambda (\beta-\epsilon) - \lambda(\ell-1,\beta-\gep)\right|
\end{equation}
and
\begin{equation}
\Delta_2(n,\ell):=\frac{n}{N}|g(T_\ell) -\frac{\pi^2}{2 T_\ell^2}|
\end{equation}
so that
\begin{equation}
\max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} |\tilde{G}_n^{\beta-\epsilon}(\ell)-G_n^{\beta-\epsilon}(\ell)| \leq \max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \Delta_1(n,\ell) + \max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \Delta_2(n,\ell).
\end{equation}
We first deal with $\Delta_2$. According to \eqref{eq:g} there exists some $C>0$ such that for all $\ell$, $|g(T_\ell) -\frac{\pi^2}{2 T_\ell^2}| \leq \frac{C}{T_\ell^4} $.
We can thus deduce that
\begin{equation}
{\ensuremath{\mathbb P}} ( \max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \Delta_2(n,\ell) \geq \eta, \Omega_n) \leq {\ensuremath{\mathbb P}} \left(\frac{C\ n}{N (\gep_0^{\frac{3}{2 \gamma }}N^{1/\gamma})^4} \geq \eta , A^{(5)}_n(\gep_0,\gep) \right),
\end{equation}
and this last term is $0$ for $n$ large enough.
We turn to the control of $\Delta_1$.
For $n\ge 1$,
\begin{equation}
\begin{aligned}
\label{eq:useAn10}
{\ensuremath{\mathbb P}} (\max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \Delta_1(n,\ell) \geq \eta, \Omega_n) &\leq {\ensuremath{\mathbb P}} (\max_{\ell\in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \Delta_1(n,\ell) \geq \eta, A_n^{(10)}(\gep_0,\gep,\eta)),
\end{aligned}
\end{equation}
and again the last term is $0$ for $n$ large enough.
Let us come back to the first term in \eqref{eq:mainub}, ${\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} } \tilde{G}_n^{\beta-\epsilon}(\ell)< u +2 \eta, \Omega_n)$.
As $\ell$ ranges ${\ensuremath{\mathcal R}} $ we may write
\begin{equation}
\min_{\ell \in {\ensuremath{\mathcal R}} } \tilde{G}_n^{\beta-\epsilon}(\ell)=\Psi^{\lambda(\beta-\epsilon)} \left(\Pi_N\right).
\end{equation}
We thus obtain
\begin{equation}
\label{eq:ubGtilde}
\begin{aligned}
{\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} }\ \tilde{G}_n^{\beta-\epsilon}(\ell)< u +2 \eta, \Omega_n ) &\leq {\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta-\epsilon)} \left(\Pi_N\right)< u +2 \eta, A^{(11)}_n(\gep_0) )\\
& \leq {\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta-\epsilon)}_K \left(\Pi_N\right) < u +2 \eta),
\end{aligned}
\end{equation}
where $K:=A^{\lambda(\beta-\gep)}_{u+2\eta} \cap \{y\leq 1/\gep_0\}$ with the set $A$ defined in \eqref{eq:defA}. As $K$ is compact, Proposition \ref{prop:convergenceP} and Lemma \ref{prop:continuity} assure that
\begin{equation}
{\ensuremath{\mathbb P}} (\Psi_K^{\lambda(\beta-\epsilon)} \left(\Pi_N\right) < u +2 \eta) \to {\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta-\epsilon)}_K(\Pi)< u+2\eta)
\end{equation}
when $n$ goes to infinity (we recall that $\Psi^{\lambda(\beta-\epsilon)}_K(\Pi)$ is continuous).
Using that $\Psi^{\lambda(\beta-\epsilon)} \leq \Psi^{\lambda(\beta-\epsilon)}_K$ we obtain
\begin{equation}
\begin{aligned}
{\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta-\epsilon)}_K(\Pi)< u+2\eta) &\leq {\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta-\epsilon)}(\Pi)< u+2\eta)\\
&={\ensuremath{\mathbb P}} (F^{\beta-\epsilon}< u+2\eta).
\end{aligned}
\end{equation}
Finally, we have proven
\begin{equation}
\limsup_{n\to +\infty}{\ensuremath{\mathbb P}} (F_n \leq u) \leq {\ensuremath{\mathbb P}} (F^{\beta-\epsilon}< u+2\eta)+{ \theta}.
\end{equation}
As $u\mapsto {\ensuremath{\mathbb P}} (F^{\beta-\epsilon}\leq u)$ is right-continuous,
\begin{equation}
\label{eq:presk}
\lim_{\eta \to 0} {\ensuremath{\mathbb P}} (F^{\beta-\epsilon}< u+2\eta) = {\ensuremath{\mathbb P}} (F^{\beta-\epsilon}\leq u).
\end{equation}
From Lemma \ref{lem:Fgep}, $F^{\beta-\epsilon} {\nearrow} F^{\beta}$ almost surely when $\gep$ goes to $0$ so that
\begin{equation}
{\ensuremath{\mathbb P}} (F^{\beta-\epsilon}\leq u){\to} {\ensuremath{\mathbb P}} (F^{\beta}\leq u) \qquad \epsilon \to 0.
\end{equation}
Finally,
\begin{equation}
\limsup_{n\to +\infty}{\ensuremath{\mathbb P}} (F_n \leq u) \leq {\ensuremath{\mathbb P}} (F^{\beta}\leq u)+{ \theta},
\end{equation}
and, as $\theta$ can be chosen arbitrarily small, we obtain the upper bound
\begin{equation}
\limsup_{n\to +\infty}{\ensuremath{\mathbb P}} (F_n \leq u) \leq {\ensuremath{\mathbb P}} (F^{\beta}\leq u).
\end{equation}
\par We turn now to the \textbf{lower bound} in \eqref{main_eq}:
\begin{equation}
\liminf_{n\to\infty} {\ensuremath{\mathbb P}} (F_n \le u) \geq {\ensuremath{\mathbb P}} (F \le u),
\end{equation}
or, equivalently,
\begin{equation}
\limsup_{n\to\infty} {\ensuremath{\mathbb P}} (F_n > u) \leq {\ensuremath{\mathbb P}} (F > u).
\end{equation}
The proof works essentially in the same way as for the upper bound. Again fix $\theta,\gep,\eta>0$ and, using Proposition~ \ref{lem:ge}, $\delta>0$ and $\gep_1(\delta)$ so that for $\gep_0<\gep_1(\delta)$
\begin{equation}
\label{eq:choice}
\liminf_{n}{\ensuremath{\mathbb P}} (\Omega_n(\delta,\gep_0,\gep,\eta))>1-\theta.
\end{equation}
We choose $\gep_0<\gep_1(\delta)$ small enough so that
\begin{enumerate}
\item the conclusion of Proposition \ref{prop:lb} is satisfied;
\item the following inequality holds
\begin{equation}
\label{eq:condlb}
\frac{\lambda(\beta)}{\gep_0}> 2\lambda(\beta)+ \frac{\pi^2}{2\delta^2}.
\end{equation}
\end{enumerate}
Using Proposition \ref{prop:lb}, for $n$ large enough,
\begin{equation}
\label{eq:presque}
\begin{aligned}
{\ensuremath{\mathbb P}} (F_n > u) &\leq {\ensuremath{\mathbb P}} (\min_{1< \ell \le N^{1+\kappa}} G^{\beta}_n(\ell)> u-\eta, \Omega_n) + {\ensuremath{\mathbb P}} (\Omega_n^c)\\
&\leq {\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} _{\gep_0}(n)} G^{\beta}_n(\ell)> u-\eta, \Omega_n) + {\ensuremath{\mathbb P}} (\Omega_n^c) \\
&\leq {\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \tilde{G}^{\beta}_n(\ell)> u-2\eta, \Omega_n)+{\ensuremath{\mathbb P}} (\max_{\ell \in {\ensuremath{\mathcal R}} _{\gep_0}(n) } |\tilde{G}_n^{\beta}(\ell)-G^{\beta}_n(\ell)| \geq \eta, \Omega_n)+{\ensuremath{\mathbb P}} (\Omega_n^c).
\end{aligned}
\end{equation}
The second term in this last equation is treated exactly in the same way as the second term in \eqref{eq:mainub} and is thus zero for $n$ large enough. The third one is smaller than $\theta$ by \eqref{eq:choice} for $n$ large enough. We thus focus on the first one. The choice of $\gep_0$ in \eqref{eq:condlb} implies that
\begin{equation}
\Omega_n\subset \{\argmin \tilde{G}_n < N/\gep_0\}.
\end{equation}
Indeed, as $\Omega_n\subset A^{(4)}_n(\delta,\gep)$, it holds that, on $\Omega_n$,
\begin{equation}
\min_{N\leq \ell \leq 2N} \tilde{G}^{\beta}_n < 2\lambda(\beta) + \frac{\pi^2}{2\delta^2}<\frac{\lambda(\beta)}{\gep_0}<\min_{\ell>N/\gep_0}\tilde{G}^{\beta}_n(\ell).
\end{equation}
Therefore, for any compact set $K$ in $E$,
\begin{equation}
\begin{aligned}
{\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} _{\gep_0}(n)} \tilde{G}^{\beta}_n(\ell)> u-2\eta, \Omega_n) &= {\ensuremath{\mathbb P}} (\min_{\ell \in {\ensuremath{\mathcal R}} } \tilde{G}^{\beta}_n(\ell)> u-2\eta, \Omega_n)\\
&={\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta)}(\Pi_N)> u-2\eta,\Omega_n)\\
&\leq {\ensuremath{\mathbb P}} (\Psi_K^{\lambda(\beta)}(\Pi_N)> u-2\eta).
\end{aligned}
\end{equation}
By Lemma \ref{prop:continuity}, ${\ensuremath{\mathbb P}} (\Psi_K^{\lambda(\beta)}(\Pi_N)> u-2\eta)$ converges to ${\ensuremath{\mathbb P}} (\Psi_K^{\lambda(\beta)}(\Pi)> u-2\eta)$ when $N$ goes to infinity.
Finally,
\begin{equation}
\limsup_{n\to\infty} {\ensuremath{\mathbb P}} (F_n > u) \leq {\ensuremath{\mathbb P}} (\Psi_K^{\lambda(\beta)}(\Pi)> u-2\eta) + \theta.
\end{equation}
By letting $K$ increase to $E$, we obtain
\begin{equation}
\limsup_{n\to\infty} {\ensuremath{\mathbb P}} (F_n > u) \leq {\ensuremath{\mathbb P}} (\Psi^{\lambda(\beta)}(\Pi)> u-2\eta) + \theta,
\end{equation}
and we conclude as for the upper bound by letting $\eta$ and $\theta$ go to $0$.
|
\section{%
\@startsection
{section}%
{1}%
{\z@}%
{0.8cm \@plus1ex \@minus .2ex}%
{0.5cm}%
{%
\normalfont\small\bfseries
\centering
}%
}%
\def\@hangfrom@section#1#2#3{\@hangfrom{#1#2}\MakeTextUppercase{#3}}%
\def\subsection{%
\@startsection
{subsection}%
{2}%
{\z@}%
{.8cm \@plus1ex \@minus .2ex}%
{.5cm}%
{%
\normalfont\small\bfseries
\centering
}%
}%
\def\subsubsection{%
\@startsection
{subsubsection}%
{3}%
{\z@}%
{.8cm \@plus1ex \@minus .2ex}%
{.5cm}%
{%
\normalfont\small\itshape
\centering
}%
}%
\def\paragraph{%
\@startsection
{paragraph}%
{4}%
{\parindent}%
{\z@}%
{-1em}%
{\normalfont\normalsize\itshape}%
}%
\def\subparagraph{%
\@startsection
{subparagraph}%
{5}%
{\parindent}%
{3.25ex \@plus1ex \@minus .2ex}%
{-1em}%
{\normalfont\normalsize\bfseries}%
}%
\def\section@preprintsty{%
\@startsection
{section}%
{1}%
{\z@}%
{0.8cm \@plus1ex \@minus .2ex}%
{0.5cm}%
{%
\normalfont\small\bfseries
}%
}%
\def\subsection@preprintsty{%
\@startsection
{subsection}%
{2}%
{\z@}%
{.8cm \@plus1ex \@minus .2ex}%
{.5cm}%
{%
\normalfont\small\bfseries
}%
}%
\def\subsubsection@preprintsty{%
\@startsection
{subsubsection}%
{3}%
{\z@}%
{.8cm \@plus1ex \@minus .2ex}%
{.5cm}%
{%
\normalfont\small\itshape
}%
}%
\@ifxundefined\frontmatter@footnote@produce{%
\let\frontmatter@footnote@produce\frontmatter@footnote@produce@endnote
}{}%
\def\@pnumwidth{1.55em}
\def\@tocrmarg {2.55em}
\def\@dotsep{4.5pt}
\setcounter{tocdepth}{3}
\def\tableofcontents{%
\addtocontents{toc}{\string\tocdepth@munge}%
\print@toc{toc}%
\addtocontents{toc}{\string\tocdepth@restore}%
}%
\def\tocdepth@munge{%
\let\l@section@saved\l@section
\let\l@section\@gobble@tw@
}%
\def\@gobble@tw@#1#2{}%
\def\tocdepth@restore{%
\let\l@section\l@section@saved
}%
\def\l@part#1#2{\addpenalty{\@secpenalty}%
\begingroup
\set@tocdim@pagenum{#2}%
\parindent \z@
\rightskip\tocleft@pagenum plus 1fil\relax
\skip@\parfillskip\parfillskip\z@
\addvspace{2.25em plus\p@}%
\large \bf %
\leavevmode\ignorespaces#1\unskip\nobreak\hskip\skip@
\hb@xt@\rightskip{\hfil\unhbox\z@}\hskip-\rightskip\hskip\z@skip
\par
\nobreak %
\endgroup
}%
\def\tocleft@{\z@}%
\def\tocdim@min{5\p@}%
\def\l@section{%
\l@@sections{}{section
}%
\def\l@f@section{%
\addpenalty{\@secpenalty}%
\addvspace{1.0em plus\p@}%
\bf
}%
\def\l@subsection{%
\l@@sections{section}{subsection
}%
\def\l@subsubsection{%
\l@@sections{subsection}{subsubsection
}%
\def\l@paragraph#1#2{}%
\def\l@subparagraph#1#2{}%
\let\toc@pre\toc@pre@auto
\let\toc@post\toc@post@auto
\def\listoffigures{\print@toc{lof}}%
\def\l@figure{\@dottedtocline{1}{1.5em}{2.3em}}
\def\listoftables{\print@toc{lot}}%
\let\l@table\l@figure
\appdef\class@documenthook{%
\@ifxundefined\raggedcolumn@sw{\@booleantrue\raggedcolumn@sw}{}%
\raggedcolumn@sw{\raggedbottom}{\flushbottom}%
}%
\def\tableft@skip@float{\z@ plus\hsize}%
\def\tabmid@skip@float{\@flushglue}%
\def\tabright@skip@float{\z@ plus\hsize}%
\def\array@row@pre@float{\hline\hline\noalign{\vskip\doublerulesep}}%
\def\array@row@pst@float{\noalign{\vskip\doublerulesep}\hline\hline}%
\def\@makefntext#1{%
\def\baselinestretch{1}%
\reset@font
\footnotesize
\leftskip1em
\parindent1em
\noindent\nobreak\hskip-\leftskip
\hb@xt@\leftskip{%
\Hy@raisedlink{\hyper@anchorstart{footnote@\the\c@footnote}\hyper@anchorend}%
\hss\@makefnmark\
}%
#1%
\par
}%
\prepdef
\section{Introduction}
The current accelerated expansion of the universe can be well explained either by introducing an exotic cosmic fluid with a sufficiently negative pressure dubbed dark energy (DE), or by modifying the
standard theory of gravity on extragalactic scales \citep{Riess1998,Perlmutter1999,Kowalski2008}. A combined analysis of cosmological observations indicates that the current universe is spatially flat and dark energy occupies about $2/3$ of the total energy budget of it \citep{Bennett:2003bz,Spergel:2003cb,Peiris:2003ff}.
The first theoretical candidate of dark energy is the well known cosmological constant $\Lambda$ in which the equation of state (EoS) parameter is equal to $-1$. Although the concordance $\Lambda$ cosmology is consistent with the cosmological observations, it always suffers from two puzzles: the fine-tuning and the cosmic coincidence \citep{Weinberg1989,Padmanabhan2003,Copeland2006}. Alternatively, in the last two decades, numerous other candidates for dark energy with a time evolving energy density have also been proposed in the literatures in order to solve or at least alleviate the above cosmological problems \citep{Caldwell:1997ii,Armendariz2001,Caldwell2002,Elizalde:2004mq}. In these models, the EoS parameter of dark energy varies a function of cosmic redshift. Unfortunately, most of the proposed models for DE are phenomenological and the nature of DE is unknown. Therefore, some cosmologists were motivated to propose a model in which the origin of DE is based on physical principles,
namely it is related with the effects of quantum gravity. In this regard, the first model was proposed by \cite{Li:2004rb} by applying the holographic principle, which is the fundamental principle in the quantum gravity scenario \citep{tHooft1993,Susskind1995}, on cosmological scale to propose a model for DE called the holographic dark energy model. The holographic principle indicates that all physical information inside
in a space-volume can be interpreted as a hologram which corresponds to a theory locating on the boundary of that space \cite{tHooft1993,Susskind1995}. It has been shown that the effective local quantum field theories
greatly over-count degrees of freedom because the entropy $S$, in a box of size $L$ with UV cut-off $\Lambda_c$ scales extensively for an effective quantum field theory, $S\sim L^3\Lambda_c^3$ \cite{tHooft1993,Susskind1995}. Historically, the peculiar thermodynamics of black hole has led Bekenstein to postulate that the maximum entropy in a box of volume $L^3$ behaves non extensively, growing only as the area of the box,
i.e. there is a so-called Bekenstein entropy bound, $S \leq S_{BH} \equiv\pi M_p^2L^2$ \citep{Bekenstein:1973ur,Bekenstein:1974ax}. This non-extensive scaling indicates that quantum field theory breaks down in large volume. To conciliate this
breakdown with the success of local quantum field theory in describing observed particle
phenomenology, Cohen et al.\citep{Cohen1999} suggested a more restrictive bound, the energy bound.
They pointed out that in quantum field theory a short distance (UV) cutoff is related to a
long distance (IR) cutoff due to the limit set by forming a black hole. In other words, If we have a system with size $L$, its total energy should not exceeds the mass of a black hole with the same size, i.e., $L^3\rho _\Lambda\leq L M_{\rm p}^2$, where $\rho _\Lambda$ is the quantum zero-point
energy density caused by UV cutoff.
In cosmological contexts, when the whole of the universe is taken into account, the vacuum
energy related to the holographic principle can be viewed as holographic dark energy (HDE) with
energy density given by
\begin{eqnarray}\label{hde}
\rho_{\rm d}=3 c^2M_{\rm p}^2L^{-2}\;.
\end{eqnarray}
where $c^2$ is a dimensionless numerical parameter and the coefficient $3$ is for convenience. Very often, for the sake of simplicity, the $c^2$ parameter is assumed constant. Depending on length scale $L$, we can define the following types of HDE models \citep[see also][]{Bousso:1999xy,Horava2000,Thomas2002,Shen:2004ck,Enqvist:2004ny,Huang:2004ai,Gao:2007ep,Duran:2010ky,Radicella:2010vf,Sheykhi:2011cn,Mehrabi:2015kta}. ({\it i}) The simplest choice is the Hubble length, i.e., $L = H^{-1}$. In this case the density of DE will be close to the observational data, but the current accelerated expansion of the universe cannot be recovered \citep{Horava2000,Thomas2002,Pavon2005}. ({\it ii}) The other choice is particle horizon. In this case, it is impossible for HDE model to provide an accelerated expansion of the universe \citep{Li:2004rb}. ({\it iii}) Another simple choice for $L$ is the future event horizon \citep{Li:2004rb, Wang:2005ph}. It has been shown that this HDE model accommodates the
late time accelerated expansion \citep{Pavon2005,Zimdahl2007}. Also, the coincidence and the fine-tuning problems are well
alleviated at this length scale \citep{Li:2004rb}. The HDE model defined on the basis of future event horizon has been extensively investigated in recent years \citep{Huang:2004wt,Kao:2005xp,Zhang:2005hs,Wang:2005ph,Chang:2005ph,Zhang:2007sh,Micheletti:2009jy,Xu:2012aw,Zhang:2013mca,Li:2013dha,Zhang:2014ija,Zhang:2015rha}. ({\it iv}) The length scale $L$ can be chosen as the curvature of spacetime, namely the Ricci scalar $R$
\citep{Gao:2007ep,Zhang:2009un}. It has been shown that the Ricci HDE model is consistent
with the observations of supernova type Ia \citep{Zhang:2009un,Easson:2010av}. ({\it v}) Another choice for length scale $L$ is proposed by Granda \& Oliveros (GO) \citep{Granda:2008dk}. GO cut-off defined based on the combination of the Hubble parameter together with its time derivative. In this proposal the late time accelerated phase of expansion is well achieved \citep{Granda:2008dk} and also this model is consistent with supernova type Ia observations \citep{Wang:2016och}. Except cases ({\it i \& ii}) in which the current accelerated phase of expansion is not achieved, other cases ,i.e., cases ({\it iii, iv \& v}), are the most popular types of HDE models studied extensively in the literature. Recently, it has been shown that the HDE models defined based on future even horizon, Ricci scale and GO cut-off, i.e., cases ({\it iii, iv \& v}), are fitted to observational growth rate data in perturbation level as equally as concordance $\Lambda$CDM universe \citep{Akhlaghi:2018knk}. In addition, as was mentioned above, these models are consistent with the observations of SNIa. However, by combination of all observational data in expansion level including SnIa, baryonic acoustic oscillation (BAO), cosmic microwave background (CMB) radiation, Big Bang Nucluesenthsis (BBN) and Hubble expansion data, \cite{Akhlaghi:2018knk} showed that the Ricci and GO HDE models have a strong tension with observations and consequently disfavored \citep[see also][]{Wang:2016och}. This result is unchanged when we combine all observational data in expansion and perturbation levels \citep{Akhlaghi:2018knk}. Moreover, the HDE model defined based on event horizon, i.e., case ({\it iii}), is in a mild tension with expansion and expansion+growth rate data \citep{Akhlaghi:2018knk}. Also from theoretical point of view, since the HDE model ({\it iii}) is defined based on the event horizon length scale, an obvious drawback concerning causality appears in this scenario. The above statements motivate us to consider another possibility for the definition of HDE model. In a more general case, we can consider $c^2$ parameter in Eq.(\ref{hde}) as a slowly varying function of time. In fact there are no strong evidences supporting that $c^2$ parameter should be constant. Adopting the time slowly varying of $c^2$, we can reconsider the first choice, i.e. $H^{-1}$, as an IR cut-off for length scale $L$. Interestingly, it has been shown that the HDE model with slowly varying $c^2$ term defined in Hubble length $H^{-1}$ can simultaneously drive accelerated expansion and solve the coincidence problem \citep[see][]{Pavon2005,Pavon:2006qm,Guberina:2006qh,Radicella:2010vf}. Also the significant advantage of this model is that the density of HDE is not depending on the future or the past evolution of the universe.
In this paper we investigate the HDE models defined in Hubble length with varying $c^2$ term in both expansion and perturbation levels. We first examine the model against the latest observational data in expansion level including those of SNIa, BAO, CMB shift parameter, BBN and expansion Hubble data. In next step, we study the model in perturbation level using the latest growth rate data, i.e., $f(z)\sigma_8$ data. In fact, DE not only accelerates the expansion rate of the universe but also changes the evolution of matter perturbations and consequently the formation epochs of large scale structures of the universe \citep{ArmendarizPicon:2000dh,Tegmark:2003ud,Pace2010}.
Moreover, in the case of dynamical DE models, in which the EoS parameter evolves with cosmic time, the growth of large scale structures are also affected by perturbations of DE \citep{Hu:2004yd,Ballesteros:2008qk,mota3,Basilakos:2010fb,Sapone:2012nh,Basse:2013zua,Pace:2013pea,Batista:2014uoa,Basilakos:2014yda,Pace:2014taa,Nesseris:2014mfa,Malekjani:2015pza,Mehrabi:2015kta,Malekjani:2016edh,Rezaei:2017hon}. In the context of Markov Chain Monte Carlo (MCMC) analysis, we setup an extended formalism in which the background expansion data are joined with the growth rate data to put tight constraints on the parameters of HDE model and evaluate it against combined observational data \citep[for similar studies, see][]{Cooray:2003hd,Corasaniti:2005pq,Blake:2011rj,Nesseris:2011pc,Basilakos:2012uu,Yang:2013hra,mota6,Contreras:2013bol,Chuang:2013hya,Li:2014mua,Mehrabi:2015hva,Basilakos:2016xob,mota7,Rivera:2016zzr,Rezaei:2017yyj}.
The paper is organized as follows. In Sect. \ref{sect:hde}, firstly, we introduce the FRW cosmology in the context of HDE models with varying $c^2$ parameter defined in Hubble length and secondly investigate four different parameterizations of $c^2$ parameter. In Sect.\ref{sect:bg}, we implement the likelihood analysis in the context of MCMC method using the geometrical expansion data to place constraints on the parameters of HDE models and compare the validity of models against observations. Then the evolution of main cosmological quantities in background level according to the dynamics of HDE models is studied.
In Sect.\ref{growth}, the growth of perturbations in our models is studied. We then perform another likelihood analysis using the solely growth rate data to examine the HDE models in perturbation level. Eventually,
we perform an overall likelihood analysis using the background + growth rate data to test the HDE models against the combined cosmological data. In Sect.\ref{conlusion}, the paper is concluded.
\section{HDE with varying $c^2$ term}\label{sect:hde}
Here we investigate the background evolution of HDE with the slowly varying $c^2$ term by taking in to account the current value of Hubble horizon as the IR cutoff \citep[see also][]{Pavon2005,Pavon:2006qm,Guberina:2006qh,Radicella:2010vf}. We are motivated to consider why one should consider the time varying function for $c^2$ parameter in HDE models defined in Hubble length scale. In fact in DE cosmologies, DE dominates the universe at very late times ($\rho_{\rm total}=\rho_{\rm d}$) . So from the Freidmann equation, we have $\rho_{\rm d}(z\rightarrow -1)=3M_{\rm p}^2H^2(z\rightarrow -1)$. On the other hand, from Eq.\ref{hde}, we can say $\rho_{\rm d}(z\rightarrow -1)=3c^2(z\rightarrow -1)M_{\rm p}^2H_0^2$. Hence we obtain the value of $c^2$ parameter at very late times as $c^2=H^2/H_0^2=E^2(z\rightarrow -1)$. It is obvious, at earlier times, $c^2$ cannot be equal to its value at far future ($E^2(z\rightarrow -1)$), because it would not leave room for dark matter and radiation. Therefore, to describe the evolution of energy densities at both matter dominated and dark energy dominated epochs, time varying $c^2$ is inevitable. In order to satisfy the slowly varying condition, we shall use four wellknown parameterizations to describe $c(z)$. These parameterizations are Chevalier-Polarski-Linder (CPL) parameterization (Model 1) \citep{Chevallier2001}, Jassal-Bagla-Padmanabhan (JBP) parameterization (Model2) \citep{Jassal:2004ej}, Wetterrich parameterization (Model 3) \citep{Wetterich:2004pv}, and Ma-Zhang parameterization (Model 4) \citep{Ma:2011nc}, described in terms of redshift as follows:
\begin{eqnarray}\label{par1}
{\rm Molde (1):}~~& c(z)=c_0+c_1\frac{z}{1+z}\;.
\end{eqnarray}
\begin{eqnarray}\label{par2}
{\rm Model (2):}~~& c(z)=c_0+c_1\frac{z}{(1+z)^2}\;.
\end{eqnarray}
\begin{eqnarray}\label{par3}
{\rm Model (3):}~~& c(z)=\frac{c_0}{1+c_1\ln (1+z)}\;.
\end{eqnarray}
\begin{eqnarray}\label{par4}
{\rm Model (4):}~~& c(z)=c_0+c_1\left( \frac{\ln (2+z)}{1+z}-\ln 2\right) \;.
\end{eqnarray}
One can see that in all of the above equations, setting $c_1=0$ leads to the original HDE model with constant $c$ parameter. Notice that the original HDE models in Hubble length cannot explain the acceleration of the universe \citep{Horava2000,Thomas2002,Hsu2004}.
For isotropic and homogeneous spatially flat FRW cosmologies, driven by radiation,
non-relativistic pressure-less matter and DE, the first Friedmann equation can be written as
\begin{eqnarray}\label{frid1}
H^2=\frac{1}{3M^2_{\rm p}}(\rho_{\rm r}+\rho_{\rm m}+\rho_{\rm d})\;,
\end{eqnarray}
where $H\equiv {\dot a}/a$ is the Hubble parameter, $\rho_{\rm r}$, $\rho_{\rm m}$ and $\rho_{\rm d}$ are
the energy densities of radiation, pressureless matter and DE, respectively. Introducing the density parameters $\Omega_{\rm i}$, for radiation, non-relativistic pressureless matter and DE, we obtain
\begin{eqnarray}\label{omega}
\Omega_{\rm r}=\frac{\rho_{\rm r}}{3M^2_{\rm p}H^2}~~& \Omega_{\rm m}=\frac{\rho_{\rm m}}{3M^2_{\rm p}H^2}&~~ \Omega_{\rm d}=\frac{\rho_{\rm d}}{3M^2_{\rm p}H^2}\;.
\end{eqnarray}
By using these parameters, we can write Eq.\ref{frid1} as
\begin{eqnarray}\label{omegasum}
\Omega_{\rm r}+ \Omega_{\rm m}+\Omega_{\rm d}=1\;.
\end{eqnarray}
In the absence of interactions among the three fluids the conservation equations for corresponding
energy densities are given by
\begin{eqnarray}\label{continuity}
&& \dot{\rho_{\rm r}}+4H\rho_{\rm r}=0\;,\label{radiation}\\
&&\dot{\rho_{\rm m}}+3H\rho_{\rm m}=0\;,\label{matter}\\
&&\dot{\rho_{\rm d}}+3H(1+w_{\rm d})\rho_{\rm d}=0\;\label{de},
\end{eqnarray}
where the over-dot denotes a derivative with respect to cosmic time $t$. From Eq.\ref{hde} and using current value of Hubble parameter as IR cutoff the energy density of HDE models can be written as
\begin{eqnarray}\label{GHDE}
\rho_{\rm d}=3 c^2M^2_{\rm p}H^2_0\;.
\end{eqnarray}
Now using Eq.\ref{omega} we can obtain energy density for the HDE as
\begin{eqnarray}\label{omeghde}
\Omega_{\rm d}=\frac{c^2(z)}{E^2(z)}\;.
\end{eqnarray}
where $E(z)=H(z)/H_0$. Replacing Eq.\ref{omeghde} into Eq.\ref{omegasum} and using the evolution functions of $\Omega_{\rm r}$ and $\Omega_{\rm m}$ we arrive
\begin{eqnarray}\label{e20}
\frac{\Omega^0_{\rm m}(1+z)^3}{E^2(z)}+\frac{\Omega^0_{\rm r}(1+z)^4}{E^2(z)}+\frac{c^2(z)}{E^2(z)}=1\;,
\end{eqnarray}
where $\Omega^0_{\rm m}$ and $\Omega^0_{\rm r}$ are the present values of matter and radiation density parameters respectively. Notice that we utilize $\Omega_{\rm r0}=2.469\times 10^{-5}h^{-2}(1.6903)$ to obtain the current value of the energy density of radiation, while $h=H_{0}/100$\citep{Hinshaw:2012aka}. Therefor the Hubble parameter, $E(z)$ takes the form
\begin{eqnarray}\label{e2}
E(z)=\sqrt{\Omega^0_{\rm m}(1+z)^3+\Omega^0_{\rm r}(1+z)^4+c^2(z)}\;.
\end{eqnarray}
If we take the time derivative of Eq.\ref{GHDE}, we obtain
\begin{eqnarray}\label{GHDEdot}
\dot{\rho}_{\rm d}=2\rho_{\rm d}\frac{\dot{c}(z)}{c(z)}\;,
\end{eqnarray}
Inserting Eq.\ref{GHDEdot} into Eq.\ref{de}, we can obtain the EoS parameter of HDE models as
\begin{eqnarray}\label{EOS}
w_{\rm d}=-1-\frac{2}{3}\frac{c'(z)}{c(z)}\;,
\end{eqnarray}
where prime means derivative with respect to $x=\ln{a}$. Clearly, for constant $c$ parameter we have $c'(z)=0$ which leads to $w_{\rm d}=-1$ representing concordance $\Lambda$CDM universe. We see from Eq.\ref{EOS} that the evolution of EoS of HDE models depends on the evolution of $c(z)$. When we have $c(z)c'(z) < 0$, HDE models evolves in quintessence regime, i.e. $w_{\rm d} >-1$ and in other hand when $c(z)c'(z) > 0$, HDE evolves in phantom regime, i.e. $w_{\rm d} <-1$. In what follow, we obtain the EoS parameter of HDE models using four different parameterizations of $c(z)$ introduced in Eqs.(\ref{par1}-\ref{par4}).
\subsection{Model (1), The CPL parameterization}
Based on Eq.\ref{par1},one can see that at early times $(z\rightarrow\infty)$, $c \rightarrow c_0+c_1$ while at the present time we have $z\rightarrow 0$ and thus $c\rightarrow c_0$. This means that parameter $c$ changes slowly from $c_0+c_1$ at early times to $c_0$ at present time. Taking derivatives of Eq. \ref{par1} with respect to $x=\ln a$ we obtain
\begin{eqnarray}\label{par1driv}
c'(z)=-\frac{c_1}{1+z}\;.
\end{eqnarray}
Replacing Eqs.(\ref{par1driv} \& \ref{par1}) in Eq.\ref{EOS}, the EoS parameter of model (1) takes the form
\begin{eqnarray}\label{EOS1}
w_{\rm d}=-1+\frac{2}{3}\frac{c_1}{c_0(1+z)+c_1z}\;,
\end{eqnarray}
\subsection{Model (2), The JBP parameterization}
In this case the parameterization for $c(z)$ is given by Eq.(\ref{par2}). Taking derivative of Eq.\ref{par2}, we have
\begin{eqnarray}\label{par2prim}
c'(z)=-c_1\frac{1-z}{(1+z)^2}\;.
\end{eqnarray}
Substituting Eqs. \ref{par2} and \ref{par2prim} into Eq. \ref{EOS}, we obtain the EoS parameter for model (2) as
\begin{eqnarray}\label{EOS2}
w_{\rm d}=-1+\frac{2}{3}\frac{c_1(1-z)}{c_0(1+z)^2+c_1z}\;.
\end{eqnarray}
In this form, at $z=0$ we have $c=c_0$ and in other hand at early times, when $z\rightarrow\infty$ we have $c=c_0$. But among these two epochs and also in the future, we could have $c\neq c_0$.
\subsection{Model (3), The Wetterich parameterization}
The other parametrization we consider in this work, is Wetterich-type which is given by Eq. (\ref{par3}). In this form, at present time, $z=0$, we have $c=c_0$ while at the early times where $z\rightarrow\infty$ we have $c=0$. Thus based on Eq. \ref{hde}, at the early universe the HDE model did not have considerable role in
the evolution of the universe. Like previous parameterizations in order to obtain $c'(z)$ and $w_{\rm d}$ we should take derivatives of Eq. \ref{par3} with respect to $\ln a$ which leads to
\begin{eqnarray}\label{par3prim}
c'(z)=\frac{c_0c_1}{\left( 1+c_1\ln(1+z)\right)^2}\;.
\end{eqnarray}
Substituting Eqs. \ref{par3} and \ref{par3prim} into Eq. \ref{EOS}, one gets the EoS parameter for model (3) as
\begin{eqnarray}\label{EOS3}
w_{\rm d}=\frac{2}{3}\frac{c_1\left( 1+c_1\ln(1+z)\right) }{c^2_0-\left( 1+c_1\ln(1+z)\right)^2}\;.
\end{eqnarray}
\subsection{Model (4), The Ma-Zhang parameterization}
The last parameterization we choose for $c(z)$ is the Ma-Zhang parameterization which reads Eq. (\ref{par4}). For this parameterization, at the present time we have $c(z)=c_0$, while at the early time where $z\rightarrow\infty$, one can find $c(z)= c_0 -c_1\ln 2$.
If we consider this form for $c(z)$, we could not investigate the future behavior of $c(z)$, because it diverges when $z \rightarrow-1$. Taking derivative of Eq.\ref{par4} we have
\begin{eqnarray}\label{par4prim}
c'(z)=\frac{c_1\ln(2+z)}{(1+z)}-\frac{c_1}{(2+z)}\;.
\end{eqnarray}
Finally, inserting Eqs. \ref{par4} and \ref{par4prim} in Eq. \ref{EOS}, we obtain
\begin{eqnarray}\label{EOS4}
w_{\rm d}=\frac{2}{3}\frac{c_1(1+z)^2\left[ (1+z)+(2+z)\ln(2+z)\right]}{(2+z)\left[c_0(1+z)+c_1\ln(2+z)-c_1(1+z)\ln2 \right]}\times \nonumber \\
\frac{1}{(1+z)^2-\left(c_0(1+z)+c_1\ln(2+z)-c_1(1+z)\ln2\right)^2}\;
\end{eqnarray}
\section{HDE models against expansion observational data}\label{sect:bg}
Bellow, we investigate the HDE models using the above parameterizations for $c(z)$, against the
latest observational data in background expansion level.
Specifically, we perform a statistical analysis using the background expansion data
including those of SnIa (we utilize here the catalog with a binned sample data which
consists of $31$ SNe Ia events from joint light-curve analysis (JLA) \citep{Betoule:2014frx,Escamilla-Rivera:2016aca}), BAO \citep{Beutler:2011hx,Blake:2011en,Padmanabhan:2012hf,
Anderson:2012sa,Hinshaw:2012aka}, the position of the acoustic peak in the Planck CMB data \citep{Shafer:2013pxa}, BBN data point which constrains mostly $\Omega_{\rm b0}$ \citep{Serra:2009yp,Burles:2000zk}, Hubble data from the redshift evolution of cosmic chronometers \citep{Moresco:2012jh,Gaztanaga:2008xz,Blake:2012pj,Anderson:2013zyy}, and the recent data point of Hubble constant $H_0$ \citep{2018ApJ...855..136R}. Concerning the Planck CMB experiment, in our analysis we use the method of distance priors which are proposed to be a compressed likelihood to substitute the full CMB power spectrum analysis \citep[see][]{Bond:1997wr,Efstathiou:1998xx,Wang:2007mza,Chen:2018dbv}. In these studies, CMB data are incorporated by using constraints on parameters$(R,l_a,\Omega_bh^2)$ instead of using the full CMB power spectra. It has been shown that measuring the parameters$(R,l_a,\Omega_bh^2)$ provide an efficient and intuitive summary of CMB data as far as dark energy constraints are concerned. Also in REFF.\citep{Chen:2018dbv}, the authors compared the distance prior method with the full CMB power spectra analysis by constraining some dark energy models and showed that the results from both methods are in full agreement.
Considering these points, the total chi-square $\chi^2_{\rm tot}$ is written as:
\begin{equation}\label{eq:like-tot_chi}
\chi^2_{\rm tot}({\bf p})=\chi^2_{\rm sn_{JLA}}+\chi^2_{\rm bao}+\chi^2_{\rm cmb}+\chi^2_{\rm h}+\chi^2_{\rm bbn}+\chi^2_{\rm H_0}\;,
\end{equation}
where the statistical vector ${\bf p}$ consists of the free parameters we have consider in the Markov chain Monte Carlo (MCMC) analysis.
In this section the above parameters are $\lbrace\Omega_{\rm DM0},\Omega_{\rm b0}, h, c_1\rbrace$ for various HDE models and $\lbrace\Omega_{\rm DM0},\Omega_{\rm b0}, h\rbrace$ for $\Lambda CDM$ model. By setting $z=0$ in Eq.\ref{e20}, it is easy to show that $c_0$ is a dependent parameter which equals $\sqrt{1-\Omega^0_{\rm m}-\Omega^0_{\rm r}}$. Based on the Maximum Likelihood Principle, maximizing ${\cal L}_{\rm tot}({\bf p})$ or equivalently minimizing $\chi^2_{\rm tot}({\bf p})$ leads to the best compatibility between the models under study and the observational data points \citep{Trotta:2017wnx}. Thus we use MCMC analysis to find the best values of free parameters which minimize $\chi^2_{\rm tot}({\bf p})$. To compare HDE models considered in this work, we use the well known information criteria, namely
AIC \citep{Akaike:1974} and BIC \citep{Schwarz:1974}.
In particular, AIC and BIC are given by
\begin{eqnarray}
{\rm AIC} = \chi^2_{\rm min}+2k\;,\nonumber\\
{\rm BIC} = \chi^2_{\rm min}+k\ln N\;.
\end{eqnarray}
where $\chi^2_{\rm min}$ is the minimum value of $\chi^2_{\rm tot}$, $k$ is the number of free parameters and $N$ is the total number of observational data points. In this paper, at the background level we have $N=79$ and $k=3$ for $\Lambda$CDM and $k=4$ for HDE models respectively. Reader can find more details about computing of the $\chi^2(\textbf{p})$ function, the MCMC analysis, the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) in \citep{Mehrabi:2015hva} \citep[see also][]{Basilakos:2009wi,Hinshaw:2012aka,Mehrabi:2015kta,Mehrabi:2016exz,Malekjani:2016edh}. In the current step of our analysis we present the statistical results in the first column of Table\ref{tab:best}. We find that all the HDE models provide the values of $\chi^2_{\rm min}$ close to that of the usual $\Lambda$ cosmology. But since HDE models have one free parameter more than those of $\Lambda$CDM, we should compare their AIC and BIC values with that of in $\Lambda$CDM cosmology. We find $\Delta {\rm AIC}={\rm AIC}-{\rm AIC}_{\rm \Lambda}<2$ which indicates that the HDE models considered in this study are consistent with the expansion data as equally as $\Lambda$CDM cosmology. In the case of BIC, since the $\Delta {\rm BIC}={\rm BIC}-{\rm BIC}_{\rm \Lambda}$ for all of HDE models is smaller than $6$, so there is no strong evidence against HDE models. Note that the HDE models defined based on the event horizon, Ricci and GO length scales have mild and strong tensions with background expansion data \citep[see][]{Akhlaghi:2018knk}. In the left panels of Fig.\ref{fig:AIC} we present the numerical values of $\Delta {\rm AIC}$ (up-left) and $\Delta {\rm BIC}$ (bottom-panel) for different HDE models studied in this work.
\begin{table*}
\centering
\caption{The values of $\chi^2_{\rm min}(AIC, BIC)$ for the different HDE models considered in this work. These results are based on the background expansion data (exp), growth rate data (gr) and combination of them (exp+gr).
The concordance $\Lambda$CDM model is shown for comparison.
}
\begin{tabular}{c c c c c c}
\hline \hline
Model / Data & exp & gr (Homogeneous) & gr (Clustered)& exp+gr (Homogeneous) & exp+gr (Clustered)\\
\hline
Model (1) & $72.78(80.78,90.26)$ & $7.71(17.71,22.16)$ & $7.71(17.71,22.16)$ & $79.80(89.80,102.67)$ & $79.77(89.77,102.64)$\\
\hline
Model (2)& $73.35(81.35,90.83)$&$7.66(17.66,22.11)$ &$7.65(17.65,22.10)$ &$80.68(90.68,103.55)$ &$80.88(90.88,103.75)$\\
\hline
Model (3) & $72.52(80.52,90.00)$&$7.71(17.71,22.16)$ & $7.74(17.74,22.19)$& $79.24(89.24,102.11)$& $79.17(89.17,102.04)$\\
\hline
Model (4) & $72.39(80.39,89.87)$&$7.92(17.92,22.37)$ & $7.80(17.80,22.25)$& $78.87(88.87,101.74)$& $79.15(89.15,102.02)$\\
\hline
$\Lambda$CDM & $73.93(79.93,87.04)$&$7.98(15.98,19.54)$ & $-$& $81.94(89.94,100.24)$& $-$\\
\hline \hline
\end{tabular}\label{tab:best}
\end{table*}
\begin{table*}
\centering
\caption{A summary of the best-fit parameters for the HDE models using expansion data.
}
\begin{tabular}{c c c c c c}
\hline \hline
Model & Model (1) & Model (2) & Model (3)& Model (4) & $\Lambda$CDM\\
\hline
$\Omega_{\rm m}^{(0)}$ & $0.2892^{+0.0047,+0.0088}_{-0.0047,-0.0085}$ & $0.2876^{+0.0049,+0.0089}_{-0.0049,-0.0085}$ & $0.2901^{+0.0048,+0.0090}_{-0.0048,-0.0085}$ & $0.2904^{+0.0048,+0.0086}_{-0.0048,-0.0087}$ & $0.2893^{+0.0045,+0.0085}_{-0.0045,-0.0083}$\\
\hline
$ h $& $0.7057^{+0.0042,+0.0090}_{-0.0046,-0.0085}$& $0.7062^{+0.0052,+0.0095}_{-0.0052,-0.010}$ &$0.7059^{+0.0047,+0.0092}_{-0.0047,-0.0091}$ &$0.7051^{+0.0041,+0.0077}_{-0.0041,-0.0074}$ & $0.7020^{+0.0034,+0.0083}_{-0.0034,-0.0082}$\\
\hline
$c_1 $ & $-0.061^{+0.057,+0.10}_{-0.050,-0.11}$ & $-0.085^{+ 0.073,+0.16}_{- 0.086,-0.15}$ & $0.073^{+0.049,+0.12}_{-0.064,-0.11}$& $0.25^{+0.10,+0.35}_{-0.16,-0.24}$ & --\\
\hline \hline
\end{tabular}\label{tab:bestfitbg}
\end{table*}
\begin{table*}
\centering
\caption{A summary of the best-fit parameters for the homogeneous HDE models using growth rate data.
}
\begin{tabular}{c c c c c c}
\hline \hline
Model & Model (1) & Model (2) & Model (3)& Model (4) & $\Lambda$CDM\\
\hline
$\Omega_{\rm m}^{(0)}$ & $0.280^{+0.037,+0.051}_{-0.019,-0.062}$ & $0.267^{+0.045+0.060}_{-0.025-0.073}$ &$0.267^{+0.046,+0.061}_{-0.024,-0.077}$ & $0.262^{+0.048,+0.065}_{-0.027,-0.079}$ & $0.278^{+0.0094,+0.011}_{-0.0094,-0.011}$\\
\hline
$ h $& $1.197^{+0.031,+2.7}_{-1.2,-1.2}$ & $3.19^{+1.4,+1.6}_{-0.30,-3.1}$ & $0.60^{+0.33,+0.80}_{-0.43,-0.71}$ & $2.6^{+2.7,+2.8}_{-2.3,-2.4}$ & $0.986^{+0.009,+0.012}_{-0.009,-0.012}$\\
\hline
$c_1 $ & $-0.75^{+ 0.56,+1.3}_{- 0.56,-1.2}$ & $0.58^{+0.92,+1.3}_{-0.78,-1.4}$ & $0.14^{+0.20,+0.78}_{-0.56,-0.60}$ & $-1.3^{+1.2,+1.5}_{-2.2,-2.7}$ & -- \\
\hline
$ \sigma_{8}(z=0) $ & $0.693^{+0.040,+0.14}_{-0.068,-0.11}$ & $0.814^{+0.051,+0.15}_{-0.089,-0.12}$ &$0.779^{+0.028,+0.19}_{-0.096,-0.13}$ & $0.813^{+0.028,+0.15}_{-0.076,-0.10}$ & $0.803^{+0.021,+0.034}_{-0.021,-0.034}$\\
\hline \hline
\end{tabular}\label{tab:bestfitgr1}
\end{table*}
\begin{table*}
\centering
\caption{A summary of the best-fit parameters for the clustered HDE models using growth rate data.
}
\begin{tabular}{c c c c c }
\hline \hline
Model & Model (1) & Model (2) & Model (3)& Model (4) \\
\hline
$\Omega_{\rm m}^{(0)}$ & $0.268^{+0.049,+0.062}_{-0.020,-0.085}$& $0.270^{+0.044,+0.060}_{-0.023-0.075}$ & $0.277^{+0.038,+0.053}_{-0.021,-0.064}$ & $0.276^{+0.040,+0.054}_{-0.021,-0.067}$ \\
\hline
$ h $&$0.89^{+0.60,+0.76}_{-0.80,-0.85}$ & $0.97^{+0.65,+1.3}_{-0.92,-0.95}$ & $1.02^{+0.63,+1.3}_{-0.73,-1.2}$ & $1.78^{+0.86,+1.5}_{-0.86,-1.7}$ \\
\hline
$c_1 $ & $0.04^{+1.0,+1.4}_{-0.82,-0.86}$&$0.14^{+0.40,+1.5}_{-0.80,-0.96}$ & $2.9^{+2.3,+2.4}_{-3.0,-3.1}$ & $0.26^{+0.40,+0.70}_{-0.40,-0.71}$ \\
\hline
$ \sigma_{8}(z=0) $ &$0.781^{+0.022,+0.23}_{-0.11,-0.13}$ & $0.780^{+0.034,+0.14}_{-0.073,-0.11}$ & $0.6606^{+0.0067,+0.12}_{-0.056,-0.067}$& $0.755^{+0.023,+0.071}_{-0.040,-0.063}$ \\
\hline \hline
\end{tabular}\label{tab:bestfitgr2}
\end{table*}
\begin{table*}
\centering
\caption{A summary of the best-fit parameters for the homogeneous HDE models using expansion + growth rate data.
}
\begin{tabular}{c c c c c c}
\hline \hline
Model & Model (1) & Model (2) & Model (3)& Model (4) & $\Lambda$CDM\\
\hline
$\Omega_{\rm m}^{(0)}$ & $0.2823^{+0.0082,+0.016}_{-0.0082,-0.016}$ & $0.2822^{+0.0084,+0.016}_{-0.0084,-0.016}$ & $0.2822^{+0.0079,+0.015}_{-0.0079,-0.015}$ &$0.2831^{+0.0079,+0.015}_{-0.0079,-0.016}$ &$0.2861^{+0.0078,+0.014}_{-0.0078,-0.014}$ \\
\hline
$ h $ &$0.7135^{+0.0092,+0.019}_{-0.0092,-0.017}$ & $0.7125^{+0.0074,+0.020}_{-0.011,-0.017}$ & $0.7146^{+0.0092,+0.018}_{-0.0092,-0.017}$ &$0.7132^{+0.0094,+0.018}_{-0.0094,-0.017}$ & $0.7049^{+0.0067,+0.012}_{-0.0067,-0.012}$\\
\hline
$c_1 $ &$-0.086^{+0.066,+0.12}_{-0.053,-0.12}$ & $-0.123^{+0.12,+0.19}_{-0.060,-0.27}$ & $0.098^{+0.066,+0.13}_{-0.066,-0.13}$ & $0.33^{+0.29,+0.38}_{-0.42,-0.48}$& -- \\
\hline
$ \sigma_{8}(z=0) $ & $0.742^{+0.020,+0.041}_{-0.020,-0.037}$ & $0.745^{+0.019,+0.041}_{-0.021,-0.037}$ &$0.741^{+0.020,+0.038}_{-0.020,-0.038}$ & $0.741
^{+0.022,+0.045}_{-0.022,-0.044}$ & $0.749^{+0.022,+0.043}_{-0.022,-0.043}$\\
\hline \hline
\end{tabular}\label{tab:bestfittot1}
\end{table*}
\begin{table*}
\centering
\caption{A summary of the best-fit parameters for the clustered HDE models using expansion + growth rate data.}
\begin{tabular}{c c c c c }
\hline \hline
Model & Model (1) & Model (2) & Model (3)& Model (4) \\
\hline
$\Omega_{\rm m}^{(0)}$ &$0.2828^{+0.0081,+0.016}_{-0.0081,-0.016}$ & $ 0.2844^{+0.0084,+0.016}_{-0.0084,-0.016}$ & $0.2823^{+0.0077,+0.015}_{-0.0077,-0.015}$& $0.2854^{+0.0079,+0.015}_{-0.0079,-0.015}$ \\
\hline
$ h $ &$0.7130^{+0.0079,+0.017}_{-0.0091,-0.016}$ & $0.7083^{+0.0092,+0.018}_{-0.0092,-0.017}$& $0.7144^{+0.0083,+0.016}_{-0.0083,-0.016}$& $0.7068^{+0.0080,+0.015}_{-0.0080,-0.016}$ \\
\hline
$c_1 $ &$-0.084^{+ 0.056,+0.11}_{- 0.056,-0.12}$ & $-0.055^{+0.075,+0.15}_{-0.075,-0.16}$ &$0.095^{+0.052,+0.10}_{-0.052,-0.10}$ & $0.08^{+0.21,+0.34}_{-0.25,-0.31}$ \\
\hline
$ \sigma_{8}(z=0) $ & $0.746^{+0.022,+0.042}_{-0.022,-0.044}$ & $0.748^{+0.021,+0.042}_{-0.021,-0.041}$ & $0.745^{+0.021,+0.040}_{-0.021,-0.042}$ & $0.749^{+0.020,+0.038}_{-0.020,-0.040}$ \\
\hline \hline
\end{tabular}\label{tab:bestfittot2}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=5.6cm]{aicbg.eps}
\includegraphics[width=5.6cm]{aicgr.eps}
\includegraphics[width=5.6cm]{aiccom.eps}
\includegraphics[width=5.6cm]{bicbg.eps}
\includegraphics[width=5.6cm]{bicgr.eps}
\includegraphics[width=5.6cm]{biccom.eps}
\caption{ The values of $\Delta {\rm AIC}$ (upper panels) and $\Delta {\rm BIC}$ (lower panels) obtained from MCMC analysis using expansion data (right panels), growth rate data (midle panels) and all of observational data (left panels) for different HDE models. }
\label{fig:AIC}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=8cm]{wD.eps}
\includegraphics[width=8cm]{deltaE.eps}
\includegraphics[width=8cm]{deltaOmegaD.eps}
\caption{ The redshift evolution of different cosmological quantities, namely dark energy EoS parameter $w_{\rm d}(z)$ ( top panel), $\Delta E(\%)$ (middle panel) and $\Delta \Omega_{\rm d}(\%)$ ( bottom panel).
The different HDE models are characterized by the colors and line-types presented in the inner panels of the figure.}
\label{fig:back}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8cm]{q.eps}
\caption{ The redshift evolution of deceleration parameter $q(z)$ using best fit parameters obtained from MCMC analysis. The different HDE models are characterized by the colors and line-types presented in the inner panel of the figure.}
\label{fig:cz}
\end{figure}
We also present the best fit values and mean $1\sigma$ and $2\sigma$ errors of free parameters for different HDE models in Table (\ref{tab:bestfitbg}). Using these best fit parameters in Fig.\ref{fig:back} we plot the redshift evolution of main cosmological quantities in background level including the EoS parameter, $w_{\rm d}$ (upper panel), the percentage of relative difference of Hubble parameter $\Delta E(\%)=[(E-E_{\rm \Lambda})/E_{\rm \Lambda}]\times 100$ (middle panel) and the percentage of relative difference of DE density parameter $\Delta \Omega_{\rm d}(\%)=[(\Omega_{\rm d}-\Omega_{\rm \Lambda})/\Omega_{\rm \Lambda}]\times 100$ (lower panel) for different HDE models considered in this work. We observe that the present value of EoS of HDE models are smaller than $-1$. This result shows that the EoS parameter of HDE models evolves in the phantom region ($w_{\rm d}<-1$) at low redshifts. Particularly, in the case of Model (2), the EoS parameter of DE evolves from quintessence regime at higher redshifts and enter the phantom regime at $z\sim1$. While for other HDE models considered in this work, the EoS parameter varies slowly in the phantom regime until relatively higher redshifts. Notice that at very high rdshifts, the EoS parameter of Models (1, 2 \& 4) coincide to $w_{\Lambda}=-1$. From the middle panel of Fig.\ref{fig:back}, we observe that the value of $E_{\rm d}(z)$ for HDE models is smaller than $E_{\rm \Lambda}(z)$ at low redshifts. This means that compared to $\Lambda$ cosmology, in the presence of HDE models the universe experiences smaller expansion rate at lower redshifts. The maximum value of relative difference $\Delta E$ for different HDE models varies between $ 0.9-1.2\%$ and occurs at redshifts $z \sim 0.5$. Lastly, from the bottom panel of Fig.\ref{fig:back} we find that the value of $\Omega_{\rm d}$ for all HDE models is smaller than $\Omega_{\rm \Lambda}$.
Now let us switch to the deceleration parameter $q$, which is one of the main cosmological parameters in background level representing the phase expansion of the universe. It is defined in terms of scale factor, as
\begin{eqnarray}
q=-\dfrac{a\ddot{a}}{\dot{a}^2}=-1-\dfrac{\dot{H}}{H^2}\;.\label{qqq}
\end{eqnarray}
Following standard lines, one can obtain \citep{Rezaei:2017yyj}
\begin{eqnarray}
\dfrac{\dot{H}}{H^2}=-\dfrac{3}{2}(1+w_{\rm d}\Omega_{\rm d})\;.\label{qq}
\end{eqnarray}
Replacing Eqs.(\ref{omeghde} \& \ref{EOS}) in Eq.\ref{qq} and inserting the result in Eq.\ref{qqq} we would have the deceleration parameter $q$ in HDE models as follows
\begin{eqnarray}
q=\dfrac{1}{2}-\dfrac{3}{2}\dfrac{c^2(z)}{E^2(z)}-\dfrac{c'(z)c(z)}{E^2(z)}\;.\label{q}
\end{eqnarray}
Now using best-fit values of free parameters we plot the evolution of deceleration parameter as a function of $z$ in Fig.\ref{fig:cz}. We observe that in all cases, including that of $\Lambda$CDM, universe starts to accelerated expansion ($q<0$) at $z_{\rm tr}\sim0.7$. Similar results for transition redshift from early decelerated to current accelerated expansion in the framework of DE and modified gravity theories \citep{Capozziello:2014zda,Capozziello:2015rda,Farooq:2016zwm,Rezaei:2017yyj}. As expected, at high redshifts, the epoch at which DE has no significant effect on the evolution of universe, $q$ tends to $1/2$.
\section{HDE models against Growth rate observational data }\label{growth}
In this section, we investigate the linear growth of matter perturbations in the presence of HDE cosmologies. In order to find the effects of DE on the linear growth of matter fluctuations we have two different scenarios which have been widely investigated in literature \citep{ArmendarizPicon:1999rj,Garriga:1999vw,ArmendarizPicon:2000dh,Abramo2007,Abramo:2008ip,Basilakos:2009mz,Mehrabi:2015hva,Mehrabi:2015kta,Malekjani:2016edh,Rezaei:2017yyj}. We limit our analysis to sub-horizon scales, where the results of Pseudo Newtonian dynamics
are well consistent with those of General Relativity paradigm \citep[see ][]{Abramo2007}. In the first scenario, clustering DE, perturbations of DE can grow same as matter perturbations \citep[see also][]{Abramo:2008ip,Batista:2013oca,Batista:2014uoa}. For this approach, DE perturbations are affected by the negative pressure which implies that the amplitude of DE perturbations is smaller than dark matter fluctuations. In the other scenario, the homogeneous DE case, DE perturbations cannot grow significantly in sub-Hubble scales. In homogeneous DE scenario we have $\delta_{\rm d}\equiv 0$ and only the corresponding non-relativistic matter is allowed to cluster. For both of these approaches, we refer the reader to follow our previous articles \citep{Mehrabi:2015kta,Malekjani:2016edh,Rezaei:2017yyj} in which we have provided the basic differential equations which describe the evolution of matter and DE perturbations. Concerning the initial conditions, we use those provided by \citep{Batista:2013oca} \citep[see also][]{Malekjani:2016edh,Rezaei:2017hon}. In fact using these initial conditions we verify that matter perturbations always stay in the linear regime.
Now we can obtain the evolution of fluctuations ($\delta_{\rm m}, \delta_{\rm d}$) and using them we can calculate the growth rate $f(z)={d\ln{\delta_{\rm m}}}/{d\ln{a}}$ of large scale structures in the presence of different HDE models considered in this work. We also calculate the rms mass variance at $R=8h^{-1}Mpc$ as $\sigma_8(z)=\frac{\delta_{\rm m}(z)}{\delta_{\rm m}(z=0)}\sigma_8(z=0)$ in HDE models to compare the theoretical prediction of $f(z)\sigma_8(z)$ quantity in HDE cosmologies with growth rate observational data measured from redshift space distortion (RSD) of galaxies.
In this way we perform another likelihood analysis using the growth rate ($f(z)\sigma_8$) data. In this step we have $N=18$ (number of data points) and $k=5$ for HDE models (${\bf p}=\lbrace\Omega_{\rm DM0},\Omega_{\rm b0}, h, c_1, \sigma_8\rbrace$). The numerical results form $\chi^2_{\rm min}$ analysis obtained in this step are shown in the second and third columns of Table (\ref{tab:best}) respectively for homogeneous and clustered HDE models. Furthermore, we present the relevant AIC and BIC values in the parentheses. We observe that for all HDE models $\Delta {\rm AIC}<2$ (see also middle-up panel of Fig.\ref{fig:AIC}) which implies, as well as $\Lambda$CDM, all the HDE models are consistent with the growth rate data. On the other hand, for all models we have $\Delta {\rm BIC}<6$ (see also middle-bottom panel of Fig.\ref{fig:AIC}) which supports the results obtained from AIC criteria. Thus using growth rate data only, we find that there is no strong evidence against HDE models compare to standard $\Lambda$CDM cosmology.
Same as previous section we present the best fit values and mean $1\sigma$ and $2\sigma$ errors of free parameters for different HDE models in Tables (\ref{tab:bestfitgr1} \& \ref{tab:bestfitgr2}) respectively for homogeneous and clustered DE scenarios. From the best fit value of free parameters and their relatively large errors obtained in this step, we find that growth rate data cannot solely put strong constraints on DE parameters, especially on parameter $h$ which has the greatest error value.
We now perform an overall likelihood analysis using the expansion data combined with growth rate ones. In this case, the total chi-square $\chi^2_{\rm tot}$ is given by
\begin{equation}\label{eq:like-tot_chi}
\chi^2_{\rm tot}({\bf p})=\chi^2_{\rm sn_{JLA}}+\chi^2_{\rm bao}+\chi^2_{\rm cmb}+\chi^2_{\rm h}+\chi^2_{\rm bbn}+\chi^2_{\rm H_0}+\chi^2_{\rm gr}\;,
\end{equation}
As mentioned before, the vector ${\bf p}$ contains the free parameters of the particular cosmological model. In this step, the relevant parameters are $\lbrace\Omega_{\rm DM0},\Omega_{\rm b0}, h, c_1,\sigma_8\rbrace$, so we have $k=5$ and $N=97$. Same as the previous step, we consider two different homogeneous and clustered HDE scenarios. The results of our analysis for different HDE models, are shown in Table \ref{tab:best} (two last columns) and Tables (\ref{tab:bestfittot1} \& \ref{tab:bestfittot2}) for homogeneous and clustered DE cases respectively.
We observe that performing overall likelihood using both background and growth rate data leads to relatively smaller values for $\Delta {\rm AIC}$ and $\Delta {\rm BIC}$. This result is more significant for Models (1,3 \& 4)where we have $\Delta {\rm AIC} < 0$ . This result indicates that the value of AIC parameter in these HDE models is smaller than the same parameter in concordance $\Lambda$CDM cosmology. However, the difference is smaller than $2$ and in this comparison we can not reject the $\Lambda$CDM model. Comparing the results obtained from combined analysis with those of Sect. \ref{sect:hde} we conclude that in the both of homogeneous and clustered DE scenarios, adding background data to the growth rate ones, leads to smaller $1\sigma$ and $2\sigma$ errors. Also, the best fit values of free parameters for different HDE models are coming closer to those found for $\Lambda$CDM model. Using the best fit values of cosmological parameters presented in Tables (\ref{tab:bestfittot1} \& \ref{tab:bestfittot2}), we plot the fractional difference growth rate $f(z)$ with respect to that of $\Lambda$CMD model ($\Delta f(\%)=100\times [f(z)-f_{\rm \Lambda}(z)]/f_{\rm \Lambda}(z)$) in the upper panel of Fig.\ref{fig:growthfunction}. In the lower panel, the observed $f(z)\sigma_8(z)$ is compared to the theoretical predicted growth rate for different HDE models. We see that all HDE models are fitted to observational growth rate data as well as concordance $\Lambda$CDM universe.
\begin{figure}
\centering
\includegraphics[width=9cm]{delta_f.eps}
\includegraphics[width=9cm]{fs8.eps}
\caption{ The corresponding fractional difference $\Delta f(\%)=100\times [f(z)-f_{\rm \Lambda}(z)]/f_{\rm \Lambda}(z)$ (upper panel) and comparison of the observed and theoretical evolution of the growth rate $f (z)\sigma_8 (z)$ as a function of redshift $z$.
The different HDE models are characterized by the colors and line-types presented
in the inner panels of the figure}.
\label{fig:growthfunction}
\end{figure}
We observe that for all HDE models, the evolution of $\Delta f$ has an maximum at low redshifts. As expected, this feature in the evolution of $\Delta f$ is related to the evolution of $\Delta E$. Indeed, we verify that higher values of the normalized Hubble parameter $E(z)$ correspond to smaller values of the growth rate. Thus, when $\Delta E$ has a minimum we expect the growth rate $\Delta f$ to have a maximum and vice versa (see middle panel of Fig. \ref{fig:back} and upper panel of Fig.\ref{fig:growthfunction}). At high redshifts the relative difference $\Delta f\rightarrow 0$, since at high redshifts, the role of DE becomes negligible and universe is matter dominated, namely $\delta_{\rm m}\propto a$. Thus the growth rate for all HDE models as well as $\Lambda$CDM model tends
to unity and therefore the relative difference tends to zero.
The present value of $\Delta f$, for homogeneous ( clustered) Model (1) is
$\sim [-0.5\%,1.2 \%]$ ( $[-0.8\%,1.2 \%]$). In the case of Model (2) we have
$\Delta f\sim [-0.5\%,1.1 \%]$ and $[-0.2\%,0.5 \%]$ for homogeneous and clustered
DE respectively. For homogeneous (clustered) Model (3) the relative deviation lies in the interval $\sim [-0.5\%,1.6\%]$ ( $[-1.3\%,1.4 \%]$). Finally, for homogeneous ( clustered) Model (4) we obtain $\Delta f\sim [-0.4\%, 1.4 \%]$ ( $[-0.3\%, 0.3 \%]$).
\section{Conclusions}
\label{conlusion}
In this work we investigated the cosmological properties of holographic DE with varying $c^2$ term in the context of Hubble distance, $H_0$ as the IR cutoff. We considered four different well known parameterizations to describe the evolution of $c(z)$ parameter. After putting constraints on the free parameters of the models using background observational data, we studied the behavior of the basic cosmological quantities include $w_d(z), \Omega_d(z), E(z)$ and $q(z)$ in the presence of different HDE models. In the perturbation level we used latest growth rate data and consider homogeneous and clustered DE scenarios. We showed that all of HDE models under study, are well fitted to cosmological data like $\Lambda$CDM model, both at background and perturbation levels.
In particular, our main results may be summarized as follows:
\textit{(i)} Initially, using the latest background observational data
we performed a likelihood analysis for different HDE cosmologies in the context of MCMC method. Based on this analysis we placed constraints on the free parameters of models and we showed that all HDE models are consistent with the background data as equally as concordance $\Lambda$CDM universe. Using the best fit values we plotted the evolution of $w_{\rm d}, \Delta E$ and $\Delta \Omega_{\rm d}$ in Fig.\ref{fig:back}. We found that the present value of $w_{\rm d}$ in all HDE models is in the phantom region. At $z\sim1$, Model (2) crosses the phantom line $w=-1$ while rest of the HDE models remains in the phantom regime until relatively higher redshifts. At early enough times, the EoS parameter of HDE models (1,2 and 4) mimics the constant EoS $w_{\Lambda}=-1$ of $\Lambda$CDM cosmology. We found that the Hubble parameter in HDE cosmologies is $ \sim 0.9-1.2\%$ smaller than the $\Lambda$CDM model at low redshifts. We also showed that in HDE, cosmologies the universe changes its phase from decelerating to accelerating expansion at $z_{\rm tr}\sim0.7$ which in $1\sigma$ error is consistent with observations \citep[see also][]{Capozziello:2014zda,Capozziello:2015rda,Farooq:2016zwm,Rezaei:2017yyj}.
\textit{(ii)}
We performed a statistical analysis using the growth rate data in order to put constraints on free parameters of models. In this step we obtained best fit parameters with relatively large error bars. This means that the growth rate data could not put tight constraints on the cosmological parameters. However the results for $\chi^2_{min}$, $\Delta AIC$ and $\Delta BIC$ showed that HDE models considered in this work are well consistent with recent growth rate data.
\textit{(iii)}
Finally, we performed a joint statistical analysis using the combined expansion and growth rate data in order to compare the models and put constraints on their free parameters. Based on the best fit parameters obtained in this step, we plotted the evolution of the fractional difference $\Delta f(\%)$ for HDE models. The maximum value of $\Delta f(\%)$ occurs at relatively low redshifts, when the role of DE becomes more significant, while the differences among HDE models are negligible at higher redshifts.We found that the absolute value of the difference between AIC (BIC) criteria of HDE models with that of obtained in $\Lambda$CDM cosmology is smaller than $2$ ($4$). Hence we concluded that the HDE models with time varying model parameter defined on Hubble length are well fitted to observational data as equally as concordance $\Lambda$CDM model. In an other word, by current cosmological data there is no even weak evidence against HDE models proposed in this work.
\section{Acknowledgements}
The work of MR has been supported financially by Research Institute for Astronomy \& Astrophysics of Maragha (RIAAM) under research project No. 1/5440-38.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
\begin{adjustwidth}{8mm}{}
\dictentry{read}{ri\textlengthmark d}{verb}{Look at and comprehend the meaning
of (written or printed matter) by interpreting the characters or symbols of which it is composed.}\\
\dictentry{spell}{sp\textepsilon l}{verb}{Write or name the letters that form (a word) in correct sequence.}\\
\null\hfill --- \emph{Oxford Dictionary of English}
\end{adjustwidth}
\noindent Text recognition, namely the problem of reading text in images,
is a classic problem in pattern recognition and computer vision that has enjoyed
continued interest over the years, owing to its many practical applications,
such as recognising printed~\cite{tesseract,Smith07} or handwritten~\cite{Lecun89,bunke04}
documents, or more recently, text in natural images~\cite{Jaderberg16,Mishra12,Neumann12}.
Consequently, many different and increasingly accurate methods have been developed.
Yet, all such methods adopt the same \emph{supervised learning} approach that requires
example images of text annotated with the corresponding strings.
Annotations are expensive because they must be \emph{aligned} to individual training images.
For example, for a text-image of \textim{\textbf{\emph{cats}}}, the corresponding
annotation is the string \texttt{\{c,a,t,s\}}.
A straightforward but tedious approach is to collect such annotations
manually~\cite{Karatzas13,netzer2011reading,Wang10b};
however, since datasets often comprise several million
examples~\cite{Krizhevsky12,Jaderberg14c}, this scales poorly.
Another, perhaps more pragmatic, approach is to engineer highly-sophisticated
synthetic data generators to mimic real images~\cite{Gupta16,Jaderberg14c,Wang12}.
However, this requires developing new generators for each new textual domain,
and could be problematic for special cases such as text in ancient manuscripts.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/method}
\vspace{-5mm}
\caption{Unsupervised text recognition \textmd{can be factored into two sub-problems:
(1)~\emph{visual}:~segmentation at the character-level, followed by
clustering (or recognition) into a known number of classes, and
(2)~\emph{linguistic}:~determining the character identity of these clusters
based on language constraints. Three character classes corresponding to
\texttt{\{a,$\phi$,r\}} are visualised above ($\phi$ stands for \text{\{space\}}).
A given text image is mapped to a sequence of characters using a fully-convolutional network;
the predicted sequences are compared against linguistically valid text-strings
using an \emph{adversarial discriminator}, which guides the mapping of
the characters to the correct identity.
The two networks trained end-to-end jointly, enable text recognition
without any labelled training data.}}
\label{fig:method}
\vspace{-5mm}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figs/synth}
\vspace{-5mm}
\caption{Synthetic text-image samples.
\textmd{A few synthetically generated samples of different lengths,
used in the controlled experiments (see~\cref{sec:exp}).
Our model attains ${\approx} 99\%$ \emph{character accuracy} and ${\approx} 95\%$
\emph{word accuracy} on such samples (\cref{sec:wlen-generalise}),
after training on only \emph{unaligned} image and text examples.}}
\label{fig:synth}
\end{figure*}
We propose instead to develop learning algorithms that can work with \emph{unaligned}
annotations. In this paradigm, images containing text can be extracted \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot from
scanned documents or by mining online image collections~\cite{Jaderberg14}.
Independently, strings containing the same \emph{type} of text (but not exactly
the same text) can be readily harvested from machine readable text corpora~(\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot WMT datasets~\cite{wmt}).
Both steps can be implemented economically in a fully-automated manner,
making such an approach highly desirable.
More specifically, we demonstrate visual text recognition by only providing
examples of valid textual strings, but without requiring them to be aligned to
the example images. In this manner, the method is almost unsupervised, as by only
knowing how to \textbf{spell} correctly, it learns to \textbf{read}.
The method works by learning a predictor that converts images into strings that
\emph{statistically} match the target corpora, implicitly reproducing quantities
such as letter and word frequencies, and n-grams. We show empirically that this seemingly
weak principle is in fact sufficient to drive learning successfully (\cref{sec:exp}).
Text recognition can be factored into two sub-problems (see~\cref{fig:method}):
(1) \emph{visual}:~segmenting the text-image into characters and clustering
the different characters into a known number of distinct classes,
and (2) \emph{linguistic}:~assigning these clusters to the correct character identity.
Indeed, earlier attempts at unsupervised text recognition proposed two-stage solutions
corresponding to the two sub-problems~\cite{Huang07c,Kae09,Knight11,Aldarrab17}.
We address the first problem by exploiting the properties of standard fully-convolutional
networks~\cite{Long15} --- namely locality and translation invariance of the network's filters.
The second problem is equivalent to solving for the correct permutation,
or breaking a 1:1--substitution cipher~\cite{Peleg79}.
The latter problem is NP-hard under a bi-gram language model~\cite{Nuhn13}.
While several solutions like aligning uni-gram (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot frequency matching) or n-gram
statistics~\cite{sutskever16,liu17} have been proposed traditionally for breaking
ciphers~\cite{dooley2013brief}, we instead adopt an adversarial approach~\cite{goodfellow2014generative}.
The result is a compact fully-convolutional sequence (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot multiple words/text-string)
recognition network which is trained against a discriminator in an end-to-end fashion.
The discriminator uses as input only unaligned examples of valid text strings.
We study various factors which affect training convergence, and use
synthetically-generated data for these controlled experiments.
We also show excellent recognition performance on real text images from the
Google1000 dataset~\cite{Google1k}, given \emph{no aligned labelled data}.
The rest of the paper is structured as follows.
\Cref{sec:priorart} reviews related work, \cref{sec:method} describes our
technical approach, \cref{sec:impl} gives the implementation details, \cref{sec:exp} evaluates the method on the aforementioned
data, and \cref{sec:conclude} summarises our findings.
\section{Related Work}\label{sec:priorart}
\parag{Supervised Text Recognition.} Distinct paradigms have emerg- ed and
evolved in text recognition. Traditional character-level methods adopt either
sliding-window classifiers~\cite{Wang11,Mishra12,Yao14,Jaderberg14,Wang12},
or over / under segment into parts~\cite{Bissacco13,Alsharif14,Neumann12},
followed by grouping through classification. Words or sentences are then
inferred using language models~\cite{Wang11,Wang10b,Lee14,
Alsharif14,Jaderberg14,Wang12,Mishra12,Mishra12a,Shi13,Novikova12}. Another
set of methods process a whole word image, modelling it either as retrieval in a
collection of word images from a fixed lexicon~\cite{Goel13,Almazan14,Rodriguez15, Gordo15}
or as learning multiple position dependent classifiers~\cite{Jaderberg14c,Jaderberg15a,Poznanski16}.
Our recognition model is similar to these character-sequence classifiers in that
we train with a fixed number of output characters; but there is an important
difference: we discard their fully-connected (hence, position sensitive)
classifier layers and replace them with fully-convolutional layers.
This drastically reduces the number of model parameters, and lends generalisation
ability to inputs of arbitrary length during inference. More recent methods treat
the text-recognition problem as one of sequence prediction in an encoder-decoder
framework~\cite{Cho14,Sutskever14}. \cite{Su14} adopted this framework first,
using HOG features with Connectionist Temporal Classification (CTC)~\cite{Graves06}
to align the predicted characters with the image features.
\cite{He16DTRN,ShiBY15} replaced HOG features with stronger CNN features,
while~\cite{Lee16,Shi16} have adopted the soft-attention~\cite{Bahdanau15} based
recurrent decoders. Note, all these methods learn from labelled training examples.\\
\parag{Unsupervised Text Recognition.}
Unsupervised methods for text recognition can be classified into two categories.
First, category includes generative models for document images. A prime example
is the \emph{Ocular} system~\cite{berg2013unsupervised}, which jointly models the text content,
as well as the noisy rendering process for historical documents, and infers the
parameters through the EM-algorithm~\cite{Dempster77}, aided by an n-gram language model.
The second category includes methods for automatic decipherment.
Decipherment, is the process of mapping unintelligible symbols
(ciphertext) to known alphabet/language (plaintext).
When the input is visual symbols, it becomes equivalent to text recognition.
Some early works~\cite{Nagy86,Casey86} for optical character recognition (OCR),
indeed model it as such. \cite{Ho00} cluster connected components in binarised
document images and assign them to characters by maximising overlap with
a fixed lexicon of words based on character frequencies and co-occurrence;
\cite{Huang07c,Kae09} also follow the same general approach.
\cite{Aldarrab17} break the Borg cipher, a 17th century 408-pages manuscript,
by also first clustering symbols but decipher using
the noisy-channel framework of~\cite{Knight06} through finite-state-machines.
\cite{Lee02} learn mappings from hidden-states
of an HMM with their transition probabilities initialised with
conditional bi-gram distributions.
\cite{Kozielski14} propose an iterative scheme for
bootstrapping predictions for learning HMMs models, and recognise handwritten text.
However, their approach is limited to (1) word images,
(2) fixed lexicon ($\approx$44K words) to facilitate exhaustive tree search,
whereas, our method is applicable to full \emph{text strings},
does not require a pre-defined lexicon of words.\\
\parag{Unsupervised Learning by Matching Distributions.}
Output Distribution Matching (ODM) which aligns the \emph{distributions}
of predictions with \emph{distributions} of labels was proposed in~\cite{sutskever16}
for ``principled'' unsupervised learning; although similar ideas for learning by
matching statistics have been explored earlier, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot for decipherment (see above),
and also for machine translation~\cite{Snyder10,Ravi08}.
\cite{liu17} extend ODM to sequences, and apply it to OCR with known character
segmentations and pre-trained image features.
In essence, ODM~\cite{sutskever16}, or Empirical-ODM~\cite{liu17} minimises the
KL-divergence cost between the empirical predicted and ground-truth
n-gram distributions.
Our learning principle is the same, however, we do not explicitly
formulate the matching cost, instead learn it online using an adversary~\cite{goodfellow2014generative}.
Recent works~\cite{Artetxe17,Lample17} have demonstrated
unsupervised machine translation using such adversarial losses, however they
closely follow the \emph{CycleGAN} framework~\cite{Zhu17}
which learns a bidirectional mapping between the input and target
domains to enforce bijection. This framework has also been applied recently in
\emph{CipherGAN} to break ciphers~\cite{Gomez18}.
The \emph{CycleGAN} framework learns a bi-directional mapping to enforce strong
correlation between the input and the generated output to avoid the
degenerate failure mode of collapsing to the same output instance regardless of
the input. We, however, dispense with back-translation/reconstruction, and
instead enforce correlation directly in the structure of the recogniser by
limiting the receptive-field of convolutional layers.
Hence, our method is an instantiation of the original (single)
generator--discriminator framework of \emph{GANs}~\cite{goodfellow2014generative}.
However, our method is perhaps the first to decode sequences of discrete symbols
from images using an adversarial framework; these two domains have only been
explored independently in \emph{CycleGAN} and \emph{CipherGAN} respectively.
\section{Method}\label{sec:method}
The aim of text recognition is to predict a sequence of characters given an image
of text. Let the image be a tensor $\mathbf{x} \in \mathcal{X} = \mathbb{R}^{H\times W\times C}$,
where $H$, $W$, $C$ are its height, width, and number of colour channel(s) respectively.
Furthermore, let $\mathbf{y} = (y_1, y_2, \hdots, y_n) \in \mathcal{Y}$ denote the
corresponding character string where each $y_i$ is a character from an alphabet
$\mathcal{A}$ containing $K$ symbols, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $|\mathcal{A}| = K$.
For later convenience, a character $y_i$ is represented as a $K$-dimensional
one-hot vector. Since such vectors are elements of the $K$-dimensional simplex
$\Delta^K$, we set $\mathcal{Y} =(\Delta^K)^n \subset \mathbb{R}^{K\times n}$.
Without loss of generality, we consider strings of a fixed length $n \in \mathbb{N}$.
The objective of \emph{unsupervised} text recognition, then, is to learn the
mapping $\Phi(\mathbf{x}) = \mathbf{y}$, given only \emph{unpaired} examples from the two
domains $\{\mathbf{x}_i\}_{i=1}^N$ where $\mathbf{x}_i \in \mathcal{X}$ and $\{\mathbf{y}_j\}_{j=1}^M$
where $\mathbf{y}_j \in \mathcal{Y}$.
We cast this in an adversarial learning framework based on Goodfellow~\emph{et al}\onedot~\cite{goodfellow2014generative}.
We view the \emph{text recogniser} $\Phi: \mathcal{X} \rightarrow \mathcal{Y}$
as a \emph{conditional generator} of strings $\mathbf{y}$.
The recogniser competes against an adversarial discriminator $D_{\mathcal{Y}}$, which aims to
distinguish between \emph{real} strings $\{\mathbf{y}\}$ and \emph{generated} strings $\{\Phi(\mathbf{x})\}$.
In other words, $\Phi$ and $D_{\mathcal{Y}}$ are optimised simultaneously to play the following
two-player minimax game~\cite{goodfellow2014generative}
$\min_{\Phi}\max_{D_{\mathcal{Y}}} \mathcal{L}(\Phi, D_{\mathcal{Y}})$ where the value function is given by:
\begin{equation*}\label{eq:gan}
\mathcal{L}(\Phi, D_{\mathcal{Y}}) = \mathop{\mathbb{E}}_{\mathbf{y}\thicksim \mathcal{Y}}[\logD_{\mathcal{Y}}(\mathbf{y})] + \mathop{\mathbb{E}}_{\mathbf{x} \thicksim \mathcal{X}}[\log(1 - D_{\mathcal{Y}}(\Phi(\mathbf{x}))].
\end{equation*}
The recogniser learns the visual problem of segmenting characters in
images, and organising them into distinct categories;
while the discriminator, by checking the predicted sequence of characters
against linguistically valid strings,
guides the assignment of these categories into the respective
correct character classes (see \cref{fig:method}).\\
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/real-quotes}
\vspace{-5mm}
\caption{Real text-image samples.
\textmd{Randomly selected samples from a \emph{real} scanned book's test set
along with the ``ground-truth'' (\texttt{`GT'}) and the predicted strings (\texttt{PRED});
punctuations are not modelled.
Our model achieves excellent recognition performance ---
$96.2\%$ \emph{character}, and $84.8\%$ \emph{word} accuracy (see~\cref{sec:real}) without
using any aligned/labelled training examples.
Note the ``ground-truth'' (\texttt{`GT'}) comes from Google's OCR engine output,
hence is not perfect (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot second and third image above).}}
\label{fig:real}
\vspace{-6mm}
\end{figure}
\parag{Grounding.} A potential pitfall is that the string generator network
(or recogniser $\Phi$) may learn to use the input image as a mere source of noise,
using it to generate the correct distribution of strings, without learning to
recognise the string represented in the image.
A useful mapping, instead, must be \emph{grounded}, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot the generated string $\mathbf{y} = \Phi(\mathbf{x})$
should correspond to the text represented in the input image $\mathbf{x}$.
A possible way to encourage grounding is to ensure that the image $\mathbf{x}$ can be recovered
back from the string $\mathbf{y}$.
Both \emph{CycleGAN}~\cite{Zhu17} and \emph{CipherGAN}~\cite{Gomez18} achieve this by
learning a second inverse mapping $\Psi: \mathcal{Y} \rightarrow \mathcal{X}$
from the target domain back to the input and complete the cycle $\Psi(\Phi(\mathbf{x})) \triangleq \mathbf{x}$.
However, learning a mapping from character strings to images is highly ambiguous:
rendering a given string requires sampling the background image, font style,
font colour, geometry of the glyphs, shadows, noise etc.
This ambiguity arises because text recognition requires translating between
two very different \emph{modalities}, \emph{viz}\onedot text and images,
which is much harder than translating within the same modality,
\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot between images in \emph{CycleGAN}~\cite{Zhu17} where only local texture is
modified, or between character strings in \emph{CipherGAN}~\cite{Gomez18},
where the characters are permuted.
Instead of enforcing cycle-consistency, we encourage grounding via
the following two key architectural modifications in the recogniser $\Phi$ (architectural details are given in~\cref{sec:impl}):
\begin{enumerate}
\item \textbf{Prediction Locality.} The character predictor is local, with a
receptive field large enough to contain at most two or three characters in the image.
While this may sound simple, it embodies a powerful constraint.
Namely, such local predictors can generate a string which is globally
consistent only if they correctly transduce the structure of the underlying image.
Otherwise, local predictors may be able to match local text statistics such as
$n$-grams, but would not be able to match global text statistics, such as
forming proper words and sentences~(see also section~6.1 of~\cite{sutskever16} for similar ideas).
Global consistency is enforced by the adversarial discrminiator which has a large
receptive field over the predicted characters (see~\cref{sec:impl}).
\item \textbf{Reduced Stochasticity.}
We also make the generated strings a deterministic function of the input.
We achieve this by removing the noise input from $\Phi$ which is normally used in generator networks.
Furthermore, we do not use dropout regularization~\cite{Srivastava15}.
\end{enumerate}
\parag{Training Objective.} The discriminator $D_{\mathcal{Y}}$ operates in the domain of
\emph{discrete} symbols. While the real symbols are represented as one-hot vectors
or \emph{vertices} $\operatorname{Vert}(\Delta^K)$ of the standard simplex, the
generated symbols are output of a $\operatorname{SoftMax}$ operator over predicted
logits, and hence typically belong to the \emph{interior} of the simplex $\Delta^K$.
This was identified, as the cause for \emph{uninformative discrimination} in
\emph{CipherGAN}~\cite{Gomez18}, where the discriminator distinguishes using this
unimportant difference, rather than soundness of the generated strings.
To mitigate this, we adopt their proposed solution and learn a $d$-dimensional
embedding for each of the $K$ symbols in the alphabet, collectively represented
by a matrix $W \in \mathbb{R}^{K\times d}$.
Furthermore, we replace the log-likelihood loss with a squared difference loss,
as proposed by~\cite{Mao17}. Hence, we optimise the following revised training objective:
\begin{equation*}\label{eq:gan1}
\mathcal{L}(\Phi, D_{\mathcal{Y}}, W) = \mathop{\mathbb{E}}_{\mathbf{y}\thicksim \mathcal{Y}}[D_{\mathcal{Y}}(W^T\mathbf{y})^2] + \mathop{\mathbb{E}}_{\mathbf{x} \thicksim \mathcal{X}}[(1 - D_{\mathcal{Y}}(W^T\Phi(\mathbf{x})))^2].
\end{equation*}
The embeddings $W$ are trained to aid discrimination among symbols by solving
$\min_{\Phi}\max_{D_{\mathcal{Y}},W} \mathcal{L}(\Phi, D_{\mathcal{Y}}, W)$. Learning such embeddings
improved the speed of convergence and final accuracy, as also noted in~\cite{Gomez18},
while using square differences improved numerical stability.\\
\parag{Discussion: Why is this a feasible learning problem?} While learning to
recognise visual symbols without any paired data seems unattainable, the tight
structure of natural language provides sufficient constraints to enable learning.
First, lexically valid text strings form a tiny sub-space of all possible
permutations of symbols, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot there are only ${\approx}13k$ valid English words of
length 7, as opposed to almost 8 billion permutations of the 26 English letters.
Second, the relative frequencies of the characters and their co-occurrence patterns
impose further constraints (see~\cref{sec:learn-order} for correlation between
character-frequency and learning).
These constraints combined with strong correlation between the input image and
the predicted characters are sufficient to drive learning successfully.
\begin{figure*}[t]
\centering
\includegraphics[width=0.97\textwidth]{figs/char-freq}\vspace{-1mm}
\includegraphics[width=\textwidth]{figs/learn-order-wl07}
\caption{Learning order for different characters.
\textmd{The order in which the various characters are learnt is strongly correlated
(Spearman's rank correlation coefficient $\rho = 0.80$, p-value $< 1e{-}5$)
to their frequency in the English language \textbf{[top]}.
Ranking for the learning order is based on the training iteration number
at which the model achieves $50\%$ accuracy for a given character.
\textbf{[bottom]} Rankings from four different training runs are presented to show the variance
--- bright colours signify high variance in rank across runs,
while dark colours correspond to low variance.
The character \texttt{\{g\}} is a curious exception to the trend, as it is sometimes
learnt first (runs $2, 3$);
see~\cref{sec:learn-order} for the reason and further discussion.}}
\label{fig:torder-var}
\end{figure*}
\section{Implementation}\label{sec:impl}
Both, the recogniser ($\Phi$) and the discriminator ($D_{\mathcal{Y}}$) are implemented as fully-convolutional networks~\cite{Long15}.
The recogniser ingests an image of text and produces a sequence of character logits.
The discriminator operates instead on character strings
represented as sequence of character vectors, and produces a scalar discrimination score as output.
The discriminator acts as a \emph{spell-checker}, pointing out the errors
in the generated strings.
We describe their architecture and optimisation details below.\\
\begin{figure}[b]
\centering
\includegraphics[width=\linewidth]{figs/tickertape}
\vspace{-5mm}
\caption{Character sequence representation.
\textmd{Text strings are represented as sequences of $n$ one-hot (for \emph{real} strings)
or $\operatorname{SoftMax}$ normalised logits (for \emph{predictions}) over
$|\mathcal{A}| = K$ character classes.
A sample image and the model's prediction are visualised above
(one-hot \emph{real} strings look similar); here $K = 29$ and $n = 50$.}}
\label{fig:tickertape}
\vspace{-6mm}
\end{figure}
\parag{Recogniser $\Phi$.}
We train our models for strings of a maximum fixed number of characters $=n$.
To this end, the input image dimensions are held fixed at $32 \times (n\cdot 2^4)$ pixels
($= \text{height} \times \text{width}$).
Hence, an image of size $H\times W$ is scaled to
$H'\times W' = 32\times \min(\lceil W\cdot\frac{32}{H}\rceil, n\cdot 2^4)$;
if the $W' < n\cdot 2^4$, it is padded on the right with the mean channel intensity.
The recogniser employs four blocks, each consisting of two convolution layers,
followed by a $2\times 2$ max-pooling layer.
Each convolutional layer comprises of 32 filters of $3\times 3$ dimensions,
and is followed by batch-normalisation~\cite{Ioffe15} and leaky-ReLU activation
(slope$=0.2$)~\cite{Maas13}.
Since max-pooling in each block downsamples the input by a factor of two,
final output dimensions are $2\times n \times D$ (where, $D=32$ is the number of features).
The height is collapsed using average-pooling, and each of the $n$ $D$-dimensional
feature vectors are mapped to $|\mathcal{A}| = K$ dimensional logits
through linear projection, yielding a $K \times n$ dimensional tensor.
Note the receptive field of the final (prediction) layer is small to encourage \emph{locality};
specifically, it is $76$ pixels wide which corresponds to ${\approx}2.5$ characters in the image.
Although we train our recogniser on fixed-length strings, yet it generalises to
different other lengths due to its fully-convolutional architecture (see~\cref{sec:wlen-generalise}).\\
\begin{figure}[t]
\centering
\includegraphics[width=0.97\linewidth]{figs/wlen-learn}
\vspace{-5mm}
\caption{Effect of text length on convergence. \textmd{Training with longer
words leads to faster convergence: the order of convergence
\texttt{\{13,11,9,7\}} mirrors the word lengths (see~\cref{sec:wlen}).
No convergence is seen for models trained on shorter words of length 3 and 5.
For each word-length, the run with largest area-under-curve (AUC)
from eight trials is plotted.}}
\label{fig:wlenconv}
\vspace{-4.5mm}
\end{figure}
\parag{Discriminator (or spell-checker) $D_{\mathcal{Y}}$.}
The input $\mathbf{y}$ to the discriminator are $K\times n$ dimensional tensors of
predicted and real strings containing $n$ characters, represented as logits and
one-hot vectors, respectively (see~\cref{fig:tickertape}).
The predicted logits are first normalised through $\operatorname{SoftMax}$ to
a valid probability distribution over the $K$ characters for each of the $n$ positions.
Next, embeddings $\mathbf{y}_e \in \mathbb{R}^{d\times n}$ for both, the real and predicted strings are obtained:
$\mathbf{y}_e = W^T \mathbf{y} $, where $W\in \mathbb{R}^{K\times d}$ are
the character embeddings ($d = 256$).
We adopt the fully-convolutional \emph{PatchGAN} discriminator architecture~\cite{Isola17,Li16,Liu17ii}, where patches correspond to sub-strings here.
The embedded input $\mathbf{y}_e$ is fed to a stack of five 1D-convolutional layers,
each with~512 filters of size~$5$.
This amounts to a final receptive field of~21 characters which helps to enforce long-range structure.
Each layer is followed by layer-normalisation~\cite{Ba16} and leaky-ReLU ($\text{slope} = 0.2$);
zero padding is used to maintain the size.
The resulting $d \times n$ dimensional output is linearly projected to $1 \times n$,
and average-pooled to obtain the final scalar score $D_{\mathcal{Y}}(\mathbf{y})$.\\
\parag{Optimization details.}
Recogniser, discriminator and character embeddings are trained jointly end-to-end.
The parameters are initialised with Xavier initialization~\cite{Glorot10}.
We use the RMSProp optimizer~\cite{Tieleman12} with a constant learning rate of 0.001.
The two-part discriminator loss objective is multiplied with $\frac{1}{2}$ as in~\cite{Isola17}.
The models are implemented in TensorFlow~\cite{Abadi16}.
\section{Experiments}\label{sec:exp}
Our experiments have two primary goals.
First, is an extensive analysis of various factors which affect the training:
we study ---
(1)~the impact of the length of training sequences on convergence (\cref{sec:wlen});
(2)~the order in which various characters are learnt and its correlation with
their frequencies (\cref{sec:learn-order}),
(3)~generalisation ability of the fully-convolutional recogniser to different
sequence lengths (\cref{sec:wlen-generalise}), and
(4) impact of varying the text corpus on recognition accuracy (\cref{sec:corpus}).
For these experiments we use synthetically generated text data
as it provides fine control over various nuisance factors.
The second objective is to show applicability of the proposed method
to \emph{real} document images (\cref{sec:real}).
We first describe the datasets used in our experiments in \cref{sec:data},
and then present the results.
\subsection{Datasets}\label{sec:data}
\parag{Synthetic data.} We generate synthetic text data to simulate
old printed documents. Synthetic data aids the controlled ablation studies,
as it provides tight control over the various factors,
\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot text content, font style and glyph geometry, background, colours, and other
noise parameters.
We sample the text content from two different sources depending on the
experimental setting --- (1) \emph{words}:
individual English words are sourced from a lexicon of 90K words used in the Hunspell
spell-checker~\cite{hunspell}, and
(2) \emph{lines}: these are full valid English language text strings
extracted from the 2011 news-crawl corpus provided by WMT~\cite{wmt}.
Note, these text sources are used for rendering images, as well as
for providing examples of valid strings to the discriminator.
To limit the variance in position of characters, we use the \texttt{VerilySerifMono}
fixed-width font. The background image data is sampled from the margins of historical
books~\cite{Antonacopoulos13} to simulate various noise effects.
The font colour is sampled from a $k$-means colour model learnt from the same dataset.
The character set consists of the 26 English letters, one space character,
and one additional \texttt{null} class for padding smaller strings,
\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $|\mathcal{A}| = K = 28$. Punctuations, and other symbols in the text are
ignored; lower and upper case letters are mapped to the same class.
Different synthetic datasets are generated as required by the experiments;
the training sets consist of 100k image samples, while the tests set contain 1k samples.
\Cref{fig:synth} visualises some synthetically generated samples.\\
\begin{figure}[t]
\centering
\includegraphics[height=38mm]{figs/cmat-wl05}
\includegraphics[height=38mm]{figs/cmat-wl07}
\vspace{-3mm}
\caption{Confusion matrices for models trained on words of length 5 \& 7.
\textmd{The model trained on length-5 words does not converge to a high accuracy,
while the one trained on length-7 words does (see~\cref{fig:wlenconv} and
\cref{sec:wlen}).
Further, the accuracy for a character depends on its frequency:
the length-5 model \textbf{[left]} confuses most characters,
yet it is quite accurate for the common character \texttt{\{s\}};
while, the length-7 model \textbf{[right]} recognises most characters with
high accuracy, yet it confuses the two least common characters \texttt{\{j,x\}}
(see~\cref{sec:learn-order}).}}
\label{fig:confmat}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/learn-order-curve-wl07}
\vspace{-5mm}
\caption{Temporal learning order.
\textmd{Training iterations for a model trained on length-7 words,
annotated at the steps when it becomes at least $50\%$ accurate for each character.
This curve corresponds to \texttt{run-\#4} in~\cref{fig:torder-var}.
The characters are learnt in the order of their frequencies
(see~\cref{sec:learn-order}).}}
\label{fig:torder}
\vspace{-4mm}
\end{figure}
\parag{Real data.} For testing the validity of our method on real text images,
we use a scanned historical printed book from the
Google1000 dataset~\cite{Google1k}. Specifically, we use the book titled
\emph{Observations on the Nature and Cure of Gout} by James Parkinson~\cite{Parkinson1805}.
For simplicity, we discard cover, title and start-of-chapter pages, and pages
with significant number of footnotes; we only work with the remaining 140 pages
(total 200 pages) which contain text in a relatively uniform font.
Nevertheless, this data is still challenging due to:
(1)~non-fixed-width font which makes character segmentation difficult,
(2)~varying spacing between words due to fully justified alignment,
(3)~varying case (lower/upper) and italics,
(4)~different background colours and textures,
(5)~show-through from the back of the page,
(6)~fading and other noise elements, and
(7)~presence of various punctuations and other symbols.
We use the localisation output of the provided OCR engine output to
segment the pages into lines;
first 300 lines are assigned to the test set, while the remaining
3000 form the training set (no page is shared between the splits).
We use the provided OCR text output for lines in the training split,
as examples of valid text strings for the discriminator.
Note, these strings are sampled uniformly at random during training,
and hence, do not have any direct correspondence to images in the training batch.
The text lines typically consist of ${\approx}50$ characters.
The character-set consists of 26 English letters, one space character,
one unknown \texttt{<UNK>} character, and one \texttt{null} class for padding,
for a total of $|\mathcal{A}| = K = 29$ characters.
We do not distinguish between upper and lower cases; the following
symbols and punctuations: \texttt{, . ? ! ` " * ( )} are suppressed (ignored),
and any other character is mapped to \texttt{<UNK>}.
\Cref{fig:real} visualises some sample text-lines.\\\\
\parag{Metrics.} We measure accuracy at the \emph{character} and \emph{word} levels:
\begin{itemize}
\item\textbf{character accuracy}: this is computed as\\
$1 - \dfrac{1}{N}\sum\limits_{i=1}^N \frac{\operatorname{EditDist}\left(\mathbf{y}_{\text{gt}}^{(i)}, \mathbf{y}_{\text{pred}}^{(i)}\right)}{\operatorname{Length}\left(\mathbf{y}_{\text{gt}}^{(i)}\right)}$,
where $\mathbf{y}_{\text{gt}}^{(i)}$ and $\mathbf{y}_{\text{pred}}^{(i)}$ are the $i^{\text{th}}$ ground-truth and predicted
strings respectively in a dataset containing $N$ strings;
$\operatorname{EditDist}$ is the \emph{character-level} \emph{Levenshtein} distance~\cite{Levenshtein66};
$\operatorname{Length}\left(\mathbf{y}_{\text{gt}}\right)$ is the number of characters in $\mathbf{y}_{\text{gt}}$.
\item \textbf{word accuracy}: computed as \emph{character accuracy} above, but
here the \emph{Levenshtein} distance uses \emph{words} (contiguous strings
demarcated by space) as tokens, and is normalized by number of ground-truth words in $\mathbf{y}_{\text{gt}}$.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/wlen-acc}
\vspace{-5mm}
\caption{Generalisation to different sequence lengths.
\textmd{A model trained on text-strings of length 24,
is evaluated on images containing both shorter and longer strings of lengths ---
$\{3,5,7,9,11,13,32,48\}$.
Word and character accuracies are plotted.
The fully-convolutional architecture of the recognition network
enables significant generalisation to lengths not in the training set,
with small variance in performance (see~\cref{sec:wlen-generalise}).}}
\label{fig:wlen-eval}
\end{figure}
\subsection{Effect of text length on convergence}\label{sec:wlen}
Although earlier works use low-order, namely uni/bi-gram statistics for alignment
~\cite{Lee02,Kozielski14}, higher-order $n$-grams could be more informative.
In this experiment we examine the impact of the \emph{length} of the training
text-sequences on convergence.
We train separate models on synthetic datasets containing \emph{one} word of a given length,
namely --- \texttt{\{3,5,7,9,11,13\}}.
\Cref{fig:wlenconv} tracks \emph{character accuracy} as the training progresses;
due to instabilities in training GANs, we train on each word-length eight times, and
plot the run with the maximum area-under-curve (AUC) (earliest ``take-off'').
Note, models trained on longer words converge faster, achieving ${\approx} 99\%$ \emph{character~accuracy}.
In detail, the model trained on length-13 words converges the fastest,
followed by those trained on 11, 9, and 7 (in order).
No convergence is seen for shorter lengths 3 and 5 (although the accuracy is higher for 5).
This confirms that longer text-sequences impose stronger structural constraints on
the possible outputs, leading to faster convergence.
\Cref{fig:confmat} visualises the confusion matrices for models trained
on lengths 5 and 7; the length-5 model confuses most characters,
whereas the length-7 model recognises most characters almost perfectly.
Note, convergence is independent of the \emph{number of distinct word-instances}
for a given length:
there are more length-5 words (${\approx} 7000$) than length-13 (${\approx} 2500$)
in the lexicon, yet training with length-5 words does not converge.
Further, words of length 7 are the most in number (${\approx}13000$),
yet it converges last.
\begin{figure*}
\centering
\resizebox{0.95\textwidth}{!}{
\includegraphics[height=58mm]{figs/real-good-acc}
\includegraphics[height=58mm]{figs/cmat-real-good}}
\vspace{-3mm}
\caption{Recognising historical printed books.
\textmd{\textbf{[left]} \emph{Character \& word accuracy}
on the test split of the \emph{real} dataset (see~\cref{fig:real})
for both the fully-convolutional and the \emph{skip-RNN} recognition models.
\emph{Skip-RNN} dramatically improves: \emph{character~accuracy} from $85.6\%$ to $96.2\%$,
and \emph{word~accuracy} from $45.0\%$ to $84.8\%$.
\emph{Character accuracy} on the test-set is also visualised against the training iterations.
\textbf{[right]}~Confusion matrix on the test set:
all characters are predicted with high accuracy, except for the low-frequency~\texttt{\{z\}}.
\{$\upvarphi$\} stands for the \texttt{\{space\}} character (see~\cref{sec:real}).}}
\label{fig:real-acc}
\end{figure*}
\subsection{Which character is learnt first?}\label{sec:learn-order}
We examine the dynamics of learning, more specifically, we probe the order
in which the model learns about different symbols --- is there a pattern?
\Cref{fig:torder-var} visualises the order in which models (trained on
synthetic word images of length 7) achieve an accuracy of at least $50\%$ for each character.
We note that this ranking is highly correlated with the frequency of the characters
in the English language --- Spearman's rank correlation coefficient $\rho = 0.80$,~p-value~$< 1e{-}5$.
It further visualises the variance in the ranking of the characters across
multiple runs. The characters at the extremities of the frequency distribution
have low variance --- common characters (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot \texttt{e,s,i,a}) are almost always
learnt first, and the least common characters (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot \texttt{z,x,j,q}) are learnt
last; while characters in the middle, \emph{viz}\onedot \texttt{\{g,p\}} show the highest variance.
The character \texttt{\{g\}} is a curious exception as it is sometimes learnt first.
This is because of $8.54\%$ of the training (length-7) words end in the suffix \texttt{`-ing'}.
Hence, \texttt{\{g\}} appears at the last position quite frequently, and becomes
relatively easy to learn.
\Cref{fig:torder} annotates the training steps at which the accuracy for a character
first reaches $50\%$. After the model becomes confident about the first symbol \texttt{\{a\}},
it quickly learns the other most commons ones; then it slowly learns
the less frequent symbols in the order of their frequencies.
Further, \cref{fig:confmat} visualises the confusion-matrices for models
trained on word-lengths 5 and 7. Again, we can note the dependence on
character frequencies --- even though model for length-5 words does not converge
(see~\cref{sec:wlen}), it is somewhat accurate about the frequent character \texttt{\{s\}},
while the length-7 model is almost perfect at recognising most characters,
yet it confuses two of the least common characters \texttt{\{j,x\}}.
\subsection{Generalisation to different lengths}\label{sec:wlen-generalise}
The fully-convolutional architecture of our recognition network generalises
to images of lengths significantly different from those it was trained on.
To demonstrate this, we train a model on synthetic \emph{text-strings} of
length 24 (containing multiple words), and evaluate on synthetic images of different lengths:
(1) \emph{shorter} single-word images of lengths --- $\{3,5,7,9,11,13\}$, and
(2) \emph{longer} text-string (multiple words) images of lengths --- $\{32, 48\}$.
\Cref{fig:wlen-eval} plots the recognition accuracy against the word lengths.
We note excellent and consistent \emph{character} (${\approx}99\%$)
and \emph{word} accuracies (${\approx} 95\%$) for both, shorter and longer lengths (5 --- 32).
Note, this demonstrates significant generalisation ability, as the model
is never trained on such images.
There is a drop in the character accuracy (${\approx}95\%$) for length-48
text-strings, as the model does not learn a long-range language model.
Performance suffers for words of length 3 due to
image-edges being close in short images, which is not encountered during training with
images of long words.
\begin{table}[b]
\begin{tabular}{@{}lccc@{}}
\toprule
\multirow{2}{*}{corpus $\rightarrow$} & \multirow{2}{*}{(1) WMT} & (2) WMT & \multirow{2}{*}{(3) War \& Peace} \\
& & no overlap & \\ \midrule
char & 98.98 & 99.13 & 98.43 \\
word & 95.33 & 96.08 & 92.53 \\ \bottomrule
\end{tabular}
\vspace{3mm}
\caption{Effect of varying the text corpus on recognition accuracy.
\textmd{Text-strings are sampled from three increasingly distant text corpora.
This has a small adverse effect on recognition \emph{word} and \emph{character} accuracies
(in $\%$) (see~\cref{sec:corpus}).}}
\label{tab:corpus}
\end{table}
\subsection{Varying the text corpus}\label{sec:corpus}
We examine the impact of varying the text corpus from which samples of text strings
are obtained, on the recognition accuracy.
We examine the following three different sources for sampling the strings.
The synthetic text-images are held constant across the three settings, and
contain up to 24 characters with the text content in them sampled from
WMT newscrawl (as before).
(1)~same strings as in text-images but randomly sampled for each batch,
(2)~strings from the same corpus (WMT newscrawl) but with no overlap with
text-image strings, and
(3)~strings sampled from a very different corpus, namely, Tolstoy's \emph{War and Peace}.
\Cref{tab:corpus} summarizes the \emph{character} and \emph{word} accuracies.
Training with the completely unrelated lexicon ($\#3$) does have a small adverse effect
(\emph{word} accuracy drops to $92.53\%$ from $\approx 95\%$), while using a related
lexicon $(\#2)$ does not have such an effect.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{figs/skip_rnn}
\vspace{-3mm}
\caption{Skip-RNN architecture for real text images.
\textmd{Non-uniform spacing and non-fixed width fonts pose a significant
challenge to the fully-convolutional recogniser.
We augment the recognition network with a \emph{skip-RNN}, which acts on the
convolutional features, and predicts residual updates to the inputs
(the residual predictions are \emph{added} to the inputs).
This improves the \emph{word accuracy} from ${\approx}45\%$ to ${\approx}85\%$
(see~\cref{fig:real-acc}).}}
\label{fig:skip-rnn}
\end{figure}
\subsection{Recognising a historical printed book}\label{sec:real}
Finally, we apply our model to \emph{real} text-line images extracted
from a historical printed book (see \cref{sec:data} for dataset details).
As noted in \cref{sec:data}, non-fixed width fonts and fully-justified text alignment introduce
non-uniform spacing between characters and words. This poses a significant
challenge to the fully-convolutional recogniser, making segmentation of the
text-image into individual characters difficult (see \cref{fig:real} for example images).
Hence, we augment the penultimate layer of the fully-convolutional recogniser
with a \emph{skip-RNN} --- a \emph{uni}-directional (left to right)
RNN (256-dimensional LSTM) with a residual skip-connection~\cite{He15} (see~\cref{fig:skip-rnn}).
The RNN lends pliability to the convolutional features, thereby aids character segmentation.
All other model parameters are as those used for the synthetic data experiments
(see \cref{sec:impl}), except: (1)~the discriminator
filter size is increased from 5 to 11, and (2)~number of layers is doubled to 8
to exploit the long-term structure in the much longer text strings
(${\approx}50$ characters each).
\Cref{fig:real-acc} visualises the \emph{character} and \emph{word} accuracies for
both the fully-convolutional and \emph{skip-RNN} recognition models.
\emph{Skip-RNN} dramatically improves the recognition performance:
\emph{word~accuracy} improves from $45.0\%$ to $84.8\%$, while
\emph{character~accuracy} improves from $85.6\%$ to $96.2\%$.
\Cref{fig:real} visualises randomly selected examples from the test set
and shows the model's predictions;
the predictions are comparable to the ``ground-truth'' annotations
obtained from Google's OCR engine.
\Cref{fig:real-acc} also visualises the confusion matrix for the character classes:
all characters are predicted with high accuracy, except for the low-frequency~\texttt{\{z\}}.
Full page read-outs from our model are visualised in \cref{sec:appendix}.
\section{Conclusion}\label{sec:conclude}
We have developed a method for training a text recognition network
using only \emph{unaligned} examples of text-images and valid text strings.
We have presented detailed analysis for various aspects of the proposed method.
We have established ---
(1) positive correlation between the length of the input text and convergence rates;
(2) the order in which the characters are learnt is strongly dependent on their
relative frequencies in the text;
(3) the generalisation ability of our method to input images of different lengths,
specifically our recognition model trained on strings of length 24 generalises
to both much shorter and longer strings (3 -- 48) without drastic degradation in performance;
(4) the effect of varying the text corpus used as the source of valid sentences
on the recognition accuracy.
Finally, we have shown successful recognition on real text images, without using
any labelled supervisory data.
These results open up a new and promising direction for training sequence recognition models
for structured domains (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot language) given no labelled training data.
The proposed method is applicable not just to text images,
but other modalities as well, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot speech and gestures.
\begin{acks}
We thank Iasonas Kokkinos, Triantafyllos Afouras, and Weidi Xie for insightful discussions,
and anonymous reviewers for their detailed feedback.
Financial support was provided by \grantsponsor{AIMS}{UK EPSRC AIMS CDT}{} \grantnum{AIMS}{EP/L015987/2},
\grantsponsor{SEEBIBYTE}{EPSRC Seebibyte Grant}{} \grantnum{SEEBIBYTE}{EP/M013774/1}, and \grantsponsor{CLARENDON}{Clarendon Fund scholarship}{}.
\end{acks}
\bibliographystyle{abbrvnat}
|
\section{Introduction}
\label{sec:introduction}
In order to analyse complex \gls*{EM} problems, two main routes can be identified.
First, a numerical approach for solving the full set of \textsc{Maxwell}\xspace's field equations offers the advantage of capturing all relevant effects but can be very demanding in terms of computational resources.
One of the first methods for \gls*{EM} analysis was the \gls*{FDTD} scheme proposed by Yee~\cite{Yee_1966aa} in the 1960s.
This method became popular and is still a standard approach for high-frequency \gls*{EM} simulations~\cite{Taflove_1998aa}.
About ten years later, in the 1970s, Weiland~\cite{Weiland_1977aa} introduced the \gls*{FIT} as an extension to the \gls*{FDTD} scheme using integral unknowns and allowing for non-Cartesian and unstructured grids~\cite{Rienen_1985aa}.
Additionally, later developments of the method led to the usage of integral unknowns to obtain an exact implementation of \textsc{Maxwell}\xspace's equations~\cite{Weiland_1996aa}.
The method proved to be more efficient in terms of memory requirements and computing time.
While the \gls*{FEM} was mainly used in structural mechanics for many years, the introduction of edge elements made it applicable for \gls*{EM} problems as well~\cite{Bossavit_1988ab}.
Secondly, one may use compact models to obtain an efficient representation of a complex system.
For example, electrical engineers employ circuits to model and describe the behaviour of complex devices.
Nevertheless, the generation of such compact models can be a tedious task requiring empirical know-how to apply the appropriate approximations.
To generate such circuit models, different techniques are available.
For instance, mathematical analysis and physical insight allows to construct circuits representing the problem at hand as is done by Choi et al~\cite{Choi_2000aa}.
A circuit's topology and the required component values can also be obtained from experimental results as is done by Moumouni and Baker~\cite{Moumouni_2015aa} and other groups.
Another approach, e.g. followed by Codecasa et al~\cite{Codecasa_2016aa}, Eller~\cite{Eller_2017ae} and Wittig et al~\cite{Wittig_2002aa}, is to apply \gls*{MOR} techniques directly to the field formulation of the problem from which a circuit description can be found more easily~\cite{Codecasa_2016aa}.
However, the resulting elements may have non-physical values.
Now, if one is able to represent an \gls*{EM} problem by means of an electric circuit, one can use any circuit simulator to obtain the solution.
The most popular representatives are SPICE\xspace programs that were introduced in the 1970s~\cite{Nagel_1973aa} and are still used as a synonym for circuit solvers.
Later, extensions to deal with \gls*{ET} simulations using circuits were developed within the SPICE\xspace framework~\cite{Vogelsong_1989aa,Hefner_1993aa}.
The mathematical tool employed by most SPICE\xspace-like programs is still the \gls*{MNA} presented by Ho et al~\cite{Ho_1975aa} in the 1970s.
First approaches to combine numerical field simulation with circuit elements were proposed in the 1990s.
These were based on the \gls*{FDTD} scheme~\cite{Sui_1992aa,Tsuei_1993aa,Piket-May_1994aa,Thomas_1994aa} because of the topological similarities between circuits and the finite difference scheme.
Later, the insertion of lumped elements into \gls*{FEM} schemes in time or frequency domain was developed by Guillouard et al~\cite{Guillouard_1996aa,Guillouard_1999aa}.
These approaches became known as field-circuit coupling and have evolved into an important research topic~\cite{Kettunen_2001aa,Benderskaya_2004aa,Schops_2013aa,Schops_2011ac}.
One possible field-circuit coupling method is the direct insertion of lumped circuit elements into the field model by applying these to edges of the discretisation grid~\cite{Witting_1997aa}.
To extract compact circuits representing \gls*{EM} problems in a generic way, numerous approaches can be found in the literature.
In the \gls*{PEEC} method presented by Ruehli~\cite{Ruehli_1974aa,Ruehli_2015aa}, equivalent circuits are derived from integral equations, allowing for a combined \gls*{EM}-circuit solution both in frequency and time domain.
However, \gls*{PEEC} requires empirical approximations in addition to the applied discretisation.
For quasistatic approximations, automated circuit generation based on the boundary element method was presented by Milsom~\cite{Milsom_1999aa}.
The method presented therein yields circuits whose size depends on the electrical dimensions of the problem.
Many methods for application-specific circuit extraction based on device or system responses are also available~\cite{Antonini_2003aa,Russer_1994aa,Choi_2002aa}.
A methodology for the generation of equivalent circuits based on the semi-discrete \textsc{Maxwell}\xspace's field equations was also proposed by Ramachadran et al~\cite{Ramachandran_2008aa}.
Therein, \textsc{Yee}\xspace's discretisation scheme is employed to cast \textsc{Maxwell}\xspace's curl equations as the concatenation of interacting fundamental circuits in which voltages and currents model the sought electric and magnetic fields.
In this manner, circuit stamps are required for both primal and dual edges.
Their interaction is organised by \glspl*{VCVS} and \glspl*{CCCS}, respectively.
In the design of electronic devices, enhancing their functionalities is always of high interest.
The according volume shrinking may give rise to high power densities that can lead to thermal issues.
As an example, the introduction of stacked 3D chips intensifies the heat issue since the heat can be trapped between the stacked layers.
Therefore, \gls*{ET} modelling is of great importance for device engineers.
To handle the \gls*{ET} coupling in a circuit simulation framework, two general approaches are mainly available:
the relaxation method which consists of the iterative coupling of an electric circuit with an external thermal-only field simulator~\cite{Chvala_2014aa,Wunsche_1997aa,Van-Petegem_1994aa}, and the monolithic approach which consists of the direct coupling of the electric circuit with a thermal circuit~\cite{Simpson_2014ab}.
The latter allows to run the simulation directly on the full \gls*{ET} circuit without any software package and thereby avoids the weak coupling between solvers.
To extract \gls*{ET} circuits from a given 3D problem, various methods were proposed.
Some of them were based on existing \gls*{EM} simulation methods and have been extended in functionality to also cover the \gls*{ET} case, as done by Lombardi et al~\cite{Lombardi_2017aa} for the \gls*{PEEC}.
Generating compact models from the calculated or measured response function is another popular approach and has been followed by Evans et al and Bernardoni et al~\cite{Evans_2013aa,Bernardoni_2018aa}.
For thermal problems, methods that derive an equivalent circuit directly from the mesh can also be found~\cite{Culpo_2010aa,Wilkerson_2004aa}, but none of them accounts for the \gls*{ET} coupling.
Karagol and Bikdash~\cite{Karagol_2010aa} presented an approximate representation obtained from a graph-partitioning algorithm of an \gls*{FEM} mesh resulting in a medium-sized \gls*{ET} circuit.
A lumped-element representation of every \gls*{FEM} element has also been proposed for \gls*{ET} simulations~\cite{Hsu_1996aa}.
For an exact representation of a semi-discretised 3D \gls*{ET} field problem, Casper et al~\cite{Casper_2016ab} developed an automatic netlist generation method based on the \gls*{FIT}.
In this paper, we present a method to automatically generate netlists representing general 3D \gls*{ET} and \gls*{EM} coupled field problems.
In our approach, neighbouring cells in the primal grid interact via the parallel connection of circuit elements as illustrated in Figure~\ref{fig:lumpedElementsOnCube}.
Thus, the size of the resulting circuit depends on the geometrical size of the problem and the fineness of the discretisation grid.
This allows to use any available circuit simulator.
Hence, the need for a dedicated field solver with a custom time integrator is alleviated.
To accomplish this, we employ the \gls*{FIT} for discretising the relevant continuous field equations. In contrast to \textsc{Yee}\xspace's finite difference scheme employed by Ramachandran et al~\cite{Ramachandran_2008aa}, \gls*{FIT} is a structure-preserving discretisation strategy which does not require further approximations in dealing with the field and material quantities.
Moreover, the concept of integral quantities used in the \gls*{FIT} translates naturally into the framework of circuit descriptions.
In this manner, we obtain an exact grid representation of the field equations that we map transparently into circuit stamps.
In these stamps, which are associated only with primal edges, lumped elements are directly taken from the entries of the material matrices.
These entries are properly integrated constitutive parameters.
In Figure~\ref{fig:methodology}, we summarise our approach.
In addition to previous work on \gls*{ET} problems~\cite{Casper_2016ab}, we also present a more elaborated description and implementation of boundary conditions and excitations.
We also point out that the methodology presented herein allows for straightforward field-circuit coupling and is especially useful for an accurate representation of small devices in a larger circuit.
In order to simulate wave propagation problems on a finite computational domain by means of circuit simulation, we present \glspl*{ABC}~\cite{Engquist_1977aa,Mur_1981ab}.
\begin{figure}[t]
\centering
\subfloat[\label{fig:lumpedElementsOnCube}]{\includegraphics{./figs/fig1a}}
\hspace{2em}
\subfloat[\label{fig:methodology}]{\includegraphics{./figs/fig1b}}
\caption{(a) One grid cell is represented by a circuit node. Connections to neighbouring cells are carried out via lumped elements. (b) The left branch of the diagram shows the standard solution approach by using a field solver. The right branch illustrates the approach described herein to generate a netlist that is then fed to a circuit simulator.}
\end{figure}
The outline of this paper is as follows.
In Section~\ref{sec:FIT}, we provide the required basics of the \gls*{FIT}.
The fundamentals of the \gls*{MNA} are summarised in Section~\ref{sec:MNA}.
Then, the main part of the paper starts with the circuit representation of \gls*{ET} field problems in Section~\ref{sec:circuitsET}.
How to extract circuit stamps for \gls*{EM} field problems is presented in Section~\ref{sec:circuitsEM}.
Finally, we show numerical examples in Section~\ref{sec:numerics} and conclude the paper in Section~\ref{sec:conclusion}.
\section{Continuous Thermal and Electromagnetic Formulations}
\label{subsec:EMproblem}
Let us consider a domain $\mathcal{D}$ with boundary $\partial\mathcal{D}$ and characterised by the constitutive parameters $\lbrace\varepsilon,\nu,\sigma\rbrace$, where $\varepsilon$ is the electric permittivity, $\nu$ is the magnetic reluctivity and $\sigma$ is the electric conductivity.
For every facet $\ensuremath{A}$, every volume $\ensuremath{V}$ and with impressed electric sources $\ensuremath{\mathbf{J}_{\mathrm{i}}}$, the \gls*{EM} field $\lbrace\ensuremath{\mathbf{E}},\ensuremath{\mathbf{H}}\rbrace$ in $\mathcal{D}$ is given by \textsc{Maxwell}\xspace's equations
\begin{subequations}
\begin{align}
-\oint\limits_{\partial\ensuremath{A}}\ensuremath{\mathbf{E}}\cdot\mathrm{d}\ensuremath{\mathbf{\ensuremath{L}}}&=
\frac{\mathrm{d}}{\mathrm{d}t}\int_{\ensuremath{A}}\ensuremath{\mathbf{B}} \cdot\mathrm{d}\ensuremath{\mathbf{S}},\\
\oint\limits_{\partial\ensuremath{A}}\ensuremath{\mathbf{H}}\cdot\mathrm{d}\ensuremath{\mathbf{\ensuremath{L}}}&=
\frac{\mathrm{d}}{\mathrm{d}t}\int_{\ensuremath{A}}\ensuremath{\mathbf{D}}+\ensuremath{\mathbf{J}_{\mathrm{c}}}+\ensuremath{\mathbf{J}_{\mathrm{i}}}\cdot\mathrm{d}\ensuremath{\mathbf{S}},\\
\int_{\partial\ensuremath{V}}\ensuremath{\mathbf{D}} \cdot\mathrm{d}\ensuremath{\mathbf{S}}&=\int_{\ensuremath{V}}\varrho\,\mathrm{d} V,\\
\int_{\partial\ensuremath{V}}\ensuremath{\mathbf{B}} \cdot\mathrm{d}\ensuremath{\mathbf{S}}&=0.
\end{align}
\label{eq:MaxwellContInt}\end{subequations}
We call henceforth~\eqref{eq:MaxwellContInt} the E-H formulation for conciseness.
Above, $\ensuremath{\mathbf{D}}$ and $\ensuremath{\mathbf{B}}$ are the electric and magnetic flux density, respectively, $\ensuremath{\mathbf{J}_{\mathrm{c}}}$ is the electric conduction current and $\varrho$ is the electric charge density.
To guarantee the uniqueness of the solution, \textsc{Maxwell}\xspace's equations are supplemented with the constitutive relations, viz.
\begin{alignat*}{4}
\ensuremath{\mathbf{D}} &= \varepsilon\ensuremath{\mathbf{E}},\qquad
& \ensuremath{\mathbf{H}} &= \nu\ensuremath{\mathbf{B}},\qquad
& \ensuremath{\mathbf{J}_{\mathrm{c}}} &= \sigma\ensuremath{\mathbf{E}},
\end{alignat*}
with suitable initial and \glspl*{BC} on $\partial\mathcal{D}$.
We also deem it convenient to obtain \textsc{Maxwell}\xspace's equations involving the auxiliary magnetic vector potential $\ensuremath{\mathbf{A}}$.
To this end, we recall that $\ensuremath{\mathbf{E}}=-\nabla\varphi-\partial\ensuremath{\mathbf{A}}/\partial t$ and $\nabla \cdot(\sigma_{\text{g}}\ensuremath{\mathbf{A}})=f$, with an auxiliary scalar potential $\varphi$, a gauging material parameter $\sigma_{\text{g}}$ and an arbitrary \emph{scalar} gauging function $f$.
Substitution of these definitions in \eqref{eq:MaxwellContInt} and applying \textsc{Stokes}\xspace' theorem yields
\begin{subequations}
\begin{align}
-\int_{\ensuremath{A}}\nabla \times\ensuremath{\mathbf{E}}\cdot\mathrm{d}\ensuremath{\mathbf{S}}&=\frac{\mathrm{d}}{\mathrm{d}t}\int_{\ensuremath{A}}\nabla \times\ensuremath{\mathbf{A}}\cdot\mathrm{d}\ensuremath{\mathbf{S}},\\
\int_{\ensuremath{A}}\nabla \times\left(\nu\nabla \times\ensuremath{\mathbf{A}}\right)\cdot\mathrm{d}\ensuremath{\mathbf{S}}&=
\frac{\mathrm{d}}{\mathrm{d}t}\int_{\ensuremath{A}}\ensuremath{\mathbf{D}}+\ensuremath{\mathbf{J}_{\mathrm{c}}}+\ensuremath{\mathbf{J}_{\mathrm{i}}}\cdot\mathrm{d}\ensuremath{\mathbf{S}},\\
\int_{\partial\ensuremath{V}}\ensuremath{\mathbf{D}}\cdot\mathrm{d}\ensuremath{\mathbf{S}}&=\int_{\ensuremath{V}}\varrho\,\mathrm{d} V,\\
\int_{\partial\ensuremath{V}}\ensuremath{\mathbf{B}}\cdot\mathrm{d}\ensuremath{\mathbf{S}}&=0,
\end{align}
\label{eq:FaradayAmpere}\end{subequations}
which we refer to as the E-A formulation.
Whenever conducting materials are involved, electric currents result in \textsc{Joule}\xspace losses $Q_{\text{J}}=\sigma\left(\nabla\varphi\right)^{2}$ that enter as source term into the heat equation which is given by
\begin{equation}
\int_{\ensuremath{V}}\left(\rho c\dot{T}-Q_{\text{J}}\right)\,\mathrm{d} V
=\int_{\partial\ensuremath{V}}\left(\lambda\nabla T-\ensuremath{\mathbf{q}_{\mathrm{i}}}\right)\cdot\mathrm{d}\ensuremath{\mathbf{S}},
\label{eq:thermalContInt}
\end{equation}
where $\rho c$ is the volumetric heat capacity, $T$ is the temperature, $\lambda$ is the thermal conductivity and $\ensuremath{\mathbf{q}_{\mathrm{i}}}$ represent any impressed thermal flux density.
In general, when thermal effects are considered, all constitutive parameters are also a function of the temperature.
To perform circuit extraction, we shall consider the above \gls*{ET} and \gls*{EM} formulations separately as described in Sections~\ref{sec:circuitsET} and \ref{sec:circuitsEM}, respectively.
\section{Discretising the Thermal and Electromagnetic Formulations}
\label{sec:FIT}
We discretise the domain $\mathcal{D}$ into a pair of orthogonal grids given by the primal grid $G$ and its dual $\protect\widetilde{G}$.
The grid $G$ consists of primal points $\ensuremath{P}_{i},\,i=1,\dots,N_{\text{P}}$, primal edges (lines) $\ensuremath{L}_{n},\,n=1,\ldots,N_{\text{E}}$, primal facets (areas) $\ensuremath{A}_{n},\,n=1,\ldots,N_{\text{F}}$, and primal volumes $\ensuremath{V}_{i},\,i=1,\ldots,N_{\text{V}}$.
Similarly, the grid $\protect\widetilde{G}$ consists of dual points $\ensuremath{\protect\widetilde{P}}_{i},\,i=1,\dots,\protect\widetilde{N}_{\text{P}}$, dual edges $\ensuremath{\protect\widetilde{L}}_{n},\,n=1,\ldots,\protect\widetilde{N}_{\text{E}}$, dual facets $\ensuremath{\protect\widetilde{A}}_{n},\,n=1,\ldots,\protect\widetilde{N}_{\text{F}}$, and dual volumes $\ensuremath{\protect\widetilde{V}}_{i},\,i=1,\ldots,\protect\widetilde{N}_{\text{V}}$.
The grids $G$ and $\protect\widetilde{G}$ are dual to each other in the sense that a primal edge $\ensuremath{L}_{n}$ intersects a dual facet $\ensuremath{\protect\widetilde{A}}_{n}$ and a primal point $\ensuremath{P}_{i}$ is located inside a dual volume $\ensuremath{\protect\widetilde{V}}_{i}$ and vice versa.
For a regular hexahedral grid, the grid staggering is depicted in Figure~\ref{fig:gridStaggering}.
Due to this duality, the number of primal and dual grid objects fulfils
\begin{equation*}
N_{\text{E}}=\protect\widetilde{N}_{\text{F}},\quad
N_{\text{F}}=\protect\widetilde{N}_{\text{E}},\quad
N_{\text{V}}=\protect\widetilde{N}_{\text{P}},\quad
N_{\text{P}}=\protect\widetilde{N}_{\text{V}}.
\end{equation*}
As mentioned above, we write $\ensuremath{L}_{n}$ for the $n$-th edge of the primal grid and we emphasise the duality of edges and facets by using the same index.
Thus, $\ensuremath{\protect\widetilde{A}}_{n}$ is the dual facet corresponding to the primal edge $\ensuremath{L}_{n}$ and $\ensuremath{\protect\widetilde{L}}_{n}$ is the dual edge corresponding to the primal facet $\ensuremath{A}_{n}$.
Let us further introduce a short (index) notation for geometric objects.
If $\ensuremath{L}_{n}$ or $\ensuremath{A}_{n}$ are used as an index, we simply write $n$ instead.
Whether $n$ refers to an edge or a facet should become clear from the context.
For the dual objects, we use $\protect\tilde{n}$ instead of $\ensuremath{\protect\widetilde{L}}_{n}$ or $\ensuremath{\protect\widetilde{A}}_{n}$.
The notation for points and volumes and their duals is done accordingly.
In Table~\ref{tab:gridNotation}, we summarise this notation.
\begin{figure}
\centering
\includegraphics{./figs/fig2}
\caption{Staggered pair of primal and dual cell for a regular hexahedral grid with the allocation of electric, magnetic and thermal quantities.}
\label{fig:gridStaggering}
\end{figure}
\begin{table}
\centering
\begin{tabular}{ccc}\toprule
Description & Normal & Index\\\bottomrule\toprule
$i$-th primal point & $\ensuremath{P}_{i}$ & $i$\\
$i$-th dual point & $\ensuremath{\protect\widetilde{P}}_{i}$ & $\protect\tilde{i}$\\
$n$-th primal edge & $\ensuremath{L}_{n}$ & $n$\\
$n$-th dual edge & $\ensuremath{\protect\widetilde{L}}_{n}$ & $\protect\tilde{n}$\\
$n$-th primal facet & $\ensuremath{A}_{n}$ & $n$\\
$n$-th dual facet & $\ensuremath{\protect\widetilde{A}}_{n}$ & $\protect\tilde{n}$\\
$i$-th primal volume & $\ensuremath{V}_{i}$ & $i$\\
$i$-th dual volume & $\ensuremath{\protect\widetilde{V}}_{i}$ & $\protect\tilde{i}$\\\bottomrule
\end{tabular}
\caption{Normal and index notation format for different entities of the grid.}
\label{tab:gridNotation}
\end{table}
The grid counterparts of the field quantities are allocated to points, edges, facets or volumes and collected in column vectors.
Typical examples from \gls*{EMs} are the discrete electric potentials \ensuremath{\boldsymbol{\mathrm{\varphi}}}, fields \ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace, currents \ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace and charges $\bbbow{\ensuremath{\mathrm{\mathbf{q}}}\xspace}$ that are allocated to primal points, edges, facets and volumes, respectively.
The number of bows indicates the dimension of the corresponding geometric object.
Nevertheless, we typically write \ensuremath{\mathrm{\mathbf{q}}}\xspace instead of $\bbbow{\ensuremath{\mathrm{\mathbf{q}}}\xspace}$ for conciseness.
Defining all other grid quantities accordingly, their allocation used in this paper is shown in Figure~\ref{fig:gridStaggering}.
Furthermore, if we want to indicate a grid quantity, e.g. an electric field, to be allocated to an edge $\ensuremath{L}_{n}$ that is part of the boundary of a facet $\ensuremath{A}_{k}$ ($\ensuremath{L}_{n}\cap\partial\ensuremath{A}_{k}=\ensuremath{L}_{n}$), we use the indexed notation $\ensuremath{\protect\bow{e}}_{k;n}$.
To relate grid quantities on either the primal or dual mesh, topological matrices, corresponding to the continuous topological operators, have to be defined.
The incidence between grid points and edges is given by the discrete gradient matrix $\ensuremath{\mathbf{G}}\xspace$.
For an oriented edge $\ensuremath{L}_{n}$, $G_{ni}=-1$ if $\ensuremath{P}_{i}$ is the starting point of the edge, $G_{ni}=1$ if $\ensuremath{P}_{i}$ is the ending point of the edge and $G_{ni}=0$ if $\ensuremath{P}_{i}$ is neither starting nor ending point of the edge ($\ensuremath{P}_{i}\cap\partial\ensuremath{L}_{n}=\emptyset$).
To express the incidence between grid edges and facets, we use the discrete curl matrix $\ensuremath{\mathbf{C}}\xspace$.
Given a primal facet $\ensuremath{A}_{k}$ and its oriented boundary $\partial\ensuremath{A}_{k}$, $\ensuremath{C}_{kn}=-1$ if edge $\ensuremath{L}_{n}$ is oriented in opposite direction than $\partial\ensuremath{A}_{k}$, $\ensuremath{C}_{kn}=1$ if edge $\ensuremath{L}_{n}$ is oriented in the same way as $\partial\ensuremath{A}_{k}$ and $\ensuremath{C}_{kn}=0$ if $\ensuremath{L}_{n}$ does not touch $\ensuremath{A}_{k}$ ($\ensuremath{L}_{n}\cap\partial\ensuremath{A}_{k}=\emptyset$).
Finally, the discrete divergence matrix $\ensuremath{\mathbf{S}}\xspace$ denotes the incidence between grid facets and volumes.
For a primal volume $\ensuremath{V}_{i}$ and its oriented boundary $\partial\ensuremath{V}_{k}$, the entries of $\ensuremath{\mathbf{S}}\xspace$ are defined in analogy to those of $\ensuremath{\mathbf{G}}\xspace$ and $\ensuremath{\mathbf{C}}\xspace$.
Additionally, we have the dual gradient, curl and divergence matrices \ensuremath{\protect\widetilde{\mathbf{G}}}\xspace, \ensuremath{\protect\widetilde{\mathbf{C}}}\xspace and \ensuremath{\protect\widetilde{\mathbf{S}}}\xspace, respectively, defined accordingly.
Useful relations between the topological matrices on the primal and dual grids are given by $\ensuremath{\mathbf{G}}\xspace=-\ensuremath{\protect\widetilde{\mathbf{S}}^\top}$, $\ensuremath{\protect\widetilde{\mathbf{G}}}\xspace=-\ensuremath{\mathbf{S}}\xspace^{\top}$ and $\ensuremath{\mathbf{C}}\xspace=\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace^{\top}$~\cite{Clemens_2001aa}.
To relate quantities on the primal grid to quantities on the dual grid and vice versa, material relations are employed.
For the problem formulated in Section~\ref{subsec:EMproblem}, the following three different kind of relations can be identified:
\begin{itemize}
\item Quantities allocated to primal facets must be related to quantities allocated to dual edges.
\item Quantities allocated to primal edges must be related to quantities allocated to dual facets.
\item Quantities allocated to primal points must be related to quantities allocated to dual volumes.
\end{itemize}
As representatives for the above listed constitutive relations, we formulate three discrete material laws as
\begin{equation*}
\ensuremath{\protect\bow{\mathrm{\mathbf{h}}}}\xspace=\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{b}}}}\xspace,\quad\ensuremath{\protect\bbow{\mathrm{\mathbf{d}}}}\xspace=\ensuremath{\mathbf{M}_{\varepsilon}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace,\quad \ensuremath{ \mathbf{Q} }=\ensuremath{\mathbf{M}_{\rho c}}\xspace\dot{\ensuremath{\mathbf{T}}},\\
\end{equation*}
where $\ensuremath{\mathbf{M}_{\nu}}\xspace$ is the magnetic reluctance matrix mapping the discrete magnetic flux $\ensuremath{\protect\bbow{\mathrm{\mathbf{b}}}}\xspace$ allocated to primal facets to the discrete magnetic field $\ensuremath{\protect\bow{\mathrm{\mathbf{h}}}}\xspace$ allocated to dual edges, $\ensuremath{\mathbf{M}_{\varepsilon}}\xspace$ is the electric capacitance matrix mapping the discrete electric field $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace$ allocated to primal edges to the discrete electric flux density $\ensuremath{\protect\bbow{\mathrm{\mathbf{d}}}}\xspace$ allocated to dual facets, and
$\ensuremath{\mathbf{M}_{\rho c}}\xspace$ is the thermal capacitance matrix mapping the time derivative of the grid temperature $\dot{\ensuremath{\mathbf{T}}}$ allocated to primal points to the discrete heat power $\ensuremath{ \mathbf{Q} }$ allocated to dual volumes.
For other constitutive parameters, the material matrices are defined following these three cases.
Having established the grid constructs $G$ and $\protect\widetilde{G}$ and the corresponding topological and material matrices, the E-H formulation~\eqref{eq:MaxwellContInt} of \textsc{Maxwell}\xspace's equations upon such a grid pair can be written as~\cite{Weiland_1996aa}
\begin{subequations}
\begin{align}
-\ensuremath{\mathbf{C}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace &=\ensuremath{\mathbf{M}_{\nu}}\xspace^{-1}\frac{\mathrm{d}\ensuremath{\protect\bow{\mathrm{\mathbf{h}}}}\xspace}{\mathrm{d}t},\label{eq:MGEmagChargeFaraday}\\
\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{h}}}}\xspace &=\ensuremath{\mathbf{M}_{\varepsilon}}\xspace\frac{\mathrm{d}\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace}{\mathrm{d}t}+\ensuremath{\mathbf{M}_{\sigma}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace+\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}},\label{eq:MGEmagChargeAmpere}\\
\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace \ensuremath{\protect\bbow{\mathrm{\mathbf{d}}}}\xspace &=\ensuremath{\mathrm{\mathbf{q}}}\xspace,\\
\qquad\ensuremath{\mathbf{S}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{b}}}}\xspace&=\ensuremath{\mathbf{0}},
\end{align}
\label{eq:MGEmagCharge}\end{subequations}
where \ensuremath{\mathbf{M}_{\sigma}}\xspace is the electric conductance matrix, \ensuremath{\protect\bbow{\mathrm{\mathbf{b}}}}\xspace is the discrete magnetic flux density, $\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}}$ is the discrete impressed electric current density and \ensuremath{\mathrm{\mathbf{q}}}\xspace is the discrete electric charge.
Similarly, with the discrete vector potential \ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace, the discrete E-A formulation is given by
\begin{subequations}
\begin{align}
\ensuremath{\mathbf{C}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace&=-\ensuremath{\mathbf{C}}\xspace\frac{\mathrm{d}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace}{\mathrm{d}t},\label{eq:MGEfaraday}\\
\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\mathbf{C}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace&=\ensuremath{\mathbf{M}_{\varepsilon}}\xspace\frac{\mathrm{d}\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace}{\mathrm{d}t}+\ensuremath{\mathbf{M}_{\sigma}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace+\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}},\label{eq:MGEampere}
\end{align}
\label{eq:MGE}\end{subequations}
together with the discrete form of the gauging
\begin{equation}
\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\mathbf{M}_{\text{G}}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace=\ensuremath{ \mathbf{F} },
\label{eq:gaugeAfit}
\end{equation}
where $\ensuremath{ \mathbf{F} }$ is the discrete counterpart of the gauging function $f$ and $\mathbf{M}_{\text{G}}$ is a gauging matrix.
Since $\mathbf{M}_{\text{G}}$ maps from quantities on primal edges to dual facets, it shares properties with the material matrices (e.g. $\ensuremath{\mathbf{M}_{\varepsilon}}\xspace$) and thus can be interpreted as a material matrix with a material value equal to $\sigma_{\text{g}}$.
Casting also the heat equation \eqref{eq:thermalContInt} into a spatially discrete form, we obtain
\begin{equation*}
\ensuremath{\mathbf{M}_{\rho c}}\xspace\dot{\ensuremath{\mathbf{T}}}+\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\mathbf{M}_{\lambda}}\xspace\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace^{\top}\ensuremath{\mathbf{T}}=\mathbf{Q}_{\text{J}}-\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{q}}}}\xspace_{\text{i}},
\end{equation*}
where $\ensuremath{\mathbf{M}_{\lambda}}\xspace$ is the thermal conductance matrix and $\mathbf{Q}_{\text{J}}$ is the discrete vector of the \textsc{Joule}\xspace losses.
For details on the computation of $\mathbf{Q}_{\text{J}}$, we refer the reader to the work by Casper et al~\cite{Casper_2016ab}.
The impressed thermal fluxes are given by their discrete representative $\ensuremath{\protect\bbow{\mathrm{\mathbf{q}}}}\xspace_{\text{i}}$.
\subsection{Finite Integration Technique and its Relation to Other Discretisation Schemes}
For the circuit extraction from 3D field models as presented in this paper, we require diagonal, symmetric and positive definite material matrices.
To fulfil this requirement, we choose to use the \gls*{FIT} and assume the material to coincide with the primal grid cells, see Figure~\ref{fig:gridStaggering}.
While the \gls*{FIT} has also been formulated for anisotropic materials~\cite{Kruger_2001aa}, diagonal material matrices are obtained only in the isotropic case or in the case when the principal axes of anisotropy coincide with the coordinate axes.
Then, the entries of the different material matrices are given by
\begin{alignat*}{2}
M_{\varepsilon;\protect\tilde{n} n}=\varepsilon_{\protect\tilde{n} n}\frac{|\ensuremath{\protect\widetilde{A}}_{n}|}{|\ensuremath{L}_{n}|},\quad
M_{\lambda;\protect\tilde{n} n}&=\lambda_{\protect\tilde{n} n}\frac{|\ensuremath{\protect\widetilde{A}}_{n}|}{|\ensuremath{L}_{n}|},\quad
M_{\sigma;\protect\tilde{n} n}&&=\sigma_{\protect\tilde{n} n}\frac{|\ensuremath{\protect\widetilde{A}}_{n}|}{|\ensuremath{L}_{n}|},\\
M_{\nu;\protect\tilde{n} n}&=\nu_{\protect\tilde{n} n}\frac{|\ensuremath{\protect\widetilde{L}}_{n}|}{|\ensuremath{A}_{n}|},\quad
M_{\rho c;\protect\tilde{i} i}&&=\rho c_{\protect\tilde{i} i}|\ensuremath{\protect\widetilde{V}}_{i}|,
\end{alignat*}
where $\varepsilon_{\protect\tilde{n} n}$, $\sigma_{\protect\tilde{n} n}$, $\lambda_{\protect\tilde{n} n}$, $\nu_{\protect\tilde{n} n}$, and $\rho c_{\protect\tilde{i} i}$ are obtained by a suitable averaging scheme~\cite{Clemens_2001aa,Casper_2016aa} and $|\cdot|$ represents the measure (i.e. area, length or volume) of the corresponding geometrical object.
When using the \gls*{FIT} as the discretisation scheme with a canonical numbering of the grid nodes, the curl and material matrices exhibit the block structure
\begin{align}
\ensuremath{\mathbf{C}}\xspace=
\begin{pmatrix}
\ensuremath{\mathbf{0}} & -\ensuremath{\mathbf{P}}\xspace_{z} & \ensuremath{\mathbf{P}}\xspace_{y} \\
\ensuremath{\mathbf{P}}\xspace_{z} & \ensuremath{\mathbf{0}} & -\ensuremath{\mathbf{P}}\xspace_{x}\\
-\ensuremath{\mathbf{P}}\xspace_{y} & \ensuremath{\mathbf{P}}\xspace_{x} & \ensuremath{\mathbf{0}}
\end{pmatrix},\quad
\ensuremath{\mathbf{M}_{\varepsilon}}\xspace&=
\begin{pmatrix}
\ensuremath{\mathbf{M}}_{\varepsilon;x} & \ensuremath{\mathbf{0}} & \ensuremath{\mathbf{0}}\\
\ensuremath{\mathbf{0}} & \ensuremath{\mathbf{M}}_{\varepsilon;y} & \ensuremath{\mathbf{0}}\\
\ensuremath{\mathbf{0}} & \ensuremath{\mathbf{0}} & \ensuremath{\mathbf{M}}_{\varepsilon;z}
\end{pmatrix},
\label{eq:fitMatBlockStruct}
\end{align}
where $\ensuremath{\mathbf{P}}\xspace_{\xi}\in\{0,\pm 1\}^{N_{\text{E}}\times N_{\text{P}}}$, $\xi\in\{x,y,z\}$ are the grid differential operators for the different coordinate directions.
Here, \ensuremath{\mathbf{M}_{\varepsilon}}\xspace was used as an example whereas an equivalent block structure applies for all other material matrices.
Within the theory of the \gls*{FIT}, the discrete quantities \ensuremath{\boldsymbol{\mathrm{\varphi}}}, \ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace, \ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace, \ensuremath{\mathrm{\mathbf{q}}}\xspace that were introduced in Section~\ref{sec:FIT} are defined by means of integration with respect to their corresponding geometrical object, such that
\begin{align*}
\left(\ensuremath{\boldsymbol{\mathrm{\varphi}}}\right)_{i}:=\varphi(\ensuremath{P}_{i}),\quad
\ensuremath{\protect\bow{e}}_{n}:=\int_{\ensuremath{L}_{n}}\ensuremath{\mathbf{E}}\cdot\mathrm{d}\ensuremath{\mathbf{\ensuremath{L}}},\quad
\ensuremath{\protect\bbow{j}}_{\protect\tilde{n}}:=\int_{\ensuremath{\protect\widetilde{A}}_{n}}\ensuremath{\mathbf{J}}\cdot\mathrm{d}\ensuremath{\mathbf{S}},\quad
q_{i}:=\int_{\ensuremath{V}_i}\varrho\ \mathrm{d} V,
\end{align*}
where an analogous definition applies for \ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace, \ensuremath{\protect\bow{\mathrm{\mathbf{h}}}}\xspace, \ensuremath{\protect\bbow{\mathrm{\mathbf{d}}}}\xspace and \ensuremath{\protect\bbow{\mathrm{\mathbf{b}}}}\xspace.
Due to the applied integration, one speaks of grid voltages instead of discrete fields, of grid currents (fluxes) instead of discrete current (flux) densities and of grid charges instead of discrete charge densities.
As an alternative to the \gls*{FIT}, equivalent approaches such as the cell method~\cite{Alotto_2006aa} or the \gls*{FEM} can be used as long as the orthogonality and one-to-one relation between primal and dual grid objects is guaranteed.
When using \gls*{FEM}, linear basis functions together with an appropriate mass lumping for the material matrices must be used to obtain an equivalent scheme~\cite{Bondeson_2005aa}.
\section{A Primer on Circuit Theory and the Modified Nodal Analysis}
\label{sec:MNA}
In this section, we briefly review some fundamentals about circuit theory and the \gls*{MNA}~\cite{Ho_1975aa,Gunther_2005aa}.
\textsc{Kirchhoff}\xspace's current and voltage laws are derived and form the basics for circuit analysis.
Any circuit can be understood as a directed graph consisting of interconnected nodes and branches.
Let $v_{i}$ be one of the $N_{\text{n}}$ nodal potentials in a circuit and $b_{n}$ one of $N_{\text{b}}$ directed branches.
With the incidence matrix $\ensuremath{ \mathbf{A} }\in\{-1,0,1\}^{N_{\text{n}}\times N_{\text{b}}}$ linking nodes and branches, the voltage-potential relation is given by
\begin{equation*}
\ensuremath{ \mathbf{V} } = \ensuremath{ \mathbf{A} }^{\top}\mathbf{v}.
\end{equation*}
The entries of $\ensuremath{ \mathbf{A} }$ are defined such that $a_{in}=+1$ if the branch $b_{n}$ is directed away from node $n_{i}$ and $a_{in}=-1$ if $b_{n}$ is directed towards $n_{i}$.
If $n_{i}$ is neither starting nor ending point of $b_{n}$, then $a_{in}=0$.
With this definition, the exemplary voltage $V_{n}$ on the branch $b_{n}$ directed from $n_{i}$ to $n_{j}$ is given by $V_{n}=v_{i}-v_{j}$.
For time invariant geometries, the current continuity equation reads
\begin{equation}
\int_{\partial V}\ensuremath{\mathbf{J}}\cdot\ \mathrm{d}\mathbf{A}+\int_V\dot{\varrho}\ \mathrm{d} V=0,
\label{eq:continuityEquation}
\end{equation}
for an arbitrary volume $V$.
By considering a volume $\ensuremath{\protect\widetilde{V}}_{i}$ around an arbitrary circuit node $n_{i}$ and assuming that capacitive charges are located either fully inside or outside of $\ensuremath{\protect\widetilde{V}}_{i}$, the total charge and also the charge's change rate in $\ensuremath{\protect\widetilde{V}}_{i}$ is zero.
Therefore, \eqref{eq:continuityEquation} becomes
\begin{equation*}
\int_{\partial\ensuremath{\protect\widetilde{V}}_{i}}\ensuremath{\mathbf{J}}\cdot\ \mathrm{d}\mathbf{A} = 0.
\end{equation*}
If $\partial\ensuremath{\protect\widetilde{V}}_{i}$ is composed by a finite number $s$ of conductors with cross-sectional areas $\ensuremath{\protect\widetilde{A}}_{n}$, \gls*{KCL} is obtained as
\begin{equation}
\sum_{n=1}^{s}I_{n}=\sum_{n=1}^{s}\int_{\ensuremath{\protect\widetilde{A}}_{n}}\ensuremath{\mathbf{J}}\cdot\ \mathrm{d}\mathbf{A}=0,
\label{eq:KCLoneNode}
\end{equation}
where $I_{n}$ is the total current through the facet $\ensuremath{\protect\widetilde{A}}_{n}$.
This relation is also depicted in Figure~\ref{fig:KCLonenode}.
To express~\eqref{eq:KCLoneNode} for all nodes in the circuit (cf. Figure~\ref{fig:KCLmanynodes}), the incidence matrix $\ensuremath{ \mathbf{A} }$ can be used such that
\begin{equation}
\ensuremath{ \mathbf{A} }\ensuremath{ \mathbf{I} }=\ensuremath{\mathbf{0}},
\label{eq:KCLallNodes}
\end{equation}
where $\ensuremath{ \mathbf{I} }\in\mathbb{R}^{N_{\text{b}}}$ is a vector of all currents allocated to the branches and $\ensuremath{\mathbf{0}}$ is a vector of zeros of suitable dimension.
\begin{figure}[t]
\centering
\subfloat[\label{fig:KCLonenode}]{\raisebox{3ex}{\includegraphics[width=0.23\columnwidth]{./figs/fig3a}}}
\subfloat[\label{fig:KCLmanynodes}]{\raisebox{3ex}{\includegraphics[width=0.23\columnwidth]{./figs/fig3b}}}
\subfloat[\label{fig:circuitExample}]{\includegraphics[width=0.53\columnwidth]{./figs/fig3c}}
\caption{KCL example for (a) one node and (b) several nodes. The net current flow into a node must be zero. (c) Circuit example and its branch matrices.}
\end{figure}
In a circuit, the basic branch elements are conductors, capacitors, inductors as well as voltage and current sources.
As these elements are allocated to branches, sets of $N_{\text{G}}$, $N_{\text{C}}$, $N_{\text{L}}$, $N_{\text{V}}$ and $N_{\text{I}}$ branches are defined.
Hence, $\ensuremath{ \mathbf{A} }$ can be arranged into a block matrix with sub-blocks for these elements~\cite{Gunther_2005aa}, viz.
\begin{equation*}
\ensuremath{ \mathbf{A} } =
\begin{bmatrix}
\ensuremath{\ensuremath{\mathbf{A}}_\text{G}} & \ensuremath{\ensuremath{\mathbf{A}}_\text{C}} & \ensuremath{\ensuremath{\mathbf{A}}_\text{L}} & \ensuremath{\ensuremath{\mathbf{A}}_\text{V}} & \ensuremath{\ensuremath{\mathbf{A}}_\text{I}}
\end{bmatrix}.
\end{equation*}
A circuit example consisting of a current source, two resistors and two capacitors together with the corresponding block structure of $\ensuremath{\mathbf{A}}$ is shown in Figure~\ref{fig:circuitExample}.
A similar subdivision is done for the current and voltage vectors,
\begin{equation*}
\ensuremath{\mathbf{I}}^{\top} =
\begin{bmatrix}
\ensuremath{\ensuremath{\mathbf{I}}_\text{R}}^{\top} & \ensuremath{\ensuremath{\mathbf{I}}_\text{C}}^{\top} & \ensuremath{\ensuremath{\mathbf{I}}_\text{L}}^{\top} & \ensuremath{\ensuremath{\mathbf{I}}_\text{V}}^{\top} & \ensuremath{\ensuremath{\mathbf{I}}_\text{I}}^{\top}
\end{bmatrix},\quad
\ensuremath{\mathbf{V}}^{\top} =
\begin{bmatrix}
\ensuremath{\ensuremath{\mathbf{V}}_\text{G}}^{\top} & \ensuremath{\ensuremath{\mathbf{V}}_\text{C}}^{\top} & \ensuremath{\ensuremath{\mathbf{V}}_\text{L}}^{\top} & \ensuremath{\ensuremath{\mathbf{V}}_\text{V}}^{\top} & \ensuremath{\ensuremath{\mathbf{V}}_\text{I}}^{\top}
\end{bmatrix}.
\end{equation*}
With these definitions, the voltages are given by
\begin{equation}
\ensuremath{\ensuremath{\mathbf{V}}_\text{G}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}^{\top}\ensuremath{\mathbf{v}},\quad
\ensuremath{\ensuremath{\mathbf{V}}_\text{C}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}^{\top}\ensuremath{\mathbf{v}},\quad
\ensuremath{\ensuremath{\mathbf{V}}_\text{L}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{L}}^{\top}\ensuremath{\mathbf{v}},\quad
\ensuremath{\ensuremath{\mathbf{V}}_\text{V}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}^{\top}\ensuremath{\mathbf{v}},\quad
\ensuremath{\ensuremath{\mathbf{V}}_\text{I}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{I}}^{\top}\ensuremath{\mathbf{v}},
\label{eq:currentVoltageBlock}
\end{equation}
and~\eqref{eq:KCLallNodes} becomes
\begin{equation}
\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}\ensuremath{\ensuremath{\mathbf{I}}_\text{R}} + \ensuremath{\ensuremath{\mathbf{A}}_\text{C}}\ensuremath{\ensuremath{\mathbf{I}}_\text{C}} + \ensuremath{\ensuremath{\mathbf{A}}_\text{L}}\ensuremath{\ensuremath{\mathbf{I}}_\text{L}} + \ensuremath{\ensuremath{\mathbf{A}}_\text{V}}\ensuremath{\ensuremath{\mathbf{I}}_\text{V}} + \ensuremath{\ensuremath{\mathbf{A}}_\text{I}}\ensuremath{\ensuremath{\mathbf{I}}_\text{I}} = \ensuremath{\mathbf{0}}.
\label{eq:KCLblock}
\end{equation}
The relation between voltages and currents for the different branches is established by constitutive diagonal matrices that contain the element-wise material parameters.
These are the conductance, capacitance, and inductance matrices $\ensuremath{\mathbf{G}}$, $\ensuremath{\mathbf{C}}$ and $\ensuremath{\mathbf{L}}$, respectively.
Expressing the corresponding source branches by means of the source voltages $\ensuremath{\ensuremath{\mathbf{V}}_\text{s}}$ and source currents $\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}$, this relation becomes
\begin{equation}
\ensuremath{\ensuremath{\mathbf{I}}_\text{R}}=\ensuremath{\mathbf{G}}\ensuremath{\ensuremath{\mathbf{V}}_\text{G}},\quad
\ensuremath{\ensuremath{\mathbf{I}}_\text{C}}=\ensuremath{\mathbf{C}}\ensuremath{\dot{\ensuremath{\mathbf{V}}}_\text{C}},\quad
\ensuremath{\ensuremath{\mathbf{I}}_\text{L}}=\ensuremath{\mathbf{L}}^{-1}\int\ensuremath{\ensuremath{\mathbf{V}}_\text{L}}\ \mathrm{d} t,\quad
\ensuremath{\ensuremath{\mathbf{V}}_\text{V}}=\ensuremath{\ensuremath{\mathbf{V}}_\text{s}},\quad
\ensuremath{\ensuremath{\mathbf{I}}_\text{I}}=\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}.
\label{eq:branchRelations}
\end{equation}
Combining~\eqref{eq:currentVoltageBlock}, \eqref{eq:KCLblock} and~\eqref{eq:branchRelations}, we obtain the \gls*{MNA} formulation
\begin{equation}
\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}\ensuremath{\mathbf{C}}\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}^{\top}\dot{\ensuremath{\mathbf{v}}}+\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}\ensuremath{\mathbf{G}}\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}^{\top}\ensuremath{\mathbf{v}}+\ensuremath{\ensuremath{\mathbf{A}}_\text{L}}\ensuremath{\ensuremath{\mathbf{I}}_\text{L}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{I}}\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}-\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}\ensuremath{\ensuremath{\mathbf{I}}_\text{V}},\quad
\ensuremath{\mathbf{L}}\dot{\ensuremath{\mathbf{I}}}_{\text{L}}=\ensuremath{\ensuremath{\mathbf{A}}_\text{L}}^{\top}\ensuremath{\mathbf{v}},\quad
\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}^{\top}\ensuremath{\mathbf{v}}=\ensuremath{\ensuremath{\mathbf{V}}_\text{s}}.
\label{eq:MNA}
\end{equation}
Note that in contrast to the standard \gls*{MNA} theory, \eqref{eq:MNA} still requires regularisation, typically done by the introduction of a reference (\emph{ground}) node.
\section{Circuit Representation of Electrothermal Field Problems}
\label{sec:circuitsET}
In this section, we derive the circuit representation for transient \gls*{ET} field problems.
We apply the \gls*{EQS} approximation~\cite{Clemens_2004ab} to \textsc{Maxwell}\xspace's equations given by \eqref{eq:MaxwellContInt} and consider the coupling with the transient heat equation given by~\eqref{eq:thermalContInt}.
Then, the bi-directionally coupled system in differential form reads
\begin{subequations}
\begin{align}
-\nabla \cdot\left(\varepsilon\nabla\dot{\varphi}\right)-\nabla \cdot\left(\sigma(T)\nabla\varphi\right) &= -\nabla \cdot\ensuremath{\mathbf{J}_{\mathrm{i}}},\label{eq:EQScont}\\
\rho c \dot{T} - \nabla \cdot\left(\lambda(T)\nabla T\right) &= Q_{\text{J}}(\varphi,T)-\nabla \cdot\ensuremath{\mathbf{q}_{\mathrm{i}}},\label{eq:thermalCont}
\end{align}
\label{eq:ETcont}\end{subequations}
with suitable initial and boundary conditions.
The coupling is manifested by the \textsc{Joule}\xspace heating given by $Q_{\text{J}}=\sigma(\nabla\varphi)^{2}$ in one way and by the temperature dependent conductivity $\sigma(T)$ in the opposite way.
Due to the \gls*{EQS} approximation, this formulation does not account for inductive effects but does consider resistive and capacitive effects.
For simplicity, we neglect the temperature dependency of the permittivity $\varepsilon$ and of the volumetric heat capacity $\rho c$.
Applying \gls*{FIT} upon the \gls*{ET} system of~\eqref{eq:ETcont}, the semi-discrete formulation reads
\begin{subequations}
\begin{align}
\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\mathbf{M}_{\varepsilon}}\xspace\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace^{\top}\dot{\ensuremath{\boldsymbol{\mathrm{\varphi}}}}+\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\mathbf{M}_{\sigma}}\xspace(\ensuremath{\mathbf{T}})\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace^{\top}\ensuremath{\boldsymbol{\mathrm{\varphi}}}&=-\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}},\label{eq:EQSdisc}\\
\ensuremath{\mathbf{M}_{\rho c}}\xspace\dot{\ensuremath{\mathbf{T}}}+\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\mathbf{M}_{\lambda}}\xspace(\ensuremath{\mathbf{T}})\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace^{\top}\ensuremath{\mathbf{T}}&=\mathbf{Q}_{\text{J}}(\ensuremath{\boldsymbol{\mathrm{\varphi}}},\ensuremath{\mathbf{T}})-\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{q}}}}\xspace_{\text{i}},\label{eq:thermalDisc}
\end{align}
\label{eq:ETdisc}\end{subequations}
with initial and boundary conditions yet to be applied.
Note that the system \eqref{eq:EQSdisc} requires a regularisation which is typically done by choosing a reference (\emph{ground}) node $\varphi_{\text{gnd}}=0$, where $\varphi_{\text{gnd}}$ is one of the entries of $\ensuremath{\boldsymbol{\mathrm{\varphi}}}$.
To generate the netlist for formulation~\eqref{eq:ETdisc}, the electric and thermal sub-problems are considered separately.
The connection is subsequently established by the \textsc{Joule}\xspace losses and the temperature dependent electric conductivity.
Next, in Section~\ref{subsec:EQSextract} and Section~\ref{subsec:thermalExtract}, the netlist generation for the \gls*{EQS} case and the thermal case is presented, respectively.
Temperature dependent materials are discussed in Section~\ref{subsec:ETnonlinearMaterials} and finally, the implementation of initial and boundary conditions is described in Section~\ref{subsec:ETinitialBoundary}.
This allows us to formulate an algorithm for the \gls*{ET} netlist generation as presented in Section~\ref{subsec:netlistGeneration}.
\subsection{Electroquasistatic Circuit Representation}
\label{subsec:EQSextract}
Let us now concentrate on the \gls*{EQS} sub-problem given by~\eqref{eq:EQScont}.
Since inductances are neglected in the \gls*{EQS} case, \eqref{eq:MNA} simplifies to
\begin{align}
\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}\ensuremath{\mathbf{C}}\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}^{\top}\dot{\ensuremath{\mathbf{v}}}+\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}\ensuremath{\mathbf{G}}\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}^{\top}\ensuremath{\mathbf{v}}=-\ensuremath{\ensuremath{\mathbf{A}}_\text{I}}\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}-\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}\ensuremath{\ensuremath{\mathbf{I}}_\text{V}},\quad
\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}^{\top}\ensuremath{\mathbf{v}}=\ensuremath{\ensuremath{\mathbf{V}}_\text{s}}.
\label{eq:MNA4EQS}
\end{align}
Thus, by inspection of~\eqref{eq:EQSdisc} and~\eqref{eq:MNA4EQS}, we are led to the following equivalences:
\begin{itemize}
\item The incidence matrices $\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}$ and $\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}$ coincide with the \gls*{FIT} divergence matrix $\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace$.
\item The capacitance matrix $\ensuremath{\mathbf{C}}$ coincides with the \gls*{FIT} capacitance matrix $\ensuremath{\mathbf{M}_{\varepsilon}}\xspace$.
\item The incidence matrices $\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}$ and $\ensuremath{\ensuremath{\mathbf{A}}_\text{I}}$ coincide with the identity matrix.
\item The conductance matrix $\ensuremath{\mathbf{G}}$ coincides with the \gls*{FIT} conductance matrix $\ensuremath{\mathbf{M}_{\sigma}}\xspace$.
\item The nodal voltages $\ensuremath{\mathbf{v}}$ correspond to the \gls*{FIT} degrees of freedom $\ensuremath{\boldsymbol{\mathrm{\varphi}}}$.
\item The source currents $\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}$ are given by the divergence of the impressed currents $\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}}$.
\item The source voltages $\ensuremath{\ensuremath{\mathbf{V}}_\text{s}}$ correspond to the \textsc{Dirichlet}\xspace potentials $\ensuremath{\boldsymbol{\mathrm{\varphi}}}_{\text{Dir}}$, which are related to the reference node $\varphi_{\text{gnd}}$.
\item These equivalences also prevail themselves in the physical units.
\end{itemize}
Summarised, the field-circuit relations for \gls*{EQS} read
\begin{subequations}
\begin{align}
\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}\mathrel{\widehat{=}}\ensuremath{\ensuremath{\mathbf{A}}_\text{C}} &\mathrel{\widehat{=}}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace, \label{eq:EQSequivAGC}\\
\ensuremath{\mathbf{G}}\mathrel{\widehat{=}}\ensuremath{\mathbf{M}_{\sigma}}\xspace,\quad\ensuremath{\mathbf{C}} &\mathrel{\widehat{=}}\ensuremath{\mathbf{M}_{\varepsilon}}\xspace, \label{eq:EQSequivMats}\\
\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}\mathrel{\widehat{=}}\ensuremath{\ensuremath{\mathbf{A}}_\text{I}} &\mathrel{\widehat{=}}\mathbbm{I}, \label{eq:EQSequivAVI}\\
\ensuremath{\mathbf{v}} &\mathrel{\widehat{=}}\ensuremath{\boldsymbol{\mathrm{\varphi}}}, \label{eq:EQSequivV}\\
\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}&\mathrel{\widehat{=}}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}},\label{eq:EQSequivI}\\
\ensuremath{\ensuremath{\mathbf{V}}_\text{s}}&\mathrel{\widehat{=}}\ensuremath{\boldsymbol{\mathrm{\varphi}}}_{\text{Dir}} \label{eq:EQSequivVdir},
\end{align}
\label{eq:EQSequivalences}\end{subequations}
where $\mathbbm{I}$ is the identity matrix of corresponding size and $\ensuremath{\boldsymbol{\mathrm{\varphi}}}_{\text{Dir}}$ represents the potentials on the \textsc{Dirichlet}\xspace boundary nodes.
To find the circuit stamp of each edge in the grid upon which~\eqref{eq:EQSdisc} holds, we employ the equivalences of~\eqref{eq:EQSequivalences} from which the circuit topology is derived.
From~\eqref{eq:EQSequivAGC}, conductors and capacitors are placed along the branches of the circuit.
According to~\eqref{eq:EQSequivMats}, the values of the conductors and capacitors are directly taken from the corresponding FIT material matrices.
\begin{figure}[t]
\centering
\includegraphics{./figs/fig4}
\caption{Equivalent electric circuit stamp for the exemplary edge $\ensuremath{L}_{m}$ between points $\ensuremath{P}_{i}$ and $\ensuremath{P}_{j}$.}
\label{fig:EQScircuitStamp}
\end{figure}
Current and voltage sources are connected between a circuit node and ground as indicated by~\eqref{eq:EQSequivAVI}.
Furthermore, \eqref{eq:EQSequivV} shows that the circuit's nodal potentials are equal to the potentials at the grid points.
According to~\eqref{eq:EQSequivI}, the current sources in the circuit represent the divergence of the \gls*{FIT} impressed currents.
Finally, if \textsc{Dirichlet}\xspace \glspl*{BC} are imposed, voltage sources in the circuit represent \gls*{FIT} \textsc{Dirichlet}\xspace potentials as given by~\eqref{eq:EQSequivVdir}.
We further discuss \glspl*{BC} in Section~\ref{subsec:ETinitialBoundary}.
To summarise, if we consider an exemplary grid edge $\ensuremath{L}_{m}$, we obtain a representative \gls*{EQS} circuit stamp as shown in Figure~\ref{fig:EQScircuitStamp}.
Note that the temperature dependence of the materials is neglected for now and will be discussed in Section~\ref{subsec:ETnonlinearMaterials}
\subsection{Thermal Circuit Representation}
\label{subsec:thermalExtract}
In this section, we describe the circuit representation of the sub-problem described by~\eqref{eq:thermalCont}.
By comparing~\eqref{eq:thermalDisc} to~\eqref{eq:EQSdisc}, we observe a slightly different equation structure.
Thermal capacities are not subject to spatial differences and thus do not link to neighbouring nodes.
Instead, a thermal capacitance influences the change rate of the absolute temperature of a node.
Thus, thermal capacitances are placed on branches connecting the nodes to a reference node at zero temperature.
This reference node is an additional non-physical node that is introduced to obtain a consistent circuit representation.
In the literature, this approach is also referred to as the \textsc{Cauer}\xspace model representing a discretised image of the heat flow~\cite{Gerstenmaier_2007aa,Bechtold_2006ab}.
An equivalent approach is the \textsc{Foster}\xspace model, in which the capacitances are placed between the circuit nodes and the parameters are adjusted accordingly.
In the \textsc{Foster}\xspace model, the heat propagation is instantaneous and does not account for the fact that an object requires some delay before changing its temperature.
The \gls*{MNA} formulation of~\eqref{eq:MNA} must be extended by this additional reference ground node such that
\begin{equation}
\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{C}}\ensuremath{\mathbf{C}}\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{C}}^{\top}\dot{\ensuremath{\widehat{\ensuremath{\mathbf{v}}}}}+\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{G}}\ensuremath{\mathbf{G}}\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{G}}^{\top}\ensuremath{\widehat{\ensuremath{\mathbf{v}}}}=-\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{I}}\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}-\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{V}}\ensuremath{\ensuremath{\mathbf{I}}_\text{V}},\quad
\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{V}}^{\top}\ensuremath{\widehat{\ensuremath{\mathbf{v}}}}=\ensuremath{\ensuremath{\mathbf{V}}_\text{s}},
\label{eq:MNA4thermal}
\end{equation}
with
\begin{equation*}
\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{C}}:=
\begin{bmatrix}
\mathbbm{I} & -\mathbbm{1}
\end{bmatrix},\quad
\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{G}}:=
\begin{bmatrix}
\ensuremath{\ensuremath{\mathbf{A}}_\text{G}} & \ensuremath{\mathbf{0}}
\end{bmatrix},\quad
\ensuremath{\ensuremath{\widehat{\ensuremath{\mathbf{A}}}}_\text{I}}:=
\begin{bmatrix}
\mathbbm{I} & -\mathbbm{1}
\end{bmatrix},\quad
\ensuremath{\widehat{\ensuremath{\mathbf{v}}}}^{\top}:=
\begin{bmatrix}
\ensuremath{\mathbf{v}}^{\top} & v_{\text{gnd}}
\end{bmatrix},
\end{equation*}
where $\mathbbm{1}$ is a column vector of ones of appropriate size and $v_{\text{gnd}}=T_{\text{gnd}}=0$.
With these definitions, the equivalences between the \gls*{MNA} formulation of~\eqref{eq:MNA4thermal} and the thermal formulation of~\eqref{eq:thermalDisc} are readily obtained as
\begin{subequations}
\begin{align}
\ensuremath{\ensuremath{\mathbf{A}}_\text{C}}\mathrel{\widehat{=}}\ensuremath{\ensuremath{\mathbf{A}}_\text{V}}\mathrel{\widehat{=}} \ensuremath{\ensuremath{\mathbf{A}}_\text{I}}&\mathrel{\widehat{=}}\mathbbm{I}, \label{eq:thermalEquivAVIC}\\
\ensuremath{\ensuremath{\mathbf{A}}_\text{G}}&\mathrel{\widehat{=}}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace, \label{eq:thermalEquivAR}\\
\ensuremath{\mathbf{G}}\mathrel{\widehat{=}}\ensuremath{\mathbf{M}_{\lambda}}\xspace,\quad\ensuremath{\mathbf{C}} &\mathrel{\widehat{=}}\ensuremath{\mathbf{M}_{\rho c}}\xspace, \label{eq:thermalEquivMats}\\
\ensuremath{\mathbf{v}} &\mathrel{\widehat{=}}\ensuremath{\mathbf{T}}, \label{eq:thermalEquivT}\\
\ensuremath{\ensuremath{\mathbf{I}}_\text{s}}&\mathrel{\widehat{=}}-\mathbf{Q}_{\text{J}}+\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{q}}}}\xspace_{\text{i}}\label{eq:thermalEquivQ}\\
\ensuremath{\ensuremath{\mathbf{V}}_\text{s}}&\mathrel{\widehat{=}}\ensuremath{\mathbf{T}}_{\text{Dir}} \label{eq:thermalEquivTdir},
\end{align}
\label{eq:thermalEquivalences}\end{subequations}
where $\ensuremath{\mathbf{T}}_{\text{Dir}}$ represents the temperatures at \textsc{Dirichlet}\xspace boundary nodes.
To derive the circuit stamp of each edge in the grid upon which~\eqref{eq:thermalDisc} holds, we employ the equivalences in~\eqref{eq:thermalEquivalences} from which the circuit topology is derived.
From~\eqref{eq:thermalEquivAVIC}, we infer that capacitances, voltage and current sources are directly connected between a circuit node and the thermal ground.
As in the \gls*{EQS} case, conductors connect two neighbouring nodes in the grid as seen from~\eqref{eq:thermalEquivAR}.
The values of the conductors and capacitors are taken directly from the matrices $\ensuremath{\mathbf{M}_{\lambda}}\xspace$ and $\ensuremath{\mathbf{M}_{\rho c}}\xspace$ according to~\eqref{eq:thermalEquivMats}.
In this manner, the nodal potentials represent the sought temperatures as seen from~\eqref{eq:thermalEquivT}.
The current source is the sum of \textsc{Joule}\xspace losses, in fact represented by \glspl*{CCCS}, and the impressed heat flux according to~\eqref{eq:thermalEquivQ}.
Finally, if \textsc{Dirichlet}\xspace \glspl*{BC} are given, these are modelled by voltage sources in the circuit as stated in~\eqref{eq:thermalEquivTdir}.
We further comment on \glspl*{BC} in Section~\ref{subsec:ETinitialBoundary}.
To summarise, if we consider an exemplary grid edge $\ensuremath{L}_{m}$, we obtain a representative thermal circuit stamp as shown in Figure~\ref{fig:thermalCircuitStamp}.
Note that the temperature dependence of the materials is neglected until now and will be discussed in Section~\ref{subsec:ETnonlinearMaterials}.
\begin{figure}
\centering
\subfloat[\label{fig:thermalCircuitStamp}]{\includegraphics{./figs/fig5a}}
\hspace{2em}
\subfloat[\label{fig:RobinBoundary}]{\raisebox{3ex}{\includegraphics{./figs/fig5b}}}
\caption{(a) Equivalent thermal circuit stamp for the exemplary edge $\ensuremath{L}_{m}$ between points $\ensuremath{P}_{i}$ and $\ensuremath{P}_{j}$. (b) Visualisation of \textsc{Robin}\xspace \glspl*{BC}}
\end{figure}
To further highlight the equivalences between electric and thermal circuits, we would like to briefly comment on this.
Many quantities in electrical circuits can find their equivalent in thermal circuits.
For example, electric potentials are equivalent to temperatures while electric currents are equivalent to heat fluxes.
In Table~\ref{tab:electrothermalEquivalence}, some of these equivalences are summarised.
\begin{table}[t]
\centering
\begin{tabular}{r r}\toprule
Electrical Circuit & Thermal Circuit\\\bottomrule\toprule
electric potential (\si{V}) & temperature (\si{K})\\
electric voltage (\si{V}) & temperature difference (\si{K})\\
electric current (\si{A}) & thermal heat flux (\si{W})\\
electric charge (\si{C})& thermal energy (\si{J})\\
electric conductor (\si{S})& thermal conductor (\si{W/K})\\
electric capacitance (\si{F})& thermal capacitance (\si{J/K})\\\bottomrule
\end{tabular}
\caption{Equivalences between electric and thermal quantities in circuit representations.}
\label{tab:electrothermalEquivalence}
\end{table}
\subsection{Temperature Dependent Materials}
\label{subsec:ETnonlinearMaterials}
Most materials exhibit temperature dependent behaviour.
For the kind of considered materials, the temperature mainly influences the electric and thermal conductivity of the involved materials.
In this section, we describe an approach to account for a temperature-dependent electric conductivity when generating the corresponding SPICE\xspace netlist.
Nevertheless, the presented approach can be applied accordingly to other material's temperature dependencies as well, e.g.\ the thermal capacity.
Since the SPICE\xspace elements that represent the electric conductivities are conductors placed in circuit branches (cf. Section~\ref{subsec:EQSextract}), we first need to define the temperature of a branch $b_{m}$.
To this end, we take the average temperature $\overline{T}_{m}$ of the nodes interconnected by the branch $b_{m}$.
Assuming that the temperature dependence of the electric conductivity is known, the temperature-dependent electric conductance of branch $b_{m}$ is given by
\begin{equation}
G_{m}\left(\overline{T}_{m}\right)=\sigma_{\protect\tilde{m} m}\left(\overline{T}_{m}\right)\frac{|\ensuremath{\protect\widetilde{A}}_{m}|}{|\ensuremath{L}_{m}|}.
\label{eq:Gel}
\end{equation}
We implement~\eqref{eq:Gel} by means of behavioural sources in the SPICE\xspace language.
\subsection{Initial Conditions and Boundary Conditions}
\label{subsec:ETinitialBoundary}
When a coupled problem of more than one transient differential equation is considered, each sub-problem requires its own initial conditions and \glspl*{BC}.
Therefore, we impose these conditions on the \gls*{EQS} and thermal sub-problems separately.
However, certain equivalences allow to follow the same procedure for both sub-problems.
For any kind of transient problem, initial conditions are required.
As every SPICE\xspace dialect supports specifying initial conditions, these can be directly imposed by the corresponding syntax in the netlist.
For electric problems, different types of \glspl*{BC} are of interest.
In low-frequency problems as the \gls*{EQS} case, \textsc{Dirichlet}\xspace or \textsc{Neumann}\xspace conditions are typically used.
\textsc{Dirichlet}\xspace \glspl*{BC} conditions correspond to a fixed potential enforced at the boundary, while \textsc{Neumann}\xspace conditions prescribe the electric current through the boundary.
Similarly, in thermal problems, \textsc{Dirichlet}\xspace \glspl*{BC} correspond to a prescribed temperature at the boundary, while \textsc{Neumann}\xspace \glspl*{BC} prescribe thermal fluxes through the boundary.
Additionally, thermal problems commonly also involve \textsc{Robin}\xspace \glspl*{BC} that describe convective and radiative boundaries.
\textsc{Dirichlet}\xspace \glspl*{BC} are represented in the circuit by voltage sources between the ground node and the \textsc{Dirichlet}\xspace nodes.
Homogeneous \textsc{Neumann}\xspace conditions are automatically fulfilled since no edge or branch leaves the domain.
For simplicity, we do not consider inhomogeneous \textsc{Neumann}\xspace conditions in this paper.
A \textsc{Robin}\xspace \gls*{BC} can be understood as a conduction between a boundary node and an external node $n_{\infty}$ representing the fixed ambient temperature $T_{\infty}$.
Therefore, \textsc{Robin}\xspace \glspl*{BC} are represented in the circuit by conductors connected between the boundary nodes and $n_{\infty}$ as shown in Figure~\ref{fig:RobinBoundary}.
We collect the relevant \glspl*{BC} in Table~\ref{tab:bndConditions}.
\begin{table}[b]
\centering
\begin{tabular}{>{\centering}p{3.8cm}>{\centering\arraybackslash}p{5.2cm}}\toprule
Boundary Condition & Implementation\\\bottomrule\toprule
\textsc{Dirichlet}\xspace & \multicolumn{1}{c}{lumped voltage sources}\\
hom. \textsc{Neumann}\xspace & \multicolumn{1}{c}{no edges leaving the circuit}\\
\textsc{Robin}\xspace & additional non-physical ground\\\bottomrule
\end{tabular}
\caption{Considered \glspl*{BC} and their implementation.}
\label{tab:bndConditions}
\end{table}
\subsection{Electrothermal Netlist Generation}
\label{subsec:netlistGeneration}
\begin{algorithm}[h]
\caption{\gls{ET} SPICE netlist generation.}
\label{alg:netlistET}
\begin{algorithmic}[1]
\For{edge $\ensuremath{L}_{m}\in G$ between primal nodes $\ensuremath{P}_{i}$ and $\ensuremath{P}_{j}$}
\State write BGel$m$\quad{}n$i$\quad{}n$j$\quad{}$\text{I=}V_{ij}G_{m}(\overline{T}_{m})$
\State write Cel$m$\quad{}n$i$\quad{}n$j$\quad{}$M_{\varepsilon;\protect\tilde{m} m}$\quad{}$\text{ic}=0$
\State write Rth$m$\quad{}n$i$T\quad{}n$j$T\quad{}$M_{\lambda;\protect\tilde{m} m}^{-1}$
\EndFor
\For{node $\ensuremath{P}_{i}\in G$}
\State write Cth$i$\quad{}n$i$T\quad{}gnd\quad{}$M_{\rho c;\protect\tilde{i} i}$\quad{}$\text{ic}=T_{0}$
\State write BLoss$i$\quad{}gnd\quad{}n$i$T\quad{}$\text{I=}Q_{\text{J},i}(t)$
\If{$\ensuremath{P}_{i}$ is electric \textsc{Dirichlet}\xspace node}
\State write VDirEl$i$\quad{}n$i$\quad{}gnd\quad{}$V_{\text{Dir},i}(t)$
\EndIf
\If{$\ensuremath{P}_{i}$ is thermal \textsc{Dirichlet}\xspace node}
\State write VDirTh$i$\quad{}n$i$T\quad{}gnd\quad{}$T_{\text{Dir},i}(t)$
\EndIf
\If{an impressed current flows out of $\ensuremath{P}_{i}$}
\State write IimpEl$i$\quad{}n$i$\quad{}gnd\quad{}$(\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}}(t))_{i}$
\EndIf
\If{an impressed heat flux flows out of $\ensuremath{P}_{i}$}
\State write IimpTh$i$\quad{}n$i$T\quad{}gnd\quad{}$(\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\ensuremath{\protect\bbow{\mathrm{\mathbf{q}}}}\xspace_{\text{i}}(t))_{i}$
\EndIf
\EndFor
\end{algorithmic}
\end{algorithm}
To finalise this section, we formulate Algorithm~\ref{alg:netlistET} to automatically generate \gls*{ET} SPICE\xspace netlists.
For every grid edge $\ensuremath{L}_{m}$ that connects grid points $\ensuremath{P}_{i}$ and $\ensuremath{P}_{j}$, where $i<j$, the thermal conductance, electric capacitance and the temperature dependent electric conductance (cf. Section~\ref{subsec:ETnonlinearMaterials}) are written to the netlist connecting nodes $n_{i}$ and $n_{j}$ of the circuit (lines 2--4).
Due to the possible non-linearity of the electric conductivity (cf.~Section~\ref{subsec:ETnonlinearMaterials}), a behavioural source is used, where $V_{ij}$ is the voltage between node $n_{i}$ and $n_{j}$.
Initial conditions (ic) for the electric part can be included using the corresponding syntax for the capacitors.
Here, we use exemplary zero initial conditions.
Additionally, for every grid point $\ensuremath{P}_{i}$, the thermal capacitance and the \glsfirst{CCCS} representing the \textsc{Joule}\xspace losses are added to the netlist connecting node $n_{i}$ and the ground node (gnd) of the circuit (lines 5 and 6)\footnote{For a straight-forward implementation, we use behavioural sources instead of \glspl*{CCCS}}.
Initial conditions (ic) for the thermal part are specified by pre-charging the thermal capacitors with the initial temperature $T_{0}$.
To specify a \gls*{CCCS} in the SPICE\xspace language, a behavioural source is used.
If $\ensuremath{P}_{i}$ is specified as an electric (thermal) \textsc{Dirichlet}\xspace node, an additional voltage source connecting node $n_{i}$ and the ground node of the circuit is inserted (lines 9--14).
Furthermore, if an impressed current (heat flux) flows out of $\ensuremath{P}_{i}$, an additional current source is added to the netlist connecting node $n_{i}$ and the ground node of the circuit (lines 15--20).
\section{Circuit Representation of Electromagnetic Field Problems}
\label{sec:circuitsEM}
In this section, we neglect thermal effects and describe the circuit representation of general 3D \gls*{EM} field problems as given by \eqref{eq:MaxwellContInt}--\eqref{eq:FaradayAmpere}.
However, the thermo-\gls*{EM} coupling can be established analogously.
First, in Section~\ref{subsec:setsEdgesFacets}, we introduce an auxiliary set notion to collect specific edges and facets of the grid.
In Section~\ref{subsec:EMcircuitRepresentation}, the E-H formulation \eqref{eq:MGEmagCharge} and the E-A formulation \eqref{eq:MGE}--\eqref{eq:gaugeAfit} of the \gls*{MGEs} are transparently mapped into an electric circuit that fully describes the problem at hand.
Finally, in Section~\ref{subsec:theoryABC}, we extend our analysis in order to realise \glspl*{ABC} as circuit stamps.
These are typically needed to limit the computational domain while minimising unphysical reflections caused by the domain truncation.
\subsection{Auxiliary Sets of Edges and Facets}
\label{subsec:setsEdgesFacets}
In the following sections, we will take sums over specific edges or facets of the grid.
Since the edges and facets in the neighbourhood of a specific edge $\ensuremath{L}_{m}$ are of interest, we introduce sets containing collections of these edges and facets and label them with the superscript~$^{m}$.
Let $\mathcal{A}^{m}$ be the set of all facets in which $\ensuremath{L}_{m}$ is embedded.
For a regular hexahedral grid, these facets are shown in Figure~\ref{fig:edgeSets}.
All edges that are embedded in the facets contained in $\mathcal{A}^{m}$ are collected in the set $\mathcal{L}^{m}$, where this definition also implies $\ensuremath{L}_{m}\in\mathcal{L}^{m}$.
We denote by $\mathcal{L}^{m,0}$ the resulting set after extracting the very edge $\ensuremath{L}_{m}$ from $\mathcal{L}^{m}$, that is $\mathcal{L}^{m,0}:=\mathcal{L}^{m}\setminus\{\ensuremath{L}_{m}\}$.
Additionally, we denote the edges that are embedded in facet $\ensuremath{A}_{k}\in\mathcal{A}^{m}$ by the sets $\mathcal{L}_{k}^{m}$ and $\mathcal{L}_{k}^{m,0}$, respectively.
The definitions of the sets $\mathcal{L}^{m}$, $\mathcal{L}^{m,0}$, $\mathcal{L}_{k}^{m}$ and $\mathcal{L}_{k}^{m,0}$ are visualised in Figure~\ref{fig:edgeSets} for the case of a regular hexahedral grid.
For such a grid, $\mathcal{L}^{m}$, $\mathcal{L}^{m,0}$, $\mathcal{L}_{k}^{m}$ and $\mathcal{L}_{k}^{m,0}$ contain \num{13}, \num{12}, \num{4} and \num{3} edges, respectively.
In Section~\ref{subsec:circuitVectorPotForm}, an additional tree and cotree splitting is introduced.
Thereby, edges can either belong to the tree or the cotree.
This motivates the introduction of the subscripts $_{t}$ and $_{c}$ to denote tree and cotree, respectively.
The additional auxiliary sets that are used due to this splitting are denoted by $\mathcal{L}_{k;\text{t}}^{m}$, $\mathcal{L}_{k;\text{c}}^{m}$, $\mathcal{L}_{k;\text{t}}^{m,0}$ and $\mathcal{L}_{k;\text{c}}^{m,0}$.
Lastly, the dual edges that are embedded in the dual facet $\ensuremath{\protect\widetilde{A}}_{m}$ are collected in the set $\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}$.
\begin{figure}[h]
\subfloat[$\mathcal{L}^{m}$\label{fig:edgeSetLm}]{\includegraphics[width=0.249\columnwidth]{./figs/fig6a}}
\subfloat[$\mathcal{L}^{m,0}$\label{fig:edgeSetLm0}]{\includegraphics[width=0.249\columnwidth]{./figs/fig6b}}
\subfloat[$\mathcal{L}_{k}^{m}$\label{fig:edgeSetLm1}]{\includegraphics[width=0.249\columnwidth]{./figs/fig6c}}
\subfloat[$\mathcal{L}_{k}^{m,0}$\label{fig:edgeSetLm10}]{\includegraphics[width=0.249\columnwidth]{./figs/fig6d}}
\caption{Edge $\ensuremath{L}_{m}$ and the four facets collected in $\mathcal{A}^{m}=\{\ensuremath{A}_{1},\dots,\ensuremath{A}_{4}\}$ for a regular hexahedral grid. The highlighted edges illustrate the sets (a) $\mathcal{L}^{m}$, (b) $\mathcal{L}^{m,0}$, (c) $\mathcal{L}_{k}^{m}$ and (d) $\mathcal{L}_{k}^{m,0}$.}
\label{fig:edgeSets}
\end{figure}
\subsection{Circuit Representation of the \textsc{Maxwell}\xspace Grid Equations}
\label{subsec:EMcircuitRepresentation}
In this section, circuit representations of the \gls*{MGEs} given by \eqref{eq:MGEmagCharge}--\eqref{eq:gaugeAfit} are derived.
First, we start with the E-H formulation of~\eqref{eq:MGEmagCharge} to find a corresponding circuit description which is presented in Section~\ref{subsec:circuitFieldForm}.
Subsequently, in Section~\ref{subsec:circuitVectorPotForm} a circuit description based on the E-A formulation of~\eqref{eq:MGE}--\eqref{eq:gaugeAfit} is presented.
For this formulation, a \emph{tree-cotree} decomposition is used.
\subsubsection{Circuit Representation Based on the E-H Formulation}
\label{subsec:circuitFieldForm}
A first electric circuit representing the \gls*{MGEs} is realised from the E-H formulation of~\eqref{eq:MGEmagCharge}.
The goal is to find an expression for the electric voltage on one primal edge such that a circuit stamp for each edge is obtained.
Let $\ensuremath{L}_{m}$ be the edge of interest with its corresponding dual facet $\ensuremath{\protect\widetilde{A}}_{m}$.
We collect all relevant quantities on the grid objects in the neighbourhood of $\ensuremath{L}_{m}$ by using the notation introduced in Section~\ref{subsec:setsEdgesFacets}.
To find the voltage $\ensuremath{\protect\bow{e}}_{m}$ on $\ensuremath{L}_{m}$, let us consider the pair of interlocked facets $\ensuremath{A}_{k}$ and $\ensuremath{\protect\widetilde{A}}_{m}$ as illustrated in Figure~\ref{fig:facetsInterlocked}.
For this part of the grid, it suffices to consider only the $k$-th row of \eqref{eq:MGEmagChargeFaraday} and the $\protect\tilde{m}$-th row of \eqref{eq:MGEmagChargeAmpere} giving
\begin{subequations}
\begin{align}
\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{C}_{kn}\ensuremath{\protect\bow{e}}_{k;n}&=-M_{\nu;\protect\tilde{k} k}^{-1}\frac{\mathrm{d}\ensuremath{\protect\bow{h}}_{\protect\tilde{k}}}{\mathrm{d}t},\label{eq:FaradayOneFacet}\\
\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}\ensuremath{\protect\bow{h}}_{\protect\tilde{m};\protect\tilde{k}}&=
M_{\varepsilon;\protect\tilde{m} m}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t}+M_{\sigma;\protect\tilde{m} m}\ensuremath{\protect\bow{e}}_{m}+\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}.\label{eq:AmpereOneFacet}
\end{align}
\end{subequations}
From~\eqref{eq:FaradayOneFacet}, the magnetic grid voltage $\ensuremath{\protect\bow{h}}_{\protect\tilde{k}}$ allocated at edge $\ensuremath{\protect\widetilde{L}}_{k}$ (which happens to be also embedded in facet $\ensuremath{\protect\widetilde{A}}_{m}$) reads
\begin{equation}
\ensuremath{\protect\bow{h}}_{\protect\tilde{k}}=-M_{\nu;\protect\tilde{k} k}\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{C}_{kn}\int\ensuremath{\protect\bow{e}}_{k;n}\ \mathrm{d} t.
\label{eq:Faradayhfitlock}
\end{equation}
By inserting~\eqref{eq:Faradayhfitlock} in~\eqref{eq:AmpereOneFacet} for all $\ensuremath{\protect\bow{h}}_{\protect\tilde{m};\protect\tilde{k}}$, the voltage $\ensuremath{\protect\bow{e}}_{m}$ on edge $\ensuremath{L}_{m}$ is implicitly given by
\begin{equation*}
-\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\int\ensuremath{\protect\bow{e}}_{k;n}\ \mathrm{d} t=M_{\varepsilon;\protect\tilde{m} m}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t}+M_{\sigma;\protect\tilde{m} m}\ensuremath{\protect\bow{e}}_{m}+\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}.
\end{equation*}
Since $\ensuremath{L}_{m}$ is contained in $\mathcal{L}_{k}^{m}$, we can extract the contribution of $\ensuremath{L}_{m}$ from the sum on the left hand side such that
\begin{multline}
\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{km}\int\ensuremath{\protect\bow{e}}_{m}\,\mathrm{d} t
+\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k}^{m,0}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\int\ensuremath{\protect\bow{e}}_{k;n}\ \mathrm{d} t\\
+M_{\varepsilon;\protect\tilde{m} m}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t}+M_{\sigma;\protect\tilde{m} m}\ensuremath{\protect\bow{e}}_{m}+\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}=0.
\label{eq:KCLfromMGEsLong}
\end{multline}
We further define
\begin{subequations}
\begin{align}
M_{\nu;\protect\tilde{m} m}^{\Sigma}&:=\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{km}=\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}M_{\nu;\protect\tilde{k} k},\\
\ensuremath{\protect\bbow{j}}_{\text{c};\protect\tilde{m}\protect\tilde{k} n}&:=\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\int\ensuremath{\protect\bow{e}}_{k;n}\,\mathrm{d} t.
\label{eq:abbrevCircuit}
\end{align}
\end{subequations}
Thanks to the properties $\ensuremath{\mathbf{C}}\xspace=\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace^{\top}$, $\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}\in\{-1,1\}$ and $\ensuremath{C}_{km}\in\{-1,1\}$, we have that $\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}=\ensuremath{C}_{km}$ and thus their product equals to unity.
However, the product $\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}\ensuremath{C}_{kn}$ can be either $-1$ or $1$ as illustrated by Figure~\ref{fig:facetsInterlocked}.
Thus, we write~\eqref{eq:KCLfromMGEsLong} with the help of~\eqref{eq:abbrevCircuit} compactly as
\begin{equation}
M_{\nu;\protect\tilde{m} m}^{\Sigma}\int\ensuremath{\protect\bow{e}}_{m}\,\mathrm{d} t
+\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k}^{m,0}}\ensuremath{\protect\bbow{j}}_{\text{c};\protect\tilde{m}\protect\tilde{k} n}
+M_{\varepsilon;\protect\tilde{m} m}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t}+M_{\sigma;\protect\tilde{m} m}\ensuremath{\protect\bow{e}}_{m}+\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}=0.
\label{eq:KCLfromMGEs}
\end{equation}
Equation~\eqref{eq:KCLfromMGEs} is the \textsc{Kirchhoff}\xspace's current law (KCL) associated with the primal edge $\ensuremath{L}_{m}$ with $\ensuremath{\protect\bow{e}}_{m}$ representing the voltage drop along the edge.
In fact, according to the definitions of the field quantities in \eqref{eq:KCLfromMGEs}, it is easy to realise that
\begin{itemize}
\item $\ensuremath{\protect\bow{e}}_{m}$ has unit of voltage (\si{V}),
\item $\ensuremath{\protect\bbow{j}}_{\text{i},\protect\tilde{m}}$ has unit of current (\si{A}),
\item $M_{\varepsilon;\protect\tilde{m} m}$ is positive and has unit of capacitance (\si{F}),
\item $M_{\sigma;\protect\tilde{m} m}$ is positive and has unit of conductance (\si{S}),
\item $M_{\nu;\protect\tilde{m} m}^{\Sigma}$ is positive and has unit of reluctance (\si{H\tothe{-1}}),
\item $\ensuremath{\protect\bbow{j}}_{\text{c};\protect\tilde{m}\protect\tilde{k} n}$ has unit of current (\si{A}).
\end{itemize}
Thus, by using \glspl*{VCCS} to model $\ensuremath{\protect\bbow{j}}_{\text{c};\protect\tilde{m}\protect\tilde{k} n}$ accounting for the contributions from neighbouring edges, we can directly represent~\eqref{eq:KCLfromMGEs} with the circuit stamp depicted in Figure~\ref{fig:circuitStampEMelSource}, which preserves the voltage drop between the terminals of $\ensuremath{L}_{m}$.
There are as many of these stamps as primal edges in $G$, and all of them interact via \glspl*{VCCS}.
The concatenation of these elementary stamps constitutes the electric circuit representing the electromagnetic problem at hand.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{./figs/fig7}
\caption{Circuit stamp of a primal edge $\ensuremath{L}_{m}$ with material matrices mapped into RLC lumped elements as stated by \eqref{eq:KCLfromMGEs}. \glspl*{VCCS} account for the interaction with neighbouring edges.}
\label{fig:circuitStampEMelSource}
\end{figure}
In Algorithm~\ref{alg:netlistEM1}, the steps to generate the netlist representing the \gls*{EM} circuit for the entire grid are listed.
A circuit stamp, as shown in Figure~\ref{fig:circuitStampEMelSource}, needs to be created for every edge $\ensuremath{L}_{m}$ in the grid which is realised by a loop in the code.
In each iteration, a resistor, inductor and capacitor with the values taken from the material matrices is added (lines~2--4).
If an impressed current source shall be placed on $\ensuremath{L}_{m}$, an independent current source with a predefined value $\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}$ is used (lines~5--7).
Finally, an inner double loop is required to insert the controlled current sources that model the influence of the edges in the neighbourhood.
For this purpose, we choose to insert \glspl*{CCCS}\footnote{Due to the integral term in the expression for $\ensuremath{\protect\bbow{j}}_{\text{c};\protect\tilde{m}\protect\tilde{k} n}$, a direct translation into \glspl*{VCCS} is not possible. Instead, behavioural sources or \glspl*{CCCS} can be used.} being controlled by the current $M_{\nu;\protect\tilde{n} n}^{\Sigma}\int\ensuremath{\protect\bow{e}}_{k;n}\,\mathrm{d} t$ with a gain of $g_{\protect\tilde{m}\protect\tilde{k} n}^{\text{I}}:=\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}(M_{\nu;\protect\tilde{n} n}^{\Sigma})^{-1}$ (lines~8--12).
By abuse of notation, we denote the controlling device by $\ensuremath{\protect\bow{e}}_{k;n}$.
\begin{algorithm}[t]
\caption{Electromagnetic SPICE netlist generation based on the E-H formulation.}
\label{alg:netlistEM1}
\begin{algorithmic}[1]
\For{edge $\ensuremath{L}_{m}\in G$}
\State write R$m$\quad{}n$m$\quad{}gnd\quad{}$M_{\sigma;\protect\tilde{m} m}^{-1}$
\State write L$m$\quad{}n$m$\quad{}gnd\quad{}$(M_{\nu;\protect\tilde{m} m}^{\Sigma})^{-1}$
\State write C$m$\quad{}n$m$\quad{}gnd\quad{}$M_{\varepsilon;\protect\tilde{m} m}$
\If{an impressed current is placed on $\ensuremath{L}_{m}$}
\State write I$m$\quad{}gnd\quad{}n$m$\quad{}$\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}$
\EndIf
\For{edge $\ensuremath{\protect\widetilde{L}}_{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}$}
\For{edge $\ensuremath{L}_{n}\in\mathcal{L}_{k}^{m,0}$}
\State write F$mn$\quad{}gnd\quad{}n$m$\quad{}$\ensuremath{\protect\bow{e}}_{k;n}$\quad{}$g_{\protect\tilde{m}\protect\tilde{k} n}^{\text{I}}$
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
We remark that a similar analysis on~\eqref{eq:MGEmagCharge} in which magnetic conductivities and sources are considered instead of electric ones can be done\footnote{Although magnetic carriers have not been observed in nature, there may be situations in which one can profit from the inclusion of an equivalent magnetic conductivity in \textsc{Maxwell}\xspace's equations~\cite{Wait_1992aa,Kharzeev_2009aa}.}. This approach would lead to a circuit stamp that is dual to the one in Figure~\ref{fig:circuitStampEMelSource}.
Namely, in this dual stamp, the \gls*{KVL} is guaranteed for each dual edge and $\ensuremath{\protect\bow{h}}_{\protect\tilde{k}}$ represents the electric current in the circuit.
Furthermore, the resulting lumped elements, stemming from the material matrices, are placed in series with the discrete impressed magnetic current $\ensuremath{\protect\bbow{m}}_{\text{i};k}$ that plays the role of an independent voltage source exciting the circuit.
The interaction between the dual stamps is mediated via \glspl*{CCVS}.
In the general case, when both electric and magnetic sources are present, the electric circuit will consist of both the primal and dual stamps, which interact via \glspl*{CCCS} and \glspl*{VCVS}, accordingly.
\subsubsection{Circuit Representation Based on the E-A Formulation}
\label{subsec:circuitVectorPotForm}
In this section, we describe a circuit representation based on the E-A formulation of~\eqref{eq:MGE}--\eqref{eq:gaugeAfit}.
To guarantee uniqueness of the solution, the magnetic potential $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace$ is gauged by means of a \emph{tree}-\emph{cotree} decomposition~\cite{Munteanu_2002aa}.
Therefore, although the inferred circuit stamps are gauge dependent, the solution obtained for
$\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace$ is unique.
We start by considering one primal facet $\ensuremath{A}_{k}\in\mathcal{A}^{m}$ (cf. Figure~\ref{fig:facetsInterlocked}).
For this facet, it suffices to consider the $k$-th row of \eqref{eq:MGEfaraday} and the $\protect\tilde{m}$-th row of \eqref{eq:MGEampere} such that the E-A formulation of the \gls*{MGEs} for a generic edge $\ensuremath{L}_{m}$ reads
\begin{subequations}
\begin{align}
\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{C}_{kn}\ensuremath{\protect\bow{e}}_{k;n}&=-\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{C}_{kn}\frac{\mathrm{d}\ensuremath{\protect\bow{a}}_{k;n}}{\mathrm{d}t},\label{eq:MGElocalRevisitFaraday}\\
\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n}
&=M_{\varepsilon;\protect\tilde{m} m}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t}+M_{\sigma;\protect\tilde{m} m}\ensuremath{\protect\bow{e}}_{m}+\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}.
\label{eq:MGElocalRevisitAmpere}
\end{align}
\label{eq:MGElocalRevisit}\end{subequations}
As in Section~\ref{subsec:circuitFieldForm}, we aim at finding a unique circuit representation of edge $\ensuremath{L}_{m}$.
We start by observing that the system matrix $\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\mathbf{C}}\xspace$ of~\eqref{eq:MGE} is singular\footnote{Note that this is also true for the corresponding continuous operator.}.
This singularity manifests itself in the non-uniqueness of $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace$.
In fact, only the curl of $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace$ is uniquely defined.
As a remedy, we must explicitly impose the gauging~\eqref{eq:gaugeAfit} upon~\eqref{eq:MGE}.
To this end, let us assume that we have constructed a suitable \emph{tree} $G_{\text{t}}$ and a \emph{cotree} $G_{\text{c}}$ out of the primal grid $G$, as exemplified in Figure~\ref{fig:treeCotree}.
Then, we symbolically introduce the orthogonal permutation matrix $\mathbf{P}_{\text{G}}$ to partition $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace$ into its \emph{tree} $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}\in G_{\text{t}}$ and \emph{cotree} $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{c}}\in G_{\text{c}}$ components with $N_{\text{t}}$ and $N_{\text{c}}$ entries, respectively.
\begin{figure}[t]
\centering
\subfloat[\label{fig:facetsInterlocked}]{\includegraphics{./figs/fig8a}}
\hspace{3em}
\subfloat[\label{fig:fitCurlEdge}]{\includegraphics{./figs/fig8b}}
\caption{(a) A pair of interlocked primal and dual facet $\ensuremath{A}_{k}$ and $\ensuremath{\protect\widetilde{A}}_{m}$ for the case of a regular hexahedral grid, respectively. The values of $\ensuremath{C}_{kn}$ and $\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}$ for the edges $\ensuremath{L}_{n}$ of $\ensuremath{A}_{k}$ and the edges $\ensuremath{\protect\widetilde{L}}_{k}$ of $\ensuremath{\protect\widetilde{A}}_{m}$ are annotated. (b) The collection of edges $\ensuremath{L}_{n}$ and $\ensuremath{\protect\widetilde{L}}_{k}$ that contribute to the computation of $\ensuremath{\protect\bow{e}}_{m}$.}
\end{figure}
Symbolically, this reads
\begin{equation}
\mathbf{P}_{\text{G}}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace=
\begin{bmatrix}
\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{c}}\\
\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}
\end{bmatrix}.
\label{eq:treeCotreeSplit}
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{./figs/fig9}
\caption{Example of a \emph{spanning} tree $G_{\text{t}}$ (solid) and cotree $G_{\text{c}}$ (dashed) in a \emph{generic} graph $G$. The case studied here consists of an~\emph{orthogonal} pair of primal and dual grids.}
\label{fig:treeCotree}
\end{figure}
Similarly, we also divide $\mathbf{M}_{\text{G}}$ and $\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace$ into their tree and cotree components by applying the $\mathbf{P}_{\text{G}}$ matrix accordingly, viz.
\begin{equation}
\mathbf{P}_{\text{G}}\mathbf{M}_{\text{G}}\mathbf{P}_{\text{G}}^{\top}=
\begin{bmatrix}
\mathbf{M}_{\text{G}_{\text{c}}} & \ensuremath{\mathbf{0}} \\
\ensuremath{\mathbf{0}} & \mathbf{M}_{\text{G}_{\text{t}}}
\end{bmatrix},
\quad\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace\mathbf{P}_{\text{G}}^{\top}=
\begin{bmatrix}
\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{c}} & \ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{t}}
\end{bmatrix}.
\label{eq:DGsplit}
\end{equation}
The identities of~\eqref{eq:treeCotreeSplit} and~\eqref{eq:DGsplit} together with the orthogonality property of $\mathbf{P}_{\text{G}}$ enable us to rewrite the gauging of~\eqref{eq:gaugeAfit} as
\begin{equation}
\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{c}}\mathbf{M}_{\text{G}_{\text{c}}}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{c}}+
\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{t}}\mathbf{M}_{\text{G}_{\text{t}}}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}=\ensuremath{ \mathbf{F} }.
\label{eq:gaugeAfitSplit}
\end{equation}
Additionally, since $\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{t}}\mathbf{M}_{\text{G}_{\text{t}}}$ is a square and invertible matrix\footnote{These two properties come from two facts: the squareness is a consequence of removing the row associated with the ground node required in circuit analysis.
The invertibility is due to removing the non-null kernel space of the matrix $\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\mathbf{C}}\xspace$ by means of the tree-cotree decomposition.},
we may finally express the tree component $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}$ as
\begin{equation*}
\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}=\mathbf{M}_{\text{G}_{\text{t}}}^{-1}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{t}}^{-1}\ensuremath{ \mathbf{F} }-\mathbf{M}_{\text{G}_{\text{t}}}^{-1}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{t}}^{-1}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{c}}\mathbf{M}_{\text{G}_{\text{c}}}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{c}}.
\end{equation*}
The column vector $\ensuremath{ \mathbf{F} }$, which represents the grid counterpart of the scalar function $f$ quantifying the divergence of $\ensuremath{\mathbf{A}}$, is of free choice.
Therefore, we conveniently choose $\ensuremath{ \mathbf{F} }=\ensuremath{\mathbf{0}}$ impressing the \textsc{Coulomb}\xspace gauge~\cite{Balanis_2012aa} to straightforwardly arrive at
\begin{equation}
\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}=-\underbrace{\mathbf{M}_{\text{G}_{\text{t}}}^{-1}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{t}}^{-1}\ensuremath{\protect\widetilde{\mathbf{S}}}\xspace_{\mathbbm{1}\text{c}}\mathbf{M}_{\text{G}_{\text{c}}}}_{\mathbf{E}_{\text{tc}}}\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{c}}.
\label{fit_cir_eq24}
\end{equation}
The matrix $\mathbf{E}_{\text{tc}}\in\mathbb{R}^{N_{\text{t}}\times N_{\text{c}}}$ is known as the \emph{essential} incidence matrix~\cite{Munteanu_2002aa} and establishes a direct relation between the tree and cotree components $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}$ and $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{c}}$, respectively\footnote{Owing to this property, we identify in the rows of $\mathbf{E}_{\text{tc}}\in\mathbb{R}^{N_{\text{t}}\times N_{\text{c}}}$ the collection of all fundamental cut-sets associated with the selected tree and cotree.
We recall that a fundamental cut-set is a set formed by the union of a single tree edge and the unique set of adjoining cotree edges. In this manner, the fundamental cut-sets are used to express \textsc{Kirchhoff}\xspace's current law in the general form of~\eqref{fit_cir_eq24}.}.
By inspection and expansion of~\eqref{fit_cir_eq24}, we express the $m$-th component of the column vector $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace_{\text{t}}$ as
\begin{equation}
\ensuremath{\protect\bow{a}}_{\text{t};m}=-\sum_{n\in G_{\text{c}}}E_{\text{tc};mn}\ensuremath{\protect\bow{a}}_{\text{c};n},
\label{eq:afittm}\end{equation}
with
\begin{equation*}
E_{\text{tc};mn}:=\sum_{i\in G}M_{G_{\text{t}};m\protect\tilde{m}}^{-1}\ensuremath{\protect\widetilde{S}}_{\mathbbm{1}\text{t};\protect\tilde{m}\protect\tilde{i}}^{-1}\ensuremath{\protect\widetilde{S}}_{\mathbbm{1}\text{c};\protect\tilde{i}\protect\tilde{n}} M_{G_{\text{c}};\protect\tilde{n} n}
\end{equation*}
and $i$ spanning over the nodes in $G$.
In this manner, \eqref{eq:afittm} removes the redundancy associated with $\ensuremath{\protect\bow{\mathrm{\mathbf{a}}}}\xspace$ and guarantees a unique solution of~\eqref{eq:MGElocalRevisit}.
Let us now get back to the generic primal edge $\ensuremath{L}_{m}$ once again.
As introduced in Section~\ref{subsec:setsEdgesFacets}, we use specific sets to refer to edges in the tree and cotree denoted by $\mathcal{L}_{k;\text{t}}^{m}$, $\mathcal{L}_{k;\text{t}}^{m,0}$ and $\mathcal{L}_{k;\text{c}}^{m}$, $\mathcal{L}_{k;\text{c}}^{m,0}$, respectively.
Isolating the voltage $\ensuremath{\protect\bow{e}}_{m}$ along edge $\ensuremath{L}_{m}$ embedded in facet $\ensuremath{A}_{k}$ from~\eqref{eq:MGElocalRevisitFaraday}, we arrive at
\begin{equation*}
\ensuremath{\protect\bow{e}}_{m}+\sum_{n\in\mathcal{L}_{k}^{m,0}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\ensuremath{\protect\bow{e}}_{k;n}
=-\sum_{n\in\mathcal{L}_{k}^{m}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\frac{\mathrm{d}\ensuremath{\protect\bow{a}}_{k;n}}{\mathrm{d}t}.
\end{equation*}
We may then take the sum over all facets $\ensuremath{A}_{k}\in\mathcal{A}^{m}$ (cf. Figure~\ref{fig:fitCurlEdge}). Then, by splitting $\lbrace\ensuremath{\protect\bow{a}}_{k;n}\rbrace$ into tree and cotree components, we arrive at
\begin{multline}
\ensuremath{\protect\bow{e}}_{m}+\frac{1}{N_{\text{F};m}}\sum_{k\in\mathcal{A}^{m}}\sum_{n\in\mathcal{L}_{k}^{m,0}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\ensuremath{\protect\bow{e}}_{k;n}\\
\quad=-\frac{1}{N_{\text{F};m}}\sum_{k\in\mathcal{A}^{m}}\sum_{n\in\mathcal{L}_{k;\text{c}}^{m}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\frac{\mathrm{d}\ensuremath{\protect\bow{a}}_{k;n}}{\mathrm{d}t}
-\frac{1}{N_{\text{F};m}}\sum_{k\in\mathcal{A}^{m}}\sum_{n\in\mathcal{L}_{k;\text{t}}^{m}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\frac{\mathrm{d}\ensuremath{\protect\bow{a}}_{k;n}}{\mathrm{d}t},
\label{eq:MGElocalSplitAfit}
\end{multline}
where $N_{\text{F};m}$ is the number of facets in $G$ containing the edge $\ensuremath{L}_{m}$.
Note that $N_{\text{F};m}=4$ for a regular hexahedral grid.
Let us now introduce the auxiliary definitions
\begin{equation*}
V_{e;m}:=\sum_{k\in\mathcal{A}^{m}}\sum_{n\in\mathcal{L}_{k}^{m,0}}V_{e;mkn},\quad
V_{\text{c};m}:=\sum_{k\in\mathcal{A}^{m}}\sum_{n\in\mathcal{L}_{k;\text{c}}^{m}}V_{\text{c};mkn},\quad
V_{\text{t};m}:=\sum_{k\in\mathcal{A}^{m}}\sum_{n\in\mathcal{L}_{k;\text{t}}^{m}}V_{\text{t};mkn},
\end{equation*}
where
\begin{align*}
V_{e;mkn}&:=\frac{1}{N_{\text{F};m}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\ensuremath{\protect\bow{e}}_{k;n},\\
V_{\text{c};mkn}&:=\frac{1}{N_{\text{F};m}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\frac{\mathrm{d}\ensuremath{\protect\bow{a}}_{k;n}}{\mathrm{d}t},\\
V_{\text{t};mkn}&:=\frac{1}{N_{\text{F};m}}\left(\frac{\ensuremath{C}_{kn}}{\ensuremath{C}_{km}}\right)\frac{\mathrm{d}\ensuremath{\protect\bow{a}}_{k;n}}{\mathrm{d}t},
\end{align*}
which allow to express \eqref{eq:MGElocalSplitAfit} as
\begin{equation}
\ensuremath{\protect\bow{e}}_{m}+V_{e;m}+V_{\text{c};m}+V_{\text{t};m}=0.
\label{eq:MGElocalSplitAfitCompact}
\end{equation}
Similarly, we expand the left-hand side of~\eqref{eq:MGElocalRevisitAmpere} as
\begin{multline*}
\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n}\\
=\left(\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{km}\right)\ensuremath{\protect\bow{a}}_{m}
+\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k}^{m,0}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n},
\end{multline*}
which, upon substitution in~\eqref{eq:MGElocalRevisitAmpere} and by splitting $\lbrace\ensuremath{\protect\bow{a}}_{k;n}\rbrace$ into tree and cotree components, yields
\begin{multline}
\ensuremath{\protect\bow{a}}_{m}+\frac{\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k;\text{c}}^{m,0}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}
+\frac{\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k;\text{t}}^{m,0}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}\\
-\frac{M_{\varepsilon;\protect\tilde{m} m}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t}-\frac{M_{\sigma;\protect\tilde{m} m}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}\ensuremath{\protect\bow{e}}_{m}-\frac{\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}=0,
\label{eq:MGEampereAfit2cir}
\end{multline}
where
\begin{equation*}
M_{\nu;\protect\tilde{m} m}^{\Sigma}:=\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{km}=\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}M_{\nu;\protect\tilde{k} k}
\end{equation*}
since $\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}=\ensuremath{C}_{km}$.
By means of the auxiliary definitions
\begin{alignat*}{3}
I_{R;m}&:=\frac{M_{\sigma;\protect\tilde{m} m}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}\ensuremath{\protect\bow{e}}_{m},\qquad
&&I_{\text{c};m}&&:=\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k;\text{c}}^{m,0}}I_{\text{c};m\protect\tilde{k} n},\\
I_{C;m}&:=\frac{M_{\varepsilon;\protect\tilde{m} m}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}}\frac{\mathrm{d}\ensuremath{\protect\bow{e}}_{m}}{\mathrm{d}t},
&&I_{\text{t};m}&&:=\sum_{\protect\tilde{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}}\sum_{n\in\mathcal{L}_{k;\text{t}}^{m,0}}I_{\text{t};m\protect\tilde{k} n},\\
I_{\text{i};m}&:=\frac{\ensuremath{\protect\bbow{j}}_{\text{i};\protect\tilde{m}}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}},\\
\end{alignat*}
where
\begin{equation*}
I_{\text{c};m\protect\tilde{k} n}:=\frac{\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}},\quad
I_{\text{t};m\protect\tilde{k} n}:=\frac{\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}\ensuremath{\protect\bow{a}}_{k;n}}{M_{\nu;\protect\tilde{m} m}^{\Sigma}},
\end{equation*}
we express~\eqref{eq:MGEampereAfit2cir} as
\begin{equation}
\ensuremath{\protect\bow{a}}_{m}+I_{\text{c};m}+I_{\text{t};m}-I_{C;m}-I_{R;m}-I_{\text{i};m}=0.
\label{eq:MGEampereAfit2cirCompact}
\end{equation}
\begin{figure}[t]
\centering
\subfloat[\label{fig:circuitStampAfitCotree}]{\includegraphics[width=\columnwidth]{figs/fig10a}
}\\
\subfloat[\label{fig:circuitStampAfitTree}]{\includegraphics[width=\columnwidth]{figs/fig10b}
}
\caption{Circuit stamp of (a) a cotree edge and (b) a tree edge.}
\end{figure}
We observe that~\eqref{eq:MGElocalSplitAfitCompact} and~\eqref{eq:MGEampereAfit2cirCompact} can be interpreted as the \gls*{KVL} and \gls*{KCL} of an arbitrary primal edge $\ensuremath{L}_{m}$ provided that $\ensuremath{\protect\bow{e}}_{m}$ and $\ensuremath{\protect\bow{a}}_{m}$ represent the sought voltage and current, respectively.
As a matter of fact, we observe that
\begin{itemize}
\item $\ensuremath{\protect\bow{e}}_{m}$ has unit of \si{V} and represents the voltage drop along $\ensuremath{L}_{m}$ that is either in the tree or cotree set.
\item $\ensuremath{\protect\bow{a}}_{m}$ has unit of \textsc{Weber}\xspace (\si{Wb}) and represents the current along $\ensuremath{L}_{m}$. If $\ensuremath{L}_{m}\in G_{\text{c}}$, this current is a degree of freedom. Otherwise, it is modelled by a \gls*{CCCS} as stated in~\eqref{eq:afittm}. For compact notation, we label this \gls*{CCCS} as ${I_{\Sigma;m}:=\sum_{n\in G_{\text{c}}}I_{\Sigma;mn}}$, where ${I_{\Sigma;mn}:=-E_{\text{tc};mn}\ensuremath{\protect\bow{a}}_{\text{c};n}}$.
\item $M_{\nu;\protect\tilde{m} m}^{\Sigma}$ has unit of \si{H\tothe{-1}} and is expected to be positive since $M_{\nu;\protect\tilde{k} k}>0$ for all materials. This term scales the displacement, conduction and impressed current as seen in~\eqref{eq:MGEampereAfit2cir}.
\end{itemize}
With these observations, we may depict~\eqref{eq:MGElocalSplitAfit} and~\eqref{eq:MGEampereAfit2cir} by means of the circuit stamps of Figure~\ref{fig:circuitStampAfitCotree} and Figure~\ref{fig:circuitStampAfitTree} for a cotree and tree edge, respectively.
As we can see therein, $\ensuremath{\protect\bow{a}}_{m}$ is regarded as the electric current along the primal edge $\ensuremath{L}_{m}$, while the voltage drop between its terminals is established by the \gls*{VCVS} $V_{e;m}$ and the \glspl*{CCVS} $V_{\text{c};m}$ and $V_{\text{t};m}$ that mediate the interaction with neighbouring edges.
When the edge $\ensuremath{L}_{m}$ belongs to the tree $G_{\text{t}}$, then the current $\ensuremath{\protect\bow{a}}_{m}$ is modelled by the \gls*{CCCS} $I_{\Sigma;m}$ as demanded by~\eqref{eq:afittm}.
Otherwise, for cotree edges, $\ensuremath{\protect\bow{a}}_{m}$ is a degree of freedom.
The current $\ensuremath{\protect\bow{a}}_{m}$ is then split into several branches where a resistance, capacitance, the impressed current source, and \glspl*{CCCS} are connected.
The concatenation of these fundamental stamps forms the electric circuit representing the magnetic vector potential formulation of the electromagnetic problem.
In Algorithm~\ref{alg:netlistEM2}, we show the pseudocode to generate the netlist of the circuit stamps depicted in Figures~\ref{fig:circuitStampAfitCotree} and \ref{fig:circuitStampAfitTree}.
The iteration over all edges $\ensuremath{L}_{m}$ in the grid is done in the outermost loop.
In each iteration, one resistor and capacitor connecting the node $n_{m}$ to ground (gnd) must be added to the netlist (lines~2 and 3).
If an impressed current is present at $\ensuremath{L}_{m}$, an independent current source with value $I_{\text{i};m}$ is added between $n_{m}$ and gnd (lines~4--6).
For every tree edge, the currents $I_{\Sigma;mn}$ from the cotree edges in the neighbourhood are modelled by a parallel connection of \glspl*{CCCS} between $n_{e;m}$ and $n_{m}$ (lines~7--11).
Using a double loop, the inductive currents $I_{\text{t};m\protect\tilde{k} n}$ and $I_{\text{c};m\protect\tilde{k} n}$ from the tree and cotree branches, respectively, are added as a parallel connection of \glspl*{CCCS} between $n_{m}$ and gnd (lines~12--19).
The controlling current is given by the current through device FIc$n11$ and the gain is ${g_{m\protect\tilde{k} n}^{\text{I}}=\ensuremath{\protect\widetilde{C}}_{\protect\tilde{m}\protect\tilde{k}}M_{\nu;\protect\tilde{k} k}\ensuremath{C}_{kn}(M_{\nu;\protect\tilde{m} m}^{\Sigma})^{-1}}$.
Finally, the voltages $V_{e;mkn}$, $V_{\text{c};mkn}$ and $V_{\text{t};mkn}$ are added as a series connection between $n_{e;m}$ and gnd with intermediate nodes indexed by $k$, $n$ and $\hat{n}=n+1$ (lines~20--30).
For the voltage $V_{e;mkn}$, the voltage between node n$n$ and gnd controls a \gls*{VCVS} with a gain of $g_{mkn}^{\text{V}}=N_{\text{F};m}^{-1}\ensuremath{C}_{kn}\ensuremath{C}_{km}^{-1}$
On the other hand, the voltages $V_{\text{c};mkn}$ and $V_{\text{t};mkn}$ are added as behavioural sources using DDT as the SPICE\xspace syntax for time derivatives and a gain of $g_{mkn}^{\text{I}}=N_{\text{F};m}^{-1}\ensuremath{C}_{kn}\ensuremath{C}_{km}^{-1}$.
Note that for a cotree edge, the node $n_{e;m}$ coincides with the node $n_{m}$.
\begin{algorithm}
\caption{Electromagnetic SPICE netlist generation based on the E-A formulation.}
\label{alg:netlistEM2}
\begin{algorithmic}[1]
\For{edge $\ensuremath{L}_{m}\in G$}
\State write R$m$\quad{}n$m$\quad{}gnd\quad{}$M_{\nu;\protect\tilde{m} m}^{\Sigma}M_{\sigma;\protect\tilde{m} m}^{-1}$
\State write C$m$\quad{}n$m$\quad{}gnd\quad{}$M_{\varepsilon;\protect\tilde{m} m}(M_{\nu;\protect\tilde{m} m}^{\Sigma})^{-1}$
\If{an impressed current is placed on $\ensuremath{L}_{m}$}
\State write I$m$\quad{}n$m$\quad{}gnd\quad{}$I_{\text{i};m}$
\EndIf
\If{$\ensuremath{L}_{m}\in G_{\text{t}}$}
\For{edge $\ensuremath{L}_{n}\in G_{\text{c}}$}
\State write FIsum$mn$\quad{}ne$m$\quad{}n$m$\quad{}FIc$n11$\quad{}$-E_{\text{tc};mn}$
\EndFor
\EndIf
\For{edge $\ensuremath{\protect\widetilde{L}}_{k}\in\protect\widetilde{\mathcal{L}}_{\protect\tilde{m}}^{m}$}
\For{edge $\ensuremath{L}_{n}\in\mathcal{L}_{k;\text{c}}^{m,0}$}
\State write FIc$mkn$\quad{}gnd\quad{}n$m$\quad{}FIc$n11$\quad{}$g_{m\protect\tilde{k} n}^{\text{I}}$
\EndFor
\For{edge $\ensuremath{L}_{n}\in\mathcal{L}_{k;\text{t}}^{m,0}$}
\State write FIt$mkn$\quad{}gnd\quad{}n$m$\quad{}FIc$n11$\quad{}$g_{m\protect\tilde{k} n}^{\text{I}}$
\EndFor
\EndFor
\For{edge $\ensuremath{A}_l\in\mathcal{A}^{m}$ with $k=1,\dots,K$}
\For{edge $\ensuremath{L}_n\in\mathcal{L}_{k}^{m,0}$ with $n=1,\dots,N-1$}
\State write EVe$mkn$\;\;{}ne$mk\hat{n}$\;\;{}ne$mkn$\;\;{}n$n$\;\;{}gnd\;\;{}$g_{mkn}^{\text{V}}$
\EndFor
\For{edge $\ensuremath{L}_n\in\mathcal{L}_{k;\text{c}}^{m}$ with $n=1,\dots,N-1$}
\State write BVc$mkn$\quad{}nc$mk\hat{n}$\quad{}nc$mkn$\quad{}\dots\\\hspace{17em}V=$g_{mkn}^{\text{I}}\text{DDT}(\ensuremath{\protect\bow{a}}_{k;n})$
\EndFor
\For{edge $\ensuremath{L}_n\in\mathcal{L}_{k;\text{t}}^{m}$ with $n=1,\dots,N-1$}
\State write BVt$mkn$\quad{}nt$mk\hat{n}$\quad{}nt$mkn$\quad{}\dots\\\hspace{17em}V=$g_{mkn}^{\text{I}}\text{DDT}(\ensuremath{\protect\bow{a}}_{k;n})$
\EndFor
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
We remark that a similar analysis on~\eqref{eq:MaxwellContInt} by considering only magnetic conductivities and sources instead of electric ones is also possible.
Thus, with the auxiliary electric potential $\ensuremath{\mathbf{D}}=\nabla \times\ensuremath{\mathbf{F}}$ and $\nabla \cdot\ensuremath{\mathbf{F}}=a$, where $a$ is an arbitrary gauging function, circuit stamps which are dual to those shown in Figure~\ref{fig:circuitStampAfitCotree} and Figure~\ref{fig:circuitStampAfitTree} can be found.
In these dual stamps, $\ensuremath{\protect\bow{h}}_{m}$ represents an electric current while $\ensuremath{\protect\bow{f}}_{m}$, namely the grid counterpart of $\ensuremath{\mathbf{F}}$, would be regarded as a voltage drop.
Finally, if both electric and magnetic sources were present, then the entire circuit would consist of the aggregate of interacting primal and dual stamps.
\subsection{Absorbing Boundary Conditions}
\label{subsec:theoryABC}
In many electromagnetic field simulation set-ups, the computational domain must be bounded.
To simulate free wave propagation, one must impose according conditions for the fields at the boundaries of the domain.
These conditions are known as \glspl*{ABC}~\cite{Engquist_1977aa} and aim at minimising (ideally cancelling) unphysical incoming reflections.
At the boundary of the domain, a distinction is made between normal and tangential components and between longitudinal and transverse derivatives\footnote{In this regard, the longitudinal (transverse) derivative coincides with the normal (tangent) derivative at the boundary.}.
For starters, let $\partial_{r}$, $\nabla_{\text{t}}$ and $\partial_{t}$ denote the longitudinal, transversal and temporal derivative operators, respectively.
Furthermore, within the context of this analysis, we define $\nabla:=\left(\partial^{2}_{r}+\nabla_{\text{t}}^{2}\right)^{1/2}$.
We start the realisation of circuit stamps for \glspl*{ABC} by considering the time-domain wave equation for the electric field $\ensuremath{\mathbf{E}}$ in a homogeneous and isotropic medium\footnote{Although we restrict ourselves to homogeneous and isotropic media, the analysis can also be extended to more general cases.}, viz.
\begin{equation}
\nabla^{2}\ensuremath{\mathbf{E}}=\mu\varepsilon\frac{\partial^2\ensuremath{\mathbf{E}}}{\partial t^2}.
\label{eq:waveEqTime}
\end{equation}
Equation~\eqref{eq:waveEqTime} can be expanded in terms of the so-called travelling wave operators as
\begin{equation}
\left(\nabla+\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\left(\nabla-\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\ensuremath{\mathbf{E}}=\mathbf{0}.
\label{eq:waveEqTimeExpand}
\end{equation}
Above, we observe that in an arbitrary point in space, the field $\ensuremath{\mathbf{E}}$ can in general be considered as the superposition of an inward and outward travelling wave $\ensuremath{\mathbf{E}}^{+}$ and $\ensuremath{\mathbf{E}}^{-}$, respectively, viz.
\begin{equation*}
\ensuremath{\mathbf{E}}=\ensuremath{\mathbf{E}}^{+}+\ensuremath{\mathbf{E}}^{-},
\end{equation*}
which upon substitution in \eqref{eq:waveEqTimeExpand} straightforwardly leads to the following set of travelling wave equations,
\begin{equation*}
\left(\nabla+\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\left(\nabla-\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\ensuremath{\mathbf{E}}^{+}=\mathbf{0},\quad
\left(\nabla+\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\left(\nabla-\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\ensuremath{\mathbf{E}}^{-}=\mathbf{0},
\end{equation*}
inasmuch as both wave components $\ensuremath{\mathbf{E}}^{\pm}$ are linearly independent.
Owing to their definition, the wave components $\ensuremath{\mathbf{E}}^{\pm}$ satisfy independently and simultaneously the following\footnote{This property can be easily visualised if we consider, for a moment, one-dimensional wave propagation along the $x$-axis. In this circumstance, we have that $\nabla\equiv\partial_{x}$ and $e^{\phi\left(x\pm\nu t\right)}$ representing backward and forward travelling waves
at speed $\nu=1/\sqrt{\mu\varepsilon}$ with $\phi$ an arbitrary function.},
\begin{equation}
\left(\nabla\mp\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\ensuremath{\mathbf{E}}^{\pm}=\mathbf{0},
\label{eq:inoutwardsim}
\end{equation}
which tells us explicitly that the outward (inward) travelling wave operator cancels out the inward (outward) travelling wave at any point of interest.
Now, let us get back to~\eqref{eq:waveEqTimeExpand} to further expand the operators therein to obtain
\begin{equation}
\left(\frac{\partial}{\partial r}+\frac{\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,}{\sqrt{1+\frac{\nabla_{\text{t}}^{2}}{\frac{\partial^2}{\partial r^2}}}}\right)
\left(\frac{\partial}{\partial r}-\frac{\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,}{\sqrt{1+\frac{\nabla_{\text{t}}^{2}}{\frac{\partial^2}{\partial r^2}}}}\right)\ensuremath{\mathbf{E}}=\mathbf{0}.
\label{eq:waveEqTimeExpandMore}
\end{equation}
As we can see, the above wave operators entail the calculation of the inverse of $\sqrt{1+\nabla_{\text{t}}^{2}/\partial_{r}^{2}}$, which generally translates into a global integral operator~\cite{Zhukovsky_2014aa}.
In principle, the expansion of this integral operator around the observation point yields the exact explicit representation of the travelling wave operators in \eqref{eq:waveEqTimeExpandMore}.
However, this approach is contrary to the idea of realising simple and efficient \glspl*{ABC} for circuit simulations.
As a remedy, we may expand $1/\sqrt{1+\nabla_{\text{t}}^{2}/\partial_{r}^{2}}$ as in a Taylor series to yield
\begin{equation}
\begin{split}
&\left(\frac{\partial}{\partial r}+\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}-\frac{1}{2}\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\frac{\nabla_{\text{t}}^{2}}{\frac{\partial^2}{\partial r^2}}+\ensuremath{{\mathcal{O}}}\left(\nabla_{\text{t}}^{4}\right)\right)\\
&\left(\frac{\partial}{\partial r}-\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}+\frac{1}{2}\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\frac{\nabla_{\text{t}}^{2}}{\frac{\partial^2}{\partial r^2}}+\ensuremath{{\mathcal{O}}}\left(\nabla_{\text{t}}^{4}\right)\right)\ensuremath{\mathbf{E}}=\mathbf{0}.
\end{split}
\label{eq:waveEqTimeExpandMoreSimple}\end{equation}
The above equation holds for any component of $\ensuremath{\mathbf{E}}$.
Furthermore, if the field $\ensuremath{\mathbf{E}}$ propagates in~\emph{free} space, we may assume that the operation $\nabla^{2}_{\text{t}}$ is negligible in the neighbourhood of an observation point within the spherical wavefront.
Hence, we may rewrite~\eqref{eq:waveEqTimeExpandMoreSimple} in terms of simplified inward and outward travelling wave propagators as
\begin{equation}
\left(\frac{\partial}{\partial r}+\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)
\left(\frac{\partial}{\partial r}-\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\ensuremath{\mathbf{E}}=\mathbf{0},
\label{eq:waveEqTimeExpandMoreSimpleMore}
\end{equation}
since the general expression of $\ensuremath{\mathbf{E}}$ in~\eqref{eq:waveEqTimeExpandMoreSimpleMore} admits the superposition of inward and outward travelling waves $\ensuremath{\mathbf{E}}^{+}$ and $\ensuremath{\mathbf{E}}^{-}$.
Then, to minimise unwanted incoming reflections at a certain boundary given by $r=r_{\text{B}}$, we may impose on $\ensuremath{\mathbf{E}}$ the condition
\begin{equation}
\left.\left(\frac{\partial}{\partial r}-\sqrt{\mu\varepsilon}\frac{\partial}{\partial{}t}\,\right)\ensuremath{\mathbf{E}}\right|_{r=r_{\text{B}}}=\mathbf{0},
\label{eq:EngquistMadja}
\end{equation}
in agreement with~\eqref{eq:inoutwardsim}.
The condition given by~\eqref{eq:EngquistMadja} is a first-order \gls*{ABC} known as Engquist-Madja condition~\cite{Engquist_1977aa}.
It is a local condition because it is evaluated pointwise taking only into account the wave component propagating perpendicularly to the boundary.
Thereby, the condition in~\eqref{eq:EngquistMadja} states that a practical transparent boundary at $r=r_{\text{B}}$ can be realised if it is guaranteed that the phase-amplitude of the field on the boundary at a certain time $t_{\text{B}}$ is equal to that one the field had at some previous instant $t_{\text{B}}-\Delta t$ at a point $r=r_{\text{B}}-\Delta t/\sqrt{\mu\varepsilon}$.
The practical relevance of~\eqref{eq:EngquistMadja} comes from its simplicity.
Let us now interpret~\eqref{eq:EngquistMadja} within the context of the \textsc{Maxwell}\xspace grid equations in order to realise circuit stamps associated with the \glspl*{ABC} of~\eqref{eq:EngquistMadja}.
To this end, let us consider the grid wave equation for the electric grid voltage $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace$, which for time-independent constitutive parameters and non-conducting source-free regions can be obtained from the grid curl equations of~\eqref{eq:MGEmagCharge}, viz.
\begin{equation*}
\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\mathbf{C}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace=-\ensuremath{\mathbf{M}_{\varepsilon}}\xspace\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace.
\end{equation*}
With the structure of the matrices \ensuremath{\mathbf{C}}\xspace, \ensuremath{\mathbf{M}_{\varepsilon}}\xspace and \ensuremath{\mathbf{M}_{\nu}}\xspace given by \eqref{eq:fitMatBlockStruct} and the subdivision $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace=\left(\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{x},\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{y},\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{z}\right)^{\top}$, the expression
\begin{equation}
\left(\ensuremath{\mathbf{P}}\xspace_{z}^{\top}\ensuremath{ \mathbf{M} }_{\nu;y}\ensuremath{\mathbf{P}}\xspace_{z}
+\ensuremath{\mathbf{P}}\xspace_{y}^{\top}\ensuremath{ \mathbf{M} }_{\nu;z}\ensuremath{\mathbf{P}}\xspace_{y}\right)\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{x}
-\ensuremath{\mathbf{P}}\xspace_{y}^{\top}\ensuremath{ \mathbf{M} }_{\nu;z}\ensuremath{\mathbf{P}}\xspace_{x}\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{y}
-\ensuremath{\mathbf{P}}\xspace_{z}^{\top}\ensuremath{ \mathbf{M} }_{\nu;y}\ensuremath{\mathbf{P}}\xspace_{x}\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{z}
=-\ensuremath{ \mathbf{M} }_{\varepsilon;x}\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{x}
\label{eq:waveEqMGEoneComp}
\end{equation}
for the component $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{x}$ is obtained.
Similar expressions can be also obtained for the other two components $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{y}$ and $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{z}$.
The grid counterpart of the \gls*{ABC} in~\eqref{eq:EngquistMadja} is obtained from~\eqref{eq:waveEqMGEoneComp} by extracting the grid wave equation associated with a generic primal edge $\ensuremath{L}_{x;m}\in G$ oriented along the $x$-direction.
This yields
\begin{multline}
\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}_{x;n}
+\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{y;\protect\tilde{m}\protect\tilde{k}}M_{\nu;z;\protect\tilde{k} k}\ensuremath{P}_{y;kn}\ensuremath{\protect\bow{e}}_{x;n}\\
-\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{y;\protect\tilde{m}\protect\tilde{k}}M_{\nu;z;\protect\tilde{k} k}\ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}_{y;n}
-\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k}\ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}_{z;n}
=-M_{\varepsilon;x;\protect\tilde{m} m}\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{e}}_{x;m},
\label{eq:gridwaveEq}
\end{multline}
with $n$ and $\protect\tilde{k}$ spanning over the corresponding primal and dual edges oriented along the indicated directions.
We make the following observations upon the above grid wave equation.
\begin{itemize}
\item Two main grid wave components contribute to the time variation of $\ensuremath{\protect\bow{e}}_{x;m}$ along $\ensuremath{L}_{x;m}$. The first of these grid wave components propagates along the $z$-direction and stems from the spatial variation of $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{x}$ along this direction as stated by the grid derivatives in $\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k} \ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}_{x;n}$. The second one propagates along the $y$-direction and stems from the spatial variation of $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{x}$ along this direction as stated by $\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{y;\protect\tilde{m}\protect\tilde{k}}M_{\nu;z;\protect\tilde{k} k}\ensuremath{P}_{y;kn}\ensuremath{\protect\bow{e}}_{x;n}$.
\item Two secondary grid wave components contribute to the time variation of $\ensuremath{\protect\bow{e}}_{x;m}$ along $\ensuremath{L}_{x;m}$. The first one propagates along the $y$-direction and stems from the spatial variation of $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{y}$ along the $x$-direction as stated by $\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{y;\protect\tilde{m}\protect\tilde{k}}M_{\nu;z;\protect\tilde{k} k} \ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}_{y;n}$. The second one propagates along the $z$-direction and stems from the spatial variation of $\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace_{z}$ along the $x$-direction as stated by $\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k} \ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}_{z;n}$.
\item Owing to both the structure of~\eqref{eq:gridwaveEq} and the grid-like embedding where the propagation takes place, we may apply superposition to treat each propagation direction separately. Therefore, each grid wave component satisfies its own one-dimensional grid wave equation.
\end{itemize}
\begin{figure}
\centering
\includegraphics{./figs/fig11}
\caption{The \gls*{ABC} is imposed at edge $\ensuremath{L}_{x;m}$ in the grid $G$. The impinging propagating grid wave coming from edge $\ensuremath{L}_{x;m-M_{z}}$ is then absorbed. The two facets $\ensuremath{A}_{y;k-M_{z}}$ and $\ensuremath{A}_{y;k}$ that contribute to the averaging of $\overline{M}_{\nu;y;\protect\tilde{k} k}$ are shown.}
\label{fig:waveBndEdges}
\end{figure}
We construct the grid counterpart of~\eqref{eq:EngquistMadja} by adhering to the same principle that led to it. Namely, we take only into account grid wave propagation perpendicular to the boundary of interest.
Thereby, let us assume that our boundary is located for example at $z=z_{\text{B}}$.
Thus, by invoking the superposition principle, we may write the grid wave equation associated with the component $\ensuremath{\protect\bow{e}}^{(z)}_{x;m}$ that propagates along the $z$-direction as
\begin{equation}
\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}^{(z)}_{x;n}
-\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k}\ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}^{(x)}_{z;n}
=-M_{\varepsilon;x;\protect\tilde{m} m}\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{e}}^{(z)}_{x;m},
\label{eq:simpgridwaveEq1}
\end{equation}
where we have used the labels $^{(z)}$ and $^{(x)}$ to explicitly indicate directions of propagation.
By considering that for sufficiently well-refined grids $G$ and $\protect\widetilde{G}$ we may expect $M_{\nu;y;\protect\tilde{k} k}$ to not vary significantly for the two relevant facets $\ensuremath{A}_{y;k}$ (cf. Figure~\ref{fig:waveBndEdges}), we may define an average value $\overline{M}_{\nu;y;k\protect\tilde{k}}$ that enables us to write the leftmost term on the left-hand side of \eqref{eq:simpgridwaveEq1} as
\begin{equation*}
\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}M_{\nu;y;\protect\tilde{k} k}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}^{(z)}_{x;n}
=\overline{M}_{\nu;y;\protect\tilde{k} k}\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{P}^{\top}_{z;\protect\tilde{m}\protect\tilde{k}}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}^{(z)}_{x;n},
\end{equation*}
with a similar result for the rightmost term on the left-hand side of \eqref{eq:simpgridwaveEq1}.
Thus, together with the property $\ensuremath{P}^{\top}_{\xi;nk}=-\ensuremath{\widetilde{P}}_{\xi;nk}$, we may proceed to write \eqref{eq:simpgridwaveEq1} as
\begin{equation*}
\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{\widetilde{P}}_{z;\protect\tilde{m}\protect\tilde{k}}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}^{(z)}_{x;n}
-\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{\widetilde{P}}_{z;\protect\tilde{m}\protect\tilde{k}}\ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}^{(x)}_{z;n}
=\overline{M}_{\nu;y;\protect\tilde{k} k}^{-1}M_{\varepsilon;x;\protect\tilde{m} m}\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{e}}^{(z)}_{x;m}.
\end{equation*}
Above, the term $\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{\widetilde{P}}_{z;\protect\tilde{m}\protect\tilde{k}}\ensuremath{P}_{x;kn}\ensuremath{\protect\bow{e}}^{(x)}_{z;n}$ is the grid counterpart of the continuous operator in the wave equation of \eqref{eq:waveEqTime}.
Analogously, we may assume that in the vicinity of $\ensuremath{L}_{x;m}$ at the boundary $z=z_{\text{B}}$, the wavefront of the impinging grid wave is plane.
Therefore, we may neglect transverse variations and we arrive at
\begin{equation*}
\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{\widetilde{P}}_{z;\protect\tilde{m}\protect\tilde{k}}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}^{(z)}_{x;n}
=\overline{M}_{\nu;y;\protect\tilde{k} k}^{-1}M_{\varepsilon;x;\protect\tilde{m} m}\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{e}}^{(z)}_{x;m}.
\end{equation*}
\begin{figure}
\centering
\includegraphics{./figs/fig12}
\caption{The space separating $\ensuremath{L}_{x;m-M_{z}}$ and $\ensuremath{L}_{x;m}$ can be regarded as a homogeneous transmission line with characteristic impedance $Z_{0;m}$. Implementing \glspl*{ABC} then entails stamping an impedance of value $Z_{0;m}$ for edge $\ensuremath{L}_{x;m}$ in the netlist.}
\label{fig:circuitABC}
\end{figure}
We observe that $\sum_{\protect\tilde{k}}\sum_{n}\ensuremath{\widetilde{P}}_{z;\protect\tilde{m}\protect\tilde{k}}\ensuremath{P}_{z;kn}\ensuremath{\protect\bow{e}}^{(z)}_{x;n}\equiv\mathrm{d}^{2}_{z}\ensuremath{\protect\bow{e}}^{(z)}_{x;m}$ is formally the second-order grid derivative of the scalar field $\ensuremath{\protect\bow{e}}^{(z)}_{x;m}$ along the $z$-direction.
This directly leads us to the grid counterpart of \eqref{eq:waveEqTime} in the vicinity of $\ensuremath{L}_{x;m}$, viz.
\begin{equation*}
\frac{\mathrm{d}^2}{\mathrm{d}z^2}\ensuremath{\protect\bow{e}}^{(z)}_{x;m}
=\overline{M}_{\nu;y;\protect\tilde{k} k}^{-1}M_{\varepsilon;x;\protect\tilde{m} m}\frac{\mathrm{d}^2}{\mathrm{d}t^2}\ensuremath{\protect\bow{e}}^{(z)}_{x;m}.
\end{equation*}
Again, this equation can be factorised in terms of grid wave propagators similar to~\eqref{eq:waveEqTimeExpandMoreSimpleMore} to yield the grid version of \eqref{eq:EngquistMadja}, viz.
\begin{equation}
\left.\left(\frac{\mathrm{d}}{\mathrm{d}z}-\sqrt{\overline{M}_{\nu;y;\protect\tilde{k} k}^{-1}M_{\varepsilon;x;\protect\tilde{m} m}}\frac{\mathrm{d}}{\mathrm{d}t}\right)\ensuremath{\protect\bow{e}}^{(z)}_{x;m}\right|_{z=z_{\text{B}}}=0,
\label{eq:EngquistMadjaGrid}
\end{equation}
which can be easily extended to other field components and to other boundary orientations.
It also states that the grid wave impinging perpendicularly to the boundary $z=z_{\text{B}}$ at the edge $\ensuremath{L}_{x;m}$ propagates at a speed $\nu=(\overline{M}_{\nu;y;\protect\tilde{k} k}^{-1}M_{\varepsilon;x;\protect\tilde{m} m})^{-\frac{1}{2}}$.
A similar result can be obtained if we had used the grid wave equation of the magnetic field $\ensuremath{\protect\bow{\mathrm{\mathbf{h}}}}\xspace$.
Having said this, we can think of the space between edge $\ensuremath{L}_{x;m}$ and $\ensuremath{L}_{x;m-M_{z}}$, namely the preceding $x$-edge in $z$-direction, as a homogeneous transmission line with characteristic impedance
\begin{equation*}
Z_{0;m}:=\left(\overline{M}_{\nu;y;\protect\tilde{k} k}M_{\varepsilon;x;\protect\tilde{m} m}\right)^{-1/2},
\end{equation*}
and length $|\ensuremath{L}_{z;m-M_{z}}|$, see Figure~\ref{fig:circuitABC}.
The voltage that excites the line is given by the voltage on edge $\ensuremath{L}_{x;m-M_z}$, namely $\ensuremath{\protect\bow{e}}_{x;m-M_{z}}$.
Therefore, if one wants to implement the \gls*{ABC} at edge $\ensuremath{L}_{m}$, an impedance $Z_{0;m}$ must be assigned to $\ensuremath{L}_{m}$.
A subsequent circuit extraction can be carried out by implementing either Algorithm~\ref{alg:netlistEM1} or \ref{alg:netlistEM2}.
\section{Numerical Examples}
\label{sec:numerics}
In this section, the presented methodology to generate electric circuit stamps representing 3D field problems is applied to several representative numerical examples.
In Section~\ref{subsec:resultsET}, we use our netlist extraction method as described in Section~\ref{sec:circuitsET} on an \gls*{ET} problem.
The considered \gls*{ET} problem is a 3D field problem corresponding to the series connection of a capacitor and a resistor.
While applying an external voltage, the transient heating due to the resulting current is simulated using SPICE\xspace and then compared to a field solver reference solution.
Additionally, a circuit representation for the \gls*{ET} field problem of a microelectronic chip package is obtained and used for circuit simulation.
In Section~\ref{subsec:resultsEM}, we apply our method of circuit extraction for \gls*{EM} field problems as described in Section~\ref{sec:circuitsEM} to compute the resonant frequencies of a rectangular cavity with \gls*{PEC} boundaries.
This example is quite illustrative and easy to implement because of the required \glspl*{BC} on the cavity walls.
It simply suffices not to print the circuit stamp associated with those edges on the wall, meaning that the associated stamps are short-circuited.
Furthermore, the availability of an analytic formula for the resonant frequencies permits a direct error assessment.
Finally, in Section~\ref{subsec:resultsABC}, the implementation of \glspl*{ABC} as discussed in Section~\ref{subsec:theoryABC} is carried out to investigate reflections at the end of a rectangular coaxial waveguide.
For all presented examples, the Matlab\textsuperscript{\textregistered}\xspace code to generate the corresponding netlists from the discretised 3D field problem is openly available~\cite{Casper_2018ab}.
\subsection{Electrothermal Circuit Validation}
\label{subsec:resultsET}
To validate our netlist extraction method on an \gls*{ET} problem, we consider the \textsc{Joule}\xspace heating in a 3D field problem represented by a series connection of an electric resistor and a capacitor.
The temperature dependence of the electric conductivity is manifested via the temperature coefficient $\alpha=\SI{3.9e-3}{1\per K}$.
The relevant configuration is realised by a brick of two different materials as shown in Figure~\ref{fig:benchmarkStructure}.
The brick is of dimension $0.4\times 0.1\times\SI{0.1}{\micro\metre\tothe{3}}$, with the resistive part having a length of $\ell=\SI{0.3}{\micro\metre}$ and the capacitive part having a length of $d=\SI{0.1}{\micro\metre}$.
At $x=0$ and $x=\ell+d$, \gls*{PEC} electrodes are used.
In Table~\ref{tab:materialPropET}, all material properties are summarised for a reference temperature of $T_{0}=\SI{293}{K}$.
\begin{figure}
\centering
\includegraphics{./figs/fig13}
\caption{Geometry of the \gls*{ET} validation example. The series connection of a resistive part and a capacitive part is excited with a voltage source $V_{\text{app}}$ imposed as a \textsc{Dirichlet}\xspace condition.}
\label{fig:benchmarkStructure}
\end{figure}
A spatial grid with $9\times 9\times 9$ cells is employed and the field problem is solved by using an in-house implementation of the \gls*{FIT} method with a first-order implicit \textsc{Euler}\xspace scheme as time integrator.
The simulation time amounts to $t_{\text{end}}=\SI{13}{\micro\second}$.
For the simulation of the extracted electric circuit, we use the freely available LTspice\xspace software\footnote{All circuit simulations in this paper have been done using LTspice\xspace in its version 4.22x with default settings.}.
LTspice\xspace uses adaptive refinement in time for which an initial time step of $\Delta t_{\text{init}}=\SI{0.13}{\mu s}$ is used.
The resulting non-equidistant time axis is refined by a factor of three and then used for the \gls*{FIT} solver.
A voltage $V_{\text{app}}=\SI{1}{kV}(1-\exp(-t/\tau))$ with $\tau=0.1 t_{\text{end}}$ is applied at the electrodes as shown in Figure~\ref{fig:benchmarkStructure}.
Using this setting, two simulations are run.
The first neglects the temperature dependence of the conductivities and thus a linear setting ensues.
The second neglects only the temperature dependence of the thermal conductivity but accounts for that of the electric conductivity via the temperature coefficient $\alpha$ entailing a non-linear setting.
To observe the transient behaviour, we select the resistor-capacitor interface point $\mathbf{x}_{0}=(\ell,0,0)$ as observation point and plot the results in Figure~\ref{fig:resultsET}.
\begin{figure}[t]
\centering
\subfloat[\label{fig:testcaseETresultsPotential}]{\includegraphics[width=0.48\columnwidth]{./figs/fig14a}}
\hspace{0.02\columnwidth}
\subfloat[\label{fig:testcaseETresultsTemperature}]{\includegraphics[width=0.48\columnwidth]{./figs/fig14b}}
\caption{\gls*{ET} validation example results at $\mathbf{x}=\mathbf{x}_{0}$ for (a) the electric potential $\ensuremath{\boldsymbol{\mathrm{\varphi}}}$ and (b) the temperature $\ensuremath{\mathbf{T}}$. A comparison between the \gls*{FIT} and circuit simulation results is shown for the linear and non-linear cases.}
\label{fig:resultsET}
\end{figure}
\begin{table}[t]
\centering
\begin{tabular}{cccc}\toprule
Symbol & Description & $0<x<\ell$ & $\ell<x<\ell+d$ \\\bottomrule\toprule
$\sigma$ $\left(\si{S\per m}\right)$ & electric conductivity & \num{1e-4} & \num{0} \\
$\varepsilon_\text{r}$ & relative permittivity & \num{1} & \num{3.9} \\
$\lambda$ $\left(\si{W/K/m}\right)$ & thermal conductivity & \num{401} & \num{1400} \\
$\rho c$ $\left(\si{J\per K\per cm^3}\right)$& volumetric heat density & \num{3.48} & \num{2.10} \\
$\alpha (\si{1\per K})$ & temperature coefficient & \num{3.9e-3} & -- \\\bottomrule
\end{tabular}
\caption{Material properties at reference temperature $T_{0}=\SI{293}{K}$ for the \gls*{ET} test case.}
\label{tab:materialPropET}
\end{table}
For a quantitative error assessment of the solution, we define the measures
\begin{equation}
\Delta_{\ensuremath{\boldsymbol{\mathrm{\varphi}}}}=\frac{\max_{i}\lVert\ensuremath{\boldsymbol{\mathrm{\varphi}}}^{\text{cir}}(t_{i})-\ensuremath{\boldsymbol{\mathrm{\varphi}}}^\text{FIT}(t_{i})\rVert_{2}}{\max_{i}\lVert\ensuremath{\boldsymbol{\mathrm{\varphi}}}^{\text{FIT}}(t_{i})\rVert_{2}},\quad
\Delta_{\ensuremath{\mathbf{T}}}=\frac{\max_{i}\lVert\ensuremath{\mathbf{T}}^{\text{cir}}(t_{i})-\ensuremath{\mathbf{T}}^\text{FIT}(t_{i})\rVert_{2}}{\max_{i}\lVert\ensuremath{\mathbf{T}}^{\text{FIT}}(t_{i})\rVert_{2}},
\label{eq:solDiffs}
\end{equation}
where $\ensuremath{\boldsymbol{\mathrm{\varphi}}}^{\text{cir}}$, $\ensuremath{\boldsymbol{\mathrm{\varphi}}}^{\text{FIT}}$, $\ensuremath{\mathbf{T}}^{\text{cir}}$ and $\ensuremath{\mathbf{T}}^{\text{FIT}}$ are the potential and temperature solution vectors obtained via circuit and \gls*{FIT} simulation, respectively.
To calculate these errors appropriately, the circuit solution is interpolated to the time axis employed by the \gls*{FIT} solution using cubic spline interpolation.
We want to remark that the quantities in \eqref{eq:solDiffs} are not errors in the classical sense since none of the solutions is exact.
The computed differences amount to $\Delta_{\ensuremath{\boldsymbol{\mathrm{\varphi}}}}^{\text{lin}}\approx\SI{0.36}{\percent}$ and $\Delta_{\ensuremath{\mathbf{T}}}^{\text{lin}}\approx\SI{0.48}{\percent}$ for the linear case and $\Delta_{\ensuremath{\boldsymbol{\mathrm{\varphi}}}}^{\text{nlin}}\approx\SI{0.42}{\percent}$ and $\Delta_{\ensuremath{\mathbf{T}}}^{\text{nlin}}\approx\SI{0.44}{\percent}$ for the non-linear case.
The remaining error is attributed mainly to the different time integrators.
\begin{figure}
\centering
\subfloat[\label{fig:chip}]{\includegraphics[width=0.28\textwidth]{./figs/fig15a}}
\hspace{0.02\textwidth}
\subfloat[\label{fig:testcaseChip}]{\includegraphics[width=0.33\textwidth]{./figs/fig15b}}
\hspace{0.02\textwidth}
\subfloat[\label{fig:chipResults3D}]{\includegraphics[width=0.3\textwidth]{figs/fig15c}}
\caption{(a) 3D microelectronic chip package with an attached bond wire. (b) Temperature at the hottest point of the chip package obtained by \gls*{FIT} and circuit simulation. (c) Temperature distribution in the chip package obtained by circuit simulation.}
\label{fig:resultsChip}
\end{figure}
For an industry-relevant example, the proposed method is applied to the 3D microelectronic chip package~\cite{Casper_2016ab} as shown in Figure~\ref{fig:chip}.
The field problem is discretised as described in Section~\ref{sec:FIT} and Algorithm~\ref{alg:netlistET} is used to generate the corresponding netlist.
This netlist uses \num{101147} circuit elements to describe a field problem that has been discretised using a grid with \num{9660} nodes.
Running a transient analysis on this netlist, an error of $\Delta_{\ensuremath{\boldsymbol{\mathrm{\varphi}}}}\approx\SI{0.23}{\percent}$ and $\Delta_{\ensuremath{\mathbf{T}}}\approx\SI{0.17}{\percent}$ compared to the field simulation is achieved.
Figure~\ref{fig:testcaseChip} shows the temperature of the hottest point in the chip package obtained by \gls*{FIT} and circuit simulation.
The temperature distribution in the chip package resulting from circuit simulation is shown in Figure~\ref{fig:chipResults3D}.
Thus, a good agreement of circuit simulation results for a 3D \gls{ET} problem is achieved when compared to the corresponding field solver results.
\subsection{Electromagnetic Circuit Validation}
\label{subsec:resultsEM}
In this section, we validate the method presented in Section~\ref{sec:circuitsEM} for the circuit representation of \gls*{EM} field problems.
To this end, a lossless rectangular resonant cavity with \gls*{PEC} boundaries and outer dimensions of $a\times b\times d=0.1\times 0.2\times \SI{0.2}{\cubic\metre}$ is simulated and its resonant frequencies are computed.
The homogeneous material within the cavity is specified by the relative permittivity $\ensuremath{\varepsilon}_{\mathrm{r}}=\num{2}$ and the relative permeability $\ensuremath{\mu}_{\mathrm{r}}=\num{1}$ for which the resonant frequencies can also be calculated by means of the formula~\cite{Griffiths_1999aa}
\begin{equation*}
f_{\text{r}}^{mnp}=\frac{c_{0}}{2\sqrt{\ensuremath{\mu}_{\mathrm{r}}\ensuremath{\varepsilon}_{\mathrm{r}}}}\sqrt{\left(\frac{m}{a}\right)^2+\left(\frac{n}{b}\right)^2+\left(\frac{p}{d}\right)^2},
\end{equation*}
where $c_{0}$ is the speed of light and $\left\{m,n,p\right\}$ are the indices of the resonant modes and are given by natural numbers including zero.
For these resonant frequencies, the longitudinal \gls*{TE} and \gls*{TM} field components are given by
\begin{subequations}
\begin{align}
H_{z}^{mnp}=H_{0}^{mnp}\cos\left(\frac{m\pi}{a}x\right)\cos\left(\frac{n\pi}{b}y\right)\sin\left(\frac{p\pi}{d}z\right),\\
E_{z}^{mnp}=E_{0}^{mnp}\sin\left(\frac{m\pi}{a}x\right)\sin\left(\frac{n\pi}{b}y\right)\cos\left(\frac{p\pi}{d}z\right),
\end{align}
\label{eq:fieldLongitud}\end{subequations}
respectively, where $H_{0}^{mnp}$ and $E_{0}^{mnp}$ are the corresponding field amplitudes.
For a \gls*{TE} (\gls*{TM}) mode $mnp$ to exist, $H_{z}$ ($E_{z}$) must not become zero.
\begin{figure}[t]
\centering
\subfloat[\label{fig:testcaseEMresultsTE}]{\includegraphics[width=0.49\columnwidth]{./figs/fig16a}}
\subfloat[\label{fig:testcaseEMresultsTM}]{\includegraphics[width=0.49\columnwidth]{./figs/fig16b}}
\caption{(a) \gls*{TE} and (b) \gls*{TM} computed resonant frequencies by using \gls*{FIT} and by using circuit simulation. The plotted fields (voltages) $\ensuremath{\protect\bow{e}}_{x}$ and $\ensuremath{\protect\bow{e}}_{z}$ are evaluated at the edges connected to the central grid point.}
\label{fig:resultsEM}
\end{figure}
To simulate the excitation of modes within the cavity, we discretise the interior of the cavity using a regular grid of $10$ cells in each direction and apply \glspl*{PEC} upon all cavity walls.
Then, we obtain the resonant frequencies by solving the generalised eigenvalue problem given by
\begin{equation*}
\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\mathbf{C}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace = \left(2\pi f_{\text{r};\text{E}}^{\text{FIT}}\right)^{2}\ensuremath{\mathbf{M}_{\varepsilon}}\xspace\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace,
\end{equation*}
where a resonant frequency is denoted by $f_{\text{r};\text{E}}^{\text{FIT}}$.
Alternatively, we can use appropriate excitations to analyse the resulting field at a set of given frequencies.
For example, we can use an electric current source oriented along the positive $z$-direction and attached to the central grid point to excite \gls*{TM} modes.
In a similar manner, \gls*{TE} modes are excited by means of a looping electric current source located in the cavity centre.
We then solve the discretised problem given by
\begin{equation*}
(\ensuremath{\protect\widetilde{\mathbf{C}}}\xspace\ensuremath{\mathbf{M}_{\nu}}\xspace\ensuremath{\mathbf{C}}\xspace-\omega^{2}\ensuremath{\mathbf{M}_{\varepsilon}}\xspace)\ensuremath{\protect\bow{\mathrm{\mathbf{e}}}}\xspace = -j\omega\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}}
\end{equation*}
for a set of angular frequencies $\omega$, where $\ensuremath{\protect\bbow{\mathrm{\mathbf{j}}}}\xspace_{\text{i}}$ is the current source vector whose entries are all zero except at the corresponding source edges.
The frequency axis from \num{0.5} to \SI{3}{GHz} is discretised using \num{2000} points for \gls*{TM} excitation and \num{3000} points for \gls*{TE} excitation.
The results are evaluated on one edge for each excitation type.
For the \gls*{TM} case, an edge in positive $z$-direction connected to the central grid point is used while for the \gls*{TE} case, an edge in positive $x$-direction connected to the point (5,6,10)\SI{}{cm} is used.
In Figure~\ref{fig:resultsEM}, the voltages $\ensuremath{\protect\bow{e}}_{x}(\omega)$ and $\ensuremath{\protect\bow{e}}_{z}(\omega)$ along these edges are plotted.
From the peaks in the plots, the corresponding resonant frequencies $f_{\text{r;TE}}^{\text{FIT}}$ and $f_{\text{r;TM}}^{\text{FIT}}$ are identified\footnote{We have used the function \texttt{findpeaks} of Matlab\textsuperscript{\textregistered}\xspace R2017a to identify the peaks in the plot.}.
To validate our circuit extraction method for \gls*{EM} problems, we generate the netlist of the resonant cavity according to Algorithm~\ref{alg:netlistEM1} and simulate the resulting circuit in LTspice\xspace by performing an AC analysis in the same frequency range as before.
We then identify the circuit voltages corresponding to $\ensuremath{\protect\bow{e}}_{x}(\omega)$ and $\ensuremath{\protect\bow{e}}_{z}(\omega)$ and plot them also directly in Figure~\ref{fig:resultsEM} for a comparison.
The circuit resonant frequencies $f_{\text{r;TE}}^{\text{Cir}}$ and $f_{\text{r;TM}}^{\text{Cir}}$ are again identified by means of the peaks and we collect the computed resonant frequencies for the first few modes in Table~\ref{tab:resultsEMerrors}.
Note that, according to~\eqref{eq:fieldLongitud}, the \gls*{TM}-mode does not exist for $m=1$.
Additionally, the errors
\begin{equation*}
\epsilon_{\text{r;TE}} := \frac{\left|f_{\text{r;TE}}^{\text{FIT}}-f_{\text{r;TE}}^{\text{Cir}}\right|}{f_{\text{r;TE}}^{\text{FIT}}},\quad \epsilon_{\text{r;TM}} := \frac{\left|f_{\text{r;TM}}^{\text{FIT}}-f_{\text{r;TM}}^{\text{Cir}}\right|}{f_{\text{r;TM}}^{\text{FIT}}}
\end{equation*}
are presented.
For all modes, these errors are much smaller than $\SI{1}{\%}$.
\begin{table}[t]
\centering
\begin{tabular}{ccccccccc}\toprule
Mode $mnp$ & $f_{\text{r};\text{E}}^{\text{FIT}}$ & $f_{\text{r;TE}}^{\text{Cir}}$ & $f_{\text{r;TE}}^{\text{FIT}}$ & $f_{\text{r;TM}}^{\text{Cir}}$ & $f_{\text{r;TM}}^{\text{FIT}}$ & $f_{\text{r}}^{mnp}$ & $\epsilon_{\text{r;TE}}$ (\%) & $\epsilon_{\text{r;TM}}$ (\%)\\\bottomrule\toprule
011 & \num{0.746} & \num{0.746} & \num{0.747} & --- & --- & \num{0.749} & \num{0.0471} & --- \\
110/101/012/021 & \num{1.180} & \num{1.169} & \num{1.169} & \num{1.180} & \num{1.180} & \num{1.185} & \num{0.0030} & \num{0.0086} \\
111 & \num{1.293} & \num{1.293} & \num{1.293} & \num{1.293} & \num{1.293} & \num{1.298} & \num{0.0270} & \num{0.0216} \\
121 & \num{1.575} & \num{1.575} & \num{1.575} & \num{1.574} & \num{1.574} & \num{1.590} & \num{0.0354} & \num{0.0262} \\
013 & \num{1.620} & \num{1.621} & \num{1.620} & --- & --- & \num{1.676} & \num{0.0050} & --- \\
122 & \num{1.853} & \num{1.853} & \num{1.853} & \num{1.860} & \num{1.861} & \num{1.836} & \num{0.0244} & \num{0.0465} \\\bottomrule
\end{tabular}
\caption{Analytic and computed resonant frequencies in \si{GHz} for several resonant modes and the corresponding relative errors. As mode degeneracy in the cavity is relevant, an exact identification of the mode indices from the plots in Figure~\ref{fig:resultsEM} is not possible.}
\label{tab:resultsEMerrors}
\end{table}
\subsection{Signal Transmission Using Absorbing Boundary Conditions}
\label{subsec:resultsABC}
In this section, based on the method described in Section~\ref{subsec:theoryABC}, we present a simple validation example for \glspl*{ABC} in the context of circuit simulation.
To this end, let us consider a coaxial transmission line of rectangular cross section oriented along the $z$-direction as depicted in Figure~\ref{fig:coaxialLine}.
For the simulation of an infinitely long line using a finite computational domain, the implementation of \glspl*{ABC} is required to counteract unwanted incoming reflections.
We use an excitation signal at port 1 and simulate its propagation in time until it has reached port~2.
Thus, we generate two simulation results in time domain.
The first one corresponds to the case when port~2 is terminated with a perfect magnetic wall, that is an open port ($Z_{2}\to\infty$).
The second one corresponds to the case when port~2 is terminated with the characteristic line impedance ($Z_{2}=Z_{0}$).
For both cases, a \gls*{PMC} ($Z_{1}\to\infty$) at port 1 is applied\footnote{According to image theory, the perfect magnetic wall at port 1 serves as a mirror which reflects uprightly the otherwise backward travelling wave.}.
\begin{figure}
\centering
\includegraphics{./figs/fig17}
\caption{Geometry of a coaxial transmission line oriented along the $z$-axis and excited at port 1 by a Gaussian current pulse. A propagating wave is thus generated arriving at port~2 that is connected to an external impedance. For reasons of visibility, the annotations of the inner conductor's dimensions are not shown.}
\label{fig:coaxialLine}
\end{figure}
The length of the coaxial line is $l=\SI{150}{cm}$, the width and height of the outer conductor are $w_{\text{o}}=h_{\text{o}}=\SI{3}{cm}$ and of the inner conductor $w_{\text{i}}=h_{\text{i}}=\SI{1}{cm}$.
While the conductors are modelled as \gls*{PEC}, the material between them is vacuum.
Due to the expected propagation in $z$-direction, the longitudinal direction requires a finer discretisation compared to the transversal direction.
Thus, we choose a grid of $3\times 3\times 150$ cells.
According to \eqref{eq:EngquistMadjaGrid} for such a discretisation grid, the characteristic impedance for the edges connecting the inner and outer conductor along the plane of port~2 should amount to\footnote{Note that in the calculation of $Z_{0;m}$, the value employed for $M_{\ensuremath{\varepsilon};x;\protect\tilde{m} m}$ is taken directly from the parallel edge just in front of the boundary edge in accordance with the impinging grid wave front speed.} $Z_{0;m}=(\overline{M}_{\nu;y;k\protect\tilde{k}}M_{\varepsilon;x;\protect\tilde{m} m})^{-1/2}\approx\SI{376.7}{\Omega}$.
For the given grid, there are eight such edges giving eight parallel conductances such that the total resistance at port~2 equals $Z_{2}=Z_{0}=8Z_{0;m}\approx\SI{47.09}{\Omega}$, which is also the characteristic impedance of the line.
To excite the signal at port 1, the edges connecting the inner and outer conductors along the plane of port 1 are impressed with a current such that the total current from inner to outer conductor is
\begin{equation*}
I_{\text{i}}(t)=\hat{I}\exp\left(\frac{(t-t_{0})^2}{2\sigma_{\text{G}}^2}\right),
\end{equation*}
with $\hat{I}=\SI{1}{A}$, which is a \textsc{Gauss}\xspace pulse with a maximal frequency\footnote{Confining the excitation to this maximal frequency component, we assure that the TEM mode is the only propagating mode on the line} component of $f_{\text{max}}=\SI{1}{GHz}$.
The used constants are given by ${\sigma_{\text{G}}=\sqrt{\ln(10)}/(\pi f_{\text{max}})}$ and by ${t_{0}=\sqrt{6\sigma_{\text{G}}^2\ln(10)}}$.
The simulation time $t_{\text{end}}=\SI{10}{ns}$ is chosen such that the excited pulse can reach port~2.
Having defined the geometry, the excitation and the simulation time, we also generate the corresponding netlist using Algorithm~\ref{alg:netlistEM1}.
\begin{figure}[t]
\centering
\subfloat[\label{fig:testcaseABCresultsWavePropReflect}]{\includegraphics[width=0.499\columnwidth]{./figs/fig18a}}
\subfloat[\label{fig:testcaseABCresultsWavePropNoReflect}]{\includegraphics[width=0.499\columnwidth]{./figs/fig18b}}
\caption{Snapshots of the wave propagating along the $z$-axis at times $t_{1}=\SI{3}{ns}$, $t_{2}=\SI{5}{ns}$, $t_{3}=\SI{7}{ns}$ and $t_{4}=\SI{8}{ns}$ computed by \gls*{FIT} and circuit simulation. (a) shows the case $Z_{2}\to\infty$ to realise total reflection. (b) shows the case $Z_{2}=Z_{0}$ to realise \glspl*{ABC}.}
\label{fig:testcaseABCresultsWaveProp}
\end{figure}
\begin{figure}[t]
\centering
\subfloat[\label{fig:testcaseABCresultsV2vsTimeReflect}]{\includegraphics[width=0.49\columnwidth]{./figs/fig19a}}
\hspace{0.01\columnwidth}
\subfloat[\label{fig:testcaseABCresultsV2vsTimeNoReflect}]{\includegraphics[width=0.49\columnwidth]{./figs/fig19b}}
\caption{Comparison of the output voltage $V_{2}$ with respect to time computed by \gls*{FIT} and circuit simulation for the case of (a) $Z_{2}\to\infty$ to realise total reflection and (b) $Z_{2}=Z_{0}$ to realise \glspl*{ABC}.}
\label{fig:testcaseABCresultsV2vsTime}
\end{figure}
The generated netlist is simulated by means of a transient analysis in LTspice\xspace.
On the other hand, the Leapfrog scheme is used as a time integrator within the \gls*{FIT} framework to carry out the simulation directly on the 3D grid.
In both cases, we use the time axis generated by the adaptive time stepping algorithm provided by LTspice\xspace, which satisfies the \gls*{CFL}-condition being a stability requirement for the explicit Leapfrog scheme~\cite{Yee_1966aa}.
In the following, we compare the voltage $V_{\text{oi}}(z,t)$ between outer and inner conductor and the voltage $V_{2}(t)=V_{\text{oi}}(l,t)$ at port~2.
For $Z_{2}\to\infty$ and $Z_{2}=Z_{0}$, Figure~\ref{fig:testcaseABCresultsWaveProp} shows $V_{\text{oi}}(z,t)$ at different times computed by means of \gls*{FIT} and circuit simulation.
Figure~\ref{fig:testcaseABCresultsWavePropReflect} shows the case in which port~2 is terminated by a perfect magnetic boundary (standard homogeneous \textsc{Neumann}\xspace) condition while Figure~\ref{fig:testcaseABCresultsWavePropNoReflect} shows the case when a matching impedance $Z_{2}=Z_{0}$ according to \eqref{eq:EngquistMadjaGrid} is applied at port~2.
As predicted by the theory in Section~\ref{subsec:theoryABC}, we observe that the matching impedance at port~2 counteracts incoming reflections effectively.
In Figure~\ref{fig:testcaseABCresultsV2vsTime}, we show $V_{2}(t)$ computed by means of \gls*{FIT} and circuit simulation for the two already considered cases.
We observe therein that incoming reflections at port~2 result in an undesired overshooting of the voltage.
For a quantitative comparison, we define the relative error of $V_{2}(t)$ between \gls*{FIT} and circuit results as
\begin{equation*}
\Delta_{V_{2}}^{Z_{2}}=\frac{\max_{i}\lVert V_{2}^{\text{cir}}(t_{i})-V_{2}^{\text{FIT}}(t_{i})\rVert_{2}}{\max_{i}\lVert V_{2}^{\text{FIT}}(t_{i})\rVert_{2}},
\end{equation*}
and obtain $\Delta_{V_{2}}^{Z_{0}}\approx\SI{1.046}{\%}$ and $\Delta_{V_{2}}^{\infty}\approx\SI{1.136}{\%}$.
\section{Conclusion and Future Work}
\label{sec:conclusion}
A method for the automatic netlist generation of general 3D \gls*{ET} and \gls*{EM} problems has been presented.
The topology of each circuit stamp associated with edges in the regular primal grid has been derived by using \gls*{FIT} for spatial discretisation.
Using the \gls*{MNA}, the \gls*{FIT}-discretised \gls*{ET} formulation has been mapped into a circuit that can be solved by any SPICE\xspace-like program.
It has been shown that initial conditions can be easily prescribed as initial potentials for the lumped capacitances in the SPICE\xspace language.
Furthermore, the implementation of mixed boundary conditions of \textsc{Dirichlet}\xspace, homogeneous \textsc{Neumann}\xspace and \textsc{Robin}\xspace type has been discussed.
We have also shown that temperature dependent material models result in non-linearities in the lumped resistances requiring the implementation of behavioural \glspl*{VCCS} in SPICE\xspace.
From the standard E-H formulation and the E-A formulation, we have derived circuit stamps representing general \gls*{EM} problems.
In both circuit representations, the integrated electric field models the voltage between the stamp terminals while the integrated magnetic vector potential models the electric current in the E-A formulation.
To guarantee uniqueness of the solution in the latter, we have employed \textsc{Coulomb}\xspace's gauge on the magnetic vector potential, that has been implemented by means of a tree-cotree decomposition of the primal discretisation grid.
Thereby, the electric current along edges in the cotree are degrees of freedom whereas those along edges in the tree are modelled by \glspl*{CCCS} being controlled by currents in the cotree.
For both representations, a dual circuit formulation exists if magnetic sources instead of electrical sources are considered.
In the dual case, an auxiliary electric potential would be used instead of the magnetic vector potential.
To demonstrate the correctness of our formulations, several numerical examples have been shown for the primal circuits involving electric sources only.
The formulation of inhomogeneous \textsc{Neumann}\xspace \glspl*{BC} could be a further extension to the presented approach.
Furthermore, the method can also be applied to extract circuits from \gls*{FEM} models.
To account for thermal effects in \gls*{EM} problems, the methods for the extraction of \gls*{ET} and \gls*{EM} circuit stamps can be combined to generate a thermo-\gls*{EM} circuit stamp.
Methods to account for non-linear material characteristics in the \gls*{EM} case are still to be developed.
However, in principle one can follow similar ideas to those presented in the \gls*{ET} case.
For large field models, the resulting circuit can become very large.
Therefore, to efficiently simulate such circuits, dedicated \gls*{MOR} techniques for circuits can be applied.
The first of these techniques is known as the \gls*{AWE} proposed by Pillage and Rohrer~\cite{Pillage_1990aa} and extensions developed afterwards.
The most prominent ones are the \gls*{MPVL} by Feldmann and Freund~\cite{Feldmann_1995aa} and the \gls*{PRIMA}~\cite{Odabasioglu_1998aa}.
More recent approaches are based on the proper orthogonal decomposition~\cite{Hinze_2012ad} and other well-known general \gls*{MOR} techniques.
\section*{Acknowledgements}
This is a pre-print of an article published in the Journal of Computational Electronics.
The final authenticated version is available online at: https://doi.org/10.1007/s10825-019-01368-6.
The authors thank Abdul Moiz and Victoria Heinz for their passionate work on implementing the automated electrothermal netlist generation. The work is supported by the European Union within FP7-ICT-2013 in the context of the \emph{Nano-electronic COupled Problems Solutions} (nanoCOPS) project (grant no. 619166), by the \emph{Excellence Initiative} of the German Federal and State Governments and the Graduate School of Computational Engineering at Technische Universität Darmstadt.
|
\section{Introduction}
Near-Extremal black holes have a universal structure near their horizons:
there is an $AdS_2$ throat with a slowly varying internal space.
Its low energy gravitational dynamics is captured universally by the following effective action in two dimensions \cite{Strominger:1994tn}:
\begin{equation}\label{full action}
I=\underbrace{-{\phi_0\over 2
}\left(\int R+2\int_{\partial_M}K\right)}_{\text{Einstein-Hilbert Action}}\underbrace{-{1\over 2
}\left(\int_M \phi(R+2)+2\int_{\partial M}\phi_b K\right)}_{\text{Jackiw-Teitelboim action}}+S_{matter}(g,\psi),
\end{equation}
where the dilaton field $\phi+\phi_0$ represents the size of internal space. We have separated the size of internal space into two parts: $\phi_0$ is its value at extremality. It sets the value of the extremal entropy
which comes from the first term in \nref{full action}.
$\phi$ is the deviaton from this value.
We have also added matter that only couples to the metric. This is a reasonable assumption when matter comes
from Kaluza Klein reduction, where the coupling to the dilaton would involve $\phi/\phi_0 \ll 1$.
The action $\int \phi(R+2)+2\int\phi_b K$ is the so-called Jackiw-Teitelboim action \cite{JACKIW1985343,TEITELBOIM198341}, and will be the main focus of our paper.
This action is one of the simplest nontrivial gravitational actions in two dimensions.\footnote{Another nontrivial action is the CGHS model which could be written as $\int (\phi R+C)$, and that characterizes the horizon structure of general black holes.}. It is simple because the bulk geometry is a rigid $AdS_2$ space fixed by the equation of motion of the dilaton field. Its nontriviality arises from the remaining boundary action. Schematically, the gravitational action is reduced to the following form:
\begin{equation}\label{schematic action}
I=-{ 2 \pi \phi_0
}\chi(M)-{\phi_b
}\int_{\partial M}K +S_{matter}(g,\psi).
\end{equation}
And the motion of the boundary is controlled by its extrinsic curvature.
Our goal will be to quantize this action and to provide expression for the full quantum gravity correlators
of \nref{full action}.
This problem was considered before in \cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Kitaev:2018wpr} from various points of view. Here we will add one other point of view where we
reduce the problem to the motion of a relativistic particle in an electric field, building on a suggestion in
Kitaev's talk at IAS \cite{kitaevIAS}. More precisely, one can consider a relativistic particle in a Lorentzian $AdS_2$ target space moving under the influence of an electric field. The coupling to electric field can also be viewed as a coupling to a spin connection so that it becomes a particle with spin as suggested by Kitaev. Alternatively we can start from a non-relativistic particle moving in hyperbolic space, $H_2$, under the influence of a magnetic field $b$. After analytic
continuation in $b$ to imaginary values we get the problem of interest.
Using this point of view one can think of the full quantum gravity problem as the combination of two problems.
First we consider quantum fields propagating in $AdS_2$ (or $H_2$ in the Euclidean case) and then we
add the ``gravitational particle" which couples to the quantum fields by changing their boundary location in $AdS_2$.
The discussion of quantum fields will be standard and depends on the particular model one interested in, therefore we will mainly focus on solving the second problem.
Generically, solving the gravitational problem is challenging and is not exactly equivalent to a quantum mechanical particle. One needs to worry about what functional space one will integrate over. For example, in path integrals, one usually integrates over all trajectories including those with self-intersections. However self-intersecting boundaries in gravitational system have no obvious meaning. On that account, more precisely the gravitational problem is equal to a self-avioding particle.
Nevertheless, it turns out that one can take a particular limit of this model, namely large $\phi_b$, to avoid this issue and a treatment of the boundary theory as an ordinary particle is justified. It is also true that the JT gravity can be rewritten as a Schwarzian action only in this limit. We call this the Schwarzian limit and will only focus on solving the JT action in the Schwarzian limit. Solving the model away from Schwarzian limit was considered recently by Kitaev and Suh \cite{Kitaev:2018wpr}.
Our result can be summarized as follows:
\textsl{First, we will give a formula to calculate all correlation functions with quantum gravity backreaction (formula \ref{final formula}). Second, we will give the exact Wheeler-DeWitt wavefunction in the Schwarzian limit, which has been analyzed classically by Harlow and Jafferis \cite{Harlow:2018tqv}. Last, we consider the recent proposed conjecture about complexity growth in this exact Wheeler-DeWitt wavefunction and show that the complexity maintains linear growth after taking quantum gravity effects into account. This, to our knowledge, is the first test of the gravitational conjecture made by Susskind \cite{Susskind:2018fmx} that the size of ERB grows linearly for as long as quantum mechanics allows.}
This paper is organized as follows: in the first section, we will review the classical calculation of this model and introduce notations; in the second section, we will make the dictionary between the JT model and a particle in a magnetic field; in the third and fourth sections, we will solve the quantum mechanical problem and derive the propagator and WdW wavefunction in the Schwarzian limit; in the last section, we will talk about gravitaional backreaction on correlators as well as complexity growth. As a useful notion in our calculation, we introduce a notation called gravitational Feyman diagrams.
\section{Classical Solutions}
Let us first consider the classical solutions of
the Jackiw-Teitelboim model \nref{full action}. See
\cite{Almheiri:2014cka} for further discussion.
%
%
The equation of motion of the dilaton field imposes $R=-2$
and
fixes the geometry to be $AdS_2$, or $H_2$ in the Euclidean case.
This is also true if
we have additional matter coupled with metric only, as in
\nref{full action}.
The equations for the metric constrain the dilaton
\begin{equation}\label{Einstein Equation}
(\nabla_{\mu}\nabla_{\nu}\phi-g_{\mu\nu}\nabla^2\phi+g_{\mu\nu}\phi)+T^{M}_{\mu\nu}=0
\end{equation}
These equations are compatible with each other thanks to the conservation of the matter stress tensor. They do not allow
any propagating mode. In fact, setting $T^M_{\mu\nu}=0$, and using a high momentum approximation we can write
\nref{Einstein Equation} as $ (k_{\mu}k_{\nu}-g_{\mu\nu}k^2)\phi(\vec k)=0 $, which then implies $ \phi(\vec k)=0 $, for large $k$.
More precisely, after introducing the Euclidean $AdS_2$ coordinates $ds^2 = d\rho^2 + \sinh^2\rho d\varphi^2$, we can solve
(\ref{Einstein Equation}) in $AdS_2$ with no matter. Up to an SL(2) transformation the solution is
\begin{equation}
\phi=\phi_h \cosh \rho,
\end{equation}
where $\phi_h$ is a constant that is fixed by the boundary conditions.
At the boundary we fix the metric along the boundary and the value of the dilaton field
\begin{equation} \label{BCond}
ds_{\parallel} = d u { \phi_r \over \epsilon } ~,~~~~~~~~~~~ \phi= \phi_b = { \phi_r \over \epsilon }
\end{equation}
where we think of $u$ as the time of the boundary theory. It is simply a rescaled version of proper time. Similarly $\phi_r$ is a rescaled
value of the dilaton. We will be interested in taking $\epsilon \to 0$. With these rescalings the value of $\phi_h$ in the interior
remains fixed as we take $\epsilon \to 0$ as $\phi_h = 2 \pi /\beta $ where $\beta$ is the inverse temperature, $u \sim u + \beta$. Notice that due to the factor of $\phi_r$ in the first expression in \nref{BCond} we are measuring time in units of the constant $\phi_r$, which has dimensions of length. We did this for convenience.
A nice feature that appears after taking the $\epsilon \to 0$ limit is that the action (\ref{schematic action}) can be written as the Schwarzian action for the boundary curve labeled by $\varphi(u)$ \cite{Maldacena:2016upp}:
\begin{equation}
I=-\int du Sch(\tan{\varphi(u)\over 2},u)
\end{equation}
The fluctuation of the boundary shape can be understood as the fluctuation of the dilaton distribution in the bulk.
A bit more explicitly we can say that the dilaton boundary condition fixes the location of the boundary at $\rho_b$ given by
$\phi_b = \phi_h \cosh \rho_b$, and the metric at that location relates the time $\varphi$ to $u$ by $\phi_r du = \epsilon \sinh \rho_b d\varphi $.
We get the above formulas noticing that the period of $\varphi$ is $2\pi$ while that of $u$ is $\beta$, which fixes $\epsilon \sinh \rho_b$.
\section{Charged Particle in $AdS_2$}
Despite the absence of a bulk propagating mode there is still a non-trivial dynamical gravitational degree of freedom. There are various
ways to describe it. Here we will think of it as arising from the motion of the physical boundary of $AdS_2$ inside a rigid $AdS_2$ space.
This picture is most clear for finite $\epsilon$ in \nref{BCond}, but it is true even as $\epsilon \to 0$. The dynamics of the boundary
is SL(2) invariant. This SL(2) invariance is a gauge symmetry since it simply reflects the freedom we have for cutting out a piece
of $AdS_2$ space that we will call the ``inside". It is important that the dilaton field we discussed above is produced after we put in the
boundary and it moves together with the boundary under this SL(2) gauge transformation. It is a bit like the Mach principle, the location
in $AdS_2$ is only defined after we fix the boundary (or distant ``stars").
We can make this picture of a dynamical boundary more manifest as follows.
Since the bulk Jackiw-Teitelboim action \nref{full action} is linear in $\phi$, we can integrate out the dilaton field which
sets the metric to that of $AdS_2$ and removes the bulk term in the action, leaving only the term involving the extrinsic curvature
\begin{eqnarray}
I&=&-{\phi_r\over
\epsilon}\int du \sqrt{g}K
\end{eqnarray}
This action, however, is divergent as we take $\epsilon$ to zero. This divergence is simply proportional to the length of the boundary
and can be interpreted as a contribution to the ground state energy of the system.
So we introduce a counterterm proportional to the length of the boundary to cancel it. This is just a common shift of the energies
of all states.
It is also convenient to use the
Gauss-Bonnet theorem to relate the extrinsic curvature to an integral over the bulk
\begin{equation}
\int_{\partial_M} du \sqrt{g}K=2\pi\chi(M)-{1\over 2}\int_M R
\end{equation}
Since the curvature is a constant, the bulk integral is actually proportional to the total area $A$ of our space. That is we have the regularized action:
\begin{eqnarray}\label{subtraction}
I&=&-{\phi_r\over
\epsilon}\int_{\partial M} du \sqrt{g}(K-\underbrace{1}_{\text{counterterm}})=-{\phi_r\over
\epsilon}\left(2 \pi\chi(M)-{1\over 2}\int_{M}R-\int_{\partial M} du \sqrt{g}
\right) \nonumber\\
&=&-2\pi q\chi(M)-q A+q L,~~~~~~~~~~~~~~~~~~~~~~~~~~~q\equiv{\phi_r\over
\epsilon} ,~~L={\beta \phi_r\over \epsilon}
\end{eqnarray}
We now define an external gauge field $a_\mu$ as
\begin{equation}
a_{\varphi}=\cosh\rho-1,~~~~~~~a_{\rho}=0,~~~~~~~f_{\rho \varphi}=\sinh\rho=\sqrt{g},
\end{equation}
and write the action as follows
\begin{equation}\label{regularized action}
I= - 2 \pi q + q L - q \int a
\end{equation}
where we used that $ \chi(M)$ is a topological invariant equal to one, in our case, where the topology is that of a disk.
The term $q L$ is just the length of the boundary. So this action has a form somewhat
similar to the action of a relativistic charged particle moving in
$AdS_2$ in the presence of a constant electric field.
There are a couple of important differences. First we are summing only over trajectories of fixed proper length set by the inverse temperature
$\beta$. Second, in the JT theory we are treating the $SL(2)$ symmetry as a gauge symmetry. And finally, in the JT theory we identify the proper length with the boundary time, viewing configurations which differ only by a shift in proper time as inequivalent. In fact, all these changes simplify the problem: we can actually think
of the problem as a non-relativistic particle moving on $H_2$ in an electric field. In appendix
\ref{appendix:relativistic particle} we discuss in more detail the connection to the relativistic particle.
In fact, precisely the problem we are interested in has been discussed by Polyakov in \cite{Polyakov:1987ez}, Chapter 9, as an an intermediate step for the sum over paths.
Now we would also like to point out that we can directly get to the final formula by
using the discussion there, where he explicitly shows
that for a particle in flat space the sum over paths of fixed proper length that stretch between
two points $\vec x$ and $\vec x'$ gives
\begin{equation} \label{PolFlat}
\int {\cal D }\vec x e^{ - m_0 \tilde\tau } \delta( \dot { \vec x}^2 - 1 ) =
e^{ - {1\over 2} \mu^2 \tau} \langle x' | e^{ - \tau H } |x\rangle =
e^{ - {1\over 2} \mu^2 \tau} \int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot \vec x^2 }
\right)
\end{equation}
$\mu^2$ is the regularized mass and $\tilde \tau$ is related to $\tau$ by a multiplicative renormalization.
The JT model consists precisely of a functional integral of this form, where we fix the proper length along the boundary. There are two simple modifications, first the particle is in a curved $H_2$ space and second we have the coupling to the electric field. These are minor modifications, but the arguments leading
to \nref{PolFlat} continue to be valid so that the partition function of the JT model can be
written directly:
\begin{equation} \label{PolAdS}
\scaleto{\int {\cal D }\vec x e^{2\pi q - m_0 \tilde\tau +q\int a } \delta( { \dot x^2 + \dot y^2 \over y^2 }- q^2 ) =
e^{ 2\pi q- {1\over 2} \mu^2 \tau} \Tr e^{ - \tau H } = e^{2\pi q - {1\over 2} \mu^2 \tau }
\int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } - q {\dot x \over y}
\right)}{26pt}
\end{equation}
The delta function implements the first condition
in \nref{BCond} at each point along the path.
The last path integral can be done exactly by doing canonical quantization of the action (section \ref{CanonicalQuantization}) and by comparing the result with the one from the Schwarzian action \cite{Stanford:2017thb}
we can determine that $\tau$ is the inverse temperature $\beta$.
In the above discussion we have been fixing the time along the boundary. Instead we can fix the energy
at the boundary, where the energy is the variable conjugate to time.
This can be done by simply integrating \nref{PolAdS} times $e^{ \beta E}$
over $\beta$ along the imaginary axis. This fixes the energy of the non-relativistic problem by generating a $\delta( H - E)$. More precisely, we will argue that after doing a spectral decomposition we can write
the propagator at coincident points as
\begin{equation} \label{JTEn}
Z_{JT}(\beta ) = \int_0^\infty \rho(E)e^{-\beta E} dE \longrightarrow \rho(E) = \int_{-i \infty}^{i \infty}
{ d\beta \over i } e^{ E \beta } Z_{JT}(\beta)
\end{equation}
where the function $\rho(E) $ can then be interpreted as a ``density of states" in the microcannonical ensemble. We will give its explicit form in section (4.2).
For now, we only want to contrast this integral with a superficially similar one that appears when we
compute the relativistic propagator
\begin{equation} \label{RelEl}
e^{-2\pi q} \int_0^\infty e^{ E \beta }Z_{JT} (\beta) = \langle \phi(x) \phi(x) \rangle
\end{equation}
which gives the relativistic propagator of a massive particle in an electric field at coincident points
(we can also compute this at non-coincident points to get a finite answer). The total mass of the
particle is
\begin{equation}
m = q - {E\over q}
\end{equation}
For large $q$ this is above threshold for pair creation\footnote{See Appendix \ref{appendix:relativistic particle}}. The pair creation interpretation
is appropriate for the problem in \nref{RelEl}, but not for \nref{JTEn}.
In both problems we have a classical approximation to the dynamics that corresponds to a particle
describing a big circular trajectory in hyperbolic space at radius $\rho_c$ :
\begin{equation} \label{SolCir}
\tanh \rho_c = { m \over q }
\end{equation}
For the problem in \nref{RelEl}, fluctuations around this circle lead to an instability, with a single negative mode and an imaginary part in the partition function \nref{RelEl}. This single negative mode corresponds to small fluctuations of the overall size of the circular trajectory around \nref{SolCir}.
On the other hand in \nref{JTEn} we are integrating the same mode along a different contour, along the
imaginary axis, where we get a real and finite answer. Furthermore, the imaginary part in the
partition function \nref{RelEl} comes precisely from the trajectory describing pair creation, which is also the type of contribution captured in \nref{JTEn}.
Finally, in the relativistic particle problem, we expect that the pair creation amplitude should be exponentially suppressed for large $q$, while the partition function for the JT model is not. In fact,
for large $q$ the exponential suppression factor for pair creation goes as $e^{ - 2 \pi q } $, which is precisely cancelled by a similar factor in \nref{JTPa}, to obtain something finite in the large $q$ limit.
\section{Solving the Quantum Mechanical Problem}
\label{CanonicalQuantization}
As we explained above the solution of the JT theory is equivalent to considering a non-relativistic particle in $AdS_2$ or $H_2$.
We first consider the Euclidean problem, of a particle moving in $H_2$. An ordinary magnetic field in $H_2$ leads to an Euclidean action
of the form
\begin{equation} \label{magnetic field}
S= \int du {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } + i b \int d u { \dot x \over y } -{1\over 2} ( b^2 +{ 1 \over 4} ) \int d u ~,~~~~~~~b = i q
\end{equation}
If $b$ is real we will call it a magnetic field,
when $q$ is real we will call it an ``electric" field. The last term is a constant we introduced for convenience. Its only effect
will be to shift the ground state energy. It is interesting to compute the classical solutions and the corresponding action for
\nref{magnetic field}. These solutions are simplest in the $\rho$ and $\varphi $ coordinates, using the SL(2) symmetry we find that the trajectories are given by ($t=-iu$):
\begin{eqnarray}
{1\over 2}\sinh^2\rho ({d\varphi\over dt})^2+{q^2\over 2}-{1\over 8}=E,~~ \cosh\rho={q\beta\over 2\pi},~~ {d\varphi\over du}={2\pi\over \beta}.
\end{eqnarray}
In this classical limit we
get the following relations for the action and the temperature:
\begin{eqnarray}
& ~& { \beta \over 2 \pi } = {1\over \sqrt{2E+{1\over 4}}}
\cr
& ~ & -S = {2\pi^2\over \beta}+{\beta\over 8}-2\pi q \label{ClassEnt}
\end{eqnarray}
When $b$ is real, this system is fairly conventional and it was solved in \cite{Comtet:1986ki} . Its detailed spectrum depends on $b$. For very large $b$ we have a series of Landau levels and also a continuous spectrum. In fact, already the classical problem contains closed circular orbits, related to the discrete Landau levels, as well as orbits that go all the way to infinity.\footnote{See Appendix \ref{appendix:landau level}}
The number of discrete Landau levels decreases as we decrease the magnetic field and for $0 < b < {1\over 2}$ we only get a continuous spectrum. The system has a $SL(2)$ symmetry and the spectrum organizes into SL(2) representations, which are all in the
continuous series for $0 < b < 1/2$. For real $q$ we also find a continuous spectrum which we can view as the analytic continuation
of the one for this last range of $b$.
The canonical momenta of the action (\ref{magnetic field}) are:
\begin{equation}
p_x={\dot x\over y^2}+{i q\over y};~~~~~~~p_y={\dot y\over y^2}.
\end{equation}
And the Hamiltonian conjugate to $\tau_L$ is thus:
\begin{equation}
H={\dot x^2+\dot y^2\over 2y^2}+{q^2\over 2}={y^2\over 2} [(p_x-i{q\over y})^2+p_y^2] +{ q^2 \over 2 } - { 1 \over 8}
\end{equation}
Note that the Hamiltonian is not Hermitian. However, it is {\cal PT}-symmetric (here parity reflects $x$ and $p_x$) and for that
reason the spectrum is still real, see \cite{Bender:2005tb}.
The action is invariant under $SL(2,R)$ transformations generated by
\begin{eqnarray}
L_0=x p_x+y p_y;~~~~~L_{-1}=p_x;~~~~~~~~L_{1}=(y^2-x^2)p_x-2xyp_y-2iqy \label{SLtwoGen}
\end{eqnarray}
Notice the extra $q$ dependent term in $L_1$ that arises due to the presence of a magnetic field.
Up to a simple additive constant, the Hamiltonian is proportional to the Casimir operator
\begin{equation}
H={1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) + { q^2 \over 2} - { 1 \over 8 }
\end{equation}
As is common practice, let us label the states by quantum numbers $j={1\over 2}+is$ and $k$, so that $H|j,k\rangle=j(1-j)|j,k\rangle$ and $L_{-1}|j,k\rangle=k|j,k\rangle$.
We can find the eigenfunctions by
solving the $Schr\ddot{o}dinger$ equation with boundary condition that the wavefunction should vanish at the horizon $y\rightarrow \infty$ \cite{Comtet:1984mm,Comtet:1986ki,Pioline:2005pf}:
\begin{equation}
\omega_{s,k}={s^2\over 2} ,~~~~~~f_{s,k}(x,y)=
\begin{cases}
({s\sinh 2\pi s\over 4\pi^3k})^{1\over 2}|\Gamma(is-b+{1\over 2})|e^{-ik x}W_{b,is}(2 k y), ~~~k>0;\\
({s\sinh 2\pi s\over 4\pi^3|k|})^{1\over 2}|\Gamma(is+b+{1\over 2})|e^{-ik x}W_{-b,is}(2|k|y),~~~k<0.
\end{cases}
\end{equation}
where $\omega_{ks}$ is giving the energy of the states labelled by $s$ and $k$, and $ W $ is the Whittaker function. The additive constant in \nref{magnetic field} was
introduced to simplify this equation. We can think of $s$ as the quantum number of the continuous series representation of $SL(2)$ with
spin $j={1\over 2} + i s $.
After continuing $b \to i q$ we find that the gravitational system has a continuous
spectrum
\begin{equation}\label{energy relation}
E(s)={s^2 \over 2 }.
\end{equation}
\subsection{The Propagator }
It is useful to compute the propagator for the non-relativistic particle in a magnetic field,
$K(u,\boldsymbol{x_1},\boldsymbol{ x_2})=\langle \boldsymbol{x_1}| e^{-u H}|\boldsymbol{x_2}\rangle$. Here, $\boldsymbol{x}$ stands for ${x,y}$.
The propagator for a real magnetic particle was obtained in \cite{Comtet:1986ki}:
\begin{eqnarray} \label{particle heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})& =& e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} \int_0^{\infty}ds s e^{- u { s^2 \over 2}}{\sinh 2\pi s\over 2\pi (\cosh 2\pi s+\cos 2\pi b)}{1\over d^{1+2is}}
\times
\cr
& & ~~~~~ \times ~_2F_1({1\over 2}-b+is,{1\over 2}+b+is,1,1-{1\over d^2}).
\cr
d & = & \sqrt{(x_1-x_2)^2+(y_1+y_2)^2\over 4y_1 y_2}
\cr
e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} &=& e^{-2 i b \arctan {x_1-x_2\over y_1+y_2}}
\end{eqnarray}
In the case that we have a real magnetic field the prefactor is a phase and it is
gauge dependent. It is equal to the value of Wilson Line $e^{i\int a}$ stretched along the geodesic between $x_2$ and $x_1$.
Here we quoted the value in the gauge where the action is
\nref{magnetic field}. The second equation defines the parameter $d$, which is a function of the geodesic distance between the two points. Note that $d=1$ corresponds to coincident points.
We can get the answer we want by making the analytic continuation $b\to i q$ of this formula.
We can check that this is the right answer for our problem by noticing the following. First one can check that this expression is invariant under the SL(2) symmetry $L_a = L^1_a + L^2_a$ where $L_a$ are the generators \nref{SLtwoGen} acting on $\boldsymbol{x_1}$
and $L_a^2$ are similar generators as in \nref{SLtwoGen}, but with $q\to -q$. It is possible to commute
the phase $e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})}$, in \nref{particle heat kernel} past these generators which would remove the $q$ dependent terms. This implies
that the rest should be a function of the proper distance, which is the case with \nref{particle heat kernel}.
Then we can check the equation
\begin{eqnarray}
0 &=& (\partial_u + H_1 )G(u,\boldsymbol{x_1},\boldsymbol{x_2})
\end{eqnarray}
which is also indeed obeyed by this expression. The $s$ dependent prefactor is fixed by the requirement that the
propagator composes properly, or more precisely, by saying that for $u=0$ we should get a $\delta$ function.
\subsection{Partition Function}
The gravitational partition function is related with the particle partition function with inverse temperature $\beta$.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{Freeenergy.pdf}
\caption{Free Energy diagram with inverse temperature $\beta$. }
\end{figure}
The canonical partition function of the quantum mechanical system is
\begin{eqnarray}
Z_{\scaleto{Particle}{4pt}}&=&Tr e^{-\beta H}
=\int_0^{\infty} ds \int_{-\infty}^{\infty}dk \int_M {d x d y\over y^2} e^{-{\beta } { s^2 \over 2 } }f^*_{s,k}(x,y)f_{s,k}(x,y)\nonumber\\
&=&V_{AdS}\int_0^{\infty}ds e^{-\beta {s^2\over 2}}{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}.
\end{eqnarray}
The volume factor $V_{AdS}$ arises because after momentum integration there is no position dependence.
In a normal quantum mechanical system, the volume factor means that the particle can have independent configurations at different locations of our space, however for a gravitational system this should be thought as redundant and should be cancelled by the volume of $SL(2,R) $ gauge group $2\pi V_{AdS}$\footnote{There might be a multiplicative factor in the volume of gauge group, but we can always absorb that into $S_0$.}.
In gravitational system, there can also other contributions to the entropy from pure topological action.
These give a contribution to the ground state entropy $S_0$. Including the topological action in (\ref{regularized action}), we find a gravitational ``density of states" as
\begin{equation} \label{Spectral}
\rho(s)=\underbrace{e^{S_0}e^{2\pi q}}_{\text{extra terms}}\underbrace{1\over {2\pi}}_{\text{residue gauge}}\underbrace{{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}}_{\text{particle in magnetic field}}=e^{S_0}e^{2\pi q}{s\over 2\pi^2}\sum_{k=1}^{\infty}(-1)^{k-1}e^{-2\pi q k}\sinh(2\pi s k).
\end{equation}
We have not given an explicit description of these states in the Lorentzian theory. More details were discussed in \cite{Kitaev:2018wpr,Lin:2018xkj}.
This expression has some interesting features. Notice that the classical limit corresponds to large $q$ and large $s$, where we reproduce \nref{ClassEnt}. After approximating, the density of states are
$\log \rho(s) \sim S_0 + 2 \pi s $ for $ s/q < 1$ and $S_0+2\pi q$ for $s/q>1$.
We can also expand the partition function for very small and very large temperatures where we obtain
\begin{eqnarray}
Z_{JT} &\sim & e^{ S_0} e^{ 2 \pi q } { 1 \over 4\pi^2\beta} ~,~~~~~~~~~~~ \beta \ll {1\over q}
\cr
Z_{JT} & \sim & e^{S_0} { 1 \over \sqrt{2\pi}\beta^{3/2} }~,~~~~~~~~~~~~ \beta \gg 1
\end{eqnarray}
Notice that at leading order we get an almost constant entropy both at low and high temperatures, with the
high temperature one being higher. In both cases there are power law corrections in temperature.
Before we try to further elucidate the interpretation of this result, let us emphasize a couple of important defects of our discussion. First, when we replaced the partition function of the JT theory by the action of a non-relativistic
particle in an electric field, we were summing over paths in $H_2$. This includes paths that self intersect see figure \ref{fig:TwoInstanton}. Such paths do not have an obvious interpretation in the JT theory and it is not even clear that we should include them. For example, the sum over $k$ in \nref{Spectral} can be understood in terms of classical solutions which wind $k$ times around the circle. These make sense for the problem of the particle in the electric field but apparently not in the JT theory.
\begin{figure*}[h!]
\centering
\subfigure[Density of States]{%
\includegraphics[width=0.4\textwidth]{densityofstate.pdf}%
\label{fig:dos}%
}\hspace{4cm
\subfigure[Two Instantons]{%
\includegraphics[width=0.35\textwidth]{twoinstanton.pdf}%
\label{fig:TwoInstanton}%
\caption{}
\end{figure*}
Maybe such paths could be given some interpretation in the gravity theory.
Alternatively, we might want to sum over paths that do not self intersect.
A second defect is that we would be eventually interested in adding some matter fields propagating in the bulk geometry. These matter fields have boundary conditions at the boundary of the region of hyperbolic space
cut out by the boundary trajectory. The partition function of the fields with such an arbitrary boundary trajectory could also modify the results we described above. Of course, this issue does not arise if we have
the pure JT theory. It is only important if we want to introduce bulk matter fields to define more complex observables.
Instead of attempting to address the above issues, we will take an easy route, which is to consider the system only in the large $q$ (or small $\epsilon$) limit. In this regime, we address the above issues, and we can still trust the description of the particle in the electric field.
This large $q$ or small $\epsilon$ limit is the same one that isolates the Schwarzian action from the JT theory
\cite{Maldacena:2016upp,Jensen:2016pah,Engelsoy:2016xyb}.
It turns out that the limit can be taken already at the level of the mechanical system, a simple rescaled
version of the above system. This provides an alternative method for quantizing the Schwarzian theory. It has the advantage of being a straightforward second order action of a particle moving in a region near the boundary of hyperbolic. Of course, the Schwarzian theory was already quantized using a variety of methods in
\cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Luca}. We will simply provide yet another perspective, recover the old results, and write a few new expressions.
\section{Quantum Gravity at Schwarzian Limit}
Before getting into the details notice that the large $q$ limit of \nref{Spectral} gives
\begin{equation}
\rho(s) = e^{S_0}{s\over 2\pi^2}\sinh(2\pi s), ~~~E = { s^2 \over 2} ~,~~~~~~~Z_{JT} = \int_0^\infty ds
\rho(s) e^{ - \beta { s^2 \over 2 } } =e^{S_0}{1 \over\sqrt{2\pi}\beta^{3\over 2}}e^{2\pi^2 \over \beta}.
\end{equation}
This reproduces what was found in \cite{Stanford:2017thb,Mertens:2017mtv,Mertens:2018fds,Blommaert:2018oro} by
other methods. We see that we get a finite answer and also that the contributions from the $k>1$ terms in
\nref{Spectral} have disappeared. Because the $S_0$ part decouples with JT gravity, from now on, we will drop it and discuss $S_0$ only when it is necessary.
\subsection{The Propagator}
To get a limit directly at the level of the mechanical system it is useful to define a
rescaled
coordinate, $z$, via
\begin{equation}
y= z/q. \label{yAndq}
\end{equation}
After taking the large $q$ limit, the boundary particle propagator becomes \footnote{see the Appendix \ref{appendix:large q} for details}:
\begin{eqnarray}\label{heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&{ 1 \over q } e^{-2\pi q \theta(x_2-x_1) }\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2});~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~q\gg 1.\\
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{2\sqrt{z_1z_2}}{\pi^2 |x_1-x_2|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|});~~~~~~~~
\label{HeatLim}\\
&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{\sqrt{2}}{\pi^{3/2} u^{3/2}}\frac{\sqrt{z_1z_2}}{ |x_1-x_2|}\int_{-\infty}^{\infty}d\xi (\pi+i\xi)e^{-2\frac{(\xi-i\pi)^2}{u}-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}.\label{HeatLim2}
\end{eqnarray}
The original phase factor $e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_1})}$ factorizes into a product of singular ``phase" $e^{-2\pi q \theta(x_2-x_1)}$, with $\theta$ the step function,
and a regular ``phase" $e^{-2 {z_1+z_2\over x_1-x_2}}$. The singular ``phase" is the same order as the topological piece in (\ref{regularized action}).
In order to have a finite result they should cancel between each other. This can only be satisfied if the $x_is$ are in cyclic order.
As shown in figure (\ref{fig:phasefactor}), the product of singular ``phase" gives $-2\pi q$ for cyclic order $x_i$s and this would cancel with the topological action $2\pi q$.
While for other ordering of the $x_i$s, this would have $-2 \pi n q$ for $n=2,3,...$ and is highly suppressed in the limit $q$ goes to infinity.
This cyclic order is telling us where the interior of our space time is. The magnetic field produces a preferred orientation for the propagator.
After fixing the order, all our formulas only depend on $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ which has no $q$ dependence.
The residual $q$ factor in \ref{heat kernel} cancels out the additional $q$ from the measure of coordinate integral, ${ dx dy \over y^2 } \to q { dx dz \over z^2 } $.
In conclusion, after taking the limit we get a finite propagator equal to \nref{HeatLim}, which should be multiplied by a step function $ \theta( x_1 - x_2 ) $ that imposes the right order.
The final function $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ has the structure of $e^{-2{z_1+z_2\over x_1-x_2}}f(u,{z_1z_2\over (x_1-x_2)^2})$. This can be understood directly from the $SL(2)$ symmetry.
After taking the large $q$ limit, the $SL(2,R)$ charges become
\begin{equation} \label{SLTwoLq}
L_0=i (x\partial_x+z\partial_z);~~~~L_{-1}=i\partial_x;~~~~~~L_{1}=-ix^2\partial_x-2ix z\partial_z-2i z.
\end{equation}
We can check that they still satisfy the SL(2) algebra.
If we drop the last term in $L_1$, the $SL(2,R)$ charges become the usual differential operators on $EAdS_2$. And the propagator will have only dependence on the geodesic distance. When $L_1$ operator is deformed, the condition of $SL(2,R)$ invariance fixes the structure of the propagator as follows.
The $L_0$ and $L_{-1}$ charges are not deformed and they imply that the
only combinations that can appear are
\begin{equation}
v \equiv {z_1+z_2\over x_1-x_2} ~~~~~~~~~~~~~\text{and}~~~~~~~~~~~~~~ w\equiv {z_1 z_2\over(x_1-x_2)^2}.
\end{equation}
Writing the propagator as
$\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_1})=k(v,w) $
and
requiring it to be invariant under $L_1$ gives the following
equation for $\alpha$:
\begin{equation}\label{phaseequation}
\partial_v k +2 k =0 \longrightarrow k = e^{ - 2 v } h(w)
\end{equation}
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=e^{-2 {z_1+z_2\over x_1-x_2}}f(u,{z_1 z_2\over(x_1-x_2)^2}),
\end{equation}
The full function can also be determined directly as follows. Again we impose the propagator equation (or heat equation)
\begin{eqnarray}
0 &=& \left[ \partial_u + {1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) -{1\over 8} \right] \tilde K
\cr
0 & = & [ { s^2 \over 2} + {w^2\over 2}\partial_w^2+2w +{1\over 8}] K_s(w)
\end{eqnarray}
where $L_a$ are given in \nref{SLTwoLq} and are acting only on the first argument of $\tilde K$.
The solution of the last equation which is regular at short distances ($w \to \infty$) is $\sqrt{w}$ times
the Bessel K
function in \nref{heat kernel}.
We can also directly determine the measure of integration for $s$ by demanding that the propagator at $u=0$ is
a $\delta$ function or by demanding the propagator compose properly. This indeed is the case with the $s\sinh{ 2 \pi s}$ function in \nref{HeatLim}.
To explicitly show the above statement, it will be useful to use spectral decomposition of the propagator:
\begin{equation}\label{SpectralDecomposition}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=\int ds {2s\sinh(2\pi s)\over \pi^3} e^{-{s^2u\over 2}}\int dk \sqrt{z_1 z_2} e^{ik(x_1-x_2)}K_{2is}(2\sqrt{2i k z_1})K_{2is}( 2\sqrt{2ikz_2}).
\end{equation}
It can be easily checked that the special functions $f_{k,s}(x,z)=\sqrt{z}e^{ikx}K_{2is}( 2\sqrt{2ikz})$ are delta function normalizable eigenmodes of the large $q$ Hamiltonian:
\begin{equation}
\int {dx dz\over z^2}f_{k_1,s_1}f_{k_2,s_2}=\delta(k_1-k_2)\delta(s_1-s_2){\pi^3\over 2s\sinh (2\pi s)}
\end{equation}
Notice that the inner product fixes the integral measure completely in (\ref{SpectralDecomposition}), and the composition relation is manifestly true:
\begin{equation}
\int {dx dz\over z^2}\tilde{K}(u_1,\boldsymbol{x_1},\boldsymbol{x})\tilde{K}(u_2,\boldsymbol{x},\boldsymbol{x_2})=\tilde{K}(u_1+u_2,\boldsymbol{x_1},\boldsymbol{x_2})
\end{equation}
At short time the propagator has the classical behavior:
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})\sim \delta(x_1-x_2+u z_2)e^{-{(z_1-z_2)^2\over 2 u z_2}}
\end{equation}
This form of singularity is expected since we are taking the large $q$ limit first and thus the velocity in $x$ direction is fixed to be $z$. In the original picture of finite $q$ we are looking at the time scale which is large compare to AdS length but relatively small such that the quantum fluctuations are not gathered yet.
The integral structure in the propagator (\ref{HeatLim}) has an obvious meaning: integrating over $s$ represents summing over all energy states with Boltzmann distribution $e^{-E u}$, and the Bessel function stands for fixed energy propagator.
We want to stress that the argument in the Bessel function is unusual, and at short distance it approaches a funny limit:
\begin{equation}
K_{2is}(\frac{4 }{\ell})\simeq \sqrt{\pi \over 8\ell}e^{-\frac{4 }{\ell}},~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}\rightarrow 0.
\end{equation}
One should contrast this exponential suppression with the short distance divergence in QFT which is power law.
In our later discussion of exact correlation function with gravity backreaction, we will see that this effect kills UV divergence from matter fields.
To obtain the expression (\ref{HeatLim2}), we use the integral representation for the Bessel function and the final result has some interesting physical properties:
Firstly, we see that at large $u$ the time dependence and coordinate dependence factorized.
So, at large time we have a universal power law decay pointed out in \cite{Bagrets:2016cdf}.
Secondly, as we said before, the phase factor $e^{-2\frac{z_1+z_2}{x_1-x_2}}$ is equal to the Wilson line $e^{-q\int_1^2 a}$ stretched along the geodesic connection between location $1$ and $2$ (Figure \ref{fig:propagator}). The field $a$ depends on our choice of gauge, our convention corresponds to fix the minimum value of $a$ at infinity and then the Wilson line is equal to $e^{-q A}$, where $A$ is the area of a hyperbolic triangle spanned by $1,2$ and $\infty$.
Thirdly, defining $2\pi+2i \xi$ as $\theta$, then $\theta$ has the meaning of the spanned angle at the horizon (Figure \ref{fig:propagator}). Then the gaussian weight $e^{-2{(\xi-i\pi)^2\over u}}=e^{\theta^2\over 2u}$ can be understood from the classical action along the boundary with fixed span angle $\theta$. The boundary drawn in the figure represents a curve with fixed (regularized) proper length $u$ in $H_2$.
Lastly, the factor $e^{-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}=e^{-{4\cos\theta\over \ell}}$ is equal to $e^{q(\alpha+\beta)}$, which is a corner term that arise from JT gravity in geometry with jump angles. Here $\alpha$ and $\beta$ are defined as the angle spanned by the geodesic with fixed length and the ray coming from horizon to the boundary.
In summary the propagator can be understood as an integral of JT gravity partition functions over geometries \ref{fig:propagator} with different $\theta$s.
\begin{figure*}[h!]
\centering
\subfigure[singular ``phase" factor for different ordering]{%
\includegraphics[width=0.45\textwidth]{phase_factor.pdf}%
\label{fig:phasefactor}%
}\hspace{1.5cm
\subfigure[A geometric representation of the propagator. Here we fix the span angle $\theta$, the propagator is a summation over such geometries.]{%
\includegraphics[width=0.45\textwidth]{propagator.pdf}%
\label{fig:propagator}%
\caption{}
\label{fig:Propagator}
\end{figure*}
Finally, let us comment on the issues we raised in section 4.2. In the large $q$ limit we are considering the propagator at relatively large distances and in a regime where locally in $AdS$ the integration over paths that fluctuate
wildly is suppressed. Alternatively we can say that in the integration over paths we put a UV cutoff which
is large compared to $1/q$ but small compared to the AdS radius. This is the non-relativistic regime for the
boundary particle. The quantum effects are still important at much longer distances due to the large size
of $AdS$.
In addition, if we have quantum fields in AdS, then their partition functions for these fluctuating contours that
have fluctuations over distances larger than the $AdS$ radius are expected to depend on this shape in a local way. Due to the symmetries of $AdS_2$, this is simply expected to renormalize the action we already have without introducing extra terms. This can be checked explicitly for conformal field theories by using
the conformal anomaly to compute the effective action of the CFT$_2$ on a portion of $H_2$ (Appendix \ref{appendix: CFT effective action}).
\subsection{Wheeler-DeWitt Wavefunction }
\label{WdW}
In the pure JT theory we can think about quantizing the bulk theory and obtaining the
Wheeler-DeWitt wavefunction. This was discussed in the classical limit by Harlow and Jafferis \cite{Harlow:2018tqv}.
The Wheeler-DeWitt wavefunction can be created by Euclidean evolution of the boundary and hence is closely related to the propagator we have discussed above. The wavefunction in Lorentzian signature could then be obtained by analytic continuation of the boundary time.
The Euclidean evolution can be specified by either of the two parameters: the proper length $u$ or energy $E$. Choosing a different parameter corresponds to imposing a different boundary condition in JT theory. In general there are four possible choices of boundary conditions in 2d dilaton gravity, there are two sets of conjugate variables: $\lbrace \phi_b,K\rbrace$, and $\lbrace u,E\rbrace$ \footnote{Energy $E$ is proportional to the normal derivative of the dilaton field at the boundary.}. In preparing the wavefunction we fix the boundary value of dilaton and hence there are only two choices of the parameter ($u$ or $E$). We denote the corresponding wavefunction as $|u\rangle_G$ and $|E\rangle_G$ respectively.
In terms of holographic considerations, $|u\rangle_G$ represents a thermofield double state:
\begin{equation}
|u\rangle_G\sim\sum\limits_{n}e^{-E_n u}|E_n\rangle_L |\bar E_n\rangle_R
\end{equation}
and $|E\rangle_G $ is like an average of energy eigenstates in a window of energy $E$:
\begin{equation}\label{EnergyStateDefinition}
|E\rangle_G\sim {1\over \delta E}\sum\limits_{ |E-E_n|<\delta E}|E_n\rangle_L |\bar E_n\rangle_R.
\end{equation}
The width of the energy window is some coarse graining factor such that the summation contains $e^{S_0}$ states and does not show up clearly in gravity.\footnote{If one understand getting $|E\rangle$ state from integrating over thermofield double state in time direction, then a natural estimate on $\delta E$ is ${1\over T}$, where $T$ is the total time one integrate over. The validity of JT description of boundary theory is $T<e^{S_0}$, and we get $\delta E>e^{-S_0}$. For $T>e^{S_0}$, there are other possible instanton contributions. The proper gravitational theory at this regime is studied in paper \cite{Saad:2018bqo} }
With the definition of the states, one can evaluate them in terms of different basis. There are three natural bases turn out to be useful, we call them $S$, $\eta$ and $\ell$ bases. Basis $S$ corresponding to fix the horizon value of dilaton field $\phi_h$, or equivalently by Bekenstein-Hawking formula, the entropy of the system. The canonical conjugate variable of $S$ will be called $\eta$ and that characterizes the boost angle at the horizon. $\ell$ stands for fixing geodesic distance between two boundary points. To see that the horizon value of the dilaton field is a gauge invariant quantity, one can do canonical analysis of JT gravity. With ADM decomposition of the spacetime metric, one can get the canonical momenta and Hamiltonian constraints of the system \cite{LouisMartinez:1993eh}:
\begin{eqnarray}
ds^2&=&-N^2dt^2+\sigma^2(dx+N^xdt)^2;~~~~~~~~~~~~~~~~~~~~\\
\mathcal{H}&=&-\Pi_{\phi}\Pi_{\sigma}+\sigma^{-1}\phi''-\sigma^{-2}\sigma'\phi'-\sigma\phi;~~~~\mathcal{H}_x=\Pi_{\phi}\phi'-\sigma\Pi_{\sigma}';\label{Hamiltonian Constraint}\\
\Pi_{\phi}&=&N^{-1}(-\dot{\sigma}+(N^x\sigma)')=K\sigma;~~~~~~~~~~~~~\Pi_{\sigma}=N^{-1}(-\dot{\phi}+N^x\phi')=\partial_n\phi.\label{Canonical Momentum}
\end{eqnarray}
That is the dilaton field is canonically conjugate to the extrinsic curvature and boundary metric is canonical conjugate to the normal derivative of the dilaton field (both are pointing inwards).
By a linear combination of the Hamiltonian constraints (\ref{Hamiltonian Constraint}), one can construct the following gauge invariant quantity $C$:
\begin{equation}
-{1\over\sigma}(\phi'\mathcal{H}+\Pi_{\sigma}\mathcal{H}_x)={1\over 2}(\Pi_{\sigma}^2+\phi^2-{\phi'^2\over \sigma^2})'\equiv C[\Pi_{\sigma},\phi,\sigma]'\sim 0
\end{equation}
The Dirac quantization scheme then tells us that the quantity $C$ has a constant mode which is gauge invariant (commute with Hamiltonian constraint).
Choosing the gauge that normal derivative of the dilaton is zero, we can solve the Hamiltonian constraint:
\begin{equation}
\phi^2-(\partial_{X}\phi)^2=2C\equiv S^2~~~~~\rightarrow~~~~~~\phi(X)=S\cosh X,
\end{equation}
where $dX= \sigma dx$ is the proper distance along the spatial slice. Because the normal derivative of dilaton field is zero, the minimum value of dilaton at this spatial slice is actually a local extremum in both directions. Therefore, the minimal value of dilaton field, namely $S$, is a global variable. The classical geometry in this gauge is a ``Pac-Man" shape (right figure in figure \ref{fig:Pac-Man}).
Focusing on the intersection region of the spatial slice and the boundary, we have the spatial slice is orthogonal to the boundary.
This is because we are gauge fixing $\partial_n\phi=0$ on the spatial slice, and $\phi=\phi_b$ on the boundary.
The ADM mass of the system, after regularization, is then $M=\phi_b(\phi_b-\partial_{X}\phi)$ \cite{Maldacena:2016upp}. Substituting the behavior of $\phi(X)$ we get:
\begin{equation}
M={S^2\over 2}.
\end{equation}
This is the same relation in \ref{energy relation} and therefore we can interpret the $s$ variable as entropy of our system $S$.
For the purpose of fixing geodesic distance, it is convenient to think of doing the path integral up to a slice $L$ with zero extrinsic curvature. This picks out a particular slice (left figure in Figure \ref{fig:Pac-Man}) among the solutions obeying the Hamiltonian constraint.
The WdW wavefunction can be evaluated as an Euclidean path integral with fixed (rescaled) geodesic distance $d$ between the two boundary points:
\begin{eqnarray}\label{WdWPath}
\nonumber \Psi(u;d)=\int {\cal D}g{\cal D}\phi e^{{1\over 2}\int \phi(R+2)+\int_{L}\phi K+\phi_b\int_{Bdy}(K-1)}=\int {\cal D }f e^{\phi_b\int_{Bdy} (K-1)}\\
=\int {\cal D}\boldsymbol x e^{-m \int_{Bdy}\sqrt{g}+q\int_{Bdy} a+q\int_{L} a+(\pi-\alpha_1-\alpha_2)q}~~~~~~~~~~~~~~~~~~~~~~~
\end{eqnarray}
Here we are fixing the total length of $L$ to be $d$ and the proper length of the boundary to be $u$. $\alpha_1$ and $\alpha_2$ in the last expression denotes the jump angle at the corner coming from the singular contribution of the extrinsic curvature and should be integrated over. Without the $e^{q\int_L a}$ factor in (\ref{WdWPath}), the path integral corresponds to the propagator (\ref{HeatLim}). Remember that the phase factor is equal to $e^{-q\int_L a}$, so the wavefunction in $\ell$ basis is actually the propagator (\ref{HeatLim}), with the phase factor stripped out
\begin{equation}\label{wavefunction}
\Psi(u; \ell)\equiv \langle \ell|u\rangle_G = \frac{2}{\pi^2 \ell}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 }{\ell}) ,~~~~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}.
\end{equation}
$\ell$ is a function of the regularized geodesic distance $d$ between $\boldsymbol{x_1}$ and $\boldsymbol{x_2}$ : $\ell=e^{d\over 2}$.
The semiclassical of $\Psi(u;\ell)$ can be obtained using formula (\ref{HeatLim2}), in the exponent we get saddle point result:
\begin{equation}
\Psi(u;\ell)\sim \exp[-{2(\xi_*-i\pi)^2\over u}+{4\over u}{\xi_*-i\pi\over \tanh \xi_*}];~~~~~~~~{\xi_*-i\pi\over u}=-{\sinh \xi_*\over \ell}.
\end{equation}
The same saddle point equation and classical action was obtained in \cite{Harlow:2018tqv} by a direct evaluation in JT gravity.
The wavefunction with fixed energy boundary condition can obtained by multiplying $\Psi(u;\ell)$ by
$e^{ + E u} $ and integrating over $u$ along the imaginary axis. This sets $E = {s^2 \over 2} $
in the above integral over $s$. So this wavefunction has a very simple expression:
\begin{equation}\label{fixed energy wavefunction}
\Psi(E;\ell)\equiv \langle \ell|E\rangle_G = \rho(E)\frac{4}{ \ell} K_{i \sqrt{ 8 E} }(\frac{4 }{\ell}).~~~~~~~~
\end{equation}
The classical geometry for $\Psi(E;\ell)$ is the same as the left figure in Figure \ref{fig:Pac-Man}, with fixing energy on the boundary.
We want to stress that it is important to have the $\rho(E)$ factor in (\ref{fixed energy wavefunction}) for a classical geometry description since we are averaging over the states.
We can roughly think of ${4\over \ell}K_{i\sqrt{8E}}({4\over \ell})$ as a gravitational ``microstate" $|\mathcal{E}\rangle$ with fixed energy $E$. Such a ``microstate" will not have a classical geometry representation and therefore is just a formal definition.
The inner product between wavefunctions is defined as $\langle\Psi_1|\Psi_2\rangle=\int_0^{\infty}d\ell \ell \Psi_1^*(\ell)\Psi_2(\ell)$.
Going to the entropy basis $S$, it is easy to start with $\Psi(E)$.
Because of the identity $E={S^2\over 2}$, expanding $\Psi(E)$ in the $S$ basis is diagonal:
\begin{equation}
\Psi(E;S)\equiv \langle S|E\rangle_G=\sqrt{\rho(S)}\delta(E-{S^2\over 2})
\end{equation}
We put this square root of $\rho(S)$ factor in the definition of $S$ basis such that inner product between different $S$ state is a delta function $\langle S|S'\rangle=\delta(S-S')$. This factor is also required such that the classical limit matches with gravity calculation.
Integrating over energy with Boltzman distribution, we can get the expression of thermofield double state in the $S$ basis:
\begin{equation}
\Psi(u;S)\equiv \langle S|u\rangle_G=\int dE e^{-u E}\langle S|E\rangle_G=\sqrt{\rho(S)}e^{-{uS^2\over 2}}
\end{equation}
In the semiclassical limit, the wavefunction becomes gaussian and coincides with the on shell evaluation of the ``Pac-Man" geometry (Figure \ref{fig:Pac-Man}):
\begin{equation}\label{PsiS}
\Psi(u,S)\sim \sqrt{S}e^{\pi S -{u S^2\over 2}}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.5]{classicalgeometry.pdf}
\caption{Classical Geometry in $\ell$ and $S$ basis}
\label{fig:Pac-Man}
\end{figure}
The on shell calculation is straightforward: JT action in this geometry contains two parts: the Schwarzian action $\int(K-1)$ on the boundary and a corner contribution at the center: $S(\pi-\theta)$, where $\theta$ is the span angle at the horizon (Figure \ref{fig:Pac-Man}). The Schwarzian action simply gives $E u={S^2 u\over 2}$ by direct evaluation. We can determine $\theta$ from $u$ since they are related with redshift: $\theta=u S$. Therefore the corner term gives: $\pi S-S^2 u$. Adding them up then gives us the classical action.
We can also expand $S$ in terms of the $\ell$ basis, and relate $\Psi(\ell)$ with $\Psi(S)$ by a change of basis:
\begin{equation}
\langle\ell|S\rangle=\sqrt{\rho(S)}{4\over \ell}K_{2iS}({4\over \ell});~~~~~\Psi(E;\ell)=\int_0^{\infty} dS \langle \ell|S\rangle \langle S|E\rangle_G;
\end{equation}
Before discussing our last basis, we want to stress the simplicity of the wavefunction in $S$ basis (\ref{PsiS}) and the Gaussian factor resembles an ordinary particle wavefunction in momentum basis. We introduce our last basis $\eta$ as canonical conjugate variable of $S$, with an analog of going to position space of the particle picture in mind:
\begin{eqnarray}
|\eta\rangle &=&\int_0^{\infty} dS \cos(\eta S)|S\rangle;~~~~~~~~~~~~~\langle \ell|\eta\rangle=\int_0^{\infty} dS\cos(\eta S)\sqrt{\rho(S)}K_{2iS}({4\over \ell}){4\over \ell};~~~~~\\
\Psi(E,\eta) &=&\scaleto{\sqrt{\sinh(2\pi \sqrt{2E})\over 2\pi^2\sqrt{2E}}\cos(\eta \sqrt{2E})}{30pt};~~~~~\Psi(u,\eta)=\int_0^{\infty} dS \sqrt{\rho(S)}\cos(\eta S)e^{-{uS^2\over 2}}.~~~~
\end{eqnarray}
To understand the meaning of $\eta$ better, we can look at the classical behavior of $\Psi(u;\eta)$:
\begin{equation}
\Psi(u;\eta)\sim {1\over u}\exp[{\pi^2\over 2 u}-{\eta^2\over 2 u}]\left(e^{i\eta\pi\over u}\sqrt{\pi+i\eta}+e^{-{i\eta\pi\over u}}\sqrt{\pi-i\eta}\right)
\end{equation}
When $u$ is real, the wavefunction is concentrated at $\eta=0$ and has classical action of a half disk in the exponent.
When $u={\beta\over 2}+it$ which corresponds to the case of analytically continuing into Lorentzian signature, the density of the wavefunction $|\Psi(u,\eta)|^2$ is dominated by:
\begin{equation}
|\Psi({\beta\over 2}+it,\eta)|^2\sim {\sqrt{\pi^2+\eta^2}\over \beta^2+4t^2}\exp[{2\pi^2\over \beta}]\left(\exp[-{2\beta(\eta-{2\pi t\over \beta})^2\over \beta^2+4t^2}]+\exp[-{2\beta(\eta+{2\pi t\over \beta})^2\over \beta^2+4t^2}]\right)
\end{equation}
showing the fact that $\eta$ is peaked at the Rindler time ${2\pi t\over \beta}$. We can therefore think of fixing $\eta$ as fixing the IR time or the boost angle at the horizon.
The classical intuition for the boost angle is most clear in Euclidean geometry, where for fixed boundary proper time there can be different cusps at the horizon (Figure \ref{fig:Boostangle}).
\begin{figure}
\centering
\includegraphics[scale=0.4]{boostangle.pdf}
\caption{Euclidean geometries with different cusps.}
\label{fig:Boostangle}
\end{figure}
One application of those wavefunctions is that we can take an inner product and get the partition function.
However, there are also other ways to get the partition function.
For example, we can concatenate three propagators and integrate over their locations. This also gives the partition function by the composition rule of propagator.
By the relation between propagator and wavefunction, we can also view this as taking an inner product of three wavefunctions with an interior state as in figure \ref{fig:threewavefunction}, where the interior state can be understood as an entangled state for three universes.
To be more precise, we can view the wavefunction as the result of integrating the bulk up to the geodesics with zero extrinsic curvature. Then the interior state is given
by the area of the hyperbolic triangle in figure \ref{fig:threewavefunction}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{threewavefunction.pdf}
\caption{Partition function from inner product of three wavefunctions.}
\label{fig:threewavefunction}
\end{figure}
The path integral for the hyperbolic triangle (denoted as $I(\ell_{12},\ell_{23},\ell_{31})$, where $\ell_{ij}={|x_i-x_j|\over \sqrt{z_iz_j}}$), is a product of three phase factors, which satisfies a nontrivial equality (with ordering $x_1>x_2>x_3$):
\begin{equation}\label{GHZ}
\scaleto{ I(\ell_{12},\ell_{23},\ell_{31})=e^{-2({z_1+z_2\over x_1-x_2}+{z_2+z_3\over x_2-x_3}+{z_3+z_1\over x_3-x_1})}={16\over\pi^2}\int_0^{\infty}d\tau \tau \sinh (2\pi \tau) K_{2i\tau}({4\over \ell_{12}})K_{2i\tau}({4\over \ell_{23}})K_{2i\tau}({4\over \ell_{31}}).}{26pt}
\end{equation}
Recalling that the Bessel function represents the fixed energy ``microstate" $|\mathcal{E}\rangle$ (\ref{fixed energy wavefunction}) and ${\tau\over 2\pi^2} \sinh(2\pi \tau)$ is the density of state, this formula tells us that the interior state is a GHZ state for three universe:
\begin{equation}
I_{123}\sim\sum_n|\mathcal{E}_n\rangle_1|\mathcal{E}_n\rangle_2|\mathcal{E}_n\rangle_3.
\end{equation}
$I$ can also been viewed as a scattering amplitude from two universes into one universe. It constrains the SL(2,R) representation of the three wavefunctions to be the same.\footnote{Some thing similar happens for 2d Yang-Mills theory \cite{Witten:1992xu,Cordes:1994fc,Luca}.}
We can write down the partition function as:
\begin{equation}
Z_{JT}=\int_0^{\infty} \prod\limits_{\lbrace i j \rbrace \in \lbrace 12,23,31\rbrace}d\ell_{ij} \Psi(u_{12},\ell_{12})\Psi(u_{23},\ell_{23})\Psi(u_{31},\ell_{31})I_{\ell_{12},\ell_{23},\ell_{31}}.
\end{equation}
This same result also holds if we repeat the process $n$ times. It is interesting that we can view the full disk amplitude in these various ways.
One can also extend our analysis to include matter field. One type of such wavefunction can be created by inserting operator during Euclidean evolution, and is analysed in appendix B. Note that because of the SL(2,R) symmetry is a gauge symmetry, our final state has to be a gauge singlet including matter field.
\section{Correlation Functions in Quantum Gravity}
\subsection{Gravitational Feynman Diagram}
The propagator enables us to ``dress'' quantum field theory correlators to produce quantum gravity ones.
Namely, we imagine that we have some quantum field theory in $H_2$ and we compute correlation functions of
operators as we take the points close to the boundary where they take the form
\begin{equation}
\langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} = q^{ - \sum \Delta_i }z_1^{\Delta_1}..z_n^{\Delta_n}\langle O_1(x_1)...O_n(x_n)\rangle_{CFT}
\end{equation}
The factor of $q$ arises from \nref{yAndq}, and the last factor is simply defined as the function
that results after extracting the $z$ dependence. For example, for a two point function we get
\begin{equation}
\langle O_1(\boldsymbol{x_1})O_2(\boldsymbol{x_2})\rangle_{\rm QFT}=
q^{ - 2\Delta } z_1^{\Delta} z_2^{\Delta}{1\over |x_1-x_2|^{2\Delta}}.
\end{equation}
We can now use the propagator \nref{HeatLim} to couple the motion of the boundary and thus obtain the full
quantum gravity expression for the correlator. The factors of $q$ are absorbed as part of the renormalization
procedure for defining the full quantum gravity correlators.
In this way we obtain
\begin{figure*}[h!]
\centering
\subfigure[Witten Diagram]{%
\includegraphics[width=0.3\textwidth]{wittendiagram.pdf}%
\label{fig:Witten Diagram}%
}\hspace{4cm
\subfigure[Gravitational Feynman Diagram]{%
\includegraphics[width=0.3\textwidth]{dynamicalwittendiagram.pdf}%
\label{fig:Dynamical Witten Diagram}%
\caption{Summation of ${1\over N}$ effects fluctuates the boundary of Witten Diagram}
\label{fig:witten diagram}
\end{figure*}
\begin{eqnarray}\label{Higher Point Function}
\small{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{\rm QG}}}&=&e^{2\pi q}\int {\prod_{i=1}^n{dx_i dy_i\over y_i^2}\over \text{V(SL(2,R))}} G(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})G(u_{23},\boldsymbol{x_2},\boldsymbol{x_3})...G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})\times
\cr
& ~& \times \langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} q^{ \sum_ i \Delta_i }
\end{eqnarray}
where the left hand side is the full quantum gravity correlator by definition. The last factor is the usual
renormalization necessary to get something finite.
The factor of $e^{ 2 \pi q } $ cancels with the $q$ dependent ``phase'' factors in \nref{heat kernel} to give one if we order
the points cyclically ($x_1>x_2...>x_n$). This requires that we define more carefully the last propagator
$G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})$ as:
\begin{equation}
e^{-2\pi q}\tilde K(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})=e^{-2\pi q}e^{-2 \frac{z_n+z_1}{x_n-x_1}}\frac{2\sqrt{z_nz_1}}{\pi^2 |x_n-x_1|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2} u_{n1}}K_{2is}(\frac{4\sqrt{z_nz_1}}{|x_n-x_1|}).
\end{equation}
The factor ${1\over \text{V(SL(2,R))}}$ in \nref{Higher Point Function} means that we should fix the $SL(2,R)$ gauge symmetry (Appendix \ref{gauge fix}).
In the end we can write down an expression where we have already taken the $q \to \infty $ limit
\begin{equation}\label{final formula}
\boxed{\scaleto{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{QG}}=\int_{ x_1>x_2..>x_n}{\prod_{i=1}^n{dx_i dz_i}\over \text{V(SL(2,R))}} \tilde{K}(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})...\tilde{K}(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})z_1^{\Delta_1-2}..z_n^{\Delta_n-2}\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}.}{26pt}}
\end{equation}
This is one of the main results of our paper and it gives a detailed expression for correlation function in 2 dimensional quantum gravity in terms of the correlation functions of the QFT in hyperbolic space, or $AdS_2$.
Notice that in usual $AdS/CFT$ the correlators $\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}$ are an approximation to the full answer. This is sometimes computed by Witten diagrams.
We get a better approximation by integrating over the metric fluctuations.
In this case, the non-trivial gravitational mode is captured by the boundary propagator. The formula \nref{final formula} includes all the effects of quantum gravity in the JT theory (in the Schwarzian limit).
The final diagrams consist of the Witten diagrams for the field theory in $AdS$ plus the propagators for the boundary particle and we can call them ``Gravitational Feynman Diagrams", see figure \ref{fig:witten diagram}.
\subsection{Two Point Function}
Using formula \nref{final formula}, we can study gravitational effects on bulk fields such as its two point function:
\begin{eqnarray}
\langle O_1(u)O_2(0)\rangle_{QG}=
\includegraphics[width=50mm,trim=0 9cm 0 0]{2ptfunction.pdf}\\
~~~~~~~~~~~\nonumber\\
~~~~~~~~~\nonumber
\end{eqnarray}
The explicit expression for $\langle O_1(u)O_2(0)\rangle_{QG}$ with dimension $\Delta$ at temperature ${1\over \beta}$ is
\footnote{We will just keep the $\Delta$ dependent constant
since at last we will normalized with respect of partition function which corresponds to set $\Delta=0$.}:
\begin{equation}
\scaleto{{1\over \text{V(SL(2,R))}}\int_{x_1>x_2} {dx_1 dx_2 dz_1 dz_2 \over z_1^2 z_2^2}\int_0^{\infty}ds_1 ds_2 \rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}K_{2is_1}({4\sqrt{z_1z_2}\over |x_1-x_2|})K_{2is_2}({4\sqrt{z_1z_2}\over |x_1-x_2|})({\sqrt{z_1z_2}\over |x_1-x_2|})^{2\Delta+2}}{25pt}.
\end{equation}
To fix the $SL(2,R)$ gauge, we can choose $z_1=z_2=1$ and $x_2=0$.
Then the integral over $H_2$ space is reduced to a single integral over $x_1$,
with a Jacobian factor $2x_1$ (Appendix \ref{gauge fix}):
\begin{equation}\label{space integral}
\int_0^{\infty}ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\int_0^{\infty} dx_1 ({1\over x_1})^{2\Delta+1}K_{2is_1}({4\over x_1})K_{2is_2}({4\over x_1}).
\end{equation}
the last integral can be interpreted as a matrix element of two point operator $O_1O_2$ between states $|E_1,\psi\rangle$ and $|E_2,\psi\rangle$,
where $|\psi\rangle$ is the wavefunction of quantum field theory
and $|E\rangle_G$ represents the fixed energy gravitational state.
Integrating over $x$ can be thought as integrating over a particular gravitational basis, and we can see that the gravity wavefunction
suppress the UV contributions from quantum field theory ($K_{2is}({4\over x})\sim \sqrt{\pi x\over 8}e^{-4/x}$ for $x\sim 0$).
The final expression for the two point function is:
\begin{eqnarray}\label{2point function}
\langle O_1(u) O_2(0)\rangle_{QG}={1\over \mathcal{N}}\int ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\frac{|\Gamma(\Delta-i(s_1+s_2))\Gamma(\Delta+i(s_1-s_2))|^2}{2^{2\Delta+1} \Gamma(2\Delta)};~~\\
={1\over \mathcal{N}}{\Gamma(2\Delta)\over u^{3/2} (\beta-u)^{3/2} 2^{4\Delta+4}\pi^3} \int_{c-i\infty}^{c+i\infty} d\theta_1d\theta_2 \theta_1\theta_2 e^{{\theta_1^2\over 2u}+{\theta_2^2\over 2 (\beta-u)}}{1\over (\cos{\theta_1\over 2}+\cos{\theta_2\over 2})^{2\Delta}}.~~~~~
\end{eqnarray}
In the second expression we write the integral in terms of variable $\theta$ using the second integral representation of the propagator (\ref{HeatLim2}).
The normalization constant can be determined by taking the $\Delta=0$ limit: $\mathcal{N}=Z_{JT}$.
If we contemplate the result (\ref{2point function}) a little bit, then we find that the two integrals of $s_1$ and $s_2$ just represent the spectral decomposition of the two point function. Indeed, under spectral decomposition we have $\langle O(u)O(0)\rangle=\sum\limits_{n,m} e^{-E_n u-E_m (\beta-u)}|\langle E_n|O|E_m\rangle|^2$. Compare with (\ref{2point function}), we can read out the square of matrix element of operator $O$:
\begin{equation}
\scaleto{_G\langle E_1|O_L O_R|E_2\rangle_G=\delta E^{-2}\sum\limits_{\substack{|E_n-E_1|<\delta E\\|E_m-E_2|<\delta E}}|\langle E_n|O|E_m\rangle|^2=\rho(E_1)\rho(E_2){|\Gamma(\Delta-i(\sqrt{2E_1}+\sqrt{2E_2}))\Gamma(\Delta+i(\sqrt{2E_1}-\sqrt{2E_2}))|^2\over 2^{2\Delta+1}\Gamma(2\Delta)}.}{38pt}
\end{equation}
Remember the notation is that $|E\rangle_G$ stands for a gravitational state with energy $E$ and $|E_n\rangle$ stands for one side microstate (\ref{EnergyStateDefinition}).
We have put the measure $\rho(E)={1\over 2\pi^2}\sinh(2\pi\sqrt{2 E})$ in the definition of matrix element for the reason that in gravity it is more natural to consider an average of energy states as a bulk state.
To understand this formula a little bit better, we can consider the classical limit, namely large $E$. In this limit the matrix element squared can be approximated as a nonanalytic function:
\begin{equation}
_G\langle E_1|O_L O_R|E_2\rangle_G\propto |E_1-E_2|^{2\Delta-1}e^{2\pi \text{min}(\sqrt{2E_1},\sqrt{2E_2})}.
\end{equation}
If we fix $E_1$ and varying $E_2$ from $0$ to infinity, the matrix element changes from $|E_1-E_2|^{2\Delta-1}\rho(E_2)$ to $|E_1-E_2|^{2\Delta-1}\rho(E_1)$ after $E_2$ cross $E_1$. We can understand this behavior qualitatively as a statistical effect: the mapping from energy subspace $E_1$ to $E_2$ by operator $O$ is surjective when the Hilbert space dimension of $E_2$ is less that $E_1$ and is injective otherwise.
Another understanding is the following: the two point function is finite in a fixed energy state $|E_n\rangle$, which means the following summation of intermediate states $|E_m\rangle$ is order one: $\sum\limits_{m}|\langle E_n|O|E_m\rangle|^2$. Looking at the case $E_m>E_n$, because of the density of states grows rapidly, the matrix element squared has to be proportional to ${1\over \rho(E_m)}$ to get a finite result. Multiplied by $\rho(E_n)\rho(E_m)$, we have $\rho(E_n)\rho(E_m)|\langle E_n|O|E_m\rangle|^2\sim \rho(\text{min}(E_n,E_m))$.
\subsection{ETH and the KMS condition}
The Eigenstate Thermalization Hypothesis (ETH) is a general expectation for chaotic system. It expresses that the operator expectation value in an energy eigenstate can be approximated by thermal expectation value with effective temperature determined from the energy. Such hypothesis can be tested with the knowledge of operator matrix elements. The two point function in microcanonical essemble is:
\begin{eqnarray}
{1\over \delta E}\sum\limits_{|E_n-E|<\delta E}\langle E_n|O(u) O(0)|E_n\rangle &=&
\rho(E)e^{uE}\langle E|O e^{-u H}O|E\rangle\nonumber\\
\langle E|O e^{-u H}O|E\rangle &=&\int_0^{\infty} ds \rho(s)e^{-{s^2\over 2}u}\scaleto{{|\Gamma(\Delta+i(s+\sqrt{2E}))|^2|\Gamma(\Delta+i(s-\sqrt{2E}))|^2\over 2^{2\Delta+1} \Gamma(2\Delta)}}{25pt}.~~~~~~~~~\label{TwoPFMicro}
\end{eqnarray}
Notice that $|E\rangle$ is not $|E\rangle_G$, the former represents a one side microstate, while the later is a gravitational state. Accordingly $ \langle E|O(u)O(0)|E\rangle$ stands for a two point function in a microstate.
To study ETH, we will consider the case of a heavy black hole $E={S^2\over 2}={2\pi^2 \over\beta^2}\gg 1$. From the discussion in last section, we know that the matrix element tries to concentrate $s$ around $\sqrt{2E}$ and thus we can approximate $\rho(E)\rho(s)|\Gamma(\Delta+i(s+S))|^2$ as proportional to $sS^{2\Delta-1}e^{\pi(s+S)}$. Using integral representation for $|\Gamma(\Delta+i(s-S))|^2 $ we derive the two point function in microcanonical essemble with energy $E$ is proportional to:
\begin{equation}
{\rho(E)S^{2\Delta-1}\over u^{3/2}}\int d\xi (\pi+i\xi)e^{-{2\over u}(\xi-i{\pi\over 2})^2-(\pi+2i\xi) S+u{S^2\over 2}}{1\over(\cosh \xi)^{2\Delta}}.
\end{equation}
The $\xi$ variable can be understood as the measure of time separation in units of effective temperature between two operators and its fluctuation represents the fluctuation of the effective temperature.
And the final integral can be understood as a statistical average of correlation functions with different temperatures.
If we put back the Newton Constant $G_N={1\over N}$, we have $S\sim N$ and $u\sim N^{-1}$ (\ref{BCond}).
As can be seen from the probability distribution, the fluctuation is of order ${1\over \sqrt{N}}$, and hence for large $N$ system we can use saddle point approximation:
\begin{equation}
\xi=i({\pi\over 2}-{S\over 2}u) -{u\Delta \tanh \xi\over 2}=i({\pi\over 2}-{\pi u\over \beta})-{u\Delta\tanh \xi\over 2}.
\end{equation}
The first piece gives the typical temperature of the external state,
while the last piece comes from the backreaction of operator on the geometry. If we first take the limit of large N, one simply get that the two point function in microcanonical essemble is the same as canonical essemble.
However, the euclidean correlator in canonical essemble is divergent as euclidean time approach to inverse temperature $\beta$ because of KMS condition.
Such singular behavior plays no role in the microcanonical essemble so is called a ``forbidden singularity" in ETH \cite{Fitzpatrick:2016ive, Faulkner:2017hll}. In our situation we can see directly how the forbidden singularity disappears in the microcanonical essemble.
When $\xi$ approach $-i{\pi\over 2}$ at the forbidden singularity the backreaction on the geometry becomes large and hence the effective temperature becomes lower:
\begin{equation}
{2\pi\over\beta*}\rightarrow {2\pi\over\beta}-{\Delta \over \pi-{\pi u\over \beta}}.
\end{equation}
At the time ${\beta-u\over\beta}\sim{\Delta\over N}$, the backreaction is important and we expect to see deviation from thermal correlators.
Therefore the correlation function in microcanonical essemble will never have singularity away from coincide point.
\subsection{Three Point Function}
The bulk diagram of the three point function will be like Figure \ref{fig:threewavefunction} with additional operator inserting at the intersection points (See Figure \ref{fig:threepointfunction}).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{threepointfunction.pdf}
\caption{Bulk Diagram for Three Point Function}
\label{fig:threepointfunction}
\end{figure}
The QFT three point correlation function in $AdS_2$ is fixed by conformal symmetry and we can write it down as:
\begin{equation}
\langle O_1 O_2 O_3\rangle=C_{123}{z_1^{\Delta_1} z_2^{\Delta_2} z_3^{\Delta_3}\over |x_{12}|^{\Delta_1+\Delta_2-\Delta_3}|x_{23}|^{\Delta_2+\Delta_3-\Delta_1}|x_{13}|^{\Delta_1+\Delta_3-\Delta_2}}={C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
$\Delta_i$ is the conformal dimension of $O_i$. Putting them in formula \ref{final formula} and rewrite the propagator in terms of the wavefunction (\ref{wavefunction}), we have the quantum gravitational three point function:
\begin{equation}
\langle O_1 O_2 O_3\rangle_{QG}=\int_{x_1>x_2>x_3}{\prod_{i=1}^3{dx_i dz_i}\over \text{V(SL(2,R))}}\Psi_{u_{12},\ell_{12}}\Psi_{u_{23},\ell_{23}}\Psi_{u_{13},\ell_{13}}I_{\ell_{12},\ell_{23},\ell_{13}}{C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
We can view this expression as an inner product of three universe wavefunction with the interior, inserting three bilocal operators $\tilde O_{ij;k}\tilde O_{ij;k}$ with dimension $\tilde\Delta_{ij;k}={1\over 2}(\Delta_i+\Delta_j-\Delta_k)$ between them. One can fix the $SL(2,R)$ symmetry and express the integral in terms of $\ell_{ij}$, it is the same exercise as in open string calculation to find the Jacobian factor (Appendix \ref{gauge fix}). Here we can just argue that in order to get the partition function at $\Delta=0$, the measure has to be flat. Therefore the three point function factorizes into form:
\begin{equation}
\small{\langle O_1(u_{1})O_2(u_{2})O_3(u_{3})\rangle_{\text{\rm QG}}}\propto C_{123}\int_0^{\infty}d\tau \rho(\tau)\mathcal{I}_{\tau}(u_{12},\tilde\Delta_{12;3})\mathcal{I}_{\tau}(u_{23},\tilde\Delta_{23;1})\mathcal{I}_{\tau}(u_{31},\tilde\Delta_{31;2})
\end{equation}
while $\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})$ is an integral of $\ell_{ij}$ which gives the two point function in microstate $E_{\tau}$ (\ref{TwoPFMicro}) with the $e^{uE_{\tau}}$ factor stripped off:
\begin{equation}
\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})={1\over 2}\int_{0}^{\infty} d\ell_{ij}\Psi_{u_{ij},\ell_{ij}}{1\over \ell_{ij}^{\Delta_i+\Delta_j-\Delta_k}}K_{2i\tau}({4\over \ell_{ij}})=\langle E_{\tau} |\tilde O_{ij;k}e^{-u_{ij}H}\tilde O_{ij;k}|E_{\tau}\rangle;~~~~~E_{\tau}={\tau^2\over 2}.
\end{equation}
Again the normalization constant can be fixed by choosing $O_i$ to be identity.
\subsection{Einstein-Rosen Bridge}
The Einstein-Rosen Bridge in a classical wormhole keeps growing linearly with time and this behavior was conjectured to related with the growth of computational complexity of the dual quantum state \cite{Stanford:2014jda}.
Based on the universal behavior of complexity growth, Susskind proposed a gravitational conjecture in a recent paper \cite{Susskind:2018fmx} about the limitation of classical general relativity description of black hole interior. The conjecture was stated as follows:
\textsl{Classical general relativity governs the behavior of an ERB for as long as possible.}
In this section, we will test this conjecture using the exact quantum wavefunction of JT gravity (\ref{wavefunction}). We will in particular focusing on the behavior of ERB at time bigger than $1$. The size of Einstein-Rosen Bridge $\mathcal{V}$ in two dimensions is the geodesic distance $d$ between two boundaries, and can be calculated in thermofield double state $|u\rangle$ as:
\begin{equation}
\mathcal{V}=\langle u|d|u\rangle
\end{equation}
We want to focus on the dependence of volume on Lorentzian time evolution. Therefore we do analytic continuation of $u$ in Lorentzian time: $u={\beta\over 2}+it$. Using the WdW wavefunction in $\ell$ basis (\ref{wavefunction}) and the relation between $d$ and $\ell$, we can calculate the expectation value exactly. This can be done by taking the derivative of the two point function (\ref{2point function}) with respect to $\Delta$ at $\Delta=0$. Using the integral representation for $|\Gamma(\Delta+i(s_1-s_2))|^2$, the only time dependence of volume is given by:
\begin{equation}
\mathcal{V}(t)={1\over \mathcal{N}}\int_{-\infty}^{\infty}d\xi\int_0^{\infty} ds_1ds_2\rho_1\rho_2 e^{i(s_1-s_2)\xi-i(s_1-s_2){(s_1+s_2) \over 2}t-{\beta\over 4}(s_1^2+s_2^2)}\log(2\cosh{\xi\over 2})|\Gamma(is_1+is_2)|^2.
\end{equation}
The limit we are interested in is $\beta\ll 1\ll t$ \footnote{Remember that we are measuring time in units of $\phi_r$ (\ref{BCond}), so time order $1$ is a quantum gravity region.}, in which case the integral has a saddle point at \footnote{Actually this saddle point is valid for any range of $t$ as long as $\beta\ll 1$.}:
\begin{equation}\label{SaddlePoint}
s_{1,2}={2\pi\over \beta};~~~~~~~~~\xi={2\pi t\over \beta}.
\end{equation}
Therefore the volume has linear dependence in time:
\begin{equation}
\mathcal{V}(t)\sim{2\pi t\over \beta}.
\end{equation}
Using the complexity equal to volume conjecture \cite{Stanford:2014jda,Brown:2015bva}, the complexity of thermofield double state is proportional to the maximum volume:
\begin{equation}
\mathcal{C}(t)=\#\mathcal{V}(t)=\#{2\pi t \over\beta} .
\end{equation}
The proportionality constant is suggested in \cite{Brown:2018kvn} to be $S_0$ based on classical calculation of near extremal black hole. This, however, is not very clear in our model since $S_0$ is the coupling constant of the pure Einstein-Hilbert action and decouples with JT theory (\ref{full action}).
Since the saddle point (\ref{SaddlePoint}) is actually valid from early time to late time, the proportionality constant can be fixed at classical level and once we fix it we can conclude that \textsf{the length of Einstein-Rosen Bridge (or complexity of the state) keeps linearly growing even considering quantum gravity effects in JT theory.}
We want to comment that this is not an obvious result that one can expect from classical observables.
For example, one might argue that we can extract the information of the ERB from two sided correlators for the reason that semiclassically we can approximate the correlator as $e^{-m d}$.
Therefore one can conclude the ERB has linear growth from the quasinormal behavior of the correlator.
However such observables can only give us information of ERB up to time order $1$, which is the same time scale we can trust the classical general relativity calculation.
After that the correlation function changes from exponential decay into universal power law decay ${1\over t^{3}}$ as one can directly derive from analytic continuation of result (\ref{2point function}).
If we still use such correlator to extract information about ERB we would get the wrong conclusion that it stops its linear growth after time order $1$.
The reason why it is incorrect is that at this time scale the operator disturbs the state and causes different energy states interfere each other strongly. It is simply that the correlator can no longer be described by the classical geometry, rather than the interior stops to behave classically.
From our calculation, we see that if we probe of the state in a weaker and weaker way, we are still able to see the classical geometry.
Lastly, we want to talk a little about when JT gravity needs to be modified. A naive estimate can be made from the partition function that when $\beta$ approaches $e^{{2\over 3}S_0}$, the partition function becomes less than one and definitely at this time scale we need new physics. A recently study of gravitational physics at this time scale was discussed in \cite{Saad:2018bqo}.
\section{Conclusion}
Our result gives an explicit formula (\ref{final formula}) to calculate all order corrections to correlation functions from quantum gravity in two dimensions.
The formula can be understood diagrammatically and we call it Gravitational Feynman Diagram.
We also give the exact Wheeler-DeWitt wavefunction and discuss the growth of its complexity quantum mechanically.
Although we are focusing on theoretical description of two dimensional black holes, the near-extremal black holes in nature should contain these features.
Both Reissner–Nordström black holes and Kerr black holes have an $AdS_2$ throat near their extremality.
For those black holes, the gravitational effects are enhanced by the their near extremal entropies (the coupling constant is $\phi_h$ rather than $\phi_0+\phi_h$) and therefore are better backgrounds to test gravitational effects.
We should however point out that the observational black holes all have large near extremal entropies and thus are very classical \cite{Preskill:1991tb}.
In addition, the Thorne limit of Kerr black hole sets a lower bound on the near extremal entropy in nature.
But for the Primordial black holes in early universe, our story might play a role and it will be interesting to study the physical consequence in that situation.
Another application of our result is to connect with SYK type models \cite{Maldacena:2016hyu,Kitaev:2017awl,Witten:2016iux,Klebanov:2016xxf} since those models all have an emergent Schwarzian action at low energy. On that account, the exact Schwarzian quantization can be used to test ${1\over N}$ corrections to those models. For example, one can try to directly test the two point function with SYK numerics \cite{BKobrin} or one can use SYK models to understand the microscopic description of WdW wavefunction and its complexity. It might also be interesting to consider the black hole information paradox \cite{Fiola:1994ir} and late time traversable wormholes including the quantum fluctuations of the boundary \cite{Gao:2016bin,Maldacena:2017axo,Susskind:2017nto,Yoshida:2017non, Maldacena:2018lmt}.
{\bf Acknowledgments }
The author wants to give special thanks to J.Maldacena and D.Stanford for patient guidance and help all through the project.
We also want to thank A.Almheiri, L.Iliesiu, J.Jiang, J.Ripley, A.Kitaev, B.Kobrin, R.Mahajan, A.M.Polyakov, S.J.Suh, G.Turiaci and H.Verlinde for discussions.
Z.Y is supported by Charlotte Elizabeth Procter Fellowship from Princeton University.
\section{Introduction}
Near-Extremal black holes have a universal structure near their horizons:
there is an $AdS_2$ throat with a slowly varying internal space.
Its low energy gravitational dynamics is captured universally by the following effective action in two dimensions \cite{Strominger:1994tn}:
\begin{equation}\label{full action}
I=\underbrace{-{\phi_0\over 2
}\left(\int R+2\int_{\partial_M}K\right)}_{\text{Einstein-Hilbert Action}}\underbrace{-{1\over 2
}\left(\int_M \phi(R+2)+2\int_{\partial M}\phi_b K\right)}_{\text{Jackiw-Teitelboim action}}+S_{matter}(g,\psi),
\end{equation}
where the dilaton field $\phi+\phi_0$ represents the size of internal space. We have separated the size of internal space into two parts: $\phi_0$ is its value at extremality. It sets the value of the extremal entropy
which comes from the first term in \nref{full action}.
$\phi$ is the deviaton from this value.
We have also added matter that only couples to the metric. This is a reasonable assumption when matter comes
from Kaluza Klein reduction, where the coupling to the dilaton would involve $\phi/\phi_0 \ll 1$.
The action $\int \phi(R+2)+2\int\phi_b K$ is the so-called Jackiw-Teitelboim action \cite{JACKIW1985343,TEITELBOIM198341}, and will be the main focus of our paper.
This action is one of the simplest nontrivial gravitational actions in two dimensions.\footnote{Another nontrivial action is the CGHS model which could be written as $\int (\phi R+C)$, and that characterizes the horizon structure of general black holes.}. It is simple because the bulk geometry is a rigid $AdS_2$ space fixed by the equation of motion of the dilaton field. Its nontriviality arises from the remaining boundary action. Schematically, the gravitational action is reduced to the following form:
\begin{equation}\label{schematic action}
I=-{ 2 \pi \phi_0
}\chi(M)-{\phi_b
}\int_{\partial M}K +S_{matter}(g,\psi).
\end{equation}
And the motion of the boundary is controlled by its extrinsic curvature.
Our goal will be to quantize this action and to provide expression for the full quantum gravity correlators
of \nref{full action}.
This problem was considered before in \cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Kitaev:2018wpr} from various points of view. Here we will add one other point of view where we
reduce the problem to the motion of a relativistic particle in an electric field, building on a suggestion in
Kitaev's talk at IAS \cite{kitaevIAS}. More precisely, one can consider a relativistic particle in a Lorentzian $AdS_2$ target space moving under the influence of an electric field. The coupling to electric field can also be viewed as a coupling to a spin connection so that it becomes a particle with spin as suggested by Kitaev. Alternatively we can start from a non-relativistic particle moving in hyperbolic space, $H_2$, under the influence of a magnetic field $b$. After analytic
continuation in $b$ to imaginary values we get the problem of interest.
Using this point of view one can think of the full quantum gravity problem as the combination of two problems.
First we consider quantum fields propagating in $AdS_2$ (or $H_2$ in the Euclidean case) and then we
add the ``gravitational particle" which couples to the quantum fields by changing their boundary location in $AdS_2$.
The discussion of quantum fields will be standard and depends on the particular model one interested in, therefore we will mainly focus on solving the second problem.
Generically, solving the gravitational problem is challenging and is not exactly equivalent to a quantum mechanical particle. One needs to worry about what functional space one will integrate over. For example, in path integrals, one usually integrates over all trajectories including those with self-intersections. However self-intersecting boundaries in gravitational system have no obvious meaning. On that account, more precisely the gravitational problem is equal to a self-avioding particle.
Nevertheless, it turns out that one can take a particular limit of this model, namely large $\phi_b$, to avoid this issue and a treatment of the boundary theory as an ordinary particle is justified. It is also true that the JT gravity can be rewritten as a Schwarzian action only in this limit. We call this the Schwarzian limit and will only focus on solving the JT action in the Schwarzian limit. Solving the model away from Schwarzian limit was considered recently by Kitaev and Suh \cite{Kitaev:2018wpr}.
Our result can be summarized as follows:
\textsl{First, we will give a formula to calculate all correlation functions with quantum gravity backreaction (formula \ref{final formula}). Second, we will give the exact Wheeler-DeWitt wavefunction in the Schwarzian limit, which has been analyzed classically by Harlow and Jafferis \cite{Harlow:2018tqv}. Last, we consider the recent proposed conjecture about complexity growth in this exact Wheeler-DeWitt wavefunction and show that the complexity maintains linear growth after taking quantum gravity effects into account. This, to our knowledge, is the first test of the gravitational conjecture made by Susskind \cite{Susskind:2018fmx} that the size of ERB grows linearly for as long as quantum mechanics allows.}
This paper is organized as follows: in the first section, we will review the classical calculation of this model and introduce notations; in the second section, we will make the dictionary between the JT model and a particle in a magnetic field; in the third and fourth sections, we will solve the quantum mechanical problem and derive the propagator and WdW wavefunction in the Schwarzian limit; in the last section, we will talk about gravitaional backreaction on correlators as well as complexity growth. As a useful notion in our calculation, we introduce a notation called gravitational Feyman diagrams.
\section{Classical Solutions}
Let us first consider the classical solutions of
the Jackiw-Teitelboim model \nref{full action}. See
\cite{Almheiri:2014cka} for further discussion.
%
%
The equation of motion of the dilaton field imposes $R=-2$
and
fixes the geometry to be $AdS_2$, or $H_2$ in the Euclidean case.
This is also true if
we have additional matter coupled with metric only, as in
\nref{full action}.
The equations for the metric constrain the dilaton
\begin{equation}\label{Einstein Equation}
(\nabla_{\mu}\nabla_{\nu}\phi-g_{\mu\nu}\nabla^2\phi+g_{\mu\nu}\phi)+T^{M}_{\mu\nu}=0
\end{equation}
These equations are compatible with each other thanks to the conservation of the matter stress tensor. They do not allow
any propagating mode. In fact, setting $T^M_{\mu\nu}=0$, and using a high momentum approximation we can write
\nref{Einstein Equation} as $ (k_{\mu}k_{\nu}-g_{\mu\nu}k^2)\phi(\vec k)=0 $, which then implies $ \phi(\vec k)=0 $, for large $k$.
More precisely, after introducing the Euclidean $AdS_2$ coordinates $ds^2 = d\rho^2 + \sinh^2\rho d\varphi^2$, we can solve
(\ref{Einstein Equation}) in $AdS_2$ with no matter. Up to an SL(2) transformation the solution is
\begin{equation}
\phi=\phi_h \cosh \rho,
\end{equation}
where $\phi_h$ is a constant that is fixed by the boundary conditions.
At the boundary we fix the metric along the boundary and the value of the dilaton field
\begin{equation} \label{BCond}
ds_{\parallel} = d u { \phi_r \over \epsilon } ~,~~~~~~~~~~~ \phi= \phi_b = { \phi_r \over \epsilon }
\end{equation}
where we think of $u$ as the time of the boundary theory. It is simply a rescaled version of proper time. Similarly $\phi_r$ is a rescaled
value of the dilaton. We will be interested in taking $\epsilon \to 0$. With these rescalings the value of $\phi_h$ in the interior
remains fixed as we take $\epsilon \to 0$ as $\phi_h = 2 \pi /\beta $ where $\beta$ is the inverse temperature, $u \sim u + \beta$. Notice that due to the factor of $\phi_r$ in the first expression in \nref{BCond} we are measuring time in units of the constant $\phi_r$, which has dimensions of length. We did this for convenience.
A nice feature that appears after taking the $\epsilon \to 0$ limit is that the action (\ref{schematic action}) can be written as the Schwarzian action for the boundary curve labeled by $\varphi(u)$ \cite{Maldacena:2016upp}:
\begin{equation}
I=-\int du Sch(\tan{\varphi(u)\over 2},u)
\end{equation}
The fluctuation of the boundary shape can be understood as the fluctuation of the dilaton distribution in the bulk.
A bit more explicitly we can say that the dilaton boundary condition fixes the location of the boundary at $\rho_b$ given by
$\phi_b = \phi_h \cosh \rho_b$, and the metric at that location relates the time $\varphi$ to $u$ by $\phi_r du = \epsilon \sinh \rho_b d\varphi $.
We get the above formulas noticing that the period of $\varphi$ is $2\pi$ while that of $u$ is $\beta$, which fixes $\epsilon \sinh \rho_b$.
\section{Charged Particle in $AdS_2$}
Despite the absence of a bulk propagating mode there is still a non-trivial dynamical gravitational degree of freedom. There are various
ways to describe it. Here we will think of it as arising from the motion of the physical boundary of $AdS_2$ inside a rigid $AdS_2$ space.
This picture is most clear for finite $\epsilon$ in \nref{BCond}, but it is true even as $\epsilon \to 0$. The dynamics of the boundary
is SL(2) invariant. This SL(2) invariance is a gauge symmetry since it simply reflects the freedom we have for cutting out a piece
of $AdS_2$ space that we will call the ``inside". It is important that the dilaton field we discussed above is produced after we put in the
boundary and it moves together with the boundary under this SL(2) gauge transformation. It is a bit like the Mach principle, the location
in $AdS_2$ is only defined after we fix the boundary (or distant ``stars").
We can make this picture of a dynamical boundary more manifest as follows.
Since the bulk Jackiw-Teitelboim action \nref{full action} is linear in $\phi$, we can integrate out the dilaton field which
sets the metric to that of $AdS_2$ and removes the bulk term in the action, leaving only the term involving the extrinsic curvature
\begin{eqnarray}
I&=&-{\phi_r\over
\epsilon}\int du \sqrt{g}K
\end{eqnarray}
This action, however, is divergent as we take $\epsilon$ to zero. This divergence is simply proportional to the length of the boundary
and can be interpreted as a contribution to the ground state energy of the system.
So we introduce a counterterm proportional to the length of the boundary to cancel it. This is just a common shift of the energies
of all states.
It is also convenient to use the
Gauss-Bonnet theorem to relate the extrinsic curvature to an integral over the bulk
\begin{equation}
\int_{\partial_M} du \sqrt{g}K=2\pi\chi(M)-{1\over 2}\int_M R
\end{equation}
Since the curvature is a constant, the bulk integral is actually proportional to the total area $A$ of our space. That is we have the regularized action:
\begin{eqnarray}\label{subtraction}
I&=&-{\phi_r\over
\epsilon}\int_{\partial M} du \sqrt{g}(K-\underbrace{1}_{\text{counterterm}})=-{\phi_r\over
\epsilon}\left(2 \pi\chi(M)-{1\over 2}\int_{M}R-\int_{\partial M} du \sqrt{g}
\right) \nonumber\\
&=&-2\pi q\chi(M)-q A+q L,~~~~~~~~~~~~~~~~~~~~~~~~~~~q\equiv{\phi_r\over
\epsilon} ,~~L={\beta \phi_r\over \epsilon}
\end{eqnarray}
We now define an external gauge field $a_\mu$ as
\begin{equation}
a_{\varphi}=\cosh\rho-1,~~~~~~~a_{\rho}=0,~~~~~~~f_{\rho \varphi}=\sinh\rho=\sqrt{g},
\end{equation}
and write the action as follows
\begin{equation}\label{regularized action}
I= - 2 \pi q + q L - q \int a
\end{equation}
where we used that $ \chi(M)$ is a topological invariant equal to one, in our case, where the topology is that of a disk.
The term $q L$ is just the length of the boundary. So this action has a form somewhat
similar to the action of a relativistic charged particle moving in
$AdS_2$ in the presence of a constant electric field.
There are a couple of important differences. First we are summing only over trajectories of fixed proper length set by the inverse temperature
$\beta$. Second, in the JT theory we are treating the $SL(2)$ symmetry as a gauge symmetry. And finally, in the JT theory we identify the proper length with the boundary time, viewing configurations which differ only by a shift in proper time as inequivalent. In fact, all these changes simplify the problem: we can actually think
of the problem as a non-relativistic particle moving on $H_2$ in an electric field. In appendix
\ref{appendix:relativistic particle} we discuss in more detail the connection to the relativistic particle.
In fact, precisely the problem we are interested in has been discussed by Polyakov in \cite{Polyakov:1987ez}, Chapter 9, as an an intermediate step for the sum over paths.
Now we would also like to point out that we can directly get to the final formula by
using the discussion there, where he explicitly shows
that for a particle in flat space the sum over paths of fixed proper length that stretch between
two points $\vec x$ and $\vec x'$ gives
\begin{equation} \label{PolFlat}
\int {\cal D }\vec x e^{ - m_0 \tilde\tau } \delta( \dot { \vec x}^2 - 1 ) =
e^{ - {1\over 2} \mu^2 \tau} \langle x' | e^{ - \tau H } |x\rangle =
e^{ - {1\over 2} \mu^2 \tau} \int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot \vec x^2 }
\right)
\end{equation}
$\mu^2$ is the regularized mass and $\tilde \tau$ is related to $\tau$ by a multiplicative renormalization.
The JT model consists precisely of a functional integral of this form, where we fix the proper length along the boundary. There are two simple modifications, first the particle is in a curved $H_2$ space and second we have the coupling to the electric field. These are minor modifications, but the arguments leading
to \nref{PolFlat} continue to be valid so that the partition function of the JT model can be
written directly:
\begin{equation} \label{PolAdS}
\scaleto{\int {\cal D }\vec x e^{2\pi q - m_0 \tilde\tau +q\int a } \delta( { \dot x^2 + \dot y^2 \over y^2 }- q^2 ) =
e^{ 2\pi q- {1\over 2} \mu^2 \tau} \Tr e^{ - \tau H } = e^{2\pi q - {1\over 2} \mu^2 \tau }
\int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } - q {\dot x \over y}
\right)}{26pt}
\end{equation}
The delta function implements the first condition
in \nref{BCond} at each point along the path.
The last path integral can be done exactly by doing canonical quantization of the action (section \ref{CanonicalQuantization}) and by comparing the result with the one from the Schwarzian action \cite{Stanford:2017thb}
we can determine that $\tau$ is the inverse temperature $\beta$.
In the above discussion we have been fixing the time along the boundary. Instead we can fix the energy
at the boundary, where the energy is the variable conjugate to time.
This can be done by simply integrating \nref{PolAdS} times $e^{ \beta E}$
over $\beta$ along the imaginary axis. This fixes the energy of the non-relativistic problem by generating a $\delta( H - E)$. More precisely, we will argue that after doing a spectral decomposition we can write
the propagator at coincident points as
\begin{equation} \label{JTEn}
Z_{JT}(\beta ) = \int_0^\infty \rho(E)e^{-\beta E} dE \longrightarrow \rho(E) = \int_{-i \infty}^{i \infty}
{ d\beta \over i } e^{ E \beta } Z_{JT}(\beta)
\end{equation}
where the function $\rho(E) $ can then be interpreted as a ``density of states" in the microcannonical ensemble. We will give its explicit form in section (4.2).
For now, we only want to contrast this integral with a superficially similar one that appears when we
compute the relativistic propagator
\begin{equation} \label{RelEl}
e^{-2\pi q} \int_0^\infty e^{ E \beta }Z_{JT} (\beta) = \langle \phi(x) \phi(x) \rangle
\end{equation}
which gives the relativistic propagator of a massive particle in an electric field at coincident points
(we can also compute this at non-coincident points to get a finite answer). The total mass of the
particle is
\begin{equation}
m = q - {E\over q}
\end{equation}
For large $q$ this is above threshold for pair creation\footnote{See Appendix \ref{appendix:relativistic particle}}. The pair creation interpretation
is appropriate for the problem in \nref{RelEl}, but not for \nref{JTEn}.
In both problems we have a classical approximation to the dynamics that corresponds to a particle
describing a big circular trajectory in hyperbolic space at radius $\rho_c$ :
\begin{equation} \label{SolCir}
\tanh \rho_c = { m \over q }
\end{equation}
For the problem in \nref{RelEl}, fluctuations around this circle lead to an instability, with a single negative mode and an imaginary part in the partition function \nref{RelEl}. This single negative mode corresponds to small fluctuations of the overall size of the circular trajectory around \nref{SolCir}.
On the other hand in \nref{JTEn} we are integrating the same mode along a different contour, along the
imaginary axis, where we get a real and finite answer. Furthermore, the imaginary part in the
partition function \nref{RelEl} comes precisely from the trajectory describing pair creation, which is also the type of contribution captured in \nref{JTEn}.
Finally, in the relativistic particle problem, we expect that the pair creation amplitude should be exponentially suppressed for large $q$, while the partition function for the JT model is not. In fact,
for large $q$ the exponential suppression factor for pair creation goes as $e^{ - 2 \pi q } $, which is precisely cancelled by a similar factor in \nref{JTPa}, to obtain something finite in the large $q$ limit.
\section{Solving the Quantum Mechanical Problem}
\label{CanonicalQuantization}
As we explained above the solution of the JT theory is equivalent to considering a non-relativistic particle in $AdS_2$ or $H_2$.
We first consider the Euclidean problem, of a particle moving in $H_2$. An ordinary magnetic field in $H_2$ leads to an Euclidean action
of the form
\begin{equation} \label{magnetic field}
S= \int du {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } + i b \int d u { \dot x \over y } -{1\over 2} ( b^2 +{ 1 \over 4} ) \int d u ~,~~~~~~~b = i q
\end{equation}
If $b$ is real we will call it a magnetic field,
when $q$ is real we will call it an ``electric" field. The last term is a constant we introduced for convenience. Its only effect
will be to shift the ground state energy. It is interesting to compute the classical solutions and the corresponding action for
\nref{magnetic field}. These solutions are simplest in the $\rho$ and $\varphi $ coordinates, using the SL(2) symmetry we find that the trajectories are given by ($t=-iu$):
\begin{eqnarray}
{1\over 2}\sinh^2\rho ({d\varphi\over dt})^2+{q^2\over 2}-{1\over 8}=E,~~ \cosh\rho={q\beta\over 2\pi},~~ {d\varphi\over du}={2\pi\over \beta}.
\end{eqnarray}
In this classical limit we
get the following relations for the action and the temperature:
\begin{eqnarray}
& ~& { \beta \over 2 \pi } = {1\over \sqrt{2E+{1\over 4}}}
\cr
& ~ & -S = {2\pi^2\over \beta}+{\beta\over 8}-2\pi q \label{ClassEnt}
\end{eqnarray}
When $b$ is real, this system is fairly conventional and it was solved in \cite{Comtet:1986ki} . Its detailed spectrum depends on $b$. For very large $b$ we have a series of Landau levels and also a continuous spectrum. In fact, already the classical problem contains closed circular orbits, related to the discrete Landau levels, as well as orbits that go all the way to infinity.\footnote{See Appendix \ref{appendix:landau level}}
The number of discrete Landau levels decreases as we decrease the magnetic field and for $0 < b < {1\over 2}$ we only get a continuous spectrum. The system has a $SL(2)$ symmetry and the spectrum organizes into SL(2) representations, which are all in the
continuous series for $0 < b < 1/2$. For real $q$ we also find a continuous spectrum which we can view as the analytic continuation
of the one for this last range of $b$.
The canonical momenta of the action (\ref{magnetic field}) are:
\begin{equation}
p_x={\dot x\over y^2}+{i q\over y};~~~~~~~p_y={\dot y\over y^2}.
\end{equation}
And the Hamiltonian conjugate to $\tau_L$ is thus:
\begin{equation}
H={\dot x^2+\dot y^2\over 2y^2}+{q^2\over 2}={y^2\over 2} [(p_x-i{q\over y})^2+p_y^2] +{ q^2 \over 2 } - { 1 \over 8}
\end{equation}
Note that the Hamiltonian is not Hermitian. However, it is {\cal PT}-symmetric (here parity reflects $x$ and $p_x$) and for that
reason the spectrum is still real, see \cite{Bender:2005tb}.
The action is invariant under $SL(2,R)$ transformations generated by
\begin{eqnarray}
L_0=x p_x+y p_y;~~~~~L_{-1}=p_x;~~~~~~~~L_{1}=(y^2-x^2)p_x-2xyp_y-2iqy \label{SLtwoGen}
\end{eqnarray}
Notice the extra $q$ dependent term in $L_1$ that arises due to the presence of a magnetic field.
Up to a simple additive constant, the Hamiltonian is proportional to the Casimir operator
\begin{equation}
H={1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) + { q^2 \over 2} - { 1 \over 8 }
\end{equation}
As is common practice, let us label the states by quantum numbers $j={1\over 2}+is$ and $k$, so that $H|j,k\rangle=j(1-j)|j,k\rangle$ and $L_{-1}|j,k\rangle=k|j,k\rangle$.
We can find the eigenfunctions by
solving the $Schr\ddot{o}dinger$ equation with boundary condition that the wavefunction should vanish at the horizon $y\rightarrow \infty$ \cite{Comtet:1984mm,Comtet:1986ki,Pioline:2005pf}:
\begin{equation}
\omega_{s,k}={s^2\over 2} ,~~~~~~f_{s,k}(x,y)=
\begin{cases}
({s\sinh 2\pi s\over 4\pi^3k})^{1\over 2}|\Gamma(is-b+{1\over 2})|e^{-ik x}W_{b,is}(2 k y), ~~~k>0;\\
({s\sinh 2\pi s\over 4\pi^3|k|})^{1\over 2}|\Gamma(is+b+{1\over 2})|e^{-ik x}W_{-b,is}(2|k|y),~~~k<0.
\end{cases}
\end{equation}
where $\omega_{ks}$ is giving the energy of the states labelled by $s$ and $k$, and $ W $ is the Whittaker function. The additive constant in \nref{magnetic field} was
introduced to simplify this equation. We can think of $s$ as the quantum number of the continuous series representation of $SL(2)$ with
spin $j={1\over 2} + i s $.
After continuing $b \to i q$ we find that the gravitational system has a continuous
spectrum
\begin{equation}\label{energy relation}
E(s)={s^2 \over 2 }.
\end{equation}
\subsection{The Propagator }
It is useful to compute the propagator for the non-relativistic particle in a magnetic field,
$K(u,\boldsymbol{x_1},\boldsymbol{ x_2})=\langle \boldsymbol{x_1}| e^{-u H}|\boldsymbol{x_2}\rangle$. Here, $\boldsymbol{x}$ stands for ${x,y}$.
The propagator for a real magnetic particle was obtained in \cite{Comtet:1986ki}:
\begin{eqnarray} \label{particle heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})& =& e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} \int_0^{\infty}ds s e^{- u { s^2 \over 2}}{\sinh 2\pi s\over 2\pi (\cosh 2\pi s+\cos 2\pi b)}{1\over d^{1+2is}}
\times
\cr
& & ~~~~~ \times ~_2F_1({1\over 2}-b+is,{1\over 2}+b+is,1,1-{1\over d^2}).
\cr
d & = & \sqrt{(x_1-x_2)^2+(y_1+y_2)^2\over 4y_1 y_2}
\cr
e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} &=& e^{-2 i b \arctan {x_1-x_2\over y_1+y_2}}
\end{eqnarray}
In the case that we have a real magnetic field the prefactor is a phase and it is
gauge dependent. It is equal to the value of Wilson Line $e^{i\int a}$ stretched along the geodesic between $x_2$ and $x_1$.
Here we quoted the value in the gauge where the action is
\nref{magnetic field}. The second equation defines the parameter $d$, which is a function of the geodesic distance between the two points. Note that $d=1$ corresponds to coincident points.
We can get the answer we want by making the analytic continuation $b\to i q$ of this formula.
We can check that this is the right answer for our problem by noticing the following. First one can check that this expression is invariant under the SL(2) symmetry $L_a = L^1_a + L^2_a$ where $L_a$ are the generators \nref{SLtwoGen} acting on $\boldsymbol{x_1}$
and $L_a^2$ are similar generators as in \nref{SLtwoGen}, but with $q\to -q$. It is possible to commute
the phase $e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})}$, in \nref{particle heat kernel} past these generators which would remove the $q$ dependent terms. This implies
that the rest should be a function of the proper distance, which is the case with \nref{particle heat kernel}.
Then we can check the equation
\begin{eqnarray}
0 &=& (\partial_u + H_1 )G(u,\boldsymbol{x_1},\boldsymbol{x_2})
\end{eqnarray}
which is also indeed obeyed by this expression. The $s$ dependent prefactor is fixed by the requirement that the
propagator composes properly, or more precisely, by saying that for $u=0$ we should get a $\delta$ function.
\subsection{Partition Function}
The gravitational partition function is related with the particle partition function with inverse temperature $\beta$.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{Freeenergy.pdf}
\caption{Free Energy diagram with inverse temperature $\beta$. }
\end{figure}
The canonical partition function of the quantum mechanical system is
\begin{eqnarray}
Z_{\scaleto{Particle}{4pt}}&=&Tr e^{-\beta H}
=\int_0^{\infty} ds \int_{-\infty}^{\infty}dk \int_M {d x d y\over y^2} e^{-{\beta } { s^2 \over 2 } }f^*_{s,k}(x,y)f_{s,k}(x,y)\nonumber\\
&=&V_{AdS}\int_0^{\infty}ds e^{-\beta {s^2\over 2}}{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}.
\end{eqnarray}
The volume factor $V_{AdS}$ arises because after momentum integration there is no position dependence.
In a normal quantum mechanical system, the volume factor means that the particle can have independent configurations at different locations of our space, however for a gravitational system this should be thought as redundant and should be cancelled by the volume of $SL(2,R) $ gauge group $2\pi V_{AdS}$\footnote{There might be a multiplicative factor in the volume of gauge group, but we can always absorb that into $S_0$.}.
In gravitational system, there can also other contributions to the entropy from pure topological action.
These give a contribution to the ground state entropy $S_0$. Including the topological action in (\ref{regularized action}), we find a gravitational ``density of states" as
\begin{equation} \label{Spectral}
\rho(s)=\underbrace{e^{S_0}e^{2\pi q}}_{\text{extra terms}}\underbrace{1\over {2\pi}}_{\text{residue gauge}}\underbrace{{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}}_{\text{particle in magnetic field}}=e^{S_0}e^{2\pi q}{s\over 2\pi^2}\sum_{k=1}^{\infty}(-1)^{k-1}e^{-2\pi q k}\sinh(2\pi s k).
\end{equation}
We have not given an explicit description of these states in the Lorentzian theory. More details were discussed in \cite{Kitaev:2018wpr,Lin:2018xkj}.
This expression has some interesting features. Notice that the classical limit corresponds to large $q$ and large $s$, where we reproduce \nref{ClassEnt}. After approximating, the density of states are
$\log \rho(s) \sim S_0 + 2 \pi s $ for $ s/q < 1$ and $S_0+2\pi q$ for $s/q>1$.
We can also expand the partition function for very small and very large temperatures where we obtain
\begin{eqnarray}
Z_{JT} &\sim & e^{ S_0} e^{ 2 \pi q } { 1 \over 4\pi^2\beta} ~,~~~~~~~~~~~ \beta \ll {1\over q}
\cr
Z_{JT} & \sim & e^{S_0} { 1 \over \sqrt{2\pi}\beta^{3/2} }~,~~~~~~~~~~~~ \beta \gg 1
\end{eqnarray}
Notice that at leading order we get an almost constant entropy both at low and high temperatures, with the
high temperature one being higher. In both cases there are power law corrections in temperature.
Before we try to further elucidate the interpretation of this result, let us emphasize a couple of important defects of our discussion. First, when we replaced the partition function of the JT theory by the action of a non-relativistic
particle in an electric field, we were summing over paths in $H_2$. This includes paths that self intersect see figure \ref{fig:TwoInstanton}. Such paths do not have an obvious interpretation in the JT theory and it is not even clear that we should include them. For example, the sum over $k$ in \nref{Spectral} can be understood in terms of classical solutions which wind $k$ times around the circle. These make sense for the problem of the particle in the electric field but apparently not in the JT theory.
\begin{figure*}[h!]
\centering
\subfigure[Density of States]{%
\includegraphics[width=0.4\textwidth]{densityofstate.pdf}%
\label{fig:dos}%
}\hspace{4cm
\subfigure[Two Instantons]{%
\includegraphics[width=0.35\textwidth]{twoinstanton.pdf}%
\label{fig:TwoInstanton}%
\caption{Density of States and the Two Instantons configuration}
\end{figure*}
Maybe such paths could be given some interpretation in the gravity theory.
Alternatively, we might want to sum over paths that do not self intersect.
A second defect is that we would be eventually interested in adding some matter fields propagating in the bulk geometry. These matter fields have boundary conditions at the boundary of the region of hyperbolic space
cut out by the boundary trajectory. The partition function of the fields with such an arbitrary boundary trajectory could also modify the results we described above. Of course, this issue does not arise if we have
the pure JT theory. It is only important if we want to introduce bulk matter fields to define more complex observables.
Instead of attempting to address the above issues, we will take an easy route, which is to consider the system only in the large $q$ (or small $\epsilon$) limit. In this regime, we address the above issues, and we can still trust the description of the particle in the electric field.
This large $q$ or small $\epsilon$ limit is the same one that isolates the Schwarzian action from the JT theory
\cite{Maldacena:2016upp,Jensen:2016pah,Engelsoy:2016xyb}.
It turns out that the limit can be taken already at the level of the mechanical system, a simple rescaled
version of the above system. This provides an alternative method for quantizing the Schwarzian theory. It has the advantage of being a straightforward second order action of a particle moving in a region near the boundary of hyperbolic. Of course, the Schwarzian theory was already quantized using a variety of methods in
\cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Luca}. We will simply provide yet another perspective, recover the old results, and write a few new expressions.
\section{Quantum Gravity at Schwarzian Limit}
Before getting into the details notice that the large $q$ limit of \nref{Spectral} gives
\begin{equation}
\rho(s) = e^{S_0}{s\over 2\pi^2}\sinh(2\pi s), ~~~E = { s^2 \over 2} ~,~~~~~~~Z_{JT} = \int_0^\infty ds
\rho(s) e^{ - \beta { s^2 \over 2 } } =e^{S_0}{1 \over\sqrt{2\pi}\beta^{3\over 2}}e^{2\pi^2 \over \beta}.
\end{equation}
This reproduces what was found in \cite{Stanford:2017thb, Mertens:2017mtv, Mertens:2018fds} by
other methods. We see that we get a finite answer and also that the contributions from the $k>1$ terms in
\nref{Spectral} have disappeared. Because the $S_0$ part decouples with JT gravity, from now on, we will drop it and discuss $S_0$ only when it is necessary.
\subsection{The Propagator}
To get a limit directly at the level of the mechanical system it is useful to define a
rescaled
coordinate, $z$, via
\begin{equation}
y= z/q. \label{yAndq}
\end{equation}
After taking the large $q$ limit, the boundary particle propagator becomes \footnote{see the Appendix \ref{appendix:large q} for details}:
\begin{eqnarray}\label{heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&{ 1 \over q } e^{-2\pi q \theta(x_2-x_1) }\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2});~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~q\gg 1.\\
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{2\sqrt{z_1z_2}}{\pi^2 |x_1-x_2|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|});~~~~~~~~
\label{HeatLim}\\
&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{\sqrt{2}}{\pi^{3/2} u^{3/2}}\frac{\sqrt{z_1z_2}}{ |x_1-x_2|}\int_{-\infty}^{\infty}d\xi (\pi+i\xi)e^{-2\frac{(\xi-i\pi)^2}{u}-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}.\label{HeatLim2}
\end{eqnarray}
The original phase factor $e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_1})}$ factorizes into a product of singular ``phase" $e^{-2\pi q \theta(x_2-x_1)}$, with $\theta$ the step function,
and a regular ``phase" $e^{-2 {z_1+z_2\over x_1-x_2}}$. The singular ``phase" is the same order as the topological piece in (\ref{regularized action}).
In order to have a finite result they should cancel between each other. This can only be satisfied if the $x_is$ are in cyclic order.
As shown in figure (\ref{fig:phasefactor}), the product of singular ``phase" gives $-2\pi q$ for cyclic order $x_i$s and this would cancel with the topological action $2\pi q$.
While for other ordering of the $x_i$s, this would have $-2 \pi n q$ for $n=2,3,...$ and is highly suppressed in the limit $q$ goes to infinity.
This cyclic order is telling us where the interior of our space time is. The magnetic field produces a preferred orientation for the propagator.
After fixing the order, all our formulas only depend on $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ which has no $q$ dependence.
The residual $q$ factor in \ref{heat kernel} cancels out the additional $q$ from the measure of coordinate integral, ${ dx dy \over y^2 } \to q { dx dz \over z^2 } $.
In conclusion, after taking the limit we get a finite propagator equal to \nref{HeatLim}, which should be multiplied by a step function $ \theta( x_1 - x_2 ) $ that imposes the right order.
The final function $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ has the structure of $e^{-2{z_1+z_2\over x_1-x_2}}f(u,{z_1z_2\over (x_1-x_2)^2})$. This can be understood directly from the $SL(2)$ symmetry.
After taking the large $q$ limit, the $SL(2,R)$ charges become
\begin{equation} \label{SLTwoLq}
L_0=i (x\partial_x+z\partial_z);~~~~L_{-1}=i\partial_x;~~~~~~L_{1}=-ix^2\partial_x-2ix z\partial_z-2i z.
\end{equation}
We can check that they still satisfy the SL(2) algebra.
If we drop the last term in $L_1$, the $SL(2,R)$ charges become the usual differential operators on $EAdS_2$. And the propagator will have only dependence on the geodesic distance. When $L_1$ operator is deformed, the condition of $SL(2,R)$ invariance fixes the structure of the propagator as follows.
The $L_0$ and $L_{-1}$ charges are not deformed and they imply that the
only combinations that can appear are
\begin{equation}
v \equiv {z_1+z_2\over x_1-x_2} ~~~~~~~~~~~~~\text{and}~~~~~~~~~~~~~~ w\equiv {z_1 z_2\over(x_1-x_2)^2}.
\end{equation}
Writing the propagator as
$\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_1})=k(v,w) $
and
requiring it to be invariant under $L_1$ gives the following
equation for $\alpha$:
\begin{equation}\label{phaseequation}
\partial_v k +2 k =0 \longrightarrow k = e^{ - 2 v } h(w)
\end{equation}
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=e^{-2 {z_1+z_2\over x_1-x_2}}f(u,{z_1 z_2\over(x_1-x_2)^2}),
\end{equation}
The full function can also be determined directly as follows. Again we impose the propagator equation (or heat equation)
\begin{eqnarray}
0 &=& \left[ \partial_u + {1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) -{1\over 8} \right] \tilde K
\cr
0 & = & [ { s^2 \over 2} + {w^2\over 2}\partial_w^2+2w +{1\over 8}] K_s(w)
\end{eqnarray}
where $L_a$ are given in \nref{SLTwoLq} and are acting only on the first argument of $\tilde K$.
The solution of the last equation which is regular at short distances ($w \to \infty$) is $\sqrt{w}$ times
the Bessel K
function in \nref{heat kernel}.
We can also directly determine the measure of integration for $s$ by demanding that the propagator at $u=0$ is
a $\delta$ function or by demanding the propagator compose properly. This indeed is the case with the $s\sinh{ 2 \pi s}$ function in \nref{HeatLim}.
To explicitly show the above statement, it will be useful to use spectral decomposition of the propagator:
\begin{equation}\label{SpectralDecomposition}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=\int ds {2s\sinh(2\pi s)\over \pi^3} e^{-{s^2u\over 2}}\int dk \sqrt{z_1 z_2} e^{ik(x_1-x_2)}K_{2is}(2\sqrt{2i k z_1})K_{2is}( 2\sqrt{2ikz_2}).
\end{equation}
It can be easily checked that the special functions $f_{k,s}(x,z)=\sqrt{z}e^{ikx}K_{2is}( 2\sqrt{2ikz})$ are delta function normalizable eigenmodes of the large $q$ Hamiltonian:
\begin{equation}
\int {dx dz\over z^2}f_{k_1,s_1}f_{k_2,s_2}=\delta(k_1-k_2)\delta(s_1-s_2){\pi^3\over 2s\sinh (2\pi s)}
\end{equation}
Notice that the inner product fixes the integral measure completely in (\ref{SpectralDecomposition}), and the composition relation is manifestly true:
\begin{equation}
\int {dx dz\over z^2}\tilde{K}(u_1,\boldsymbol{x_1},\boldsymbol{x})\tilde{K}(u_2,\boldsymbol{x},\boldsymbol{x_2})=\tilde{K}(u_1+u_2,\boldsymbol{x_1},\boldsymbol{x_2})
\end{equation}
At short time the propagator has the classical behavior:
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})\sim \delta(x_1-x_2+u z_2)e^{-{(z_1-z_2)^2\over 2 u z_2}}
\end{equation}
This form of singularity is expected since we are taking the large $q$ limit first and thus the velocity in $x$ direction is fixed to be $z$. In the original picture of finite $q$ we are looking at the time scale which is large compare to AdS length but relatively small such that the quantum fluctuations are not gathered yet.
The integral structure in the propagator (\ref{HeatLim}) has an obvious meaning: integrating over $s$ represents summing over all energy states with Boltzmann distribution $e^{-E u}$, and the Bessel function stands for fixed energy propagator.
We want to stress that the argument in the Bessel function is unusual, and at short distance it approaches a funny limit:
\begin{equation}
K_{2is}(\frac{4 }{\ell})\simeq \sqrt{\pi \over 8\ell}e^{-\frac{4 }{\ell}},~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}\rightarrow 0.
\end{equation}
One should contrast this exponential suppression with the short distance divergence in QFT which is power law.
In our later discussion of exact correlation function with gravity backreaction, we will see that this effect kills UV divergence from matter fields.
To obtain the expression (\ref{HeatLim2}), we use the integral representation for the Bessel function and the final result has some interesting physical properties:
Firstly, we see that at large $u$ the time dependence and coordinate dependence factorized.
So, at large time we have a universal power law decay pointed out in \cite{Bagrets:2016cdf}.
Secondly, as we said before, the phase factor $e^{-2\frac{z_1+z_2}{x_1-x_2}}$ is equal to the (regularized) Wilson line $e^{-q\int_1^2 a}$ stretched along the geodesic connection between location $1$ and $2$ (Figure \ref{fig:propagator}). The field $a$ depends on our choice of gauge, our convention corresponds to fix the minimum value of $a$ at infinity and then the Wilson line is equal to $e^{-q A}$, where $A$ is the area of a hyperbolic triangle spanned by $1,2$ and $\infty$.
Thirdly, defining $2\pi+2i \xi$ as $\theta$, then $\theta$ has the meaning of the spanned angle at the horizon (Figure \ref{fig:propagator}). Then the gaussian weight $e^{-2{(\xi-i\pi)^2\over u}}=e^{\theta^2\over 2u}$ can be understood from the classical action along the boundary with fixed span angle $\theta$. The boundary drawn in the figure represents a curve with fixed (regularized) proper length $u$ in $H_2$.
Lastly, the factor $e^{-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}=e^{{4\cos\theta/2\over \ell}}$ is equal to $e^{-q(\alpha+\beta)}$, which is a corner term that arise from JT gravity in geometry with jump angles. Here $\alpha$ and $\beta$ are defined as the angle spanned by the geodesic with fixed length and the ray coming from horizon to the boundary.
In summary the propagator can be understood as an integral of JT gravity partition functions over geometries \ref{fig:propagator} with different $\theta$s.
\begin{figure*}[h!]
\centering
\subfigure[singular ``phase" factor for different ordering]{%
\includegraphics[width=0.45\textwidth]{phase_factor.pdf}%
\label{fig:phasefactor}%
}\hspace{1.5cm
\subfigure[A geometric representation of the propagator. Here we fix the span angle $\theta$, the propagator is a summation over such geometries.]{%
\includegraphics[width=0.45\textwidth]{propagator.pdf}%
\label{fig:propagator}%
\caption{The singular ``phase" for different ordering and the geometric representation of the propagator}
\label{fig:Propagator}
\end{figure*}
Finally, let us comment on the issues we raised in section 4.2. In the large $q$ limit we are considering the propagator at relatively large distances and in a regime where locally in $AdS$ the integration over paths that fluctuate
wildly is suppressed. Alternatively we can say that in the integration over paths we put a UV cutoff which
is large compared to $1/q$ but small compared to the AdS radius. This is the non-relativistic regime for the
boundary particle. The quantum effects are still important at much longer distances due to the large size
of $AdS$.
In addition, if we have quantum fields in AdS, then their partition functions for these fluctuating contours that
have fluctuations over distances larger than the $AdS$ radius are expected to depend on this shape in a local way. Due to the symmetries of $AdS_2$, this is simply expected to renormalize the action we already have without introducing extra terms. This can be checked explicitly for conformal field theories by using
the conformal anomaly to compute the effective action of the CFT$_2$ on a portion of $H_2$ (Appendix \ref{appendix: CFT effective action}).
\subsection{Wheeler-DeWitt Wavefunction }
\label{WdW}
In the pure JT theory we can think about quantizing the bulk theory and obtaining the
Wheeler-DeWitt wavefunction. This was discussed in the classical limit by Harlow and Jafferis \cite{Harlow:2018tqv}.
The Wheeler-DeWitt wavefunction can be created by Euclidean evolution of the boundary and hence is closely related to the propagator we have discussed above. The wavefunction in Lorentzian signature could then be obtained by analytic continuation of the boundary time.
The Euclidean evolution can be specified by either of the two parameters: the proper length $u$ or energy $E$. Choosing a different parameter corresponds to imposing a different boundary condition in JT theory. In general there are four possible choices of boundary conditions in 2d dilaton gravity, there are two sets of conjugate variables: $\lbrace \phi_b,K\rbrace$, and $\lbrace u,E\rbrace$ \footnote{Energy $E$ is proportional to the normal derivative of the dilaton field at the boundary.}. In preparing the wavefunction we fix the boundary value of dilaton and hence there are only two choices of the parameter ($u$ or $E$). We denote the corresponding wavefunction as $|u\rangle_G$ and $|E\rangle_G$ respectively.
In terms of holographic considerations, $|u\rangle_G$ represents a thermofield double state:
\begin{equation}
|u\rangle_G\sim\sum\limits_{n}e^{-E_n u}|E_n\rangle_L |\bar E_n\rangle_R
\end{equation}
and $|E\rangle_G $ is like an average of energy eigenstates in a window of energy $E$:
\begin{equation}\label{EnergyStateDefinition}
|E\rangle_G\sim {1\over \delta E}\sum\limits_{ |E-E_n|<\delta E}|E_n\rangle_L |\bar E_n\rangle_R.
\end{equation}
The width of the energy window is some coarse graining factor such that the summation contains $e^{S_0}$ states and does not show up clearly in gravity.\footnote{If one understand getting $|E\rangle$ state from integrating over thermofield double state in time direction, then a natural estimate on $\delta E$ is ${1\over T}$, where $T$ is the total time one integrate over. The validity of JT description of boundary theory is $T<e^{S_0}$, and we get $\delta E>e^{-S_0}$. For $T>e^{S_0}$, there are other possible instanton contributions. The proper gravitational theory at this regime is studied in paper \cite{Saad:2018bqo} }
With the definition of the states, one can evaluate them in terms of different basis. There are three natural bases turn out to be useful, we call them $S$, $\eta$ and $\ell$ bases. Basis $S$ corresponding to fix the horizon value of dilaton field $\phi_h$, or equivalently by Bekenstein-Hawking formula, the entropy of the system. The canonical conjugate variable of $S$ will be called $\eta$ and that characterizes the boost angle at the horizon. $\ell$ stands for fixing geodesic distance between two boundary points. To see that the horizon value of the dilaton field is a gauge invariant quantity, one can do canonical analysis of JT gravity. With ADM decomposition of the spacetime metric, one can get the canonical momenta and Hamiltonian constraints of the system \cite{LouisMartinez:1993eh}:
\begin{eqnarray}
ds^2&=&-N^2dt^2+\sigma^2(dx+N^xdt)^2;~~~~~~~~~~~~~~~~~~~~\\
\mathcal{H}&=&-\Pi_{\phi}\Pi_{\sigma}+\sigma^{-1}\phi''-\sigma^{-2}\sigma'\phi'-\sigma\phi;~~~~\mathcal{H}_x=\Pi_{\phi}\phi'-\sigma\Pi_{\sigma}';\label{Hamiltonian Constraint}\\
\Pi_{\phi}&=&N^{-1}(-\dot{\sigma}+(N^x\sigma)')=K\sigma;~~~~~~~~~~~~~\Pi_{\sigma}=N^{-1}(-\dot{\phi}+N^x\phi')=\partial_n\phi.\label{Canonical Momentum}
\end{eqnarray}
That is the dilaton field is canonically conjugate to the extrinsic curvature and boundary metric is canonical conjugate to the normal derivative of the dilaton field (both are pointing inwards).
By a linear combination of the Hamiltonian constraints (\ref{Hamiltonian Constraint}), one can construct the following gauge invariant quantity $C$:
\begin{equation}
-{1\over\sigma}(\phi'\mathcal{H}+\Pi_{\sigma}\mathcal{H}_x)={1\over 2}(\Pi_{\sigma}^2+\phi^2-{\phi'^2\over \sigma^2})'\equiv C[\Pi_{\sigma},\phi,\sigma]'\sim 0
\end{equation}
The Dirac quantization scheme then tells us that the quantity $C$ has a constant mode which is gauge invariant (commute with Hamiltonian constraint).
Choosing the gauge that normal derivative of the dilaton is zero, we can solve the Hamiltonian constraint:
\begin{equation}
\phi^2-(\partial_{X}\phi)^2=2C\equiv S^2~~~~~\rightarrow~~~~~~\phi(X)=S\cosh X,
\end{equation}
where $dX= \sigma dx$ is the proper distance along the spatial slice. Because the normal derivative of dilaton field is zero, the minimum value of dilaton at this spatial slice is actually a local extremum in both directions. Therefore, the minimal value of dilaton field, namely $S$, is a global variable. The classical geometry in this gauge is a ``Pac-Man" shape (right figure in figure \ref{fig:Pac-Man}).
Focusing on the intersection region of the spatial slice and the boundary, we have the spatial slice is orthogonal to the boundary.
This is because we are gauge fixing $\partial_n\phi=0$ on the spatial slice, and $\phi=\phi_b$ on the boundary.
The ADM mass of the system, after regularization, is then $M=\phi_b(\phi_b-\partial_{X}\phi)$ \cite{Maldacena:2016upp}. Substituting the behavior of $\phi(X)$ we get:
\begin{equation}
M={S^2\over 2}.
\end{equation}
This is the same relation in \ref{energy relation} and therefore we can interpret the $s$ variable as entropy of our system $S$.
For the purpose of fixing geodesic distance, it is convenient to think of doing the path integral up to a slice $L$ with zero extrinsic curvature. This picks out a particular slice (left figure in Figure \ref{fig:Pac-Man}) among the solutions obeying the Hamiltonian constraint.
The WdW wavefunction can be evaluated as an Euclidean path integral with fixed (rescaled) geodesic distance $d$ between the two boundary points:
\begin{equation}\label{WdWPath}
\begin{split}
\Psi(u;d)&=\int {\cal D}g{\cal D}\phi e^{{1\over 2}\int \phi(R+2)+\int_{L}\phi K+\phi_b\int_{Bdy}(K-1)+\phi_b(\alpha_1+\alpha_2-\pi)}=\int {\cal D }f e^{\phi_b\int_{Bdy} (K-1)+\phi_b(\alpha_1+\alpha_2-\pi)}\\
&=\int {\cal D}\boldsymbol x e^{-m \int_{Bdy}\sqrt{g}+q\int_{Bdy} a+q\int_{L} a+\pi q}
\end{split}
\end{equation}
Here we are fixing the total length of $L$ to be $d$ and the proper length of the boundary to be $u$. $\alpha_1$ and $\alpha_2$ in the last expression denotes the jump angle at the corner coming from the singular contribution of the extrinsic curvature and should be integrated over. Without the $e^{q\int_L a}$ factor in (\ref{WdWPath}), the path integral corresponds to the propagator (\ref{HeatLim}). Remember that the phase factor is equal to $e^{-q\int_L a}$, so the wavefunction in $\ell$ basis is actually the propagator (\ref{HeatLim}), with the phase factor stripped out
\begin{equation}\label{wavefunction}
\Psi(u; \ell)\equiv \langle \ell|u\rangle_G = \frac{2}{\pi^2 \ell}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 }{\ell}) ,~~~~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}.
\end{equation}
$\ell$ is a function of the regularized geodesic distance $d$ between $\boldsymbol{x_1}$ and $\boldsymbol{x_2}$ : $\ell=e^{d\over 2}$.
The semiclassical of $\Psi(u;\ell)$ can be obtained using formula (\ref{HeatLim2}), in the exponent we get saddle point result:
\begin{equation}
\Psi(u;\ell)\sim \exp[-{2(\xi_*-i\pi)^2\over u}+{4\over u}{\xi_*-i\pi\over \tanh \xi_*}];~~~~~~~~{\xi_*-i\pi\over u}=-{\sinh \xi_*\over \ell}.
\end{equation}
The same saddle point equation and classical action was obtained in \cite{Harlow:2018tqv} by a direct evaluation in JT gravity.
The wavefunction with fixed energy boundary condition can obtained by multiplying $\Psi(u;\ell)$ by
$e^{ + E u} $ and integrating over $u$ along the imaginary axis. This sets $E = {s^2 \over 2} $
in the above integral over $s$. So this wavefunction has a very simple expression:
\begin{equation}\label{fixed energy wavefunction}
\Psi(E;\ell)\equiv \langle \ell|E\rangle_G = \rho(E)\frac{4}{ \ell} K_{i \sqrt{ 8 E} }(\frac{4 }{\ell}).~~~~~~~~
\end{equation}
The classical geometry for $\Psi(E;\ell)$ is the same as the left figure in Figure \ref{fig:Pac-Man}, with fixing energy on the boundary.
We want to stress that it is important to have the $\rho(E)$ factor in (\ref{fixed energy wavefunction}) for a classical geometry description since we are averaging over the states.
We can roughly think of ${4\over \ell}K_{i\sqrt{8E}}({4\over \ell})$ as a gravitational ``microstate" $|\mathcal{E}\rangle$ with fixed energy $E$. Such a ``microstate" will not have a classical geometry representation and therefore is just a formal definition.
The inner product between wavefunctions is defined as $\langle\Psi_1|\Psi_2\rangle=\int_0^{\infty}d\ell \ell \Psi_1^*(\ell)\Psi_2(\ell)$.
Going to the entropy basis $S$, it is easy to start with $\Psi(E)$.
Because of the identity $E={S^2\over 2}$, expanding $\Psi(E)$ in the $S$ basis is diagonal:
\begin{equation}
\Psi(E;S)\equiv \langle S|E\rangle_G=\sqrt{\rho(S)}\delta(E-{S^2\over 2})
\end{equation}
We put this square root of $\rho(S)$ factor in the definition of $S$ basis such that inner product between different $S$ state is a delta function $\langle S|S'\rangle=\delta(S-S')$. This factor is also required such that the classical limit matches with gravity calculation.
Integrating over energy with Boltzman distribution, we can get the expression of thermofield double state in the $S$ basis:
\begin{equation}
\Psi(u;S)\equiv \langle S|u\rangle_G=\int dE e^{-u E}\langle S|E\rangle_G=\sqrt{\rho(S)}e^{-{uS^2\over 2}}
\end{equation}
In the semiclassical limit, the wavefunction becomes gaussian and coincides with the on shell evaluation of the ``Pac-Man" geometry (Figure \ref{fig:Pac-Man}):
\begin{equation}\label{PsiS}
\Psi(u,S)\sim \sqrt{S}e^{\pi S -{u S^2\over 2}}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.5]{classicalgeometry.pdf}
\caption{Classical Geometry in $\ell$ and $S$ basis}
\label{fig:Pac-Man}
\end{figure}
The on shell calculation is straightforward: JT action in this geometry contains two parts: the Schwarzian action $\int(K-1)$ on the boundary and a corner contribution at the center: $S(\pi-\theta)$, where $\theta$ is the span angle at the horizon (Figure \ref{fig:Pac-Man}). The Schwarzian action simply gives $E u={S^2 u\over 2}$ by direct evaluation. We can determine $\theta$ from $u$ since they are related with redshift: $\theta=u S$. Therefore the corner term gives: $\pi S-S^2 u$. Adding them up then gives us the classical action.
We can also expand $S$ in terms of the $\ell$ basis, and relate $\Psi(\ell)$ with $\Psi(S)$ by a change of basis:
\begin{equation}
\langle\ell|S\rangle=\sqrt{\rho(S)}{4\over \ell}K_{2iS}({4\over \ell});~~~~~\Psi(E;\ell)=\int_0^{\infty} dS \langle \ell|S\rangle \langle S|E\rangle_G;
\end{equation}
Before discussing our last basis, we want to stress the simplicity of the wavefunction in $S$ basis (\ref{PsiS}) and the Gaussian factor resembles an ordinary particle wavefunction in momentum basis. We introduce our last basis $\eta$ as canonical conjugate variable of $S$, with an analog of going to position space of the particle picture in mind:
\begin{eqnarray}
|\eta\rangle &=&\int_0^{\infty} dS \cos(\eta S)|S\rangle;~~~~~~~~~~~~~\langle \ell|\eta\rangle=\int_0^{\infty} dS\cos(\eta S)\sqrt{\rho(S)}K_{2iS}({4\over \ell}){4\over \ell};~~~~~\\
\Psi(E,\eta) &=&\scaleto{\sqrt{\sinh(2\pi \sqrt{2E})\over 2\pi^2\sqrt{2E}}\cos(\eta \sqrt{2E})}{30pt};~~~~~\Psi(u,\eta)=\int_0^{\infty} dS \sqrt{\rho(S)}\cos(\eta S)e^{-{uS^2\over 2}}.~~~~
\end{eqnarray}
To understand the meaning of $\eta$ better, we can look at the classical behavior of $\Psi(u;\eta)$:
\begin{equation}
\Psi(u;\eta)\sim {1\over u}\exp[{\pi^2\over 2 u}-{\eta^2\over 2 u}]\left(e^{i\eta\pi\over u}\sqrt{\pi+i\eta}+e^{-{i\eta\pi\over u}}\sqrt{\pi-i\eta}\right)
\end{equation}
When $u$ is real, the wavefunction is concentrated at $\eta=0$ and has classical action of a half disk in the exponent.
When $u={\beta\over 2}+it$ which corresponds to the case of analytically continuing into Lorentzian signature, the density of the wavefunction $|\Psi(u,\eta)|^2$ is dominated by:
\begin{equation}
|\Psi({\beta\over 2}+it,\eta)|^2\sim {\sqrt{\pi^2+\eta^2}\over \beta^2+4t^2}\exp[{2\pi^2\over \beta}]\left(\exp[-{2\beta(\eta-{2\pi t\over \beta})^2\over \beta^2+4t^2}]+\exp[-{2\beta(\eta+{2\pi t\over \beta})^2\over \beta^2+4t^2}]\right)
\end{equation}
showing the fact that $\eta$ is peaked at the Rindler time ${2\pi t\over \beta}$. We can therefore think of fixing $\eta$ as fixing the IR time or the boost angle at the horizon.
The classical intuition for the boost angle is most clear in Euclidean geometry, where for fixed boundary proper time there can be different cusps at the horizon (Figure \ref{fig:Boostangle}).
\begin{figure}
\centering
\includegraphics[scale=0.4]{boostangle.pdf}
\caption{Euclidean geometries with different cusps.}
\label{fig:Boostangle}
\end{figure}
One application of those wavefunctions is that we can take an inner product and get the partition function.
However, there are also other ways to get the partition function.
For example, we can concatenate three propagators and integrate over their locations. This also gives the partition function by the composition rule of propagator.
By the relation between propagator and wavefunction, we can also view this as taking an inner product of three wavefunctions with an interior state as in figure \ref{fig:threewavefunction}, where the interior state can be understood as an entangled state for three universes.
To be more precise, we can view the wavefunction as the result of integrating the bulk up to the geodesics with zero extrinsic curvature. Then the interior state is given
by the area of the hyperbolic triangle in figure \ref{fig:threewavefunction}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{threewavefunction.pdf}
\caption{Partition function from inner product of three wavefunctions.}
\label{fig:threewavefunction}
\end{figure}
The path integral for the hyperbolic triangle (denoted as $I(\ell_{12},\ell_{23},\ell_{31})$, where $\ell_{ij}={|x_i-x_j|\over \sqrt{z_iz_j}}$), is a product of three phase factors, which satisfies a nontrivial equality (with ordering $x_1>x_2>x_3$):
\begin{equation}\label{GHZ}
\scaleto{ I(\ell_{12},\ell_{23},\ell_{31})=e^{-2({z_1+z_2\over x_1-x_2}+{z_2+z_3\over x_2-x_3}+{z_3+z_1\over x_3-x_1})}={16\over\pi^2}\int_0^{\infty}d\tau \tau \sinh (2\pi \tau) K_{2i\tau}({4\over \ell_{12}})K_{2i\tau}({4\over \ell_{23}})K_{2i\tau}({4\over \ell_{31}}).}{26pt}
\end{equation}
Recalling that the Bessel function represents the fixed energy ``microstate" $|\mathcal{E}\rangle$ (\ref{fixed energy wavefunction}) and ${\tau\over 2\pi^2} \sinh(2\pi \tau)$ is the density of state, this formula tells us that the interior state is a GHZ state for three universe:
\begin{equation}
I_{123}\sim\sum_n|\mathcal{E}_n\rangle_1|\mathcal{E}_n\rangle_2|\mathcal{E}_n\rangle_3.
\end{equation}
$I$ can also been viewed as a scattering amplitude from two universes into one universe. It constrains the SL(2,R) representation of the three wavefunctions to be the same.\footnote{Some thing similar happens for 2d Yang-Mills theory \cite{Witten:1992xu,Cordes:1994fc,Luca}.}
We can write down the partition function as:
\begin{equation}
Z_{JT}=\int_0^{\infty} \prod\limits_{\lbrace i j \rbrace \in \lbrace 12,23,31\rbrace}d\ell_{ij} \Psi(u_{12},\ell_{12})\Psi(u_{23},\ell_{23})\Psi(u_{31},\ell_{31})I_{\ell_{12},\ell_{23},\ell_{31}}.
\end{equation}
This same result also holds if we repeat the process $n$ times. It is interesting that we can view the full disk amplitude in these various ways.
One can also extend our analysis to include matter field. One type of such wavefunction can be created by inserting operator during Euclidean evolution, and is analysed in appendix B. Note that because of the SL(2,R) symmetry is a gauge symmetry, our final state has to be a gauge singlet including matter field.
\section{Correlation Functions in Quantum Gravity}
\subsection{Gravitational Feynman Diagram}
The propagator enables us to ``dress'' quantum field theory correlators to produce quantum gravity ones.
Namely, we imagine that we have some quantum field theory in $H_2$ and we compute correlation functions of
operators as we take the points close to the boundary where they take the form
\begin{equation}
\langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} = q^{ - \sum \Delta_i }z_1^{\Delta_1}..z_n^{\Delta_n}\langle O_1(x_1)...O_n(x_n)\rangle_{CFT}
\end{equation}
The factor of $q$ arises from \nref{yAndq}, and the last factor is simply defined as the function
that results after extracting the $z$ dependence. For example, for a two point function we get
\begin{equation}
\langle O_1(\boldsymbol{x_1})O_2(\boldsymbol{x_2})\rangle_{\rm QFT}=
q^{ - 2\Delta } z_1^{\Delta} z_2^{\Delta}{1\over |x_1-x_2|^{2\Delta}}.
\end{equation}
We can now use the propagator \nref{HeatLim} to couple the motion of the boundary and thus obtain the full
quantum gravity expression for the correlator. The factors of $q$ are absorbed as part of the renormalization
procedure for defining the full quantum gravity correlators.
In this way we obtain
\begin{figure*}[h!]
\centering
\subfigure[Witten Diagram]{%
\includegraphics[width=0.3\textwidth]{wittendiagram.pdf}%
\label{fig:Witten Diagram}%
}\hspace{4cm
\subfigure[Gravitational Feynman Diagram]{%
\includegraphics[width=0.3\textwidth]{dynamicalwittendiagram.pdf}%
\label{fig:Dynamical Witten Diagram}%
\caption{Summation of ${1\over N}$ effects fluctuates the boundary of Witten Diagram}
\label{fig:witten diagram}
\end{figure*}
\begin{eqnarray}\label{Higher Point Function}
\small{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{\rm QG}}}&=&e^{2\pi q}\int {\prod_{i=1}^n{dx_i dy_i\over y_i^2}\over \text{V(SL(2,R))}} G(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})G(u_{23},\boldsymbol{x_2},\boldsymbol{x_3})...G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})\times
\cr
& ~& \times \langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} q^{ \sum_ i \Delta_i }
\end{eqnarray}
where the left hand side is the full quantum gravity correlator by definition. The last factor is the usual
renormalization necessary to get something finite.
The factor of $e^{ 2 \pi q } $ cancels with the $q$ dependent ``phase'' factors in \nref{heat kernel} to give one if we order
the points cyclically ($x_1>x_2...>x_n$). This requires that we define more carefully the last propagator
$G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})$ as:
\begin{equation}
e^{-2\pi q}\tilde K(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})=e^{-2\pi q}e^{-2 \frac{z_n+z_1}{x_n-x_1}}\frac{2\sqrt{z_nz_1}}{\pi^2 |x_n-x_1|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2} u_{n1}}K_{2is}(\frac{4\sqrt{z_nz_1}}{|x_n-x_1|}).
\end{equation}
The factor ${1\over \text{V(SL(2,R))}}$ in \nref{Higher Point Function} means that we should fix the $SL(2,R)$ gauge symmetry (Appendix \ref{gauge fix}).
In the end we can write down an expression where we have already taken the $q \to \infty $ limit
\begin{equation}\label{final formula}
\boxed{\scaleto{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{QG}}=\int_{ x_1>x_2..>x_n}{\prod_{i=1}^n{dx_i dz_i}\over \text{V(SL(2,R))}} \tilde{K}(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})...\tilde{K}(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})z_1^{\Delta_1-2}..z_n^{\Delta_n-2}\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}.}{26pt}}
\end{equation}
This is one of the main results of our paper and it gives a detailed expression for correlation function in 2 dimensional quantum gravity in terms of the correlation functions of the QFT in hyperbolic space, or $AdS_2$.
Notice that in usual $AdS/CFT$ the correlators $\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}$ are an approximation to the full answer. This is sometimes computed by Witten diagrams.
We get a better approximation by integrating over the metric fluctuations.
In this case, the non-trivial gravitational mode is captured by the boundary propagator. The formula \nref{final formula} includes all the effects of quantum gravity in the JT theory (in the Schwarzian limit).
The final diagrams consist of the Witten diagrams for the field theory in $AdS$ plus the propagators for the boundary particle and we can call them ``Gravitational Feynman Diagrams", see figure \ref{fig:witten diagram}.
\subsection{Two Point Function}
Using formula \nref{final formula}, we can study gravitational effects on bulk fields such as its two point function:
\begin{eqnarray}
\langle O_1(u)O_2(0)\rangle_{QG}=
\includegraphics[width=50mm,trim=0 9cm 0 0]{2ptfunction.pdf}\\
~~~~~~~~~~~\nonumber\\
~~~~~~~~~\nonumber
\end{eqnarray}
The explicit expression for $\langle O_1(u)O_2(0)\rangle_{QG}$ with dimension $\Delta$ at temperature ${1\over \beta}$ is
\footnote{We will just keep the $\Delta$ dependent constant
since at last we will normalized with respect of partition function which corresponds to set $\Delta=0$.}:
\begin{equation}
\scaleto{{1\over \text{V(SL(2,R))}}\int_{x_1>x_2} {dx_1 dx_2 dz_1 dz_2 \over z_1^2 z_2^2}\int_0^{\infty}ds_1 ds_2 \rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}K_{2is_1}({4\sqrt{z_1z_2}\over |x_1-x_2|})K_{2is_2}({4\sqrt{z_1z_2}\over |x_1-x_2|})({\sqrt{z_1z_2}\over |x_1-x_2|})^{2\Delta+2}}{25pt}.
\end{equation}
To fix the $SL(2,R)$ gauge, we can choose $z_1=z_2=1$ and $x_2=0$.
Then the integral over $H_2$ space is reduced to a single integral over $x_1$,
with a Jacobian factor $2x_1$ (Appendix \ref{gauge fix}):
\begin{equation}\label{space integral}
\int_0^{\infty}ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\int_0^{\infty} dx_1 ({1\over x_1})^{2\Delta+1}K_{2is_1}({4\over x_1})K_{2is_2}({4\over x_1}).
\end{equation}
the last integral can be interpreted as a matrix element of two point operator $O_1O_2$ between states $|E_1,\psi\rangle$ and $|E_2,\psi\rangle$,
where $|\psi\rangle$ is the wavefunction of quantum field theory
and $|E\rangle_G$ represents the fixed energy gravitational state.
Integrating over $x$ can be thought as integrating over a particular gravitational basis, and we can see that the gravity wavefunction
suppress the UV contributions from quantum field theory ($K_{2is}({4\over x})\sim \sqrt{\pi x\over 8}e^{-4/x}$ for $x\sim 0$).
The final expression for the two point function is:
\begin{eqnarray}\label{2point function}
\langle O_1(u) O_2(0)\rangle_{QG}={1\over \mathcal{N}}\int ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\frac{|\Gamma(\Delta-i(s_1+s_2))\Gamma(\Delta+i(s_1-s_2))|^2}{2^{2\Delta+1} \Gamma(2\Delta)};~~\\
={1\over \mathcal{N}}{\Gamma(2\Delta)\over u^{3/2} (\beta-u)^{3/2} 2^{4\Delta+4}\pi^3} \int_{c-i\infty}^{c+i\infty} d\theta_1d\theta_2 \theta_1\theta_2 e^{{\theta_1^2\over 2u}+{\theta_2^2\over 2 (\beta-u)}}{1\over (\cos{\theta_1\over 2}+\cos{\theta_2\over 2})^{2\Delta}}.~~~~~
\end{eqnarray}
In the second expression we write the integral in terms of variable $\theta$ using the second integral representation of the propagator (\ref{HeatLim2}).
The normalization constant can be determined by taking the $\Delta=0$ limit: $\mathcal{N}=Z_{JT}$.
If we contemplate the result (\ref{2point function}) a little bit, then we find that the two integrals of $s_1$ and $s_2$ just represent the spectral decomposition of the two point function. Indeed, under spectral decomposition we have $\langle O(u)O(0)\rangle=\sum\limits_{n,m} e^{-E_n u-E_m (\beta-u)}|\langle E_n|O|E_m\rangle|^2$. Compare with (\ref{2point function}), we can read out the square of matrix element of operator $O$:
\begin{equation}
\scaleto{_G\langle E_1|O_L O_R|E_2\rangle_G=\delta E^{-2}\sum\limits_{\substack{|E_n-E_1|<\delta E\\|E_m-E_2|<\delta E}}|\langle E_n|O|E_m\rangle|^2=\rho(E_1)\rho(E_2){|\Gamma(\Delta-i(\sqrt{2E_1}+\sqrt{2E_2}))\Gamma(\Delta+i(\sqrt{2E_1}-\sqrt{2E_2}))|^2\over 2^{2\Delta+1}\Gamma(2\Delta)}.}{38pt}
\end{equation}
Remember the notation is that $|E\rangle_G$ stands for a gravitational state with energy $E$ and $|E_n\rangle$ stands for one side microstate (\ref{EnergyStateDefinition}).
We have put the measure $\rho(E)={1\over 2\pi^2}\sinh(2\pi\sqrt{2 E})$ in the definition of matrix element for the reason that in gravity it is more natural to consider an average of energy states as a bulk state.
To understand this formula a little bit better, we can consider the classical limit, namely large $E$. In this limit the matrix element squared can be approximated as a nonanalytic function:
\begin{equation}
_G\langle E_1|O_L O_R|E_2\rangle_G\propto |E_1-E_2|^{2\Delta-1}e^{2\pi \text{min}(\sqrt{2E_1},\sqrt{2E_2})}.
\end{equation}
If we fix $E_1$ and varying $E_2$ from $0$ to infinity, the matrix element changes from $|E_1-E_2|^{2\Delta-1}\rho(E_2)$ to $|E_1-E_2|^{2\Delta-1}\rho(E_1)$ after $E_2$ cross $E_1$. We can understand this behavior qualitatively as a statistical effect: the mapping from energy subspace $E_1$ to $E_2$ by operator $O$ is surjective when the Hilbert space dimension of $E_2$ is less that $E_1$ and is injective otherwise.
Another understanding is the following: the two point function is finite in a fixed energy state $|E_n\rangle$, which means the following summation of intermediate states $|E_m\rangle$ is order one: $\sum\limits_{m}|\langle E_n|O|E_m\rangle|^2$. Looking at the case $E_m>E_n$, because of the density of states grows rapidly, the matrix element squared has to be proportional to ${1\over \rho(E_m)}$ to get a finite result. Multiplied by $\rho(E_n)\rho(E_m)$, we have $\rho(E_n)\rho(E_m)|\langle E_n|O|E_m\rangle|^2\sim \rho(\text{min}(E_n,E_m))$.
\subsection{ETH and the KMS condition}
The Eigenstate Thermalization Hypothesis (ETH) is a general expectation for chaotic system. It expresses that the operator expectation value in an energy eigenstate can be approximated by thermal expectation value with effective temperature determined from the energy. Such hypothesis can be tested with the knowledge of operator matrix elements. The two point function in microcanonical essemble is:
\begin{eqnarray}
{1\over \delta E}\sum\limits_{|E_n-E|<\delta E}\langle E_n|O(u) O(0)|E_n\rangle &=&
\rho(E)e^{uE}\langle E|O e^{-u H}O|E\rangle\nonumber\\
\langle E|O e^{-u H}O|E\rangle &=&\int_0^{\infty} ds \rho(s)e^{-{s^2\over 2}u}\scaleto{{|\Gamma(\Delta+i(s+\sqrt{2E}))|^2|\Gamma(\Delta+i(s-\sqrt{2E}))|^2\over 2^{2\Delta+1} \Gamma(2\Delta)}}{25pt}.~~~~~~~~~\label{TwoPFMicro}
\end{eqnarray}
Notice that $|E\rangle$ is not $|E\rangle_G$, the former represents a one side microstate, while the later is a gravitational state. Accordingly $ \langle E|O(u)O(0)|E\rangle$ stands for a two point function in a microstate.
To study ETH, we will consider the case of a heavy black hole $E={S^2\over 2}={2\pi^2 \over\beta^2}\gg 1$. From the discussion in last section, we know that the matrix element tries to concentrate $s$ around $\sqrt{2E}$ and thus we can approximate $\rho(E)\rho(s)|\Gamma(\Delta+i(s+S))|^2$ as proportional to $sS^{2\Delta-1}e^{\pi(s+S)}$. Using integral representation for $|\Gamma(\Delta+i(s-S))|^2 $ we derive the two point function in microcanonical essemble with energy $E$ is proportional to:
\begin{equation}
{\rho(E)S^{2\Delta-1}\over u^{3/2}}\int d\xi (\pi+i\xi)e^{-{2\over u}(\xi-i{\pi\over 2})^2-(\pi+2i\xi) S+u{S^2\over 2}}{1\over(\cosh \xi)^{2\Delta}}.
\end{equation}
The $\xi$ variable can be understood as the measure of time separation in units of effective temperature between two operators and its fluctuation represents the fluctuation of the effective temperature.
And the final integral can be understood as a statistical average of correlation functions with different temperatures.
If we put back the Newton Constant $G_N={1\over N}$, we have $S\sim N$ and $u\sim N^{-1}$ (\ref{BCond}).
As can be seen from the probability distribution, the fluctuation is of order ${1\over \sqrt{N}}$, and hence for large $N$ system we can use saddle point approximation:
\begin{equation}
\xi=i({\pi\over 2}-{S\over 2}u) -{u\Delta \tanh \xi\over 2}=i({\pi\over 2}-{\pi u\over \beta})-{u\Delta\tanh \xi\over 2}.
\end{equation}
The first piece gives the typical temperature of the external state,
while the last piece comes from the backreaction of operator on the geometry. If we first take the limit of large N, one simply get that the two point function in microcanonical essemble is the same as canonical essemble.
However, the euclidean correlator in canonical essemble is divergent as euclidean time approach to inverse temperature $\beta$ because of KMS condition.
Such singular behavior plays no role in the microcanonical essemble so is called a ``forbidden singularity" in ETH \cite{Fitzpatrick:2016ive, Faulkner:2017hll}. In our situation we can see directly how the forbidden singularity disappears in the microcanonical essemble.
When $\xi$ approach $-i{\pi\over 2}$ at the forbidden singularity the backreaction on the geometry becomes large and hence the effective temperature becomes lower:
\begin{equation}
{2\pi\over\beta*}\rightarrow {2\pi\over\beta}-{\Delta \over \pi-{\pi u\over \beta}}.
\end{equation}
At the time ${\beta-u\over\beta}\sim{\Delta\over N}$, the backreaction is important and we expect to see deviation from thermal correlators.
Therefore the correlation function in microcanonical essemble will never have singularity away from coincide point.
\subsection{Three Point Function}
The bulk diagram of the three point function will be like Figure \ref{fig:threewavefunction} with additional operator inserting at the intersection points (See Figure \ref{fig:threepointfunction}).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{threepointfunction.pdf}
\caption{Bulk Diagram for Three Point Function}
\label{fig:threepointfunction}
\end{figure}
The QFT three point correlation function in $AdS_2$ is fixed by conformal symmetry and we can write it down as:
\begin{equation}
\langle O_1 O_2 O_3\rangle=C_{123}{z_1^{\Delta_1} z_2^{\Delta_2} z_3^{\Delta_3}\over |x_{12}|^{\Delta_1+\Delta_2-\Delta_3}|x_{23}|^{\Delta_2+\Delta_3-\Delta_1}|x_{13}|^{\Delta_1+\Delta_3-\Delta_2}}={C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
$\Delta_i$ is the conformal dimension of $O_i$. Putting them in formula \ref{final formula} and rewrite the propagator in terms of the wavefunction (\ref{wavefunction}), we have the quantum gravitational three point function:
\begin{equation}
\langle O_1 O_2 O_3\rangle_{QG}=\int_{x_1>x_2>x_3}{\prod_{i=1}^3{dx_i dz_i}\over \text{V(SL(2,R))}}\Psi_{u_{12},\ell_{12}}\Psi_{u_{23},\ell_{23}}\Psi_{u_{13},\ell_{13}}I_{\ell_{12},\ell_{23},\ell_{13}}{C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
We can view this expression as an inner product of three universe wavefunction with the interior, inserting three bilocal operators $\tilde O_{ij;k}\tilde O_{ij;k}$ with dimension $\tilde\Delta_{ij;k}={1\over 2}(\Delta_i+\Delta_j-\Delta_k)$ between them. One can fix the $SL(2,R)$ symmetry and express the integral in terms of $\ell_{ij}$, it is the same exercise as in open string calculation to find the Jacobian factor (Appendix \ref{gauge fix}). Here we can just argue that in order to get the partition function at $\Delta=0$, the measure has to be flat. Therefore the three point function factorizes into form:
\begin{equation}
\small{\langle O_1(u_{1})O_2(u_{2})O_3(u_{3})\rangle_{\text{\rm QG}}}\propto C_{123}\int_0^{\infty}d\tau \rho(\tau)\mathcal{I}_{\tau}(u_{12},\tilde\Delta_{12;3})\mathcal{I}_{\tau}(u_{23},\tilde\Delta_{23;1})\mathcal{I}_{\tau}(u_{31},\tilde\Delta_{31;2})
\end{equation}
while $\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})$ is an integral of $\ell_{ij}$ which gives the two point function in microstate $E_{\tau}$ (\ref{TwoPFMicro}) with the $e^{uE_{\tau}}$ factor stripped off:
\begin{equation}
\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})={1\over 2}\int_{0}^{\infty} d\ell_{ij}\Psi_{u_{ij},\ell_{ij}}{1\over \ell_{ij}^{\Delta_i+\Delta_j-\Delta_k}}K_{2i\tau}({4\over \ell_{ij}})=\langle E_{\tau} |\tilde O_{ij;k}e^{-u_{ij}H}\tilde O_{ij;k}|E_{\tau}\rangle;~~~~~E_{\tau}={\tau^2\over 2}.
\end{equation}
Again the normalization constant can be fixed by choosing $O_i$ to be identity.
\subsection{Einstein-Rosen Bridge}
The Einstein-Rosen Bridge in a classical wormhole keeps growing linearly with time and this behavior was conjectured to related with the growth of computational complexity of the dual quantum state \cite{Stanford:2014jda}.
Based on the universal behavior of complexity growth, Susskind proposed a gravitational conjecture in a recent paper \cite{Susskind:2018fmx} about the limitation of classical general relativity description of black hole interior. The conjecture was stated as follows:
\textsl{Classical general relativity governs the behavior of an ERB for as long as possible.}
In this section, we will test this conjecture using the exact quantum wavefunction of JT gravity (\ref{wavefunction}). We will in particular focusing on the behavior of ERB at time bigger than $1$. The size of Einstein-Rosen Bridge $\mathcal{V}$ in two dimensions is the geodesic distance $d$ between two boundaries, and can be calculated in thermofield double state $|u\rangle$ as:
\begin{equation}
\mathcal{V}=\langle u|d|u\rangle
\end{equation}
We want to focus on the dependence of volume on Lorentzian time evolution. Therefore we do analytic continuation of $u$ in Lorentzian time: $u={\beta\over 2}+it$. Using the WdW wavefunction in $\ell$ basis (\ref{wavefunction}) and the relation between $d$ and $\ell$, we can calculate the expectation value exactly. This can be done by taking the derivative of the two point function (\ref{2point function}) with respect to $\Delta$ at $\Delta=0$. Using the integral representation for $|\Gamma(\Delta+i(s_1-s_2))|^2$, the only time dependence of volume is given by:
\begin{equation}
\mathcal{V}(t)={1\over \mathcal{N}}\int_{-\infty}^{\infty}d\xi\int_0^{\infty} ds_1ds_2\rho_1\rho_2 e^{i(s_1-s_2)\xi-i(s_1-s_2){(s_1+s_2) \over 2}t-{\beta\over 4}(s_1^2+s_2^2)}\log(2\cosh{\xi\over 2})|\Gamma(is_1+is_2)|^2.
\end{equation}
The limit we are interested in is $\beta\ll 1\ll t$ \footnote{Remember that we are measuring time in units of $\phi_r$ (\ref{BCond}), so time order $1$ is a quantum gravity region.}, in which case the integral has a saddle point at \footnote{Actually this saddle point is valid for any range of $t$ as long as $\beta\ll 1$.}:
\begin{equation}\label{SaddlePoint}
s_{1,2}={2\pi\over \beta};~~~~~~~~~\xi={2\pi t\over \beta}.
\end{equation}
Therefore the volume has linear dependence in time:
\begin{equation}
\mathcal{V}(t)\sim{2\pi t\over \beta}.
\end{equation}
Using the complexity equal to volume conjecture \cite{Stanford:2014jda,Brown:2015bva}, the complexity of thermofield double state is proportional to the maximum volume:
\begin{equation}
\mathcal{C}(t)=\#\mathcal{V}(t)=\#{2\pi t \over\beta} .
\end{equation}
The proportionality constant is suggested in \cite{Brown:2018kvn} to be $S_0$ based on classical calculation of near extremal black hole. This, however, is not very clear in our model since $S_0$ is the coupling constant of the pure Einstein-Hilbert action and decouples with JT theory (\ref{full action}).
Since the saddle point (\ref{SaddlePoint}) is actually valid from early time to late time, the proportionality constant can be fixed at classical level and once we fix it we can conclude that \textsf{the length of Einstein-Rosen Bridge (or complexity of the state) keeps linearly growing even considering quantum gravity effects in JT theory.}
We want to comment that this is not an obvious result that one can expect from classical observables.
For example, one might argue that we can extract the information of the ERB from two sided correlators for the reason that semiclassically we can approximate the correlator as $e^{-m d}$.
Therefore one can conclude the ERB has linear growth from the quasinormal behavior of the correlator.
However such observables can only give us information of ERB up to time order $1$, which is the same time scale we can trust the classical general relativity calculation.
After that the correlation function changes from exponential decay into universal power law decay ${1\over t^{3}}$ as one can directly derive from analytic continuation of result (\ref{2point function}).
If we still use such correlator to extract information about ERB we would get the wrong conclusion that it stops its linear growth after time order $1$.
The reason why it is incorrect is that at this time scale the operator disturbs the state and causes different energy states interfere each other strongly. It is simply that the correlator can no longer be described by the classical geometry, rather than the interior stops to behave classically.
From our calculation, we see that if we probe of the state in a weaker and weaker way, we are still able to see the classical geometry.
Lastly, we want to talk a little about when JT gravity needs to be modified. A naive estimate can be made from the partition function that when $\beta$ approaches $e^{{2\over 3}S_0}$, the partition function becomes less than one and definitely at this time scale we need new physics. A recently study of gravitational physics at this time scale was discussed in \cite{Saad:2018bqo}.
\section{Conclusion}
Our result gives an explicit formula (\ref{final formula}) to calculate all order corrections to correlation functions from quantum gravity in two dimensions.
The formula can be understood diagrammatically and we call it Gravitational Feynman Diagram.
We also give the exact Wheeler-DeWitt wavefunction and discuss the growth of its complexity quantum mechanically.
Although we are focusing on theoretical description of two dimensional black holes, the near-extremal black holes in nature should contain these features.
Both Reissner–Nordström black holes and Kerr black holes have an $AdS_2$ throat near their extremality.
For those black holes, the gravitational effects are enhanced by the their near extremal entropies (the coupling constant is $\phi_h$ rather than $\phi_0+\phi_h$) and therefore are better backgrounds to test gravitational effects.
We should however point out that the observational black holes all have large near extremal entropies and thus are very classical \cite{Preskill:1991tb}.
In addition, the Thorne limit of Kerr black hole sets a lower bound on the near extremal entropy in nature.
But for the Primordial black holes in early universe, our story might play a role and it will be interesting to study the physical consequence in that situation.
Another application of our result is to connect with SYK type models \cite{Maldacena:2016hyu,Kitaev:2017awl,Witten:2016iux,Klebanov:2016xxf} since those models all have an emergent Schwarzian action at low energy. On that account, the exact Schwarzian quantization can be used to test ${1\over N}$ corrections to those models. For example, one can try to directly test the two point function with SYK numerics \cite{BKobrin} or one can use SYK models to understand the microscopic description of WdW wavefunction and its complexity. It might also be interesting to consider the black hole information paradox \cite{Fiola:1994ir} and late time traversable wormholes including the quantum fluctuations of the boundary \cite{Gao:2016bin,Maldacena:2017axo,Susskind:2017nto,Yoshida:2017non, Maldacena:2018lmt}.
{\bf Acknowledgments }
The author wants to give special thanks to J.Maldacena and D.Stanford for patient guidance and help all through the project.
We also want to thank A.Almheiri, L.Iliesiu, J.Jiang, J.Ripley, A.Kitaev, B.Kobrin, R.Mahajan, A.M.Polyakov, S.J.Suh, G.Turiaci and H.Verlinde for discussions.
Z.Y is supported by Charlotte Elizabeth Procter Fellowship from Princeton University.
\section{Introduction}
Near-Extremal black holes have a universal structure near their horizons:
there is an $AdS_2$ throat with a slowly varying internal space.
Its low energy gravitational dynamics is captured universally by the following effective action in two dimensions \cite{Strominger:1994tn}:
\begin{equation}\label{full action}
I=\underbrace{-{\phi_0\over 2
}\left(\int R+2\int_{\partial_M}K\right)}_{\text{Einstein-Hilbert Action}}\underbrace{-{1\over 2
}\left(\int_M \phi(R+2)+2\int_{\partial M}\phi_b K\right)}_{\text{Jackiw-Teitelboim action}}+S_{matter}(g,\psi),
\end{equation}
where the dilaton field $\phi+\phi_0$ represents the size of internal space. We have separated the size of internal space into two parts: $\phi_0$ is its value at extremality. It sets the value of the extremal entropy
which comes from the first term in \nref{full action}.
$\phi$ is the deviaton from this value.
We have also added matter that only couples to the metric. This is a reasonable assumption when matter comes
from Kaluza Klein reduction, where the coupling to the dilaton would involve $\phi/\phi_0 \ll 1$.
The action $\int \phi(R+2)+2\int\phi_b K$ is the so-called Jackiw-Teitelboim action \cite{JACKIW1985343,TEITELBOIM198341}, and will be the main focus of our paper.
This action is one of the simplest nontrivial gravitational actions in two dimensions.\footnote{Another nontrivial action is the CGHS model which could be written as $\int (\phi R+C)$, and that characterizes the horizon structure of general black holes.}. It is simple because the bulk geometry is a rigid $AdS_2$ space fixed by the equation of motion of the dilaton field. Its nontriviality arises from the remaining boundary action. Schematically, the gravitational action is reduced to the following form:
\begin{equation}\label{schematic action}
I=-{ 2 \pi \phi_0
}\chi(M)-{\phi_b
}\int_{\partial M}K +S_{matter}(g,\psi).
\end{equation}
And the motion of the boundary is controlled by its extrinsic curvature.
Our goal will be to quantize this action and to provide expression for the full quantum gravity correlators
of \nref{full action}.
This problem was considered before in \cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Kitaev:2018wpr} from various points of view. Here we will add one other point of view where we
reduce the problem to the motion of a relativistic particle in an electric field, building on a suggestion in
Kitaev's talk at IAS \cite{kitaevIAS}. More precisely, one can consider a relativistic particle in a Lorentzian $AdS_2$ target space moving under the influence of an electric field. The coupling to electric field can also be viewed as a coupling to a spin connection so that it becomes a particle with spin as suggested by Kitaev. Alternatively we can start from a non-relativistic particle moving in hyperbolic space, $H_2$, under the influence of a magnetic field $b$. After analytic
continuation in $b$ to imaginary values we get the problem of interest.
Using this point of view one can think of the full quantum gravity problem as the combination of two problems.
First we consider quantum fields propagating in $AdS_2$ (or $H_2$ in the Euclidean case) and then we
add the ``gravitational particle" which couples to the quantum fields by changing their boundary location in $AdS_2$.
The discussion of quantum fields will be standard and depends on the particular model one interested in, therefore we will mainly focus on solving the second problem.
Generically, solving the gravitational problem is challenging and is not exactly equivalent to a quantum mechanical particle. One needs to worry about what functional space one will integrate over. For example, in path integrals, one usually integrates over all trajectories including those with self-intersections. However self-intersecting boundaries in gravitational system have no obvious meaning. On that account, more precisely the gravitational problem is equal to a self-avioding particle.
Nevertheless, it turns out that one can take a particular limit of this model, namely large $\phi_b$, to avoid this issue and a treatment of the boundary theory as an ordinary particle is justified. It is also true that the JT gravity can be rewritten as a Schwarzian action only in this limit. We call this the Schwarzian limit and will only focus on solving the JT action in the Schwarzian limit. Solving the model away from Schwarzian limit was considered recently by Kitaev and Suh \cite{Kitaev:2018wpr}.
Our result can be summarized as follows:
\textsl{First, we will give a formula to calculate all correlation functions with quantum gravity backreaction (formula \ref{final formula}). Second, we will give the exact Wheeler-DeWitt wavefunction in the Schwarzian limit, which has been analyzed classically by Harlow and Jafferis \cite{Harlow:2018tqv}. Last, we consider the recent proposed conjecture about complexity growth in this exact Wheeler-DeWitt wavefunction and show that the complexity maintains linear growth after taking quantum gravity effects into account. This, to our knowledge, is the first test of the gravitational conjecture made by Susskind \cite{Susskind:2018fmx} that the size of ERB grows linearly for as long as quantum mechanics allows.}
This paper is organized as follows: in the first section, we will review the classical calculation of this model and introduce notations; in the second section, we will make the dictionary between the JT model and a particle in a magnetic field; in the third and fourth sections, we will solve the quantum mechanical problem and derive the propagator and WdW wavefunction in the Schwarzian limit; in the last section, we will talk about gravitaional backreaction on correlators as well as complexity growth. As a useful notion in our calculation, we introduce a notation called gravitational Feyman diagrams.
\section{Classical Solutions}
Let us first consider the classical solutions of
the Jackiw-Teitelboim model \nref{full action}. See
\cite{Almheiri:2014cka} for further discussion.
%
%
The equation of motion of the dilaton field imposes $R=-2$
and
fixes the geometry to be $AdS_2$, or $H_2$ in the Euclidean case.
This is also true if
we have additional matter coupled with metric only, as in
\nref{full action}.
The equations for the metric constrain the dilaton
\begin{equation}\label{Einstein Equation}
(\nabla_{\mu}\nabla_{\nu}\phi-g_{\mu\nu}\nabla^2\phi+g_{\mu\nu}\phi)+T^{M}_{\mu\nu}=0
\end{equation}
These equations are compatible with each other thanks to the conservation of the matter stress tensor. They do not allow
any propagating mode. In fact, setting $T^M_{\mu\nu}=0$, and using a high momentum approximation we can write
\nref{Einstein Equation} as $ (k_{\mu}k_{\nu}-g_{\mu\nu}k^2)\phi(\vec k)=0 $, which then implies $ \phi(\vec k)=0 $, for large $k$.
More precisely, after introducing the Euclidean $AdS_2$ coordinates $ds^2 = d\rho^2 + \sinh^2\rho d\varphi^2$, we can solve
(\ref{Einstein Equation}) in $AdS_2$ with no matter. Up to an SL(2) transformation the solution is
\begin{equation}
\phi=\phi_h \cosh \rho,
\end{equation}
where $\phi_h$ is a constant that is fixed by the boundary conditions.
At the boundary we fix the metric along the boundary and the value of the dilaton field
\begin{equation} \label{BCond}
ds_{\parallel} = d u { \phi_r \over \epsilon } ~,~~~~~~~~~~~ \phi= \phi_b = { \phi_r \over \epsilon }
\end{equation}
where we think of $u$ as the time of the boundary theory. It is simply a rescaled version of proper time. Similarly $\phi_r$ is a rescaled
value of the dilaton. We will be interested in taking $\epsilon \to 0$. With these rescalings the value of $\phi_h$ in the interior
remains fixed as we take $\epsilon \to 0$ as $\phi_h = 2 \pi /\beta $ where $\beta$ is the inverse temperature, $u \sim u + \beta$. Notice that due to the factor of $\phi_r$ in the first expression in \nref{BCond} we are measuring time in units of the constant $\phi_r$, which has dimensions of length. We did this for convenience.
A nice feature that appears after taking the $\epsilon \to 0$ limit is that the action (\ref{schematic action}) can be written as the Schwarzian action for the boundary curve labeled by $\varphi(u)$ \cite{Maldacena:2016upp}:
\begin{equation}
I=-\int du Sch(\tan{\varphi(u)\over 2},u)
\end{equation}
The fluctuation of the boundary shape can be understood as the fluctuation of the dilaton distribution in the bulk.
A bit more explicitly we can say that the dilaton boundary condition fixes the location of the boundary at $\rho_b$ given by
$\phi_b = \phi_h \cosh \rho_b$, and the metric at that location relates the time $\varphi$ to $u$ by $\phi_r du = \epsilon \sinh \rho_b d\varphi $.
We get the above formulas noticing that the period of $\varphi$ is $2\pi$ while that of $u$ is $\beta$, which fixes $\epsilon \sinh \rho_b$.
\section{Charged Particle in $AdS_2$}
Despite the absence of a bulk propagating mode there is still a non-trivial dynamical gravitational degree of freedom. There are various
ways to describe it. Here we will think of it as arising from the motion of the physical boundary of $AdS_2$ inside a rigid $AdS_2$ space.
This picture is most clear for finite $\epsilon$ in \nref{BCond}, but it is true even as $\epsilon \to 0$. The dynamics of the boundary
is SL(2) invariant. This SL(2) invariance is a gauge symmetry since it simply reflects the freedom we have for cutting out a piece
of $AdS_2$ space that we will call the ``inside". It is important that the dilaton field we discussed above is produced after we put in the
boundary and it moves together with the boundary under this SL(2) gauge transformation. It is a bit like the Mach principle, the location
in $AdS_2$ is only defined after we fix the boundary (or distant ``stars").
We can make this picture of a dynamical boundary more manifest as follows.
Since the bulk Jackiw-Teitelboim action \nref{full action} is linear in $\phi$, we can integrate out the dilaton field which
sets the metric to that of $AdS_2$ and removes the bulk term in the action, leaving only the term involving the extrinsic curvature
\begin{eqnarray}
I&=&-{\phi_r\over
\epsilon}\int du \sqrt{g}K
\end{eqnarray}
This action, however, is divergent as we take $\epsilon$ to zero. This divergence is simply proportional to the length of the boundary
and can be interpreted as a contribution to the ground state energy of the system.
So we introduce a counterterm proportional to the length of the boundary to cancel it. This is just a common shift of the energies
of all states.
It is also convenient to use the
Gauss-Bonnet theorem to relate the extrinsic curvature to an integral over the bulk
\begin{equation}
\int_{\partial_M} du \sqrt{g}K=2\pi\chi(M)-{1\over 2}\int_M R
\end{equation}
Since the curvature is a constant, the bulk integral is actually proportional to the total area $A$ of our space. That is we have the regularized action:
\begin{eqnarray}\label{subtraction}
I&=&-{\phi_r\over
\epsilon}\int_{\partial M} du \sqrt{g}(K-\underbrace{1}_{\text{counterterm}})=-{\phi_r\over
\epsilon}\left(2 \pi\chi(M)-{1\over 2}\int_{M}R-\int_{\partial M} du \sqrt{g}
\right) \nonumber\\
&=&-2\pi q\chi(M)-q A+q L,~~~~~~~~~~~~~~~~~~~~~~~~~~~q\equiv{\phi_r\over
\epsilon} ,~~L={\beta \phi_r\over \epsilon}
\end{eqnarray}
We now define an external gauge field $a_\mu$ as
\begin{equation}
a_{\varphi}=\cosh\rho-1,~~~~~~~a_{\rho}=0,~~~~~~~f_{\rho \varphi}=\sinh\rho=\sqrt{g},
\end{equation}
and write the action as follows
\begin{equation}\label{regularized action}
I= - 2 \pi q + q L - q \int a
\end{equation}
where we used that $ \chi(M)$ is a topological invariant equal to one, in our case, where the topology is that of a disk.
The term $q L$ is just the length of the boundary. So this action has a form somewhat
similar to the action of a relativistic charged particle moving in
$AdS_2$ in the presence of a constant electric field.
There are a couple of important differences. First we are summing only over trajectories of fixed proper length set by the inverse temperature
$\beta$. Second, in the JT theory we are treating the $SL(2)$ symmetry as a gauge symmetry. And finally, in the JT theory we identify the proper length with the boundary time, viewing configurations which differ only by a shift in proper time as inequivalent. In fact, all these changes simplify the problem: we can actually think
of the problem as a non-relativistic particle moving on $H_2$ in an electric field. In appendix
\ref{appendix:relativistic particle} we discuss in more detail the connection to the relativistic particle.
In fact, precisely the problem we are interested in has been discussed by Polyakov in \cite{Polyakov:1987ez}, Chapter 9, as an an intermediate step for the sum over paths.
Now we would also like to point out that we can directly get to the final formula by
using the discussion there, where he explicitly shows
that for a particle in flat space the sum over paths of fixed proper length that stretch between
two points $\vec x$ and $\vec x'$ gives
\begin{equation} \label{PolFlat}
\int {\cal D }\vec x e^{ - m_0 \tilde\tau } \delta( \dot { \vec x}^2 - 1 ) =
e^{ - {1\over 2} \mu^2 \tau} \langle x' | e^{ - \tau H } |x\rangle =
e^{ - {1\over 2} \mu^2 \tau} \int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot \vec x^2 }
\right)
\end{equation}
$\mu^2$ is the regularized mass and $\tilde \tau$ is related to $\tau$ by a multiplicative renormalization.
The JT model consists precisely of a functional integral of this form, where we fix the proper length along the boundary. There are two simple modifications, first the particle is in a curved $H_2$ space and second we have the coupling to the electric field. These are minor modifications, but the arguments leading
to \nref{PolFlat} continue to be valid so that the partition function of the JT model can be
written directly:
\begin{equation} \label{PolAdS}
\scaleto{\int {\cal D }\vec x e^{2\pi q - m_0 \tilde\tau +q\int a } \delta( { \dot x^2 + \dot y^2 \over y^2 }- q^2 ) =
e^{ 2\pi q- {1\over 2} \mu^2 \tau} \Tr e^{ - \tau H } = e^{2\pi q - {1\over 2} \mu^2 \tau }
\int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } - q {\dot x \over y}
\right)}{26pt}
\end{equation}
The delta function implements the first condition
in \nref{BCond} at each point along the path.
The last path integral can be done exactly by doing canonical quantization of the action (section \ref{CanonicalQuantization}) and by comparing the result with the one from the Schwarzian action \cite{Stanford:2017thb}
we can determine that $\tau$ is the inverse temperature $\beta$.
In the above discussion we have been fixing the time along the boundary. Instead we can fix the energy
at the boundary, where the energy is the variable conjugate to time.
This can be done by simply integrating \nref{PolAdS} times $e^{ \beta E}$
over $\beta$ along the imaginary axis. This fixes the energy of the non-relativistic problem by generating a $\delta( H - E)$. More precisely, we will argue that after doing a spectral decomposition we can write
the propagator at coincident points as
\begin{equation} \label{JTEn}
Z_{JT}(\beta ) = \int_0^\infty \rho(E)e^{-\beta E} dE \longrightarrow \rho(E) = \int_{-i \infty}^{i \infty}
{ d\beta \over i } e^{ E \beta } Z_{JT}(\beta)
\end{equation}
where the function $\rho(E) $ can then be interpreted as a ``density of states" in the microcannonical ensemble. We will give its explicit form in section (4.2).
For now, we only want to contrast this integral with a superficially similar one that appears when we
compute the relativistic propagator
\begin{equation} \label{RelEl}
e^{-2\pi q} \int_0^\infty e^{ E \beta }Z_{JT} (\beta) = \langle \phi(x) \phi(x) \rangle
\end{equation}
which gives the relativistic propagator of a massive particle in an electric field at coincident points
(we can also compute this at non-coincident points to get a finite answer). The total mass of the
particle is
\begin{equation}
m = q - {E\over q}
\end{equation}
For large $q$ this is above threshold for pair creation\footnote{See Appendix \ref{appendix:relativistic particle}}. The pair creation interpretation
is appropriate for the problem in \nref{RelEl}, but not for \nref{JTEn}.
In both problems we have a classical approximation to the dynamics that corresponds to a particle
describing a big circular trajectory in hyperbolic space at radius $\rho_c$ :
\begin{equation} \label{SolCir}
\tanh \rho_c = { m \over q }
\end{equation}
For the problem in \nref{RelEl}, fluctuations around this circle lead to an instability, with a single negative mode and an imaginary part in the partition function \nref{RelEl}. This single negative mode corresponds to small fluctuations of the overall size of the circular trajectory around \nref{SolCir}.
On the other hand in \nref{JTEn} we are integrating the same mode along a different contour, along the
imaginary axis, where we get a real and finite answer. Furthermore, the imaginary part in the
partition function \nref{RelEl} comes precisely from the trajectory describing pair creation, which is also the type of contribution captured in \nref{JTEn}.
Finally, in the relativistic particle problem, we expect that the pair creation amplitude should be exponentially suppressed for large $q$, while the partition function for the JT model is not. In fact,
for large $q$ the exponential suppression factor for pair creation goes as $e^{ - 2 \pi q } $, which is precisely cancelled by a similar factor in \nref{JTPa}, to obtain something finite in the large $q$ limit.
\section{Solving the Quantum Mechanical Problem}
\label{CanonicalQuantization}
As we explained above the solution of the JT theory is equivalent to considering a non-relativistic particle in $AdS_2$ or $H_2$.
We first consider the Euclidean problem, of a particle moving in $H_2$. An ordinary magnetic field in $H_2$ leads to an Euclidean action
of the form
\begin{equation} \label{magnetic field}
S= \int du {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } + i b \int d u { \dot x \over y } -{1\over 2} ( b^2 +{ 1 \over 4} ) \int d u ~,~~~~~~~b = i q
\end{equation}
If $b$ is real we will call it a magnetic field,
when $q$ is real we will call it an ``electric" field. The last term is a constant we introduced for convenience. Its only effect
will be to shift the ground state energy. It is interesting to compute the classical solutions and the corresponding action for
\nref{magnetic field}. These solutions are simplest in the $\rho$ and $\varphi $ coordinates, using the SL(2) symmetry we find that the trajectories are given by ($t=-iu$):
\begin{eqnarray}
{1\over 2}\sinh^2\rho ({d\varphi\over dt})^2+{q^2\over 2}-{1\over 8}=E,~~ \cosh\rho={q\beta\over 2\pi},~~ {d\varphi\over du}={2\pi\over \beta}.
\end{eqnarray}
In this classical limit we
get the following relations for the action and the temperature:
\begin{eqnarray}
& ~& { \beta \over 2 \pi } = {1\over \sqrt{2E+{1\over 4}}}
\cr
& ~ & -S = {2\pi^2\over \beta}+{\beta\over 8}-2\pi q \label{ClassEnt}
\end{eqnarray}
When $b$ is real, this system is fairly conventional and it was solved in \cite{Comtet:1986ki} . Its detailed spectrum depends on $b$. For very large $b$ we have a series of Landau levels and also a continuous spectrum. In fact, already the classical problem contains closed circular orbits, related to the discrete Landau levels, as well as orbits that go all the way to infinity.\footnote{See Appendix \ref{appendix:landau level}}
The number of discrete Landau levels decreases as we decrease the magnetic field and for $0 < b < {1\over 2}$ we only get a continuous spectrum. The system has a $SL(2)$ symmetry and the spectrum organizes into SL(2) representations, which are all in the
continuous series for $0 < b < 1/2$. For real $q$ we also find a continuous spectrum which we can view as the analytic continuation
of the one for this last range of $b$.
The canonical momenta of the action (\ref{magnetic field}) are:
\begin{equation}
p_x={\dot x\over y^2}+{i q\over y};~~~~~~~p_y={\dot y\over y^2}.
\end{equation}
And the Hamiltonian conjugate to $\tau_L$ is thus:
\begin{equation}
H={\dot x^2+\dot y^2\over 2y^2}+{q^2\over 2}={y^2\over 2} [(p_x-i{q\over y})^2+p_y^2] +{ q^2 \over 2 } - { 1 \over 8}
\end{equation}
Note that the Hamiltonian is not Hermitian. However, it is {\cal PT}-symmetric (here parity reflects $x$ and $p_x$) and for that
reason the spectrum is still real, see \cite{Bender:2005tb}.
The action is invariant under $SL(2,R)$ transformations generated by
\begin{eqnarray}
L_0=x p_x+y p_y;~~~~~L_{-1}=p_x;~~~~~~~~L_{1}=(y^2-x^2)p_x-2xyp_y-2iqy \label{SLtwoGen}
\end{eqnarray}
Notice the extra $q$ dependent term in $L_1$ that arises due to the presence of a magnetic field.
Up to a simple additive constant, the Hamiltonian is proportional to the Casimir operator
\begin{equation}
H={1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) + { q^2 \over 2} - { 1 \over 8 }
\end{equation}
As is common practice, let us label the states by quantum numbers $j={1\over 2}+is$ and $k$, so that $H|j,k\rangle=j(1-j)|j,k\rangle$ and $L_{-1}|j,k\rangle=k|j,k\rangle$.
We can find the eigenfunctions by
solving the $Schr\ddot{o}dinger$ equation with boundary condition that the wavefunction should vanish at the horizon $y\rightarrow \infty$ \cite{Comtet:1984mm,Comtet:1986ki,Pioline:2005pf}:
\begin{equation}
\omega_{s,k}={s^2\over 2} ,~~~~~~f_{s,k}(x,y)=
\begin{cases}
({s\sinh 2\pi s\over 4\pi^3k})^{1\over 2}|\Gamma(is-b+{1\over 2})|e^{-ik x}W_{b,is}(2 k y), ~~~k>0;\\
({s\sinh 2\pi s\over 4\pi^3|k|})^{1\over 2}|\Gamma(is+b+{1\over 2})|e^{-ik x}W_{-b,is}(2|k|y),~~~k<0.
\end{cases}
\end{equation}
where $\omega_{ks}$ is giving the energy of the states labelled by $s$ and $k$, and $ W $ is the Whittaker function. The additive constant in \nref{magnetic field} was
introduced to simplify this equation. We can think of $s$ as the quantum number of the continuous series representation of $SL(2)$ with
spin $j={1\over 2} + i s $.
After continuing $b \to i q$ we find that the gravitational system has a continuous
spectrum
\begin{equation}\label{energy relation}
E(s)={s^2 \over 2 }.
\end{equation}
\subsection{The Propagator }
It is useful to compute the propagator for the non-relativistic particle in a magnetic field,
$K(u,\boldsymbol{x_1},\boldsymbol{ x_2})=\langle \boldsymbol{x_1}| e^{-u H}|\boldsymbol{x_2}\rangle$. Here, $\boldsymbol{x}$ stands for ${x,y}$.
The propagator for a real magnetic particle was obtained in \cite{Comtet:1986ki}:
\begin{eqnarray} \label{particle heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})& =& e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} \int_0^{\infty}ds s e^{- u { s^2 \over 2}}{\sinh 2\pi s\over 2\pi (\cosh 2\pi s+\cos 2\pi b)}{1\over d^{1+2is}}
\times
\cr
& & ~~~~~ \times ~_2F_1({1\over 2}-b+is,{1\over 2}+b+is,1,1-{1\over d^2}).
\cr
d & = & \sqrt{(x_1-x_2)^2+(y_1+y_2)^2\over 4y_1 y_2}
\cr
e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} &=& e^{-2 i b \arctan {x_1-x_2\over y_1+y_2}}
\end{eqnarray}
In the case that we have a real magnetic field the prefactor is a phase and it is
gauge dependent. It is equal to the value of Wilson Line $e^{i\int a}$ stretched along the geodesic between $x_2$ and $x_1$.
Here we quoted the value in the gauge where the action is
\nref{magnetic field}. The second equation defines the parameter $d$, which is a function of the geodesic distance between the two points. Note that $d=1$ corresponds to coincident points.
We can get the answer we want by making the analytic continuation $b\to i q$ of this formula.
We can check that this is the right answer for our problem by noticing the following. First one can check that this expression is invariant under the SL(2) symmetry $L_a = L^1_a + L^2_a$ where $L_a$ are the generators \nref{SLtwoGen} acting on $\boldsymbol{x_1}$
and $L_a^2$ are similar generators as in \nref{SLtwoGen}, but with $q\to -q$. It is possible to commute
the phase $e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})}$, in \nref{particle heat kernel} past these generators which would remove the $q$ dependent terms. This implies
that the rest should be a function of the proper distance, which is the case with \nref{particle heat kernel}.
Then we can check the equation
\begin{eqnarray}
0 &=& (\partial_u + H_1 )G(u,\boldsymbol{x_1},\boldsymbol{x_2})
\end{eqnarray}
which is also indeed obeyed by this expression. The $s$ dependent prefactor is fixed by the requirement that the
propagator composes properly, or more precisely, by saying that for $u=0$ we should get a $\delta$ function.
\subsection{Partition Function}
The gravitational partition function is related with the particle partition function with inverse temperature $\beta$.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{Freeenergy.pdf}
\caption{Free Energy diagram with inverse temperature $\beta$. }
\end{figure}
The canonical partition function of the quantum mechanical system is
\begin{eqnarray}
Z_{\scaleto{Particle}{4pt}}&=&Tr e^{-\beta H}
=\int_0^{\infty} ds \int_{-\infty}^{\infty}dk \int_M {d x d y\over y^2} e^{-{\beta } { s^2 \over 2 } }f^*_{s,k}(x,y)f_{s,k}(x,y)\nonumber\\
&=&V_{AdS}\int_0^{\infty}ds e^{-\beta {s^2\over 2}}{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}.
\end{eqnarray}
The volume factor $V_{AdS}$ arises because after momentum integration there is no position dependence.
In a normal quantum mechanical system, the volume factor means that the particle can have independent configurations at different locations of our space, however for a gravitational system this should be thought as redundant and should be cancelled by the volume of $SL(2,R) $ gauge group $2\pi V_{AdS}$\footnote{There might be a multiplicative factor in the volume of gauge group, but we can always absorb that into $S_0$.}.
In gravitational system, there can also other contributions to the entropy from pure topological action.
These give a contribution to the ground state entropy $S_0$. Including the topological action in (\ref{regularized action}), we find a gravitational ``density of states" as
\begin{equation} \label{Spectral}
\rho(s)=\underbrace{e^{S_0}e^{2\pi q}}_{\text{extra terms}}\underbrace{1\over {2\pi}}_{\text{residue gauge}}\underbrace{{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}}_{\text{particle in magnetic field}}=e^{S_0}e^{2\pi q}{s\over 2\pi^2}\sum_{k=1}^{\infty}(-1)^{k-1}e^{-2\pi q k}\sinh(2\pi s k).
\end{equation}
We have not given an explicit description of these states in the Lorentzian theory. More details were discussed in \cite{Kitaev:2018wpr,Lin:2018xkj}.
This expression has some interesting features. Notice that the classical limit corresponds to large $q$ and large $s$, where we reproduce \nref{ClassEnt}. After approximating, the density of states are
$\log \rho(s) \sim S_0 + 2 \pi s $ for $ s/q < 1$ and $S_0+2\pi q$ for $s/q>1$.
We can also expand the partition function for very small and very large temperatures where we obtain
\begin{eqnarray}
Z_{JT} &\sim & e^{ S_0} e^{ 2 \pi q } { 1 \over 4\pi^2\beta} ~,~~~~~~~~~~~ \beta \ll {1\over q}
\cr
Z_{JT} & \sim & e^{S_0} { 1 \over \sqrt{2\pi}\beta^{3/2} }~,~~~~~~~~~~~~ \beta \gg 1
\end{eqnarray}
Notice that at leading order we get an almost constant entropy both at low and high temperatures, with the
high temperature one being higher. In both cases there are power law corrections in temperature.
Before we try to further elucidate the interpretation of this result, let us emphasize a couple of important defects of our discussion. First, when we replaced the partition function of the JT theory by the action of a non-relativistic
particle in an electric field, we were summing over paths in $H_2$. This includes paths that self intersect see figure \ref{fig:TwoInstanton}. Such paths do not have an obvious interpretation in the JT theory and it is not even clear that we should include them. For example, the sum over $k$ in \nref{Spectral} can be understood in terms of classical solutions which wind $k$ times around the circle. These make sense for the problem of the particle in the electric field but apparently not in the JT theory.
\begin{figure*}[h!]
\centering
\subfigure[Density of States]{%
\includegraphics[width=0.4\textwidth]{densityofstate.pdf}%
\label{fig:dos}%
}\hspace{4cm
\subfigure[Two Instantons]{%
\includegraphics[width=0.35\textwidth]{twoinstanton.pdf}%
\label{fig:TwoInstanton}%
\caption{}
\end{figure*}
Maybe such paths could be given some interpretation in the gravity theory.
Alternatively, we might want to sum over paths that do not self intersect.
A second defect is that we would be eventually interested in adding some matter fields propagating in the bulk geometry. These matter fields have boundary conditions at the boundary of the region of hyperbolic space
cut out by the boundary trajectory. The partition function of the fields with such an arbitrary boundary trajectory could also modify the results we described above. Of course, this issue does not arise if we have
the pure JT theory. It is only important if we want to introduce bulk matter fields to define more complex observables.
Instead of attempting to address the above issues, we will take an easy route, which is to consider the system only in the large $q$ (or small $\epsilon$) limit. In this regime, we address the above issues, and we can still trust the description of the particle in the electric field.
This large $q$ or small $\epsilon$ limit is the same one that isolates the Schwarzian action from the JT theory
\cite{Maldacena:2016upp,Jensen:2016pah,Engelsoy:2016xyb}.
It turns out that the limit can be taken already at the level of the mechanical system, a simple rescaled
version of the above system. This provides an alternative method for quantizing the Schwarzian theory. It has the advantage of being a straightforward second order action of a particle moving in a region near the boundary of hyperbolic. Of course, the Schwarzian theory was already quantized using a variety of methods in
\cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Luca}. We will simply provide yet another perspective, recover the old results, and write a few new expressions.
\section{Quantum Gravity at Schwarzian Limit}
Before getting into the details notice that the large $q$ limit of \nref{Spectral} gives
\begin{equation}
\rho(s) = e^{S_0}{s\over 2\pi^2}\sinh(2\pi s), ~~~E = { s^2 \over 2} ~,~~~~~~~Z_{JT} = \int_0^\infty ds
\rho(s) e^{ - \beta { s^2 \over 2 } } =e^{S_0}{1 \over\sqrt{2\pi}\beta^{3\over 2}}e^{2\pi^2 \over \beta}.
\end{equation}
This reproduces what was found in \cite{Stanford:2017thb,Mertens:2017mtv,Mertens:2018fds,Blommaert:2018oro} by
other methods. We see that we get a finite answer and also that the contributions from the $k>1$ terms in
\nref{Spectral} have disappeared. Because the $S_0$ part decouples with JT gravity, from now on, we will drop it and discuss $S_0$ only when it is necessary.
\subsection{The Propagator}
To get a limit directly at the level of the mechanical system it is useful to define a
rescaled
coordinate, $z$, via
\begin{equation}
y= z/q. \label{yAndq}
\end{equation}
After taking the large $q$ limit, the boundary particle propagator becomes \footnote{see the Appendix \ref{appendix:large q} for details}:
\begin{eqnarray}\label{heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&{ 1 \over q } e^{-2\pi q \theta(x_2-x_1) }\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2});~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~q\gg 1.\\
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{2\sqrt{z_1z_2}}{\pi^2 |x_1-x_2|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|});~~~~~~~~
\label{HeatLim}\\
&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{\sqrt{2}}{\pi^{3/2} u^{3/2}}\frac{\sqrt{z_1z_2}}{ |x_1-x_2|}\int_{-\infty}^{\infty}d\xi (\pi+i\xi)e^{-2\frac{(\xi-i\pi)^2}{u}-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}.\label{HeatLim2}
\end{eqnarray}
The original phase factor $e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_1})}$ factorizes into a product of singular ``phase" $e^{-2\pi q \theta(x_2-x_1)}$, with $\theta$ the step function,
and a regular ``phase" $e^{-2 {z_1+z_2\over x_1-x_2}}$. The singular ``phase" is the same order as the topological piece in (\ref{regularized action}).
In order to have a finite result they should cancel between each other. This can only be satisfied if the $x_is$ are in cyclic order.
As shown in figure (\ref{fig:phasefactor}), the product of singular ``phase" gives $-2\pi q$ for cyclic order $x_i$s and this would cancel with the topological action $2\pi q$.
While for other ordering of the $x_i$s, this would have $-2 \pi n q$ for $n=2,3,...$ and is highly suppressed in the limit $q$ goes to infinity.
This cyclic order is telling us where the interior of our space time is. The magnetic field produces a preferred orientation for the propagator.
After fixing the order, all our formulas only depend on $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ which has no $q$ dependence.
The residual $q$ factor in \ref{heat kernel} cancels out the additional $q$ from the measure of coordinate integral, ${ dx dy \over y^2 } \to q { dx dz \over z^2 } $.
In conclusion, after taking the limit we get a finite propagator equal to \nref{HeatLim}, which should be multiplied by a step function $ \theta( x_1 - x_2 ) $ that imposes the right order.
The final function $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ has the structure of $e^{-2{z_1+z_2\over x_1-x_2}}f(u,{z_1z_2\over (x_1-x_2)^2})$. This can be understood directly from the $SL(2)$ symmetry.
After taking the large $q$ limit, the $SL(2,R)$ charges become
\begin{equation} \label{SLTwoLq}
L_0=i (x\partial_x+z\partial_z);~~~~L_{-1}=i\partial_x;~~~~~~L_{1}=-ix^2\partial_x-2ix z\partial_z-2i z.
\end{equation}
We can check that they still satisfy the SL(2) algebra.
If we drop the last term in $L_1$, the $SL(2,R)$ charges become the usual differential operators on $EAdS_2$. And the propagator will have only dependence on the geodesic distance. When $L_1$ operator is deformed, the condition of $SL(2,R)$ invariance fixes the structure of the propagator as follows.
The $L_0$ and $L_{-1}$ charges are not deformed and they imply that the
only combinations that can appear are
\begin{equation}
v \equiv {z_1+z_2\over x_1-x_2} ~~~~~~~~~~~~~\text{and}~~~~~~~~~~~~~~ w\equiv {z_1 z_2\over(x_1-x_2)^2}.
\end{equation}
Writing the propagator as
$\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_1})=k(v,w) $
and
requiring it to be invariant under $L_1$ gives the following
equation for $\alpha$:
\begin{equation}\label{phaseequation}
\partial_v k +2 k =0 \longrightarrow k = e^{ - 2 v } h(w)
\end{equation}
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=e^{-2 {z_1+z_2\over x_1-x_2}}f(u,{z_1 z_2\over(x_1-x_2)^2}),
\end{equation}
The full function can also be determined directly as follows. Again we impose the propagator equation (or heat equation)
\begin{eqnarray}
0 &=& \left[ \partial_u + {1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) -{1\over 8} \right] \tilde K
\cr
0 & = & [ { s^2 \over 2} + {w^2\over 2}\partial_w^2+2w +{1\over 8}] K_s(w)
\end{eqnarray}
where $L_a$ are given in \nref{SLTwoLq} and are acting only on the first argument of $\tilde K$.
The solution of the last equation which is regular at short distances ($w \to \infty$) is $\sqrt{w}$ times
the Bessel K
function in \nref{heat kernel}.
We can also directly determine the measure of integration for $s$ by demanding that the propagator at $u=0$ is
a $\delta$ function or by demanding the propagator compose properly. This indeed is the case with the $s\sinh{ 2 \pi s}$ function in \nref{HeatLim}.
To explicitly show the above statement, it will be useful to use spectral decomposition of the propagator:
\begin{equation}\label{SpectralDecomposition}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=\int ds {2s\sinh(2\pi s)\over \pi^3} e^{-{s^2u\over 2}}\int dk \sqrt{z_1 z_2} e^{ik(x_1-x_2)}K_{2is}(2\sqrt{2i k z_1})K_{2is}( 2\sqrt{2ikz_2}).
\end{equation}
It can be easily checked that the special functions $f_{k,s}(x,z)=\sqrt{z}e^{ikx}K_{2is}( 2\sqrt{2ikz})$ are delta function normalizable eigenmodes of the large $q$ Hamiltonian:
\begin{equation}
\int {dx dz\over z^2}f_{k_1,s_1}f_{k_2,s_2}=\delta(k_1-k_2)\delta(s_1-s_2){\pi^3\over 2s\sinh (2\pi s)}
\end{equation}
Notice that the inner product fixes the integral measure completely in (\ref{SpectralDecomposition}), and the composition relation is manifestly true:
\begin{equation}
\int {dx dz\over z^2}\tilde{K}(u_1,\boldsymbol{x_1},\boldsymbol{x})\tilde{K}(u_2,\boldsymbol{x},\boldsymbol{x_2})=\tilde{K}(u_1+u_2,\boldsymbol{x_1},\boldsymbol{x_2})
\end{equation}
At short time the propagator has the classical behavior:
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})\sim \delta(x_1-x_2+u z_2)e^{-{(z_1-z_2)^2\over 2 u z_2}}
\end{equation}
This form of singularity is expected since we are taking the large $q$ limit first and thus the velocity in $x$ direction is fixed to be $z$. In the original picture of finite $q$ we are looking at the time scale which is large compare to AdS length but relatively small such that the quantum fluctuations are not gathered yet.
The integral structure in the propagator (\ref{HeatLim}) has an obvious meaning: integrating over $s$ represents summing over all energy states with Boltzmann distribution $e^{-E u}$, and the Bessel function stands for fixed energy propagator.
We want to stress that the argument in the Bessel function is unusual, and at short distance it approaches a funny limit:
\begin{equation}
K_{2is}(\frac{4 }{\ell})\simeq \sqrt{\pi \over 8\ell}e^{-\frac{4 }{\ell}},~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}\rightarrow 0.
\end{equation}
One should contrast this exponential suppression with the short distance divergence in QFT which is power law.
In our later discussion of exact correlation function with gravity backreaction, we will see that this effect kills UV divergence from matter fields.
To obtain the expression (\ref{HeatLim2}), we use the integral representation for the Bessel function and the final result has some interesting physical properties:
Firstly, we see that at large $u$ the time dependence and coordinate dependence factorized.
So, at large time we have a universal power law decay pointed out in \cite{Bagrets:2016cdf}.
Secondly, as we said before, the phase factor $e^{-2\frac{z_1+z_2}{x_1-x_2}}$ is equal to the Wilson line $e^{-q\int_1^2 a}$ stretched along the geodesic connection between location $1$ and $2$ (Figure \ref{fig:propagator}). The field $a$ depends on our choice of gauge, our convention corresponds to fix the minimum value of $a$ at infinity and then the Wilson line is equal to $e^{-q A}$, where $A$ is the area of a hyperbolic triangle spanned by $1,2$ and $\infty$.
Thirdly, defining $2\pi+2i \xi$ as $\theta$, then $\theta$ has the meaning of the spanned angle at the horizon (Figure \ref{fig:propagator}). Then the gaussian weight $e^{-2{(\xi-i\pi)^2\over u}}=e^{\theta^2\over 2u}$ can be understood from the classical action along the boundary with fixed span angle $\theta$. The boundary drawn in the figure represents a curve with fixed (regularized) proper length $u$ in $H_2$.
Lastly, the factor $e^{-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}=e^{-{4\cos\theta\over \ell}}$ is equal to $e^{q(\alpha+\beta)}$, which is a corner term that arise from JT gravity in geometry with jump angles. Here $\alpha$ and $\beta$ are defined as the angle spanned by the geodesic with fixed length and the ray coming from horizon to the boundary.
In summary the propagator can be understood as an integral of JT gravity partition functions over geometries \ref{fig:propagator} with different $\theta$s.
\begin{figure*}[h!]
\centering
\subfigure[singular ``phase" factor for different ordering]{%
\includegraphics[width=0.45\textwidth]{phase_factor.pdf}%
\label{fig:phasefactor}%
}\hspace{1.5cm
\subfigure[A geometric representation of the propagator. Here we fix the span angle $\theta$, the propagator is a summation over such geometries.]{%
\includegraphics[width=0.45\textwidth]{propagator.pdf}%
\label{fig:propagator}%
\caption{}
\label{fig:Propagator}
\end{figure*}
Finally, let us comment on the issues we raised in section 4.2. In the large $q$ limit we are considering the propagator at relatively large distances and in a regime where locally in $AdS$ the integration over paths that fluctuate
wildly is suppressed. Alternatively we can say that in the integration over paths we put a UV cutoff which
is large compared to $1/q$ but small compared to the AdS radius. This is the non-relativistic regime for the
boundary particle. The quantum effects are still important at much longer distances due to the large size
of $AdS$.
In addition, if we have quantum fields in AdS, then their partition functions for these fluctuating contours that
have fluctuations over distances larger than the $AdS$ radius are expected to depend on this shape in a local way. Due to the symmetries of $AdS_2$, this is simply expected to renormalize the action we already have without introducing extra terms. This can be checked explicitly for conformal field theories by using
the conformal anomaly to compute the effective action of the CFT$_2$ on a portion of $H_2$ (Appendix \ref{appendix: CFT effective action}).
\subsection{Wheeler-DeWitt Wavefunction }
\label{WdW}
In the pure JT theory we can think about quantizing the bulk theory and obtaining the
Wheeler-DeWitt wavefunction. This was discussed in the classical limit by Harlow and Jafferis \cite{Harlow:2018tqv}.
The Wheeler-DeWitt wavefunction can be created by Euclidean evolution of the boundary and hence is closely related to the propagator we have discussed above. The wavefunction in Lorentzian signature could then be obtained by analytic continuation of the boundary time.
The Euclidean evolution can be specified by either of the two parameters: the proper length $u$ or energy $E$. Choosing a different parameter corresponds to imposing a different boundary condition in JT theory. In general there are four possible choices of boundary conditions in 2d dilaton gravity, there are two sets of conjugate variables: $\lbrace \phi_b,K\rbrace$, and $\lbrace u,E\rbrace$ \footnote{Energy $E$ is proportional to the normal derivative of the dilaton field at the boundary.}. In preparing the wavefunction we fix the boundary value of dilaton and hence there are only two choices of the parameter ($u$ or $E$). We denote the corresponding wavefunction as $|u\rangle_G$ and $|E\rangle_G$ respectively.
In terms of holographic considerations, $|u\rangle_G$ represents a thermofield double state:
\begin{equation}
|u\rangle_G\sim\sum\limits_{n}e^{-E_n u}|E_n\rangle_L |\bar E_n\rangle_R
\end{equation}
and $|E\rangle_G $ is like an average of energy eigenstates in a window of energy $E$:
\begin{equation}\label{EnergyStateDefinition}
|E\rangle_G\sim {1\over \delta E}\sum\limits_{ |E-E_n|<\delta E}|E_n\rangle_L |\bar E_n\rangle_R.
\end{equation}
The width of the energy window is some coarse graining factor such that the summation contains $e^{S_0}$ states and does not show up clearly in gravity.\footnote{If one understand getting $|E\rangle$ state from integrating over thermofield double state in time direction, then a natural estimate on $\delta E$ is ${1\over T}$, where $T$ is the total time one integrate over. The validity of JT description of boundary theory is $T<e^{S_0}$, and we get $\delta E>e^{-S_0}$. For $T>e^{S_0}$, there are other possible instanton contributions. The proper gravitational theory at this regime is studied in paper \cite{Saad:2018bqo} }
With the definition of the states, one can evaluate them in terms of different basis. There are three natural bases turn out to be useful, we call them $S$, $\eta$ and $\ell$ bases. Basis $S$ corresponding to fix the horizon value of dilaton field $\phi_h$, or equivalently by Bekenstein-Hawking formula, the entropy of the system. The canonical conjugate variable of $S$ will be called $\eta$ and that characterizes the boost angle at the horizon. $\ell$ stands for fixing geodesic distance between two boundary points. To see that the horizon value of the dilaton field is a gauge invariant quantity, one can do canonical analysis of JT gravity. With ADM decomposition of the spacetime metric, one can get the canonical momenta and Hamiltonian constraints of the system \cite{LouisMartinez:1993eh}:
\begin{eqnarray}
ds^2&=&-N^2dt^2+\sigma^2(dx+N^xdt)^2;~~~~~~~~~~~~~~~~~~~~\\
\mathcal{H}&=&-\Pi_{\phi}\Pi_{\sigma}+\sigma^{-1}\phi''-\sigma^{-2}\sigma'\phi'-\sigma\phi;~~~~\mathcal{H}_x=\Pi_{\phi}\phi'-\sigma\Pi_{\sigma}';\label{Hamiltonian Constraint}\\
\Pi_{\phi}&=&N^{-1}(-\dot{\sigma}+(N^x\sigma)')=K\sigma;~~~~~~~~~~~~~\Pi_{\sigma}=N^{-1}(-\dot{\phi}+N^x\phi')=\partial_n\phi.\label{Canonical Momentum}
\end{eqnarray}
That is the dilaton field is canonically conjugate to the extrinsic curvature and boundary metric is canonical conjugate to the normal derivative of the dilaton field (both are pointing inwards).
By a linear combination of the Hamiltonian constraints (\ref{Hamiltonian Constraint}), one can construct the following gauge invariant quantity $C$:
\begin{equation}
-{1\over\sigma}(\phi'\mathcal{H}+\Pi_{\sigma}\mathcal{H}_x)={1\over 2}(\Pi_{\sigma}^2+\phi^2-{\phi'^2\over \sigma^2})'\equiv C[\Pi_{\sigma},\phi,\sigma]'\sim 0
\end{equation}
The Dirac quantization scheme then tells us that the quantity $C$ has a constant mode which is gauge invariant (commute with Hamiltonian constraint).
Choosing the gauge that normal derivative of the dilaton is zero, we can solve the Hamiltonian constraint:
\begin{equation}
\phi^2-(\partial_{X}\phi)^2=2C\equiv S^2~~~~~\rightarrow~~~~~~\phi(X)=S\cosh X,
\end{equation}
where $dX= \sigma dx$ is the proper distance along the spatial slice. Because the normal derivative of dilaton field is zero, the minimum value of dilaton at this spatial slice is actually a local extremum in both directions. Therefore, the minimal value of dilaton field, namely $S$, is a global variable. The classical geometry in this gauge is a ``Pac-Man" shape (right figure in figure \ref{fig:Pac-Man}).
Focusing on the intersection region of the spatial slice and the boundary, we have the spatial slice is orthogonal to the boundary.
This is because we are gauge fixing $\partial_n\phi=0$ on the spatial slice, and $\phi=\phi_b$ on the boundary.
The ADM mass of the system, after regularization, is then $M=\phi_b(\phi_b-\partial_{X}\phi)$ \cite{Maldacena:2016upp}. Substituting the behavior of $\phi(X)$ we get:
\begin{equation}
M={S^2\over 2}.
\end{equation}
This is the same relation in \ref{energy relation} and therefore we can interpret the $s$ variable as entropy of our system $S$.
For the purpose of fixing geodesic distance, it is convenient to think of doing the path integral up to a slice $L$ with zero extrinsic curvature. This picks out a particular slice (left figure in Figure \ref{fig:Pac-Man}) among the solutions obeying the Hamiltonian constraint.
The WdW wavefunction can be evaluated as an Euclidean path integral with fixed (rescaled) geodesic distance $d$ between the two boundary points:
\begin{eqnarray}\label{WdWPath}
\nonumber \Psi(u;d)=\int {\cal D}g{\cal D}\phi e^{{1\over 2}\int \phi(R+2)+\int_{L}\phi K+\phi_b\int_{Bdy}(K-1)}=\int {\cal D }f e^{\phi_b\int_{Bdy} (K-1)}\\
=\int {\cal D}\boldsymbol x e^{-m \int_{Bdy}\sqrt{g}+q\int_{Bdy} a+q\int_{L} a+(\pi-\alpha_1-\alpha_2)q}~~~~~~~~~~~~~~~~~~~~~~~
\end{eqnarray}
Here we are fixing the total length of $L$ to be $d$ and the proper length of the boundary to be $u$. $\alpha_1$ and $\alpha_2$ in the last expression denotes the jump angle at the corner coming from the singular contribution of the extrinsic curvature and should be integrated over. Without the $e^{q\int_L a}$ factor in (\ref{WdWPath}), the path integral corresponds to the propagator (\ref{HeatLim}). Remember that the phase factor is equal to $e^{-q\int_L a}$, so the wavefunction in $\ell$ basis is actually the propagator (\ref{HeatLim}), with the phase factor stripped out
\begin{equation}\label{wavefunction}
\Psi(u; \ell)\equiv \langle \ell|u\rangle_G = \frac{2}{\pi^2 \ell}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 }{\ell}) ,~~~~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}.
\end{equation}
$\ell$ is a function of the regularized geodesic distance $d$ between $\boldsymbol{x_1}$ and $\boldsymbol{x_2}$ : $\ell=e^{d\over 2}$.
The semiclassical of $\Psi(u;\ell)$ can be obtained using formula (\ref{HeatLim2}), in the exponent we get saddle point result:
\begin{equation}
\Psi(u;\ell)\sim \exp[-{2(\xi_*-i\pi)^2\over u}+{4\over u}{\xi_*-i\pi\over \tanh \xi_*}];~~~~~~~~{\xi_*-i\pi\over u}=-{\sinh \xi_*\over \ell}.
\end{equation}
The same saddle point equation and classical action was obtained in \cite{Harlow:2018tqv} by a direct evaluation in JT gravity.
The wavefunction with fixed energy boundary condition can obtained by multiplying $\Psi(u;\ell)$ by
$e^{ + E u} $ and integrating over $u$ along the imaginary axis. This sets $E = {s^2 \over 2} $
in the above integral over $s$. So this wavefunction has a very simple expression:
\begin{equation}\label{fixed energy wavefunction}
\Psi(E;\ell)\equiv \langle \ell|E\rangle_G = \rho(E)\frac{4}{ \ell} K_{i \sqrt{ 8 E} }(\frac{4 }{\ell}).~~~~~~~~
\end{equation}
The classical geometry for $\Psi(E;\ell)$ is the same as the left figure in Figure \ref{fig:Pac-Man}, with fixing energy on the boundary.
We want to stress that it is important to have the $\rho(E)$ factor in (\ref{fixed energy wavefunction}) for a classical geometry description since we are averaging over the states.
We can roughly think of ${4\over \ell}K_{i\sqrt{8E}}({4\over \ell})$ as a gravitational ``microstate" $|\mathcal{E}\rangle$ with fixed energy $E$. Such a ``microstate" will not have a classical geometry representation and therefore is just a formal definition.
The inner product between wavefunctions is defined as $\langle\Psi_1|\Psi_2\rangle=\int_0^{\infty}d\ell \ell \Psi_1^*(\ell)\Psi_2(\ell)$.
Going to the entropy basis $S$, it is easy to start with $\Psi(E)$.
Because of the identity $E={S^2\over 2}$, expanding $\Psi(E)$ in the $S$ basis is diagonal:
\begin{equation}
\Psi(E;S)\equiv \langle S|E\rangle_G=\sqrt{\rho(S)}\delta(E-{S^2\over 2})
\end{equation}
We put this square root of $\rho(S)$ factor in the definition of $S$ basis such that inner product between different $S$ state is a delta function $\langle S|S'\rangle=\delta(S-S')$. This factor is also required such that the classical limit matches with gravity calculation.
Integrating over energy with Boltzman distribution, we can get the expression of thermofield double state in the $S$ basis:
\begin{equation}
\Psi(u;S)\equiv \langle S|u\rangle_G=\int dE e^{-u E}\langle S|E\rangle_G=\sqrt{\rho(S)}e^{-{uS^2\over 2}}
\end{equation}
In the semiclassical limit, the wavefunction becomes gaussian and coincides with the on shell evaluation of the ``Pac-Man" geometry (Figure \ref{fig:Pac-Man}):
\begin{equation}\label{PsiS}
\Psi(u,S)\sim \sqrt{S}e^{\pi S -{u S^2\over 2}}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.5]{classicalgeometry.pdf}
\caption{Classical Geometry in $\ell$ and $S$ basis}
\label{fig:Pac-Man}
\end{figure}
The on shell calculation is straightforward: JT action in this geometry contains two parts: the Schwarzian action $\int(K-1)$ on the boundary and a corner contribution at the center: $S(\pi-\theta)$, where $\theta$ is the span angle at the horizon (Figure \ref{fig:Pac-Man}). The Schwarzian action simply gives $E u={S^2 u\over 2}$ by direct evaluation. We can determine $\theta$ from $u$ since they are related with redshift: $\theta=u S$. Therefore the corner term gives: $\pi S-S^2 u$. Adding them up then gives us the classical action.
We can also expand $S$ in terms of the $\ell$ basis, and relate $\Psi(\ell)$ with $\Psi(S)$ by a change of basis:
\begin{equation}
\langle\ell|S\rangle=\sqrt{\rho(S)}{4\over \ell}K_{2iS}({4\over \ell});~~~~~\Psi(E;\ell)=\int_0^{\infty} dS \langle \ell|S\rangle \langle S|E\rangle_G;
\end{equation}
Before discussing our last basis, we want to stress the simplicity of the wavefunction in $S$ basis (\ref{PsiS}) and the Gaussian factor resembles an ordinary particle wavefunction in momentum basis. We introduce our last basis $\eta$ as canonical conjugate variable of $S$, with an analog of going to position space of the particle picture in mind:
\begin{eqnarray}
|\eta\rangle &=&\int_0^{\infty} dS \cos(\eta S)|S\rangle;~~~~~~~~~~~~~\langle \ell|\eta\rangle=\int_0^{\infty} dS\cos(\eta S)\sqrt{\rho(S)}K_{2iS}({4\over \ell}){4\over \ell};~~~~~\\
\Psi(E,\eta) &=&\scaleto{\sqrt{\sinh(2\pi \sqrt{2E})\over 2\pi^2\sqrt{2E}}\cos(\eta \sqrt{2E})}{30pt};~~~~~\Psi(u,\eta)=\int_0^{\infty} dS \sqrt{\rho(S)}\cos(\eta S)e^{-{uS^2\over 2}}.~~~~
\end{eqnarray}
To understand the meaning of $\eta$ better, we can look at the classical behavior of $\Psi(u;\eta)$:
\begin{equation}
\Psi(u;\eta)\sim {1\over u}\exp[{\pi^2\over 2 u}-{\eta^2\over 2 u}]\left(e^{i\eta\pi\over u}\sqrt{\pi+i\eta}+e^{-{i\eta\pi\over u}}\sqrt{\pi-i\eta}\right)
\end{equation}
When $u$ is real, the wavefunction is concentrated at $\eta=0$ and has classical action of a half disk in the exponent.
When $u={\beta\over 2}+it$ which corresponds to the case of analytically continuing into Lorentzian signature, the density of the wavefunction $|\Psi(u,\eta)|^2$ is dominated by:
\begin{equation}
|\Psi({\beta\over 2}+it,\eta)|^2\sim {\sqrt{\pi^2+\eta^2}\over \beta^2+4t^2}\exp[{2\pi^2\over \beta}]\left(\exp[-{2\beta(\eta-{2\pi t\over \beta})^2\over \beta^2+4t^2}]+\exp[-{2\beta(\eta+{2\pi t\over \beta})^2\over \beta^2+4t^2}]\right)
\end{equation}
showing the fact that $\eta$ is peaked at the Rindler time ${2\pi t\over \beta}$. We can therefore think of fixing $\eta$ as fixing the IR time or the boost angle at the horizon.
The classical intuition for the boost angle is most clear in Euclidean geometry, where for fixed boundary proper time there can be different cusps at the horizon (Figure \ref{fig:Boostangle}).
\begin{figure}
\centering
\includegraphics[scale=0.4]{boostangle.pdf}
\caption{Euclidean geometries with different cusps.}
\label{fig:Boostangle}
\end{figure}
One application of those wavefunctions is that we can take an inner product and get the partition function.
However, there are also other ways to get the partition function.
For example, we can concatenate three propagators and integrate over their locations. This also gives the partition function by the composition rule of propagator.
By the relation between propagator and wavefunction, we can also view this as taking an inner product of three wavefunctions with an interior state as in figure \ref{fig:threewavefunction}, where the interior state can be understood as an entangled state for three universes.
To be more precise, we can view the wavefunction as the result of integrating the bulk up to the geodesics with zero extrinsic curvature. Then the interior state is given
by the area of the hyperbolic triangle in figure \ref{fig:threewavefunction}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{threewavefunction.pdf}
\caption{Partition function from inner product of three wavefunctions.}
\label{fig:threewavefunction}
\end{figure}
The path integral for the hyperbolic triangle (denoted as $I(\ell_{12},\ell_{23},\ell_{31})$, where $\ell_{ij}={|x_i-x_j|\over \sqrt{z_iz_j}}$), is a product of three phase factors, which satisfies a nontrivial equality (with ordering $x_1>x_2>x_3$):
\begin{equation}\label{GHZ}
\scaleto{ I(\ell_{12},\ell_{23},\ell_{31})=e^{-2({z_1+z_2\over x_1-x_2}+{z_2+z_3\over x_2-x_3}+{z_3+z_1\over x_3-x_1})}={16\over\pi^2}\int_0^{\infty}d\tau \tau \sinh (2\pi \tau) K_{2i\tau}({4\over \ell_{12}})K_{2i\tau}({4\over \ell_{23}})K_{2i\tau}({4\over \ell_{31}}).}{26pt}
\end{equation}
Recalling that the Bessel function represents the fixed energy ``microstate" $|\mathcal{E}\rangle$ (\ref{fixed energy wavefunction}) and ${\tau\over 2\pi^2} \sinh(2\pi \tau)$ is the density of state, this formula tells us that the interior state is a GHZ state for three universe:
\begin{equation}
I_{123}\sim\sum_n|\mathcal{E}_n\rangle_1|\mathcal{E}_n\rangle_2|\mathcal{E}_n\rangle_3.
\end{equation}
$I$ can also been viewed as a scattering amplitude from two universes into one universe. It constrains the SL(2,R) representation of the three wavefunctions to be the same.\footnote{Some thing similar happens for 2d Yang-Mills theory \cite{Witten:1992xu,Cordes:1994fc,Luca}.}
We can write down the partition function as:
\begin{equation}
Z_{JT}=\int_0^{\infty} \prod\limits_{\lbrace i j \rbrace \in \lbrace 12,23,31\rbrace}d\ell_{ij} \Psi(u_{12},\ell_{12})\Psi(u_{23},\ell_{23})\Psi(u_{31},\ell_{31})I_{\ell_{12},\ell_{23},\ell_{31}}.
\end{equation}
This same result also holds if we repeat the process $n$ times. It is interesting that we can view the full disk amplitude in these various ways.
One can also extend our analysis to include matter field. One type of such wavefunction can be created by inserting operator during Euclidean evolution, and is analysed in appendix B. Note that because of the SL(2,R) symmetry is a gauge symmetry, our final state has to be a gauge singlet including matter field.
\section{Correlation Functions in Quantum Gravity}
\subsection{Gravitational Feynman Diagram}
The propagator enables us to ``dress'' quantum field theory correlators to produce quantum gravity ones.
Namely, we imagine that we have some quantum field theory in $H_2$ and we compute correlation functions of
operators as we take the points close to the boundary where they take the form
\begin{equation}
\langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} = q^{ - \sum \Delta_i }z_1^{\Delta_1}..z_n^{\Delta_n}\langle O_1(x_1)...O_n(x_n)\rangle_{CFT}
\end{equation}
The factor of $q$ arises from \nref{yAndq}, and the last factor is simply defined as the function
that results after extracting the $z$ dependence. For example, for a two point function we get
\begin{equation}
\langle O_1(\boldsymbol{x_1})O_2(\boldsymbol{x_2})\rangle_{\rm QFT}=
q^{ - 2\Delta } z_1^{\Delta} z_2^{\Delta}{1\over |x_1-x_2|^{2\Delta}}.
\end{equation}
We can now use the propagator \nref{HeatLim} to couple the motion of the boundary and thus obtain the full
quantum gravity expression for the correlator. The factors of $q$ are absorbed as part of the renormalization
procedure for defining the full quantum gravity correlators.
In this way we obtain
\begin{figure*}[h!]
\centering
\subfigure[Witten Diagram]{%
\includegraphics[width=0.3\textwidth]{wittendiagram.pdf}%
\label{fig:Witten Diagram}%
}\hspace{4cm
\subfigure[Gravitational Feynman Diagram]{%
\includegraphics[width=0.3\textwidth]{dynamicalwittendiagram.pdf}%
\label{fig:Dynamical Witten Diagram}%
\caption{Summation of ${1\over N}$ effects fluctuates the boundary of Witten Diagram}
\label{fig:witten diagram}
\end{figure*}
\begin{eqnarray}\label{Higher Point Function}
\small{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{\rm QG}}}&=&e^{2\pi q}\int {\prod_{i=1}^n{dx_i dy_i\over y_i^2}\over \text{V(SL(2,R))}} G(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})G(u_{23},\boldsymbol{x_2},\boldsymbol{x_3})...G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})\times
\cr
& ~& \times \langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} q^{ \sum_ i \Delta_i }
\end{eqnarray}
where the left hand side is the full quantum gravity correlator by definition. The last factor is the usual
renormalization necessary to get something finite.
The factor of $e^{ 2 \pi q } $ cancels with the $q$ dependent ``phase'' factors in \nref{heat kernel} to give one if we order
the points cyclically ($x_1>x_2...>x_n$). This requires that we define more carefully the last propagator
$G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})$ as:
\begin{equation}
e^{-2\pi q}\tilde K(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})=e^{-2\pi q}e^{-2 \frac{z_n+z_1}{x_n-x_1}}\frac{2\sqrt{z_nz_1}}{\pi^2 |x_n-x_1|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2} u_{n1}}K_{2is}(\frac{4\sqrt{z_nz_1}}{|x_n-x_1|}).
\end{equation}
The factor ${1\over \text{V(SL(2,R))}}$ in \nref{Higher Point Function} means that we should fix the $SL(2,R)$ gauge symmetry (Appendix \ref{gauge fix}).
In the end we can write down an expression where we have already taken the $q \to \infty $ limit
\begin{equation}\label{final formula}
\boxed{\scaleto{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{QG}}=\int_{ x_1>x_2..>x_n}{\prod_{i=1}^n{dx_i dz_i}\over \text{V(SL(2,R))}} \tilde{K}(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})...\tilde{K}(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})z_1^{\Delta_1-2}..z_n^{\Delta_n-2}\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}.}{26pt}}
\end{equation}
This is one of the main results of our paper and it gives a detailed expression for correlation function in 2 dimensional quantum gravity in terms of the correlation functions of the QFT in hyperbolic space, or $AdS_2$.
Notice that in usual $AdS/CFT$ the correlators $\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}$ are an approximation to the full answer. This is sometimes computed by Witten diagrams.
We get a better approximation by integrating over the metric fluctuations.
In this case, the non-trivial gravitational mode is captured by the boundary propagator. The formula \nref{final formula} includes all the effects of quantum gravity in the JT theory (in the Schwarzian limit).
The final diagrams consist of the Witten diagrams for the field theory in $AdS$ plus the propagators for the boundary particle and we can call them ``Gravitational Feynman Diagrams", see figure \ref{fig:witten diagram}.
\subsection{Two Point Function}
Using formula \nref{final formula}, we can study gravitational effects on bulk fields such as its two point function:
\begin{eqnarray}
\langle O_1(u)O_2(0)\rangle_{QG}=
\includegraphics[width=50mm,trim=0 9cm 0 0]{2ptfunction.pdf}\\
~~~~~~~~~~~\nonumber\\
~~~~~~~~~\nonumber
\end{eqnarray}
The explicit expression for $\langle O_1(u)O_2(0)\rangle_{QG}$ with dimension $\Delta$ at temperature ${1\over \beta}$ is
\footnote{We will just keep the $\Delta$ dependent constant
since at last we will normalized with respect of partition function which corresponds to set $\Delta=0$.}:
\begin{equation}
\scaleto{{1\over \text{V(SL(2,R))}}\int_{x_1>x_2} {dx_1 dx_2 dz_1 dz_2 \over z_1^2 z_2^2}\int_0^{\infty}ds_1 ds_2 \rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}K_{2is_1}({4\sqrt{z_1z_2}\over |x_1-x_2|})K_{2is_2}({4\sqrt{z_1z_2}\over |x_1-x_2|})({\sqrt{z_1z_2}\over |x_1-x_2|})^{2\Delta+2}}{25pt}.
\end{equation}
To fix the $SL(2,R)$ gauge, we can choose $z_1=z_2=1$ and $x_2=0$.
Then the integral over $H_2$ space is reduced to a single integral over $x_1$,
with a Jacobian factor $2x_1$ (Appendix \ref{gauge fix}):
\begin{equation}\label{space integral}
\int_0^{\infty}ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\int_0^{\infty} dx_1 ({1\over x_1})^{2\Delta+1}K_{2is_1}({4\over x_1})K_{2is_2}({4\over x_1}).
\end{equation}
the last integral can be interpreted as a matrix element of two point operator $O_1O_2$ between states $|E_1,\psi\rangle$ and $|E_2,\psi\rangle$,
where $|\psi\rangle$ is the wavefunction of quantum field theory
and $|E\rangle_G$ represents the fixed energy gravitational state.
Integrating over $x$ can be thought as integrating over a particular gravitational basis, and we can see that the gravity wavefunction
suppress the UV contributions from quantum field theory ($K_{2is}({4\over x})\sim \sqrt{\pi x\over 8}e^{-4/x}$ for $x\sim 0$).
The final expression for the two point function is:
\begin{eqnarray}\label{2point function}
\langle O_1(u) O_2(0)\rangle_{QG}={1\over \mathcal{N}}\int ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\frac{|\Gamma(\Delta-i(s_1+s_2))\Gamma(\Delta+i(s_1-s_2))|^2}{2^{2\Delta+1} \Gamma(2\Delta)};~~\\
={1\over \mathcal{N}}{\Gamma(2\Delta)\over u^{3/2} (\beta-u)^{3/2} 2^{4\Delta+4}\pi^3} \int_{c-i\infty}^{c+i\infty} d\theta_1d\theta_2 \theta_1\theta_2 e^{{\theta_1^2\over 2u}+{\theta_2^2\over 2 (\beta-u)}}{1\over (\cos{\theta_1\over 2}+\cos{\theta_2\over 2})^{2\Delta}}.~~~~~
\end{eqnarray}
In the second expression we write the integral in terms of variable $\theta$ using the second integral representation of the propagator (\ref{HeatLim2}).
The normalization constant can be determined by taking the $\Delta=0$ limit: $\mathcal{N}=Z_{JT}$.
If we contemplate the result (\ref{2point function}) a little bit, then we find that the two integrals of $s_1$ and $s_2$ just represent the spectral decomposition of the two point function. Indeed, under spectral decomposition we have $\langle O(u)O(0)\rangle=\sum\limits_{n,m} e^{-E_n u-E_m (\beta-u)}|\langle E_n|O|E_m\rangle|^2$. Compare with (\ref{2point function}), we can read out the square of matrix element of operator $O$:
\begin{equation}
\scaleto{_G\langle E_1|O_L O_R|E_2\rangle_G=\delta E^{-2}\sum\limits_{\substack{|E_n-E_1|<\delta E\\|E_m-E_2|<\delta E}}|\langle E_n|O|E_m\rangle|^2=\rho(E_1)\rho(E_2){|\Gamma(\Delta-i(\sqrt{2E_1}+\sqrt{2E_2}))\Gamma(\Delta+i(\sqrt{2E_1}-\sqrt{2E_2}))|^2\over 2^{2\Delta+1}\Gamma(2\Delta)}.}{38pt}
\end{equation}
Remember the notation is that $|E\rangle_G$ stands for a gravitational state with energy $E$ and $|E_n\rangle$ stands for one side microstate (\ref{EnergyStateDefinition}).
We have put the measure $\rho(E)={1\over 2\pi^2}\sinh(2\pi\sqrt{2 E})$ in the definition of matrix element for the reason that in gravity it is more natural to consider an average of energy states as a bulk state.
To understand this formula a little bit better, we can consider the classical limit, namely large $E$. In this limit the matrix element squared can be approximated as a nonanalytic function:
\begin{equation}
_G\langle E_1|O_L O_R|E_2\rangle_G\propto |E_1-E_2|^{2\Delta-1}e^{2\pi \text{min}(\sqrt{2E_1},\sqrt{2E_2})}.
\end{equation}
If we fix $E_1$ and varying $E_2$ from $0$ to infinity, the matrix element changes from $|E_1-E_2|^{2\Delta-1}\rho(E_2)$ to $|E_1-E_2|^{2\Delta-1}\rho(E_1)$ after $E_2$ cross $E_1$. We can understand this behavior qualitatively as a statistical effect: the mapping from energy subspace $E_1$ to $E_2$ by operator $O$ is surjective when the Hilbert space dimension of $E_2$ is less that $E_1$ and is injective otherwise.
Another understanding is the following: the two point function is finite in a fixed energy state $|E_n\rangle$, which means the following summation of intermediate states $|E_m\rangle$ is order one: $\sum\limits_{m}|\langle E_n|O|E_m\rangle|^2$. Looking at the case $E_m>E_n$, because of the density of states grows rapidly, the matrix element squared has to be proportional to ${1\over \rho(E_m)}$ to get a finite result. Multiplied by $\rho(E_n)\rho(E_m)$, we have $\rho(E_n)\rho(E_m)|\langle E_n|O|E_m\rangle|^2\sim \rho(\text{min}(E_n,E_m))$.
\subsection{ETH and the KMS condition}
The Eigenstate Thermalization Hypothesis (ETH) is a general expectation for chaotic system. It expresses that the operator expectation value in an energy eigenstate can be approximated by thermal expectation value with effective temperature determined from the energy. Such hypothesis can be tested with the knowledge of operator matrix elements. The two point function in microcanonical essemble is:
\begin{eqnarray}
{1\over \delta E}\sum\limits_{|E_n-E|<\delta E}\langle E_n|O(u) O(0)|E_n\rangle &=&
\rho(E)e^{uE}\langle E|O e^{-u H}O|E\rangle\nonumber\\
\langle E|O e^{-u H}O|E\rangle &=&\int_0^{\infty} ds \rho(s)e^{-{s^2\over 2}u}\scaleto{{|\Gamma(\Delta+i(s+\sqrt{2E}))|^2|\Gamma(\Delta+i(s-\sqrt{2E}))|^2\over 2^{2\Delta+1} \Gamma(2\Delta)}}{25pt}.~~~~~~~~~\label{TwoPFMicro}
\end{eqnarray}
Notice that $|E\rangle$ is not $|E\rangle_G$, the former represents a one side microstate, while the later is a gravitational state. Accordingly $ \langle E|O(u)O(0)|E\rangle$ stands for a two point function in a microstate.
To study ETH, we will consider the case of a heavy black hole $E={S^2\over 2}={2\pi^2 \over\beta^2}\gg 1$. From the discussion in last section, we know that the matrix element tries to concentrate $s$ around $\sqrt{2E}$ and thus we can approximate $\rho(E)\rho(s)|\Gamma(\Delta+i(s+S))|^2$ as proportional to $sS^{2\Delta-1}e^{\pi(s+S)}$. Using integral representation for $|\Gamma(\Delta+i(s-S))|^2 $ we derive the two point function in microcanonical essemble with energy $E$ is proportional to:
\begin{equation}
{\rho(E)S^{2\Delta-1}\over u^{3/2}}\int d\xi (\pi+i\xi)e^{-{2\over u}(\xi-i{\pi\over 2})^2-(\pi+2i\xi) S+u{S^2\over 2}}{1\over(\cosh \xi)^{2\Delta}}.
\end{equation}
The $\xi$ variable can be understood as the measure of time separation in units of effective temperature between two operators and its fluctuation represents the fluctuation of the effective temperature.
And the final integral can be understood as a statistical average of correlation functions with different temperatures.
If we put back the Newton Constant $G_N={1\over N}$, we have $S\sim N$ and $u\sim N^{-1}$ (\ref{BCond}).
As can be seen from the probability distribution, the fluctuation is of order ${1\over \sqrt{N}}$, and hence for large $N$ system we can use saddle point approximation:
\begin{equation}
\xi=i({\pi\over 2}-{S\over 2}u) -{u\Delta \tanh \xi\over 2}=i({\pi\over 2}-{\pi u\over \beta})-{u\Delta\tanh \xi\over 2}.
\end{equation}
The first piece gives the typical temperature of the external state,
while the last piece comes from the backreaction of operator on the geometry. If we first take the limit of large N, one simply get that the two point function in microcanonical essemble is the same as canonical essemble.
However, the euclidean correlator in canonical essemble is divergent as euclidean time approach to inverse temperature $\beta$ because of KMS condition.
Such singular behavior plays no role in the microcanonical essemble so is called a ``forbidden singularity" in ETH \cite{Fitzpatrick:2016ive, Faulkner:2017hll}. In our situation we can see directly how the forbidden singularity disappears in the microcanonical essemble.
When $\xi$ approach $-i{\pi\over 2}$ at the forbidden singularity the backreaction on the geometry becomes large and hence the effective temperature becomes lower:
\begin{equation}
{2\pi\over\beta*}\rightarrow {2\pi\over\beta}-{\Delta \over \pi-{\pi u\over \beta}}.
\end{equation}
At the time ${\beta-u\over\beta}\sim{\Delta\over N}$, the backreaction is important and we expect to see deviation from thermal correlators.
Therefore the correlation function in microcanonical essemble will never have singularity away from coincide point.
\subsection{Three Point Function}
The bulk diagram of the three point function will be like Figure \ref{fig:threewavefunction} with additional operator inserting at the intersection points (See Figure \ref{fig:threepointfunction}).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{threepointfunction.pdf}
\caption{Bulk Diagram for Three Point Function}
\label{fig:threepointfunction}
\end{figure}
The QFT three point correlation function in $AdS_2$ is fixed by conformal symmetry and we can write it down as:
\begin{equation}
\langle O_1 O_2 O_3\rangle=C_{123}{z_1^{\Delta_1} z_2^{\Delta_2} z_3^{\Delta_3}\over |x_{12}|^{\Delta_1+\Delta_2-\Delta_3}|x_{23}|^{\Delta_2+\Delta_3-\Delta_1}|x_{13}|^{\Delta_1+\Delta_3-\Delta_2}}={C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
$\Delta_i$ is the conformal dimension of $O_i$. Putting them in formula \ref{final formula} and rewrite the propagator in terms of the wavefunction (\ref{wavefunction}), we have the quantum gravitational three point function:
\begin{equation}
\langle O_1 O_2 O_3\rangle_{QG}=\int_{x_1>x_2>x_3}{\prod_{i=1}^3{dx_i dz_i}\over \text{V(SL(2,R))}}\Psi_{u_{12},\ell_{12}}\Psi_{u_{23},\ell_{23}}\Psi_{u_{13},\ell_{13}}I_{\ell_{12},\ell_{23},\ell_{13}}{C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
We can view this expression as an inner product of three universe wavefunction with the interior, inserting three bilocal operators $\tilde O_{ij;k}\tilde O_{ij;k}$ with dimension $\tilde\Delta_{ij;k}={1\over 2}(\Delta_i+\Delta_j-\Delta_k)$ between them. One can fix the $SL(2,R)$ symmetry and express the integral in terms of $\ell_{ij}$, it is the same exercise as in open string calculation to find the Jacobian factor (Appendix \ref{gauge fix}). Here we can just argue that in order to get the partition function at $\Delta=0$, the measure has to be flat. Therefore the three point function factorizes into form:
\begin{equation}
\small{\langle O_1(u_{1})O_2(u_{2})O_3(u_{3})\rangle_{\text{\rm QG}}}\propto C_{123}\int_0^{\infty}d\tau \rho(\tau)\mathcal{I}_{\tau}(u_{12},\tilde\Delta_{12;3})\mathcal{I}_{\tau}(u_{23},\tilde\Delta_{23;1})\mathcal{I}_{\tau}(u_{31},\tilde\Delta_{31;2})
\end{equation}
while $\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})$ is an integral of $\ell_{ij}$ which gives the two point function in microstate $E_{\tau}$ (\ref{TwoPFMicro}) with the $e^{uE_{\tau}}$ factor stripped off:
\begin{equation}
\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})={1\over 2}\int_{0}^{\infty} d\ell_{ij}\Psi_{u_{ij},\ell_{ij}}{1\over \ell_{ij}^{\Delta_i+\Delta_j-\Delta_k}}K_{2i\tau}({4\over \ell_{ij}})=\langle E_{\tau} |\tilde O_{ij;k}e^{-u_{ij}H}\tilde O_{ij;k}|E_{\tau}\rangle;~~~~~E_{\tau}={\tau^2\over 2}.
\end{equation}
Again the normalization constant can be fixed by choosing $O_i$ to be identity.
\subsection{Einstein-Rosen Bridge}
The Einstein-Rosen Bridge in a classical wormhole keeps growing linearly with time and this behavior was conjectured to related with the growth of computational complexity of the dual quantum state \cite{Stanford:2014jda}.
Based on the universal behavior of complexity growth, Susskind proposed a gravitational conjecture in a recent paper \cite{Susskind:2018fmx} about the limitation of classical general relativity description of black hole interior. The conjecture was stated as follows:
\textsl{Classical general relativity governs the behavior of an ERB for as long as possible.}
In this section, we will test this conjecture using the exact quantum wavefunction of JT gravity (\ref{wavefunction}). We will in particular focusing on the behavior of ERB at time bigger than $1$. The size of Einstein-Rosen Bridge $\mathcal{V}$ in two dimensions is the geodesic distance $d$ between two boundaries, and can be calculated in thermofield double state $|u\rangle$ as:
\begin{equation}
\mathcal{V}=\langle u|d|u\rangle
\end{equation}
We want to focus on the dependence of volume on Lorentzian time evolution. Therefore we do analytic continuation of $u$ in Lorentzian time: $u={\beta\over 2}+it$. Using the WdW wavefunction in $\ell$ basis (\ref{wavefunction}) and the relation between $d$ and $\ell$, we can calculate the expectation value exactly. This can be done by taking the derivative of the two point function (\ref{2point function}) with respect to $\Delta$ at $\Delta=0$. Using the integral representation for $|\Gamma(\Delta+i(s_1-s_2))|^2$, the only time dependence of volume is given by:
\begin{equation}
\mathcal{V}(t)={1\over \mathcal{N}}\int_{-\infty}^{\infty}d\xi\int_0^{\infty} ds_1ds_2\rho_1\rho_2 e^{i(s_1-s_2)\xi-i(s_1-s_2){(s_1+s_2) \over 2}t-{\beta\over 4}(s_1^2+s_2^2)}\log(2\cosh{\xi\over 2})|\Gamma(is_1+is_2)|^2.
\end{equation}
The limit we are interested in is $\beta\ll 1\ll t$ \footnote{Remember that we are measuring time in units of $\phi_r$ (\ref{BCond}), so time order $1$ is a quantum gravity region.}, in which case the integral has a saddle point at \footnote{Actually this saddle point is valid for any range of $t$ as long as $\beta\ll 1$.}:
\begin{equation}\label{SaddlePoint}
s_{1,2}={2\pi\over \beta};~~~~~~~~~\xi={2\pi t\over \beta}.
\end{equation}
Therefore the volume has linear dependence in time:
\begin{equation}
\mathcal{V}(t)\sim{2\pi t\over \beta}.
\end{equation}
Using the complexity equal to volume conjecture \cite{Stanford:2014jda,Brown:2015bva}, the complexity of thermofield double state is proportional to the maximum volume:
\begin{equation}
\mathcal{C}(t)=\#\mathcal{V}(t)=\#{2\pi t \over\beta} .
\end{equation}
The proportionality constant is suggested in \cite{Brown:2018kvn} to be $S_0$ based on classical calculation of near extremal black hole. This, however, is not very clear in our model since $S_0$ is the coupling constant of the pure Einstein-Hilbert action and decouples with JT theory (\ref{full action}).
Since the saddle point (\ref{SaddlePoint}) is actually valid from early time to late time, the proportionality constant can be fixed at classical level and once we fix it we can conclude that \textsf{the length of Einstein-Rosen Bridge (or complexity of the state) keeps linearly growing even considering quantum gravity effects in JT theory.}
We want to comment that this is not an obvious result that one can expect from classical observables.
For example, one might argue that we can extract the information of the ERB from two sided correlators for the reason that semiclassically we can approximate the correlator as $e^{-m d}$.
Therefore one can conclude the ERB has linear growth from the quasinormal behavior of the correlator.
However such observables can only give us information of ERB up to time order $1$, which is the same time scale we can trust the classical general relativity calculation.
After that the correlation function changes from exponential decay into universal power law decay ${1\over t^{3}}$ as one can directly derive from analytic continuation of result (\ref{2point function}).
If we still use such correlator to extract information about ERB we would get the wrong conclusion that it stops its linear growth after time order $1$.
The reason why it is incorrect is that at this time scale the operator disturbs the state and causes different energy states interfere each other strongly. It is simply that the correlator can no longer be described by the classical geometry, rather than the interior stops to behave classically.
From our calculation, we see that if we probe of the state in a weaker and weaker way, we are still able to see the classical geometry.
Lastly, we want to talk a little about when JT gravity needs to be modified. A naive estimate can be made from the partition function that when $\beta$ approaches $e^{{2\over 3}S_0}$, the partition function becomes less than one and definitely at this time scale we need new physics. A recently study of gravitational physics at this time scale was discussed in \cite{Saad:2018bqo}.
\section{Conclusion}
Our result gives an explicit formula (\ref{final formula}) to calculate all order corrections to correlation functions from quantum gravity in two dimensions.
The formula can be understood diagrammatically and we call it Gravitational Feynman Diagram.
We also give the exact Wheeler-DeWitt wavefunction and discuss the growth of its complexity quantum mechanically.
Although we are focusing on theoretical description of two dimensional black holes, the near-extremal black holes in nature should contain these features.
Both Reissner–Nordström black holes and Kerr black holes have an $AdS_2$ throat near their extremality.
For those black holes, the gravitational effects are enhanced by the their near extremal entropies (the coupling constant is $\phi_h$ rather than $\phi_0+\phi_h$) and therefore are better backgrounds to test gravitational effects.
We should however point out that the observational black holes all have large near extremal entropies and thus are very classical \cite{Preskill:1991tb}.
In addition, the Thorne limit of Kerr black hole sets a lower bound on the near extremal entropy in nature.
But for the Primordial black holes in early universe, our story might play a role and it will be interesting to study the physical consequence in that situation.
Another application of our result is to connect with SYK type models \cite{Maldacena:2016hyu,Kitaev:2017awl,Witten:2016iux,Klebanov:2016xxf} since those models all have an emergent Schwarzian action at low energy. On that account, the exact Schwarzian quantization can be used to test ${1\over N}$ corrections to those models. For example, one can try to directly test the two point function with SYK numerics \cite{BKobrin} or one can use SYK models to understand the microscopic description of WdW wavefunction and its complexity. It might also be interesting to consider the black hole information paradox \cite{Fiola:1994ir} and late time traversable wormholes including the quantum fluctuations of the boundary \cite{Gao:2016bin,Maldacena:2017axo,Susskind:2017nto,Yoshida:2017non, Maldacena:2018lmt}.
{\bf Acknowledgments }
The author wants to give special thanks to J.Maldacena and D.Stanford for patient guidance and help all through the project.
We also want to thank A.Almheiri, L.Iliesiu, J.Jiang, J.Ripley, A.Kitaev, B.Kobrin, R.Mahajan, A.M.Polyakov, S.J.Suh, G.Turiaci and H.Verlinde for discussions.
Z.Y is supported by Charlotte Elizabeth Procter Fellowship from Princeton University.
\section{Introduction}
Near-Extremal black holes have a universal structure near their horizons:
there is an $AdS_2$ throat with a slowly varying internal space.
Its low energy gravitational dynamics is captured universally by the following effective action in two dimensions \cite{Strominger:1994tn}:
\begin{equation}\label{full action}
I=\underbrace{-{\phi_0\over 2
}\left(\int R+2\int_{\partial_M}K\right)}_{\text{Einstein-Hilbert Action}}\underbrace{-{1\over 2
}\left(\int_M \phi(R+2)+2\int_{\partial M}\phi_b K\right)}_{\text{Jackiw-Teitelboim action}}+S_{matter}(g,\psi),
\end{equation}
where the dilaton field $\phi+\phi_0$ represents the size of internal space. We have separated the size of internal space into two parts: $\phi_0$ is its value at extremality. It sets the value of the extremal entropy
which comes from the first term in \nref{full action}.
$\phi$ is the deviaton from this value.
We have also added matter that only couples to the metric. This is a reasonable assumption when matter comes
from Kaluza Klein reduction, where the coupling to the dilaton would involve $\phi/\phi_0 \ll 1$.
The action $\int \phi(R+2)+2\int\phi_b K$ is the so-called Jackiw-Teitelboim action \cite{JACKIW1985343,TEITELBOIM198341}, and will be the main focus of our paper.
This action is one of the simplest nontrivial gravitational actions in two dimensions.\footnote{Another nontrivial action is the CGHS model which could be written as $\int (\phi R+C)$, and that characterizes the horizon structure of general black holes.}. It is simple because the bulk geometry is a rigid $AdS_2$ space fixed by the equation of motion of the dilaton field. Its nontriviality arises from the remaining boundary action. Schematically, the gravitational action is reduced to the following form:
\begin{equation}\label{schematic action}
I=-{ 2 \pi \phi_0
}\chi(M)-{\phi_b
}\int_{\partial M}K +S_{matter}(g,\psi).
\end{equation}
And the motion of the boundary is controlled by its extrinsic curvature.
Our goal will be to quantize this action and to provide expression for the full quantum gravity correlators
of \nref{full action}.
This problem was considered before in \cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Kitaev:2018wpr} from various points of view. Here we will add one other point of view where we
reduce the problem to the motion of a relativistic particle in an electric field, building on a suggestion in
Kitaev's talk at IAS \cite{kitaevIAS}. More precisely, one can consider a relativistic particle in a Lorentzian $AdS_2$ target space moving under the influence of an electric field. The coupling to electric field can also be viewed as a coupling to a spin connection so that it becomes a particle with spin as suggested by Kitaev. Alternatively we can start from a non-relativistic particle moving in hyperbolic space, $H_2$, under the influence of a magnetic field $b$. After analytic
continuation in $b$ to imaginary values we get the problem of interest.
Using this point of view one can think of the full quantum gravity problem as the combination of two problems.
First we consider quantum fields propagating in $AdS_2$ (or $H_2$ in the Euclidean case) and then we
add the ``gravitational particle" which couples to the quantum fields by changing their boundary location in $AdS_2$.
The discussion of quantum fields will be standard and depends on the particular model one interested in, therefore we will mainly focus on solving the second problem.
Generically, solving the gravitational problem is challenging and is not exactly equivalent to a quantum mechanical particle. One needs to worry about what functional space one will integrate over. For example, in path integrals, one usually integrates over all trajectories including those with self-intersections. However self-intersecting boundaries in gravitational system have no obvious meaning. On that account, more precisely the gravitational problem is equal to a self-avioding particle.
Nevertheless, it turns out that one can take a particular limit of this model, namely large $\phi_b$, to avoid this issue and a treatment of the boundary theory as an ordinary particle is justified. It is also true that the JT gravity can be rewritten as a Schwarzian action only in this limit. We call this the Schwarzian limit and will only focus on solving the JT action in the Schwarzian limit. Solving the model away from Schwarzian limit was considered recently by Kitaev and Suh \cite{Kitaev:2018wpr}.
Our result can be summarized as follows:
\textsl{First, we will give a formula to calculate all correlation functions with quantum gravity backreaction (formula \ref{final formula}). Second, we will give the exact Wheeler-DeWitt wavefunction in the Schwarzian limit, which has been analyzed classically by Harlow and Jafferis \cite{Harlow:2018tqv}. Last, we consider the recent proposed conjecture about complexity growth in this exact Wheeler-DeWitt wavefunction and show that the complexity maintains linear growth after taking quantum gravity effects into account. This, to our knowledge, is the first test of the gravitational conjecture made by Susskind \cite{Susskind:2018fmx} that the size of ERB grows linearly for as long as quantum mechanics allows.}
This paper is organized as follows: in the first section, we will review the classical calculation of this model and introduce notations; in the second section, we will make the dictionary between the JT model and a particle in a magnetic field; in the third and fourth sections, we will solve the quantum mechanical problem and derive the propagator and WdW wavefunction in the Schwarzian limit; in the last section, we will talk about gravitaional backreaction on correlators as well as complexity growth. As a useful notion in our calculation, we introduce a notation called gravitational Feyman diagrams.
\section{Classical Solutions}
Let us first consider the classical solutions of
the Jackiw-Teitelboim model \nref{full action}. See
\cite{Almheiri:2014cka} for further discussion.
%
%
The equation of motion of the dilaton field imposes $R=-2$
and
fixes the geometry to be $AdS_2$, or $H_2$ in the Euclidean case.
This is also true if
we have additional matter coupled with metric only, as in
\nref{full action}.
The equations for the metric constrain the dilaton
\begin{equation}\label{Einstein Equation}
(\nabla_{\mu}\nabla_{\nu}\phi-g_{\mu\nu}\nabla^2\phi+g_{\mu\nu}\phi)+T^{M}_{\mu\nu}=0
\end{equation}
These equations are compatible with each other thanks to the conservation of the matter stress tensor. They do not allow
any propagating mode. In fact, setting $T^M_{\mu\nu}=0$, and using a high momentum approximation we can write
\nref{Einstein Equation} as $ (k_{\mu}k_{\nu}-g_{\mu\nu}k^2)\phi(\vec k)=0 $, which then implies $ \phi(\vec k)=0 $, for large $k$.
More precisely, after introducing the Euclidean $AdS_2$ coordinates $ds^2 = d\rho^2 + \sinh^2\rho d\varphi^2$, we can solve
(\ref{Einstein Equation}) in $AdS_2$ with no matter. Up to an SL(2) transformation the solution is
\begin{equation}
\phi=\phi_h \cosh \rho,
\end{equation}
where $\phi_h$ is a constant that is fixed by the boundary conditions.
At the boundary we fix the metric along the boundary and the value of the dilaton field
\begin{equation} \label{BCond}
ds_{\parallel} = d u { \phi_r \over \epsilon } ~,~~~~~~~~~~~ \phi= \phi_b = { \phi_r \over \epsilon }
\end{equation}
where we think of $u$ as the time of the boundary theory. It is simply a rescaled version of proper time. Similarly $\phi_r$ is a rescaled
value of the dilaton. We will be interested in taking $\epsilon \to 0$. With these rescalings the value of $\phi_h$ in the interior
remains fixed as we take $\epsilon \to 0$ as $\phi_h = 2 \pi /\beta $ where $\beta$ is the inverse temperature, $u \sim u + \beta$. Notice that due to the factor of $\phi_r$ in the first expression in \nref{BCond} we are measuring time in units of the constant $\phi_r$, which has dimensions of length. We did this for convenience.
A nice feature that appears after taking the $\epsilon \to 0$ limit is that the action (\ref{schematic action}) can be written as the Schwarzian action for the boundary curve labeled by $\varphi(u)$ \cite{Maldacena:2016upp}:
\begin{equation}
I=-\int du Sch(\tan{\varphi(u)\over 2},u)
\end{equation}
The fluctuation of the boundary shape can be understood as the fluctuation of the dilaton distribution in the bulk.
A bit more explicitly we can say that the dilaton boundary condition fixes the location of the boundary at $\rho_b$ given by
$\phi_b = \phi_h \cosh \rho_b$, and the metric at that location relates the time $\varphi$ to $u$ by $\phi_r du = \epsilon \sinh \rho_b d\varphi $.
We get the above formulas noticing that the period of $\varphi$ is $2\pi$ while that of $u$ is $\beta$, which fixes $\epsilon \sinh \rho_b$.
\section{Charged Particle in $AdS_2$}
Despite the absence of a bulk propagating mode there is still a non-trivial dynamical gravitational degree of freedom. There are various
ways to describe it. Here we will think of it as arising from the motion of the physical boundary of $AdS_2$ inside a rigid $AdS_2$ space.
This picture is most clear for finite $\epsilon$ in \nref{BCond}, but it is true even as $\epsilon \to 0$. The dynamics of the boundary
is SL(2) invariant. This SL(2) invariance is a gauge symmetry since it simply reflects the freedom we have for cutting out a piece
of $AdS_2$ space that we will call the ``inside". It is important that the dilaton field we discussed above is produced after we put in the
boundary and it moves together with the boundary under this SL(2) gauge transformation. It is a bit like the Mach principle, the location
in $AdS_2$ is only defined after we fix the boundary (or distant ``stars").
We can make this picture of a dynamical boundary more manifest as follows.
Since the bulk Jackiw-Teitelboim action \nref{full action} is linear in $\phi$, we can integrate out the dilaton field which
sets the metric to that of $AdS_2$ and removes the bulk term in the action, leaving only the term involving the extrinsic curvature
\begin{eqnarray}
I&=&-{\phi_r\over
\epsilon}\int du \sqrt{g}K
\end{eqnarray}
This action, however, is divergent as we take $\epsilon$ to zero. This divergence is simply proportional to the length of the boundary
and can be interpreted as a contribution to the ground state energy of the system.
So we introduce a counterterm proportional to the length of the boundary to cancel it. This is just a common shift of the energies
of all states.
It is also convenient to use the
Gauss-Bonnet theorem to relate the extrinsic curvature to an integral over the bulk
\begin{equation}
\int_{\partial_M} du \sqrt{g}K=2\pi\chi(M)-{1\over 2}\int_M R
\end{equation}
Since the curvature is a constant, the bulk integral is actually proportional to the total area $A$ of our space. That is we have the regularized action:
\begin{eqnarray}\label{subtraction}
I&=&-{\phi_r\over
\epsilon}\int_{\partial M} du \sqrt{g}(K-\underbrace{1}_{\text{counterterm}})=-{\phi_r\over
\epsilon}\left(2 \pi\chi(M)-{1\over 2}\int_{M}R-\int_{\partial M} du \sqrt{g}
\right) \nonumber\\
&=&-2\pi q\chi(M)-q A+q L,~~~~~~~~~~~~~~~~~~~~~~~~~~~q\equiv{\phi_r\over
\epsilon} ,~~L={\beta \phi_r\over \epsilon}
\end{eqnarray}
We now define an external gauge field $a_\mu$ as
\begin{equation}
a_{\varphi}=\cosh\rho-1,~~~~~~~a_{\rho}=0,~~~~~~~f_{\rho \varphi}=\sinh\rho=\sqrt{g},
\end{equation}
and write the action as follows
\begin{equation}\label{regularized action}
I= - 2 \pi q + q L - q \int a
\end{equation}
where we used that $ \chi(M)$ is a topological invariant equal to one, in our case, where the topology is that of a disk.
The term $q L$ is just the length of the boundary. So this action has a form somewhat
similar to the action of a relativistic charged particle moving in
$AdS_2$ in the presence of a constant electric field.
There are a couple of important differences. First we are summing only over trajectories of fixed proper length set by the inverse temperature
$\beta$. Second, in the JT theory we are treating the $SL(2)$ symmetry as a gauge symmetry. And finally, in the JT theory we identify the proper length with the boundary time, viewing configurations which differ only by a shift in proper time as inequivalent. In fact, all these changes simplify the problem: we can actually think
of the problem as a non-relativistic particle moving on $H_2$ in an electric field. In appendix
\ref{appendix:relativistic particle} we discuss in more detail the connection to the relativistic particle.
In fact, precisely the problem we are interested in has been discussed by Polyakov in \cite{Polyakov:1987ez}, Chapter 9, as an an intermediate step for the sum over paths.
Now we would also like to point out that we can directly get to the final formula by
using the discussion there, where he explicitly shows
that for a particle in flat space the sum over paths of fixed proper length that stretch between
two points $\vec x$ and $\vec x'$ gives
\begin{equation} \label{PolFlat}
\int {\cal D }\vec x e^{ - m_0 \tilde\tau } \delta( \dot { \vec x}^2 - 1 ) =
e^{ - {1\over 2} \mu^2 \tau} \langle x' | e^{ - \tau H } |x\rangle =
e^{ - {1\over 2} \mu^2 \tau} \int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot \vec x^2 }
\right)
\end{equation}
$\mu^2$ is the regularized mass and $\tilde \tau$ is related to $\tau$ by a multiplicative renormalization.
The JT model consists precisely of a functional integral of this form, where we fix the proper length along the boundary. There are two simple modifications, first the particle is in a curved $H_2$ space and second we have the coupling to the electric field. These are minor modifications, but the arguments leading
to \nref{PolFlat} continue to be valid so that the partition function of the JT model can be
written directly:
\begin{equation} \label{PolAdS}
\scaleto{\int {\cal D }\vec x e^{2\pi q - m_0 \tilde\tau +q\int a } \delta( { \dot x^2 + \dot y^2 \over y^2 }- q^2 ) =
e^{ 2\pi q- {1\over 2} \mu^2 \tau} \Tr e^{ - \tau H } = e^{2\pi q - {1\over 2} \mu^2 \tau }
\int { \cal D}{\bf x} \exp\left( - \int_0^\tau d\tau' {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } - q {\dot x \over y}
\right)}{26pt}
\end{equation}
The delta function implements the first condition
in \nref{BCond} at each point along the path.
The last path integral can be done exactly by doing canonical quantization of the action (section \ref{CanonicalQuantization}) and by comparing the result with the one from the Schwarzian action \cite{Stanford:2017thb}
we can determine that $\tau$ is the inverse temperature $\beta$.
In the above discussion we have been fixing the time along the boundary. Instead we can fix the energy
at the boundary, where the energy is the variable conjugate to time.
This can be done by simply integrating \nref{PolAdS} times $e^{ \beta E}$
over $\beta$ along the imaginary axis. This fixes the energy of the non-relativistic problem by generating a $\delta( H - E)$. More precisely, we will argue that after doing a spectral decomposition we can write
the propagator at coincident points as
\begin{equation} \label{JTEn}
Z_{JT}(\beta ) = \int_0^\infty \rho(E)e^{-\beta E} dE \longrightarrow \rho(E) = \int_{-i \infty}^{i \infty}
{ d\beta \over i } e^{ E \beta } Z_{JT}(\beta)
\end{equation}
where the function $\rho(E) $ can then be interpreted as a ``density of states" in the microcannonical ensemble. We will give its explicit form in section (4.2).
For now, we only want to contrast this integral with a superficially similar one that appears when we
compute the relativistic propagator
\begin{equation} \label{RelEl}
e^{-2\pi q} \int_0^\infty e^{ E \beta }Z_{JT} (\beta) = \langle \phi(x) \phi(x) \rangle
\end{equation}
which gives the relativistic propagator of a massive particle in an electric field at coincident points
(we can also compute this at non-coincident points to get a finite answer). The total mass of the
particle is
\begin{equation}
m = q - {E\over q}
\end{equation}
For large $q$ this is above threshold for pair creation\footnote{See Appendix \ref{appendix:relativistic particle}}. The pair creation interpretation
is appropriate for the problem in \nref{RelEl}, but not for \nref{JTEn}.
In both problems we have a classical approximation to the dynamics that corresponds to a particle
describing a big circular trajectory in hyperbolic space at radius $\rho_c$ :
\begin{equation} \label{SolCir}
\tanh \rho_c = { m \over q }
\end{equation}
For the problem in \nref{RelEl}, fluctuations around this circle lead to an instability, with a single negative mode and an imaginary part in the partition function \nref{RelEl}. This single negative mode corresponds to small fluctuations of the overall size of the circular trajectory around \nref{SolCir}.
On the other hand in \nref{JTEn} we are integrating the same mode along a different contour, along the
imaginary axis, where we get a real and finite answer. Furthermore, the imaginary part in the
partition function \nref{RelEl} comes precisely from the trajectory describing pair creation, which is also the type of contribution captured in \nref{JTEn}.
Finally, in the relativistic particle problem, we expect that the pair creation amplitude should be exponentially suppressed for large $q$, while the partition function for the JT model is not. In fact,
for large $q$ the exponential suppression factor for pair creation goes as $e^{ - 2 \pi q } $, which is precisely cancelled by a similar factor in \nref{JTPa}, to obtain something finite in the large $q$ limit.
\section{Solving the Quantum Mechanical Problem}
\label{CanonicalQuantization}
As we explained above the solution of the JT theory is equivalent to considering a non-relativistic particle in $AdS_2$ or $H_2$.
We first consider the Euclidean problem, of a particle moving in $H_2$. An ordinary magnetic field in $H_2$ leads to an Euclidean action
of the form
\begin{equation} \label{magnetic field}
S= \int du {1\over 2} { \dot x^2 + \dot y^2 \over y^2 } + i b \int d u { \dot x \over y } -{1\over 2} ( b^2 +{ 1 \over 4} ) \int d u ~,~~~~~~~b = i q
\end{equation}
If $b$ is real we will call it a magnetic field,
when $q$ is real we will call it an ``electric" field. The last term is a constant we introduced for convenience. Its only effect
will be to shift the ground state energy. It is interesting to compute the classical solutions and the corresponding action for
\nref{magnetic field}. These solutions are simplest in the $\rho$ and $\varphi $ coordinates, using the SL(2) symmetry we find that the trajectories are given by ($t=-iu$):
\begin{eqnarray}
{1\over 2}\sinh^2\rho ({d\varphi\over dt})^2+{q^2\over 2}-{1\over 8}=E,~~ \cosh\rho={q\beta\over 2\pi},~~ {d\varphi\over du}={2\pi\over \beta}.
\end{eqnarray}
In this classical limit we
get the following relations for the action and the temperature:
\begin{eqnarray}
& ~& { \beta \over 2 \pi } = {1\over \sqrt{2E+{1\over 4}}}
\cr
& ~ & -S = {2\pi^2\over \beta}+{\beta\over 8}-2\pi q \label{ClassEnt}
\end{eqnarray}
When $b$ is real, this system is fairly conventional and it was solved in \cite{Comtet:1986ki} . Its detailed spectrum depends on $b$. For very large $b$ we have a series of Landau levels and also a continuous spectrum. In fact, already the classical problem contains closed circular orbits, related to the discrete Landau levels, as well as orbits that go all the way to infinity.\footnote{See Appendix \ref{appendix:landau level}}
The number of discrete Landau levels decreases as we decrease the magnetic field and for $0 < b < {1\over 2}$ we only get a continuous spectrum. The system has a $SL(2)$ symmetry and the spectrum organizes into SL(2) representations, which are all in the
continuous series for $0 < b < 1/2$. For real $q$ we also find a continuous spectrum which we can view as the analytic continuation
of the one for this last range of $b$.
The canonical momenta of the action (\ref{magnetic field}) are:
\begin{equation}
p_x={\dot x\over y^2}+{i q\over y};~~~~~~~p_y={\dot y\over y^2}.
\end{equation}
And the Hamiltonian conjugate to $\tau_L$ is thus:
\begin{equation}
H={\dot x^2+\dot y^2\over 2y^2}+{q^2\over 2}={y^2\over 2} [(p_x-i{q\over y})^2+p_y^2] +{ q^2 \over 2 } - { 1 \over 8}
\end{equation}
Note that the Hamiltonian is not Hermitian. However, it is {\cal PT}-symmetric (here parity reflects $x$ and $p_x$) and for that
reason the spectrum is still real, see \cite{Bender:2005tb}.
The action is invariant under $SL(2,R)$ transformations generated by
\begin{eqnarray}
L_0=x p_x+y p_y;~~~~~L_{-1}=p_x;~~~~~~~~L_{1}=(y^2-x^2)p_x-2xyp_y-2iqy \label{SLtwoGen}
\end{eqnarray}
Notice the extra $q$ dependent term in $L_1$ that arises due to the presence of a magnetic field.
Up to a simple additive constant, the Hamiltonian is proportional to the Casimir operator
\begin{equation}
H={1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) + { q^2 \over 2} - { 1 \over 8 }
\end{equation}
As is common practice, let us label the states by quantum numbers $j={1\over 2}+is$ and $k$, so that $H|j,k\rangle=j(1-j)|j,k\rangle$ and $L_{-1}|j,k\rangle=k|j,k\rangle$.
We can find the eigenfunctions by
solving the $Schr\ddot{o}dinger$ equation with boundary condition that the wavefunction should vanish at the horizon $y\rightarrow \infty$ \cite{Comtet:1984mm,Comtet:1986ki,Pioline:2005pf}:
\begin{equation}
\omega_{s,k}={s^2\over 2} ,~~~~~~f_{s,k}(x,y)=
\begin{cases}
({s\sinh 2\pi s\over 4\pi^3k})^{1\over 2}|\Gamma(is-b+{1\over 2})|e^{-ik x}W_{b,is}(2 k y), ~~~k>0;\\
({s\sinh 2\pi s\over 4\pi^3|k|})^{1\over 2}|\Gamma(is+b+{1\over 2})|e^{-ik x}W_{-b,is}(2|k|y),~~~k<0.
\end{cases}
\end{equation}
where $\omega_{ks}$ is giving the energy of the states labelled by $s$ and $k$, and $ W $ is the Whittaker function. The additive constant in \nref{magnetic field} was
introduced to simplify this equation. We can think of $s$ as the quantum number of the continuous series representation of $SL(2)$ with
spin $j={1\over 2} + i s $.
After continuing $b \to i q$ we find that the gravitational system has a continuous
spectrum
\begin{equation}\label{energy relation}
E(s)={s^2 \over 2 }.
\end{equation}
\subsection{The Propagator }
It is useful to compute the propagator for the non-relativistic particle in a magnetic field,
$K(u,\boldsymbol{x_1},\boldsymbol{ x_2})=\langle \boldsymbol{x_1}| e^{-u H}|\boldsymbol{x_2}\rangle$. Here, $\boldsymbol{x}$ stands for ${x,y}$.
The propagator for a real magnetic particle was obtained in \cite{Comtet:1986ki}:
\begin{eqnarray} \label{particle heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})& =& e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} \int_0^{\infty}ds s e^{- u { s^2 \over 2}}{\sinh 2\pi s\over 2\pi (\cosh 2\pi s+\cos 2\pi b)}{1\over d^{1+2is}}
\times
\cr
& & ~~~~~ \times ~_2F_1({1\over 2}-b+is,{1\over 2}+b+is,1,1-{1\over d^2}).
\cr
d & = & \sqrt{(x_1-x_2)^2+(y_1+y_2)^2\over 4y_1 y_2}
\cr
e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})} &=& e^{-2 i b \arctan {x_1-x_2\over y_1+y_2}}
\end{eqnarray}
In the case that we have a real magnetic field the prefactor is a phase and it is
gauge dependent. It is equal to the value of Wilson Line $e^{i\int a}$ stretched along the geodesic between $x_2$ and $x_1$.
Here we quoted the value in the gauge where the action is
\nref{magnetic field}. The second equation defines the parameter $d$, which is a function of the geodesic distance between the two points. Note that $d=1$ corresponds to coincident points.
We can get the answer we want by making the analytic continuation $b\to i q$ of this formula.
We can check that this is the right answer for our problem by noticing the following. First one can check that this expression is invariant under the SL(2) symmetry $L_a = L^1_a + L^2_a$ where $L_a$ are the generators \nref{SLtwoGen} acting on $\boldsymbol{x_1}$
and $L_a^2$ are similar generators as in \nref{SLtwoGen}, but with $q\to -q$. It is possible to commute
the phase $e^{ i \varphi(\boldsymbol{x_1},\boldsymbol{x_2})}$, in \nref{particle heat kernel} past these generators which would remove the $q$ dependent terms. This implies
that the rest should be a function of the proper distance, which is the case with \nref{particle heat kernel}.
Then we can check the equation
\begin{eqnarray}
0 &=& (\partial_u + H_1 )G(u,\boldsymbol{x_1},\boldsymbol{x_2})
\end{eqnarray}
which is also indeed obeyed by this expression. The $s$ dependent prefactor is fixed by the requirement that the
propagator composes properly, or more precisely, by saying that for $u=0$ we should get a $\delta$ function.
\subsection{Partition Function}
The gravitational partition function is related with the particle partition function with inverse temperature $\beta$.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.3]{Freeenergy.pdf}
\caption{Free Energy diagram with inverse temperature $\beta$. }
\end{figure}
The canonical partition function of the quantum mechanical system is
\begin{eqnarray}
Z_{\scaleto{Particle}{4pt}}&=&Tr e^{-\beta H}
=\int_0^{\infty} ds \int_{-\infty}^{\infty}dk \int_M {d x d y\over y^2} e^{-{\beta } { s^2 \over 2 } }f^*_{s,k}(x,y)f_{s,k}(x,y)\nonumber\\
&=&V_{AdS}\int_0^{\infty}ds e^{-\beta {s^2\over 2}}{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}.
\end{eqnarray}
The volume factor $V_{AdS}$ arises because after momentum integration there is no position dependence.
In a normal quantum mechanical system, the volume factor means that the particle can have independent configurations at different locations of our space, however for a gravitational system this should be thought as redundant and should be cancelled by the volume of $SL(2,R) $ gauge group $2\pi V_{AdS}$\footnote{There might be a multiplicative factor in the volume of gauge group, but we can always absorb that into $S_0$.}.
In gravitational system, there can also other contributions to the entropy from pure topological action.
These give a contribution to the ground state entropy $S_0$. Including the topological action in (\ref{regularized action}), we find a gravitational ``density of states" as
\begin{equation} \label{Spectral}
\rho(s)=\underbrace{e^{S_0}e^{2\pi q}}_{\text{extra terms}}\underbrace{1\over {2\pi}}_{\text{residue gauge}}\underbrace{{s\over 2\pi}{\sinh(2\pi s)\over \cosh(2\pi q)+\cosh(2\pi s)}}_{\text{particle in magnetic field}}=e^{S_0}e^{2\pi q}{s\over 2\pi^2}\sum_{k=1}^{\infty}(-1)^{k-1}e^{-2\pi q k}\sinh(2\pi s k).
\end{equation}
We have not given an explicit description of these states in the Lorentzian theory. More details were discussed in \cite{Kitaev:2018wpr,Lin:2018xkj}.
This expression has some interesting features. Notice that the classical limit corresponds to large $q$ and large $s$, where we reproduce \nref{ClassEnt}. After approximating, the density of states are
$\log \rho(s) \sim S_0 + 2 \pi s $ for $ s/q < 1$ and $S_0+2\pi q$ for $s/q>1$.
We can also expand the partition function for very small and very large temperatures where we obtain
\begin{eqnarray}
Z_{JT} &\sim & e^{ S_0} e^{ 2 \pi q } { 1 \over 4\pi^2\beta} ~,~~~~~~~~~~~ \beta \ll {1\over q}
\cr
Z_{JT} & \sim & e^{S_0} { 1 \over \sqrt{2\pi}\beta^{3/2} }~,~~~~~~~~~~~~ \beta \gg 1
\end{eqnarray}
Notice that at leading order we get an almost constant entropy both at low and high temperatures, with the
high temperature one being higher. In both cases there are power law corrections in temperature.
Before we try to further elucidate the interpretation of this result, let us emphasize a couple of important defects of our discussion. First, when we replaced the partition function of the JT theory by the action of a non-relativistic
particle in an electric field, we were summing over paths in $H_2$. This includes paths that self intersect see figure \ref{fig:TwoInstanton}. Such paths do not have an obvious interpretation in the JT theory and it is not even clear that we should include them. For example, the sum over $k$ in \nref{Spectral} can be understood in terms of classical solutions which wind $k$ times around the circle. These make sense for the problem of the particle in the electric field but apparently not in the JT theory.
\begin{figure*}[h!]
\centering
\subfigure[Density of States]{%
\includegraphics[width=0.4\textwidth]{densityofstate.pdf}%
\label{fig:dos}%
}\hspace{4cm
\subfigure[Two Instantons]{%
\includegraphics[width=0.35\textwidth]{twoinstanton.pdf}%
\label{fig:TwoInstanton}%
\caption{}
\end{figure*}
Maybe such paths could be given some interpretation in the gravity theory.
Alternatively, we might want to sum over paths that do not self intersect.
A second defect is that we would be eventually interested in adding some matter fields propagating in the bulk geometry. These matter fields have boundary conditions at the boundary of the region of hyperbolic space
cut out by the boundary trajectory. The partition function of the fields with such an arbitrary boundary trajectory could also modify the results we described above. Of course, this issue does not arise if we have
the pure JT theory. It is only important if we want to introduce bulk matter fields to define more complex observables.
Instead of attempting to address the above issues, we will take an easy route, which is to consider the system only in the large $q$ (or small $\epsilon$) limit. In this regime, we address the above issues, and we can still trust the description of the particle in the electric field.
This large $q$ or small $\epsilon$ limit is the same one that isolates the Schwarzian action from the JT theory
\cite{Maldacena:2016upp,Jensen:2016pah,Engelsoy:2016xyb}.
It turns out that the limit can be taken already at the level of the mechanical system, a simple rescaled
version of the above system. This provides an alternative method for quantizing the Schwarzian theory. It has the advantage of being a straightforward second order action of a particle moving in a region near the boundary of hyperbolic. Of course, the Schwarzian theory was already quantized using a variety of methods in
\cite{Bagrets:2016cdf,Bagrets:2017pwq,Stanford:2017thb,Mertens:2017mtv,Luca}. We will simply provide yet another perspective, recover the old results, and write a few new expressions.
\section{Quantum Gravity at Schwarzian Limit}
Before getting into the details notice that the large $q$ limit of \nref{Spectral} gives
\begin{equation}
\rho(s) = e^{S_0}{s\over 2\pi^2}\sinh(2\pi s), ~~~E = { s^2 \over 2} ~,~~~~~~~Z_{JT} = \int_0^\infty ds
\rho(s) e^{ - \beta { s^2 \over 2 } } =e^{S_0}{1 \over\sqrt{2\pi}\beta^{3\over 2}}e^{2\pi^2 \over \beta}.
\end{equation}
This reproduces what was found in \cite{Stanford:2017thb,Mertens:2017mtv,Mertens:2018fds,Blommaert:2018oro} by
other methods. We see that we get a finite answer and also that the contributions from the $k>1$ terms in
\nref{Spectral} have disappeared. Because the $S_0$ part decouples with JT gravity, from now on, we will drop it and discuss $S_0$ only when it is necessary.
\subsection{The Propagator}
To get a limit directly at the level of the mechanical system it is useful to define a
rescaled
coordinate, $z$, via
\begin{equation}
y= z/q. \label{yAndq}
\end{equation}
After taking the large $q$ limit, the boundary particle propagator becomes \footnote{see the Appendix \ref{appendix:large q} for details}:
\begin{eqnarray}\label{heat kernel}
G(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&{ 1 \over q } e^{-2\pi q \theta(x_2-x_1) }\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2});~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~q\gg 1.\\
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{2\sqrt{z_1z_2}}{\pi^2 |x_1-x_2|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|});~~~~~~~~
\label{HeatLim}\\
&=&e^{-2 \frac{z_1+z_2}{x_1-x_2} }\frac{\sqrt{2}}{\pi^{3/2} u^{3/2}}\frac{\sqrt{z_1z_2}}{ |x_1-x_2|}\int_{-\infty}^{\infty}d\xi (\pi+i\xi)e^{-2\frac{(\xi-i\pi)^2}{u}-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}.\label{HeatLim2}
\end{eqnarray}
The original phase factor $e^{i \varphi(\boldsymbol{x_1},\boldsymbol{x_1})}$ factorizes into a product of singular ``phase" $e^{-2\pi q \theta(x_2-x_1)}$, with $\theta$ the step function,
and a regular ``phase" $e^{-2 {z_1+z_2\over x_1-x_2}}$. The singular ``phase" is the same order as the topological piece in (\ref{regularized action}).
In order to have a finite result they should cancel between each other. This can only be satisfied if the $x_is$ are in cyclic order.
As shown in figure (\ref{fig:phasefactor}), the product of singular ``phase" gives $-2\pi q$ for cyclic order $x_i$s and this would cancel with the topological action $2\pi q$.
While for other ordering of the $x_i$s, this would have $-2 \pi n q$ for $n=2,3,...$ and is highly suppressed in the limit $q$ goes to infinity.
This cyclic order is telling us where the interior of our space time is. The magnetic field produces a preferred orientation for the propagator.
After fixing the order, all our formulas only depend on $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ which has no $q$ dependence.
The residual $q$ factor in \ref{heat kernel} cancels out the additional $q$ from the measure of coordinate integral, ${ dx dy \over y^2 } \to q { dx dz \over z^2 } $.
In conclusion, after taking the limit we get a finite propagator equal to \nref{HeatLim}, which should be multiplied by a step function $ \theta( x_1 - x_2 ) $ that imposes the right order.
The final function $\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})$ has the structure of $e^{-2{z_1+z_2\over x_1-x_2}}f(u,{z_1z_2\over (x_1-x_2)^2})$. This can be understood directly from the $SL(2)$ symmetry.
After taking the large $q$ limit, the $SL(2,R)$ charges become
\begin{equation} \label{SLTwoLq}
L_0=i (x\partial_x+z\partial_z);~~~~L_{-1}=i\partial_x;~~~~~~L_{1}=-ix^2\partial_x-2ix z\partial_z-2i z.
\end{equation}
We can check that they still satisfy the SL(2) algebra.
If we drop the last term in $L_1$, the $SL(2,R)$ charges become the usual differential operators on $EAdS_2$. And the propagator will have only dependence on the geodesic distance. When $L_1$ operator is deformed, the condition of $SL(2,R)$ invariance fixes the structure of the propagator as follows.
The $L_0$ and $L_{-1}$ charges are not deformed and they imply that the
only combinations that can appear are
\begin{equation}
v \equiv {z_1+z_2\over x_1-x_2} ~~~~~~~~~~~~~\text{and}~~~~~~~~~~~~~~ w\equiv {z_1 z_2\over(x_1-x_2)^2}.
\end{equation}
Writing the propagator as
$\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_1})=k(v,w) $
and
requiring it to be invariant under $L_1$ gives the following
equation for $\alpha$:
\begin{equation}\label{phaseequation}
\partial_v k +2 k =0 \longrightarrow k = e^{ - 2 v } h(w)
\end{equation}
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=e^{-2 {z_1+z_2\over x_1-x_2}}f(u,{z_1 z_2\over(x_1-x_2)^2}),
\end{equation}
The full function can also be determined directly as follows. Again we impose the propagator equation (or heat equation)
\begin{eqnarray}
0 &=& \left[ \partial_u + {1\over 2}\left( L_0^2+ { 1 \over 2} L_{-1}L_{1} + { 1 \over 2} L_{1}L_{-1} \right) -{1\over 8} \right] \tilde K
\cr
0 & = & [ { s^2 \over 2} + {w^2\over 2}\partial_w^2+2w +{1\over 8}] K_s(w)
\end{eqnarray}
where $L_a$ are given in \nref{SLTwoLq} and are acting only on the first argument of $\tilde K$.
The solution of the last equation which is regular at short distances ($w \to \infty$) is $\sqrt{w}$ times
the Bessel K
function in \nref{heat kernel}.
We can also directly determine the measure of integration for $s$ by demanding that the propagator at $u=0$ is
a $\delta$ function or by demanding the propagator compose properly. This indeed is the case with the $s\sinh{ 2 \pi s}$ function in \nref{HeatLim}.
To explicitly show the above statement, it will be useful to use spectral decomposition of the propagator:
\begin{equation}\label{SpectralDecomposition}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})=\int ds {2s\sinh(2\pi s)\over \pi^3} e^{-{s^2u\over 2}}\int dk \sqrt{z_1 z_2} e^{ik(x_1-x_2)}K_{2is}(2\sqrt{2i k z_1})K_{2is}( 2\sqrt{2ikz_2}).
\end{equation}
It can be easily checked that the special functions $f_{k,s}(x,z)=\sqrt{z}e^{ikx}K_{2is}( 2\sqrt{2ikz})$ are delta function normalizable eigenmodes of the large $q$ Hamiltonian:
\begin{equation}
\int {dx dz\over z^2}f_{k_1,s_1}f_{k_2,s_2}=\delta(k_1-k_2)\delta(s_1-s_2){\pi^3\over 2s\sinh (2\pi s)}
\end{equation}
Notice that the inner product fixes the integral measure completely in (\ref{SpectralDecomposition}), and the composition relation is manifestly true:
\begin{equation}
\int {dx dz\over z^2}\tilde{K}(u_1,\boldsymbol{x_1},\boldsymbol{x})\tilde{K}(u_2,\boldsymbol{x},\boldsymbol{x_2})=\tilde{K}(u_1+u_2,\boldsymbol{x_1},\boldsymbol{x_2})
\end{equation}
At short time the propagator has the classical behavior:
\begin{equation}
\tilde{K}(u,\boldsymbol{x_1},\boldsymbol{x_2})\sim \delta(x_1-x_2+u z_2)e^{-{(z_1-z_2)^2\over 2 u z_2}}
\end{equation}
This form of singularity is expected since we are taking the large $q$ limit first and thus the velocity in $x$ direction is fixed to be $z$. In the original picture of finite $q$ we are looking at the time scale which is large compare to AdS length but relatively small such that the quantum fluctuations are not gathered yet.
The integral structure in the propagator (\ref{HeatLim}) has an obvious meaning: integrating over $s$ represents summing over all energy states with Boltzmann distribution $e^{-E u}$, and the Bessel function stands for fixed energy propagator.
We want to stress that the argument in the Bessel function is unusual, and at short distance it approaches a funny limit:
\begin{equation}
K_{2is}(\frac{4 }{\ell})\simeq \sqrt{\pi \over 8\ell}e^{-\frac{4 }{\ell}},~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}\rightarrow 0.
\end{equation}
One should contrast this exponential suppression with the short distance divergence in QFT which is power law.
In our later discussion of exact correlation function with gravity backreaction, we will see that this effect kills UV divergence from matter fields.
To obtain the expression (\ref{HeatLim2}), we use the integral representation for the Bessel function and the final result has some interesting physical properties:
Firstly, we see that at large $u$ the time dependence and coordinate dependence factorized.
So, at large time we have a universal power law decay pointed out in \cite{Bagrets:2016cdf}.
Secondly, as we said before, the phase factor $e^{-2\frac{z_1+z_2}{x_1-x_2}}$ is equal to the Wilson line $e^{-q\int_1^2 a}$ stretched along the geodesic connection between location $1$ and $2$ (Figure \ref{fig:propagator}). The field $a$ depends on our choice of gauge, our convention corresponds to fix the minimum value of $a$ at infinity and then the Wilson line is equal to $e^{-q A}$, where $A$ is the area of a hyperbolic triangle spanned by $1,2$ and $\infty$.
Thirdly, defining $2\pi+2i \xi$ as $\theta$, then $\theta$ has the meaning of the spanned angle at the horizon (Figure \ref{fig:propagator}). Then the gaussian weight $e^{-2{(\xi-i\pi)^2\over u}}=e^{\theta^2\over 2u}$ can be understood from the classical action along the boundary with fixed span angle $\theta$. The boundary drawn in the figure represents a curve with fixed (regularized) proper length $u$ in $H_2$.
Lastly, the factor $e^{-\frac{4 \sqrt{z_1z_2}}{|x_1-x_2|}\cosh \xi}=e^{-{4\cos\theta\over \ell}}$ is equal to $e^{q(\alpha+\beta)}$, which is a corner term that arise from JT gravity in geometry with jump angles. Here $\alpha$ and $\beta$ are defined as the angle spanned by the geodesic with fixed length and the ray coming from horizon to the boundary.
In summary the propagator can be understood as an integral of JT gravity partition functions over geometries \ref{fig:propagator} with different $\theta$s.
\begin{figure*}[h!]
\centering
\subfigure[singular ``phase" factor for different ordering]{%
\includegraphics[width=0.45\textwidth]{phase_factor.pdf}%
\label{fig:phasefactor}%
}\hspace{1.5cm
\subfigure[A geometric representation of the propagator. Here we fix the span angle $\theta$, the propagator is a summation over such geometries.]{%
\includegraphics[width=0.45\textwidth]{propagator.pdf}%
\label{fig:propagator}%
\caption{}
\label{fig:Propagator}
\end{figure*}
Finally, let us comment on the issues we raised in section 4.2. In the large $q$ limit we are considering the propagator at relatively large distances and in a regime where locally in $AdS$ the integration over paths that fluctuate
wildly is suppressed. Alternatively we can say that in the integration over paths we put a UV cutoff which
is large compared to $1/q$ but small compared to the AdS radius. This is the non-relativistic regime for the
boundary particle. The quantum effects are still important at much longer distances due to the large size
of $AdS$.
In addition, if we have quantum fields in AdS, then their partition functions for these fluctuating contours that
have fluctuations over distances larger than the $AdS$ radius are expected to depend on this shape in a local way. Due to the symmetries of $AdS_2$, this is simply expected to renormalize the action we already have without introducing extra terms. This can be checked explicitly for conformal field theories by using
the conformal anomaly to compute the effective action of the CFT$_2$ on a portion of $H_2$ (Appendix \ref{appendix: CFT effective action}).
\subsection{Wheeler-DeWitt Wavefunction }
\label{WdW}
In the pure JT theory we can think about quantizing the bulk theory and obtaining the
Wheeler-DeWitt wavefunction. This was discussed in the classical limit by Harlow and Jafferis \cite{Harlow:2018tqv}.
The Wheeler-DeWitt wavefunction can be created by Euclidean evolution of the boundary and hence is closely related to the propagator we have discussed above. The wavefunction in Lorentzian signature could then be obtained by analytic continuation of the boundary time.
The Euclidean evolution can be specified by either of the two parameters: the proper length $u$ or energy $E$. Choosing a different parameter corresponds to imposing a different boundary condition in JT theory. In general there are four possible choices of boundary conditions in 2d dilaton gravity, there are two sets of conjugate variables: $\lbrace \phi_b,K\rbrace$, and $\lbrace u,E\rbrace$ \footnote{Energy $E$ is proportional to the normal derivative of the dilaton field at the boundary.}. In preparing the wavefunction we fix the boundary value of dilaton and hence there are only two choices of the parameter ($u$ or $E$). We denote the corresponding wavefunction as $|u\rangle_G$ and $|E\rangle_G$ respectively.
In terms of holographic considerations, $|u\rangle_G$ represents a thermofield double state:
\begin{equation}
|u\rangle_G\sim\sum\limits_{n}e^{-E_n u}|E_n\rangle_L |\bar E_n\rangle_R
\end{equation}
and $|E\rangle_G $ is like an average of energy eigenstates in a window of energy $E$:
\begin{equation}\label{EnergyStateDefinition}
|E\rangle_G\sim {1\over \delta E}\sum\limits_{ |E-E_n|<\delta E}|E_n\rangle_L |\bar E_n\rangle_R.
\end{equation}
The width of the energy window is some coarse graining factor such that the summation contains $e^{S_0}$ states and does not show up clearly in gravity.\footnote{If one understand getting $|E\rangle$ state from integrating over thermofield double state in time direction, then a natural estimate on $\delta E$ is ${1\over T}$, where $T$ is the total time one integrate over. The validity of JT description of boundary theory is $T<e^{S_0}$, and we get $\delta E>e^{-S_0}$. For $T>e^{S_0}$, there are other possible instanton contributions. The proper gravitational theory at this regime is studied in paper \cite{Saad:2018bqo} }
With the definition of the states, one can evaluate them in terms of different basis. There are three natural bases turn out to be useful, we call them $S$, $\eta$ and $\ell$ bases. Basis $S$ corresponding to fix the horizon value of dilaton field $\phi_h$, or equivalently by Bekenstein-Hawking formula, the entropy of the system. The canonical conjugate variable of $S$ will be called $\eta$ and that characterizes the boost angle at the horizon. $\ell$ stands for fixing geodesic distance between two boundary points. To see that the horizon value of the dilaton field is a gauge invariant quantity, one can do canonical analysis of JT gravity. With ADM decomposition of the spacetime metric, one can get the canonical momenta and Hamiltonian constraints of the system \cite{LouisMartinez:1993eh}:
\begin{eqnarray}
ds^2&=&-N^2dt^2+\sigma^2(dx+N^xdt)^2;~~~~~~~~~~~~~~~~~~~~\\
\mathcal{H}&=&-\Pi_{\phi}\Pi_{\sigma}+\sigma^{-1}\phi''-\sigma^{-2}\sigma'\phi'-\sigma\phi;~~~~\mathcal{H}_x=\Pi_{\phi}\phi'-\sigma\Pi_{\sigma}';\label{Hamiltonian Constraint}\\
\Pi_{\phi}&=&N^{-1}(-\dot{\sigma}+(N^x\sigma)')=K\sigma;~~~~~~~~~~~~~\Pi_{\sigma}=N^{-1}(-\dot{\phi}+N^x\phi')=\partial_n\phi.\label{Canonical Momentum}
\end{eqnarray}
That is the dilaton field is canonically conjugate to the extrinsic curvature and boundary metric is canonical conjugate to the normal derivative of the dilaton field (both are pointing inwards).
By a linear combination of the Hamiltonian constraints (\ref{Hamiltonian Constraint}), one can construct the following gauge invariant quantity $C$:
\begin{equation}
-{1\over\sigma}(\phi'\mathcal{H}+\Pi_{\sigma}\mathcal{H}_x)={1\over 2}(\Pi_{\sigma}^2+\phi^2-{\phi'^2\over \sigma^2})'\equiv C[\Pi_{\sigma},\phi,\sigma]'\sim 0
\end{equation}
The Dirac quantization scheme then tells us that the quantity $C$ has a constant mode which is gauge invariant (commute with Hamiltonian constraint).
Choosing the gauge that normal derivative of the dilaton is zero, we can solve the Hamiltonian constraint:
\begin{equation}
\phi^2-(\partial_{X}\phi)^2=2C\equiv S^2~~~~~\rightarrow~~~~~~\phi(X)=S\cosh X,
\end{equation}
where $dX= \sigma dx$ is the proper distance along the spatial slice. Because the normal derivative of dilaton field is zero, the minimum value of dilaton at this spatial slice is actually a local extremum in both directions. Therefore, the minimal value of dilaton field, namely $S$, is a global variable. The classical geometry in this gauge is a ``Pac-Man" shape (right figure in figure \ref{fig:Pac-Man}).
Focusing on the intersection region of the spatial slice and the boundary, we have the spatial slice is orthogonal to the boundary.
This is because we are gauge fixing $\partial_n\phi=0$ on the spatial slice, and $\phi=\phi_b$ on the boundary.
The ADM mass of the system, after regularization, is then $M=\phi_b(\phi_b-\partial_{X}\phi)$ \cite{Maldacena:2016upp}. Substituting the behavior of $\phi(X)$ we get:
\begin{equation}
M={S^2\over 2}.
\end{equation}
This is the same relation in \ref{energy relation} and therefore we can interpret the $s$ variable as entropy of our system $S$.
For the purpose of fixing geodesic distance, it is convenient to think of doing the path integral up to a slice $L$ with zero extrinsic curvature. This picks out a particular slice (left figure in Figure \ref{fig:Pac-Man}) among the solutions obeying the Hamiltonian constraint.
The WdW wavefunction can be evaluated as an Euclidean path integral with fixed (rescaled) geodesic distance $d$ between the two boundary points:
\begin{eqnarray}\label{WdWPath}
\nonumber \Psi(u;d)=\int {\cal D}g{\cal D}\phi e^{{1\over 2}\int \phi(R+2)+\int_{L}\phi K+\phi_b\int_{Bdy}(K-1)}=\int {\cal D }f e^{\phi_b\int_{Bdy} (K-1)}\\
=\int {\cal D}\boldsymbol x e^{-m \int_{Bdy}\sqrt{g}+q\int_{Bdy} a+q\int_{L} a+(\pi-\alpha_1-\alpha_2)q}~~~~~~~~~~~~~~~~~~~~~~~
\end{eqnarray}
Here we are fixing the total length of $L$ to be $d$ and the proper length of the boundary to be $u$. $\alpha_1$ and $\alpha_2$ in the last expression denotes the jump angle at the corner coming from the singular contribution of the extrinsic curvature and should be integrated over. Without the $e^{q\int_L a}$ factor in (\ref{WdWPath}), the path integral corresponds to the propagator (\ref{HeatLim}). Remember that the phase factor is equal to $e^{-q\int_L a}$, so the wavefunction in $\ell$ basis is actually the propagator (\ref{HeatLim}), with the phase factor stripped out
\begin{equation}\label{wavefunction}
\Psi(u; \ell)\equiv \langle \ell|u\rangle_G = \frac{2}{\pi^2 \ell}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2 } u}K_{2is}(\frac{4 }{\ell}) ,~~~~~~~~\ell={|x_1-x_2|\over\sqrt{z_1 z_2}}.
\end{equation}
$\ell$ is a function of the regularized geodesic distance $d$ between $\boldsymbol{x_1}$ and $\boldsymbol{x_2}$ : $\ell=e^{d\over 2}$.
The semiclassical of $\Psi(u;\ell)$ can be obtained using formula (\ref{HeatLim2}), in the exponent we get saddle point result:
\begin{equation}
\Psi(u;\ell)\sim \exp[-{2(\xi_*-i\pi)^2\over u}+{4\over u}{\xi_*-i\pi\over \tanh \xi_*}];~~~~~~~~{\xi_*-i\pi\over u}=-{\sinh \xi_*\over \ell}.
\end{equation}
The same saddle point equation and classical action was obtained in \cite{Harlow:2018tqv} by a direct evaluation in JT gravity.
The wavefunction with fixed energy boundary condition can obtained by multiplying $\Psi(u;\ell)$ by
$e^{ + E u} $ and integrating over $u$ along the imaginary axis. This sets $E = {s^2 \over 2} $
in the above integral over $s$. So this wavefunction has a very simple expression:
\begin{equation}\label{fixed energy wavefunction}
\Psi(E;\ell)\equiv \langle \ell|E\rangle_G = \rho(E)\frac{4}{ \ell} K_{i \sqrt{ 8 E} }(\frac{4 }{\ell}).~~~~~~~~
\end{equation}
The classical geometry for $\Psi(E;\ell)$ is the same as the left figure in Figure \ref{fig:Pac-Man}, with fixing energy on the boundary.
We want to stress that it is important to have the $\rho(E)$ factor in (\ref{fixed energy wavefunction}) for a classical geometry description since we are averaging over the states.
We can roughly think of ${4\over \ell}K_{i\sqrt{8E}}({4\over \ell})$ as a gravitational ``microstate" $|\mathcal{E}\rangle$ with fixed energy $E$. Such a ``microstate" will not have a classical geometry representation and therefore is just a formal definition.
The inner product between wavefunctions is defined as $\langle\Psi_1|\Psi_2\rangle=\int_0^{\infty}d\ell \ell \Psi_1^*(\ell)\Psi_2(\ell)$.
Going to the entropy basis $S$, it is easy to start with $\Psi(E)$.
Because of the identity $E={S^2\over 2}$, expanding $\Psi(E)$ in the $S$ basis is diagonal:
\begin{equation}
\Psi(E;S)\equiv \langle S|E\rangle_G=\sqrt{\rho(S)}\delta(E-{S^2\over 2})
\end{equation}
We put this square root of $\rho(S)$ factor in the definition of $S$ basis such that inner product between different $S$ state is a delta function $\langle S|S'\rangle=\delta(S-S')$. This factor is also required such that the classical limit matches with gravity calculation.
Integrating over energy with Boltzman distribution, we can get the expression of thermofield double state in the $S$ basis:
\begin{equation}
\Psi(u;S)\equiv \langle S|u\rangle_G=\int dE e^{-u E}\langle S|E\rangle_G=\sqrt{\rho(S)}e^{-{uS^2\over 2}}
\end{equation}
In the semiclassical limit, the wavefunction becomes gaussian and coincides with the on shell evaluation of the ``Pac-Man" geometry (Figure \ref{fig:Pac-Man}):
\begin{equation}\label{PsiS}
\Psi(u,S)\sim \sqrt{S}e^{\pi S -{u S^2\over 2}}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.5]{classicalgeometry.pdf}
\caption{Classical Geometry in $\ell$ and $S$ basis}
\label{fig:Pac-Man}
\end{figure}
The on shell calculation is straightforward: JT action in this geometry contains two parts: the Schwarzian action $\int(K-1)$ on the boundary and a corner contribution at the center: $S(\pi-\theta)$, where $\theta$ is the span angle at the horizon (Figure \ref{fig:Pac-Man}). The Schwarzian action simply gives $E u={S^2 u\over 2}$ by direct evaluation. We can determine $\theta$ from $u$ since they are related with redshift: $\theta=u S$. Therefore the corner term gives: $\pi S-S^2 u$. Adding them up then gives us the classical action.
We can also expand $S$ in terms of the $\ell$ basis, and relate $\Psi(\ell)$ with $\Psi(S)$ by a change of basis:
\begin{equation}
\langle\ell|S\rangle=\sqrt{\rho(S)}{4\over \ell}K_{2iS}({4\over \ell});~~~~~\Psi(E;\ell)=\int_0^{\infty} dS \langle \ell|S\rangle \langle S|E\rangle_G;
\end{equation}
Before discussing our last basis, we want to stress the simplicity of the wavefunction in $S$ basis (\ref{PsiS}) and the Gaussian factor resembles an ordinary particle wavefunction in momentum basis. We introduce our last basis $\eta$ as canonical conjugate variable of $S$, with an analog of going to position space of the particle picture in mind:
\begin{eqnarray}
|\eta\rangle &=&\int_0^{\infty} dS \cos(\eta S)|S\rangle;~~~~~~~~~~~~~\langle \ell|\eta\rangle=\int_0^{\infty} dS\cos(\eta S)\sqrt{\rho(S)}K_{2iS}({4\over \ell}){4\over \ell};~~~~~\\
\Psi(E,\eta) &=&\scaleto{\sqrt{\sinh(2\pi \sqrt{2E})\over 2\pi^2\sqrt{2E}}\cos(\eta \sqrt{2E})}{30pt};~~~~~\Psi(u,\eta)=\int_0^{\infty} dS \sqrt{\rho(S)}\cos(\eta S)e^{-{uS^2\over 2}}.~~~~
\end{eqnarray}
To understand the meaning of $\eta$ better, we can look at the classical behavior of $\Psi(u;\eta)$:
\begin{equation}
\Psi(u;\eta)\sim {1\over u}\exp[{\pi^2\over 2 u}-{\eta^2\over 2 u}]\left(e^{i\eta\pi\over u}\sqrt{\pi+i\eta}+e^{-{i\eta\pi\over u}}\sqrt{\pi-i\eta}\right)
\end{equation}
When $u$ is real, the wavefunction is concentrated at $\eta=0$ and has classical action of a half disk in the exponent.
When $u={\beta\over 2}+it$ which corresponds to the case of analytically continuing into Lorentzian signature, the density of the wavefunction $|\Psi(u,\eta)|^2$ is dominated by:
\begin{equation}
|\Psi({\beta\over 2}+it,\eta)|^2\sim {\sqrt{\pi^2+\eta^2}\over \beta^2+4t^2}\exp[{2\pi^2\over \beta}]\left(\exp[-{2\beta(\eta-{2\pi t\over \beta})^2\over \beta^2+4t^2}]+\exp[-{2\beta(\eta+{2\pi t\over \beta})^2\over \beta^2+4t^2}]\right)
\end{equation}
showing the fact that $\eta$ is peaked at the Rindler time ${2\pi t\over \beta}$. We can therefore think of fixing $\eta$ as fixing the IR time or the boost angle at the horizon.
The classical intuition for the boost angle is most clear in Euclidean geometry, where for fixed boundary proper time there can be different cusps at the horizon (Figure \ref{fig:Boostangle}).
\begin{figure}
\centering
\includegraphics[scale=0.4]{boostangle.pdf}
\caption{Euclidean geometries with different cusps.}
\label{fig:Boostangle}
\end{figure}
One application of those wavefunctions is that we can take an inner product and get the partition function.
However, there are also other ways to get the partition function.
For example, we can concatenate three propagators and integrate over their locations. This also gives the partition function by the composition rule of propagator.
By the relation between propagator and wavefunction, we can also view this as taking an inner product of three wavefunctions with an interior state as in figure \ref{fig:threewavefunction}, where the interior state can be understood as an entangled state for three universes.
To be more precise, we can view the wavefunction as the result of integrating the bulk up to the geodesics with zero extrinsic curvature. Then the interior state is given
by the area of the hyperbolic triangle in figure \ref{fig:threewavefunction}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{threewavefunction.pdf}
\caption{Partition function from inner product of three wavefunctions.}
\label{fig:threewavefunction}
\end{figure}
The path integral for the hyperbolic triangle (denoted as $I(\ell_{12},\ell_{23},\ell_{31})$, where $\ell_{ij}={|x_i-x_j|\over \sqrt{z_iz_j}}$), is a product of three phase factors, which satisfies a nontrivial equality (with ordering $x_1>x_2>x_3$):
\begin{equation}\label{GHZ}
\scaleto{ I(\ell_{12},\ell_{23},\ell_{31})=e^{-2({z_1+z_2\over x_1-x_2}+{z_2+z_3\over x_2-x_3}+{z_3+z_1\over x_3-x_1})}={16\over\pi^2}\int_0^{\infty}d\tau \tau \sinh (2\pi \tau) K_{2i\tau}({4\over \ell_{12}})K_{2i\tau}({4\over \ell_{23}})K_{2i\tau}({4\over \ell_{31}}).}{26pt}
\end{equation}
Recalling that the Bessel function represents the fixed energy ``microstate" $|\mathcal{E}\rangle$ (\ref{fixed energy wavefunction}) and ${\tau\over 2\pi^2} \sinh(2\pi \tau)$ is the density of state, this formula tells us that the interior state is a GHZ state for three universe:
\begin{equation}
I_{123}\sim\sum_n|\mathcal{E}_n\rangle_1|\mathcal{E}_n\rangle_2|\mathcal{E}_n\rangle_3.
\end{equation}
$I$ can also been viewed as a scattering amplitude from two universes into one universe. It constrains the SL(2,R) representation of the three wavefunctions to be the same.\footnote{Some thing similar happens for 2d Yang-Mills theory \cite{Witten:1992xu,Cordes:1994fc,Luca}.}
We can write down the partition function as:
\begin{equation}
Z_{JT}=\int_0^{\infty} \prod\limits_{\lbrace i j \rbrace \in \lbrace 12,23,31\rbrace}d\ell_{ij} \Psi(u_{12},\ell_{12})\Psi(u_{23},\ell_{23})\Psi(u_{31},\ell_{31})I_{\ell_{12},\ell_{23},\ell_{31}}.
\end{equation}
This same result also holds if we repeat the process $n$ times. It is interesting that we can view the full disk amplitude in these various ways.
One can also extend our analysis to include matter field. One type of such wavefunction can be created by inserting operator during Euclidean evolution, and is analysed in appendix B. Note that because of the SL(2,R) symmetry is a gauge symmetry, our final state has to be a gauge singlet including matter field.
\section{Correlation Functions in Quantum Gravity}
\subsection{Gravitational Feynman Diagram}
The propagator enables us to ``dress'' quantum field theory correlators to produce quantum gravity ones.
Namely, we imagine that we have some quantum field theory in $H_2$ and we compute correlation functions of
operators as we take the points close to the boundary where they take the form
\begin{equation}
\langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} = q^{ - \sum \Delta_i }z_1^{\Delta_1}..z_n^{\Delta_n}\langle O_1(x_1)...O_n(x_n)\rangle_{CFT}
\end{equation}
The factor of $q$ arises from \nref{yAndq}, and the last factor is simply defined as the function
that results after extracting the $z$ dependence. For example, for a two point function we get
\begin{equation}
\langle O_1(\boldsymbol{x_1})O_2(\boldsymbol{x_2})\rangle_{\rm QFT}=
q^{ - 2\Delta } z_1^{\Delta} z_2^{\Delta}{1\over |x_1-x_2|^{2\Delta}}.
\end{equation}
We can now use the propagator \nref{HeatLim} to couple the motion of the boundary and thus obtain the full
quantum gravity expression for the correlator. The factors of $q$ are absorbed as part of the renormalization
procedure for defining the full quantum gravity correlators.
In this way we obtain
\begin{figure*}[h!]
\centering
\subfigure[Witten Diagram]{%
\includegraphics[width=0.3\textwidth]{wittendiagram.pdf}%
\label{fig:Witten Diagram}%
}\hspace{4cm
\subfigure[Gravitational Feynman Diagram]{%
\includegraphics[width=0.3\textwidth]{dynamicalwittendiagram.pdf}%
\label{fig:Dynamical Witten Diagram}%
\caption{Summation of ${1\over N}$ effects fluctuates the boundary of Witten Diagram}
\label{fig:witten diagram}
\end{figure*}
\begin{eqnarray}\label{Higher Point Function}
\small{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{\rm QG}}}&=&e^{2\pi q}\int {\prod_{i=1}^n{dx_i dy_i\over y_i^2}\over \text{V(SL(2,R))}} G(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})G(u_{23},\boldsymbol{x_2},\boldsymbol{x_3})...G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})\times
\cr
& ~& \times \langle O_1(\boldsymbol{x_1})...O_n(\boldsymbol{x_n})\rangle_{\rm QFT} q^{ \sum_ i \Delta_i }
\end{eqnarray}
where the left hand side is the full quantum gravity correlator by definition. The last factor is the usual
renormalization necessary to get something finite.
The factor of $e^{ 2 \pi q } $ cancels with the $q$ dependent ``phase'' factors in \nref{heat kernel} to give one if we order
the points cyclically ($x_1>x_2...>x_n$). This requires that we define more carefully the last propagator
$G(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})$ as:
\begin{equation}
e^{-2\pi q}\tilde K(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})=e^{-2\pi q}e^{-2 \frac{z_n+z_1}{x_n-x_1}}\frac{2\sqrt{z_nz_1}}{\pi^2 |x_n-x_1|}\int_0^\infty ds s\sinh(2\pi s)e^{-\frac{s^2}{2} u_{n1}}K_{2is}(\frac{4\sqrt{z_nz_1}}{|x_n-x_1|}).
\end{equation}
The factor ${1\over \text{V(SL(2,R))}}$ in \nref{Higher Point Function} means that we should fix the $SL(2,R)$ gauge symmetry (Appendix \ref{gauge fix}).
In the end we can write down an expression where we have already taken the $q \to \infty $ limit
\begin{equation}\label{final formula}
\boxed{\scaleto{\langle O_1(u_1)...O_n(u_n)\rangle_{\text{QG}}=\int_{ x_1>x_2..>x_n}{\prod_{i=1}^n{dx_i dz_i}\over \text{V(SL(2,R))}} \tilde{K}(u_{12},\boldsymbol{x_1},\boldsymbol{x_2})...\tilde{K}(u_{n1},\boldsymbol{x_n},\boldsymbol{x_1})z_1^{\Delta_1-2}..z_n^{\Delta_n-2}\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}.}{26pt}}
\end{equation}
This is one of the main results of our paper and it gives a detailed expression for correlation function in 2 dimensional quantum gravity in terms of the correlation functions of the QFT in hyperbolic space, or $AdS_2$.
Notice that in usual $AdS/CFT$ the correlators $\langle O_1(x_1)...O_n(x_n)\rangle_{\rm CFT}$ are an approximation to the full answer. This is sometimes computed by Witten diagrams.
We get a better approximation by integrating over the metric fluctuations.
In this case, the non-trivial gravitational mode is captured by the boundary propagator. The formula \nref{final formula} includes all the effects of quantum gravity in the JT theory (in the Schwarzian limit).
The final diagrams consist of the Witten diagrams for the field theory in $AdS$ plus the propagators for the boundary particle and we can call them ``Gravitational Feynman Diagrams", see figure \ref{fig:witten diagram}.
\subsection{Two Point Function}
Using formula \nref{final formula}, we can study gravitational effects on bulk fields such as its two point function:
\begin{eqnarray}
\langle O_1(u)O_2(0)\rangle_{QG}=
\includegraphics[width=50mm,trim=0 9cm 0 0]{2ptfunction.pdf}\\
~~~~~~~~~~~\nonumber\\
~~~~~~~~~\nonumber
\end{eqnarray}
The explicit expression for $\langle O_1(u)O_2(0)\rangle_{QG}$ with dimension $\Delta$ at temperature ${1\over \beta}$ is
\footnote{We will just keep the $\Delta$ dependent constant
since at last we will normalized with respect of partition function which corresponds to set $\Delta=0$.}:
\begin{equation}
\scaleto{{1\over \text{V(SL(2,R))}}\int_{x_1>x_2} {dx_1 dx_2 dz_1 dz_2 \over z_1^2 z_2^2}\int_0^{\infty}ds_1 ds_2 \rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}K_{2is_1}({4\sqrt{z_1z_2}\over |x_1-x_2|})K_{2is_2}({4\sqrt{z_1z_2}\over |x_1-x_2|})({\sqrt{z_1z_2}\over |x_1-x_2|})^{2\Delta+2}}{25pt}.
\end{equation}
To fix the $SL(2,R)$ gauge, we can choose $z_1=z_2=1$ and $x_2=0$.
Then the integral over $H_2$ space is reduced to a single integral over $x_1$,
with a Jacobian factor $2x_1$ (Appendix \ref{gauge fix}):
\begin{equation}\label{space integral}
\int_0^{\infty}ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\int_0^{\infty} dx_1 ({1\over x_1})^{2\Delta+1}K_{2is_1}({4\over x_1})K_{2is_2}({4\over x_1}).
\end{equation}
the last integral can be interpreted as a matrix element of two point operator $O_1O_2$ between states $|E_1,\psi\rangle$ and $|E_2,\psi\rangle$,
where $|\psi\rangle$ is the wavefunction of quantum field theory
and $|E\rangle_G$ represents the fixed energy gravitational state.
Integrating over $x$ can be thought as integrating over a particular gravitational basis, and we can see that the gravity wavefunction
suppress the UV contributions from quantum field theory ($K_{2is}({4\over x})\sim \sqrt{\pi x\over 8}e^{-4/x}$ for $x\sim 0$).
The final expression for the two point function is:
\begin{eqnarray}\label{2point function}
\langle O_1(u) O_2(0)\rangle_{QG}={1\over \mathcal{N}}\int ds_1ds_2\rho(s_1)\rho(s_2)e^{-{s_1^2\over 2}u-{s_2^2\over 2}(\beta-u)}\frac{|\Gamma(\Delta-i(s_1+s_2))\Gamma(\Delta+i(s_1-s_2))|^2}{2^{2\Delta+1} \Gamma(2\Delta)};~~\\
={1\over \mathcal{N}}{\Gamma(2\Delta)\over u^{3/2} (\beta-u)^{3/2} 2^{4\Delta+4}\pi^3} \int_{c-i\infty}^{c+i\infty} d\theta_1d\theta_2 \theta_1\theta_2 e^{{\theta_1^2\over 2u}+{\theta_2^2\over 2 (\beta-u)}}{1\over (\cos{\theta_1\over 2}+\cos{\theta_2\over 2})^{2\Delta}}.~~~~~
\end{eqnarray}
In the second expression we write the integral in terms of variable $\theta$ using the second integral representation of the propagator (\ref{HeatLim2}).
The normalization constant can be determined by taking the $\Delta=0$ limit: $\mathcal{N}=Z_{JT}$.
If we contemplate the result (\ref{2point function}) a little bit, then we find that the two integrals of $s_1$ and $s_2$ just represent the spectral decomposition of the two point function. Indeed, under spectral decomposition we have $\langle O(u)O(0)\rangle=\sum\limits_{n,m} e^{-E_n u-E_m (\beta-u)}|\langle E_n|O|E_m\rangle|^2$. Compare with (\ref{2point function}), we can read out the square of matrix element of operator $O$:
\begin{equation}
\scaleto{_G\langle E_1|O_L O_R|E_2\rangle_G=\delta E^{-2}\sum\limits_{\substack{|E_n-E_1|<\delta E\\|E_m-E_2|<\delta E}}|\langle E_n|O|E_m\rangle|^2=\rho(E_1)\rho(E_2){|\Gamma(\Delta-i(\sqrt{2E_1}+\sqrt{2E_2}))\Gamma(\Delta+i(\sqrt{2E_1}-\sqrt{2E_2}))|^2\over 2^{2\Delta+1}\Gamma(2\Delta)}.}{38pt}
\end{equation}
Remember the notation is that $|E\rangle_G$ stands for a gravitational state with energy $E$ and $|E_n\rangle$ stands for one side microstate (\ref{EnergyStateDefinition}).
We have put the measure $\rho(E)={1\over 2\pi^2}\sinh(2\pi\sqrt{2 E})$ in the definition of matrix element for the reason that in gravity it is more natural to consider an average of energy states as a bulk state.
To understand this formula a little bit better, we can consider the classical limit, namely large $E$. In this limit the matrix element squared can be approximated as a nonanalytic function:
\begin{equation}
_G\langle E_1|O_L O_R|E_2\rangle_G\propto |E_1-E_2|^{2\Delta-1}e^{2\pi \text{min}(\sqrt{2E_1},\sqrt{2E_2})}.
\end{equation}
If we fix $E_1$ and varying $E_2$ from $0$ to infinity, the matrix element changes from $|E_1-E_2|^{2\Delta-1}\rho(E_2)$ to $|E_1-E_2|^{2\Delta-1}\rho(E_1)$ after $E_2$ cross $E_1$. We can understand this behavior qualitatively as a statistical effect: the mapping from energy subspace $E_1$ to $E_2$ by operator $O$ is surjective when the Hilbert space dimension of $E_2$ is less that $E_1$ and is injective otherwise.
Another understanding is the following: the two point function is finite in a fixed energy state $|E_n\rangle$, which means the following summation of intermediate states $|E_m\rangle$ is order one: $\sum\limits_{m}|\langle E_n|O|E_m\rangle|^2$. Looking at the case $E_m>E_n$, because of the density of states grows rapidly, the matrix element squared has to be proportional to ${1\over \rho(E_m)}$ to get a finite result. Multiplied by $\rho(E_n)\rho(E_m)$, we have $\rho(E_n)\rho(E_m)|\langle E_n|O|E_m\rangle|^2\sim \rho(\text{min}(E_n,E_m))$.
\subsection{ETH and the KMS condition}
The Eigenstate Thermalization Hypothesis (ETH) is a general expectation for chaotic system. It expresses that the operator expectation value in an energy eigenstate can be approximated by thermal expectation value with effective temperature determined from the energy. Such hypothesis can be tested with the knowledge of operator matrix elements. The two point function in microcanonical essemble is:
\begin{eqnarray}
{1\over \delta E}\sum\limits_{|E_n-E|<\delta E}\langle E_n|O(u) O(0)|E_n\rangle &=&
\rho(E)e^{uE}\langle E|O e^{-u H}O|E\rangle\nonumber\\
\langle E|O e^{-u H}O|E\rangle &=&\int_0^{\infty} ds \rho(s)e^{-{s^2\over 2}u}\scaleto{{|\Gamma(\Delta+i(s+\sqrt{2E}))|^2|\Gamma(\Delta+i(s-\sqrt{2E}))|^2\over 2^{2\Delta+1} \Gamma(2\Delta)}}{25pt}.~~~~~~~~~\label{TwoPFMicro}
\end{eqnarray}
Notice that $|E\rangle$ is not $|E\rangle_G$, the former represents a one side microstate, while the later is a gravitational state. Accordingly $ \langle E|O(u)O(0)|E\rangle$ stands for a two point function in a microstate.
To study ETH, we will consider the case of a heavy black hole $E={S^2\over 2}={2\pi^2 \over\beta^2}\gg 1$. From the discussion in last section, we know that the matrix element tries to concentrate $s$ around $\sqrt{2E}$ and thus we can approximate $\rho(E)\rho(s)|\Gamma(\Delta+i(s+S))|^2$ as proportional to $sS^{2\Delta-1}e^{\pi(s+S)}$. Using integral representation for $|\Gamma(\Delta+i(s-S))|^2 $ we derive the two point function in microcanonical essemble with energy $E$ is proportional to:
\begin{equation}
{\rho(E)S^{2\Delta-1}\over u^{3/2}}\int d\xi (\pi+i\xi)e^{-{2\over u}(\xi-i{\pi\over 2})^2-(\pi+2i\xi) S+u{S^2\over 2}}{1\over(\cosh \xi)^{2\Delta}}.
\end{equation}
The $\xi$ variable can be understood as the measure of time separation in units of effective temperature between two operators and its fluctuation represents the fluctuation of the effective temperature.
And the final integral can be understood as a statistical average of correlation functions with different temperatures.
If we put back the Newton Constant $G_N={1\over N}$, we have $S\sim N$ and $u\sim N^{-1}$ (\ref{BCond}).
As can be seen from the probability distribution, the fluctuation is of order ${1\over \sqrt{N}}$, and hence for large $N$ system we can use saddle point approximation:
\begin{equation}
\xi=i({\pi\over 2}-{S\over 2}u) -{u\Delta \tanh \xi\over 2}=i({\pi\over 2}-{\pi u\over \beta})-{u\Delta\tanh \xi\over 2}.
\end{equation}
The first piece gives the typical temperature of the external state,
while the last piece comes from the backreaction of operator on the geometry. If we first take the limit of large N, one simply get that the two point function in microcanonical essemble is the same as canonical essemble.
However, the euclidean correlator in canonical essemble is divergent as euclidean time approach to inverse temperature $\beta$ because of KMS condition.
Such singular behavior plays no role in the microcanonical essemble so is called a ``forbidden singularity" in ETH \cite{Fitzpatrick:2016ive, Faulkner:2017hll}. In our situation we can see directly how the forbidden singularity disappears in the microcanonical essemble.
When $\xi$ approach $-i{\pi\over 2}$ at the forbidden singularity the backreaction on the geometry becomes large and hence the effective temperature becomes lower:
\begin{equation}
{2\pi\over\beta*}\rightarrow {2\pi\over\beta}-{\Delta \over \pi-{\pi u\over \beta}}.
\end{equation}
At the time ${\beta-u\over\beta}\sim{\Delta\over N}$, the backreaction is important and we expect to see deviation from thermal correlators.
Therefore the correlation function in microcanonical essemble will never have singularity away from coincide point.
\subsection{Three Point Function}
The bulk diagram of the three point function will be like Figure \ref{fig:threewavefunction} with additional operator inserting at the intersection points (See Figure \ref{fig:threepointfunction}).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{threepointfunction.pdf}
\caption{Bulk Diagram for Three Point Function}
\label{fig:threepointfunction}
\end{figure}
The QFT three point correlation function in $AdS_2$ is fixed by conformal symmetry and we can write it down as:
\begin{equation}
\langle O_1 O_2 O_3\rangle=C_{123}{z_1^{\Delta_1} z_2^{\Delta_2} z_3^{\Delta_3}\over |x_{12}|^{\Delta_1+\Delta_2-\Delta_3}|x_{23}|^{\Delta_2+\Delta_3-\Delta_1}|x_{13}|^{\Delta_1+\Delta_3-\Delta_2}}={C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
$\Delta_i$ is the conformal dimension of $O_i$. Putting them in formula \ref{final formula} and rewrite the propagator in terms of the wavefunction (\ref{wavefunction}), we have the quantum gravitational three point function:
\begin{equation}
\langle O_1 O_2 O_3\rangle_{QG}=\int_{x_1>x_2>x_3}{\prod_{i=1}^3{dx_i dz_i}\over \text{V(SL(2,R))}}\Psi_{u_{12},\ell_{12}}\Psi_{u_{23},\ell_{23}}\Psi_{u_{13},\ell_{13}}I_{\ell_{12},\ell_{23},\ell_{13}}{C_{123}\over \ell_{12}^{\Delta_1+\Delta_2-\Delta_3}\ell_{23}^{\Delta_2+\Delta_3-\Delta_1}\ell_{13}^{\Delta_1+\Delta_3-\Delta_2}}.
\end{equation}
We can view this expression as an inner product of three universe wavefunction with the interior, inserting three bilocal operators $\tilde O_{ij;k}\tilde O_{ij;k}$ with dimension $\tilde\Delta_{ij;k}={1\over 2}(\Delta_i+\Delta_j-\Delta_k)$ between them. One can fix the $SL(2,R)$ symmetry and express the integral in terms of $\ell_{ij}$, it is the same exercise as in open string calculation to find the Jacobian factor (Appendix \ref{gauge fix}). Here we can just argue that in order to get the partition function at $\Delta=0$, the measure has to be flat. Therefore the three point function factorizes into form:
\begin{equation}
\small{\langle O_1(u_{1})O_2(u_{2})O_3(u_{3})\rangle_{\text{\rm QG}}}\propto C_{123}\int_0^{\infty}d\tau \rho(\tau)\mathcal{I}_{\tau}(u_{12},\tilde\Delta_{12;3})\mathcal{I}_{\tau}(u_{23},\tilde\Delta_{23;1})\mathcal{I}_{\tau}(u_{31},\tilde\Delta_{31;2})
\end{equation}
while $\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})$ is an integral of $\ell_{ij}$ which gives the two point function in microstate $E_{\tau}$ (\ref{TwoPFMicro}) with the $e^{uE_{\tau}}$ factor stripped off:
\begin{equation}
\mathcal{I}_{\tau}(u_{ij},\tilde\Delta_{ij;k})={1\over 2}\int_{0}^{\infty} d\ell_{ij}\Psi_{u_{ij},\ell_{ij}}{1\over \ell_{ij}^{\Delta_i+\Delta_j-\Delta_k}}K_{2i\tau}({4\over \ell_{ij}})=\langle E_{\tau} |\tilde O_{ij;k}e^{-u_{ij}H}\tilde O_{ij;k}|E_{\tau}\rangle;~~~~~E_{\tau}={\tau^2\over 2}.
\end{equation}
Again the normalization constant can be fixed by choosing $O_i$ to be identity.
\subsection{Einstein-Rosen Bridge}
The Einstein-Rosen Bridge in a classical wormhole keeps growing linearly with time and this behavior was conjectured to related with the growth of computational complexity of the dual quantum state \cite{Stanford:2014jda}.
Based on the universal behavior of complexity growth, Susskind proposed a gravitational conjecture in a recent paper \cite{Susskind:2018fmx} about the limitation of classical general relativity description of black hole interior. The conjecture was stated as follows:
\textsl{Classical general relativity governs the behavior of an ERB for as long as possible.}
In this section, we will test this conjecture using the exact quantum wavefunction of JT gravity (\ref{wavefunction}). We will in particular focusing on the behavior of ERB at time bigger than $1$. The size of Einstein-Rosen Bridge $\mathcal{V}$ in two dimensions is the geodesic distance $d$ between two boundaries, and can be calculated in thermofield double state $|u\rangle$ as:
\begin{equation}
\mathcal{V}=\langle u|d|u\rangle
\end{equation}
We want to focus on the dependence of volume on Lorentzian time evolution. Therefore we do analytic continuation of $u$ in Lorentzian time: $u={\beta\over 2}+it$. Using the WdW wavefunction in $\ell$ basis (\ref{wavefunction}) and the relation between $d$ and $\ell$, we can calculate the expectation value exactly. This can be done by taking the derivative of the two point function (\ref{2point function}) with respect to $\Delta$ at $\Delta=0$. Using the integral representation for $|\Gamma(\Delta+i(s_1-s_2))|^2$, the only time dependence of volume is given by:
\begin{equation}
\mathcal{V}(t)={1\over \mathcal{N}}\int_{-\infty}^{\infty}d\xi\int_0^{\infty} ds_1ds_2\rho_1\rho_2 e^{i(s_1-s_2)\xi-i(s_1-s_2){(s_1+s_2) \over 2}t-{\beta\over 4}(s_1^2+s_2^2)}\log(2\cosh{\xi\over 2})|\Gamma(is_1+is_2)|^2.
\end{equation}
The limit we are interested in is $\beta\ll 1\ll t$ \footnote{Remember that we are measuring time in units of $\phi_r$ (\ref{BCond}), so time order $1$ is a quantum gravity region.}, in which case the integral has a saddle point at \footnote{Actually this saddle point is valid for any range of $t$ as long as $\beta\ll 1$.}:
\begin{equation}\label{SaddlePoint}
s_{1,2}={2\pi\over \beta};~~~~~~~~~\xi={2\pi t\over \beta}.
\end{equation}
Therefore the volume has linear dependence in time:
\begin{equation}
\mathcal{V}(t)\sim{2\pi t\over \beta}.
\end{equation}
Using the complexity equal to volume conjecture \cite{Stanford:2014jda,Brown:2015bva}, the complexity of thermofield double state is proportional to the maximum volume:
\begin{equation}
\mathcal{C}(t)=\#\mathcal{V}(t)=\#{2\pi t \over\beta} .
\end{equation}
The proportionality constant is suggested in \cite{Brown:2018kvn} to be $S_0$ based on classical calculation of near extremal black hole. This, however, is not very clear in our model since $S_0$ is the coupling constant of the pure Einstein-Hilbert action and decouples with JT theory (\ref{full action}).
Since the saddle point (\ref{SaddlePoint}) is actually valid from early time to late time, the proportionality constant can be fixed at classical level and once we fix it we can conclude that \textsf{the length of Einstein-Rosen Bridge (or complexity of the state) keeps linearly growing even considering quantum gravity effects in JT theory.}
We want to comment that this is not an obvious result that one can expect from classical observables.
For example, one might argue that we can extract the information of the ERB from two sided correlators for the reason that semiclassically we can approximate the correlator as $e^{-m d}$.
Therefore one can conclude the ERB has linear growth from the quasinormal behavior of the correlator.
However such observables can only give us information of ERB up to time order $1$, which is the same time scale we can trust the classical general relativity calculation.
After that the correlation function changes from exponential decay into universal power law decay ${1\over t^{3}}$ as one can directly derive from analytic continuation of result (\ref{2point function}).
If we still use such correlator to extract information about ERB we would get the wrong conclusion that it stops its linear growth after time order $1$.
The reason why it is incorrect is that at this time scale the operator disturbs the state and causes different energy states interfere each other strongly. It is simply that the correlator can no longer be described by the classical geometry, rather than the interior stops to behave classically.
From our calculation, we see that if we probe of the state in a weaker and weaker way, we are still able to see the classical geometry.
Lastly, we want to talk a little about when JT gravity needs to be modified. A naive estimate can be made from the partition function that when $\beta$ approaches $e^{{2\over 3}S_0}$, the partition function becomes less than one and definitely at this time scale we need new physics. A recently study of gravitational physics at this time scale was discussed in \cite{Saad:2018bqo}.
\section{Conclusion}
Our result gives an explicit formula (\ref{final formula}) to calculate all order corrections to correlation functions from quantum gravity in two dimensions.
The formula can be understood diagrammatically and we call it Gravitational Feynman Diagram.
We also give the exact Wheeler-DeWitt wavefunction and discuss the growth of its complexity quantum mechanically.
Although we are focusing on theoretical description of two dimensional black holes, the near-extremal black holes in nature should contain these features.
Both Reissner–Nordström black holes and Kerr black holes have an $AdS_2$ throat near their extremality.
For those black holes, the gravitational effects are enhanced by the their near extremal entropies (the coupling constant is $\phi_h$ rather than $\phi_0+\phi_h$) and therefore are better backgrounds to test gravitational effects.
We should however point out that the observational black holes all have large near extremal entropies and thus are very classical \cite{Preskill:1991tb}.
In addition, the Thorne limit of Kerr black hole sets a lower bound on the near extremal entropy in nature.
But for the Primordial black holes in early universe, our story might play a role and it will be interesting to study the physical consequence in that situation.
Another application of our result is to connect with SYK type models \cite{Maldacena:2016hyu,Kitaev:2017awl,Witten:2016iux,Klebanov:2016xxf} since those models all have an emergent Schwarzian action at low energy. On that account, the exact Schwarzian quantization can be used to test ${1\over N}$ corrections to those models. For example, one can try to directly test the two point function with SYK numerics \cite{BKobrin} or one can use SYK models to understand the microscopic description of WdW wavefunction and its complexity. It might also be interesting to consider the black hole information paradox \cite{Fiola:1994ir} and late time traversable wormholes including the quantum fluctuations of the boundary \cite{Gao:2016bin,Maldacena:2017axo,Susskind:2017nto,Yoshida:2017non, Maldacena:2018lmt}.
{\bf Acknowledgments }
The author wants to give special thanks to J.Maldacena and D.Stanford for patient guidance and help all through the project.
We also want to thank A.Almheiri, L.Iliesiu, J.Jiang, J.Ripley, A.Kitaev, B.Kobrin, R.Mahajan, A.M.Polyakov, S.J.Suh, G.Turiaci and H.Verlinde for discussions.
Z.Y is supported by Charlotte Elizabeth Procter Fellowship from Princeton University.
|
\section{Introduction}
Average consensus is a fundamental problem in distributed computing and multi-agent systems. It comes up in many real world applications such as coordination of autonomous agents, estimation, rumour spreading in social networks, PageRank and distributed data fusion on ad-hoc networks and decentralized optimization. Due to its great importance there is much classical \cite{tsitsiklis1986distributed,degroot1974reaching} and recent \cite{xiao2005scheme, xiao2004fast, boyd2006randomized} work on the design of efficient algorithms/protocols for solving it.
One of the most attractive classes of protocols for solving the average consensus are gossip algorithms. The development and design of gossip algorithms was studied extensively in the last decade. The seminal 2006 paper of Boyd et al.\ \cite{boyd2006randomized} on randomized gossip algorithms motivated a fury of subsequent research and now gossip algorithms appear in many applications, including distributed data fusion in sensor networks \cite{xiao2005scheme}, load balancing \cite{cybenko1989dynamic} and clock synchronization \cite{freris2012fast}. For a survey of selected relevant work prior to 2010, we refer the reader to the work of Dimakis et al.\ \cite{dimakis2010gossip}. For more recent results on randomized gossip algorithms we suggest \cite{zouzias2015randomized, liu2013analysis,olshevsky2014linear,LoizouRichtarik, nedic2018network, aybat2017decentralized}.
See also \cite{dimakis2008geographic, aysal2009broadcast, olshevsky2009convergence,hanzely2017privacy}.
The main goal in the design of gossip protocols is for the computation and communication to be done as quickly and efficiently as possible. In this work, our focus is precisely this. We design randomized gossip protocols which converge to consensus fast.
\subsection{The average consensus problem} \label{sec:ACP}
In the average consensus (AC) problem we are given an undirected connected network $\cG=(\cV,\cE)$ with node set $\cV=\{1,2,\dots,n\}$ and edges $\cE$. Each node $i \in \cV$ ``knows'' a private value $c_i \in \mathcal{R}$. The goal of AC is for every node to compute the average of these private values, $\bar{c}:=\tfrac{1}{n}\sum_i c_i$, in a distributed fashion. That is, the exchange of information can only occur between connected nodes (neighbors).
\subsection{Main Contributions}
We present a new class of randomized gossip protocols where in each iteration all nodes of the network update their values but only a subset of them exchange their private information. Our protocols are based on recently proposed ideas for the acceleration of randomized Kaczmarz methods for solving consistent linear systems \cite{loizou2017momentum} where the addition of a momentum term was shown to provide practical speedups over the vanilla Kaczmarz methods. Further, we explain the connection between gossip algorithms for solving the average consensus problem, Kaczmarz-type methods for solving consistent linear systems, and stochastic gradient descent and stochastic heavy ball methods for solving stochastic optimization problems. We show that essentially all these algorithms behave as gossip algorithms. Finally, we explain in detail the gossip nature of two recently proposed fast Kacmzarz-type methods: the randomized Kacmzarz with momentum (mRK), and its block variant, the randomized block Kaczmarz with momentum (mRBK). We present a detailed comparison of our proposed gossip protocols with existing popular randomized gossip protocols and through numerical experiments we show the benefits of our methods.
\subsection{Structure of the paper}
This work is organized as follows. Section~\ref{background} introduces the important technical preliminaries and the necessary background for understanding of our methods. A new connection between gossip algorithms, Kaczmarz methods for solving linear systems and stochastic gradient descent (SGD) for solving stochastic optimization problems is also described. In Section~\ref{theProtocols} the two new accelerated gossip protocols are presented. Details of their behaviour and performance are also explained. Numerical evaluation of the new gossip protocols is presented in Section~\ref{experiments}.
Finally, concluding remarks are given in Section~\ref{conclusion}.
\subsection{Notation}
The following notational conventions are used in this paper. We write $[n]:= \{1,2, \dots ,n\}$. Boldface upper-case letters denote matrices; $\mathbf{I}$ is the identity matrix.
By $\cL$ we denote the solution set of the linear system $\mathbf{A} x=b$, where $\mathbf{A} \in \mathcal{R}^{m\times n}$ and $b\in \mathcal{R}^m$. Throughout the paper, $x^*$ is the projection of $x^0$ onto $\cL$ (that is, $x^*$ is the solution of the best approximation problem; see equation \eqref{best approximation}).
An explicit formula for the projection of $x$ onto set $\cL$ is given by
\begin{equation*}
\Pi_{\cL}(x):= \arg\min_{x' \in \cL} \|x'-x\| =x-\mathbf{A}^\top (\mathbf{A}\bA ^ \top )^\dagger (\mathbf{A} x-b).
\end{equation*}
A matrix that often appears in our update rules is
\begin{equation}
\label{MatrixH}
{\bf H} := {\bf S} ({\bf S}^\top {\bf A}\mA^\top {\bf S})^\dagger {\bf S}^\top,
\end{equation}
where ${\bf S} \in \mathcal{R}^{m \times q}$ is a random matrix drawn in each step of the proposed methods from a given distribution ${\cal D}$, and $\dagger$ denotes the Moore-Penrose pseudoinverse. Note that
${\bf H}$ is a random symmetric positive semi-definite matrix.
In the convergence analysis we use $\lambda_{\min}^+$ to indicate the smallest nonzero eigenvalue, and $\lambda_{\max}$ for the largest eigenvalue of matrix $\mathbf{W}=\mathbb{E}[\mathbf{A}^\top \mathbf{H} \mathbf{A}]$, where the expectation is taken over $\mathbf{S}\sim {\cal D}$. Finally, $x^k = (x^k_1,\dots,x^k_n) \in \mathcal{R}^n$ represents the vector with the private values of the $n$ nodes of the network at the $k^{th}$ iteration while with $x_i^{k}$ we denote the value of node $i \in [n]$ at the $k^{th}$ iteration.
\section{Background-Technical Preliminaries}
\label{background}
Our work is closely related to two recent papers. In \cite{LoizouRichtarik}, a new perspective on randomized gossip algorithms is presented. In particular, a new approach for the design and analysis of randomized gossip algorithms is proposed and it was shown how the Randomized Kaczmarz and Randomized Block Kaczmarz, popular methods for solving linear systems, work as gossip algorithms when applied to a special system encoding the underlying network. In \cite{loizou2017momentum}, several classes of stochastic optimization algorithms enriched with {\em heavy ball momentum} were analyzed. Among the methods studied are: stochastic gradient descent, stochastic Newton, stochastic proximal point and stochastic dual subspace ascent.
In the rest of this section we present the main results of the above papers, highlighting several connections. These results will be later used for the development of the new randomized gossip protocols.
\subsection{Kaczmarz Methods and Gossip Algorithms}
Kaczmarz-type methods are very popular for solving linear systems $\mathbf{A} x =b$ with many equations. The (deterministic) Kaczmarz method for solving consistent linear systems was originally introduced by Kaczmarz in 1937 \cite{1937angenaherte}. Despite the fact that a large volume of papers was written on the topic, the first provably linearly convergent variant of the Kaczmarz method---the randomized Kaczmarz Method (RK)---was developed more than 70 years later, by Strohmer and Vershynin \cite{RK}. This result sparked renewed interest in design of randomized methods for solving linear systems \cite{needell2010randomized, RBK, eldar2011acceleration, MaConvergence15, zouzias2013randomized, l2015randomized, schopfer2016linear, liu2016accelerated}. More recently, Gower and Richt\'{a}rik \cite{gower2015randomized} provide a unified analysis for several randomized iterative methods for solving linear systems using a sketch-and-project framework. We adopt this framework in this paper.
In particular, the sketch-and-project algorithm \cite{gower2015randomized} for solving the consistent linear system $\mathbf{A} x= b$ has the form
\begin{eqnarray}
\label{sketchproject}
x^{k+1} &=& x^k -\mathbf{A}^\top \mathbf{S}_k (\mathbf{S}_k^\top \mathbf{A} \mathbf{A}^\top \mathbf{S}_k)^\dagger \mathbf{S}_k^\top (\mathbf{A} x^k-b) \notag\\
&\overset{\eqref{MatrixH}}{=}& x^k -\mathbf{A}^\top \mathbf{H}_k (\mathbf{A} x^k-b),
\end{eqnarray}
where in each iteration matrix ${\bf S}_k$ is sampled afresh from an arbitrary distribution ${\cal D}$. In \cite{gower2015randomized} it was shown that many popular algorithms for solving linear systems, including RK method and randomized coordinate descent method can be cast as special cases of the above update by choosing\footnote{In order to recover a randomized coordinate descent method, one also needs to perform projections with respect to a more general Euclidean norm. However, for simplicity, in this work we only consider the standard Euclidean norm.} an appropriate distribution ${\cal D}$. The special cases that we are interested in are the
randomized Kaczmarz (RK) and its block variant, the randomized block Kaczmarz (RBK).
Let $e_i \in \mathcal{R}^m$ be the $i^{\text{th}}$ unit coordinate vector in $ \mathcal{R}^m$ and let $\mathbf{I}_{:C}$ be column submatrix of the $m \times m$ identity matrix with columns indexed by $C\subseteq [m]$. Then RK and RBK methods can be obtained as special cases of the update rule \eqref{sketchproject} as follows:
\begin{itemize}
\item RK: Let $\mathbf{S}_k=e_i$, where $i=i_k$ is chosen in each iteration independently, with probability $p_i>0$. In this setup the update rule \eqref{sketchproject} simplifies to
\begin{equation}
\label{RK}
x^{k+1}=x^k - \tfrac{\mathbf{A}_{i :} x^k -b_{i}}{\|\mathbf{A}_{i :}\|_2^2} \mathbf{A}_{i :}^ \top .
\end{equation}
\item RBK: Let $\mathbf{S}=\mathbf{I}_{:C}$, where $C=C_k$ is chosen in each iteration independently, with probability $p_C\geq 0$. In this setup the update rule \eqref{sketchproject} simplifies to
\begin{equation}
\label{RBK}
x^{k+1}=x^k - \mathbf{A}_{C:}^\top (\mathbf{A}_{C:}\mathbf{A}_{C:}^\top)^\dagger (\mathbf{A}_{C:}x^k-b_C).
\end{equation}
\end{itemize}
In this paper we are interested in two particular extension of the above methods: the randomized Kaczmarz method with momentum (mRK) and its block variant, the randomized block Kaczmarz with momentum (mRBK), both proposed and analyzed in \cite{loizou2017momentum}. Before we describe these two algorithms, let us summarize the main connections between the Kaczmarz methods for solving linear systems and gossip algorithms, as presented in \cite{LoizouRichtarik}.
In \cite{gower2015stochastic, ASDA, loizou2017momentum}, it was shown that even in the case of consistent linear systems with {\em multiple} solutions, Kaczmarz-type methods converge linearly to one particular solution: the projection of the initial iterate $x^0$ onto the solution set of the linear system. This naturally leads to the formulation of the {\em best approximation problem}:
\begin{equation}
\label{best approximation}
\min_{x = (x_1,\dots, x_n) \in \mathcal{R}^n} \tfrac{1}{2} \|x-x^0\|^2
\quad \text{subject to} \quad \mathbf{A} x = b.
\end{equation}
Above, $\mathbf{A}\in \mathcal{R}^{m\times n}$ and $\|\cdot\|$ is the standard Euclidean norm. By $x^*=\Pi_{\cL}(x^0)$ we denote the solution of \eqref{best approximation}.
In \cite{LoizouRichtarik} it was shown how RK and RBK work as gossip algorithms when applied to a special linear system encoding the underlying network.
\begin{defn}[\cite{LoizouRichtarik}] A linear system $\mathbf{A} x = b$ is called ``average consensus (AC) system'' when $\mathbf{A} x = b$ is equivalent to saying that $x_i = x_j$ for all $(i,j) \in \cE$.
\end{defn}
Note that many linear systems satisfy the above definition. For example, we can choose $b=0$ and $\mathbf{A} \in \mathcal{R}^{|\cE| \times n}$ to be the incidence matrix of $\cG$. In this case, the row of the system corresponding to edge $(i,j)$ directly encodes the constraint $x_i=x_j$.
A different choice is to pick $b=0$ and ${\bf A} ={\bf L}$, where ${\bf L}$ is the Laplacian of $\cG$. Note that depending on what AC system is used, RK and RBK have different interpretations as gossip protocols.
From now on we work with the AC system described in the first example. Since $b=0$, the general sketch-and-project update rule \eqref{sketchproject} simplifies to:
\begin{equation}
\label{updateSkProj}
x^{k+1}= \left[ \mathbf{I}- \mathbf{A}^\top \mathbf{H}_k\mathbf{A} \right] x^k.
\end{equation}
The convergence performance of RK and RBK for solving the best approximation problem (and as a result the average consensus problem) is described by the following theorem.
\begin{thm}[\cite{gower2015randomized,gower2015stochastic}]
Let $\{x^k\}$ be the iterates produced by \eqref{sketchproject}. Then
$\mathbb{E}[\|x^k-x^*\|^2]\leq \rho^k \|x^0-x^*\|^2,$
where $x^*$ is the solution of \eqref{best approximation},
$\rho := 1 - \lambda_{\min}^+ \in [0,1]$, and
$\lambda_{\min}^+$ denotes the minimum nonzero eigenvalue of $\mathbf{W}:= \mathbb{E}[\mathbf{A}^\top \mathbf{H} \mathbf{A}]$.
\end{thm}
In \cite{LoizouRichtarik}, the behavior of both RK and RBK as gossip algorithms was described, and a comparison with the convergence results of existing randomized gossip protocols was made. In particular, it was shown that the most basic randomized gossip algorithm \cite{boyd2006randomized} (``randomly pick an edge $(i,j)\in \cE$ and then replace the values stored at vertices $i$ and $j$ by their average'') is an instance of RK applied to the linear system $\mathbf{A} x=0 $, where the $\mathbf{A}$ is the incidence matrix of $\cG$. RBK can also be interpreted as a gossip algorithm:
\begin{thm}[\cite{LoizouRichtarik}, RBK as a Gossip Algorithm]
\label{TheoremRBK}
Each iteration of RBK for solving $\mathbf{A} x=0$ works as follows:
1) Select a random set of edges $\cS \subseteq \cE$,
2) Form subgraph $\cG_k$ of $\cG$ from the selected edges,
3) For each connected component of $\cG_k$, replace node values with their average.
\end{thm}
\subsection{The Heavy Ball momentum}
A detailed study of several (equivalent) {\em stochastic reformulations} of consistent linear systems was developed in \cite{ASDA}. This new viewpoint facilitated the development and analysis of relaxed variants (with relaxation parameter $\omega \in (0,2)$) of the sketch-and-project update \eqref{sketchproject}. In particular, one of the reformulations is the {\em stochastic optimization} problem
\begin{equation}
\label{stoch_reform}
\min_{x\in \mathcal{R}^n} f(x) := \mathbb{E}_{{\bf S}\sim {\cal D}}[f_{\bf S}(x)], \quad \text{where}
\end{equation}
\begin{equation}
\label{eq:f_s}
f_{{\bf S}}(x) := \tfrac{1}{2}\|{\bf A} x - b\|_{{\bf H}}^2 = \tfrac{1}{2}({\bf A} x - b)^\top {\bf H} ({\bf A} x - b),
\end{equation}
and ${\bf H}$ is the random symmetric positive semi-definite matrix defined in \eqref{MatrixH}.
Under certain (weak) condition on ${\cal D}$, the set of minimizers of $f$ is identical to the set of the solutions of the linear system. In \cite{ASDA}, problem \eqref{stoch_reform} was solved via Stochastic Gradient Descent (SGD):
\begin{equation}
\label{SGD}
x^{k+1}=x^k-\omega \nabla f_{\mathbf{S}_k}(x^k),
\end{equation}
and a linear rate of convergence was proved despite the fact that $f$ is not necessarily strongly convex and that a fixed stepsize $\omega>0$ is used. Observe that the gradient of the stochastic function \eqref{eq:f_s} is given by
\begin{equation}
\label{stochGradi}
\nabla f_{{\bf S}_k} (x) \overset{\eqref{eq:f_s}}{=} {\bf A}^\top {\bf H}_k ({\bf A} x - b).
\end{equation}
and as a result, it is easy to see that for $\omega=1$, the SGD update \eqref{SGD} reduces to the sketch-and-project update~\eqref{sketchproject}.
The recent works \cite{loizou2017linearly,loizou2017momentum} analyze momentum variants of SGD, with the goal to accelerate the convergence of the method for solving problem \eqref{stoch_reform}. SGD with momentum---also known as the stochastic heavy ball method (SHB)---is a well known algorithm in the optimization literature for solving stochastic optimization problems, and it is extremely popular in areas such as deep learning \cite{sutskever2013importance, szegedy2015going, krizhevsky2012imagenet, wilson2017marginal}. However, even though SHB is used extensively in practice, its theoretical convergence behavior is not well understood. To the best of our knowledge, \cite{loizou2017linearly,loizou2017momentum} are the first that prove linear convergence of SHB in any setting.
The update rule of SHB for solving problem \eqref{stoch_reform} is formally presented in the following algorithm:
\begin{algorithm}[H]
\caption{Stochastic Heavy Ball (SHB) }
\label{alg:SARAHHB}
\small \small
\begin{algorithmic}[1]
\State {\bf Parameters:} Distribution $\mathcal{D}$ from which method samples matrices; stepsize/relaxation parameter $\omega \in \mathcal{R}$; momentum parameter $\beta$.
\State {\bf Initialize:} $x^0,x^1 \in \mathcal{R}^n$
\For{$k=1,2,\dots$}
\State Draw a fresh $\mathbf{S}_k \sim {\cal D}$
\State Set
$x^{k+1}=x^k-\omega \nabla f_{\mathbf{S}_k}(x^k)+\beta (x^k-x^{k-1})
$
\EndFor
\State {\bf Output:} The last iterate $x^k$
\end{algorithmic}
\end{algorithm}
Using the expression for the stochastic gradient \eqref{stochGradi}, the update rule of SHB can be written more explicitly:
\begin{equation}
\label{SPmomentum}
x^{k+1}=x^k -\omega \mathbf{A}^\top \mathbf{H}_k(\mathbf{A} x^k-b) + \beta(x^k - x^{k-1}).
\end{equation}
Using the same choice of distribution ${\cal D}$ as in equation \eqref{RK} and \eqref{RBK}, we now obtain momentum variants of RK and RBK:
\begin{itemize}
\item RK with momentum (mRK): $$x^{k+1}=x^k -\omega \tfrac{\mathbf{A}_{i :} x^k -b_{i}}{\|\mathbf{A}_{i :}\|_2^2} \mathbf{A}_{i :}^ \top + \beta(x^k - x^{k-1}) $$
\item RBK with momentum (mRBK):
$$x^{k+1}=x^k -\omega \mathbf{A}_{C:}^\top (\mathbf{A}_{C:}\mathbf{A}_{C:}^\top)^\dagger (\mathbf{A}_{C:}x^k-b_C)+ \beta(x^k - x^{k-1}) $$
\end{itemize}
In \cite{loizou2017momentum}, two main theoretical results describing the behavior of SHB (and as a result also the special cases mRK and mRBK) were presented:
\begin{thm}[Theorem 1, \cite{loizou2017momentum}]
\label{L2}
Choose $x^0= x^1\in \mathcal{R}^n$. Let $\{x^k\}_{k=0}^\infty$ be the sequence of random iterates produced by SHB. Let $\lambda_{\min}^+$ (resp,\ $\lambda_{\max}$) be the smallest nonzero (resp.\ largest) eigenvalue of $\mathbf{W}$. Assume $0< \omega < 2$ and $\beta \geq 0$ and that the expressions
$a_1 := 1+3\beta+2\beta^2 - (\omega(2-\omega) +\omega\beta)\lambda_{\min}^+$ and
$a_2 := \beta +2\beta^2 + \omega \beta \lambda_{\max}$
satisfy $a_1+a_2<1$. Then
\begin{equation}\label{eq:nfiug582}\mathbb{E}[\|x^{k}-x^*\|^2] \leq q^k (1+\delta) \|x^0-x^*\|^2,
\end{equation}
and
$\mathbb{E}[f(x^k)] \leq q^k \tfrac{\lambda_{\max}}{2} (1+\delta) \|x^{0}-x^*\|^2,$
where $q=\tfrac{1}{2} (a_1+\sqrt{a_1^2+4a_2})$ and $\delta=q-a_1$. Moreover, $a_1+a_2 \leq q <1$.
\end{thm}
\begin{thm}[Theorem 4, \cite{loizou2017momentum}]
\label{theoremheavyball}
Let $\{x^k\}_{k=0}^{\infty}$ be the sequence of random iterates produced by SHB, started with $x^0= x^1\in \mathcal{R}^n$, with relaxation parameter (stepsize) $0<\omega \leq1/\lambda_{\max}$ and momentum parameter $(1-\sqrt{\omega \lambda_{\min}^+})^2 < \beta <1$. Let $x^* = \Pi^{\bf B}_{\cL}(x^0)$. Then there exists a constant $C >0$ such that for all $k\geq0$ we have
$\|\mathbb{E}[x^{k} -x^*]\|^2 \leq \beta^k C.$
\end{thm}
Using Theorem~\ref{theoremheavyball} and by a proper combination of the stepsize $\omega$ and the momentum parameter $\beta$, SHB enjoys an accelerated linear convergence rate in mean, \cite{loizou2017momentum}.
\begin{cor}
(i) If $ \omega= 1$ and $\beta= (1- \sqrt{0.99 \lambda_{\min}^+}) ^2$, then the iteration complexity of SHB becomes: $ \tilde{O}(\sqrt{1/ \lambda_{\min}^+})$.\\
(ii) If $ \omega= 1/\lambda_{\max}$ and $\beta= (1- \sqrt{ 0.99\lambda_{\min}^+/ \lambda_{\max}})^2$, then the iteration complexity of SHB becomes: $\tilde{O}(\sqrt{\lambda_{\max}/ \lambda_{\min}^+})$.
\end{cor}
\section{Randomized Gossip protocols with momentum}
\label{theProtocols}
Having presented SHB for solving the stochastic optimization problem \eqref{stoch_reform} and describing its sketch-and-project nature \eqref{SPmomentum}, let us now describe its behavior as a randomized gossip protocol when applied to solving the AC system $\mathbf{A} x=0$, where $\mathbf{A}\in |\cE| \times n$ is the incidence matrix of the network.
Since $b=0$, method \eqref{SPmomentum} can be simplified to:
\begin{equation}
\label{momentumupdateb0}
x^{k+1}=\left[\mathbf{I} -\omega \mathbf{A}^\top \mathbf{H}_k \mathbf{A}\right] x^k + \beta(x^k - x^{k-1}).
\end{equation}
In the rest of this section we focus on two special cases of \eqref{momentumupdateb0}: RK with momentum and RBK with momentum.
\subsection{Randomized Kaczmarz Gossip with momentum}
When RK is applied to solve an AC system $\mathbf{A} x=0$, one recovers the famous pairwise gossip algorithm~\cite{boyd2006randomized}. Algorithm~\ref{RKmomentum} describes how the relaxed variant of randomized Kaczmarz with momentum behaves as a gossip algorithm. See also Figure~\eqref{fig:mRK} for a graphical illustration of the method.
\begin{algorithm}[t!]
\caption{mRK: Randomized Kaczmarz with momentum as a gossip algorithm}
\label{RKmomentum}
\small \small
\begin{algorithmic}[1]
\State {\bf Parameters:} Distribution $\mathcal{D}$ from which method samples matrices; stepsize/relaxation parameter $\omega \in \mathcal{R}$; heavy ball/momentum parameter $\beta$.
\State {\bf Initialize:} $x^0 ,x^1 \in \mathcal{R}^n$
\For{$k=1,2,\dots$}
\State Pick an edge $e=(i,j)$ following the distribution ${\cal D}$
\State The values of the nodes are updated as follows:
\begin{itemize}
\item Node $i$: $x_i^{k+1}= \frac{2-\omega}{2}x_i^k+ \frac{\omega}{2}x_j^k+\beta (x_i^k - x_i^{k-1})$
\item Node $j$: $x_j^{k+1}= \frac{2-\omega}{2} x_j^k+\frac{\omega}{2}x_i^k+\beta (x_j^k - x_j^{k-1})$
\item Any other node $\ell$: $x_\ell^{k+1}=x_\ell^k+\beta (x_\ell^k - x_\ell^{k-1})$
\end{itemize}
\EndFor
\State {\bf Output:} The last iterate $x^k$
\end{algorithmic}
\end{algorithm}
\begin{rem}
In the special case with $\beta=0$ (zero momentum) only the two nodes of edge $e=(i,j)$ update their values. In this case the two nodes do not update their values to their exact average but to a convex combination that depends on the stepsize $\omega \in (0,2)$. To obtain the pairwise gossip algorithm of \cite{boyd2006randomized}, we should further choose $\omega=1$.
\end{rem}
\textbf{Distributed Nature of the Algorithm:} Here we highlight a few ways to implement mRK in a distributed fashion:
\emph{Asynchronous pairwise broadcast gossip:} In this protocol each node $i \in \cV$ of the network $\cG$ has a clock that ticks at the times of a rate 1 Poisson process. The inter-tick times are exponentially distributed, independent across nodes, and independent across time. This is equivalent to a global clock ticking at a rate $n$ Poisson process which wakes up an edge of the network at random. In particular, in this implementation mRK works as follows: In the $k^{th}$ iteration (time slot) the clock of node $i$ ticks and node $i$ randomly contact one of its neighbors and simultaneously broadcast a signal to inform the nodes of the whole network that is updating (this signal does not contain any private information of node $i$). The two nodes $(i,j)$ share their information and update their private values following the update rule of Algorithm~\ref{RKmomentum} while all the other nodes updating their values using their own information. In each iteration only one pair of nodes exchange their private values.
\emph{Synchronous pairwise gossip:} In this protocol a single global clock is available to all nodes. The time is assumed to be slotted commonly across nodes and in each time slot only a pair of nodes of the network is randomly activated and exchange their information following the update rule of Algorithm~\ref{RKmomentum}. The remaining not activated nodes update their values using their own last two private values. Note that this implementation of mRK comes with the disadvantage that requires a central entity which choose the activate pair of nodes in each step.
\emph{Asynchronous pairwise gossip with common counter:}
The update rule of the nodes of the active pair $(i,j)$ in Algorithm~\ref{RKmomentum} can be rewritten as follows:
$$ x_i^{k+1}= x_i^k + \beta (x_i^k - x_i^{k-1}) + \tfrac{\omega}{2} (x_j^k -x_i^k)$$
$$x_j^{k+1}= x_j^k + \beta (x_j^k - x_j^{k-1}) + \tfrac{\omega}{2} (x_i^k -x_j^k)$$
In particular observe that in their update rule they have the expression $x_i^k + \beta (x_i^k - x_i^{k-1})$ which is precisely the update of all non activate nodes of the network. Thus if we assume that the nodes share a common counter that counts how many iterations take place and each node $i$ saves also the last iterate $k_i$ that it was activated then the algorithm can work in distributed fashion as follows:
Let us denote the number of total iterations (common counter) that becomes available to the activate nodes of each step as $K$ and let us define with $i_k=K-k_{i}$ the number of iterations between the current iterate and the last time that the $i^{th}$ node is picked (iteration $k_{i}$) then the update rule of the Algorithm~\ref{RKmomentum} can be equivalently expressed as:
\begin{itemize}
\item Pick an edge $e=(i,j)$ at random following ${\cal D}$.
\item The private values of the nodes are updated as follows:
$$ x_i^{k+1}= i_k \left[x_i^k + \beta (x_i^k - x_i^{k-1}) \right]+ \tfrac{\omega}{2} (x_j^k -x_i^k)$$
$$x_j^{k+1}= j_k \left[x_j^k + \beta (x_j^k - x_j^{k-1}) \right] + \tfrac{\omega}{2} (x_i^k -x_j^k)$$
$k_i = k_j =k+1$; for any other node $\ell$: $x_\ell^{k+1}=x_\ell^k$
\end{itemize}
\begin{figure}[t!]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[scale=0.4]{mRK}}
\caption{\footnotesize Example of how mRK works as gossip algorithm. In the presented network the edge that connects nodes $6$ and $7$ is randomly selected. The pair of nodes exchange their information and update their values following the update rule of the Algorithm~\ref{RKmomentum} while the rest of the nodes, $\ell \in [5]$, update their values using only their own previous private values.}
\label{fig:mRK}
\end{minipage}
\end{figure}
\subsection{Connection with the accelerated gossip algorithm}
\label{connectionOfAcceleratedMethods}
In the randomized gossip literature there in one particular method closely related to our approach. It was first proposed in \cite{cao2006accelerated} and its analysis under strong conditions was presented in \cite{liu2013analysis}.
In this paper local memory is exploited by installing shift registers at each agent. In particular we are interested in the case of just two registers where the first stores the agent's current value and the second the agent's value before the latest update. The algorithm can be described as follows. Suppose that edge $e=(i,j)$ is chosen at time $k$. Then,
\begin{itemize}
\item Node $i$: $x_i^{k+1}= \omega(\frac{x_i^k+x_j^k}{2})+(1-\omega)x_i^{k-1}$
\item Node $j$: $x_i^{k+1}= \omega(\frac{x_i^k+x_j^k}{2})+(1-\omega)x_j^{k-1}$
\item Any other node $\ell$: $x_\ell^{k+1}=x_\ell^k$
\end{itemize}
where $\omega \in [1,2)$. The method was analyzed in \cite{liu2013analysis} under a strong assumption on the probabilities of choosing the pair of nodes that as the authors mentioned is unrealistic in practical scenarios, and for networks like the random geometric graphs. At this point we should highlight that the results presented in \cite{loizou2017momentum} hold for essentially any distribution ${\cal D}$ and as a result such a problem cannot occur.
Note also that if we choose $\beta=\omega-1$ in the update rule of Algorithm~\ref{RKmomentum}, then our method is simplified to:
\begin{itemize}
\item Node $i$: $x_i^{k+1}= \omega(\frac{x_i^k+x_j^k}{2})+(1-\omega)x_i^{k-1}$
\item Node $j$: $x_i^{k+1}= \omega(\frac{x_i^k+x_j^k}{2})+(1-\omega)x_j^{k-1}$
\item Any other node $\ell$: $x_\ell^{k+1}=\omega x_\ell^k+(1-\omega)x_\ell^{k-1}$
\end{itemize}
In order to apply Theorem~\ref{L2}, we need to assume that $0< \omega < 2$ and $\beta=\omega-1 \geq 0$ which also means that $\omega \in [1,2)$. Thus for $\omega \in [1,2)$ and momentum parameter $\beta=\omega-1$ it is easy to see that our approach is very similar to the shift-register algorithm. Both methods update the selected pair of nodes in the same way. However, in our case the other nodes of the network do not remain idle but instead also update their values using their own previous information.
Using the momentum matrix $\mathbf{B}=Diag(b_{11},b_{22},\dots, b_{nn})$, the two algorithms above can be expressed as:
\begin{equation}
\label{updatewithB}
x^{k+1} = x^k - \tfrac{\omega}{2} (x_i^k-x_j^k)(e_i-e_j) + \mathbf{B} (x^k-x^{k-1}).
\end{equation}
In particular, in our algorithm every element on the diagonal is equal to $\beta=\omega-1$, while in \cite{cao2006accelerated} all values on the diagonal are zeros except for the two values $b_{ii}=b_{jj}=\omega-1$.
\begin{rem}
The shift register case and our algorithm can be seen as two limit cases of the update rule \eqref{updatewithB}. In particular, the shift register method uses only two non-zero diagonal elements in $\mathbf{B}$, while our method has a full diagonal. We believe that further methods can be developed in the future by exploring the cases where more than two but not all elements of the diagonal matrix $\mathbf{B}$ are non-zero. It might be possible to obtain better convergence if one carefully chooses these values based on the network topology. We leave this as an open problem for future research.
\end{rem}
\subsection{Randomized block Kaczmarz gossip with momentum}
Recall that Theorem~\ref{TheoremRBK} says how RBK (with no momentum and no relaxation) can be interpreted as a gossip algorithm. Now we use this result to explain how relaxed RBK with momentum works. Note that the update rule of RBK with momentum can be rewritten as follows:
\begin{equation}
\label{updateRBK2}
x^{k+1} \overset{\eqref{momentumupdateb0}}{=} \omega (\mathbf{I}- \mathbf{A}^ \top \mathbf{H}_k \mathbf{A}) x^k+(1-\omega)x^k +\beta(x^k-x^{k-1}),
\end{equation}
where $(\mathbf{I} - \mathbf{A}^ \top \mathbf{H}_k \mathbf{A}) x^k$ is the update rule of RBK \eqref{updateSkProj}.
Thus, in analogy to the simple RBK, in the $k^{th}$ step, a random set of edges is selected and $q \leq n$ connected components are formed as a result. This includes the connected components that belong to both sub-graph $\cG_k$ and also the singleton connected components (nodes outside the $\cG_k$). Let us define the set of the nodes that belong in the $r \in [q]$ connected component at the $k^{th}$ step $\cV_r^k$, such that $\cV= \cup_{r\in [q]} \cV_r^k$ and $|\cV|=\sum_{r=1}^{q} |\cV_r^k|$ for any $k>0$.
Using the update rule \eqref{updateRBK2}, Algorithm~\ref{RBKmomentum} shows how mRBK is updating the private values of the nodes of the network (see also Figure~\ref{fig:RBK} for the graphical interpretation).
\begin{algorithm}[t!]
\caption{Randomized Block Kaczmarz Gossip with momentum}
\label{RBKmomentum}
\small \small
\begin{algorithmic}[1]
\State {\bf Parameters:} Distribution $\mathcal{D}$ from which method samples matrices; stepsize/relaxation parameter $\omega \in \mathcal{R}$; heavy ball/momentum parameter $\beta$.
\State {\bf Initialize:} $x^0,x^1 \in \mathcal{R}^n$
\For{$k=1,2,...$}
\State Select a random set of edges $\cS \subseteq \cE$
\State Form subgraph $\cG_k$ of $\cG$ from the selected edges
\State Node values are updated as follows:
\begin{itemize}
\item For each connected component $\cV_r^k$ of $\cG_k$, replace the values of its nodes with:
\begin{equation}
\label{updateruelblock}
x_i^{k+1}=\omega \tfrac{\sum_{j \in \cV_r^k} x_j^{k}}{|\cV_r^k|} +(1-\omega)x_i^k+\beta (x_i^k-x_i^{k-1})
\end{equation}
\item Any other node $\ell$: $x_\ell^{k+1}=x_\ell^k+\beta (x_\ell^k - x_\ell^{k-1})$
\end{itemize}
\EndFor
\State {\bf Output:} The last iterate $x^k$
\end{algorithmic}
\end{algorithm}
Note that in the update rule of mRBK the nodes that are not attached to a selected edge (do not belong in the sub-graph $\cG_k$) update their values via $x_\ell^{k+1}=x_\ell^k+\beta (x_\ell^k - x_\ell^{k-1})$. By considering these nodes as singleton connected components their update rule is exactly the same with the nodes of sub-graph $\cG_k$. This is easy to see as follows:
\begin{eqnarray}
x_\ell^{k+1}&=&\omega \tfrac{\sum_{j \in \cV_r^k} x_j^{k}}{|\cV_r^k|} +(1-\omega)x_\ell^k+\beta (x_\ell^k-x_\ell^{k-1})\notag\\
&=&\omega x_\ell^k +(1-\omega)x_\ell^k+\beta (x_\ell^k-x_\ell^{k-1})\notag\\
&=& x_\ell^k+\beta (x_\ell^k - x_\ell^{k-1}).
\end{eqnarray}
\begin{rem}
In the special case that only one edge is selected in each iteration ($\mathbf{S}_k \in \mathcal{R}^{m \times 1}$) the update rule of mRBK is simplified to the update rule of mRK. In this case the sub-graph $\cG_k$ is the pair of the two selected edges.
\end{rem}
\begin{rem}
In \cite{LoizouRichtarik} it was shown that several existing gossip protocols for solving the average consensus problem are special cases of the simple RBK (Theorem~\ref{TheoremRBK}). For example two gossip algorithms that can be cast as special cases of the simple RBK are the path averaging proposed in \cite{benezit2010order} and the clique gossiping \cite{liu2017clique}. In path averaging, in each iteration a path of nodes is selected and its nodes update their values to their exact average ($\omega=1$). In clique gossiping, the network is already divided into cliques and a through a random procedure a clique is activated and the nodes of it update their values to their exact average ($\omega=1$). Since mRBK contains simple RBK as a special case for $\beta=0$, we expect that these special protocols can also be accelerated with the addition of momentum parameter $\beta \in (0,1)$.
\end{rem}
\begin{figure}[t!
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[scale=0.41]{mRBK}}
\caption{\footnotesize Example of how the mRBK method works as gossip algorithm. In the presented network in the $k^{th}$ iteration the red edges are randomly chosen and they form subgraph $\cG_k$(from the red edges) and also four connected component. In this figure $V_1^k$ and $V_2^k$ are the two connected components that belong in the subgraph $\cG_k$ while $V_3^k$ and $V_4^k$ are the singleton connected components. Then the nodes update their values by communicate with the other nodes of their connected component using the update rule \eqref{updateruelblock}. For example the node number 5 that belongs in the connected component $V_2^k$ will update its value using the values of node 4 and 3 that also belong in the same component as follows:
$x_5^{k+1}=\omega \tfrac{x_3^k+x_4^k+x_5^k}{3} +(1-\omega)x_5^k+\beta (x_5^k-x_5^{k-1})$.}
\label{fig:RBK}
\end{minipage}
\end{figure}
\subsection{Mass preservation}
One of the key properties of some of the most efficient randomized gossip algorithms is mass preservation. If a gossip algorithm has this property it means that the sum (and as a result the average) of the private values of the nodes remains fixed during the iterative procedure. That is, $\textstyle \sum_{i=1}^{n}x_i^{k}=\sum_{i=1}^{n}x_i^{0}, \quad \forall k \geq 1.$ The original pairwise gossip algorithm proposed in \cite{boyd2006randomized} satisfied the mass preservation property, while exisiting accelerated gossip algorithms \cite{cao2006accelerated,liu2013analysis} preserving a scaled sum.
In this section we show that the two proposed protocols presented above also have a mass preservation property. In particular, we prove mass preservation for the case of the block randomized gossip protocol (Algorithm~\ref{RBKmomentum}) with momentum. This is sufficient since the Kaczmarz gossip with momentum (mRK) can be cast as special case.
\begin{thm}
Assume that $x^0=x^1$. That is, the two registers of each node have the same initial value. Then for the Algorithms~\ref{RKmomentum} and \ref{RBKmomentum} we have $\sum_{i=1}^{n}x_i^k=\sum_{i=1}^{n}c_i$ for any $k\geq 0$ and as a result, $\tfrac{1}{n}\sum_{i=1}^{n}x_i^k=\bar{c}$.
\end{thm}
\begin{proof}
We prove the result for the more general Algorithm~\ref{RBKmomentum}. Assume that in the $k^{th}$ step of the method $q$ connected components are formed. Let the set of the nodes of each connected component be $\cV_r^k$ so that $\cV= \cup_{r=\{1,2,...q\}} \cV_r^k$ and $|\cV|=\sum_{{r}=1}^{q} |\cV_r^k|$ for any $k>0$. Thus:
\begin{equation}
\label{generalsum}
\textstyle \sum_{i=1}^{n}x_i^{k+1}=\sum_{i \in \cV_1^k} x_i^{k+1} +\dots + \sum_{i \in \cV_q^k} x_i^{k+1}
\end{equation}
Let us first focus, without loss of generality, on connected component $r \in [q]$ and simplify the expression for the sum of its nodes:
$ \sum_{i\in \cV_r^k} x_i^{k+1}
\overset{\eqref{updateruelblock}}= \textstyle \sum_{i \in \cV_r^k} \omega \tfrac{\sum_{j \in \cV_r^k} x_j^{k}}{|\cV_r^k|} + (1-\omega) \sum_{i \in \cV_r^k} x_i^k +\beta \sum_{i \in \cV_r^k} (x_i^k-x_i^{k-1})
=|\cV_r^k| \tfrac{\omega \sum_{j \in \cV_r^k} x_j^{k}}{|\cV_r^k|}+ (1-\omega) \sum_{i \in \cV_r^k} x_i^k
+\beta \sum_{i \in \cV_r^k} (x_i^k-x_i^{k-1})
=(1+\beta) \sum_{i \in \cV_r^k}x_i^k-\beta \sum_{i \in \cV_r^k}x_i^{k-1}
$. By substituting this for all $r \in [q]$ into the right hand side of \eqref{generalsum} and from the fact that $\cV= \cup_{r\in [q]}
\cV_r^k$, we get
$ \sum_{i=1}^{n}x_i^{k+1}
= (1+\beta) \sum_{i=1}^{n}x_i^k-\beta \sum_{i=1}^{n} x_i^{k-1}.
$
Since $x^0=x^1$, we have $\sum_{i=1}^{n}x_i^{0}=\sum_{i=1}^{n}x_i^{1}$, and as a result $
\sum_{i=1}^{n}x_i^{k}
= \sum_{i=1}^{n}x_i^{0}$ for all $ k \geq 0$.
\end{proof}
\section{Numerical Evaluation}
\label{experiments}
We devote this section to experimentally evaluate the performance of the proposed gossip algorithms: mRK and mRBK. In particular we perform three experiments. In the first two we focus on the performance of the mRK, while in the last one on its block variant mRBK.
In comparing the methods with their momentum variants we use the relative error measure $\|x^k-x^*\|^2 / \|x^0-x^*\|^2 $ where the starting vectors of values $x^0=x^1=c$ are taken to be always Gaussian vectors. For all of our experiments the horizontal axis represents the number of iterations. The networks used in the experiments are the cycle (ring graph), the 2-dimension grid and the randomized geometric graph (RGG) with radius $r=\sqrt{\log(n)/n}$. Code was written in Julia 0.6.3.
\subsection{Impact of momentum parameter on mRK}
Recall that in the simple pairwise gossip algorithm the two nodes that exchange information update their values to their exact average while all the other nodes remain idle. In our framework this method can be cast as special case of mRK when $\beta=0$ and $\omega=1$. In this experiment we keep always the stepsize to be $\omega=1$ which means that the pair of the chosen nodes update their values to their exact average. We show that by choosing a suitable momentum parameter $\beta \in (0,1)$ we can have faster convergence for all networks under study. See Figure~\ref{mRKomega1} for more details.
\begin{figure}[t!]
\centering
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Grid10x10stepsize1}
\label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Grid20x20stepsize1}
\label{fig:sub2}
\end{subfigure}\\
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{RGG300stepsize1}
\label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{RGG500stepsize1}
\label{fig:sub2}
\end{subfigure}\\
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Cycle300stepsize1}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Cycle500stepsize1}
\end{subfigure}
\caption{\footnotesize Performance of mRK for fixed step-size $\omega=1$ and several momentum parameters $\beta$ in a cycle, 2-dimension grid and RGG. The choice $\beta=0$ corresponds to the randomized pairwise gossip algorithm proposed in \cite{boyd2006randomized}; The $n$ in the title of each plot indicates the number of nodes of the network. For the grid graph this is $n \times n$.}
\label{mRKomega1}
\end{figure}
\subsection{Comparison with the Shift-Register}
In this experiment we compare mRK with the shift register case when we choose the $\omega$ and $\beta$ in such a way in order to satisfy the connection establish in Section~\ref{connectionOfAcceleratedMethods}. That is, we choose $\beta=\omega-1$ for any choice of $\omega \in (1,2)$. Observe that in all plots of Figure~\ref{shiftregister} our algorithm outperform the corresponding shift-register case.
\begin{figure}[t!]
\centering
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\linewidth]{Grid30x30}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\linewidth]{Grid40x40}
\end{subfigure}\\
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\linewidth]{RGG500}
\end{subfigure}%
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\linewidth]{RGG1000}
\end{subfigure}\\
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\linewidth]{Cycle500}
\end{subfigure}
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\linewidth]{Cycle1000}
\end{subfigure}%
\caption{\footnotesize Comparison of mRK with the pairwise momentum method (Pmom), shift-register algorithm proposed in \cite{liu2013analysis}. For fair comparison we take always $\beta=\omega-1$ for our algorithm and the stepsizes are chosen to be either $\omega= 1.2$ or $\omega=1.3$. The baseline method is the simple not accelerated randomized pairwise gossip algorithm from \cite{boyd2006randomized}. The $n$ in the title of each plot indicates the number of nodes of the network. For the grid graph this is $n \times n$.}
\label{shiftregister}
\end{figure}
\subsection{Impact of momentum parameter on mRBK}
In this experiment our goal is to show that the addition of momentum accelerates the RBK gossip algorithm proposed in \cite{LoizouRichtarik}. Without loss of generality we choose the block size to be always equal to $\tau=5$. That is the random matrix $\mathbf{S}_k\sim {\cal D}$ in the update rule of mRBK is always a $m \times 5$ column submatrix of the indetity $m \times m$ matrix. Thus, in each iteration $5$ edges of the network are chosen to form the subgraph $\cG_k$ and the values of the nodes are updated according to Algorithm~\ref{RBKmomentum}. Note that similar plots can be obtained for any choice of block size. We run all algorithms with fixed stepsize $\omega=1$. It is obvious that by choosing a suitable momentum parameter $\beta \in (0,1)$ we have faster convergence than when $\beta =0$, for all networks under study. See Figure~\ref{RBKfigures} for more details.
\begin{figure}[t!]
\centering
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Grid20x20}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Grid20x20withbeta05}
\end{subfigure}\\
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{RGG500block5}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{RGG500block5withbeta05}
\end{subfigure}\\
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Cycle300block5}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=1\linewidth]{Cycle500block5}
\label{fig:sub2}
\end{subfigure}
\caption{\footnotesize Comparison of mRBK with its no momentum variant RBK ($\beta=0$) proposed in \cite{LoizouRichtarik}. The stepsize for all methods is $\omega=1$ and the block size is $\tau=5$. The baseline method in the plots denotes the simple randomized pairwise gossip algorithm (block $\tau=1$) and is plotted to highlight the benefits of having larger block sizes. The $n$ in the title of each plot indicates the number of nodes. For the grid graph this is $n \times n$.}
\label{RBKfigures}
\end{figure}
\section{Conclusion and Future research}
\label{conclusion}
In this paper we present new accelerated randomized gossip algorithms using tools from numerical linear algebra and the area of randomized Kaczmarz methods for solving linear systems. In particular, using recently developed results on the stochastic reformulation of consistent linear systems we explain how stochastic heavy ball method for solving a specific quadratic stochastic optimization problem can be interpreted as gossip algorithm. To the best of our knowledge, it is the first time that such protocols are presented for average consensus problem. We believe that this work opens up many possible future venues for research. For example, using other Kaczmarz-type methods to solve particular linear systems we can obtain novel distributed protocols for average consensus. In addition, we believe that the gossip protocols presented in this work can be extended to the more general setting of distributed optimization where the goal is to minimize the average of convex functions $(1/n) \sum_{i=1}^n f_i(x)$ in a distributed fashion.
\bibliographystyle{plain}
|
\section{Introduction}
GW170817 (\citealt{GW170817}) has just opened the era of multi-messenger astronomy (\citealt{MMA}).
The first coincident set of gravitational-waves and electromagnetic observations has already provided an extraordinary insight into
the physics of the binary neutron star mergers.
Among the key results of this revolutionary discovery is the confirmation of the association between the merger of two neutron stars (NSs) and (at least some) short gamma ray bursts (SGRBs) (\citealt{MMA} and refs. therein).
The last radio VLBI observations demonstrate
that a narrow jet was formed
and prove the association with a classical SGRB (see
\citealt{troja2017x, mooley2018mildly, margutti2018binary, d2018evolution, lyman2018optical, dobie2018turnover,mooley2018superluminal,ghirlanda2018re}).
The intense multi-wavelength follow-ups of gamma ray bursts in the last decade have revealed
new and unexpected features, such as early and late X-ray flares,
extended emission,
and X-ray plateaus (e.g., \citealt{berger2014short} and refs. therein).
The challenges posed by this rich astronomical scenario motivated
a growing interest of the community in investigating compact binary mergers from both the theoretical and observational points of view.
Intensified theoretical efforts have been
dedicated to explain these observations and coherently explore these and other possible electromagnetic signals generated by these sources.
In order to validate the variety of proposed theoretical scenarios in the context of multi-messenger astronomy with compact binary mergers, we present \texttt{saprEMo}.
We developed \texttt{saprEMo},
a {\it Simplified Algorithm for PRedicting ElectroMagnetic Observations}, to evaluate how many electromagnetic (EM) signals, characterised by a specific light curve and spectrum, should be present in a data set
given some overall characteristics of the astronomical survey and a cosmological rate of compact binary mergers.
Predictions can be used both to validate the theoretical scenarios against already collected data and to critically examine the scientific means of future missions and their observational strategies. While we use compact binary mergers as the prime multi-messenger targets, \texttt{saprEMo} can also be applied to other type of transients (e.g. core-collapse supernovae).
We describe the main features of \texttt{saprEMo} in Section \ref{S2:saprEMo_outline}.
As first case study, we use \texttt{saprEMo} to investigate the X-ray emission
from Binary NS (BNS) mergers leading to the formation of a long-lived and strongly magnetized NS, following the model of \citealt{siegel2016electromagnetic,siegel2016electromagneticII} (see Section \ref{S3_1:Xray_model}).
We apply \texttt{saprEMo} to present X-ray surveys, collected by {\it XMM-Newton} \citep{jansen2001xmm, struder2001european, turner2001european}
and {\it Chandra} \citep{weisskopf1996advanced, weisskopf2000chandra}, and study the prospectives of the mission concept {\it THESEUS} \citep{Amati2017}. Results are reported in Section \ref{S3_3:Xray_results} and discussed in Section \ref{S4:Discussion}. Finally, in Section \ref{S5:Conclusions} we draw our conclusions summarising our first results and outlining the main features and primary scopes of \texttt{saprEMo}.
Throughout the paper we assume a flat cosmology with: $H_0= 70\mathrm{~km~s^{-1}~Mpc^{-1}}$, $\Omega_{M} = 0.3$ and $\Omega_{\Lambda} = 0.7$.
\section{saprEMo outline}
\label{S2:saprEMo_outline}
\texttt{saprEMo} is a Python algorithm designed to
predict how many detectable electromagnetic signals, associated with a specific EM emission ({\it EMe}) model, are present in a survey {\it S}.
The full code and a short manual are publicly available at \texttt{https://github.com/saprEMo/source\_code}. \\
According to instrument limitations (such field of view and spectral sensitivity), \texttt{saprEMo} estimates the number of signals
whose emission flux $F$ at the observer is above the flux limit $F_{lim}$ of the {\it S} survey.
Accounting for the energy dependency of the survey sensitivity, we define detections
on a instantaneous flux-based criterion: $\exists~ g, t' ~\mid F_g\left(t'\right)>F_{lim,g}$, where $g$ labels the spectral band of the survey. We therefore simplify our analysis by treating
detectability in each band independently, i.e., a source is considered to be detected if and only if it can be independently detected in at least one instrument band. More realistic treatments include flux integration over the observation time and noise simulation (see \citealt{carbone2016calculating} and references therein for a discussion of these data analysis aspects).
\texttt{saprEMo} does not directly consider the actual sky locations observed by the survey {\it S} (even when applied to archival data) and instead focuses on accounting for cosmological distances, relying on the isotropy of the Universe.
\subsection{Core analysis}
\texttt{saprEMo} can be applied to any type of EM emission, from transients to continuous sources
emitting in any EM spectral range.
In this work,
we focus on X-ray transients associated with mergers of neutron star binaries.
The expected number of BNS mergers $N_{BNS}$ in the volume enclosed within redshift $z_{max}$, in a time $T$ at the observer, is:
\begin{eqnarray}
N_{BNS}=T\int_{0}^{z_{max}}\frac{R_V(z)}{1+z}\frac{dV_C}{dz}\,dz
\label{eq:nBNS}
\end{eqnarray}
where $R_V(z)$ is the rate of BNS mergers per unit comoving volume, per unit source time. In our case $z_{max}$ is the maximum distance at which the emission following the model of interest {\it EMe}, can be detected.
$z_{max}$ is calculated considering both the spectral shift due to the source redshift compared to the instrument energy band $E^I\sim\left[E^I_{min},E^I_{max}\right]$ and the maximum luminosity distance, set by the peak luminosity $L_p(E)$ of the {\it EMe} model
and the sensitivity $F_{lim}$ of the survey.
We only expect a fraction of $N_{BNS}$ to be observed
by a specific instrument, depending on the emission properties and the characteristics of the survey.
The number of BNS mergers, detectable by the survey {\it S} such that the peak of the considered emission {\it EMe} falls within the observing time, is given by:
\begin{eqnarray}
\begin{split}
N_{p}=
\varepsilon~ \frac{FoV}{4\pi} T \int_{0}^{z_{max}} \frac{\displaystyle R_V(z)}{1+z}\frac{dVc}{dz}\,dz
\label{eq:peaks}
\end{split}
\end{eqnarray}
The total observing time $T$ can be expressed in terms of the survey {\it S} as $T = \left<T_{obs}\right>~n_{obs}$, where $n_{obs}$ is the number of the observations and
$\left<T_{obs}\right>$ is the average exposure time for observation.
In eq. (\ref{eq:peaks}), the field of view $FoV$ of the instrument reduces the number $N_{BNS}$ of signals present in the volume enclosed within $z_{max}$ by $FoV/4\pi$.
In the specific case of BNS mergers, the efficiency factor $\varepsilon$ typically includes the
occurrence rate of a specific merger remnant ($\varepsilon_{sr}$), which are expected to generate the emission {\it EMe}, and source geometry/observational restrictions such as collimation ($\varepsilon_c = 1-\cos(\theta)$, where $\theta$ is the beaming angle), so that $\varepsilon\sim \varepsilon_{sr}\cdot \varepsilon_c$.
We designate as {\it peaks} the signals included in $N_p$ (see figure \ref{fig:1}).
This contribution only depends on the emission model by the intensity of the light curve peaks in the energy bands of the survey. This dependency is enclosed in $z_{max}$.
There is also a contribution, which we call {\it tails} (figure \ref{fig:1}), from the mergers whose emission is detected only before (first block of eq. (\ref{eq:tails})) or after (second block of eq. (\ref{eq:tails})) the luminosity peaks (i.e., $L_p$ is outside the observation period). The longer the light curve is above $F_{lim}$, the higher the probability of it being observed (see \citealt{carbone2016calculating} for a detailed discussion of signal duration in the context of transient detectability and classification).
To estimate this contribution, we need to account for the evolution in time of the emission luminosity $L(t')$, which affects the horizon of the survey:
\begin{eqnarray}
\begin{split}
N_t & =\displaystyle \varepsilon~ n_{obs}\frac{FoV}{4\pi}\left[\int_{-\infty}^{t'_p}\int_{0}^{z_t(L(t'))} \frac{\displaystyle R_V(z)}{1+z}\frac{dVc}{dz}\,dz\,dt +\right.\\ & \quad \left.\int_{t'_p}^{+\infty}\int_{0}^{z_t(L(t'))} \frac{\displaystyle R_V(z)}{1+z}\frac{dVc}{dz}\,dz\,dt \right] \\ & =\displaystyle \varepsilon~ n_{obs}\frac{FoV}{4\pi}\int_{-\infty}^{+\infty}\int_{0}^{z_t(L(t'))} \frac{\displaystyle R_V(z)}{1+z}\frac{dVc}{dz}\,dz\,dt
\label{eq:tails}
\end{split}
\end{eqnarray}
where $t$ and $t' = t/(1+z)$ are the time respectively in observer and source frames and $t'_p$ is the time corresponding to the peak luminosity. $z_t\left(L(t')\right)$ represents the horizon of the survey, given the specific intrinsic luminosity of the source $L(t')$.
The integration time of eq. (\ref{eq:tails}) is practically limited by the duration of the emission above the flux limit.
At the moment, \texttt{saprEMo} does not correctly account for the possibility of
detecting multiple times
the same event. Multiple detections of the same source might occur if the survey contains repeated observations of the same sky locations and the time interval between the different exposures is
shorter than
the considered emission {\it EMe}.
While $N_p$ would be unaffected, in these cases $N_t$ would overestimate the expected number of events by these additional detections. Under these specific conditions, $N_t$ should then be considered as an upper limit.
We refer the readers to \citealt{carbone2016calculating} for discussions of transient detectability in the context of multiple images of the same field.
The ratio between the duration of the observable emission and the typical exposure time determines the relative importance of peaks and tails.
The trade off between these two contributions, as well as their different origin, can be understood with figures \ref{fig:1} and \ref{fig:peaks_tails}.
For illustrative purposes, we consider only local events with unphysical rates and a generic triangular light curve.
Figure \ref{fig:1} shows how tails events $b$ and $c$ can be observed during one exposure of duration $\left<T_{obs}\right>$.
Figure \ref{fig:peaks_tails} shows the impact of the transient observable duration on tails. Upper and lower panels represent exactly the same scenarios (10 seconds of exposure of transients at $z = 0$ characterised by a 1 Hz rate) except for the transient duration, which is doubled in the lower panel. The number of stars, which represent the peak contribution, is the same in both upper and lower panels, demonstrating that peaks are unaffected by the change of the transient observable duration.
However the tail contribution, given by the number of pink triangles, doubles in the lower panel compared to the upper one.
On the contrary, extending the exposure to 20 seconds would double the number of peaks, while leaving unchanged the number of tails.
Given a fixed event rate, $N_p$ depends on the observing time, while $N_t$ depends on the duration of the events.
We return to this topic in section \ref{S4:Discussion}, when we discuss the results of this study.
\begin{figure}
\centering
\begin{minipage}[b]{0.5\textwidth}
\centering
\includegraphics[width=8.cm]{images/paperFig1_2.pdf}
\caption{Schematic representation of a peak ($t_p \in \left<T_{obs}\right>$) signal ($a$) and tails ($b,~c$). The solid curves represent the part of signals {\it EMe} visible during the exposure time at the observer, the dashed components are the missed (because of time or flux restrictions) part of the emissions. The upper dotted line shows the peak flux $F_p = L_p/(4\pi D^2)$, the lower line the flux limit of the survey.}
\label{fig:1}
\end{minipage}
\hfill
\begin{minipage}[b]{0.5\textwidth}
\centering
\includegraphics[width=8.cm]{images/PT.png}
\caption{Simplified examples of peaks and tails. In both upper and lower panels we report the detected flux (in arbitrary units on the y-axis) as a faction of time at the observer. We assume a transient class emitting from $z\sim 0$, characterised by the unrealistic rate $1$ Hz and a fixed triangular light curve. The flux limit $F_{lim}$ is represented by the horizontal line. The transient light curves above the flux limit are shown in black solid lines. The light gray region highlights the 10~s exposure window.
The panels show the peak contribution, i.e. the events whose luminosity peak falls in the observational window (green stars) and the tail contribution, i.e. the events detected only before or after the luminosity peak (fuchsia lines and area). The only difference between upper and lower panels is the duration of the transient light curve above the flux limit, respectively $1$~s and $2$~s.
Doubling the light curve duration results in doubling the number of tails, while leaving unchanged the number of peaks.
}
\label{fig:peaks_tails}
\end{minipage}
\end{figure}
While eq. (\ref{eq:peaks}) and (\ref{eq:tails}) explain the general concept behind the tool, they do not explicitly account for the energy (or wavelength $E=\hbar c/\lambda$) dependence of light curves, instrument sensitivities and absorption.
These effects are particularly important when exploring the Universe at high redshift.
We explain how these effects are implemented in \texttt{saprEMo} analysis in appendix \ref{App:A}.\\
\subsection{Inputs and Outputs}
We now present inputs and outputs of
\texttt{saprEMo}.\\
\noindent {\it Inputs:}
\begin{enumerate}[i]
\item \textbf{light curves} (LC), in terms of luminosity rest frame, of the EM emission {\it EMe} in different energy bins
(if predicted by the model, also including energies outside the instrument sensitivity band, as they
might be redshifted into it, after accounting for cosmology corrections);
{\es{(spin-down luminosity of the stable newly born neutron star - fiducial light curve proposed by \citealt{Siegel2016a})}}
\item \textbf{astrophysical rate} in the source frame $R_V\left(z\right)$;
{\es{( rate of BNS mergers, as derived by \citealt{Domink2013})}}
\item \textbf{efficiency} $\boldsymbol{\varepsilon}$ of the EM model {\it EMe} ($\boldsymbol{\varepsilon}$ should account for source geometry/selection effects, such as collimation, as well as the frequency with which this type of astrophysical source generates the electromagnetic emission {\it EMe});
{\es{(assuming this corresponds to the fraction of SGRBs followed by extended X-ray emission $\varepsilon\sim\%50$ \citealt{Rowlinson2013})}}
\item \textbf{main instrument and survey properties}:
\begin{itemize}
\item for each spectral band $g$ of the survey {\it S}, minimum and maximum energy included $[E_i,E_s]_g$;
\item corresponding flux limits $[F_{lim}]_g$;
\item average exposure time $\left<T_{obs}\right>$;
\item field of view $FoV$ \footnote{Or equivalently covered sky-area $f_{sky}\sim n_{obs}\frac{FOV}{4\pi}$.};
\item number of observations $n_{obs}$\footnotemark[1].
\end{itemize}
{\es{({XMM-Newton Serendipitous Source Catalog. Averaged sensitivity {\ToBeChecked{$\sim 10^{-15}$\,\flux}}, average observation time {\ToBeChecked{$\sim 10^{4}$\,s}}, instrument band $0.2-12$\,kHz, field of view $\sim 0.2$\,$deg^2$})}}
\end{enumerate}
\noindent {\it Outputs:}
\begin{itemize}
\item $\boldsymbol{\mathrm{N_p}}$ and $\boldsymbol{\mathrm{N_t}}$: numbers of peaks and tails which are expected in the survey {\it S}. The numbers of signals returned by \texttt{saprEMo} should be interpreted as the expectation value of a Poisson process.
Therefore Poisson statistical errors should be considered in addition to the systematics due to rate and emission model uncertainties;
\item $\boldsymbol{\mathrm{dN_p}}${\bf/dz} and $\boldsymbol{\mathrm{dN_t}}${\bf/dz}: distributions of tail and peak numbers as a function of redshift;
\item $\boldsymbol{\mathrm{dN_p}}${\bf/dlog(D)} and $\boldsymbol{\mathrm{dN_t}}${\bf/dlog(D)}: expected distribution of signal observed durations, obtained by convolving the approximate distribution of the survey exposure times $P_{obs}$ with the light curve span $LCS$ observable at each step in redshift.
For each redshift, the $LCS$ represents the total time of the light curve which is above the flux limits at the observer frame (for more details see appendix \ref{APP_A3}).
To estimate the distribution of the signal durations,
we analytically approximate the exposure time distribution $P_{obs}$ from the average exposure time $\left<T_{obs}\right>$ (and standard deviation, when available) with a {\it Maxwell-Boltzmann} or {\it Log-normal} function, according to the user input.
For each data point saved from the cosmic integrations,
we simulate $N_{trials}$ (for both peaks and tails)
observation durations $T_{obs}$ and define
for each of them the starting time $t_s$.
The starting time is uniformly drawn from a time interval including both the exposure time of the specific trial $T_{obs}$ and the observable emission at observer $(t'_f - t'_i)(1+z)$ (where $t'_f$ and $t'_i$ are respectively the last and first LC time at source satisfying our detection criteria at redshift $z$).
If $t_s$ is drawn in the interval $I_p =\left[t_p - T_{obs}, t_p\right]$, where $t_p$ is the time at observer correspondent to the luminosity peak, it contributes to the peak distribution, otherwise it adds up to the tail distribution. For each simulated observation the total duration is then calculated summing only the contribution of light curve intervals whose flux is above the limit;
\item $\boldsymbol{\mathrm{dN_p}}${\bf/dlog(F)} and $\boldsymbol{\mathrm{dN_t}}${\bf/dlog(F)}: distributions of peak and tail detection numbers as a function of maximum flux. At each step in redshift, necessary to compute the integral (\ref{eq:peaks}), \texttt{saprEMo} also calculates the associated flux.
The fluxes are obtained by summing the contribution of each energy and rescaling with the associated luminosity distance. From the same observations simulated for estimating the duration distributions, we obtain the expected distribution of maximum fluxes.
\end{itemize}
Distributions in redshift are useful to estimate the horizon of the survey to the emission {\it EMe} and for astrophysical interpretation.
They provide a prior on the redshift distribution when a counterpart allowing $z$ measurement is missing, or constrain cosmic rate evolution of BNS mergers when multi-wavelength observations yield the source distance.
Distributions of fluxes and durations are robust observables, which
enable comparisons with real data
\footnote{
The reported flux distribution is calculated from the maximum theoretical fluxes of detected events in our simulation (see paragraph on $dN_{p/t}/d\log(D)$-output).
Quantitative comparisons with actual data would require a more detailed analysis, including
the use of the instrument response, a realistic model for noise, the model used to convert photon counts into a light curve, etc.~(see for example \citealt{carbone2016calculating}, who modeled some of these aspects).}.
\section{Application to soft X-ray emission from long-lived binary neutron star merger remnants}
\label{S3:Xray_sky}
In the following, we consider a specific application of \texttt{saprEMo} to the case of spindown-powered X-ray transients from long-lived NS remnants of BNS mergers.
Depending on the involved masses and the NS equation of state (EOS), a BNS merger can either produce a short-lived remnant, collapsing to a black hole (BH) within a fraction of a second, or a long-lived massive NS. The latter can survive for much longer spindown timescales (up to minutes, hours or more) prior to collapse or even be stable forever \citep{lasky2014nuclear, Lu2015}.
After the discovery of NSs with a mass of $\sim2~M_\odot$ \citep{Demorest2010, Antoniadis2013}, different authors converged to the idea that the fraction of BNS mergers leading to a long-lived NS remnants should range from a few percent up to more than half (e.g., \citealt{Piro2017}).
Information extracted from the multimessenger observations of the BNS merger event GW170817 did not change this view, although more stringent constraints on the NS EOS were obtained from the GW signal \citep{LVC-EOS}, from various indications excluding the prompt collapse to a BH, and from the kilonova brightness and the relatively high mass of the merger ejecta (e.g., \citealt{margalit2017constraining,Bauswein2017,radice2018gw170817,Rezzolla2018}).
An additional supporting element in favour of long-lived remnants is given by the observation of long-lasting ($\sim$ minutes to hours) X-ray transients following a significant fraction of SGRBs (e.g., \citealt{Rowlinson2013,Gompertz2014,Lu2015}). Given the short accretion timescale of a remnant disk onto the central BH ($\lesssim1$~s), such long-lasting emission represents a challenge for the canonical BH-disk scenario of SGRBs while it can be easily explained by alternative scenarios involving a long-lived NS central engine, e.g., the magnetar \citep{Zhang2001,Metzger2008} and the time-reversal \citep{Ciolfi2015,Ciolfi2018} scenarios. According to this view, the fraction of SGRBs accompanied by long-lasting X-ray transients might reflect the fraction of BNS mergers producing a long-lived NS.
If the merger remnant is a long-lived NS, its spindown-powered electromagnetic emission represents an additional energy reservoir that can potentially result in a detectable transient. Recent studies taking into account the reprocessing of this radiation across the baryon-polluted environment surrounding the merger site have shown that the resulting signal should peak at wavelengths between optical and soft X-rays, with luminosities in the range $10^{43}-10^{48}$~erg/s and time scales of minutes to days (e.g., \citealt{Yu2013,MetzgerPiro2014,siegel2016electromagnetic,siegel2016electromagneticII}).
Besides representing an explanation for the long-lasting X-ray transients accompanying SGRBs, this spindown-powered emission is a promising counterpart to BNS mergers (e.g., \citealt{Stratta2017} and refs.~therein), having the advantage of being both very luminous and nearly isotropic.
For our first direct application of \texttt{saprEMo},
we consider the spindown-powered transient model by \citealt{siegel2016electromagnetic,siegel2016electromagneticII} (hereafter SC16), described in the next Section \ref{S3_1:Xray_model}, in which the emission is expected to peak in the soft X-ray band.
This model cannot be excluded or constrained by GW170817. The first X-ray observations in the $2-10$~keV band were performed by MAXY \citep{sugita2018maxi} 4.6 hours after the merger with a sensitivity of $8.6\times10^{-9}~\mathrm{erg~s^{-1}~cm^{-2}}$,
well above the flux that the model predicts at that time after the merger.
In Section \ref{sec:3.2}, we briefly describe the model of the BNS merger rate adopted for this first study.
Then, in Section \ref{S3_3:Xray_results} we present our results referring to three different X-ray satellites: {\it XMM-Newton}, {\it Chandra}, and the proposed {\it THESEUS}. We discuss these results in section \ref{S4:Discussion}.
\subsection{Reference emission model}
\label{S3_1:Xray_model}
The model proposed by Siegel \& Ciolfi (SC16) describes the evolution of the environment surrounding a long-lived NS formed as the result of a BNS merger. The spindown radiation from the NS injects energy into the system and interacts with the optically thick baryon-loaded wind ejected isotropically in the early post-merger phase, rapidly forming a baryon-free high-pressure cavity or ``nebula'' (with properties analogous to a pulsar wind nebula) surrounded by a spherical shell of ``ejecta'' heated and accelerated by the incoming radiation.
As long as the ejecta remain optically thick, the non-thermal radiation from the nebula is reprocessed and thermalised before eventually escaping. As soon as the ejecta become optically thin, a signal rebrightening is expected, accompanied by a transition from dominantly thermal to non-thermal spectrum. The model can also take into account the collapse of the NS to a BH at any time during the spindown phase.\footnote{
We refer to \citealt{siegel2016electromagnetic,siegel2016electromagneticII} and \citealt{Ciolfi2016} for a detailed discussion of the model and its current limitations.}
Exploring a wide range of physical parameters, Siegel \& Ciolfi found that the escaping spindown-powered signal has a delayed onset of $\sim10-100$ s and in most cases peaks $\sim100-10^4$ s after merger. Furthermore, the emission typically falls inside the soft X-ray band (peaking at $\sim0.1-1$ keV) and the peak luminosity is in the range $10^{46}-10^{48}$ erg s$^{-1}$.
\begin{figure}
\begin{centering}
\includegraphics[width=8.75cm] {images/LCcomp}
\caption{Light curve of the spindown-powered emission from a long-lived BNS merger remnant according to the model proposed by \citealt{siegel2016electromagnetic,siegel2016electromagneticII} (corresponding to their ``fiducial'' case; see text).
The solid curves represent the contributions of different energy bands to the total light curve (dashed line).}
\label{fig:LCinE}
\end{centering}
\end{figure}
In this work, we consider only one representative case, corresponding to the ``fiducial case'' of SC16 (SC16f) (this model is rescaled for the
analysis in Section \ref{sec3.3.2}).
The light curve and spectral distribution of this particular model are shown in figure~\ref{fig:LCinE}. The main parameters of the model are as follows. The early baryon-loaded wind ejects mass isotropically at an initial rate of $5\times10^{-3}~M_\odot$~s$^{-1}$, decreasing in time with a timescale of 0.5~s. The dipolar magnetic field strength at the poles of the NS is $10^{15}$~G and the initial rotational energy of the NS is $5\times10^{52}$~erg ($\sim$~ms initial spin period).
Moreover, in this case the remnant evolves without collapsing to a BH.
In figure~\ref{fig:LCinE} we can distinguish two important transitions. The first, around $\sim10$~s, marks the end of the early baryon wind phase and the beginning of the spindown phase. The second, at several times $10^{4}$~s, corresponds to the time when the ejecta become optically thin.
While the emission described by the above model is essentially isotropic, allowing us to set $\varepsilon_c\sim 1$, only a fraction of BNS mergers $\varepsilon_{LLNS}$ is expected to generate a long-lived neutron star.
The value of this fraction mainly depends on the unknown
NS EOS and distribution of component masses. Here, we assume for simplicity a one-to-one correspondence between the fraction $\varepsilon_{LLNS}$ and the fraction of SGRBs accompanied by a long-lasting X-ray transient (i.e. extended emission and/or X-ray plateau). Following the analysis presented in
\citealt{Rowlinson2013}, we set $\varepsilon_{LLNS}$ to 50\%.
Once we assign $\varepsilon_{sr}=\varepsilon_{LLNS}$, the resulting total efficiency of the emission is $\varepsilon\sim \varepsilon_{sr}\cdot \varepsilon_{c}=50\%$.
\subsection{BNS merger rate model}
\label{sec:3.2}
The dependence of the BNS merger rate on redshift is poorly observationally constrained. Several models based on different assumptions have been proposed.
For the present work we consider 4 different rate models, a simplified (default) case and three further astrophysically-motivated scenarios:
\begin{description}
\item[{\bf DEFAULT}]: a constant BNS merger rate per unit comoving volume per unit source time in the range $R_V(z) = [100-10000]~\mathrm{Gpc^{-3}yr^{-1}}$, extending up to a maximum redshift of $z = 6$;
\item[{\bf D2013}]: the Monte Carlo population synthesis model of \citealt{dominik2013double} (their cosmological standard model, high-end metallicity scenario \citealt{syntheticuniverse}.);
\item[{\bf G2016}]: the analytic approximation based on SGRB observations described in eq. (12) of \citealt{ghirlanda2016short} (adopting the average value of the parameter reported for case {\it a} with an opening angle of $4.5\deg$);
\item[{\bf MD2014}]: the analytic prescription for the star formation history proposed by \citealt{madau2014cosmic} convolved with a probability distribution of delay times between formation and merger given by the power low $P(t_{del})\propto t_{del}^{-1}$, with a minimum delay time of $20\times 10^{6}~$yr, normalised to the local BNS merger rate of $1540~$Gpc$^{-3}$yr$^{-1}$, as estimated with GW170817 ($1540^{+3200}_{-1220}~\mathrm{Gpc^{-3}yr^{-1}}$ median and uncertainties at 90\% probability,
\citealt{GW170817}).
\end{description}
The different BNS merger rates are reported in figure \ref{fig:rate}. We note that D2013 and G2016, as proposed, are inconsistent with the local rate range obtained from GW170817 (gray region).
However both inferred rate from a single gravitational-wave observation and population synthesis models rely on poorly constrained astrophysical model assumptions and are therefore highly uncertain.
We adopt these rate models for illustrative purposes to test the impact of different redshift-dependent merger rates.\\
To investigate the impact of other inputs, we adopt the DEFAULT simplified model of constant cosmological rate as a reference case.
We report distributions and results for $R_V(z) = 1000 ~\mathrm{Gpc^{-3}yr^{-1}}$.
Because the considered $R_V(z)$ is constant, the results for the upper (lower) bound of the whole range $R_V(z) = [100-10000] ~\mathrm{Gpc^{-3}yr^{-1}}$, can be obtained by scaling up (down) the output quantities by one order of magnitude.
This wide range of the BNS merger rate includes the local rate interval inferred from the detection of GW170817 and is broadly consistent with estimates obtained using Galactic BNS observations and population synthesis models
(\citealt{abadie2010predictions,paul2017binary,chruslinska2017double,vignagomez2018}).
\begin{figure}
\begin{centering}
\includegraphics[width=9.25cm]{images/RATE.png}
\caption{BNS merger rate as a function of redshift for different models: D2013 \citealt{dominik2013double} in {green}, G2016 \citealt{ghirlanda2016short} in {blue}, MD2014 \citealt{madau2014cosmic} convolved with $P(t_{del})\propto t_{del}^{-1}$ in {orange} and our default constant model in {violet}: the solid line represents the reference rate of $R_V(z) = 1000 ~\mathrm{Gpc^{-3}yr^{-1}}$, while the light shadowed area includes the whole interval $R_V(z) =[100-10000]~\mathrm{Gpc^{-3}yr^{-1}}$.
The range in redshift is divided into $z\le 1$ and $z>1$ to enhance the visibility of the constraints set by the observation of GW170817 (gray area, 90\% probability as reported in \citealt{GW170817}), which only apply to the local Universe.
}
\label{fig:rate}
\end{centering}
\end{figure}
We use the DEFAULT model in Section \ref{S3_3:Xray_results} and discuss the impact of applying different BNS merger rate models in Section \ref{S4:Discussion}.
\subsection{Results}
\label{S3_3:Xray_results}
We now proceed with using \texttt{saprEMo} and the X-ray transient model described in Section \ref{S3_1:Xray_model} to evaluate the expected number of detections of this type of signal in three present and future surveys by {\it XMM-Newton}, {\it Chandra} and {\it THESEUS}.
To emphasise the impact of the survey properties, we initially keep fixed: (i) the light curve to the SC16 model described in Section \ref{S3_1:Xray_model} (fiducial, SD16f, or fiducial-rescaled, see Section \ref{sec3.3.2}); (ii) the assumed astrophysical rate to the DEFAULT case, described in Section \ref{sec:3.2}; and (iii) the efficiency $\varepsilon\sim 50\%$.
In particular, we consider:
\begin{itemize}
\item two present surveys, collected during the decade of operation of {\it XMM-Newton}, to predict the expected number of detectable signals in these archived data;
\item the {\it Chandra} Deep Field - South (CDF-S)
data set to verify whether the transient class discovered by \citealt{bauer2017new} is statistically consistent with the SC16 model;
\item 10~ks of {\it THESEUS} observations, to explore
the sensitivity of this mission concept
to transients associated with BNS mergers, such as SC16f.
\end{itemize}
The main properties of the surveys are summarised in appendix \ref{App:Survey_Prop}.
\subsubsection{XMM-Newton}
\label{sec3.3.1}
We apply \texttt{saprEMo} to two different
collections of data obtained by
{\it XMM-Newton}; we call them SLEW and PO (Pointed Observations), their characteristics are presented in the following.
The number of signals predicted by \texttt{saprEMo} are reported
in table \ref{tab:XMM_res}
\footnote{For both the surveys, the statistical uncertainties due to the assumed Poisson distribution
are almost negligible compared to the systematics due to
uncertainty in the signal production efficiency $\varepsilon$ and the BNS merger rate.}.
In the case of {\it XMM-Newton} surveys, the sky locations of observations have been used to estimate the impact of the absorption due to the Milky Way (see appendix \ref{APP_A2} for more details on our adopted absorption model).
\begin{description}
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
&\multicolumn{2}{|c|}{{\bf {\it XMM-Newton}}} &{\bf {\it Chandra}} &\multicolumn{2}{|c|}{{\bf {\it THESEUS}}}\\
\hline
& PO & SLEW &CDF-S & Case a& Case b\\
\hline
$\mathbf{N_p}$ & 8 & 0& 0.14 & 5 (4) & 3 (2) \\
\hline
$\mathbf{N_t}$& 25 & 165 &\ding{56} & 5 (3) & 20 (11) \\
\hline
\hline
$\mathbf{FoV}~[\deg^2]$ & \multicolumn{2}{|c|}{$\sim 0.2$} & 0.08 & \multicolumn{2}{|c|}{{3300}} \\
\hline
$\mathbf{T}~[10^6\mathrm{s}]$ & $\sim 160$ & $\sim 1.06$ & $\sim6.73$ & \multicolumn{2}{|c|}{{$0.01$}}\\
\hline
\multicolumn{4}{c}{{}}\\
\\
\end{tabular}
\caption{
Average expected values for peaks ($N_p$) and tails ($N_t$) for different surveys. {\it XMM-Newton} PO and SLEW surveys$^a$, {\it Chandra} Deep Field - South (CDF-S) and 10 ks of {\it THESEUS} operation for a single exposure, {\it case a}, and 10 distinct exposures {\it case b} considering $N_H = 5\times 10^{-22}~\mathrm{cm^{-2}}$ ($N_H = 5\times 10^{-20}~\mathrm{cm^{-2}}$). $^a$ For the {\it XMM-Newton} surveys the total observing time was inferred from $T = n_{obs}\left<T_{obs}\right> = \left(4\pi~f_{sky}~FoV^{-1}\right)\left<T_{obs}\right>$, using the properties reported in tables \ref{tab:XMM_PO_par} and \ref{tab:XMM_SLEW_par}.
}
\label{tab:XMM_res}
\end{table}
\item[{\bf{The PO survey}}] is a collection of pointed observations made between 3/2/2000 and 15/12/2016.
The data belong to the {\it XMM-Newton Serendipitous Source Catalog (3XMM DR7)} (\citealt{3XMM-DR7ref,3XMM_Pref}). \\
PO exposures are longer (typically $10^3-10^4$~s) compared to the SLEW catalog (see following paragraph).
This implies an extension to lower fluxes (down to $10^{-15}-10^{-16}~\mathrm{erg~s^{-1}~cm^{-2}}$), as figure \ref{fig:XMMres} (a) shows.
The same figure shows that such low fluxes are however reached only by tails.
This is because the luminosity of the model makes the flux higher than the survey flux limits up to $z=6$, which is the artificial cut of our BNS merger rate.
The distribution of source redshift is represented in the bottom graph of figure \ref{fig:XMMres} (a) and implies that, under the assumption of a constant cosmic BNS merger rate,
the median redshift of detectable signals is $z\sim 2$.
The double bump in the tail distribution of the PO survey is explained by the blue and purple curves which respectively represent the light curve span above the threshold at a fixed redshift $z$, $LCS(z)$,
and
the time-shifted rate of events per unit redshift, $\displaystyle R_V(z)\,dVc/dz$.
Given our simplified BNS merger rate model,
the purple and black-solid lines scale like the
redshift derivative of the comoving volume,
since in both cases only constants multiply the element $dV_c/dz$.
The blue curve has instead a very peculiar behaviour which depends on the specific emission light curve compared to the limit fluxes of the survey.
The distinct trends in the blue lines of figure \ref{fig:XMMres}, are due to features very peculiar to the adopted light curve (figure \ref{fig:LCinE}).
When the flux from the non-thermal second peak drops below the limit, the overall visible duration sharply decreases; this happens at $z\sim 0.5$ and $z\sim 0.05$ for PO and SLEW, respectively.
The second turn-over at $z=4$ in the LCS, evident in PO tails (gray area and dashed-black line), is instead due to the discretisation of the light curves in energy bins and the relatively soft energy spectrum of these transients.
In particular the lowest energy bin characterising the light curve (see figure \ref{fig:LCinE}) exits the band of the instrument \footnote{According to appendix \ref{App:A} notation: $z_{exit} =[E'_{max,h=0}/E^I_{min}-1]\sim [1~\mathrm{keV}/\mathrm{0.2}~\mathrm{keV}-1] = 4$, where $z_{exit}$ is the redshift at which the energy bin of lowest energy, denoted with $h = 0$, exits the instrument band, $E'_{max,h=0} = 1~keV$ is the highest energy included in the bin $h = 0$ and $E^I_{min} = 0.2~$keV is the minimum energy included in the instrument band.}.
\item[{\bf The SLEW survey}] is composed by data collected while changing the target in the sky, according to the {\it XMM-Newton} observation program (\citealt{SLEWref}).
The tested observations are collected in the
{\it XMM-Newton Slew Survey Clean Source Catalog, v2.0}. \\
The SLEW survey is characterised by typical exposure time of only few seconds, and consequent flux limits $\gtrsim 10^{-13}~\mathrm{erg~s^{-1}~cm^{-2}}$.
Given the properties of the model SC16, this yields a predominance of tails over peaks (see first point of section \ref{S4:Discussion}).
Our results show that SLEW observations, assuming correct identification (see section \ref{S4:Discussion}), could already reveal a population of BNS merger events.
The flux limits of the survey determines the distance of most of the X-ray sources at $z<3$.
\end{description}
Although the data have been collected by the same instrument, PO and SLEW considerably differ in terms of exposure time (and therefore sensitivity), sky coverage and energy responses (as shown by tables in appendix \ref{Appendix:XMM-Newton}).
The SLEW survey is less sensitive, but it scans a much wider area of the sky compared to the PO survey, so that the total number of expected signals is actually considerably larger (see tab. \ref{tab:XMM_res}).
\begin{figure*}
\centering
\subfloat[][]
{
\includegraphics[scale=.365]{images/P.pdf}
}
\qquad
\subfloat[][]
{
\includegraphics[scale=.365]{images/S.pdf}
}
\caption
{Comparison between expected distributions of detectable signals in PO and SLEW surveys. The upper-left panel shows the total expectations on the observed maximum flux distribution obtained by adding peak (dark green) and tail (green) contributions. The upper-right panel shows the expected distributions of peak and tail durations. The redshift distributions in the bottom panel represent the differential contribution of peaks and tails throughout the scanned comoving volume of Universe. The violet and blue curves have been added to explain the trend of the tail distribution and represent respectively the time-shifted rate of events per unit redshift (scaled argument of the $z$ integral \ref{eq:tails_rev}) and the light curve span $LCS(z)$ (the duration of the light curve above the flux limits at a fixed $z$).
(a) Expected distributions of observed maximum flux (upper-left), duration of signals above the flux limits (upper-right) and redshift (bottom) of the observations in {\it XMM-Newton} PO survey.
(b) Expected distributions of observed maximum flux (upper-left), duration of signals above the flux limits (upper-right) and redshift (bottom) of the observations in {\it XMM-Newton} SLEW survey. Because of the low expected number of detections for peaks compared to tails (see tab. \ref{tab:XMM_res}), the black solid line in bottom graph is not visible.
}
\label{fig:XMMres}
\end{figure*}
\subsubsection{Chandra and the new, faint X-ray population}
\label{sec3.3.2}
\citet{bauer2017new} have recently claimed the discovery of a X-ray signal, belonging to a new, previously unobserved, transient class. The event was observed within the {\it Chandra} Deep Field - South (CDF-S), a deep survey of a 0.11 square degrees sky region composed of 102 observations collected in different periods during the last decade.
Interestingly, the main properties of the event presented by \citet{bauer2017new} are broadly consistent with the SC16 emission model.
The maximum luminosity of $\sim 10^{47}~\mathrm{erg/s}$, the spectral peak around $\sim2$~keV (source frame), the rise time of $\sim 100~$s, and the overall duration of $\sim 10^{4}~\mathrm{s}$ are all in broad agreement with the model predictions. Here we do not attempt to provide convincing evidence for a potential match, but we want to show another interesting case for exploiting the capabilities of \texttt{saprEMo}.
We assume a signal analogous to the SC16f adopted throughout this paper only rescaled to have a maximum luminosity of $\sim 10^{47}~\mathrm{erg/s}$ (referred to as ``rescaled SC16 model/signal'' in the following), consistent with
the X-ray transient at $z\sim 2.23$ \citep{bauer2017new}.
To test the rate consistency between the detected X-ray transient and the rescaled SC16 model, we apply \texttt{saprEMo} to the CDF-S, adopting a Galactic neutral column density of $n_{H_{tot},MW} \approx n_{H,MW}\sim 8.8\times 10^{19}~\mathrm{cm^{-2}}$, as reported by \citet{bauer2017new}.
Given the proposed source redshifts, the shape and fluxes of the detected transient,
we assume that the observed maximum flux corresponds to the luminosity peak of our model.
We therefore
evaluate only the expected number of peaks.
\texttt{saprEMo} predicts an expectation value of $\sim0.14$ signals in the CDF-S (see table \ref{tab:XMM_res}).
Given the adopted constant rate model, the probability of one rescaled SC16 signal being present at its luminosity peak in the $\sim 7~$Ms of the CDF-S is $\sim 12\%$ (with $\sim 87\%$ probability of $0$ signals). Considering the whole range of allowed BNS merger rates, this value ranges from $\sim 1.4\%$ (with $\sim 98.6\%$ probability of $0$ signals, correspondent to $R_V(z) = 100~\mathrm{Gpc^{-3}yr^{-1}}$) and $\sim 35\%$ (with $\sim 25\%$ probability of $0$ signals, correspondent to $R_V(z) = 10000~\mathrm{Gpc^{-3}yr^{-1}}$).
Despite the broad consistency of the transient revealed by \citealt{bauer2017new} with the rescaled SC16 emission model,
our analysis shows that a real association between the two is rather disfavoured, although not inconsistent given the uncertainties over rate and emission model.
Conversely, assuming that the detected transient is in fact the rescaled SC16 signal adopted in the above calculation, the constant BNS merger rate value
is constrained to\footnote{To estimate the posterior on the constant rate value, we assume a flat prior in the range $[100,10000]~\mathrm{Gpc^{-3}yr^{-1}}$.} $6000^{+4000}_{-3700}~\mathrm{Gpc^{-3}~yr^{-1}}$ (median with 90\% credible interval, as in \citealt{GW170817}), higher than the median inferred from GW170817, though still consistent with the claimed interval
(cf.~Section \ref{sec:3.2}).
\subsubsection{Future observations with {\it THESEUS}}
\label{sec:3.3.3}
In the last few years, different wide-FoV X-ray missions have been proposed to monitor the X-ray sky, and specifically to follow up GRBs and GWs (\citealt{feng2016extp,barcons2012athena,yuan2015einstein,merloni2012erosita}).
In particular, the mission concept
{\it THESEUS} has been recently selected by
ESA for assessment studies
\citep{ESA_THESEUS} to explore the transient high-energy sky and contribute to multi-messenger astronomy \citep{Amati2017,Stratta2017,frontera2018observing}.
We apply \texttt{saprEMo} to test the sensitivity of the {\it THESEUS} mission to BNS mergers emitting in the X-ray according to the SC16f model.
On the {\it THESEUS} payload, the Soft X-ray Imager (SXI, \citealt{o2018soft}) would be the instrument sensitive to such emission.
SXI flux sensitivities for sources in the Galactic plane ($N_H= 5 \times 10^{22}~\mathrm{cm^{-2}}$) and well outside it ($N_H = 5 \times 10^{20}~\mathrm{cm^{-2}}$) are taken from figure~4 of \citealt{Amati2017}.
With \texttt{saprEMo}, we predict detection numbers and properties for two cases
of gathering {\it THESEUS} observations, each having a total observing duration $T$ of 10~ks,
acquired with:
\begin{description}
\item[{\bf case a}] a single exposure of $\left<T_{obs}\right> = 10^4$~s;
\item[{\bf case b}] 10 exposures of non-overlapping sky regions, each lasting $\left<T_{obs}\right> = 10^3$~s.
\end{description}
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|}
&\multicolumn{3}{|c|}{
{\bf {\it XMM-Newton} PO}} \\
\hline
{\bf Rate model }& MD2014 & D2013 & G2016 \\
\hline
$\mathbf{N_p}$&20 &1&2 \\
\hline
$\mathbf{N_t}$&65&2 &5\\
\hline
\hline
$\mathbf{FoV}~[\deg^2]$ & \multicolumn{3}{|c|}{{
$\sim$~0.25
}} \\
\hline
$\mathbf{n_{obs}\left<T_{obs}\right>}~[s]$ &\multicolumn{3}{|c|}{{$\sim 234\times 10^6~$}} \\
\hline
\\
\multicolumn{4}{c}{{}}\\
&\multicolumn{3}{|c|}{{\bf {\it THESEUS} Case {\it a}}} \\
\hline
{\bf Rate model }& MD2014 & D2013 & G2016 \\
\hline
$\mathbf{N_p}$& 17 (15) & 0.54 (0.46) & 1.2 (1.0)\\
\hline
$\mathbf{N_t}$& 16 (10) & 0.54 (0.34) & 1.0 (0.6)\\
\hline
\end{tabular}
\caption{
For {\it XMM-Newton} PO (top) and {\it THESEUS} case {\it a} surveys, comparison of expectation values assuming different BNS merger rate models, from left to right (i) analytic prescription proposed by \citealt{madau2014cosmic}, assuming a power-law distribution of delay times between formation and merger $P(t_{del})\propto t_{del}^{-1}$, (ii) cosmological rate derived by the population synthesis study \citealt{dominik2013double}, standard model at high-end metallicity scenario (D2013) and (iii) model based on SGRB statistics \citealt{ghirlanda2016short} with assumed opening angle of $4.5\deg$ (G2016).
}
\label{tab:RESULTS}
\end{table}
\begin{figure}
\centering
\includegraphics[width=9cm]{images/Peaks_THES.png}
\caption{Expected redshift distribution for peak signals in 10~ks of {\it THESEUS/SXI} observations (opaque lines), assuming $N_H = 5\times 10^{20}\mathrm{cm^{-2}}$.
In the default configuration of constant BNS merger rates we test two different observing strategies: {\it case a} and {\it case b}, as described in Section \ref{sec:3.3.3}. They are shown by the solid-violet and dashed-fuchsia lines, which completely overlap for $z \lesssim 2$ and only differ in the maximum achieved redshift, $z\sim 3$ for the former and $z\sim 2$ for the latter.
We also report the peak distribution as a function of redshift for other tested BNS merger models: MD2014 in orange, D2013 in green and G2016 in blue. For comparison with figure \ref{fig:rate}, we show the results in the two different regimes: $z>1$ in linear and $z\le 1$ in logarithmic scale. For the same reason, we also add in transparency
the distribution of peaks missed because of sensitivity constraints.
}
\label{fig:THESEUS1OBS_z}
\end{figure}
The last 2 columns of table \ref{tab:XMM_res} show the expectation values for $N_p$ and $N_t$
for both case {\it a} and {\it b}.
{\it In less than 3 hours of total observing time T, we expect {\it THESEUS} to detect a number of SC16-like transients comparable to the ones inferred for years-long CDF-S {\it Chandra} and PO {\it XMM-Newton} surveys} (see also table \ref{tab:XMM_res}).
The main reason why {\it THESEUS} is capable of reaching these numbers of detections in such a short observing time $T$ (about 4 orders of magnitude shorter than for CDF-S and PO/SLEW), is the 4 orders of
magnitude difference between its FoV and
the ones in {\it Chandra} and {\it XMM-Newton}.
Albeit specific for the SC16f emission, our tabulated results demonstrate that the characteristics of the mission concept {\it THESEUS} suit the search for
X-ray emission generated during BNS mergers.
We predict that {\it THESEUS/SXI} with {\it case a} sensitivity would detect on the order of hundreds to thousands of BNS mergers in only a few years, assuming an emission with peak luminosity in the range $L_p\sim[10^{45}-10^{48}]~\mathrm{erg~cm^{-2}~s^{-1}}$ and a spectral distribution similar to SC16f.
Our analysis therefore shows that proposed large FoV instruments, such {\it THESEUS}, offer an incredible opportunity compared to present deep surveys, for the detection of rare but bright transients, such as those
of the presented model. {\it THESEUS/SXI}-like campaigns are expected to detect events generated near the peak of the cosmic star formation and BNS merger rate, at $z\sim 1-3$. \\
The redshift distribution $dN_p/dz$ of case {\it a} and {\it b} in figure \ref{fig:THESEUS1OBS_z} show that, as expected, longer exposure times decrease the flux limits and therefore probe larger redshift.
Meanwhile, multiple shorter exposures at distinct sky locations, as in case {\it b}, increase the probability of detecting tails. We find that {\it THESEUS/SXI} would make more detections, during a fixed total observing time, with an observing strategy that increased sky coverage at the cost of shorter exposures.
\section{Discussion}
\label{S4:Discussion}
With \texttt{saprEMo}, we tested the sensitivity of different astronomical surveys to the emission model SC16f.\\
\\
{\bf{
Given a fixed emission model, the tail contribution becomes more important for surveys with shorter exposures $\boldsymbol{\left<T_{obs}\right>}$.
}}
The analysis has confirmed that when the exposure $\left<T_{obs}\right>$ is considerably shorter than the duration of the emission model, as is the case of {\it XMM-Newton} SLEW observations, detections of pre-peak rises or post-peak decays will be more common than observations of peak flux.
Therefore the relation between $\left<T_{obs}\right>$ and the typical duration of the EM model sets the relative number of peaks and tails.
Given a fixed amount of total observing time, the number of expected tails increases with the number of pointings of different sky positions.
This is shown, for example, by the two cases tested with {\it THESEUS} (see tables \ref{tab:XMM_res} and \ref{tab:RESULTS}). \\
\\
{\bf{Inferences on BNS merger models}}
\begin{itemize}
\item{\bf{A few detections can already constrain the lower limit of cosmic BNS merger rate.}}
For example, assuming the emission model proposed by \citealt{siegel2016electromagneticII}, the probability of detecting more than 3 peaks in {\it XMM-Newton} PO
assuming a constant merger rate of $100~\mathrm{Gp^{-3}yr^{-1}}$
is $<1\%$; thus, a few detections could set a lower limit on the BNS merger rate.
The proposed emission model also offers the unique possibility of exploring mergers in the high-redshift Universe:
with peak luminosity as high as $\sim 10^{48}~ \mathrm{erg/s}$, {\it XMM-Newton} could detect
signals generated at redshifts as high as $z\sim 15$ with PO sensitivities. \\
Constraints on BNS merger rate can be also obtained by assuming the association of the {\it Chandra} X-ray transient with the SC16 emission model.
In this case our analysis can put a lower limit on the BNS merger rate of $\sim 2300~\mathrm{Gpc^{-3}yr^{-1}}$ at 90\% confidence interval, assuming a constant rate up to $z\sim 6$.
The peak luminosity of the rescaled SC16 emission is indeed bright enough to actually be detectable by {\it Chandra} up to $z\sim 6$, even after accounting for the Milky Way absorption.
\item{\bf{Larger PO-like or {\it THESEUS} surveys will probe $\mathbf{R_V(z)}$ and likely constrain also EM emission models.}}
In the case of bright emission, as expected by SC16,
\texttt{saprEMo} predicts that the PO campaign can detect all BNS merger peaks in the Universe, localised within {\it XMM-Newton}'s field of view.
However, table \ref{tab:XMM_res} shows that to probe a statistically significant population of BNS mergers, PO-like sensitivities must be obtained with longer time window $T$ and/or larger FoV.
Large field of view instrument, such as {\it THESEUS/SXI}, can also detect many of these events.
Although limited to smaller redshift, cosmological distances could still be achieved if we assume bright emission models. This is shown for the case of SC16 in figure \ref{fig:THESEUS1OBS_z}.
These campaigns should therefore yield the BNS merger rate distribution as
a function of redshift, providing that multi-wavelength observations will allow redshift associations. These observations would then enable us to constrain the proposed BNS merger rate scenarios, which indeed predict different z distributions (see figure \ref{fig:THESEUS1OBS_z}).
\item{\bf{Redshift measurement play a fundamental role for breaking the degeneracy between emission parameters and rate models.}}
We apply \texttt{saprEMo} to the PO and {\it THESEUS} cases, {\it case a}, for the three scenarios
introduced in section \ref{sec:3.2}: D2013, G2016 and MD2014 (see figure \ref{fig:rate}).
The absolute expectation values reported in table \ref{tab:RESULTS} and figure \ref{fig:THESEUS1OBS_z} , reflect the trend of the rate models reported in figure \ref{fig:rate}.
The overall results show that, without perfect knowledge of light curve and spectrum of the emission, measurements of source distances are necessary to constrain the redshift dependence of the BNS merger rate.
\end{itemize}
{\bf{Considerations on \texttt{saprEMo}'s results.}}
In this paragraph we highlight some general considerations, to realistically interpreting \texttt{saprEMo} results.
The main output of the analysis consists in the number of peaks and tails expected for a specific emission model in a selected survey of data.
However, depending on the purpose of the analysis,
other information, such as more accurate requirements for detectability and classification, should be taken into account. In the following we give some examples.
\begin{description}
\item {\it Challenges for detectability:}
despite this does not concern the results presented in the previous session, short transients in long exposure observations can be lost in the integrated background flux.
To overcome this issue, targeted analyses might be required (see for e.g. the work of EXTraS group \citealt{extrasWEB_transient} \citealt{de2016science}
for detection of $\sim 10^2~$s-lasting transients in PO).
\item {\it Challenges for classifications: }
because of their definition, we generally expect durations and fluxes of tails to extend to lower values compared to the peak ones (as shown in figure~\ref{fig:XMMres} (a)).
This generally worsen the performances of signal classifications and identification among more common phenomena.
In particular for our analyses, given the shape of SC16f's light curve, tails should mostly appear as simple decaying signals, which can be challenging to distinguish from other X-ray transients (e.g., tails of tidal disruption events \citealt{Lodato2011} or supernovae \citealt{dwarkadas2011published}).\\
The correct classification of X-ray events can also be challenged by short exposure times.
This is for example the case of the SLEW survey, where observations typically last only few seconds.
Indeed some emission models, as SC16, predict a long-scale time evolution of the emission properties which would likely result in detections of dissimilar signals, challenging their association to a common origin.
Campaigns characterised by typically longer observations, such as the {\it XMM-Newton} PO and {\it Chandra} CDF-S, are less affected by classification problems.
The extension of the typical exposure time to thousands of seconds and improved spectral resolution, allow for the acquisition of more informative data,
simplifying transient identification.
\end{description}
\section{Summary and Outlook}
\label{S5:Conclusions}
In this study we showed some applications of our tool \texttt{saprEMo}; we applied it on few present and possible future surveys, assuming a specific emission model and cosmological BNS merger rate $R_V(z)$.
In terms of multi-messenger astronomy, our results show that the luminosities predicted by the SC16 emission model can be detected up to cosmological distances which extend much further than the horizon of present (\citealt{aasi2016prospects,2018LRR....21....3A}) and future gravitational wave detectors (\citealt{sathyaprakash2012scientific}) to binary neutron star mergers, both in the cases of current surveys, such as CDF-S and {\it XMM-Newton} PO and SLEW, and proposed missions, such as {\it THESEUS}.
\texttt{saprEMo} provides theoretical predictions allowing us:
\begin{itemize}
\item {\it{to compare predictions with actual data.}}
E.g. we proved that some signals consistent with the model could already be detected in present surveys of data such as {\it XMM-Newton} PO and SLEW;
\item {\it{to test potential associations.}} E.g. we proved that the new transient found by \citealt{bauer2017new} is marginally consistent with the model;
\item {\it{to assess the effectiveness of proposed mission concepts for a specific type of signal. }}
We illustrate the utility of \texttt{saprEMo} for evaluating proposed missions with a case study of {\it THESEUS}.
We demonstrate that, with few years of operation, the large FoV of {\it THESEUS/SXI} could allow for the detection of up to thousands of SC16-like signals, enabling considerable constrains on both the BNS merger rate and the emission models;
\item {\it{and to compare different observational strategies.}}
\texttt{saprEMo} can be used to determine advantages and disadvantages compared to a particular emission, of adopting different observational strategies.
With the case of {\it THESEUS}, we indeed demonstrate that \texttt{saprEMo} can compare observations characterised by different values of typical exposure time $\left<T_{obs}\right>$ and point out the main properties of the relative detections.
In general, given a total observing time $T$, different observational strategies can be applied; increasing the exposure time to increase the sensitivity or decreasing the exposure time to enlarge the sky coverage.
The effect of adopting different exposures depends on several parameters, including both source and instrumental properties (such as rate, luminosity, flux limit dependence on exposure time, etc).
In this paper, we specifically prove that, given {\it THESEUS/SXI} sensitivity as a function of exposure time,
10 observations of disjoint sky areas lasting 1~ks would en-captured more SC16f-like transients than an extended single pointing of 10~ks.
\end{itemize}
In general \texttt{saprEMo} allows us to test both survey and astrophysical properties. This study has mainly focused on the former, exploring the impact
of different trade offs among such properties (including exposure time, sky localisation, and spectral sensitivity), assuming a single light curve model from SC16.
However \texttt{saprEMo} can also
test (and be used for inference on) astrophysical quantities such as emission duration, peak luminosity and spectra.\\
Once the design sensitivity of Advanced gravitational-wave interferometers is achieved, GW detections of EM bright sources, such as GW170817, will occur more and more often and very likely at lower SNRs.
In this context of multi-messenger astronomy,
\texttt{saprEMo} can be used to
optimize the analysis
by identifying specific emission model.
We conclude remarking that the flexibility of the implemented methodology allows considerations of emission model spanning the whole electromagnetic spectrum (e.g. kilonovae models can also be tested).
Moreover our analysis include no priors on nature of EM sources, so that it can be applied to a wide range of astrophysical phenomena.
With its analysis dedicated to treat high redshift effects, \texttt{saprEMo} particularly suits studies on emission of cosmological origin.
\section*{Acknowledgements}
The authors thank L. Amati for his assistance with the case of {\it THESEUS} and for the useful comments.
We thank A. Belfiore, A. De Luca, M.Marelli, D. Salvietti and A. Tiengo for the help in understanding and interpreting {\it XMM-Newton} data. We thank R. Salvaterra for the useful suggestions and discussions.
The research leading to these results has received funding from the
People Programme (Marie Curie Actions) of the European Union's Seventh
Framework Programme FP7/2007-2013/ (PEOPLE-2013-ITN) under REA grant
agreement no.~[606176]. This paper reflects only the authors' view and the European Union is not liable for any use that may be made of the information contained therein.
G.S. acknowledges EGO support through a VESF fellowship (EGO-DIR-133-2015).
This research has made use of data obtained from XMMSL2, the Second {\it XMM-Newton} Slew Survey Catalogue, produced by members of the XMM SOC, the EPIC consortium, and using work carried out in the context of the EXTraS project ("Exploring the X-ray Transient and variable Sky", funded from the EU's Seventh Framework Programme under grant agreement no.~[607452]).
This research has made use of data obtained from the 3XMM {\it XMM-Newton} serendipitous source catalogue compiled by the 10 institutes of the {\it XMM-Newton} Survey Science Centre selected by ESA.
\bibliographystyle{mnras}
|
\section{Introduction}
Inverse problems deal with the recovery of an unknown quantity of interest
${\boldsymbol{\mathsf{x}}}\in\bbR^d$ from a corrupted observation ${\boldsymbol{\mathsf{y}}}\in\bbR^m$.
In most cases the relationship between ${\boldsymbol{\mathsf{x}}}$ and ${\boldsymbol{\mathsf{y}}}$ is linear, and can be approximately
described by
\begin{equation}\label{eqn:inverse_problem}
{\boldsymbol{\mathsf{y}}} = {\textrm{A}} {\boldsymbol{\mathsf{x}}} +\sigma{\boldsymbol{\mathsf{w}}},
\end{equation}
where ${\textrm{A}}\in\bbR^{m\times d}$ is a known linear forward operator, ${\boldsymbol{\mathsf{w}}}$ is a zero-mean isotropic random
vector, modeling the noise, and $\sigma>0$ is the noise level.
Inverse problems of this type are ubiquitous in image processing, compressed sensing and other scientific fields.
In image processing applications they model tasks such as: denoising, where ${\textrm{A}}$ is the identity; deblurring, where ${\textrm{A}}$ is a convolution operator; and inpainting, where ${\textrm{A}}$ is a masking operator.
The recovery of the original signal ${\boldsymbol{\mathsf{x}}}$ from the corrupted observation ${\boldsymbol{\mathsf{y}}}$ is an ill-posed inverse problem.
Thus, theory of inverse problems suggests the use of suitable regularization techniques \cite{zbMATH00936298}.
Specifically, in case of Gaussian noise, ${\boldsymbol{\mathsf{x}}}$ is approximated with the minimizer of a regularized functional
\begin{equation}\label{eqn:regularizer}
\argmin_{ {\boldsymbol{\mathsf{z}}}\in\bbR^d} \N{{\textrm{A}} {\boldsymbol{\mathsf{z}}}- {\boldsymbol{\mathsf{y}}}}_2^2 + \lambda
J({\boldsymbol{\mathsf{z}}}),
\end{equation}
where $\N{\cdot}_2$ is the Euclidean norm modeling \emph{data-fidelity}, $J$ is a \emph{penalty term}
encoding an \emph{a priori} knowledge on the nature of the true solution, and $\lambda$ is a \emph{regularization
parameter} determining a trade-off between these two terms.
Having the penalty term fixed, a central issue concerns the selection of $\lambda$.
The optimal parameter $\lambda$ is the one
that minimizes the discrepancy between the minimizer ${\boldsymbol{\mathsf{z}}}^\lambda$
of~\eqref{eqn:regularizer} and
the exact solution~${\boldsymbol{\mathsf{x}}}$
\begin{equation}\label{eqn:lambda_ast}
\lambda_{\text{opt}} = \argmin_{\lambda \in (0,+\infty)} \N{{\boldsymbol{\mathsf{z}}}^\lambda
-{\boldsymbol{\mathsf{x}}}}_2.
\end{equation}
In the context of (semi-)supervised machine learning methods, regularization parameters are selected by evaluating \eqref{eqn:lambda_ast}, or another metric, over a training set of clean signals.
Unfortunately though, due to the curse of dimensionality accurate estimation of high-dimensional functions requires a number of samples that scales exponentially with
the ambient dimension. A common approach to mitigating these effects is to assume that the relevant data are supported on structures of substantially lower dimensionality.
On the other hand, in regularization theory the clean image is unknown and hence $\lambda_\text{opt}$ is approximated using prior knowledge about the noise, such as the noise level.
Moreover, classical regularization theory is mostly concerned with the case when the data belongs to a function space (and is thus infinite dimensional).
In this case most existing parameter selection methods focus on the recovery of the minimum least-squares norm solution.
On the other hand, our work considers finite dimensional problems that incorporate additional constraints on the recovered solution in order to ensure it has the desired structure.
In many applications ${\boldsymbol{\mathsf{x}}}$ is unknown (thus we cannot use supervised methods) and there is no available information about the noise ${\boldsymbol{\mathsf{w}}}$ or the noise level $\sigma$.
Hence, $\lambda_{\text{opt}}$ needs to be approximated.
Moreover, the lower level problem \eqref{eqn:lambda_ast} is often non-convex, even when \eqref{eqn:regularizer} is.
Choosing a good approximation to $\lambda_{\text{opt}}$ is a non-trivial, problem-dependent task that has motivated significant amounts of research over the last decades.
However, there is still no framework that allows a fast and efficient parameter selection, particularly in a completely unsupervised setting.
In this paper, we aim at (partially) closing this gap and provide a novel concept for automated parameter selection by recasting the problem to the framework of statistical learning theory.
Specifically, inspired by recent and (to our knowledge) first results in this spirit \cite{DFN16} we propose a method for learning the optimal parameter for elastic net regularization that uses a dimension reduction preprocessing step.
We emphasize that the method is unsupervised and requires minimal human interference.
\paragraph{Existing parameter selection methods.}
Parameter selection rules used in regularization theory can be broadly classified as those based on the discrepancy principle \cite{morozov2012,AR2010}, generalized cross-validation (GCV)\cite{golub79}, balancing principle \cite{lepskii1991problem,DeVito2010}, quasi-optimality \cite{tikhonov1977,hofmann1986} and various estimations of the mean-squared error (MSE) (see \cite{stein1981,deledalle2014stein} and references therein).
GCV is a particularly popular parameter rule for linear methods since it gives a closed form for the regularization parameter and does not require tuning of any additional parameters or the knowledge of the noise.
In specialized cases GCV can be extended to nonlinear problems \cite{wood2000}, but the regularization parameter is no longer given in closed form nor through an implicit equation.
Balancing principle is a stable method that has received a lot of attention in the inverse problems community and has also been studied in the framework of learning theory, but requires tuning of additional parameters.
Quasi-optimality is one of the simplest parameter choice methods. It does not require any information about the problem but it is not as stable as the balancing principle.
Discrepancy and MSE-based principles still remain the preferred methods for parameter selection for nonlinear estimators due to their simplicity and accuracy.
We refer to a recent rather comprehensive comparative study on the existing approaches \cite{zbMATH05929140}.
In order to select the regularization parameter most existing methods require the regularized solution ${\boldsymbol{\mathsf{z}}}^\lambda$ to be computed over a predefined grid of values of $\lambda$.
The regularization parameters are then chosen according to some criteria, \emph{e.g.}, loss over a validation set.
To find regularization parameters by an exhaustive search is a computationally expensive task, especially in the high-dimensional data scenario, with often no guarantees on the quality of approximation.
Moreover, most criteria presuppose that some \emph{a priori} information is available, such as an accurate estimate of the noise level (in \emph{e.g.} discrepancy principle) or bounds on the noise error (in \emph{e.g.} balancing principle) and require additional, method-specific parameters to be preselected.
The main motivation of this work is to compute an accurate regularized solution ${\boldsymbol{\mathsf{z}}}^\lambda$, with a nearly optimal $\lambda$, while ensuring low computational complexity and minimizing the need for manual intervention.
In particular, we propose an unsupervised parameter selection method by recasting the problem to the framework of statistical learning, where we are interested in learning a function ${\boldsymbol{\mathsf{y}}} \mapsto \widehat \lambda_{\text{opt}}$, from a training set of corrupted data, while ensuring that $\widehat\lambda_{\text{opt}}$ is a good approximation of the optimal parameter $\lambda_{\text{opt}}.$
\paragraph{Elastic net regularization.}
Elastic net regularization was proposed by Zou and Hastie \cite{ZH05}, as
\begin{equation}
{\boldsymbol{\mathsf{z}}}^{\lambda}({\boldsymbol{\mathsf{y}}}) = \argmin_{\boldsymbol{\mathsf{z}}} \N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\boldsymbol{\mathsf{y}}}}^2 +
\lambda \LRP{\N{{\boldsymbol{\mathsf{z}}}}_1 + \alpha\N{{\boldsymbol{\mathsf{z}}}}_2^2},
\label{eq:3}
\end{equation}
where $\alpha\geq 0 $ is a hyperparameter controlling the
trade-off between $\ell_1$ and $\ell_2$ penalty terms.
Our main motivation for considering the elastic net is that it produces sparse models comparable to the Lasso (and is thus well suited for problems with data on lower dimensional structures), while often achieving superior accuracy in real-world and simulated data.
Moreover, the elastic net overcomes the main limitations of $\ell_1$ minimization.
Namely, it encourages the grouping effect, which is relevant for many real-life applications such as microarray classification and face detection (see \cite{DDR09} and references therein).
To solve \eqref{eq:3} the authors in \cite{ZH05} rewrite the elastic net functional as Lasso regularization with augmented datum, use LARS \cite{efron2004} to reconstruct the entire solution path, and apply cross-validation to select the optimal regularization parameter.
Later work \cite{DDR09} studies theoretical properties of \eqref{eq:3} in the context of learning theory, analyzes the
underlying functional and uses \emph{iterative soft-thresholding} to compute the solution.
For the parameter choice the authors provide an adaptive version of the balancing principle \cite{lepskii1991problem,DeVito2010}.
The rule aims to balance approximation and sample errors, which have contrasting behavior with respect to the tuning parameter, but requires (potentially) many evaluations of ${\boldsymbol{\mathsf{z}}}^\lambda$.
We will rework some of the arguments from \cite{DDR09} for the computation of ${\boldsymbol{\mathsf{z}}}^{\lambda}$, while keeping our focus on an efficient approach for parameter learning.
In \cite{JLS09} the authors propose an \emph{active set} algorithm for solving \eqref{eq:3}.
Addressing the problem in the framework of classical regularization theory, the authors consider the discrepancy principle \cite{morozov2012,bohesky2009} for determining the parameter.
This requires estimations of the solution ${\boldsymbol{\mathsf{z}}}^{\lambda}$ for many parameter values, and a pre-tuning of other, method-specific parameters.
Moreover, it is assumed that the noise level is known, which is often not the case in practice.
The authors in \cite{LR10} use a hybrid alternating method for tuning parameters $\lambda$ and $\alpha$ for the model fitting problem
$
\mathsf{y}_i = {\boldsymbol{\mathsf{a}}}_i^\top {\boldsymbol{\mathsf{x}}}, i = 1,\ldots, n,
$,
where $\mathsf{y}_i \in \bbR,$ ${\boldsymbol{\mathsf{a}}}_i \in \bbR^p$ and ${\boldsymbol{\mathsf{x}}} \in \bbR^p.$
First step is to update the solution ${\boldsymbol{\mathsf{z}}}^{\lambda}$, using coordinate descent, and then to update $\lambda$ and $\alpha$ in one iteration.
The main advantage is the efficiency, as one does not need to calculate ${\boldsymbol{\mathsf{z}}}^{\lambda}$ for multiple parameters at once, but rather on a much coarser parameter grid.
The method is in spirit similar to LARS, but has better scalability. It requires that a non-convex problem is solved, and hence has inherent limitations.
Moreover, it cannot be used in the setting of general inverse problems \ref{eqn:inverse_problem}, where the design matrix ${\textrm{A}}$ is fixed and each response ${\boldsymbol{\mathsf{y}}}$ is generated by a new clean signal ${\boldsymbol{\mathsf{x}}}$.
{In summary, we are not aware of any parameter selection rule for the elastic net that allows to select $\lambda$ without \emph{a priori} assumptions and without extensive manual adjustments.}
This paper leverages the work \cite{DFN16} where the parameter selection is considered in the context of non-parametric regression with random design.
In particular, the authors propose a data-driven method for determining the optimal parameter for Tikhonov regularization, under the assumption that
a training set of independent observations
\( {\boldsymbol{\mathsf{y}}}_1,\ldots, {\boldsymbol{\mathsf{y}}}_N\) is made available,
each of them associated with an (unknown) signal
${\boldsymbol{\mathsf{x}}}_1,\ldots, {\boldsymbol{\mathsf{x}}}_N$ through
\(
{\boldsymbol{\mathsf{y}}}_i = {\textrm{A}} {\boldsymbol{\mathsf{x}}}_i +\sigma {\boldsymbol{\mathsf{w}}}_i.
\)
The starting point of the method is to find an empirical proxy $\widehat{\boldsymbol{\mathsf{x}}} $ of the real solution ${\boldsymbol{\mathsf{x}}}$ by assuming that ${\boldsymbol{\mathsf{x}}}_1,\ldots, {\boldsymbol{\mathsf{x}}}_N$ are distributed over a lower-dimensional linear subspace and then select the regularization parameter as
\begin{equation}
\widehat{\lambda}_{\text{opt}} = \argmin_{\lambda \in (0,+\infty)} \N{{\boldsymbol{\mathsf{z}}}_{\text{Tik}}^\lambda -\widehat{\boldsymbol{\mathsf{x}}}}_2,\label{eq:2}
\end{equation}
where ${\boldsymbol{\mathsf{z}}}_{\text{Tik}}^\lambda$ is the minimizer of the Tikhonov
functional
\(
\min_{{\boldsymbol{\mathsf{z}}}\in\bbR^d} \N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\boldsymbol{\mathsf{y}}}}_2^2 +
\lambda \N{{\boldsymbol{\mathsf{z}}}}_2^2.
\)
The analysis and techniques related to $\widehat{\boldsymbol{\mathsf{x}}}$ are independent of the choice of the optimization scheme, whereas the selection of $\widehat{\lambda}_{\text{opt}}$ is defined by the regularization scheme.
However, it is worthwhile to mention that if ${\textrm{A}}^\dagger$ is not injective, $\widehat{{\boldsymbol{\mathsf{x}}}}$ is not a good proxy of ${\boldsymbol{\mathsf{x}}}$. For Tikhonov regularization this is not an issue as, without loss of generality, we can always assume that ${\textrm{A}}$ is injective.
Specifically, one can replace ${\boldsymbol{\mathsf{x}}}$ in~\eqref{eqn:lambda_ast} with ${\boldsymbol{\mathsf{x}}}^\dagger= {\textrm{A}}^\dagger {\textrm{A}}{\boldsymbol{\mathsf{x}}}$ and recall that ${\boldsymbol{\mathsf{z}}}_{\text{Tik}}^\lambda$ belongs to $\Null {\textrm{A}}^{\perp}$ for all $\lambda$.
Therefore, for wider applicability of the suggested framework, it is important to address the selection of $\widehat{\lambda}_{\text{opt}}$ for a wider class of regularizers and inverse problems.
In this paper, we extend the framework of \cite{DFN16} by providing the analysis for the elastic net regularization and improving the theoretical results.
Moreover, we develop an efficient, fully automated algorithm that is extensively tested on synthetic and real-world examples.
The last point is the main practical contribution of our paper.
Namely, our goal is not to introduce a new regularization paradigm but rather to design a fast and unsupervised method for determining a near optimal regularization parameter for existing regularization methods.
To do this, we analyze our problem in {two settings}:
\begin{enumerate}
\item {\it simplified case ${\textrm{A}} = \Id$} (corresponding to image denoising):
We restate the lower level problem and show that in case of a bounded ${\boldsymbol{\mathsf{w}}}$ it admits a unique minimizer, which motivates our algorithm.
Furthermore, we provide a bound on
$| \lambda_{\text{opt}} - \widehat{\lambda}_{\text{opt}}|$ for independent Bernoulli random noise and discuss the number of samples needed for optimal learning, see Proposition~\ref{prop:error_estimate}.
Though the latter model might be oversimplified, it captures the essence of the problem and our experiments confirm the results in more general settings.
\item {\it general case:} for a general matrix ${\textrm{A}}$ we provide an unsupervised, efficient, and accurate algorithm for the computation of an approximate optimal parameter.
We study the performance of our algorithm, comparing it to state-of-the-art parameter choice methods on synthetic and image denoising problems. The obtained results show that our approach achieves superior accuracy.
\end{enumerate}
\subsection{Outline}
In Section \ref{sec:setting}, we describe the main ingredients of our approach.
We define and prove bounds regarding {empirical estimators}, discuss minimizers of the elastic net \eqref{eq:3}, and define loss functions that will be used for parameter selection.
Section \ref{sec:error_analysis} provides the main theoretical results of the paper regarding loss functionals and their minimizers.
In Section \ref{sec:general_matrices} we present an efficient and accurate algorithm for the computation of an approximate optimal parameter.
We study the performance of our method through several numerical experiments on synthetic and imaging data in Section \ref{sec:experiments}. Therein, we compare our method with state-of-the-art parameter selection criteria in terms of accuracy of the solution recovery, closeness to the optimal parameter, sparse recovery and computational time.
For imaging tasks our focus is on wavelet-based denoising where we work on synthetic images and real-world brain MRIs.
We conclude with a brief discussion about future directions in Section \ref{sec:conclusion}.
The Appendix contains proofs of auxiliary results.
\subsection{Notation}
The Euclidean and the $\ell_1$-norms of ${\boldsymbol{\mathsf{u}}}=(\mathsf{u}_1,\ldots,\mathsf{u}_d)^\top$ are denoted by $\N{{\boldsymbol{\mathsf{u}}}}_2$ and
$\N{{\boldsymbol{\mathsf{u}}}}_1$, respectively. The modulus function $\SN{\cdot}$, the \emph{sign}
function $\Sgn(\cdot)$, and the \emph{positive part} function
$(\cdot)_+$ are defined component-wise for $i=1,\ldots,d,$ by
\(
\SN{{\boldsymbol{\mathsf{u}}}}_i =\SN{\mathsf{u}_i},\) \(\Sgn({\boldsymbol{\mathsf{u}}})_i = \Sgn(\mathsf{u}_i),\) and \(
(({\boldsymbol{\mathsf{u}}})_+)_i= (\mathsf{u}_i)_+,
\)
where for any $\mathsf{u}\in\bbR$
\[\Sgn(\mathsf{u})=
\begin{cases} 1,&\quad \textrm{if } \mathsf{u}>0,\\
0,&\quad \textrm{if } \mathsf{u}=0,\\-1,&\quad \textrm{if }
\mathsf{u}<0,\end{cases}\quad\textrm{and}\quad(\mathsf{u})_+=\max\{0,\mathsf{u}\}.
\]
The canonical basis of $\bbR^d$ is denoted by $\{{\boldsymbol{\mathsf{e}}}_i\}_{i=1,\ldots,d}$.
We denote the transpose of a matrix ${\textrm{M}} $ by ${\textrm{M}}^\top,$ the Moore-Penrose pseudo inverse by ${\textrm{M}}^\dagger$, and the spectral norm by $\N{{\textrm{M}}}_2$.
Furthermore, $\range({\textrm{M}})$ and $\Null({\textrm{M}})$ are the range and the null space of ${\textrm{M}},$ respectively. For a square-matrix ${\textrm{M}}$, we use $\trace ({\textrm{M}})$ to denote its trace.
The identity matrix is denoted by $\Id$ and we use $\mathbbm{1}_{\CD}$ for the indicator function of a set $\CD\subset\bbR^d$.
For any ${\boldsymbol{\mathsf{v}}}\in\bbR^d$,
${\boldsymbol{\mathsf{v}}}\otimes {\boldsymbol{\mathsf{v}}}$ is the rank one operator acting on ${\boldsymbol{\mathsf{w}}}\in\bbR^d$ as
$({\boldsymbol{\mathsf{v}}}^\top {\boldsymbol{\mathsf{w}}}) {\boldsymbol{\mathsf{v}}}$.
A random vector $\boldsymbol{\xi}$ is called \emph{sub-Gaussian} if
\[\N{\boldsymbol{\xi}}_{\psi_2} :=\sup_{\N{{\boldsymbol{\mathsf{v}}}}=1}\sup_{q\geq 1} q^{-\frac{1}{2}} \bbE\LRS{\SN{{\boldsymbol{\mathsf{v}}}^\top \boldsymbol{\xi}}^q}^{\frac{1}{q}} < +\infty.\]
The value $\N{\boldsymbol{\xi}}_{\psi_2}$ is the sub-Gaussian norm of $\boldsymbol{\xi}$, with which the space of sub-Gaussian vectors becomes a normed vector space \cite{V18}.
The (non-centered) covariance of a random vector $\boldsymbol{\xi}$ is denoted as
\[\Sigma({\boldsymbol{\xi}}):=\cov{\boldsymbol{\xi}} = \bbE [\boldsymbol{\xi} \otimes \boldsymbol{\xi}].\]
We write $a \lesssim b$ if there exists an absolute constant $C>0$ such that $a \leq Cb.$
\section{Problem setting}\label{sec:setting}
We consider the following stochastic linear inverse problem:
given a deterministic matrix ${\textrm{A}}\in\bbR^{m\times d}$, we are interested in recovering a vector ${\boldsymbol{\mathsf{x}}}\in\bbR^d$ from a noisy observation ${\boldsymbol{\mathsf{y}}}\in\bbR^m$ obeying
\begin{equation}\label{eqn:LinearModel}
{\boldsymbol{\mathsf{y}}} = {\textrm{A}} {\boldsymbol{\mathsf{x}}} +\sigma{\boldsymbol{\mathsf{w}}},
\end{equation}
where
\begin{enumerate}[label=(\textsf{A}\arabic*)]
\item\label{ass:x} the unknown datum ${\boldsymbol{\mathsf{x}}}\in\bbR^d$ is a sub-Gaussian vector, such that $\N{{\boldsymbol{\mathsf{x}}}}_{\psi_2}=1$;
\item\label{ass:V} there exists a subfamily
$1\leq i_1<\ldots i_h\leq d$ of $h$ indices, with $h\ll d$, such that
\[ \CV:=\range\LRP{\Sigma({\boldsymbol{\mathsf{x}}})}=\Span{{\boldsymbol{\mathsf{e}}}_{i_1},\ldots,{\boldsymbol{\mathsf{e}}}_{i_h}}\]
and $\Null({\textrm{A}})\cap\CV=\{\boldsymbol0\}$;
\item\label{ass:noise} the noise ${\boldsymbol{\mathsf{w}}}\in\bbR^m$ is an
independent sub-Gaussian vector, such that
$\N{\sigma{\boldsymbol{\mathsf{w}}}}_{\psi_2}\leq 1$, $\Sigma({{\boldsymbol{\mathsf{w}}}})=\Id$ and
$\sigma>0$ is the noise level.
\end{enumerate}
Conditions~\ref{ass:x} and~\ref{ass:noise} are standard assumptions on the distributions of the exact datum ${\boldsymbol{\mathsf{x}}}$ and the noise $\sigma{\boldsymbol{\mathsf{w}}}$, ensuring that the tails have fast decay.
Note also that normalization conditions on ${\boldsymbol{\mathsf{x}}}$ and ${\boldsymbol{\mathsf{w}}}$ can always be satisfied by rescaling.
Furthermore, it follows from the definition that $\CV$ is the smallest subspace such that ${\boldsymbol{\mathsf{x}}}\in \CV$, almost surely.
Thus, by~\ref{ass:V}, the exact datum ${\boldsymbol{\mathsf{x}}}$ is almost surely $h$-sparse and since
$\Null\LRP{{\textrm{A}}}\cap \CV=\{0\}$, it is a unique vector with that property.
Define now
\( \CW=\range\LRP{\Sigma({{\textrm{A}} {\boldsymbol{\mathsf{x}}}})}.\)
The following simple result was shown for an injective ${\textrm{A}}$ in \cite{DFN16}; here we extend it to the general case.
\begin{lemma}\label{lem:AV}
Under~Assumption~\ref{ass:V} we have $\dim\CW=h$ and $\CW ={\textrm{A}}\CV$.
\end{lemma}
\begin{proof}
A direct computation gives
\[\Sigma({{\textrm{A}}{\boldsymbol{\mathsf{x}}}})=\bbE[{\textrm{A}}{\boldsymbol{\mathsf{x}}}\otimes{\textrm{A}}{\boldsymbol{\mathsf{x}}}]={\textrm{A}}\Sigma({{\boldsymbol{\mathsf{x}}}}){\textrm{A}}^\top = {\textrm{A}}{\textrm{P}}
\Sigma({{\boldsymbol{\mathsf{x}}}}) ({\textrm{A}}{\textrm{P}})^\top,\]
where ${\textrm{P}}$ denotes the orthogonal projection onto $\CV$.
Assumption \ref{ass:V} says that ${\textrm{A}}$ is injective on $\CV$, and thus $\Sigma({{\textrm{A}}{\boldsymbol{\mathsf{x}}}})$ and $\Sigma({{\boldsymbol{\mathsf{x}}}})$ have the same rank.
Furthermore $({\textrm{A}}{\textrm{P}})^\top$ maps $\bbR^d$ onto $\CV$, so that
\[ \range (\Sigma({{\textrm{A}}{\boldsymbol{\mathsf{x}}}})) = ({\textrm{A}} {\textrm{P}} \Sigma({{\boldsymbol{\mathsf{x}}}})) \CV={\textrm{A}}{\textrm{P}} \CV=\CW,\]
where $\Sigma({{\boldsymbol{\mathsf{x}}}})\CV=\CV$, since $\Sigma({{\boldsymbol{\mathsf{x}}}})$ is symmetric.
\end{proof}
\subsection{Empirical estimators}
Lemma \ref{lem:AV} suggests that $\CV$ could be directly recovered if ${\textrm{A}}$ were invertible and $\CW$ were known.
In most practical situations though, neither of those assumptions is satisfied: we only have access to noisy observations and ${\textrm{A}}$ could not only be non-invertible, but also non-injective.
We will address this issue by recasting the problem to a statistical learning framework, similar to \cite{DFN16}.
Namely, suppose we are given observation samples ${\boldsymbol{\mathsf{y}}}_1,\ldots,{\boldsymbol{\mathsf{y}}}_N$ such that ${\boldsymbol{\mathsf{y}}}_i={\textrm{A}}{\boldsymbol{\mathsf{x}}}_i+\sigma{\boldsymbol{\mathsf{w}}}_i$ for $i=1,\ldots,N$, where
${\boldsymbol{\mathsf{x}}}_i,{\boldsymbol{\mathsf{w}}}_i$ and $\sigma$ are unknown, and
let \[\widehat\Sigma({\boldsymbol{\mathsf{y}}})= \frac{1}{N}\sum_{i=1}^N
{\boldsymbol{\mathsf{y}}}_i\otimes{\boldsymbol{\mathsf{y}}}_i\]
be the empirical covariance of the observations.
Standard statistical learning theory suggests that $\widehat{\Sigma}\LRP{{\boldsymbol{\mathsf{y}}}}$ is a good approximation to $\Sigma\LRP{{\boldsymbol{\mathsf{y}}}}$ provided $N$ is large enough.
As a consequence, we will show that a vector space spanned by the first $h$ eigenvectors of $\widehat{\Sigma}\LRP{{\boldsymbol{\mathsf{y}}}}$, denoted by $\widehat{\CW}$, is a good estimator of $\CW$.
To justify the above claims, observe first that since $\Sigma({\boldsymbol{\mathsf{w}}}) = \Id$ holds by~\ref{ass:noise}, we have
\begin{equation}\label{eqn:SigmaAx}\Sigma\LRP{{\boldsymbol{\mathsf{y}}}} = \Sigma\LRP{{\textrm{A}}{\boldsymbol{\mathsf{x}}}} + \sigma^2\Id.\end{equation}
Therefore, $\Sigma\LRP{{\boldsymbol{\mathsf{y}}}}$ and $\Sigma\LRP{{\textrm{A}}{\boldsymbol{\mathsf{x}}}}$ have the same eigenvectors and the spectrum of $\Sigma\LRP{{\boldsymbol{\mathsf{y}}}}$ is just a
shift of the spectrum of $\Sigma\LRP{{\textrm{A}}{\boldsymbol{\mathsf{x}}}}$ by $\sigma^2$.
Let $\lambda_1\geq\ldots\geq\lambda_h$ be the non-zero
eigenvalues of $\Sigma({\textrm{A}}{\boldsymbol{\mathsf{x}}})$, counting for multiplicity, and
$\alpha_1\geq\ldots\geq\alpha_m$ and
$\widehat\alpha_1\geq\ldots\geq\widehat\alpha_m$ be the eigenvalues
of $\Sigma({\boldsymbol{\mathsf{y}}})$ and $\widehat\Sigma({\boldsymbol{\mathsf{y}}})$, respectively.
From \eqref{eqn:SigmaAx} it follows
\begin{equation}
\label{eq:1}
\begin{cases}
\alpha_i=\lambda_i+\sigma^2, & \text{for } i=1,\ldots, h,\\
\alpha_i=\sigma^2, & \text{for }i=h+1,\ldots,m
\end{cases}.
\end{equation}
Let $\Pi$ be the (orthogonal) projection onto $\CW$, which has rank
$h$ due to Lemma~\ref{lem:AV}, and let $\widehat{\Pi}$ be
the (orthogonal) projection onto $\widehat{\CW}$.
We now show the fundamental tool of our study: that $\widehat{\Pi}$ is an accurate and an unbiased approximation of $\Pi$.
We distinguish between bounded and unbounded ${\boldsymbol{\mathsf{y}}}$ and improve upon results in \cite{DFN16}.
\begin{lemma}\label{lem:proj_error}
Assume that $\sigma^2<\lambda_h$. Given $u>0$, with probability greater than $1-2\exp(-u)$
\begin{equation}\label{eqn:unbounded}
\N{\widehat{\Pi}-\Pi}_2 \lesssim \frac{\lambda_1}{\lambda_h}
\LRP{\sqrt{\frac{{h+\sigma^2m +u}}{N}} +
\frac{{h+\sigma^2m +u}}{N} },
\end{equation}
provided $N\gtrsim \LRP{h+\sigma^2m+u}$. Furthermore, if ${\boldsymbol{\mathsf{y}}}$ is
bounded, then with probability greater than $1 - \exp(-u)$
\begin{equation}\label{eqn:bounded}
\N{\widehat{\Pi}-\Pi}_2 \lesssim \frac{\lambda_1}{\lambda_h}
\LRP{\sqrt{\frac{{\log(h+m)+u }}{N}} + \frac{\log(h+m)+u }{N} },
\end{equation}
provided $N\gtrsim (\log(2m)+u)$.
\end{lemma}
\begin{proof}
We will first show \eqref{eqn:unbounded}.
Using Theorem 9.2.4 and Exercise 9.2.5 in \cite{V18}, we have
\begin{equation}\label{eqn:low_rank_cov_estimation}
\N{\Sigma({\boldsymbol{\mathsf{y}}})-\widehat{\Sigma}({\boldsymbol{\mathsf{y}}})}_2 \lesssim
\N{\Sigma({\boldsymbol{\mathsf{y}}})}_2\LRP{\sqrt{\frac{{r+u}}{N}} +
\frac{{r+u}}{N}},\end{equation}
with probability greater than $1-2\exp(-u)$, where
$r=\trace\LRP{\Sigma({\boldsymbol{\mathsf{y}}})}/\N{\Sigma({\boldsymbol{\mathsf{y}}})}_2$ is the \emph{stable rank} of $\Sigma({\boldsymbol{\mathsf{y}}})$.
Using \(\trace\LRP{\Sigma({\boldsymbol{\mathsf{y}}})}\leq \lambda_1 h + m\sigma^2\)
and $\N{\Sigma({\boldsymbol{\mathsf{y}}})}_2=\lambda_1+\sigma^2\leq 2\lambda_1$, we get
\begin{equation}\label{eqn:low_rank_writtenout}
\N{\Sigma({\boldsymbol{\mathsf{y}}})-\widehat{\Sigma}({\boldsymbol{\mathsf{y}}})}_2 \lesssim
\lambda_1
\LRP{\sqrt{\frac{{h+\sigma^2m+u} }{N}} + \frac{{h+\sigma^2m+u }}{N}}.
\end{equation}
Let $\alpha^*=\alpha_h>\sigma^2$. By~\eqref{eq:1} it follows that $\Pi$ is the
projection onto the linear span of those eigenvectors of $\Sigma({\boldsymbol{\mathsf{y}}})$
whose corresponding eigenvalue is greater than or equal to $\alpha^*$. Using
\(
\alpha_{h}-\alpha_{h+1} = \lambda_h
\), by \eqref{eqn:low_rank_writtenout} we have
\begin{equation}\label{eqn:N_requirement} \epsilon:=\N{\Sigma({\boldsymbol{\mathsf{y}}})-\widehat\Sigma({\boldsymbol{\mathsf{y}}})}_2 <
\frac{\alpha_h -\alpha_{h+1}}{2}=\frac{\lambda_h}{2},\end{equation}
provided $N\gtrsim \LRP{h+\sigma^2m +u}$.
Let now $\Pi_{\alpha^*}$ be the projection onto the linear span of those eigenvectors of $\widehat\Sigma({\boldsymbol{\mathsf{y}}})$
whose corresponding eigenvalue is greater than or equal to $\alpha^*$.
As a consequence of Theorem 7.3.1 in \cite{bhatia97}, there exists an
eigenvalue $\widehat{\alpha}^*$ of $\widehat\Sigma({\boldsymbol{\mathsf{y}}})$ such that
\begin{alignat}{1}
|\alpha^*-\widehat{\alpha}^*|&\leq \epsilon, \text{ and } \operatorname{dim} \Pi_{\alpha^*} = \operatorname{dim}\Pi \label{eq:6c} \\
\widehat{\alpha}_j &\leq \alpha_{h+1} + \epsilon =\sigma^2 +\epsilon, \quad \forall
\widehat{\alpha}_j<\widehat{\alpha}^* \label{eq:6b} \\
\N{\Pi_{\alpha^*}-\Pi}_2 & \leq \frac{1}{\lambda_h-\epsilon}
\N{(\Id-\Pi_{\alpha^*})(\widehat\Sigma({\boldsymbol{\mathsf{y}}})-\Sigma({\boldsymbol{\mathsf{y}}}))\Pi}_2 \label{eq:6d}.
\end{alignat}
By~\eqref{eq:6c} it follows that
$\widehat{\alpha}^*=\widehat{\alpha}_h$
so that $\Pi_{\alpha^*}=\widehat{\Pi}$ and hence
\begin{equation}
\label{eq:7}
\N{\widehat{\Pi}-\Pi}_2 \leq \frac{1}{\lambda_h-\epsilon}
\N{\widehat\Sigma({\boldsymbol{\mathsf{y}}})\Pi-\Sigma({\boldsymbol{\mathsf{y}}})\Pi}_2 \leq \frac{2}{\lambda_h} \N{\widehat\Sigma({\boldsymbol{\mathsf{y}}})\Pi-\Sigma({\boldsymbol{\mathsf{y}}})\Pi}_2 .
\end{equation}
Since $\N{\widehat\Sigma({\boldsymbol{\mathsf{y}}})\Pi-\Sigma({\boldsymbol{\mathsf{y}}})\Pi}_2\leq
\N{\widehat\Sigma({\boldsymbol{\mathsf{y}}})-\Sigma({\boldsymbol{\mathsf{y}}})}_2$, the claim follows by \eqref{eqn:low_rank_cov_estimation}.
Assume now that $\N{{\boldsymbol{\mathsf{y}}}}_2 \leq \sqrt{L}$ holds almost surely and consider a family of independent $m\times
h$ matrices
\[
{\textrm{S}}_i = {{\boldsymbol{\mathsf{y}}}_i^\top {\boldsymbol{\mathsf{y}}}_i \Pi - \Sigma({\boldsymbol{\mathsf{y}}})\Pi},\qquad i=1,\ldots,N.
\]
Since $\frac{1}{N}\sum_{i=1}^N {\textrm{S}}_i= \widehat\Sigma({\boldsymbol{\mathsf{y}}})\Pi-\Sigma({\boldsymbol{\mathsf{y}}})\Pi$ we can apply the matrix Bernstein inequality for rectangular matrices (Theorem 6.1.1. in \cite{Tro15}). Thus, for $u>0$ we have
\[
\bbP\left(\N{\widehat\Sigma({\boldsymbol{\mathsf{y}}})\Pi-\Sigma({\boldsymbol{\mathsf{y}}})\Pi}_2\geq u
\right)\leq (m+h) \exp\LRP{\dfrac{-Ns^2 }{M+ 2L u/3}}
\]
where $M>0$ is a matrix variance constant independent of $m$, $h$, and $d$, such that
\begin{alignat*}{1}
\max\left\{\bbE \N{{\textrm{S}}_i^\top {\textrm{S}}_i}_2,\bbE \N{{\textrm{S}}_i{\textrm{S}}_i^\top}_2\right\} \leq M.
\end{alignat*}
A direction computation gives $M\leq L\N{\Sigma({\boldsymbol{\mathsf{y}}})}_2$.
It follows that
\[
\N{\widehat\Sigma({\boldsymbol{\mathsf{y}}})\Pi-\Sigma({\boldsymbol{\mathsf{y}}})\Pi}_2 \lesssim \lambda_1
\LRP{\sqrt{\frac{{\log(h+m)+u }}{N}} + \frac{\log(h+m)+u }{N} },
\]
holds with probability greater than $1-\exp(-u)$ for every $u>0$.
Moreover, by analogous argumentation \eqref{eqn:N_requirement} holds provided $N\gtrsim (\log(2m) + u)$, see \ref{sec:bounded_y_covs} for details.
Thus, \eqref{eqn:bounded} follows by applying \eqref{eq:7}.
\end{proof}
The previous result comes with a certain caveat.
Namely, the proof implicitly assumes that either $h$ or the spectral gap are known (which informs the choice of the approximate projector $\widehat{\Pi}$).
In practice however, the desired eigenspace can only be detected if there is a spectral gap and if it corresponds to the eigenspace we want to recover, {\em i.e.}, if $ \lambda_h>\delta$, where
\begin{equation}\label{eqn:SPGAP}
\delta =\max_{i=1,\ldots,h-1}\LRP{\lambda_i-\lambda_{i+1}}=\max_{i=1,\ldots,h-1}\LRP{\alpha_i-\alpha_{i+1}}.
\end{equation}
\begin{proposition}\label{prop:sp_gap}
Assume \eqref{eqn:SPGAP} holds. Then the empirical covariance matrix has a spectral gap at the $h\textrm{-th}$ eigenvalue, with probability greater than $1-2\exp(-u)$, provided $\delta<\lambda_h$ and $N\gtrsim \frac{\lambda_1^2}{\LRP{\lambda_h-\delta }^2}(h+u)$.
\end{proposition}
\begin{proof}
Assume $\N{\Sigma({\boldsymbol{\mathsf{y}}})-\widehat\Sigma({\boldsymbol{\mathsf{y}}})}_2 < \epsilon$ holds for $\epsilon >0$.
Since $\sup_{i=1,\ldots, m}\SN{\widehat{\alpha}_j-{\alpha}_{j}} \leq \N{\Sigma({\boldsymbol{\mathsf{y}}})-\widehat\Sigma({\boldsymbol{\mathsf{y}}})}$ we get
\[ \SN{\widehat{\alpha}_j-\widehat{\alpha}_{j+1}} \leq 2\epsilon + \SN{{\alpha}_j-{\alpha}_{j+1}},\]
by adding and subtracting $\alpha_j$ and $\alpha_{j+1}$ inside the first term.
Thus, if $j>h$ then $\SN{\widehat{\alpha}_j-\widehat{\alpha}_{j+1}} \leq 2\epsilon$, and if $j<h$ then $\SN{\widehat{\alpha}_j-\widehat{\alpha}_{j+1}} \leq 2\epsilon + \delta $.
For $j=h$ on the other hand we have $\SN{\widehat{\alpha}_h-\widehat{\alpha}_{h+1}} > \SN{{\alpha}_h-{\alpha}_{h+1}} - 2\epsilon$.
In conclusion,
\[\argmax_{i=1,\ldots,m-1} \LRP{\widehat \alpha_i - \widehat \alpha_{i-1}}
= h\]
holds provided
provided $\epsilon<\frac{\lambda_h-\delta }{4}$.
Using \eqref{eqn:low_rank_cov_estimation} the claim follows.
\end{proof}
It is clear that if $\delta>\lambda_h$ the spectral gap of the empirical covariance matrix
is at \(\argmax_{i=1,\ldots,h-1}\LRP{\lambda_i-\lambda_{i+1}}\), which is smaller than $h$.
In practice though, the situation is not as pessimistic as this observation would suggest and we can rely on a wealth of \emph{ad hoc} remedies.
We devote more attention to this question in Section \ref{sec:estimating_sparsity}, and suggest alternative heuristics for estimating the intrinsic dimension $h$.
We are ready to define our {empirical estimator} of ${\boldsymbol{\mathsf{x}}}$.
Let ${\textrm{Q}} = {\textrm{A}}\Amat^{\dagger}$ be the orthogonal projection onto $\Range({\textrm{A}})=\Null^\perp({\textrm{A}}^\top)$, and ${\textrm{P}}={\textrm{A}}^{\dagger}{\textrm{A}}$ the orthogonal projection onto $\Range({\textrm{A}}^\top)=\Null^\perp({\textrm{A}})$.
The \emph{{empirical estimator}} of ${\boldsymbol{\mathsf{x}}}$ is defined as
\begin{equation}\label{eqn:xhat}
\widehat{{\boldsymbol{\mathsf{x}}}}={\textrm{A}}^\dagger\widehat{\Pi}{\boldsymbol{\mathsf{y}}}.
\end{equation}
For \( \widehat{\boldsymbol{\Veta}} = {\boldsymbol{\mathsf{y}}} - \widehat{\Pi}{\boldsymbol{\mathsf{y}}}\) the {empirical estimator} $\widehat{{\boldsymbol{\mathsf{x}}}}$ satisfies the (empirical) inverse problem
\begin{equation}
\label{eqn:empiricalprob}{\textrm{A}}\widehat{{\boldsymbol{\mathsf{x}}}} +
{\textrm{Q}}\widehat{\Veta} = {\textrm{Q}}{\boldsymbol{\mathsf{y}}}.
\end{equation}
A direct consequence of \eqref{eqn:empiricalprob} is that minimizers of the empirical and of the original problem coincide.
\begin{lemma}
Let
\(\widehat{{\boldsymbol{\mathsf{z}}}}^\lambda{({\boldsymbol{\mathsf{y}}})}=\argmin_{{\boldsymbol{\mathsf{z}}}} \N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\textrm{Q}}{\boldsymbol{\mathsf{y}}}}^2 + \lambda\, J({\boldsymbol{\mathsf{z}}}).\)
Then $\widehat{{\boldsymbol{\mathsf{z}}}}^\lambda{({\boldsymbol{\mathsf{y}}})}={\boldsymbol{\mathsf{z}}}^\lambda({\boldsymbol{\mathsf{y}}})$.
\end{lemma}
\begin{proof}
We compute
\( \N{{\textrm{A}} {\boldsymbol{\mathsf{z}}} - {\boldsymbol{\mathsf{y}}}}_2^2 =\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}} -{\textrm{Q}}{\boldsymbol{\mathsf{y}}} + ({\textrm{Q}}-\Id){\boldsymbol{\mathsf{y}}}}_2^2.\)
Since ${\textrm{Q}}$ is an orthogonal projection onto $\Range({\textrm{A}})$ it follows $({\textrm{Q}}-\Id){\boldsymbol{\mathsf{y}}}\in\Range^\perp({\textrm{A}})$. Using Pythagoras' theorem we thus have
\( \N{{\textrm{A}} {\boldsymbol{\mathsf{z}}} - {\boldsymbol{\mathsf{y}}}}_2^2 = \N{{\textrm{A}}{\boldsymbol{\mathsf{z}}} -{\textrm{Q}}{\boldsymbol{\mathsf{y}}}}_2^2 + \N{({\textrm{Q}}-\Id){\boldsymbol{\mathsf{y}}}}_2^2.\)
Since the second term does not depend on ${\boldsymbol{\mathsf{z}}}$ we get
\[ \argmin_{\boldsymbol{\mathsf{z}}}\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\boldsymbol{\mathsf{y}}}}^2_2 +\lambda J({\boldsymbol{\mathsf{z}}}) = \argmin_{\boldsymbol{\mathsf{z}}} \N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\textrm{Q}}{\boldsymbol{\mathsf{y}}}}^2_2 +\lambda\,J({\boldsymbol{\mathsf{z}}}).\]
\end{proof}
The definition of $\widehat{{\boldsymbol{\mathsf{x}}}}$ is independent of the choice of the optimization scheme.
In the following, we use $\widehat{{\boldsymbol{\mathsf{x}}}}$ to learn a nearly optimal regularization parameter for elastic net minimization.
Before doing so, we note that one might want to consider $\widehat{{\boldsymbol{\mathsf{x}}}}$ as an approximate solution by itself, and completely avoid regularization and thus the issue of parameter choice.
Experimental evidence in Section \ref{sec:experiments} shows that when ${\textrm{A}}$ is not injective, the training set size $N$ is small, or when the noise level is small, $\| \widehat{\boldsymbol{\mathsf{x}}} - {\boldsymbol{\mathsf{x}}}\|$ is larger than $\| {\boldsymbol{\mathsf{z}}}^\lambda - {\boldsymbol{\mathsf{x}}}\|,$ some of which has also been observed in \cite{DFN16}.
In addition, $\widehat{\boldsymbol{\mathsf{x}}}$ does not preserve the structure of the original signal, \emph{e.g.}, $\widehat{\boldsymbol{\mathsf{x}}}$ will in general not be sparse for a sparse ${\boldsymbol{\mathsf{x}}}$, and regularization is needed.
Lastly, we remind that we are interested in using an estimator $\widehat{{\boldsymbol{\mathsf{x}}}}$ for which $\SN{\lambda_{\text{opt}} -\widehat{\lambda}_{\text{opt}}}$ is small with high probability (\emph{i.e.} the one that can be used to derive an accurate parameter selection) and we are not interested in directly controlling $\N{{\boldsymbol{\mathsf{x}}}- \widehat{{\boldsymbol{\mathsf{x}}}}}$, which is the {goal in manifold learning} \cite{bel06}.
\subsection{Elastic net minimization}\label{sec:enets}
From now on we focus on the parameter choice for the elastic net, where $J({\boldsymbol{\mathsf{z}}}) = \N{{\boldsymbol{\mathsf{z}}}}_1 + \alpha\N{{\boldsymbol{\mathsf{z}}}}_2^2$,
so that
\begin{equation}\label{eq:enet_reg}
{\boldsymbol{\mathsf{z}}}^{\lambda}({\boldsymbol{\mathsf{y}}}) = \argmin_{{\boldsymbol{\mathsf{z}}}\in\bbR^m} \N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\boldsymbol{\mathsf{y}}}}_2^2 + \lambda \LRP{\N{{\boldsymbol{\mathsf{z}}}}_1 + \alpha\N{{\boldsymbol{\mathsf{z}}}}_2^2}.
\end{equation}
The term $\N{{\boldsymbol{\mathsf{z}}}}_1$, enforces the sparsity of the solution, whereas $\N{{\boldsymbol{\mathsf{z}}}}_2^2$ enforces smoothness and ensures that in case of highly correlated features we can correctly retrieve all the relevant ones.
We first recall some basic facts about existence, uniqueness and sensitivity of elastic net solutions with respect to regularization parameters \cite{JLS09}.
\begin{lemma}\label{lem:solutions}
The elastic net functional is strictly convex and coercive.
Moreover, for each $\lambda>0$ the minimizer of \eqref{eqn:regularizer} exists, is unique and the mapping $\lambda\mapsto{\boldsymbol{\mathsf{z}}}^\lambda$ is continuous with respect to $\lambda>0$.
\end{lemma}
In the remainder of this paper we will recast \eqref{eq:enet_reg} as
\begin{align}\label{eqn:enets_minimization}
{\boldsymbol{\mathsf{z}}}^{t}({\boldsymbol{\mathsf{y}}}) & = \argmin_{{\boldsymbol{\mathsf{z}}}\in\bbR^m} t\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}-{\boldsymbol{\mathsf{y}}}}_2^2 + (1-t)\LRP{\N{{\boldsymbol{\mathsf{z}}}}_1 + \alpha\N{{\boldsymbol{\mathsf{z}}}}_2^2},
\end{align}
where $t\in[0,1]$ and $\alpha>0$ is a fixed parameter.
For $t\in(0,1)$ the solutions of \eqref{eqn:enets_minimization} correspond to solutions of \eqref{eq:enet_reg} for $\lambda=\frac{1-t}{t}.$
On the other hand, for $t=0$ we get ${\boldsymbol{\mathsf{z}}}^0({\boldsymbol{\mathsf{y}}})=\boldsymbol{0}$, and for $t=1$ we define ${\boldsymbol{\mathsf{z}}}^1({\boldsymbol{\mathsf{y}}}):={\boldsymbol{\mathsf{x}}}^\alpha,$ where
\begin{equation}\label{eqn:xalpha}
{\boldsymbol{\mathsf{x}}}^\alpha=\argmin_{{\boldsymbol{\mathsf{z}}}\in\CN} \LRP{\N{{\boldsymbol{\mathsf{z}}}}_1 + \alpha\N{{\boldsymbol{\mathsf{z}}}}_2^2},
\text{ for } \CN=\{{\boldsymbol{\mathsf{z}}}\in\bbR^m\mid {\textrm{A}}^\top {\textrm{A}}{\boldsymbol{\mathsf{z}}}={\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}\}.
\end{equation}
This definition is driven by the following observations.
First, the set $\CN= {\textrm{A}}^\dagger {\boldsymbol{\mathsf{y}}} \oplus \ker\LRP{{\textrm{A}}}$ is non-empty (since ${\textrm{A}}$ has finite rank).
Furthermore, it was shown in \cite{DDR09} and \cite{JLS09} that in case of elastic nets minimization
\begin{equation}
\lim_{t\to 1} {\boldsymbol{\mathsf{z}}}^t = {{\boldsymbol{\mathsf{x}}}}^\alpha\label{eq:4}
\end{equation}
In other words, ${\boldsymbol{\mathsf{x}}}^\alpha$ plays the role of the Moore-Penrose solution in linear regularization schemes \cite{zbMATH00936298}.
By Lemma \ref{lem:solutions} the minimizer of
\eqref{eqn:enets_minimization} always exist and is unique, the map $t\mapsto {\boldsymbol{\mathsf{z}}}^t$ is continuous for $t\in(0,1)$.
Equation ~\eqref{eq:4} implies that $t\mapsto {\boldsymbol{\mathsf{z}}}^t$ is continuous at $t=1$, and later in \eqref{eq:5} we show that the continuity also holds at $t=0$.
\paragraph{Solution via soft-thresholding. }
The elastic net does not admit a closed form solution in case of a general forward matrix ${\textrm{A}}$.
In Zou and Hastie \cite{ZH05} the elastic net problem is recast as a Lasso problem with augmented data, which can then be solved by many different algorithms (\emph{e.g.} the LARS method \cite{efron2004least}).
Alternative algorithms compute the elastic net minimizer directly, and are generally either of the \emph{active set} \cite{JLS09} or the \emph{iterative soft-thresholding}-type \cite{DDR09}.
Here we adhere to iterative soft-thresholding, and rework the arguments in \cite{DDR09} to show that the solution to \eqref{eqn:enets_minimization} can be obtained through fixed point iterations for all $t\in[0,1]$.
To begin, define the soft-thresholding function by
\begin{align}\label{eqn:Stw} \CS_\tau(\mathsf{u}) &= \Sgn(\mathsf{u})\LRP{\SN{\mathsf{u}}-\frac{\tau}{2}}_+
\end{align}
and the corresponding soft-thresholding operator
$\boldsymbol{\CS}_\tau(\boldsymbol{{\boldsymbol{\mathsf{u}}}})$, acting component-wise on vectors ${\boldsymbol{\mathsf{u}}}\in\bbR^m$.
The next lemma is a direct reworking of the arguments in \cite{DDR09} and states that \eqref{eqn:enets_minimization} is a fixed point of a contractive map.
\begin{lemma}\label{lem:contraction}
The solution to \eqref{eqn:enets_minimization}, for
${\textrm{A}}\in\bbR^{m\times d}, {\boldsymbol{\mathsf{y}}}\in\bbR^m$ and $t\in (0,1)$, satisfies ${\boldsymbol{\mathsf{z}}}= \CT_t({\boldsymbol{\mathsf{z}}})$, where the map $\CT_t:\bbR^d\rightarrow\bbR^d$ is a contraction and is defined by
\begin{equation}\label{eqn:TtDefn}{\boldsymbol{\mathsf{z}}}^t =\CT_t({\boldsymbol{\mathsf{z}}}) = \frac{1}{\tau
t +(1-t)\alpha}\boldsymbol{\CS}_{1-t}\LRP{t\LRP{\theta\Id-{\textrm{A}}^\top{\textrm{A}}}{\boldsymbol{\mathsf{z}}}+t{\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}},
\end{equation}
with the Lipschitz constant
\[
\dfrac{t (\sigma^2_M- \sigma_m^2) }{t (\sigma^2_M+ \sigma_m^2)
+2\alpha (1-t)}<1,
\]
where $\theta= \frac{\sigma_m^2+\sigma_M^2}{2}$, and $\sigma_m$ and
$\sigma_M$ are the smallest and the largest singular values of the matrix~${\textrm{A}}$, respectively.
\end{lemma}
For $t=0$, the solution is ${\boldsymbol{\mathsf{z}}}^0=\boldsymbol{0}$, which is consistent with~\eqref{eqn:TtDefn}.
Furthermore, by~\eqref{eqn:Stw} and~\eqref{eqn:TtDefn}, we get
\begin{equation}
\label{eq:5}
{\boldsymbol{\mathsf{z}}}^t =0 \qquad \text{ if }\quad 0\leq t\leq \frac{1}{1+2 \N{{\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}}}.
\end{equation}
Our definition of the solution at $t=1$ in \eqref{eqn:xalpha} also satisfies ${\boldsymbol{\mathsf{z}}}= \CT_t({\boldsymbol{\mathsf{z}}})$ since $\boldsymbol{\CS}_{1-t}$ is identity and thus
${\boldsymbol{\mathsf{z}}}^{t=1}({\boldsymbol{\mathsf{y}}})={\textrm{A}}^\dagger{\boldsymbol{\mathsf{y}}}$, though $\CT_1$ is not a contraction.
In summary, the solutions are consistent with Lemma \ref{lem:solutions}, as expected.
\paragraph{Closed form solution. }
In the case of orthogonal design \cite{ZH05}, \emph{i.e.} ${\textrm{A}}^\top{\textrm{A}}=\Id$, the solution of \eqref{eqn:regularizer} is given by
\begin{equation}\label{eqn:Naive} {\boldsymbol{\mathsf{z}}}^\lambda = \frac{(\lvert{\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}\lvert -\lambda/2)_+}{1+\alpha\lambda}\Sgn({\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}).\end{equation}
Plugging $\lambda=\frac{1-t}{t}$ into (\ref{eqn:Naive}) we have
\begin{equation}\label{eqn:naive} {\boldsymbol{\mathsf{z}}}^{t}({\boldsymbol{\mathsf{y}}}) =
\frac{(t(1+2\lvert{\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}\lvert)-1)_+}{2(t(1-\alpha)+\alpha)}\Sgn\LRP{{\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}}.
\end{equation}
\subsection{Quadratic loss functionals}
\label{subsec:qlf}
To select the regularization parameters we go back to the first principles and consider quadratic loss functionals.
\begin{definition}\label{defn:loss}
Functions $R,\widehat{R}\colon[0,1]\rightarrow\bbR$, defined by
\begin{equation}\label{eqn:loss} R(t) = \N{{\boldsymbol{\mathsf{z}}}^t({\boldsymbol{\mathsf{y}}})-{\boldsymbol{\mathsf{x}}}}_2^2, \quad \widehat{R}(t) = \N{{\boldsymbol{\mathsf{z}}}^t({\boldsymbol{\mathsf{y}}})-\widehat{{\boldsymbol{\mathsf{x}}}}}_2^2,\end{equation}
are called the \emph{true} and the \emph{empirical quadratic loss}, respectively.
Furthermore, define
\begin{equation}\label{eqn:topt} t_{\text{opt}} = \argmin_{t\in[0,1]} R(t), \quad \widehat{t}_{\text{opt}}= \argmin_{t\in[0,1]} \widehat{R}(t).\end{equation}
\end{definition}
In view of Lemma \ref{lem:solutions} and the discussion in Section \ref{sec:enets}, the benefits of recasting $\lambda$ to $[0,1]$ are clear: $R$ and $\hat R$ are both continuous, defined on a bounded interval, and, hence, achieve a minimum.
Thus, our aim is to minimize $\widehat{R}$ while ensuring $\SN{t_{\text{opt}}-\widehat{t}_{\text{opt}}}$ is small.
Let us discuss some difficulties associated with elastic net minimization which need to be addressed.
On one hand, d a closed form solution of \eqref{eqn:enets_minimization} is available only when ${\textrm{A}}^\top{\textrm{A}}=\Id$ and is otherwise only approximated.
Furthermore, as we will see below, loss functionals $R$ and $\widehat{R}$ are globally neither differentiable nor convex, but rather only piecewise.
These two issues suggest that their minimizers in general cannot be analyzed in full detail.
Therefore, in the following we split the analysis into a simplified case for ${\textrm{A}}^\top{\textrm{A}} = \Id$ where can provide guarantees, and the general case where we provide an efficient algorithm.
Furthermore, we need to amend the empirical loss function $\widehat{R}$ in the case when ${\textrm{A}}$ is non-injective.
This is due to the fact that in case of non-linear methods $R(t)$ cannot be reliably estimated outside the kernel of ${\textrm{A}}$, see \cite{E09} and Figure \ref{fig:loss_functionals}.
We follow the idea of SURE-based methods \cite{GEE11}, which provide an unbiased estimate of $R(t)$ by projecting the regularized solution onto $\Null^\perp({\textrm{A}})$.
Namely, we define \emph{projected and modified loss functions} $\widehat{R}_{\FP},\widehat{R}_{\FM}\colon[0,1]\rightarrow\bbR$ by
\begin{equation}\label{eqn:loss_projeced}
\widehat{R}_{\FP}(t) =\N {{\textrm{P}} {\boldsymbol{\mathsf{z}}}^t({\boldsymbol{\mathsf{y}}})-\widehat{{\boldsymbol{\mathsf{x}}}}}^2_2, \text{ and } \widehat{R}_{\FM}(t) =\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}^t({\boldsymbol{\mathsf{y}}}) - \widehat{\Pi}{\boldsymbol{\mathsf{y}}}}_2^2,
\end{equation}
where ${\textrm{P}}={\textrm{A}}^\dagger{\textrm{A}}$ is the orthogonal projection onto $\Null^\perp({\textrm{A}})$.
Define also
\begin{equation}\label{eqn:topt_proj}
t_\FP= \argmin_{t\in[0,1]} \widehat{R}_\FP(t), \text{ and } t_\FM= \argmin_{t\in[0,1]} \widehat{R}_\FM(t)
\end{equation}
Note that to define $\widehat{R}_{\FM}(t)$ we used the fact that $\widehat{{\boldsymbol{\mathsf{x}}}} = {\textrm{A}}^\dagger \widehat{\Pi}{\boldsymbol{\mathsf{y}}}$, and thus compared to $\widehat{R}_{\FP},$ we avoid the computation of the Moore-Penrose inverse ${\textrm{A}}^\dagger$, which might be either costly to compute or indeed numerically unstable if ${\textrm{A}}$ is poorly conditioned.
As we will show in Section \ref{sec:experiments}, and can see in the right-most panel in Figure \ref{fig:loss_functionals}, using $\widehat{R}_{\FP}$ and $\widehat{R}_{\FM}$ instead of $\widehat{R}$ when ${\textrm{A}}$ is non-injective dramatically improves the performance.
Note that projecting onto $\Null^\perp({\textrm{A}})$ affects makes the loss functional smoother (dampening the gradients).
\begin{figure}[ht!]
\centering
\begin{minipage}[b]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{full_rank.pdf}
\end{minipage}
\begin{minipage}[b]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{low_rank.pdf}
\end{minipage}
\begin{minipage}[b]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{all_together_now.pdf}
\end{minipage}
\caption{Empirical and true losses for $m=500$, $d=60$, $h=5$, $N=50$, $\alpha=10^{-3}$, $\sigma =0.08$, and zero mean isotropic Gaussians ${\boldsymbol{\mathsf{x}}}$ and ${\boldsymbol{\mathsf{w}}}$. In the left panel ${\textrm{A}}$ is injective and $\widehat{t}_{\textrm{opt}}$ is a good proxy for $t_{\textrm{opt}}$. In the middle panel $\rank({\textrm{A}})=40$ and we see that $\widehat{t}_{\textrm{opt}}$ does not approximate $t_{\textrm{opt}}$ well. On the other hand, the right panel shows that in case of a non-injective matrix, $\widehat{R}_\FM$ and $\widehat{R}_\FP$ improve the performance}
\label{fig:loss_functionals}
\end{figure}
\section{Parameter error}\label{sec:error_analysis}
Since the elastic net solution is in general not available in closed form, a rigorous study of the parameter error is unfeasible in full generality.
Therefore, we restrict our attention to simplified cases, though we emphasize that our approach in practice performs well on significantly broader model assumptions, which we will show in Section \ref{sec:experiments}.
In case of orthogonal design ${\textrm{A}}^\top{\textrm{A}}=\Id$ we can, without loss of generality, assume ${\textrm{A}}=\Id$ (otherwise redefine ${\boldsymbol{\mathsf{y}}}$ as ${\textrm{A}}^\top{\boldsymbol{\mathsf{y}}}$).
Let now ${\boldsymbol{\mathsf{y}}} = {\boldsymbol{\mathsf{x}}}+\sigma{\boldsymbol{\mathsf{w}}}$ and assume $\lvert\mathsf{y}_1\lvert\geq\ldots\geq\lvert\mathsf{y}_m\lvert$.
Plugging \eqref{eqn:naive} into \eqref{eqn:loss} we get
\[ R(t) = \sum_{i=1}^m \left(\frac{(t(1+2\lvert\mathsf{y}_i\lvert)-1)_+}{2(t(1-\alpha)+\alpha)}\Sgn(\mathsf{y}_i) - \mathsf{x}_i\right)^2.\]
Define {$\edge_i= {{1+2\SN{\mathsf{y}_i}}}$, for $i =1,\ldots, m$}.
Loss function $R(t)$ is continuous on $[0,1]$, and differentiable on intervals
\begin{align*}
\CI_0 &= \Big[0, \edge_1^{-1}\Big),\quad \CI_{m} = \Big(\edge_{m}^{-1}, 1\Big], \text{ and } \CI_k = \LRP{\edge_k^{-1}, \edge_{k+1}^{-1}}, \textrm{ for } k=1,\ldots, m-1.
\end{align*}
Considering one interval at a time a direct computation yields that for $k=1,\ldots, m-1$ the minimizer of $R\lvert_{{\CI}_k}$ is
\begin{align}\label{eqn:tstark} t^{\star,k} &= \varphi_k\left(\frac{\sum_{i=1}^k a_i d_i}{\sum_{i=1}^k a_ic_i} \right), \quad
\textrm{ where }\nonumber\\
\varphi_k(t)&=\begin{cases} \edge_k, &\textrm{for } t<\edge_k \\t, & \textrm{for } t\in\CI_k \\ \edge_{k+1}, &\textrm{for } t>\edge_{k+1}
\end{cases}, \textrm{ and } \quad \begin{aligned}
a_i &= \Sgn(\mathsf{y}_i)(1+2\alpha\SN{\mathsf{y}_i}), \\
c_i &= \Sgn(\mathsf{y}_i)+2\mathsf{x}_i(\alpha-1) + 2\mathsf{y}_i, \\
d_i &= \Sgn(\mathsf{y}_i)+2\alpha\mathsf{x}_i.
\end{aligned}
\end{align}
An analogous expression holds for $k=m$, whereas $R(t)$ is constant on $\CI_0$, as argued in \eqref{eq:5}.
Therefore, the minimizer of $R(t)$ is $t_{\textrm{opt}}=\argmin_{k=0,\ldots,m} R(t^{\ast,k})$.
The empirical loss function $\widehat{R}(t)$ is also continuous on $[0,1]$ and is piecewise differentiable on the same set of intervals since they depend only on ${\boldsymbol{\mathsf{y}}}$.
Consequently, minimizers $\widehat{t}^{\star,k}$ of $\widehat{R}(t)$ are also of the form \eqref{eqn:tstark}, where we only ought to replace $\mathsf{x}_i$ by $\widehat{\mathsf{x}}_i$.
Notice that unless further assumptions are made,
minimizers $t_\textrm{opt}$ and $\widehat{t}_\textrm{opt}$ are not given explicitly: we still need to evaluate $R(t)$ and $\widehat{R}(t)$ at $m+1$ locations, and it is not clear that there are no local minima or that the minimizer is unique.
We will now show that in case of bounded sub-Gaussian noise there is indeed only one minimum and that it concentrates near $t=1.0$ for moderate noise levels.
This analysis will also give a theoretical intuition that will drive our algorithm.
Furthermore, we will show that in a simplified case of Bernoulli noise we get explicit bounds on the parameter error.
\subsection{Bounded noise}\label{sec:bounded_noise}
Consider now the case of bounded noise such that there is a gap between the noise and the signal.
We show that there exists a unique minimizer and there are no local minima.
For simplicity of computation, we let $\alpha=1$, though the results hold for all $\alpha>0$.
Let ${\boldsymbol{\mathsf{y}}}= {\boldsymbol{\mathsf{x}}}+\sigma{\boldsymbol{\mathsf{w}}}$ and assume ${\boldsymbol{\mathsf{x}}} = \LRP{\mathsf{x}_1,\ldots,\mathsf{x}_h, 0,\ldots,0}^\top$ where $\SN{\mathsf{x}_i}>2\sigma\SN{\mathsf{w}_j}$ for all $i=1,\ldots,h$ and $j=1,\ldots, m$.
Without loss of generality, we assume that ${\boldsymbol{\mathsf{y}}}$ is ordered so that
\begin{equation*}\SN{\mathsf{x}_i+\sigma\mathsf{w}_i}> \SN{\mathsf{x}_j+\sigma\mathsf{w}_j} \text{ for }1\leq i < j\leq h \text{ and } \SN{\mathsf{w}_i}\geq\SN{\mathsf{w}_j}\text{ for }1\leq i<j\leq m.
\end{equation*}
The loss functional $R(t) = \N{{\boldsymbol{\mathsf{z}}}^t - {\boldsymbol{\mathsf{x}}}}_2^2$ is thus piecewise differentiable on intervals $\CI_k$, where $\edge_i = 1+2\SN{\mathsf{x}_i+\sigma\mathsf{w}_i}$ for $i=1,\ldots,h$, and
$\edge_i = 1+2\sigma\SN{\mathsf{w}_i}$ {for } $i=h+1,\ldots,m$.
Also, we have $\edge_i\geq\edge_j$ for $i\leq j$.
We will show that $R(t)$ is decreasing\footnote{We will show that $R(t)$ can be monotonically increasing only for a large enough $t$. Thus, for all $t$ smaller than that value (denoted as $\vartheta_j$), it will be a monotonously decreasing function} for all $t\leq \mathsf{b}_{h+1}^{-1}$.
Let thus $t\in\CI_j$ for $j<h$.
The function $R(t)$ is continuously differentiable in $\CI_j$, so it is sufficient to show that $R'(t)$ is positive.
By a direct computation it follows
\begin{equation}\label{eqn:Rderiv} R'(t) \geq 0 \text{ if } t\geq \frac{\sum_{i=1}^j \edge_i\LRP{1+2\Sgn(\mathsf{y}_i)\mathsf{x}_i }}{\sum_{i=1}^j \edge_i^2}=:\vartheta_j.\end{equation}
It suffices to show $\vartheta_j\geq \edge_{j+1}^{-1}.$
Since $\Sgn(\mathsf{y}_i)=\Sgn(\mathsf{x}_i)$ for $i\leq j < h$, we have
\begin{align}\label{eqn:outside_the_edge}
\edge_{j+1}\LRP{1+2\SN{\mathsf{x}_i}} - \edge_i >4\SN{\mathsf{x}_i} \SN{\mathsf{x}_{j+1}+\sigma\mathsf{w}_{j+1}} >0.
\end{align}
Therefore, $ \edge_i\LRP{1+2\Sgn(\mathsf{y}_i)\mathsf{x}_i }\geq \edge_{j+1}^{-1}\edge_i^2$, and the claim follows.
Extending the same analysis to $t\in\CI_h$ what we ought to show is $\edge_{h+1}\LRP{1+2\SN{\mathsf{x}_i}} - \edge_i>0$ for $1\leq i \leq h$.
A direct computation gives
\[\edge_{h+1}\LRP{1+2\SN{\mathsf{x}_i}} - \edge_i > 2 \sigma\SN{\mathsf{w}_{h+1}}\LRP{1+2\SN{\mathsf{x}_i}} >0.\]
Hence, $t_{\text{opt}} > \edge_{h+1}^{-1}$, as desired.
We will now show that $R(t)$ admits only one minimizer.
Assume there exists $t^\star$ such that $t^\star\in\CI_{j^\star}$ for some $j^\star >h$ and $R'(t^\star)=0$.
This means
\[ t^\star = \frac{\sum_{i=1}^h \edge_i\LRP{1+2\Sgn(\mathsf{y}_i)\mathsf{x}_i } + \sum_{i=h+1}^{j^\star} \edge_i }{\sum_{i=1}^{j^\star} \edge_i^2}=\vartheta_{j^\star}, \text{ and } \edge_{j^\star}^{-1} < \vartheta_{j^\star}<\edge_{j^\star+1}^{-1}.\]
We proceed by induction showing that $R(t)$ is increasing for all $j>j^\star$.
For $t\in\CI_j$ with $j>h$, it follows
\[ R'(t) \leq 0 \text{ if } t\leq \frac{\sum_{i=1}^h \edge_i\LRP{1+2\Sgn(\mathsf{y}_i)\mathsf{x}_i } + \sum_{i=h+1}^{j} \edge_i }{\sum_{i=1}^{j} \edge_i^2}=:\vartheta_j.\]
Let us show $\vartheta_j<\edge_j^{-1}$ for $j=j^\star+1$.
We have
\begin{align*}
\vartheta_j &= \frac{\sum_{i=1}^h \edge_i\LRP{1+2\Sgn(\mathsf{y}_i)\mathsf{x}_i } + \sum_{i=h+1}^{j} \edge_i }{\sum_{i=1}^{j} \edge_i^2}= \frac{\vartheta_{j^\star}{\sum_{i=1}^{j^\star} \edge_i^2} + \edge_j}{\sum_{i=1}^{j} \edge_i^2}\leq \edge_{j}^{-1},
\end{align*}
where we used the fact $\vartheta_{j^\star}\leq \edge_j^{-1} = \edge_{j^\star+1}^{-1}$, and $\edge_j>\edge_{j^\star}$.
The rest of the proof then follows by mathematical induction.
Analogous computation yields the same type of a result for the empirical loss function $\widehat{R}$.
\begin{lemma}\label{lem:non_bernoulli}
Let the above assumptions hold.
Loss function $R(t)$ is then either monotonically decreasing on the entire interval $[0,1]$, or it is decreasing until some interval $\CI_{j^\star}$, for $j^\star > h+1$ where it achieves a (unique) minimum, and it is monotonically increasing on all the subsequent intervals.
The same holds for $\widehat{R}$ and all $\alpha>0$.
\end{lemma}
\subsection{Bernoulli noise}\label{sec:bernoulli_noise}
Lemma \ref{lem:non_bernoulli} states that $R(t)$ and $\widehat{R}(t)$ achieve a unique minimum in $[0,1]$, and that they are monotonically decreasing before, and monotonically increasing after this minimum.
Furthermore, the minimizer is bigger than $(1+2\sigma\SN{\mathsf{w}_{h+1}})^{-1}$, which means that for moderate noise levels, it will be close to $1$.
The issue is that minimizers $\vartheta_{j^\star}$ and $\widehat{\vartheta}_{\widehat{j}^\star}$ do not need to lie in the same interval, that is $j^\star\neq\widehat{j}^\star$, and thus they cannot be directly compared.
Instead, we consider a simplified model that still encodes the main features of the problem.
In particular, let
\[ {\boldsymbol{\mathsf{y}}} = {\boldsymbol{\mathsf{x}}} + \sigma{\boldsymbol{\mathsf{w}}}, \textrm{ where } \bbP\LRP{ \mathsf{w}_i = \pm 1} = \frac{1}{2},\]
and assume
\( {\boldsymbol{\mathsf{x}}}=(\mathsf{x}_1,\ldots,\mathsf{x}_h,0,\ldots,0)^\top,\) and \(\SN{\mathsf{x}_i}\geq 2\sigma,\) for \(i=1,\ldots, h\).
As before, without loss of generality we can assume $\lvert\mathsf{y}_1\lvert\geq\lvert\mathsf{y}_2\lvert\geq\ldots\geq\lvert\mathsf{y}_m\lvert$. It then follows
\( \edge_i=2\SN{\mathsf{x}_i\pm\sigma}+1,\) for $1\leq i \leq h$, and $\edge_i=2\sigma + 1$ otherwise. Moreover, $\edge_j>\edge_{h+1},$ for all $j=1,\ldots, h$.
In the following, we will for the sake of simplicity consider the case $\alpha=1$.
The details regarding the general case, $\alpha \neq 1$, are in the Appendix.
First, as in Section \ref{sec:bounded_noise} we know that $t_\textrm{opt}\geq\edge^{-1}_{h+1}.$
We can now explicitly compute the minimizer of $R(t)$.
\begin{lemma}\label{lem:bernoulli_minimizer}
True loss functional $R(t)$ is minimized for $t_{\textrm{opt}} = \min\{t^\star,1\}\in [\edge_{h+1}^{-1},1]$, where
\begin{equation}\label{eqn:bernoulli_tstar} t^\star = \frac{\sum_{i=1}^h\edge_i\LRP{1+2\Sgn\LRP{\mathsf{y}_i}} + \edge_{h+1}(m-h)}{\sum_{i=1}^h \edge_i^2 + (m-h)\edge_{h+1}^2}.\end{equation}
\end{lemma}
\begin{proof}
Considering \eqref{eqn:Rderiv} for $t\in \LRP{\edge_{h+1}^{-1},1}$ we get
\begin{align}\label{eqn:der_at_one}
2R'(t) &= \LRP{\sum_{i=1}^h \edge_i^2 + (m-h)\edge_{h+1}^2}t -\sum_{i=1}^h\edge_i\LRP{1+2\Sgn\LRP{\mathsf{y}_i}} - \edge_{h+1}(m-h).
\end{align}
The root of \eqref{eqn:der_at_one} is exactly \eqref{eqn:bernoulli_tstar}.
Arguing as in \eqref{eqn:outside_the_edge} we have $t^\star>\frac{1}{\edge_{h+1}}$.
Restricting to $[0,1]$ the claim follows.
\end{proof}
\begin{remark}
The minimizer given by Lemma \ref{lem:bernoulli_minimizer} will be in $[0,1]$ provided $\sum_{i=1}^h \SN{\mathsf{y}_i}\leq (m-h)\sigma$ and $h\leq m/2$.
\end{remark}
For the empirical loss function it is in general not true that $\widehat{\mathsf{x}}_i=0$ for $i>h$, nor is $\frac{\mathsf{y}_i-\widehat{\mathsf{x}}_i}{\sigma}$ a Bernoulli random variable.
However, ${\boldsymbol{\mathsf{y}}}$ and $\edge_i$'s remain the same, and an entirely analogous computation gives
\begin{equation}\label{eqn:t-t} \widehat{t}^\star = t^\star +
\frac{\sum_{i=1}^m \edge_i \Sgn\LRP{\mathsf{y}_i}(\mathsf{x}_i-\widehat{\mathsf{x}}_i)
}{\sum_{\edge_i>1/t^{\star}}\edge_i^2}.
\end{equation}
We can now bound the approximation error for the optimal regularization parameter.
\begin{theorem}\label{prop:error_estimate}
Assume that $t_{\textrm{opt}}<1$ and $\sigma\frac{h}{m}<1$. Given $u>0$, with probability of at least $1-2\exp(-u)$ we have
\begin{equation}\label{eqn:bernoulli_unbounded} \SN{t_{\textrm{opt}} - \widehat{t}_{\textrm{opt}}} \leq
\frac{\lambda_1}{\lambda_h}
\LRP{\sqrt{\frac{h+\sigma^2 m+u }{N}} +
\frac{h+\sigma^2 m+u}{N}} + \sigma\sqrt{\frac{h}{m}},\end{equation}
provided $N\gtrsim \LRP{h+\sigma^2m +u}$.
{Assume now ${\boldsymbol{\mathsf{y}}}$ is bounded. With
probability greater than $1 - 3 \exp\LRP{-u}$ we then have
\begin{equation}\label{eqn:bernoulli_bounded} \SN{t_{\textrm{opt}} - \widehat{t}_{\textrm{opt}}} \leq
\frac{\lambda_1}{\lambda_h}
\LRP{\sqrt{\frac{\log(h+m)+u}{N}} +
\frac{\log(h+m)+u}{N} } + \sigma\sqrt{\frac{h}{m}} ,\end{equation}
provided $N\gtrsim \LRP{\log{2m} + u}$. }
\end{theorem}
\begin{proof}
By assumption and Lemma \ref{lem:bernoulli_minimizer} $t_{\textrm{opt}} =t^\star$. We can now rewrite \eqref{eqn:bernoulli_tstar} and \eqref{eqn:t-t} as
\[ t^\star = \frac{\N{{\boldsymbol{\mathsf{b}}}}_1 + 2\EUSP{
(\Sgn{{\boldsymbol{\mathsf{y}}}}\cdot{\boldsymbol{\mathsf{b}}})}{{\boldsymbol{\mathsf{x}}}} } {\N{{\boldsymbol{\mathsf{b}}}}_2^2},\quad \widehat
t^\star = \frac{\N{{\boldsymbol{\mathsf{b}}}}_1 + 2\EUSP{
(\Sgn{{\boldsymbol{\mathsf{y}}}}\cdot{\boldsymbol{\mathsf{b}}})}{\widehat{\boldsymbol{\mathsf{x}}}} } {\N{{\boldsymbol{\mathsf{b}}}}_2^2}.\]
Therefore, using \eqref{eqn:t-t} we have
\begin{align*} t^\star - \widehat{t}^\star =\frac{ \EUSP{{\boldsymbol{\mathsf{v}}}}{{\boldsymbol{\mathsf{x}}}-\widehat{\boldsymbol{\mathsf{x}}}}}{\N{{\boldsymbol{\mathsf{b}}}}_2^2},
\end{align*}
with ${\boldsymbol{\mathsf{v}}}$ defined by $\mathsf{v}_j=2\Sgn(\mathsf{y}_j)\edge_j$ for $j=1,\ldots, h$, and $\mathsf{v}_j=\Sgn(\mathsf{w}_j)$ for $j>h$.
Thus, $\N{{\boldsymbol{\mathsf{v}}}}_2\leq\N{{\boldsymbol{\mathsf{b}}}}_2$ and we have
\begin{align*}
\SN{t^\star - \widehat{t}^\star} &\leq 2 \frac{\N{{\boldsymbol{\mathsf{x}}}-\widehat{\boldsymbol{\mathsf{x}}}}_2} {\N{{\boldsymbol{\mathsf{b}}}}_2} \leq 2\LRP{\frac{\N{{\boldsymbol{\mathsf{y}}}}_2}{\N{{\boldsymbol{\mathsf{b}}}}_2}\N{\Pi-\widehat\Pi}_2
+\sigma\frac{\N{\Pi{\boldsymbol{\mathsf{w}}}}_2} {\N{{\boldsymbol{\mathsf{b}}}}_2}}.
\end{align*}
Using $\N{{\boldsymbol{\mathsf{b}}}}_2^2 = m + 4\N{{\boldsymbol{\mathsf{y}}}}_1 + 4\N{{\boldsymbol{\mathsf{y}}}}_2^2$ and $\N{\Pi{\boldsymbol{\mathsf{w}}}}_2={h}$, and provided $N\gtrsim \LRP{h+u+\sigma^2m }$, by Lemma \ref{lem:proj_error} we have
\[
\SN{t^\star - \widehat{t}^\star} \leq C \frac{\lambda_1}{\lambda_h}
\LRP{\sqrt{\frac{h+\sigma^2 m+u}{N}} +
\frac{h+\sigma^2 m+u}{N}} + \sigma\sqrt{\frac{h}{m}},\]
with probability greater than $1-2\exp(-u)$, and $C>0$ is a constant.
Since $\SN{t^\star-\widehat{t}^\star}\geq \SN{t^\star - \widehat{t}_{\textrm{opt}}} $ the claim follows.
The proof for a bounded ${\boldsymbol{\mathsf{y}}}$ is entirely analogous.
\end{proof}
\begin{remark}
Theorem \ref{prop:error_estimate} is valid not only for $\alpha=1$ but for all $\alpha$.
Namely, in Appendix \ref{app:alphaneq1} we show
\[\SN{t_{\textrm{opt}} - \widehat{t}_{\textrm{opt}}} \lesssim
\frac{\lambda_1}{\lambda_h}
\LRP{\sqrt{\frac{h+u+\sigma^2 m }{N}} +
\frac{h+u+\sigma^2 m}{N}} + \sigma\sqrt{\frac{h}{m}},\]
holds for any $\alpha>0$.
\end{remark}
\section{OptEN algorithm}\label{sec:general_matrices}
Driven by insights in Section \ref{sec:error_analysis}, we are ready to present an efficient heuristic algorithm for learning the {\it Opt}imal regularization parameter for the {\it E}lastic {\it N}et (OptEN).
The algorithm is based on the minimization of a given loss function ($\widehat{R}$, $\widehat{R}_\FM$ or $\widehat{R}_\FP$).
In Section \ref{sec:error_analysis} we showed that in a simplified, yet instructive, setting that the optimal parameter tends to be in the vicinity of $t=1$, depending on the noise level and the signal-to-noise gap.
This is supported by experimental evidence in more general situations such as for non-injective ${\textrm{A}}$, as we will see in Section \ref{sec:experiments}.
Moreover, the loss function is monotonically decreasing as we get away from $t=1$.
These observations drive our algorithm which assumes that the minimizer lies in a valley not too far from $t=1,$ see Figure \ref{fig:loss_functionals}.
Therefore, we will perform a \emph{line search} on the graph of a given loss function, starting from $t=1$.
Line search methods follow iterations $t_{k+1} = t_k + s_k {\boldsymbol{\mathsf{p}}}_k$, where ${\boldsymbol{\mathsf{p}}}_k$ is the \emph{search direction} and $s_k$ the \emph{step size}:
\begin{itemize}
\item {\bf Search direction.}
We select $p_k$ by estimating $\widehat{R}'(t_k)$ with central differences,
\( \widehat{R}'(t) \sim \Delta_\epsilon R(t) := \frac{R(t+\epsilon)-R(t-\epsilon)}{2\epsilon}\)
where $\epsilon>0$.
For $t=1$ we instead use $\tilde\Delta_\epsilon R(1) := \frac{R(1)-R(1-\epsilon)}{\epsilon}$. Then set $p_k = - \Delta_\epsilon R(t).$
\item {\bf Step size.} We estimate $s_k$ with the backtracking line search (consult \cite{A06} for an overview of line search methods).
\end{itemize}
Our approach is presented in Algorithm \ref{alg:algo}, while an extensive numerical study is provided in the next section.
\begin{algorithm}[ht!]
\begin{algorithmic}
\small{
\STATE {\bfseries Input:} ${\boldsymbol{\mathsf{y}}}_1,\ldots,{\boldsymbol{\mathsf{y}}}_N,{\boldsymbol{\mathsf{y}}}\in\bbR^m,\,{\textrm{A}}\in\bbR^{m\times d}$;
\vspace{5pt}
\STATE Compute $\widehat{{\boldsymbol{\mathsf{x}}}}$ according to \eqref{eqn:xhat}.
\STATE Set a loss {function} $\rightarrow\,\,\widehat{R} \textrm{ or } \widehat{R}_\FP \textrm{ or } \widehat{R}_\FM$. In the rest of the algorithm we will refer to it as $R$;
\STATE Set $\epsilon>0$, $\texttt{tol}>0$, $\texttt{tol2}>0$, $0<\alpha<1$, and $c_1,\beta,\gamma>0$;
\STATE Set $k\leftarrow0$, $t_0 = 1$
\STATE Compute $r_1=R(1)$, $\tilde{r}_1=R(1-\epsilon)$, $\mathsf{p}_0=(r_1-\tilde r_1)/\epsilon$, and $r_2 = \varphi(\gamma_0)$;
\REPEAT
\STATE $\tilde{t} = t_k + \alpha \mathsf{p}_k$;
\STATE $\varphi_0=r_1$, $\varphi_0'=-\mathsf{p}_k^2$, $\varphi_1=R(\tilde t)$;
\IF{$\varphi_1-\varphi_0 < c_1 \varphi_0'\alpha$}
\STATE $s_k=\alpha$;
\ELSE
\STATE $s_k = -\frac{1}{2}\frac{\varphi_0'\alpha^2}{\varphi_1- \varphi_0 -\varphi_0'\alpha}$;
\ENDIF
\IF{$|s_k| < \texttt{tol2}$ or $|s_{k-1}/s_k| > \gamma$}
\STATE $s_k = s_{k-1} \cdot \beta$;
\ENDIF
\STATE Set $t_{k+1}=t_k + s_k \mathsf{p}_k$;
\STATE Compute $r_1=R(t_{k+1})$, $\mathsf{p}_{k+1} = (R(t_{k+1}+\epsilon)-R(t_{k+1}-\epsilon)/(2\epsilon)$;
\STATE $k\leftarrow k+1$;
\UNTIL{$|p_k| < \texttt{tol}$ or $k<\texttt{max\_iter}$};
}
\STATE \textbf{Output}: Approximate regularization parameter $\hat{t}:=t_k$.
\end{algorithmic}
\caption{{\bf OptEN algorithm} for approximating the optimal elastic net regularization parameter using backtracking line search}
\label{alg:algo}
\end{algorithm}
\section{Experimental results}\label{sec:experiments}
We now study the performance of our approach and show its adaptivity to different scenarios by conducting experiments on synthetic and imaging data.
In the first set of experiments we perform a thorough comparison of our method with state-of-the-art parameter selection rules by exploring their behavior with respect to noise level and other notions.
The second set of experiments deals with image denoising where we use wavelet-based thresholding with elastic nets.
We consider two data-sets: natural images and a real-world brain MRI data.
Note that we do not aim to compare our method with state-of-the art denoising methods, but rather only with state-of-the-art methods regarding the selection of the regularization parameter for the elastic net.
We start with a discussion of methods that can be used for the automatic detection of the sparsity level $h$ and show that when a sufficient amount of training points is given, we can reliably estimate $h$.
\subsection{Estimating the sparsity level}\label{sec:estimating_sparsity}
In real applications the sparsity level of a vector is either not available or is only an approximate notion, \emph{i.e.} the desired vector is sparse only when we threshold its entries.
Such regimes require $h$ to be estimated, which in our case means looking at the spectrum of the corresponding covariance matrix.
This question belongs to the class of low-rank matrix recovery problems since what we are trying to recover is the geometry (\emph{i.e.} projection onto the range) of the noiseless, lower rank matrix $\Sigma({\textrm{A}}{\boldsymbol{\mathsf{x}}})$, using only the covariance matrix of noisy observations $\widehat\Sigma({\boldsymbol{\mathsf{y}}})$, which is of full rank.
Thus, estimating $h$ boils down to thresholding singular values of the empirical covariance matrix according to some spectral criteria that exploits the underlying structure.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{hSelection.pdf}
\caption{Plot in the upper left corner shows the spectrum of three different types of matrices. The remaining plots (not including the one in the bottom right corner) consider different notions of the spectral gap for $N=100$. The plot in the bottom right considers the last criteria, $1-\frac{\lambda_k}{\lambda_{k+1}}$ but for $N=150$ samples, showing that the behavior for large $k$ changes dramatically, compared to the plot in the bottom right for $N=100$}
\label{fig:spectral_gap}
\end{figure}
For a positive definite matrix with singular values $\lambda_1\geq \ldots\geq \lambda_m\geq 0$, commonly used spectral criteria are (a) the spectral gap $\argmax_k \SN{\lambda_{k} - \lambda_{k+1}}$; (b) the relative gap $\argmax_{k}\LRP{1 - \frac{\lambda_k}{\lambda_{k+1}}}$; (c) the cumulative spectral energy $\sum_{i=1}^k \lambda_i/\sum_{i=1}^m \lambda_i$, and (d) the relative cumulative spectral energy $\sum_{i=1}^k \lambda_i/\sum_{i=1}^{k+1} \lambda_i$.
For the latter two criteria one sets a threshold, say $0.95$, and selects $\tilde h$ as the first $k$ for which the corresponding spectral energy reaches that threshold.
We study the behavior of these four criteria on three different types of forward matrices ${\textrm{A}}$: random Gaussian, random circulant Rademacher, and random Toeplitz Gaussian matrices.
These matrices were chosen because they have different spectral behavior and commonly appear in inverse problems.
In each case ${\textrm{A}}$ is a $100\times 100$ real matrix, normalized so that $\N{{\textrm{A}}}=1$, and we take $N=100$ samples ${\boldsymbol{\mathsf{x}}}_i$, sampled according {to \ref{data:D2}} for $h=20$, ${\boldsymbol{\mathsf{w}}}_i\sim\CN(\boldsymbol{0},{\textrm{I}}_{100})$ and $\sigma=0.3$. We compute $\widehat\Sigma({\boldsymbol{\mathsf{y}}}) = \frac{1}{N} \sum_{i=1}^N {\boldsymbol{\mathsf{y}}}_i\otimes{\boldsymbol{\mathsf{y}}}_i$ for ${\boldsymbol{\mathsf{y}}}_i={\textrm{A}}{\boldsymbol{\mathsf{x}}}_i+\sigma{\boldsymbol{\mathsf{w}}}_i$.
In Figure \ref{fig:spectral_gap}, we show the application of the aforementioned spectral criteria to $\widehat{\Sigma}({\boldsymbol{\mathsf{y}}})$.
It is clear from the results that all four methods would fail if used without taking further information into account.
For example, the spectral gap criterion (in the upper right panel) would dictate {the selection of $h=1$}, but a more careful look at the plot suggests that the behavior of the spectral gap changes dramatically around $h=20$, which corresponds to the true $h$.
Such \emph{ad hoc} solutions are sensible and often improve the performance but can be hard to quantify, especially on real data.
The last spectral criteria, $1-{\lambda_k}/{\lambda_{k+1}}$, is perhaps the most promising, but is also subject to demands on $N$, as shown in the bottom row of Figure \ref{fig:spectral_gap}.
Namely, if $N$ is not large enough then $1-{\lambda_k}/{\lambda_{k+1}}$ has a heavy tail on the spectrum of $\widehat{\Sigma}({\boldsymbol{\mathsf{y}}})$ and would thus suggest a large $h$, as in the bottom left corner of the Figure.
Instead, in Section \ref{sec:synth_experiments} we look for the relative gap within the first $m/2$ singular vectors.
We note that in the case of the first three spectral criteria the situation does not change as the number of samples increases.
On the other hand, for the last criterion it does: heavy tails flatten back to zero for all three choices of random matrices, see the plot in the bottom right corner of Figure \ref{fig:spectral_gap}.
In Section \ref{sec:wavelet_denoising} we will apply our algorithm to wavelet denoising where there is no natural choice of $h$ since wavelet coefficients of images are not truly sparse.
We will instead consider two scenarios; when we are given an oracle $h$ (\emph{i.e.} the $h$ giving the highest PSNR), and when $h$ has to be estimated from data.
\subsection{Synthetic examples}\label{sec:synth_experiments}
\paragraph{Experimental setting.}
We consider the inverse problem of the type ${\boldsymbol{\mathsf{y}}}={\textrm{A}}{\boldsymbol{\mathsf{x}}} + \sigma {\boldsymbol{\mathsf{w}}}$, where the data are generated according to ${\boldsymbol{\mathsf{x}}} = \boldsymbol{\xi} + {{\boldsymbol{\mathsf{v}}}}$ , {where
\begin{enumerate}[label=(\textsf{D}\arabic*)]
\item\label{data:D1} ${\textrm{A}}\in\bbR^{m\times d}$ is a random Gaussian matrix such that $\N{{\textrm{A}}}_2=1$,
\item\label{data:D2} $\xi_i\sim\CN(0, 1)$, and $\mathsf{v}_i= 4\Sgn(\xi_i)$ for $1\leq i \leq h$; $\xi_i=\mathsf{v}_i=0$ otherwise,
\item\label{data:D3} ${\boldsymbol{\mathsf{w}}}\sim\CN(\boldsymbol{0},\Id_m)$.
\end{enumerate}
}
{
The rationale behind distributional choices in \ref{data:D2} is twofold.
First, having ${\boldsymbol{\mathsf{v}}}$ be a non-constant vector ensures that $\CV=\range (\cov{{\boldsymbol{\mathsf{x}}}})$ is truly and fully $h$-dimensional (\emph{i.e.} if ${\boldsymbol{\mathsf{v}}}$ were constant there would be one dominant singular vector).
Second, forcing $\SN{\mathsf{v}_i}= 4$ for $i=1,\ldots, h$ ensures that the data are not concentrated around the origin (which happens \emph{e.g.} if $\mathsf{v}_i=0$ for all $i$ in \ref{data:D2}) and that there is a gap between the original signal and the noise.
The gap can be measured by $\operatorname{SparseSNR}:=\frac{\max_{1\leq i\leq m} \SN{\sigma \mathsf{w}_i}}{\min_{1\leq i\leq h} \SN{\mathsf{x}_i}}$. }
We add that the conclusions and the results of this section, and of Section \ref{sec:estimating_sparsity}, stay the same in case of the more usual distribution assumptions, \emph{i.e.} for ${\boldsymbol{\mathsf{x}}}\sim\CN(\boldsymbol{0}, \Id_h)$ where $\Id_h$ is a diagonal matrix with exactly $h$ entries set to $1$ and the rest to $0$, and noise levels as in Section \ref{sec:wavelet_denoising}.
\paragraph{Comparison.}
We compare our algorithm with the following parameter selection methods:
the discrepancy principle \cite{morozov2012}, monotone error rule \cite{TH99}, quasi optimality \cite{TG65}, L curve method \cite{hansen1992}, (Monte-Carlo) balancing principle \cite{lepskii1991problem} and its elastic net counterpart \cite{DDR09}, (Monte-Carlo) generalized cross-validation \cite{GEE11} and nonlinear cross validation \cite{F05}. In the remainder of this paper we refer to other methods by their acronyms, and to our method as \emph{OptEN}.
The first five methods are commonly used in inverse problems (a detailed account and an experimental study can be found in \cite{zbMATH05929140}),
whereas Monte-Carlo and nonlinear cross-validation are adaptations of generalized cross-validation for non-linear regularization methods.
Before presenting the results, we provide a concise description of considered methods.
Most of the methods require some additional information about the problem, predominantly the noise level $\sigma$, to be either known or estimated, which affects their performance.
We provide the true noise level whenever a given method requires it and furthermore, we perform judicious testing and tuning of all other quantities, taking into account recommendations from relevant literature.
We consider a regularization parameter sequence $t_n = \frac{1}{1+\mu_0 q^n}$,
where $n\in\{0,1,\ldots, N_{\max}\}$, and $\mu_0>0$, $q>0$ and $N_{\max}\in\bbN$ are preselected\footnote{This is an adaptation of the parameter sequence from \cite{zbMATH05929140} that reflects our reparametrization from $\lambda$ to $t$, as in \eqref{eqn:enets_minimization}}.
For each $n$ we denote the corresponding elastic nets solution as ${\boldsymbol{\mathsf{z}}}_n:={\boldsymbol{\mathsf{z}}}^{t_n}$.
\paragraph{Discrepancy Principle [DP]\\}
Discrepancy principle is one of the oldest parameter choice rules which selects a solution so that the norm of the residual is at the noise level.
Thus, the regularization parameter is chosen by the first $n\in\bbN$ such that
\begin{equation}\label{eqn:dp} \N{{\textrm{A}} {\boldsymbol{\mathsf{z}}}_n - {\boldsymbol{\mathsf{y}}}}_2 \leq \tau \sigma \sqrt{m},\end{equation}
where we fix $\tau=1$.
\paragraph{Monotone Error Rule [ME]\\}
This rule is based on the observation that the monotone decrease of the error $\N{{\boldsymbol{\mathsf{z}}}_n - {\boldsymbol{\mathsf{x}}}}_2$ can only be guaranteed for large values of the regularization parameter. Therefore, the best parameter $t_{n^*}$ is chosen as the first $t$-value for which one can ensure that the error is monotonically decreasing. The parameter is then chosen by the smallest $n$ such that
\begin{equation}\label{eqn:mer}
\frac{\EUSP{{\textrm{A}}{\boldsymbol{\mathsf{z}}}_n-{\boldsymbol{\mathsf{y}}}}{{\textrm{A}}^{-\top}\LRP{{\boldsymbol{\mathsf{z}}}_n - {\boldsymbol{\mathsf{z}}}_{n+1}}} }{\| {\textrm{A}}^{-\top}\LRP{{\boldsymbol{\mathsf{z}}}_n - {\boldsymbol{\mathsf{z}}}_{n+1}} \|_2} \leq \tau \sigma \sqrt{m}.
\end{equation}
We fix $\tau=1$ for our experiments.
The left hand side of \eqref{eqn:mer} is replaced with \eqref{eqn:dp} whenever the denominator is $0$.
\paragraph{Quasi-Optimality Criterion [QO]\\}
Quasi-optimality is a parameter rule that does not need the noise level, and thus has enjoyed reasonable success in practice, especially for Tikhonov regularization and truncated singular value decomposition.
The regularization parameter is chosen according to
\begin{equation}\label{eqn:qoc} n_\star = \argmin_{n\leq N_{\max}} \N{{\boldsymbol{\mathsf{z}}}_n-{\boldsymbol{\mathsf{z}}}_{n+1}}_2.\end{equation}
\paragraph{L-curve method [LC] \\}
The criterion is based on the fact that the $\log-\log$ plot of $(\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}_n-{\boldsymbol{\mathsf{y}}}}_2,\N{{\boldsymbol{\mathsf{z}}}_n}_2)$ often has a distinct L-shape.
As the points on the vertical part correspond to under-smoothed solutions, and those on the horizontal part correspond to over-smoothed solutions, the optimal parameter is chosen at the elbow of that \emph{L-curve}. There exist several versions of the method; here we use the following criterion
\begin{equation}\label{eqn:lc}n_\ast = \argmin_{n\leq N_{\max}} \{\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}_n-{\boldsymbol{\mathsf{y}}}}_2\N{{\boldsymbol{\mathsf{z}}}_n}_2\}.\end{equation}
\paragraph{(Monte-Carlo) Balancing Principle [BP]\\}
The principle aims to balance two error contributions, approximation and sampling errors, which have an opposite behavior with respect to the tuning parameter. More precisely, we select the parameter by
\[
n^* = \argmin_n \{ t_n | \N{{\boldsymbol{\mathsf{z}}}_n-{\boldsymbol{\mathsf{z}}}_k}_2 \leq 4 \kappa \sigma \rho(k), k = n,\ldots, N_{\max} \},
\]
where $\kappa>0$ is a tuning parameter.
More computationally friendly, yet equally accurate, versions of the balancing principle are also available \cite{zbMATH05929140}. As our main focus on the accuracy of the parameter choice, we will use the original and more computationally heavy version of the balancing principle.
The value of $\sigma \rho(k)$ is in general unknown but it can be estimated in case of white noise. Following \cite{zbMATH05929140}, we calculate
\( \rho(k)^2\approx \operatorname{mean} \{ \N{{\textrm{A}}_n^{-1}\boldsymbol{\xi}_i}_2^2\},\)
where $\boldsymbol{\xi}_i\sim\CN(\boldsymbol{0},\Id_m)$, $1\leq i \leq L$ (we use $L=4$), and ${\textrm{A}}_n^{-1}$ is the map that assigns ${\boldsymbol{\mathsf{y}}}$ to ${\boldsymbol{\mathsf{z}}}_n.$
\paragraph{Elastic Nets Balancing Principle [ENBP]\\}
In \cite{DDR09} the authors propose a reformulation of the balancing principle for elastic net,
\[
n^* = \argmin_n \{ t_n | \N{{\boldsymbol{\mathsf{z}}}_k-{\boldsymbol{\mathsf{z}}}_{k+1}}_2 \leq \frac{4 C}{\sqrt{d}\alpha \mu_0 q^{k+1} }, k =N_{\max} -1 ,\ldots,n \},
\]
The method stops the first time two solutions are sufficiently far apart. The constant $C$ needs to be selected, and in our experience this task requires a delicate touch.
\paragraph{(Monte-Carlo) Generalized Cross-Validation [GCV]\\}
The rule stems from the ordinary cross-validation, which considers all the \emph{leave-one-out} regularized solutions and chooses the parameter that minimizes the average of the squared prediction errors.
Specifically, GCV selects $n$ according to
\begin{equation}\label{eqn:gcv} n_\ast = \argmin_{n\leq N_{\max}} \frac{m^{-1}\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}_n - {\boldsymbol{\mathsf{y}}}}_2^2}{\LRP{m^{-1} \tr(\Id_m - {\textrm{A}}\Amat_n^{-1})}^2},\end{equation}
where ${\textrm{A}}_n^{-1}$ is the map such that ${\boldsymbol{\mathsf{z}}}_n = {\textrm{A}}_n^{-1} ({\boldsymbol{\mathsf{y}}})$.
In the case of elastic nets the map ${\textrm{A}}_n^{-1}$ is not linear and, thus, we cannot assign a meaning to its trace.
Instead, we follow the ideas of \cite{GEE11} and estimate the trace stochastically using only one data sample.
\paragraph{Nonlinear Generalized Cross-Validation [NGCV]\\}
In \cite{F05} the authors reconfigure GCV for non-linear shrinkage methods, and $n$ is selected according to
\[ n_\ast = \argmin_{n\leq N_{\max}} \frac{\N{{\textrm{A}}{\boldsymbol{\mathsf{z}}}_n - {\boldsymbol{\mathsf{y}}}}_2^2}{m^{-1}\LRP{1- ds/m}^2},\]
where $s = \frac{\N{{\boldsymbol{\mathsf{z}}}_n}_\gamma}{\N{{\boldsymbol{\mathsf{z}}}^\dagger}_\gamma}$ with $\N{\cdot}_\gamma := \N{\cdot}_1+\alpha\N{\cdot}^2_2.$
\begin{table}[th]
\centering
\ra{1.1}
\begin{tabular}{cccccc}Method & $\frac{|t_{opt} - \widehat{t}|}{t_{opt}}$& $\frac{\|{\boldsymbol{\mathsf{x}}} - {\boldsymbol{\mathsf{z}}}^{\widehat{t}}\|}{\|{\boldsymbol{\mathsf{x}}}\|}$& $\text{FDP}(\widehat{t})$& $\text{TPP}(\widehat{t})$& $\genfrac{}{}{0pt}{}{\textrm{computational}}{\textrm{time [s]}}$\\
\hline\hline
$x_{t_{opt}}$ & 0 & 0.0984 & 0.1583 & 1.000 & 0\\
\hline
$\widehat{\boldsymbol{\mathsf{x}}}$ & N/A & 0.1004 & \textbf{0.000} & \textbf{1.000} & N/A\\
OptEN & \textbf{0.0254} & \textbf{0.0994} & 0.160 & \textbf{1.000} & 3.716\\
DP & 0.1196 & 0.1118 & 0.087 & \textbf{1.000} & 0.7421\\
ME & 0.2493 & 0.1376 & 0.010 & \textbf{1.000} & \textbf{0.1757}\\
QO & 0.4543 & 0.1843 & 0.716 & \textbf{1.000} & 7.193\\
LC & 0.1414 & 0.1174 & 0.188 & \textbf{1.000} & 1.564\\
BP & 0.0877 & 0.1057 & 0.095 & \textbf{1.000} & 36.77\\
ENBP & 0.2728 & 0.1450 & \textbf{0.007} & \textbf{1.000} & 4.355\\
GCV & 0.4548 & 0.1844 & 0.716 & \textbf{1.000} & 15.91\\
NGCV & 0.3597 & 0.1577 & 0.638 & \textbf{1.000} & 7.306
\end{tabular}
\caption{Comparison of errors for regularization parameter selection methods, with an injective matrix ${\textrm{A}}\in\bbR^{500\times 100}$ and $h=10$, $\alpha=10^{-3}$, $\sigma=0.3$. The values are averages over $100$ independent runs}
\label{tab:rando_label_3119}
\end{table}
\paragraph{Comparison/Error.}
For each method we compute: the (normalized) error in approximating the optimal regularization parameter, the (normalized) error, false discovery proportion (FDP), true positive proportion (TPP), and the computational time.
TPP and FDP are measures that quantify the notions of true and false discovery of relevant features in sparsity based regression tasks \cite{SBC16}.
FDP is the ratio between false discoveries and the total number of discoveries,
\[ \operatorname{FDP}(t) = \frac{\#\left[{j:\, \mathsf{z}^t_j \neq 0 \textrm{ and } \mathsf{x}_j = 0}\right]}{\max\LRP{\#\left[{j:\, \mathsf{z}^t_j \neq 0}\right], 1}}.\]
TPP on the other hand is the ratio between true (\emph{i.e.} correct) discoveries in the reconstruction and true discoveries in the original signal,
\[ \operatorname{TPP}(t) =\frac{\#\left[{j\in\{1,\ldots, h\}:\, \mathsf{z}^t_j \neq 0 \textrm{ and } \mathsf{x}_j \neq 0}\right]}{h} .\]
Thus, to recover the structure of the original sparse data we want FDP close to $0$ and TPP close to $1$.
It is known that there is often an explicit (and sometimes even quantifiable) trade-off between FDP and TPP, in the sense that the support overestimation is an (undesirable) side-effect of full support recovery.
In other words, a consequence of true support discovery is often a non-trivial false support discovery \cite{SBC16}.
When computing FDP and TPP we will rather than demand for an entry to be exactly zero, instead threshold the values (with $0.5$ being the threshold).
\paragraph{Testing setup.}
To compute the true optimal parameter $t_{opt},$ we run a dense grid search on $[0,1]$ using the true expected loss $\N{{\boldsymbol{\mathsf{z}}}^t-{\boldsymbol{\mathsf{x}}}}_{2}^2$.
As suggested in \cite{zbMATH05929140} we use $\tau=1$ and provide the true noise level $\sigma$ for discrepancy principle and monotone error rule; balancing principle uses $\kappa = 1/4$ and true $\sigma$; elastic net balancing principle uses $C=1/2500$.
The parameter grid for DP, ME, BP, QO, LC, BP, GCV, and NGCV is defined by $\mu_0=1$, $q=0.95$, and $N_{\max}= 100$ (thus, $t_0=0.5$ and $t_{N_{\max}}= 0.9941143171)$, whereas for
ENBP, we use $t_0=0.05$, $q=1.05$, and $N_{\max}= 100$.
The tests are conducted for $m\in\{500, 900\},\,d\in\{100,200\}$ and $h\in\{10,20,30\}$, where all combinations of $\alpha\in\{10^{-5}, 10^{-3},10^{-2},10^{-1}\}$ and $\sigma\in\{0.05, 0.1, 0.2, 0.3\}$ are considered.
To compute the {empirical estimator} $\widehat{{\boldsymbol{\mathsf{x}}}}$, we generate $N = 50$ independent random samples of the training data (${\boldsymbol{\mathsf{x}}}, {\boldsymbol{\mathsf{y}}}$).
Results in Table \ref{tab:rando_label_3119} are averaged over $100$ independent runs for $\alpha=10^{-3}$, $m=500$, $d=100$, $h=10$, where $\CV=\Span{{\boldsymbol{\mathsf{e}}}_1,\ldots,{\boldsymbol{\mathsf{e}}}_h}$, and $\sigma =0.3$, which corresponds to $\operatorname{SparseSNR}\approx 0.17$.
The first row in the table, $x_{t_{opt}}$, describes the elastic net minimizer for which the \emph{true} optimal regularization parameter is provided.
\paragraph{Discussion.}
OptEN always returns the value which is the closest to the optimal regularization parameter, and its results are in general comparable to the ones provided by the minimizer with the optimal parameter.
However, one can observe that other methods, \emph{e.g.}, discrepancy principle, provide a better balance between FDP and TPP (returning solutions that are more sparse), though at a cost of a larger approximation error.
Balancing principle also provides very good results, but it is slow unless an effort is made to improve its computational time.
Moreover, we observed that the performance of all methods that require the noise level $\sigma$ to be known deteriorates if we do not provide the exact value of the noise level, but only its rough estimate.
The overall results are mostly consistent over all experimental scenarios we looked at, with a couple of exceptions.
As expected, FDP and estimation errors deteriorate not only for larger $\sigma$ but also for larger $\alpha$, though the ranking of the methods and the patterns of behavior remain the same.
This is due to the fact that as $\alpha$ increases elastic nets sacrifice sparsity for smoothness.
The {empirical estimator} $\widehat{\boldsymbol{\mathsf{x}}}$ is a very accurate estimator of the original signal, and it sometimes outperforms even the elastic net solution that uses the optimal parameter. However, as has been observed in \cite{DFN16} for Tikhonov regularization and confirmed in Figure \ref{fig:noise_levels}, the performance of the {empirical estimator} worsens in the small noise regime.
\paragraph{Comparison with {empirical estimator}: effects of $\sigma$ and $N_{\text{train}}$\\}
We study the behavior of the relative estimation error with respect to $\sigma$ and the number of training samples.
We compare OptEN with the {empirical estimator} $\widehat{\boldsymbol{\mathsf{x}}}$, DP, NGCV, and BP.
We use $m=500,\,d=100, \,h=10$ and \ref{data:D1}-\ref{data:D3} with $\sigma$ ranging from {$0.1$ to $0.5$} in the first experiment, whereas we vary the number of training samples $N$ from $20$ to $60$ in the second experiments as depicted in Figure \ref{fig:noise_levels}.
Our method again outperforms other considered parameter selection rules.
On the other hand, the {empirical estimator} performs slightly better than OptEN for larger noise levels (it is also better than the elastic nets solution with the optimal parameter), and it performs worse for lower noise levels.
This is essentially due to the fact that $\widehat{{\boldsymbol{\mathsf{x}}}}$ is never truly sparse, but has a lower (thresholded) FDP.
Namely, as a projection onto an $h$-dimensional space the non-zero entries of $\widehat{{\boldsymbol{\mathsf{x}}}}$ are very small, whereas the non-zero entries obtained with elastic net are larger and their size depending on the noise level.
The parameter $\alpha$ plays a similar role; for small $\alpha$ OptEN beats $\widehat{\boldsymbol{\mathsf{x}}}$, and for larger $\alpha$ the situation is reversed.
\begin{figure}[ht!]
\centering
\begin{minipage}[b]{0.49\linewidth}
\centering
\includegraphics[width=\textwidth]{x_error_wrt_sigma.pdf}
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\includegraphics[width=\textwidth]{x_error_wrt_N.pdf}
\end{minipage}
\caption{In the left panel is the behavior of the {empirical estimator}, OptEN, discrepancy principle, balancing principle, and nonlinear GCV with respect to $\sigma$ is shown.
In the right panel the behavior of our method for different values of $\sigma$ as the number of samples $N$ increases is shown.
Dashed lines represent the error achieved by taking the true optimal parameter for the corresponding $\sigma$}
\label{fig:noise_levels}
\end{figure}
\paragraph{Non-injective matrices\\}
We now conduct experiments with non-injective matrices.
The setting is as in Table \ref{tab:rando_label_3119}, where now ${\textrm{A}}\in\bbR^{500\times 100}$ with $\rank\LRP{{\textrm{A}}}=40$.
As mentioned in Section \ref{subsec:qlf}, we test our method by minimizing the projected loss functional $\widehat{R}_\FP$ and the modified error functional $\widehat{R}_\FM$.
The results can be found in Table \ref{tab:rando_label_7614}.
Our method (using both the projected and modified functionals) again outperforms standard parameter selection rules in terms of the precision accuracy, and loses out to some methods when it comes to FDP and TPP.
We also observe that the performance of the {empirical estimator} deteriorates and $\widehat{\boldsymbol{\mathsf{x}}}$ indeed should not be used as the solution itself but some additional regularization is required.
\begin{table}[th]
\centering
\ra{1.1}
\begin{tabular}{cccccc}Method & $\frac{|t_{opt} - \widehat{t}|}{t_{opt}}$& $\frac{\|{\boldsymbol{\mathsf{x}}} - {\boldsymbol{\mathsf{z}}}^{\widehat{t}}\|}{\|{\boldsymbol{\mathsf{x}}}\|}$& $FDP(\widehat{t})$& $TPP(\widehat{t})$& $\genfrac{}{}{0pt}{}{\textrm{computational}}{\textrm{time [s]}}$\\
\hline\hline
using $t_{opt}$ & 0 & 0.5704 & 0.4918 & 0.9450 & 0\\
\hline
{empirical estimator} $\widehat{\boldsymbol{\mathsf{x}}}$ & N/A & 0.7978 & 0.8047 & \textbf{1.000} & N/A\\
projected OptEN & \textbf{0.0718} & \textbf{0.6033} & 0.507 & 0.930 & 8.16\\
modified OptEN& 0.0763 & 0.6046 & 0.497 & 0.926 & 15.57\\
DP & 0.1316 & 0.6343 & 0.528 & 0.926 & 1.93\\
ME & 0.3203 & 0.7234 & \textbf{0.290} & 0.765 & \textbf{0.15}\\
QO & 0.3167 & 0.7857 & 0.819 & 0.997 & 7.04\\
LC & 0.3389 & 0.7426 & 0.304 & 0.749 & 7.00\\
GCV & 0.3172 & 0.7865 & 0.819 & 0.997 & 14.04\\
NGCV & 0.3172 & 0.7865 & 0.819 & 0.997 & 7.06\\
BP & 0.2636 & 0.7179 & 0.757 & 0.974 & 35.01\\
ENBP & 0.2133 & 0.6549 & 0.362 & 0.847 & 3.69
\end{tabular}
\caption{Comparison of errors for regularization parameter selection methods, with a matrix ${\textrm{A}}\in\bbR^{500\times 100}$, $\rank({\textrm{A}})=40$, and $h=10$, $\alpha=10^{-3}$, $\sigma=0.3$. The values are averages over $100$ independent runs}
\label{tab:rando_label_7614}
\end{table}
\subsection{Image denoising}\label{sec:wavelet_denoising}
The task of image denoising is to find an estimate $ \boldsymbol Z$ of an unknown image $\boldsymbol{X}$ from a noisy measurement $\boldsymbol{Y}$, where
$\boldsymbol{Y} = \boldsymbol X + \sigma\boldsymbol{\Xi}$, and $\boldsymbol{\Xi}$ denotes isotropic white noise. The goal is to improve the image quality by removing noise while preserving important image features such as edges and homogeneous regions.
There are a large number of methods addressing image denoisig , starting from 'classical' wavelet thresholding \cite{DJ94, D95} and non-linear filters, to stochastic and variational methods \cite{CL97,MR1669536}.
Since the primary goal of this paper is to evaluate how does the proposed approach perform as a parameter selection method for the elastic net, here we only compare our method with other -art parameter selection methods for elastic nets, and do not compare elastic nets with image denoising methods in general.
In particular, we compare OptEN with the discrepancy principle and the balancing principle (\emph{i.e.} the top performers from previous experiments).
In all cases the results show that OptEN has superior performance and selects nearly optimal parameters, see Table \ref{tab:iqas}.
\paragraph{Wavelet-based denoising.}
We denoise the noisy image $\boldsymbol{Y}$ by minimizing
\begin{equation}\label{eq:wavelet} (1-t)\N{\boldsymbol{Z} - \boldsymbol{Y}}_2^2 + t( \N{\bbW\boldsymbol{Z}}_1 + \alpha\N{\boldsymbol{Z}}^2_2), \textrm{ for } \boldsymbol{Z}\in[0,1]^p,\end{equation}
where $\bbW$ is the wavelet transform using the family of \texttt{db4} wavelets, and $\alpha = 10^{-3}$. %
Wavelet transform sparsify natural images, and we thus select the {empirical estimator} in the wavelet domain.
Moreover, in the limit with respect to the number of samples $N\rightarrow\infty$, the empirical projection $\widehat \Pi Y$ is for a given $h$ equivalent to a hard thresholding of $\bbW\boldsymbol{Y}$ that preserves its $h$ largest wavelet coefficients.
Thus, for image denoising we do not use samples $\boldsymbol{Y}_i$ but instead only threshold $\bbW\boldsymbol{Y}$ for a well chosen $h$.
Here $h$ cannot be chosen by searching for a gap in $\bbW\boldsymbol{Y}$, since it most often does not exist.
Instead, we say that the \emph{true} $h$ is the one that minimizes the MSE of the reconstructed image.
In our first set of experiments the {empirical estimator} $\widehat{\boldsymbol{X}}$ is chosen by hard thresholding $\boldsymbol{Y}$, where $h$ is optimal.
\subsubsection{Denoising with an oracle \it{\normalfont{h}}}\label{sec:OracleH}
\paragraph{Data and learning setup.}
We consider five grayscale images: \texttt{space shuttle}, \texttt{cherries}, \texttt{cat}, \texttt{mud flow}, and \texttt{IHC}, each of size $512\times512$ pixels.
For BP and DP the regularization parameter is selected from a sequence of parameter values $t_n = \frac{1}{1+\mu_0 q^n}$ with $\mu_0=1$, $q=0.95$, and $N_{\max} = 100,$ same as before.
Moreover, we fix $\tau=1$, $\kappa=1/4$ for BP and DP, and provide them with the true noise level.
For OptEN
the {empirical estimator} is computed with an oracle $h$, \emph{i.e.} the one returning the lowest MSE.
\paragraph{Comparison/Error.}
We use two performance metrics: peak signal-to-noise ratio (PSNR) and the similarity index (SSIM) between the original image $\boldsymbol{X}$ and the recovered version $\boldsymbol{Z}$.
PSNR is a standard pixel-based performance metric, defined through the MSE by
\[\quad \text{PSNR}(\boldsymbol{X},\boldsymbol{Z}) = 10 \log_{10}\LRP{\frac{\max_{1\leq i\leq p} X_i - \min_{1\leq i\leq p} X_i}{\text{MSE}(\boldsymbol{X},\boldsymbol{Z})}}, \text{MSE}(\boldsymbol{X},{\boldsymbol{Z}}) = \frac{1}{p} \N{\boldsymbol{X} - \boldsymbol{Z}}^2.\]
MSE and PSNR are ubiquitous in image and signal analysis due to their simplicity and suitability for optimization tasks, but are also infamous for their inability to capture features salient for human perception of image quality and fidelity \cite{WB+04}.
SSIM on the other hand, is a structure-based performance metric that tries to address this issue by using easy-to-compute structural statistics to estimate image similarity.
It is defined through
\[\text{SSIM}(\boldsymbol{X},\boldsymbol{Z})=\LRP{\frac{2{\overline{\boldsymbol{X}}}\,{\overline{\boldsymbol{Z}}} + C_1}{{\overline{\boldsymbol{X}}}^2{\overline{\boldsymbol{Z}}}^2 + C_1}}\LRP{\frac{2 std(\boldsymbol{X}) std(\boldsymbol{Z}) + C_2}{std(\boldsymbol{X})^2std(\boldsymbol{Z})^2 + C_2}},\]
where $\overline{\boldsymbol{X}},\,\overline{\boldsymbol{Z}}$ are the means, and $std(\boldsymbol{X}),\,std(\boldsymbol{Z})$ are the standard deviations of pixels of corresponding images $\boldsymbol{X}$ and $\boldsymbol{Z}$, and $C_1,C_2$ are positive constants\footnote{We take $C_1 = 0.01,\, C_2 = 0.03$ by the convention of \texttt{python}'s \texttt{skimage} package}.
\begin{table*}\centering
\ra{1.1}
{\footnotesize
\begin{tabular}{@{}cccccccccccc@{}}
& \multicolumn{5}{c}{{\normalsize PSNR}} && \multicolumn{5}{c}{{\normalsize SSIM}} \\
\cmidrule{2-6} \cmidrule{8-12} & $\genfrac{}{}{0pt}{}{\textrm{using}}{t_{\textrm{opt}}}$ & noisy
& OptEN & DP & BP && $\genfrac{}{}{0pt}{}{\textrm{using}}{t_{\textrm{opt}}}$ & noisy& OptEN & DP & BP \\
\midrule
$\genfrac{}{}{0pt}{}{\textrm{\textsf{space}}}{\textrm{\textsf{shuttle}}}$\\
$\sigma=0.05$ & 32.33& 26.27& \textbf{32.32}& 30.26& 30.63 && 0.929& 0.687& \textbf{0.929}& 0.850& 0.862\\
$\sigma=0.075$ & 30.26& 22.87& \textbf{30.25}& 27.32& 27.69 && 0.908& 0.516& \textbf{0.908}& 0.748& 0.765\\
$\sigma=0.1$ & 28.77& 20.50& \textbf{28.76}& 25.32& 25.60 && 0.892& 0.394& \textbf{0.891}& 0.660& 0.676\\
\midrule
\textsf{cherries}\\
$\sigma=0.05$ & 35.80& 26.08& \textbf{35.79}& 30.52& 31.08 && 0.973& 0.647& \textbf{0.972}& 0.837& 0.855\\
$\sigma=0.075$ & 33.59& 22.63& \textbf{33.59}& 27.40& 27.77 && 0.964& 0.463& \textbf{0.964}& 0.720& 0.737\\
$\sigma=0.1$ & 31.94& 20.24& \textbf{31.82}& 25.26& 25.54 && 0.958& 0.339& \textbf{0.942}& 0.617& 0.633\\
\midrule
\textsf{cat}\\
$\sigma=0.05$ & 29.34& 26.02& \textbf{29.33}& 28.72& 28.93 && 0.890& 0.767& \textbf{0.890}& 0.861& 0.868\\
$\sigma=0.075$ & 27.06& 22.51& \textbf{27.03}& 25.85& 26.04 && 0.825& 0.612& \textbf{0.823}& 0.758& 0.767\\
$\sigma=0.1$ & 25.69& 20.04& \textbf{25.24}& 23.88& 24.03 && 0.771& 0.486& \textbf{0.741}& 0.664& 0.671\\
\midrule
\textsf{mud flow}\\
$\sigma=0.05$ & 28.38& 26.02& \textbf{28.37}& 28.20& 28.30 && 0.870& 0.777& \textbf{0.869}& 0.856& 0.860\\
$\sigma=0.075$ & 26.07& 22.50& \textbf{26.06}& 25.46& 25.60 && 0.795& 0.629& \textbf{0.795}& 0.755& 0.762\\
$\sigma=0.1$ & 24.71& 20.01& \textbf{24.46}& 23.58& 23.70 && 0.735& 0.506& \textbf{0.717}& 0.662& 0.668\\
\midrule
\textsf{IHC}\\
$\sigma=0.05$ & 29.33& 26.02& \textbf{29.33}& 28.73& 28.92 && 0.890& 0.767& \textbf{0.889}& 0.861& 0.868\\
$\sigma=0.075$ & 27.06& 22.51& \textbf{27.04}& 25.86& 26.04 && 0.825& 0.612& \textbf{0.823}& 0.759& 0.767\\
$\sigma=0.1$ & 25.70& 20.04& \textbf{25.39}& 23.85& 24.04 && 0.772& 0.486& \textbf{0.748}& 0.662& 0.671\\
\bottomrule
\end{tabular}}
\caption{Results on wavelet denoising of noisy images with different noise levels using elastic nets minimization. Each column defines the method used to select the regularization parameter. In bold is the method that achieved the best result. Columns titled \emph{noisy} correspond to PSNR and SSIM values of the initial noisy image. Columns titled \emph{using $\lambda_{\textrm{opt}}$} correspond to the best values achievable for the selected elastic nets functional, where we find the optimal parameter by a grid search on the true loss functional}
\label{tab:iqas}
\end{table*}
\begin{figure}
\includegraphics[width=\textwidth]{7825_nongcv_with_bp.pdf}
\caption{Comparing the effects of denoising for different parameter selection rules}\label{fig:visual}
\end{figure}
Table \ref{tab:iqas} provides the PSNR and SSIM values generated by all algorithms on the considered images, while Figure \ref{fig:visual} shows the result of denoising on a $128\times128$ detail of each image for $\sigma=0.075$.
We can see that our method achieves the highest PSNR on all images and that this effect is more pronounced for larger noise values.
\subsubsection{Denoising with a heuristically chosen $h$}
{\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{h_criteria.pdf}
\caption{A visual representation of our heuristic criteria for selecting $h$. Here we choose $h=4400$. The flat line at the end corresponds to $\hat{t}_k=1$.}
\label{fig:h_criteria}
\end{figure}
In this set of experiments we study the performance of our method in a situation where the optimal $h$ is not known \emph{a priori}.
We run experiments on a real-world dataset of brain images\footnote{Obtained from \url{http://nist.mni.mcgill.ca/?page_id=672}, therein referred to as \emph{group 2}},
which consists of \emph{in vivo} MRIs of 13 patients with brain tumor, taken pre-surgery.
For each patient we took an MRI slice, isolated the area around the brain and then added additional isotropic white noise with $\sigma\in\{0.05, 0.075, 0.1\}$.
{We then select $h$ by a heuristicaly driven procedure.
Namely, for each image $\boldsymbol{Y}$ we set an initial $h_0\in\bbN$ and determine $\hat{t}_0$ by performing Algorithm \ref{alg:algo}, where $\widehat{\boldsymbol{X}}_{h_0}$ is constructed by taking $h_0$ largest coefficients of $\bbW \boldsymbol{Y}$.
We then set $h_1= h_0+h_\text{step}$, repeat the procedure, and continue iteratively for $h_k$.
The iterations are stoppped once the corresponding $\hat{t}_k$ start to decrease or become discontinuous (since heuristically this corresponds to a decrease in the PSNR of the corresponding elastic-net regularized solution)}.
$h_0$ and $h_\text{step}$ are chosen according to the size of the image.
The behaviour of this criteria can be seen in Figure \ref{fig:h_criteria}, and it shows that if $h$ is too large the empirical estimator is virtually the same as $\boldsymbol{Y}$. In other words, the minimizer of $\N{\boldsymbol{Z}^t-\widehat{\boldsymbol{X}}_h}$ is $t=1$ (\emph{i.e.} $\lambda=0$), which we observe in Figure \ref{fig:h_criteria}.}
The resulting reconstruction for $\sigma=0.1$ can be seen in Figure \ref{fig:brains} on four images.
The effects of denoising are visually not as striking as the results in Section \ref{sec:OracleH}.
We attribute this to the fact that PSNR gains with the best possible choice of parameter using elastic net are quite small, namely, PSNR of the noisy image improves only by around $5-7\%$, when taking the optimal parameter (see Table \ref{tab:brainpsnr}).
Other parameter selection rules (discrepancy principle and balancing principle) did not improve the PSNR and are thus not presented.
\begin{table}[ht!]
\centering
\ra{1.1}
\begin{tabular}{ccc}
using $t_{\textrm{opt}}$ & noisy & OptEN
\\
\hline\hline
28.608 & 27.297 & 28.432 \\
28.329 & 26.792 & 28.079 \\
28.221 & 26.735 & 27.935 \\
28.501 & 26.906 & 28.240
\end{tabular}
\caption{PSNR values for the results in Figure \ref{fig:brains}}
\label{tab:brainpsnr}
\end{table}
\section{Conclusion and future work}\label{sec:conclusion}
{In this paper, we presented an approach for the estimation of the optimal regularization parameter for elastic net.
The theoretical guarantees are possible only in simplified scenarios but we used insights gained therein to steer and create an efficient algorithm.
The algorithm exhibits excellent prediction accuracy, including in cases when there are no theoretical guarantees.
Comparison with state-of-the-art methods show a clear superiority of our method, under the studied testing scenarios.
Moreover, whereas other studied methods require adjsting a number of additional parameters in order to achieve satisfactory results, our method is entirely autonomous given a sufficient number of training samples.
\begin{figure*}[th!]
\centering
{%
\includegraphics[clip,width=0.9\textwidth]{brains_0075_1}%
}
{%
\includegraphics[clip,width=0.9\textwidth]{brains_0075_2}%
}
{%
\includegraphics[clip,width=0.9\textwidth]{brains_0075_3}%
}
{%
\includegraphics[clip,width=0.9\textwidth]{brains_0075_4}%
}
\caption{Denoising results on a brain image data set. The PSNR values are in Table \ref{tab:brainpsnr}}\label{fig:brains}
\end{figure*}
We aim to use the ideas presented in this paper in further studies.
Namely, we will study the behavior of the solution with respect to the other hyperparameter; $\alpha$, and consider other optimization schemes, predominantly focusing on imaging applications.
We will also work on developing an optimization scheme for a joint minimization of both the regularisation functional and the loss functional for the regularization parameter.
}
\section*{Acknowledgements}
V. Naumova and Z. Kereta acknowledge the support from RCN-funded FunDaHD project No 251149/O70.
E. De Vito is a member of the Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM).
|
\section{Introduction}
Automatic precision assembly is important for industrial manipulators to improve the efficiency and reduce the cost. Most of the current assembly tasks rely on dedicated manual tuning to provide trajectories for specific tasks, which requires intensive labors and is not robust to uncertainties. To reduce the human involvement and increase the robustness to uncertainties, more researches are focusing on learning the assembly skills.
There are three types of learning in Psychology~\cite{kalat2016introduction}: classical conditioning, observational learning and operant conditioning. The second and third types correspond to supervised learning and reinforcement learning, respectively.
The supervised learning is ideal when the training data is sufficient. Practically, collecting data is inefficient under various uncertainties of the environment. A Gaussian mixture model (GMM) is trained in~\cite{tang2016teach} from human demonstration to learn a peg hole insertion skill. The peg hole insertion task is simplified by constraining the policy into planar motion and the trained policy is not adaptable to different environments.
The reinforcement learning (RL) learns a sequence of optimal actions by exploring the environment to maximize the expected reward. Different types of RL methods include the direct policy gradient such as REINFORCE~\cite{williams1992simple}, Q-learning based methods such as DQN~\cite{mnih2015human}, as well as the actor-critic framework such as DDPG~\cite{lillicrap2015continuous} or PPO~\cite{schulman2017proximal}.
These methods are called model-free RL since the dynamics model is not used during exploration. Despite lack of dynamics, the model-free RL has been successfully applied to assembly tasks~\cite{vecerik2017leveraging, inoue2017deep}. The model-free RL requires considerable data to explore the state/action space and reconstruct the transitions of the environment. Consequently, it is less data-efficient and time-efficient.
Model-based RL is proposed to increase the data efficiency~\cite{levine2013guided, levine2016end}. It fits dynamics models and applies optimal control such as iLQR/iLQG~\cite{tassa2012synthesis} to compute the optimal trajectories. The exploration is conducted by adding random noise to the actions during the optimization. Then the optimized trajectories are used to train a neural network policy in a supervised manner. Compared with model-free RL, the model-based RL has larger exploit-exploration ratio, thus explores narrower space and converges faster than the model-free RL. The performance of the model-based RL depends on the behavior of the optimal controller (i.e. supervisor), which in turn is effected by the accuracy of the local dynamics model. For the rigid robot dynamics with force/torque as states, the dynamics model is less smooth\footnote{The dynamics change dramatically as the trajectory slightly changes.}, which makes the dynamics fitting not effective. Consequently, the model-based RL cannot converge consistently. In practice, people usually use soft robotics model (Baxter, PR2)~\cite{levine2016end} with position/velocity states by ignoring the force/torque feedback.
This paper proposes a learning framework to train a more natural assembly policy by incorporating both the force/torque and the positional feedback signals. The proposed framework combines the model-based RL with the model-free actor-critic to learn the manipulation skills for precision assembly tasks.
The model-based RL computes for the optimal trajectories with both positional and force/torque feedback. The performance of the controller might be affected by the smoothness of the local fitted dynamics model. To avoid the problem of inconsistency or tedious parameter tuning of optimal controller, a critic network is introduced to learn the correct critic value (Q-value). Instead of training the policy network by pure supervision, we train an actor network by combining the supervised learning with the policy gradient. To accelerate the training efficiency of the critic network, the Q-value from the optimal control is employed to train the critic network.
The contribution of this work are as follows. First, the optimal controller is able to constrain the exploration space in safe region compared with the random exploration at the first iterations of actor-critic methods.
Secondly, the optimal controller is more data-efficient when exploring in a narrower space and solving for optimal trajectory mathematically.
Thirdly, the combined critic network is able to address the potential inconsistency and instability of the optimal controller caused by the rigid robotics system and force/torque feedback, and build up a ground truth critic for the policy network.
The remainder of this paper is described as follows. The related work is stated in Section~\ref{sec:related}, followed by a detailed explanation of the proposed learning framework in Section~\ref{sec:proposed}. Simulation and experiment results are presented in Section~\ref{sec:experiments}. Section~\ref{sec:conclusions} concludes the paper and proposes future works.
\section{Related Work}
\label{sec:related}
The objective of an assembly task is to learn an optimal policy $\pi_\theta(a_t|o_t)$ to choose an action $a_t$ based on the current observation $o_t$ in order to maximize an expected reward:
\begin{equation}
\label{eq:loss_minimization}
\begin{aligned}
\min_{\pi_\theta} E_{\tau \sim \pi_\theta}(l(\tau)),
\end{aligned}
\end{equation}
where $\theta$ is the parameterization of the policy, $\tau = \{s_0,a_0, s_1, a_1, ..., s_T, a_T\}$ is the trajectory, $\pi_\theta(\tau) = p(s_0)\prod_{1}^{T}p(s_t|s_{t-1},a_{t-1})\pi_\theta(a_t|s_t)$, and $l$ is the loss of the trajectory $\tau$.
Equation~(\ref{eq:loss_minimization}) can be solved by optimization once a global dynamics model $p(x_t|x_{t-1}, u_{t-1})$ is explicitly modeled. For a contact-rich complex manipulation task, the global dynamics model is extremely difficult to obtain. Therefore, the assembly task either avoids using dynamics~\cite{inoue2017deep} or fits the a linear dynamics model~\cite{tang2016teach,levine2013guided, levine2016end}.
On one hand, the RL without dynamics requires excessively data to explore the space and locate to the optimal policy due to the potential high-dimensionality of the action space. On the other hand, the performance of the~\cite{levine2013guided, levine2016end} can be downgraded once the robotic system is rigid or the force/torque feedback is included in the optimal controller.
We propose a learning framework that combines the actor-critic framework and optimal control for efficient high-accuracy assembly. The optimal controller is adapted from the model-based RL~\cite{levine2013guided}, while the actor-critic framework is modified from the DDPG algorithm.
These two algorithms will be briefly introduced below.
\subsection{Deep Deterministic Policy Gradient (DDPG)}
The DDPG algorithm collects sample data $(s_j, a_j, s_{j+1}, r_j)$ from the replay buffer $R$ and trains a critic network ${Q}_\phi$ and actor network $u_\theta$ parameterized by $\phi$ and $\theta$.
More specifically, the critic network is updated by:
\begin{equation}
\label{eq:ddpg_critic}
\begin{aligned}
& \phi \leftarrow \argmin_\phi\frac{1}{N_{dd}}\sum_{j = 1}^{N_{dd}}\left(y_j - {Q}_{{\phi}}(s_j, a_j)\right)^2, \\
& y_j = r_j + \gamma {Q}_{\hat{\phi}}(s_{j+1}, u_{\hat{\theta}}(s_{j+1})),
\end{aligned}
\end{equation}
where $N_{dd}$ is the batch size for DDPG, $\hat{\phi}, \hat{\theta}$ are parameters of the target critic network and target actor network, and $\gamma$ is the discount for future reward.
The policy network is updated by:
\begin{equation}
\label{eq:ddpg_actor}
\begin{aligned}
\theta \leftarrow \argmax_\theta\frac{1}{N_{dd}}\sum_{j = 1}^{N_{dd}} {Q}_{\hat{\phi}}(s_j, u_\theta(s_j)),
\end{aligned}
\end{equation}
where $\theta$ is the parameters for the policy network to be optimized. Policy gradient is applied to update the parameters of the actor network:
\begin{equation}
\label{eq:ddpg_policy_gradient}
\begin{aligned}
\theta \leftarrow \theta + \alpha \frac{1}{N_{dd}}\sum_{j = 1}^{N_{dd}}\nabla_a \hat{Q}(s,a)|_{s = s_j, a=a_j} \nabla_\theta u_\theta(s)|_{s = s_j},
\end{aligned}
\end{equation}
where the $\alpha$ is the learning rate of the actor network.
The target networks are updated by
\begin{equation}
\label{eq:ddpg_target_update}
\begin{aligned}
&\hat{\phi} \leftarrow \delta \phi + (1 - \delta) \hat{\phi}, \\
&\hat{\theta} \leftarrow \delta \theta + (1 - \delta) \hat{\theta},
\end{aligned}
\end{equation}
where $\delta$ is the target update rate and is set to be small value ($\delta\approx 0.01$).
\subsection{Guided Policy Search (GPS)}
With the involvement of guiding distribution $p(\tau)$,Problem~(\ref{eq:loss_minimization}) can be rewritten as
\begin{equation}
\label{eq:gps_minimization}
\begin{aligned}
\min_{\pi_\theta, p} E_{p}(l(\tau)), \quad s.t. \ \ p(\tau) = \pi_\theta(\tau).
\end{aligned}
\end{equation}
GPS solves~(\ref{eq:gps_minimization}) by alternatively minimizing the augmented Lagrangian with respect to primal variables $p, \pi_\theta$ and updating the Lagrangian multipliers~$\lambda$. The augmented Lagrangian for $\theta$ and $p$ optimization are:
\begin{equation}
\label{eq:gps_Lagrangian}
\begin{aligned}L_p(p, \theta) = & E_p(l(\tau)) + \lambda \left(\pi_\theta(\tau)- p(\tau)\right) + \\
& \nu D_{KL}\left(p(\tau)\|\pi_\theta(\tau)\right),\\
L_\theta(p, \theta) = & E_p(l(\tau)) + \lambda \left(\pi_\theta(\tau)- p(\tau)\right) + \\
& \nu D_{KL}\left(\pi_\theta(\tau)\|p(\tau)\right),
\end{aligned}
\end{equation}
where $\lambda$ is the Lagrangian multiplier, $\nu$ is the penalty parameter for the violation of the equality constraint, and $D_{KL}$ represents the KL-divergence.
The optimization of primal variable $p$ is called trajectory optimization. It optimizes the guiding distribution $p$ with learned local dynamics. To assure the accuracy of dynamics fitting, the optimization is constrained within the trust region $\epsilon$:
\begin{equation}
\label{eq:gps_trajopt}
\begin{aligned}
\min_{p} L_p(p, \theta), \quad s.t. \ \ D_{KL}(p(\tau)\|\hat{p}(\tau)) \leq \epsilon,
\end{aligned}
\end{equation}
where $\hat{p}$ is the guiding distribution of the previous iteration. The Lagrangian of~(\ref{eq:gps_trajopt}) is:
\begin{equation}
\label{eq:gps_trajopt_lagrangian}
\begin{aligned}
\mathcal{L}(p) = L_p(p,\theta) + \eta(D_{KL}(p(\tau)\|\hat{p}(\tau)) - \epsilon),
\end{aligned}
\end{equation}
where $\eta$ is the Lagrangian multiplier for the constraint optimization. With the Gaussian assumption of the dynamics,~(\ref{eq:gps_trajopt_lagrangian}) is solved by iLQG. To avoid large derivation from the fitted dynamics, $\eta$ is adapted by comparing the predicted KL-divergence with the actual one.
The optimization of the policy parameters $\theta$ can be written as a supervised learning problem. With the Gaussian policy $\pi_\theta(a_t|o_t) = \mathcal{N}(u_\theta(o_t), \Sigma^\pi_t)$, we can rewrite $L_\theta(p,\theta)$ in~(\ref{eq:gps_Lagrangian}) as:
\begin{equation}
\label{eq:gps_polopt_objective}
\begin{aligned}
&L_\theta(\theta, p) = \frac{1}{2N_b}\sum_{i,t=1}^{N_b,T}E_{p_i(s_t,o_t)}[
\text{tr}\left(C_{ti}^{-1}\Sigma^\pi_t
\right)-\text{log}|\Sigma^\pi_t|+\\&
\left(u_\theta(o_t) - u_{ti}^p(s_t)\right)^TC_{ti}^{-1}\left(u_\theta(o_t) - u_{ti}^p(s_t)\right) + 2\lambda^T_{t}u_\theta(o_t)
],
\end{aligned}
\end{equation}
where $p_i(u_t|s_t)\sim \mathcal{N}(u_{ti}^p(s_t), C_{ti})$ is the guiding distribution.
Equation~(\ref{eq:gps_polopt_objective}) contains the decoupled form of the variance optimization and policy optimization.
Refer~\cite{levine2016end} for more details.
\begin{figure}[t]
\begin{center}
{\includegraphics[width =\linewidth]{gps_ddpg.png}}
\caption{(a) Guided Policy Search (GPS) and (b) Deep deterministic policy gradient (DDPG). }
\label{fig:GPS-DDPG}
\end{center}
\end{figure}
\subsection{Comparison of GPS and DDPG}
GPS decouples RL into a trajectory optimization (\textit{supervisor}) and a supervised policy network learning (\textit{learner}), as shown in Fig.~\ref{fig:GPS-DDPG}(a). The performance of the learner relies on the quality of the supervisor. By fitting the dynamics from sampling data and computing the supervisor with the optimal control, GPS is more efficient than the DDPG and many other model-free RL algorithms.
However, the performance of the learner would be compromised if the system has high stiffness and has force/torque feedback as states due to the less smooth dynamics and smaller trust region.
In comparison, DDPG uses rollout samples to jointly train the Q-network (\textit{critic}) and policy network (\textit{actor}), as shown in Fig.~\ref{fig:GPS-DDPG}(b). The critic gradually builds up the Q-value from physical rollouts, and the Q-value is applied to train the actor network based on policy gradient.
The actor-critic framework provides more stable policy in the tasks with non-smooth dynamics. These tasks are common in high precision industrial assembly where the system has higher stiffness and contains force/torque feedback in the states. Despite the reliable performance, the actor-critic framework is less data efficient due to the intensive exploration, which is usually unnecessary since assembly tasks only requires exploration in narrow trajectory space.
\section{Proposed Approach}
\label{sec:proposed}
Precision industrial assembly usually has large system stiffness in order to achieve precise tracking performance and reduce the vibration.
With large stiffness, small clearance and force/torque feedback, both the model-free RL and model-based method cannot accomplish the task efficiently and stably.
In this paper, we propose a learning framework that combines the actor-critic with the model-based RL for high precision industrial assembly. The framework is named as guided-deep deterministic policy gradient (guided-DDPG). Guided-DDPG behaves more efficient than the actor-critic and more stable/reliable than the model-based RL.
\begin{figure}[t]
\begin{center}
{\includegraphics[width =0.8\linewidth]{guided_ddpg.png}}
\caption{Illustration of the proposed learning framework (guided-DDPG). Trajectory optimization provides initial guidance to both actor and critic nets to avoid excessive exploration. The actor-critic nets gradually establish the evaluation system, instead of relying on pure supervised learning.}
\label{fig:guided-DDPG}
\end{center}
\end{figure}
Figure~\ref{fig:guided-DDPG} illustrates the proposed guided-DDPG algorithm.
Due to the discontinuity of the fitted dynamics in rigid precise systems, the trajectory optimization can have inconsistent behavior or requires dedicated parameter tuning.
Therefore, a pure supervised learning from trajectory optimization cannot fulfill the task consistently.
The actor-critic is incorporated to the framework to address this issue. The trajectory optimization serves as a \textit{semi-supervisor} to train the actor-critic to establish the initial critic and constrain the network in narrow task space.
The involvement of the supervision will be reduced as the training progresses and the critic network becomes more accurate, since the actor-critic exhibits superior performance than the semi-supervisor.
To be more specific, the trajectory optimization (semi-supervisor) has the following form:
\begin{equation}
\label{eq:trajopt_klcon}
\begin{aligned}
\min_{p} E_p(l(\tau)), \quad s.t. \ \ D_{KL}(p(\tau)\|\hat{p}_\theta(\tau)) \leq \epsilon,
\end{aligned}
\end{equation}
where $\hat{p}_\theta$ is set as the trajectory distribution generated by actor policy at the first sub-iteration, and is set as the previous trajectory distribution $\hat{p}$ for the successive $N_{trajopt}-1$ sub-iterations.
Equation~(\ref{eq:trajopt_klcon}) is optimized by the dual:
\begin{equation}
\label{eq:trajopt_lagrangian}
\begin{aligned}
\max_\eta \{\min_{p} E_p(l(\tau)) + \eta (D_{KL}(p(\tau)\|\hat{p}_\theta(\tau)) - \epsilon)\}.
\end{aligned}
\end{equation}
The optimization of $p$ is solved by LQG with fixed $\eta$ and dynamics, and the optimization of $\eta$ is done heuristically: decrease $\eta$ if $D_{KL}(p(
\tau)\| \hat{p}_\theta(\tau)) < \epsilon$, otherwise increase $\eta$. The trust region $\epsilon$ varies based on the expected improvement and actual one. $\epsilon$ would be reduced once the actual improvement is far smaller from the expected one, thus the network focuses on penalizing the KL divergence from $\hat{p}_\theta(\tau)$.
We collect the trajectory after $N_{trajopt}$ sub-iterations to replay buffer $R_1$ for supervised training of actor-critic nets, and feed all the sample data during $N_{trajopt}$ executions to replay buffer $R_2$. With the supervision from $R_1$, the critic is trained by:
\begin{equation}
\label{eq:guided_ddpg_critic}
\begin{aligned}
& \phi \leftarrow \argmin_\phi\frac{1}{N_{dd}}\sum_{j = 1}^{N_{dd}}\left(y_j - {Q}_{{\phi}}(s_j, a_j)\right)^2 + \\ & \qquad \quad w_{to} \frac{1}{N_{to}}\sum_{i=1}^{N_{to}} \|{Q}_\phi(s_i, a_i) - Q_i^{to}\|^2\\
\end{aligned}
\end{equation}
where $w_{to}, N_{to}$ are the weight and batch size of the semi-supervisor, $y_j$ has the same form as~(\ref{eq:ddpg_critic}). $(s_i, a_i, Q_i^{to})$ is the supervision data from $R_1$, and $(s_j, a_j, r_j, s_{j+1})$ is the sample data from $R_2$.
The actor is trained by:
\begin{equation}
\label{eq:guided_ddpg_actor}
\begin{aligned}
\theta \leftarrow & \argmax_\theta\frac{1}{N_{dd}} \sum_{j = 1}^{N_{dd}} {Q}_{\hat{\phi}}(s_j, u_\theta(s_j)) + \\ & \quad w_{to} \frac{1}{N_{to}}\sum_{i=1}^{N_{to}} \|u_\theta(s_i) - a_i\|^2\\
\end{aligned}
\end{equation}
The supervision weight $w_{to}$ decays as the number of training rollouts $N_{roll}$ increases. We use $w_{to} = \frac{c}{N_{roll} + c}$, where $c$ is a constant to control the decay speed.
The guided-DDPG algorithm is summarized in Alg.~\ref{alg:guided_ddpg}. The critic and actor are initialized in Line~\ref{alg:init}. Guided-DDPG runs for $EP$ epochs in total. In each epoch, semi-supervisor is first executed to update the trajectories for supervision. With the high stiffness, small clearance and the force/torque feedback, the fitted dynamics (Line~\ref{alg:fit_dynamics}) is discontinuous and has small trust region. Therefore, the trajectories generated from semi-supervisor might be sub-optimal. Nevertheless, they are sufficient to guide the initial training of the actor-critic.
The actor-critic is trained in Line~(\ref{alg:ddpg} - \ref{alg:end_ddpg}) following the standard procedure of DDPG with the modified objective function (Line~(\ref{alg:ddpg_modifications})).
The supervision weight $w_{to}$ is decreased as the training progresses due to the superior performance of the actor-critic than the semi-supervisor.
\begin{algorithm} [t]
\caption{Guided-DDPG}\label{alg:guided_ddpg}
\begin{algorithmic}[1]
\State \textbf{input:}$EP,N_{ddpg},N_{inc}, N_{trajopt},N_{roll}=0,R_{1/2}\leftarrow\Phi$\label{alg:input}
\State \textbf{init:} $Q_\phi(s,a), u_\theta(s)$, set target nets $\hat{\phi}\leftarrow\phi, \hat{\theta}\leftarrow\theta$ \label{alg:init}
\For {$epoch=0:EP$} \label{alg:trajopt}
\State $p_{prev}\leftarrow u_\theta$
\For {$it=0:N_{trajopt}$}
\State $\mathcal{S}\leftarrow sample\_data(p_{prev})$, $R_2\leftarrow R_2\cup\mathcal{S}$
\State $f_{dy}\leftarrow fit\_dynamics(\mathcal{S})$\label{alg:fit_dynamics}
\State $\hat{p}_\theta \leftarrow linearize\_policy(p_{prev},\mathcal{S})$
\State $p\leftarrow update\_trajectory(f_{dy},\hat{p}_\theta), p_{prev}\leftarrow p$
\EndFor
\State $S\leftarrow sample\_data(p)$, $R_1\leftarrow R_1\cup\mathcal{S}, R_2\leftarrow R_2\cup\mathcal{S}$
\For {$it=0:N_{ddpg}$} \label{alg:ddpg}
\State $\mathcal{N}_{ex}\leftarrow exploration\_noise()$
\State $s_0\leftarrow observe\_state(), w_{to} = \frac{c}{c+N_{roll}\texttt{++}}$
\For {$t=0:T$}
\State $a_t = u_\theta(s_t)+\mathcal{N}_{ex}(t)$, observe $s_{t+1},r_t$
\State $R_2\leftarrow R_2\cup(s_t,a_t,s_{t+1},r_t)$
\State sample $N_{to}, N_{dd}$ transitions from $R_1, R_2$
\State update critic and actor nets by (\ref{eq:guided_ddpg_critic}) and (\ref{eq:guided_ddpg_actor})\label{alg:ddpg_modifications}
\State update target nets by (\ref{eq:ddpg_target_update})
\EndFor
\EndFor\label{alg:end_ddpg}
\State $N_{ddpg} \leftarrow N_{ddpg} + N_{inc}$
\EndFor
\end{algorithmic}
\end{algorithm}
\section{Simulations and Experiments}
\label{sec:experiments}
This section presents both the simulation and experimental results of the guided-DDPG to verify the effectiveness of the proposed learning framework. The videos are available at~\cite{website}.
\begin{figure}[t]
\begin{center}
{\includegraphics[width =0.9\linewidth]{simulation_environment.png}}
\caption{Two simulation tasks for algorithm evaluation. (a) Lego brick insertion, (b) U-shape joint assembly. }
\label{fig:simulation_environment}
\end{center}
\end{figure}
\begin{figure*}[t]
\begin{center}
{\includegraphics[width =\linewidth]{LegoUsequence.png}}
\caption{Simulation animations of the proposed guided-DDPG on (Top) U-shape joint assembly and (Bottom) Lego brick insertion. The guided-DDPG was trained on 2$\times$2 Lego and tested on 4$\times$2 one. Snapshots are taken from left to right.}
\label{fig:Usequence}
\end{center}
\end{figure*}
To compare the performance of the guided-DDPG with other state-of-the-art RL algorithms, we built up a simulation model using the Mujoco physics engine~\cite{todorov2012mujoco}. The host computer we used was a desktop with 32GB RAM, 4.0GHz CPU and GTX 1070 GPU.
A 6-axis UR5 robot model from universal robotics was used to perform the tasks. Two different assembly tasks were simulated, the first one was the Lego brick insertion, and the second one was the U-shape joint assembly, as shown in Fig.~\ref{fig:simulation_environment}.
\subsection{Parameter Lists}
The number of the maximum epoch is set to $EP=100$, initial number of rollouts for DDPG and trajectory optimization were $N_{ddpg}=21$ and $N_{trajopt}=3$, respectively. To ensure less visit of trajectory optimization as the training progresses, we increased the number of rollouts by $N_{inc}=15$ for each DDPG iteration. The sizes of the replay buffer $R_1, R_2$ were 2000 and 1E6, respectively.
The soft update rate $\gamma=0.001$ in (\ref{eq:ddpg_target_update}). The batch size for trajectory optimization $N_{to}$ and DDPG $N_{dd}$ were both 64.
The algorithm used a cost function $l(s,a) = 0.0001\|a\|_2 + \|FK(s) - p_{tgt}(s)\|_2$, where $FK$ represents the forward kinematics and $p_{tgt}$ is the target end-effector points.
\subsection{Simulation Results}
\begin{figure}[t]
\begin{center}
{\includegraphics[width =0.9\linewidth]{score_different_guidance.png}}
\caption{Comparison of different supervisions with Lego brick insertion task. The supervision methods with performance in descending order: guided-DDPG (proposed), DDPG with supervised data in Replay buffer, pure DDPG, guided-DDPG w/o policy guidance, and guided-DDPG w/o critic guidance. }
\label{fig:score_different_guidance}
\end{center}
\end{figure}
\begin{figure*}[t]
\centering
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{score_different_weights.png}
\caption{}
\label{fig:score_different_weights}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{compare_methods_Lego.png}
\caption{}
\label{fig:score_different_methods}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.32\textwidth}
\centering
\includegraphics[width=\textwidth]{compare_methods_U}
\caption{}
\label{fig:score_different_U}
\end{subfigure}
\caption{(a) Illustration of the supervision weights on Lego brick insertion task. (b) Comparison of the algorithms for Lego brick insertion task. (c) Comparison of the algorithms for U-shape joint assembly task.}
\label{fig:three graphs}
\end{figure*}
The simulation results on U-shape joint assembly and Lego brick insertion are shown by Fig.~\ref{fig:Usequence}. Both simulations were trained with assembly clearance as 0.1 mm. Guided-DDPG takes poses and force/torque measurements of the end-effector as the states, and generates joint torques as action to drive the robot.
The U-shape joint has more complicated surface than the Lego brick, and a successful assembly requires matching the shapes twice, as shown in Fig.~\ref{fig:Usequence} (Top). Despite the difficulties, the proposed algorithm was able to train the policy within 1000 rollouts. We also visualized the adaptability of the trained policy on the Lego brick insertion task, as shown in Fig.~\ref{fig:Usequence} (Bottom). The policy was trained with a brick of size 2$\times$2 and clearance 0.1 mm and tested with a brick of size 4$\times$2 and clearance 1 $\upmu$m. Moreover, the brick position had an unknown offset (1.4 mm) to the network. The proposed network was able to address these uncertainties and successfully inserted the brick to a tighter hole with uncertain position.
\subsubsection{Comparison of Different Supervision Methods}
The proposed learning framework guides both the critic and actor. To illustrate the necessity of the proposed guidance, we compared the results of guided-DDPG with several other supervision methods, including the guided-DDPG with partial guidance, pure-DDPG with supervision data to replay buffer (no supervision on objective function) and the pure-DDPG. The result was shown in Fig.~\ref{fig:score_different_guidance}. The proposed guided-DDPG achieved the best performance.
The partial guidance without critic (Fig.~\ref{fig:score_different_guidance} Green) was able to guide the actor and realized safe exploration at the beginning. However, the actor network behaved worse as the involvement of the semi-supervisor reduced and the weight of the critic increased, since the critic is trained purely by the contaminated target actor~(\ref{eq:ddpg_critic}). In contrast, the partial guidance without actor~(Fig.~\ref{fig:score_different_guidance} Orange) had poorly behaved actor since the actor was trained purely by the policy gradient using the contaminated critic~(\ref{eq:ddpg_actor}).
The pure-DDPG with supervision data~(Fig.~\ref{fig:score_different_guidance} Purple) achieved better performance than pure-DDPG, since the trajectories obtained from semi-supervisor were better behaved than the initial rollouts of DDPG. This kind of supervision is similar with the human demonstration in~\cite{vecerik2017leveraging}.
\subsubsection{Effects of the Supervision Weight $w_{to}$}
The supervision weight $w_{to}$ balances the model-based supervision and model-free policy gradient in actor/critic updates, as shown in (\ref{eq:guided_ddpg_actor}) and (\ref{eq:guided_ddpg_critic}). The results of different weights on Lego brick insertion are shown in Fig.~\ref{fig:three graphs} (a). With $c = 1$, the supervision weight is $w_{to} = \frac{1}{1+N_{roll}}$. The weights starts with 1 and decays to 0.001 as $N_{roll}=1000$, while $c=100$ makes $w_{to}$ decay to 0.1 as $N_{roll}= 1000$. Slower decay provides excessive guidance by the semi-supervisor and contaminates the original policy gradient and makes the DDPG unstable. Empirically, $c=1\sim10$ achieves comparable results.
\subsubsection{Comparison of Different Algorithms}
The proposed learning framework was compared with other state-of-the-art algorithms, including the pure-DDPG, twin delayed deep deterministic policy gradients (TD3)~\cite{td3} and the soft actor-critic (SAC)~\cite{sac}. Default parameters were used for TD3, as shown in~\cite{rlkit}. As for SAC, we used the default parameters in~\cite{rlkit} with tuned reward scale as 10. The comparison result on Lego brick insertion task is shown in Fig.~\ref{fig:three graphs} (b). The proposed guided-DDPG passed the success threshold (shaded purple line) at the 800 rollouts and consistently succeeded the task after 2000 rollouts. In comparison, the pure DDPG passed the success threshold at the 5000 rollouts and collapsed around 10000 rollouts. The performance of pure DDPG was inconsistent in seven different trials. TD3 and SAC had the similar efficiency with pure DDPG. The comparison of the algorithms on U-shape joint assembly is shown in Fig.~\ref{fig:three graphs} (c). Similar with Lego brick insertion, the guided-DDPG achieved more stable and efficient learning. The time efficiency and data-efficiency of the DDPG and guided-DDPG are compared in Table~\ref{tab:comparison}.
\begin{table}[t]
\centering
\caption{Comparison between DDPG and guided-DDPG}
\label{tab:comparison}
\begin{tabular}{c|c|c}
items & DDPG & Guided-DDPG\\
\hline
time (min) & 83 & \textbf{37.3} \\
\hline
data (rollouts) & 7000 & \textbf{1500} \\
\hline
\end{tabular}
\end{table}
\begin{figure*}[ht]
\centering
\begin{subfigure}[b]{0.214\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{exp_setup.jpg}
\caption{}
\label{fig:exp_setup}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.77\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{exp_results.png}
\caption{}
\label{fig:exp_results}
\end{subfigure}
\caption{(a) Experimental setup, and (b) experimental results for Lego brick insertion.}
\label{fig:two_graphs}
\end{figure*}
\subsubsection{Adaptability of the Learned Policy}
\begin{figure}[t]
\begin{center}
{\includegraphics[width =0.9\linewidth]{adaptive.png}}
\caption{Different shapes of the bricks and holes for adaptability test. (a) 2$\times$2 brick used in training, (b) 4$\times$2 brick, (c) 4$\times$2 brick with incomplete hole, and (d) cylinder brick. }
\label{fig:adaptability}
\end{center}
\end{figure}
The adaptability of the learned policy is discussed in this section. Three different types of uncertainties were considered. The first type was the unknown hole position. The learned policy was able to successfully insert the brick when moving the hole to an uncalibrated position (maximum offset is 5 mm, hole has width of 16 mm). The second type of uncertainty was the shapes of peg/hole. We found that the learned policy is robust to different shapes shown in Fig.~\ref{fig:adaptability}. The third type was the different clearance. The policy was trained with clearance 0.1 mm and tested successfully on insertion tasks with clearance 10 $\upmu$m, 1 $\upmu$ and 0.
The simulation videos are available at~\cite{website}.
\subsection{Experimental Results}
Experimental results are presented in this section. The Lego brick was attached to a 3D printed stick at the end-effector of the Universal robot (UR5). A Robotiq FT 300 force torque sensor was used to collect the force/torque signal at the wrist. The experimental setup is shown in Fig.~\ref{fig:two_graphs}(a).
The policy took the estimated hole position and the force/torque reading as inputs, and generated transitional velocities for the end-effector. The velocity was tracked by a low-level tracking controller.
The clearance of the Lego brick is less than 0.2 mm. The target position of the hole had 0.5 mm uncertainty, yet the policy was able to successfully locate the hole and insert the brick, as shown in Fig.~\ref{fig:two_graphs}(b).
It took 2 hours for pure-DDPG to find a policy in the exploration space bounded within $1$ mm around the hole, and took 1.5 hours for guided-DDPG to find a policy in a larger exploration space bounded within $3$ mm around the hole.
The experimental videos are shown in~\cite{website}.
\section{Conclusions and Future Works}
\label{sec:conclusions}
This paper proposed a learning framework for high precision assembly task. The framework contains a trajectory optimization and an actor-critic structure. The trajectory optimization was served as a semi-supervisor to provide initial guidance to actor-critic, and the critic network established the ground-truth quality of the policy by learning from both the semi-supervisor and exploring with policy gradient. The actor network learned from both the supervision of the semi-supervisor and the policy gradient of the critic. The involvement of critic network successfully addressed the stability issue of the trajectory optimization caused by the high-stiffness and the force/torque feedback.
The proposed learning framework constrained the exploration in a safe narrow space, improved the consistency and reliability of the model-based RL, and reduced the data requirements to train a policy. Simulation and experimental results verified the effectiveness of the proposed learning framework.
In the future, the authors would evaluate the algorithm on more realistic industrial applications such as connector insertion, furniture assembly and tight peg-in-hole tasks.
\addtolength{\textheight}{-1cm}
\section*{Acknowledgment}
The authors would like to thank Dr. Yotto Koga and AI Lab in Autodesk Inc. for the help on experiments.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Estimating the parameter of a Bernoulli process is a fundamental problem in statistics and signal processing.
From the binary-valued outcomes of independent and identically distributed (i.i.d.) trials (generically failure (0) or success (1)), we wish to estimate the probability $p$ of success.
Among myriad applications, our primary interest is raster-scanned active imaging in which a scene patch is periodically illuminated with a pulse, and each illumination period (Bernoulli trial) either has a photon-detection event (success) or not (failure)~\cite{Shin2015}.
The probability $p$ of a photon-detection event has a monotonic relationship with the reflectivity of the scene patch, and a monotonic function of an estimate of $p$ becomes the corresponding image pixel value.
For efficiency in acquisition time or illumination energy, we are motivated to form the image from a small number of illumination pulses, under conditions where $p$ is small.\footnote{For applications using time-correlated single photon counting driven by a detector with dead time,
such as a single-photon avalanche diode (SPAD),
it is recommended to keep $p$ below $0.05$ to avoid time skew and missed detections~\cite{tcspcWahl2015}.}
Other types of raster-scanned imaging can be modeled similarly assuming that the dwell time is an integer multiple of some base time interval, during which the observations are binary.
Conventional methods are not adaptive. With a fixed number of trials $n$, the number of successes $K$ is a binomial random variable, and the maximum likelihood (ML) estimate of $p$ is $K/n$.
Though less common in active imaging, a well-known alternative in the statistics literature is to fix a number of successes $\ell$.
Repeating trials until success $\ell$ occurs results in a random number of trials $M$ that is a negative binomial random variable,%
\footnote{Note that the negative binomial distribution is defined inconsistently in the literature, with sometimes the number of failures being fixed rather than the number of successes (reversing the roles of $p$ and $1-p$).
}
and the ML estimate of $p$ is $\ell/M$.
While there may seem to be nothing to design here,
a constraint on the mean number of trials opens up possibilities for data acquisition that results in neither binomial nor negative binomial distributions.
The mean may be
over a multiplicity of (non-random) Bernoulli process parameters to estimate
(such as in active imaging with one parameter per pixel) or
over a prior for a single Bernoulli parameter.
The two cases are formally linked through the relative frequency interpretation of probability,
with the empirical distribution of the multiplicity of deterministic parameters in the former case
playing the role of the prior distribution in the latter case~\cite{Fine2006}.
For multiple deterministic parameters,
we have a resource allocation problem reminiscent of bit allocation in transform coding~\cite{Segall:76,Goyal:01a}.
As we will demonstrate,
in an oracle-aided setting, trials can be allocated to maximize a \emph{trial allocation gain} that is analogous to the coding gain of transform coding.
For a single random parameter, a simple and implementable approach~-- not requiring an oracle~-- asymptotically achieves the optimal trial allocation gain and may perform better than the oracle-aided method for moderate numbers of trials.
The focus of this paper is on allocating trials in the estimation of a single random parameter through the design of a stopping rule.
A stopping rule may~-- implicitly and stochastically~-- allocate trials differently for different values of $p$, even though $p$ is not known \emph{a priori}.
We show that any optimal stopping rule can be described by a trellis rather than a more complicated graph,
and greedy construction of the trellis is very nearly optimal.
For a rectangular array of Bernoulli processes representing a scene in an imaging problem, applying a good stopping rule allocates more trials to the pixels where they provide the most benefit.
The final image formation may include a method such as
total variation (TV) regularization for exploiting spatial correlations among neighbors.
{Regularized image formation makes it more difficult to optimize the acquisition, but it does not invalidate the advantage from adaptive acquisition.}
In simulations with parameters realistic for active optical imaging, we demonstrate a reduction in mean-squared error (MSE) by a factor of up to 2.73 (4.36~$\mathrm{dB}$) in comparison to the same regularized reconstruction approach applied without adaptation in numbers of trials.
Such gains vary based on image content, and gains without regularization are predictable from the trial allocation gain formulation.
\subsection{Related Work}
\subsubsection{Statistics Literature}
In statistics, forming a parameter estimate from a number of i.i.d.\ observations that is dependent on the observations themselves is called \emph{sequential estimation}~\cite{Anscombe1953}.
Early interest in sequential estimation of a Bernoulli process parameter was inspired by the high relative error of deterministically stopping after $n$ trials when $p$ is small.
Specifically, the standard error of the ML estimate is $\sqrt{p(1-p)/n}$, which for small $p$ is unfavorable compared to anything proportional to $p$.
This shortcoming manifests, for example, in requiring large $n$ to distinguish between two small possible values for $p$.
Haldane~\cite{Haldane1945} observed that if one stops after $\ell$ successes, the (random) number of trials $M$ is informative about $p$.
Specifically, $(\ell-1)/(M-1)$ is an unbiased estimate of $p$ (provided $\ell \geq 2$), and its standard error is proportional to $p$ (provided $\ell \geq 3$).
(The ML estimate $\ell/M$ is not unbiased, though $M/\ell$ is an unbiased estimate of $1/p$.)
Tweedie~\cite{Tweedie1945} suggested to call this inverse binomial sampling, but the resulting random variable is now commonly known as negative binomial or Pascal distributed.
More recent works have focused on
non-MSE performance metrics~\cite{Cabilio1975,Cabilio1977},
estimation of functions of $p$~\cite{Hubert2000},
estimation from imperfect observations~\cite{Djuric2000},
and composite hypothesis testing~\cite{Ciuonzo2015}.
\subsubsection{Photon-Efficient Imaging and Variable Dwell Time}
First-photon imaging~\cite{FPI2014} introduced sequential estimation to active imaging.
This method uses the number of illumination pulses until the first photon is detected to reveal information about reflectivity, setting $\ell=1$ in the concept of Haldane~\cite{Haldane1945} and thus using geometric sampling as a special case of negative binomial sampling.
A censoring method is used to approximately separate signal and background detections, and
spatial correlations are used to regularize the estimation of the full scene reflectivity image, resulting in good performance from only 1 detected photon per pixel, even when half of the detected photons are attributable to uninformative ambient light.
Subsequent work with binomial sampling (and otherwise identical experimental conditions) resulted in similar performance~\cite{Shin2015}, and
greatly increasing robustness to ambient light is largely attributable to improving the censoring step~\cite{Rapp2017}.
These works leave questions on the importance of negative binomial sampling to first-photon imaging unanswered;
comparing first-photon imaging to photon-efficient methods with deterministic dwell time~\cite{Krichel2010,Morris2015,Shin2015,Shin2015b,Altmann2016,Shin2016camera,Shin2016,Shin2016multidepth,Mertens2017,Rapp2017,Altmann2017,Halimi2017}
was an initial inspiration for the present work.
While recent works have exploited the first-photon idea in imaging techniques such as ghost imaging~\cite{Liu2018,Altmann2018} and \mbox{x-ray} tomography~\cite{Zhu2018},
previous uses of variable dwell time are not closely connected to sequential estimation or the result of optimized resource allocation.
For example, in lidar, varying dwell time to maintain approximately constant signal strength despite varying effective reflectivity (including greater radial fall-off for more distant scene patches) dates back to at least the 1970s~\cite{Lipke:1979}.
He \emph{et al.}~\cite{He2017} closely follow the technique of \cite{Shin2015}, including its background censoring, and vary the dwell time to keep the number of photon detections after censoring (i.e., photon detections attributed to signal rather than background) at each pixel approximately constant.
In scanning electron microscopy, Dahmen \emph{et al.}~\cite{Dahmen2016} increase dwell time where a measure of image detail is large.
To the best of our knowledge, no previous paper has formally optimized dwell time under a Bernoulli process measurement model.
\subsection{Main Contributions and Preview of Results}
\subsubsection{Framework}
This work discusses a novel framework for depicting and understanding stopping rules for sequential estimation of Bernoulli parameters under number of trials constraints
(Section~\ref{sec:single-bernoulli}).
In this framework, first presented in~\cite{MedinMBG:2018}, each Bernoulli trial corresponds to a transition in a trellis in which each node is identified by the number of trials and number of successes;
it is easily shown that distinct paths to reach a given node need not be distinguished.
A stopping rule is the assignment of probabilities of stopping to each node in the trellis (see Figs.~\ref{fig:GeneralTrellis}--\ref{fig:example-trellis}).
By construction, a stopping rule defined in this way is implementable because it does not depend on knowledge of $p$ or non-causally on the Bernoulli process.
This framework applies equally well under any prior for $p$.
\subsubsection{Stopping Rule Design}
Simple stopping rules lead to binomial
(Fig.~\ref{fig:stopping_rules-bin-negbin}(a))
and negative binomial
(Fig.~\ref{fig:stopping_rules-bin-negbin}(b))
sampling.
Specializing to the Beta family of priors,
which is both convenient and conventional because it is the conjugate prior for the relevant observation distributions,
methods to optimize the stopping rule are presented in decreasing order of computational complexity:
dynamic programming (Section~\ref{ssec:dynamic_prog}),
offline greedy design (Section~\ref{ssec:greedy-offline}), and
online threshold-based termination (Section~\ref{ssec:greedy-online}) first introduced in~\cite{MedinMBG:2018}.
Empirically, all three methods, including the online method requiring no storage of a precomputed stopping rule, provide very similar performance.
Thus, the easily implementable online method provides very nearly optimal performance.
\subsubsection{Analysis in Oracle-Aided Setting}
This paper introduces the concept of oracle-aided trial allocation whereby processes with different parameters are allocated different fractions of an overall trial budget (Section~\ref{sec:oracle-aided}).
This yields a readily-computed \emph{trial allocation gain} that can be arbitrarily large, though it is generally modest
(Section~\ref{sec:trial-allocation-gain}).
Furthermore, we show that under any Beta prior the threshold-based stopping asymptotically allocates trials identically to the oracle-based optimal
(Section~\ref{ssec:comparison_oracle}).
\subsubsection{Evaluation}
In simulations inspired by realistic active imaging scenarios, an MSE improvement factor of up to $4.36\,\mathrm{dB}$ is demonstrated where spatial correlations are exploited through total variation (TV) regularization
(Section~\ref{sec:image_p}).
Without TV regularization, achieved gains are close to the values predicted by the theoretical trial allocation gain.
For example, the theoretical trial allocation gain is 2.29~dB, and a gain of 2.27~dB is realized for the Shepp-Logan phantom.
\subsubsection{Estimating functions of $p$}
Inspired by applications where estimating functions of $p$ is of interest \cite{Hubert2000}, online threshold-based termination is also extended to estimating $\log p$ from Bernoulli observations
(Section~\ref{sec:estimating-functions}).
Experimental results without TV regularization demonstrate improvements of up to $1.86$~dB using the threshold-based stopping rule versus the conventional binomial sampling
(Section~\ref{sec:image_log_p}).
\section{Trial Allocation Across Multiple Processes}
\label{sec:trial-allocation}
Consider the estimation of the parameters $\{p_i\}_{i=1}^r$ of a finite number $r$ of Bernoulli processes with binomial sampling of each process.
When $m_i$ trials of process $i$ are observed,
the MSE of the ML estimate of $p_i$ is $p_i(1-p_i)/m_i$.
Suppose that we are interested in making the average of the MSEs,
\[
\frac{1}{r}\sum_{i=1}^r \frac{p_i(1-p_i)}{m_i},
\]
small under a constraint on the average of the numbers of trials
$(1/r)\sum_{i=1}^r m_i \leq \trialbudget$.
Since $p_i(1-p_i)$ varies over $[0,\,1/4]$ for $p_i \in [0,1]$, there can be an advantage to varying the $m_i$ values.
However, that allocation of trials depends on parameters that are to be estimated.
In this section, we suspend the need for implementability and instead study the optimal trial allocation as if the parameters were known.
This provides a benchmark for the implementable methods developed in the remainder of the paper,
with $\{p_i\}_{i=1}^r$ playing the role of a discrete prior on $p$.
We also consider $r \rightarrow \infty$ to reach a distributional limit.
\subsection{Oracle-Aided Optimal Allocation}
\label{sec:oracle-aided}
In optimizations such as
\begin{equation}
\min_{m_i, \ i=1,\,2,\,\ldots,\,r} \, \sum_{i=1}^r \frac{p_i(1-p_i)}{m_i} \mathrm{\ \ s.t. \ \ } \sum_{i=1}^{r} m_i \leq r \trialbudget,
\label{eq:opt_oracleaided_bin}
\end{equation}
ignoring that each $m_i$ should be a positive integer,
each MSE vs.\ number of trials trade-off should be at the same slope,
else it would be advantageous to shift trial resources to the process for which the benefit (MSE reduction) per trial is largest.
This is formalized using the method of Lagrange multipliers.
The resulting optimal allocation is
\begin{equation}
m_i^\ast = r \trialbudget \frac{\sqrt{p_i(1-p_i)}}{\sum_{j=1}^r \sqrt{p_j(1-p_j)}},
\qquad
i = 1,\,2\,\ldots,\,r.
\label{eq:oracleaided_solution}
\end{equation}
Since each process has a fixed number of trials $m_i^\ast$, independent of the experimental outcome of each trial, we call using these numbers of trials \textit{oracle-aided binomial sampling}.
\begin{myExample}[Oracle-aided allocations]
\label{ex:allocations}
\begin{enumerate}
\item[(a)] Let $p_1 = \varepsilon$ and $p_2 = 1/2$. Then the fractional oracle-aided allocations are
\[
\frac{m_1^*}{2\trialbudget} = \frac{\sqrt{\varepsilon(1-\varepsilon)}}
{\sqrt{\varepsilon(1-\varepsilon)}+\Frac{1}{2}},
\quad
\frac{m_2^*}{2\trialbudget} = \frac{\Frac{1}{2}}
{\sqrt{\varepsilon(1-\varepsilon)}+\Frac{1}{2}}.
\]
These are plotted as functions of $\varepsilon$ in Fig.~\ref{fig:oracle_allocations}(a).
\begin{figure}
\centering
\begin{tabular}{@{}cc@{}}
\includegraphics[width=0.48\linewidth]{oracle_allocations_1.pdf} &
\includegraphics[width=0.48\linewidth]{oracle_allocations_2.pdf} \\
{\small (a) Example~\ref{ex:allocations}(a)} &
{\small (b) Example~\ref{ex:allocations}(c) with $r=20$} \\
\end{tabular}
\caption{Oracle-aided optimal allocations in Examples~\ref{ex:allocations}(a) and~\ref{ex:allocations}(c).}
\label{fig:oracle_allocations}
\end{figure}
\item[(b)] Let $p_1 = p_2 = \cdots = p_{r-1} = \varepsilon$ and $p_r = 1/2$. Then the fractional oracle-aided allocations are
\begin{eqnarray*}
\frac{m_i^*}{r\trialbudget} & \!\!\!=\!\!\! & \frac{\sqrt{\varepsilon(1-\varepsilon)}}
{(r-1)\sqrt{\varepsilon(1-\varepsilon)}+\Frac{1}{2}},
\quad i {=} 1,\,2,\,\ldots,\,r{-}1, \\
\frac{m_r^*}{r\trialbudget} & \!\!\!=\!\!\! & \frac{\Frac{1}{2}}
{(r-1)\sqrt{\varepsilon(1-\varepsilon)}+\Frac{1}{2}}.
\end{eqnarray*}
\item[(c)] Let $p_i = (2i-1)/(2r)$, $i = 1,\,2,\,\ldots,\,r$.
The fractional oracle-aided allocations $m_i^*/(r \trialbudget)$ are plotted for $r = 20$ in Fig.~\ref{fig:oracle_allocations}(b).
\end{enumerate}
\end{myExample}
\subsection{Trial Allocation Gain}
\label{sec:trial-allocation-gain}
Using the oracle-aided allocations \eqref{eq:oracleaided_solution} reduces the average MSE relative to a constant allocation $m_1 = m_2 = \cdots = m_r = \trialbudget$.
The constant allocation results in the average MSE
\begin{equation}
\frac{1}{r} \sum_{i=1}^r \frac{p_i(1-p_i)}{\trialbudget},
\label{eq:mse_constant_allocation}
\end{equation}
whereas using \eqref{eq:oracleaided_solution} yields
\begin{eqnarray}
\frac{1}{r} \sum_{i=1}^r \frac{p_i(1-p_i)}{m_i^*}
& \!\!=\!\! & \frac{1}{r} \sum_{i=1}^r \frac{p_i(1-p_i)}{\sqrt{p_i(1-p_i)}} \frac{1}{r \trialbudget} {\sum_{j=1}^r \sqrt{p_j(1-p_j)}} \nonumber \\
& \!\!=\!\! & \frac{1}{r^2 \trialbudget} \Bigg( \sum_{j=1}^r \sqrt{p_j(1-p_j)} \Bigg)^2.
\label{eq:mse_optimal_allocation}
\end{eqnarray}
We define the ratio of \eqref{eq:mse_constant_allocation} and \eqref{eq:mse_optimal_allocation} as the \emph{trial allocation gain}:
\begin{equation}
\label{eq:allocation_gain_deterministic}
\allocationgain = \frac{r \sum_{i=1}^r p_i(1-p_i)}
{\left( \sum_{j=1}^r \sqrt{p_j(1-p_j)} \right)^2}.
\end{equation}
Trial allocation gain is reminiscent of the coding gain in transform coding~\cite{Segall:76,Goyal:01a}.
\begin{myExample}[Trial allocation gains]
\label{ex:gains}
\begin{enumerate}
\item[(a)] For the parameters in Example~\ref{ex:allocations}(a),
the trial allocation gain is plotted as a function of $\varepsilon$ in Fig.~\ref{fig:allocation_gains}.
Notice that in the limit of $\varepsilon \rightarrow 0$, all the trials are allocated to the nontrivial Bernoulli process,
doubling its number of trials, which halves the average MSE\@. Thus $\allocationgain \rightarrow 2$.
\begin{figure}
\centering
\begin{tabular}{@{}cc@{}}
\includegraphics[width=0.48\linewidth]{allocation_gain_1.pdf} &
\includegraphics[width=0.49\linewidth]{allocation_gain_2.pdf} \\
{\small (a) Example~\ref{ex:gains}(a)} &
{\small (b) Example~\ref{ex:gains}(c)} \\
\end{tabular}
\caption{Trial allocation gains in Examples~\ref{ex:gains}(a) and~\ref{ex:gains}(c).
}
\label{fig:allocation_gains}
\end{figure}
\item[(b)] For the parameters in Example~\ref{ex:allocations}(b), $\lim_{\varepsilon \rightarrow 0} \allocationgain = r$.
\item[(c)] For the parameters in Example~\ref{ex:allocations}(c), the trial allocation gain is plotted as a function of $r$ in Fig.~\ref{fig:allocation_gains}(b).
\item[(d)] Fig.~\ref{fig:Shepp-Logan}(a) shows the ``'Modified Shepp--Logan phantom'' provided by the Matlab {\tt phantom} command, at size $100 \times 100$ and scaled to $[0.001,\, 0.101]$.
Fig.~\ref{fig:Shepp-Logan}(b) shows a histogram of the $10^4$ intensity values of the phantom. Evaluating \eqref{eq:allocation_gain_deterministic} gives 1.6944, or 2.29 dB\@.
\end{enumerate}
\end{myExample}
One can show that $\allocationgain \in [1,\,r]$.
The upper bound is illustrated in part (b) of the example.
\begin{figure}
\centering
\begin{tabular}{@{}cc@{}}
\hspace{1mm}
\includegraphics[width=0.40\linewidth]{shepp-logan.pdf}&
\hspace{-6mm}
\includegraphics[width=0.58\linewidth]{phantom-hist.pdf} \\
\hspace{-1mm}
{\small (a) Shepp--Logan phantom} &
{\small (b) Histogram of phantom values} \\
\end{tabular}
\caption{Phantom of size $100 \times 100$ with trial allocation gain of 1.686
(Example~\ref{ex:gains}(d)).
}
\label{fig:Shepp-Logan}
\end{figure}
\subsection{Distributional Limit}
\label{sec:trial-allocation-distributional}
Suppose now that the Bernoulli process parameter is modeled with random variable $P$ and
the number of trials $M$ is to be assigned by an oracle (i.e., it is allowed to depend on the realization of $P$)
to minimize the MSE of the ML estimate of $P$ under the constraint $\E[M] \leq \trialbudget$.
By analogy to the computations giving \eqref{eq:oracleaided_solution}~--
or formally taking a limit of $r \rightarrow \infty$ with the empirical distribution of $\{p_i\}_{i=1}^r$ converging to the distribution of $P$~--
the number of trials should be assigned based on how large $\sqrt{P(1-P)}$ is relative to
$\E[\sqrt{P(1-P)}]$:
\begin{equation}
M = \trialbudget \, \frac{\sqrt{P(1-P)}}{\E\!\big[\sqrt{P(1-P)}\big]}.
\label{eq:trial-allocation-random}
\end{equation}
The resulting trial allocation gain is
\begin{equation}
\allocationgain = \frac{\E[P(1-P)]}{\left(\E\!\big[\sqrt{P(1-P)}\big]\right)^2}.
\label{eq:allocation-gain-random}
\end{equation}
This can also be written as
\begin{equation}
\allocationgain = 1 + \frac{\var{V}}{\left(\E[V]\right)^2},
\label{eq:allocation-gain-random-2}
\end{equation}
where $V = \sqrt{P(1-P)}$.
It follows that $\allocationgain \geq 1$, with equality if and only if the random variable $\sqrt{P(1-P)}$ has zero variance.
\begin{myExample}[Trial allocation gains~-- random parameter]
\label{ex:gains-random}
\begin{enumerate}
\item[(a)] Let $P$ have the continuous uniform distribution on $[0,\,1]$.
Then evaluating \eqref{eq:allocation-gain-random} gives $\allocationgain = 32/(3\pi^2) \approx 1.0808$.
This value is the asymptote in Fig.~\ref{fig:allocation_gains}(b).
\item[(b)] Let $P$ take two values: $\frac{1}{2}$ with probability $\delta$ and $0$ with probability $1-\delta$.
Then $\E[P(1-P)] = \delta/4$ and $\E[\sqrt{P(1-P)}] = \delta/2$.
Substituting in \eqref{eq:allocation-gain-random} gives $\allocationgain = 1/\delta$.
We can interpret this with relative frequencies: Since $p=0$ requires no trials,
fraction $\delta$ of the time, $p=1/2$ will occur and should be allocated $1/\delta$ times the mean number of trials.
\item[(c)] When $p \ll 1$ holds, $p(1-p) \approx p$.
Therefore, \eqref{eq:allocation-gain-random-2} becomes approximately invariant to rescaling.
For example, rescaling the phantom in Example~\ref{ex:gains}(d)
by a factor of $2$ to $[0.002,\,0.202]$
gives $\allocationgain \approx 1.6633$,
and by a factor of $\frac{1}{2}$ to $[0.0005,\,0.0505]$
gives $\allocationgain \approx 1.7096$;
these are small changes from the value in Example~\ref{ex:gains}(d).
\end{enumerate}
\end{myExample}
The first two parts of the example show that though an allocation gain may typically be modest, it may also be arbitrarily large.
The third part shows that allocation gain is approximately dependent on the coefficient of variation of the Bernoulli parameter,
provided that the parameter is known to be small.
Having established that varying the numbers of trials can be beneficial,
we now turn our attention to methods that do not depend on an oracle.
We will compare to the oracle-aided allocations in certain asymptotic settings.
\section{Observation of a Single Bernoulli Process}
\label{sec:single-bernoulli}
Let $\{X_n : n = 1,\,2,\,\ldots\}$ be a Bernoulli process with an unknown random parameter $p$,
and let $\trialbudget \in \mathbb{R}^+$ be a \emph{trial budget}.
A \emph{stopping rule} consists of a sequence of
\emph{continuation probability} functions
\begin{equation}
\pi_n: \{0,\,1\}^n \rightarrow [0,\,1], \qquad n = 0,\,1\,\ldots,
\label{eq:pi_t}
\end{equation}
that give the probability of continuing observations after trial $n$~--
based on a biased coin flip independent of the Bernoulli process~--
as a function of $(X_1,\,X_2,\,\ldots,\,X_n)$.
The result is a random number of observed trials $N$.%
\footnote{The time $N$ does not satisfy the standard definition of a stopping time when the stopping rule is randomized because randomness independent of the sequence of outcomes $\{X_n\}$ is allowed to influence the decision of whether or not to continue observations.}
The stopping rule is said to satisfy the trial budget when $\E[N] \leq \trialbudget$.
It is said to be \emph{deterministic} when every $\pi_n$ takes values only in $\{0,\,1\}$ and it is said to be \emph{randomized} otherwise.
A randomized stopping rule can be seen as stochastic multiplexing among some number of deterministic stopping rules.
Our goal is to minimize the MSE in estimation of $p$ through the design
of a stopping rule that satisfies the trial budget and
of an estimator $\widehat{p} \, (X_1,\,X_2,\,\ldots,\,X_N)$.
We will first show that the continuation probability functions can be simplified greatly with no loss of optimality.
Then, we will provide results on optimizing the stopping rule under a Beta prior on $p$.
\subsection{Framework for Data-Dependent Stopping}
Based on \eqref{eq:pi_t}, a natural representation of a stopping rule is a node-labeled binary tree representing all sample paths of the Bernoulli process,
with a probability of continuation label at each node.
This representation has
$2^{d+1}-1$
labels for observation sequences up to length
$d$.
However, the tree can be simplified to a trellis without loss of optimality.
Conditioned on observing $k$ successes in $m$ trials,
all $\binom{m}{k}$ sequences of length $m$ with $k$ successes are equally likely.
Thus,
no improvement can come from having unequal continuation probabilities for the $\binom{m}{k}$ tree nodes
that each represent having $k$ successes in $m$ trials.
Instead, these nodes should be combined, therefore reducing the tree to a trellis.
This representation has
$\frac{1}{2}(d+1)(d+2)$
labels for observation sequences up to length
$d$.
The continuation probability functions are reduced to a set of probabilities
$\{ q_{k,m} : m = 0,\,1\,\,\ldots;\ k = 0,\,1,\,\ldots,\,m \}$
for continuing after $k$ successes in $m$ trials,
as depicted in Fig.~\ref{fig:GeneralTrellis}.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{complete-trellis-depth-5.pdf}
\caption{A node-labeled trellis showing continuation probabilities for observation sequences up to length 5;
$q_{k,m}$ denotes the probability of continuing after observing $k$ successes in $m$ trials.
}
\label{fig:GeneralTrellis}
\end{figure}
Trellises alone~-- without labels~-- give a simple representation for both data-dependent and data-independent deterministic stopping rules:
Hence, we begin with some related terminology that will be used throughout this paper.
\begin{myDef}[Complete trellis]
A \emph{complete trellis} of depth $d \in \mathbb{N}$ contains all nodes $\bm{v} = (k,m)$ belonging to the set
\begin{equation}
\mathcal{T}_d = \{ (k,m): k = 0,1,\ldots,m; \ m = 0,1,\ldots,d \}.
\label{eq:trellis_set}
\end{equation}
\end{myDef}
\begin{myDef}[Strategy]
Any $T \in 2^{\mathcal{T}_d}$ is a \emph{strategy} when all nodes in $T$ are connected and $T$ contains the root node $(0,0)$.
\end{myDef}
\smallskip
Henceforth, we restrict our attention to strategies and stochastic multiplexing among strategies.
The stopping rule prescribed by the strategy $T$ is
\begin{equation}
q_{k,m}(T) =
\begin{cases}
1, \ \ \bm{v} = (k,m) \in T; \\
0, \ \ \text{otherwise.}
\end{cases}
\label{eq:q_km-T}
\end{equation}
\subsection{Standard Sampling Methods and their Representations}
\label{ssec:ExistingProtocols}
The conventional use of a fixed number of trials $n$ corresponds to continuation probabilities
\begin{equation}
q_{k,m} =
\begin{cases}
1, \ \ m < n; \\
0, \ \ \text{otherwise.}
\end{cases}
\end{equation}
Regardless of the sample path, one
observes exactly $n$ trials,
and the number of successes $K$ is a $\mathrm{Binomial}(n,p)$ random variable.
We refer to this as \emph{binomial sampling} or the \emph{binomial stopping rule}.
An example of the corresponding trellis representation for a fixed number of trials $n=5$ is shown in Fig.~\ref{fig:stopping_rules-bin-negbin}(a).
\begin{figure}
\begin{tabular}{@{}c@{}c@{}}
\includegraphics[width=0.57\linewidth,trim={2cm 0cm 0.2cm 0cm},clip]{trellis-bin-depth-5-v4.pdf}
&
\includegraphics[width=0.41\linewidth,trim={2cm 0cm 9.8cm 0cm},clip]{trellis-negbin-depth-5-v4.pdf}
\\
{\small (a) binomial} & {\small (b) negative binomial}
\end{tabular}
\caption{Green nodes form the trellis representations $T$ of (a) the binomial stopping rule with $n = 5$
and (b) the negative binomial stopping rule with $\ell = 2$;
these are nodes with continuation probability $1$.
Red nodes are not in $T$; these are nodes with continuation probability $0$.}
\label{fig:stopping_rules-bin-negbin}
\end{figure}
The technique analyzed by Haldane~\cite{Haldane1945} and employed in first-photon imaging~\cite{FPI2014} with $\ell = 1$ can be expressed with continuation probabilities
\begin{equation}
q_{k,m} =
\begin{cases}
1, \ \ k < \ell; \\
0, \ \ \text{otherwise.}
\end{cases}
\end{equation}
Observations cease with $\ell$ successes in $M$ trials, where
$M$ is a $\mathrm{NegativeBinomial}(\ell,p)$ random variable.
We call such a strategy the \emph{negative binomial stopping rule}, or \emph{geometric stopping rule} for the special case where $\ell = 1$.
The trellis representation of the negative binomial stopping rule for $\ell = 2$ is shown in Fig.~\ref{fig:stopping_rules-bin-negbin}(b).
In general, observations cease with $K$ successes in $M$ trials, where $K$ and $M$ are both random variables.
Importantly, the i.i.d.\ nature of a Bernoulli process makes the pair $(K,M)$ contain all the information that is relevant from the sequence of observations.
As noted in the reduction from tree to trellis, conditioned on $(K,M) = (k,m)$, all sequences of length $m$ with $k$ successes are equally likely, so the specific sequence among these is uninformative about $p$.
\subsection{Analysis Under Beta Prior}
\label{ssec:Beta-prior-anal}
Our method for optimizing the design of continuation probabilities is through analyzing mean Bayes risk reduction from continuation.
We define risk function $L$ as \emph{squared error} or \emph{squared loss}
\[
L(p, \widehat{p}) = (p-\widehat{p})^2,
\]
where $p$ is the unknown Bernoulli parameter and $\widehat{p}$ is the estimate of this parameter.
The Bayes risk $R$ is defined as
\[
R(\widehat{p}) = \E\!\left[L(p, \widehat{p})\right] = \E\!\left[(p-\widehat{p})^2\right],
\]
which in this case is the MSE\@.
Using the minimum MSE (MMSE) estimator, for which $\widehat{p} = \E[P]$, the Bayes risk
is the \emph{variance} of the posterior distribution.
Thus, key to the optimization is to track posterior variances through the trellis.
For any prior on $p$,
the posterior variance could be computed online or
precomputed for some fixed trellis.
Here we {provide detailed computations} only for the
convenient case of choosing a conjugate prior.
\subsubsection{{Beta Prior}}
The Beta distribution is the conjugate prior for Bernoulli, binomial, and negative binomial distributions:
When $P$ has the $\mathrm{Beta}(a,b)$ distribution with probability density function
\[
f_P(p;a,b) = \frac{\Gamma(a+b)}{\Gamma(a)\Gamma(b)} p^{a-1}(1-p)^{b-1},
\]
where $\Gamma(\cdot) = (\cdot \, - 1)!$ is the gamma function,
the posterior distribution after observing $k$ successes in $m$ trials has the $\mathrm{Beta}(a+k,b+m-k)$ distribution.
The beta distribution $P \sim \mathrm{Beta}(a,b)$
has mean
\begin{equation}
\label{eq:beta_mean}
\mu_{a,b} = \E[P] = \frac{a}{a+b}
\end{equation}
and variance
\begin{equation}
\Rbeta{a}{b} = \var{P} = \frac{ab}{(a+b)^2 (a+b+1)}.
\label{eq:beta_posterior}
\end{equation}
\subsubsection{Expected Number of Trials}
\label{sssec:ExpectedNumTrials}
For the stopping rule represented by the trellis $T$, the expected number of trials is the weighted sum of the depths of all stopping (or leaf) nodes with weights corresponding to probability of reaching that node, under the initial prior.
For a trellis $T \in 2^{\mathcal{T}_d}$ and initial prior $\mathrm{Beta}(\alpha,\beta)$, the probability of reaching any node $\bm{v} = (k,m) \in \mathcal{T}_d$
can be expressed recursively using the probabilities of reaching its parents, $(k-1, m-1)$ and $(k,m-1)$.
Conditioned on reaching $(k-1, m-1)$, the probability of reaching $(k,m)$ is the product of continuation probability $q_{k-1,m-1}(T)$
and success probability
\begin{subequations}
\begin{equation}
\mu_{\alpha+k-1,\beta+m-k} = \frac{\alpha + k - 1}{\alpha + \beta + m - 1};
\label{eq:prob_recursion_success}
\end{equation}
similarly, conditioned on reaching $(k,m-1)$,
the probability of reaching $(k,m)$ is the product of continuation probability $q_{k,m-1}(T)$ and
failure probability
\begin{equation}
1 - \mu_{\alpha+k,\beta+m-k-1} = \frac{\beta + m - k - 1}{\alpha + \beta + m - 1}.
\label{eq:prob_recursion_failure}
\end{equation}
\end{subequations}
Hence, we have the recursion
\begin{align}
u_{k,m}(T) =& u_{k{-}1,m{-}1}(T) q_{k{-}1,m{-}1}(T) \frac{\alpha + k - 1}{\alpha + \beta + m - 1} \nonumber \\
&{+} u_{k,m{-}1}(T) q_{k,m{-}1}(T) \frac{\beta + m - k - 1}{\alpha + \beta + m - 1}
\label{eq:prob_reaching_km}
\end{align}
for the probability $u_{k,m}(T)$ of reaching node $(k,m)$.
The recursion is initialized with $u_{0,0}(T) = 1$ and $u_{k,m}(T) = 0$ when $k \notin \{0,1,\ldots,m\}$.
Since $T \in 2^{\mathcal{T}_d}$, it suffices to compute up to $m = d+1$.
Using $u_{k,m}(T)$ from \eqref{eq:prob_reaching_km}, the expected number of trials incurred by a strategy $T \in 2^{\mathcal{T}_d}$ starting with a $\mathrm{Beta}(\alpha,\beta)$ prior is
\begin{equation}
{h}_{\alpha,\beta}(T) = \sum_{\bm{v} \in \mathcal{T}_{d+1} \setminus T} m \, u_{k,m}(T).
\label{eq:ExpNumTrials}
\end{equation}
The nonzero terms in the sum correspond to the reachable leaf nodes,
which are all contained in $T' = \mathcal{T}_{d+1} \setminus T$.
\subsubsection{Expected Bayes Risk}
\label{sssec:expected_bayes_risk}
Under initial prior $\mathrm{Beta}(\alpha,\beta)$,
the Bayes risk of the estimate of $p$ from observations leading to node $(k,m)$ is given by \eqref{eq:beta_posterior}, with $a = \alpha + k$ and $b = \beta + m - k$.
A strategy $T$ has expected Bayes risk $g_{\alpha,\beta}(T)$ given by the sum of the Bayes risks of nodes with zero continuation probability
weighted by the probabilities of reaching that node:
\begin{align}
g_{\alpha,\beta}&(T)
= \sum_{\bm{v} \in T'} u_{k,m}(T) \sigma^2_{\alpha+k,\beta+m-k} \nonumber \\
&= \sum_{\bm{v} \in T'} u_{k,m}(T) \frac{ (\alpha+k)(\beta+m-k)}{(\alpha+\beta+m)^2(\alpha+\beta+m+1)}.
\label{eq:EB_risk}
\end{align}
\subsubsection{Optimization Problem Statement}
With the proposed trellis-based framework,
finding an optimal deterministic stopping rule (in the MSE sense) under an average budget constraint becomes a set minimization problem:
\begin{equation}
\begin{split}
T^\ast = \, \argmin_{T \in 2^{\mathcal{T}_{d}}} g_{\alpha, \beta} (T) \\
\textrm{subject to } h_{\alpha,\beta} \left( T \right) \leq \trialbudget.
\end{split}
\label{eq:minimization-problem}
\end{equation}
Implementable solutions to \eqref{eq:minimization-problem}, with varying complexities and deviations from optimality, are presented in the subsequent section.
We seek only solutions on the lower convex hull of the trade-off between $\eta$ and $\min g_{\alpha,\beta}(T)$.
Stochastic multiplexing among these solutions gives optimal randomized stopping rules.
\section{Stopping Rule Design}
\subsection{A Dynamic Programming Solution}
\label{ssec:dynamic_prog}
For a fixed and sufficiently large $d$, total enumeration of the entire solution space is a possible approach for
solving \eqref{eq:minimization-problem} to find
an optimal deterministic stopping rule.
However, the combinatorial structure of the problem means that evaluating the Bayes risks
\eqref{eq:EB_risk}
and expected numbers of trials
\eqref{eq:ExpNumTrials}
for all possible strategies can be computationally prohibitive, even for moderate trial budgets;
this precludes full enumeration.
Conversely, one could start at the leaf nodes of a complete trellis (with depth $d$), traverse the trellis towards its root, whilst deciding whether each visited node merits inclusion in the optimized solution.
This is the basis of a dynamic programming (DP) solution:
it solves our optimization problem that involves making a sequence of decisions by determining, for each decision, subproblems that can be solved in a similar fashion \cite{bertsekas1996dynamic}.
As such, a solution of the original problem can be found from solutions of subproblems.
Precisely, we first relax \eqref{eq:minimization-problem} by writing its Lagrangian formulation:
\begin{equation}
\min_{T \in 2^{\mathcal{T}_{d}}} g_{\alpha,\beta}(T) + \lambda h_{\alpha,\beta}(T),
\label{eq:min_problem_relx}
\end{equation}
where $\lambda \in \mathbb{R}_{+}$, can be viewed as the desired MSE reduction per additional trial.
We introduce three compact notations associated with
node $\bm{v} = (k,m)$:
\begin{subequations}
\begin{equation}
\Risk{\,\alpha}{\beta}{k}{m}
= \sigma^2_{\alpha+k,\beta+m-k}
=\frac{ (\alpha+k)(\beta+m-k)}{(\alpha+\beta+m)^2(\alpha+\beta+m+1)}
\label{eq:Rstop_dp}
\end{equation}
is the mean Bayes risk conditioned on stopping at $\bm{v}$,
\begin{equation}
S_{k,m}^{\,\alpha,\beta}
= \mu_{\alpha+k,\beta+m-k}
= \frac{\alpha+k}{\alpha+\beta+m}
\label{eq:Skm}
\end{equation}
is the probability of the next trial being a success, and
\begin{equation}
F_{k,m}^{\,\alpha,\beta}
= 1 - \mu_{\alpha+k,\beta+m-k}
= \frac{\beta+m-k}{\alpha+\beta+m}
\label{eq:Fkm}
\end{equation}
\end{subequations}
is the probability of the next trial being a failure.
The dynamic program summarized in Algorithm~\ref{alg:dynamic-program} iteratively constructs a solution to \eqref{eq:min_problem_relx} by comparing
$\Risk{\,\alpha}{\beta}{k}{m}$ to the lowest cost achievable from the state that results after a single trial.
{More precisely, at each node $(k,m)$, $ \Risk{\,\alpha}{\beta}{k}{m}$ is compared to the cost of one additional trial plus the expected lowest cost achievable from the subsequent state. We keep track of the lowest of these values in $V_{k,m}$, which is the lowest achievable cost from any node $(k,m)$.}
If $\Risk{\,\alpha}{\beta}{k}{m}$ is lower, then an additional trial is not warranted and the node is eliminated, i.e. $T \leftarrow T\setminus \! \{\bm{v}\}$ and $q_{k,m} = 0$.
Because of the decomposability of the problem, the solutions are optimal.
\begin{algorithm}
\caption{Dynamic programming algorithm to find optimal deterministic stopping rule for $\mathrm{Beta}(\alpha,\beta)$ prior, Lagrange multiplier $\lambda$, and maximum depth $d$}
\begin{algorithmic}
\State \textbf{Input:} $(\alpha, \beta)$, $\lambda \in \mathbb{R}_{+}$, $d \in \mathbb{N}$
\State \textbf{Output:} $T^\ast \in 2^{\mathcal{T}_{d}}$
\State \textbf{Initialize:} $T = \mathcal{T}_{d}$ and $[V_{k,m}]_{k,m} = 0$ for all $k,m$
\For{$k = 1, \ldots, d$}
\State Set $V_{k,d} \leftarrow \Risk{\,\alpha}{\beta}{k}{d}$ using \eqref{eq:Rstop_dp}
\EndFor
\For{$m = d-1, ... , 1$}
\For{$k = 1, 2, ..., m$}
\If{$ \Risk{\,\alpha}{\beta}{k}{m} > \lambda + S_{k,m}^{\,\alpha,\beta} V_{k+1,m+1} + F_{k,m}^{\,\alpha,\beta} V_{k,m{+}1}$}
\State $V_{k,m} \gets \lambda + S_{k,m}^{\,\alpha,\beta} V_{k{+}1,m{+}1} + F_{k,m}^{\,\alpha,\beta} V_{k,m{+}1}$
\Else
\State $V_{k,m} \gets \Risk{\,\alpha}{\beta}{k}{m}$
\State $T \gets T\setminus \! \{\bm{v}\}$
\EndIf
\EndFor
\EndFor \\
\Return $T^\ast \gets T$
\end{algorithmic}
\label{alg:dynamic-program}
\end{algorithm}
\subsection{A Greedy Algorithm}
\label{ssec:greedy-offline}
The DP method (Algorithm~\ref{alg:dynamic-program}) prunes from the complete trellis $\mathcal{T}_d$.
Monotonicity of the objective $g_{\alpha,\beta}(T)$
and cost $h_{\alpha, \beta}(T)$
can be exploited to develop a lower-complexity greedy algorithm that instead builds a trellis starting from just the root node.
The scheme outlined in Algorithm~\ref{alg:greedy-offline} monotonically improves the objective function value for the minimization problem \eqref{eq:minimization-problem} with each iteration.
Specifically, at iteration $i$, the greedy decision is to add to the current trellis $T_i$ a node $\bm{v} \not\in T_i$ that yields the largest reduction in the Bayes risk \textit{per additional trial},
\begin{equation*}
\frac{g_{\alpha, \beta}(T_{i}\cup {\bm{v}}) - g_{\alpha, \beta} (T_{i})}{h_{\alpha, \beta} (T_{i}\cup {\bm{v}}) - h_{\alpha, \beta} (T_{i})},
\end{equation*}
without violating the mean number of trials constraint. The scheme terminates when no such node exists.
\begin{algorithm}
\caption{Greedy algorithm to find deterministic stopping rule for $\mathrm{Beta}(\alpha,\beta)$ prior and trial budget $\eta$}
\label{alg:greedy-offline}
\begin{algorithmic}
\State \textbf{Input:} $(\alpha, \beta)$, $\trialbudget$
\State \textbf{Output:} $T^\ast$
\State \textbf{Initialize:} $i \gets 0, \, T_0 \gets \{\}$
\Repeat
\State $\displaystyle \bm{\tilde{v}} \gets \argmin_{\bm{v} \not\in T_i} \,\frac{g_{\alpha, \beta}(T_{i}\cup {\bm{v}}) - g_{\alpha, \beta} (T_{i})}{h_{\alpha, \beta} (T_{i}\cup {\bm{v}}) - h_{\alpha, \beta} (T_{i})}$
\State $T_{i+1} \gets T_{i} \cup {\bm{\tilde{v}}}$
\State $i \gets i+1$
\Until{$h_{\alpha,\beta}(T_{i}) > \trialbudget$} \\
\Return $T^\ast \gets T_{i-1}$
\end{algorithmic}
\end{algorithm}
\subsection{Online Threshold-Based Termination}
\label{ssec:greedy-online}
Our final method applies a simple rule for termination of trials, depending on the prior parameters $(\alpha,\beta)$ and the $(k,m)$ position in the trellis.
It implies a trellis design, but it does not require storage of a designed trellis.
Suppose a sequence of trials reaches a node corresponding to the posterior distribution $\mathrm{Beta}(a,b)$.
Denote the mean Bayes risk \emph{without} performing an additional trial by
\begin{equation}
R_{\mathrm{stop}}(a,b) = \Rbeta{a}{b},
\label{eq:R_stop}
\end{equation}
using the variance given in \eqref{eq:beta_posterior}.
When one additional trial is performed, the posterior distribution is either $\mathrm{Beta}(a+1,b)$ if the outcome of the additional trial is a success,
or $\mathrm{Beta}(a,b+1)$ if the outcome of the additional trial is a failure.
Therefore, the mean Bayes risk resulting from continuing with one additional trial is
\begin{align}
R_{\mathrm{cont}}(a,b) =&\ \E\!\left[ (1-P) \, \Rbeta{a}{b+1} + P \, \Rbeta{a+1}{b} \right] \nonumber \\
=&\ \frac{ab}{(a+b)(a+b+1)^2}.
\end{align}
The Bayes risk \emph{reduction} from one additional trial is
\begin{align}
\Delta R(a,b) =&\ R_{\mathrm{stop}}(a,b) - R_{\mathrm{cont}}(a,b) \nonumber \\
=&\ \frac{ab}{(a+b)^2(a+b+1)^2}.
\end{align}
Recall that, starting from a $\mathrm{Beta}(\alpha,\beta)$ prior, upon reaching node $(k,m)$, the posterior is $\mathrm{Beta}(\alpha+k,\beta+m-k)$.
The Bayes risk reduction from an additional trial,
\begin{equation}
\label{eq:DeltaR-km}
\Delta R(k,m;\alpha,\beta) = \frac{(\alpha+k)(\beta+m-k)}{(\alpha+\beta+m)^2(\alpha+\beta+m+1)^2} ,
\end{equation}
can be the basis of an online stopping rule.
Let $\Delta_{\mathrm{min}} > 0$ denote a specified threshold value for the reduction in Bayes risk that justifies an additional trial.
Then stopping based on this threshold induces the probabilities of continuing at each node of the trellis given by
\begin{equation}
q_{k,m} =
\begin{cases}
1, \ \ \Delta R(k,m;\alpha,\beta) \geq \Delta_{\mathrm{min}}; \\
0, \ \ \Delta R(k,m;\alpha,\beta) < \Delta_{\mathrm{min}}.
\end{cases}
\label{eq:qkm-from-thresholding}
\end{equation}
Fig.~\ref{fig:example-trellis}(a) shows values of $\Delta R(k,m;1,1)$ for $m=0,1,\ldots,5$.
The choice of threshold $\Delta_{\mathrm{min}} = 0.005$ results in the trellis shown in
Fig.~\ref{fig:example-trellis}(b).
\begin{figure}
\begin{tabular}{@{}c@{}c@{}}
\includegraphics[height=4.2cm,keepaspectratio,trim={2cm 2cm 3cm 0cm},clip]{example-trellis-values-v3.pdf}
&
\includegraphics[height=4.2cm,keepaspectratio, trim={8cm 2cm 8cm 0cm},clip]{example-trellis-depth-5-v4.pdf} \\
{\small (a)} & {\small (b)}
\end{tabular}
\caption{(a) Bayes risk reductions per additional trial.
(b) Resulting trellis of continuation probabilities for $\Delta_{\mathrm{min}} = 0.005$ (right).
$\mathrm{Beta}(1,1)$ (i.e., uniform) prior for $P$ has been assumed.}
\label{fig:example-trellis}
\end{figure}
Notice that for a fixed trellis depth $m$, the denominator of \eqref{eq:DeltaR-km} is fixed, and the numerator of \eqref{eq:DeltaR-km} is a product of factors with fixed sum that is equal to $\alpha + \beta + m$.
Thus, from the arithmetic--geometric mean inequality, $\Delta R(k,m;\alpha,\beta)$ is largest where the posterior distribution is most symmetric.
This is apparent in the example in Fig. \ref{fig:example-trellis}(b);
since we have started with a uniform prior, the center of each row represents a symmetric posterior,
and additional observations are most merited near the center of each row.
Starting with a highly asymmetric prior ($\alpha \ll \beta$ or $\alpha \gg \beta$),
the same principle explains an asymmetry in the greedily optimized trellis of continuation probabilities.
\begin{myExample}[Suboptimality of binomial sampling]
\label{ex:suboptimality-binomial}
Suppose we have a $\mathrm{Beta}(1,1)$ (i.e., uniform) prior.
Then \eqref{eq:DeltaR-km} simplifies to
\begin{equation*}
\Delta R(k,m;1,1) = \frac{(k+1)(m-k+1)}{(m+2)^2(m+3)^2} .
\end{equation*}
For the threshold-based termination to induce binomial sampling with $m^*$ trials,
the incremental benefit $\Delta R$ at $(k,m) = (0,m^*)$ must be greater than
$\Delta R$ at $(k,m) = ( \lfloor\frac{1}{2}(m^*+1)\rfloor ,\,m^*+1)$:
\begin{align}
& \frac{m^*+1}
{(m^*+2)^2(m^*+3)^2} \nonumber \\
& \quad \geq
\frac{( \lfloor\frac{1}{2}(m^*+1)\rfloor +1)(m^*- \lfloor\frac{1}{2}(m^*+1)\rfloor +2)}
{(m^*+3)^2(m^*+4)^2}.
\label{eq:binomial-optimality-test}
\end{align}
Since \eqref{eq:binomial-optimality-test} fails to hold for any $m^* > 2$,
threshold-based termination induces binomial sampling only for $1$ and $2$ trials.
This is consistent with Fig.~\ref{fig:example-trellis}.
For such a small trial budget, full enumeration of stopping rules is also feasible,
and one can conclude that binomial sampling is indeed suboptimal for any trial budget
greater than 2\@.
Similar arguments can be made for non-uniform beta priors.
\end{myExample}
\subsection{Comparisons of Designs}
Sweeping $\Delta_{\mathrm{min}}$ in threshold-based termination is very similar to sweeping $\lambda$ in Algorithm~\ref{alg:dynamic-program};
it will achieve certain mean numbers of trials, similar to sweeping $\eta$ in Algorithm~\ref{alg:greedy-offline}.
Intermediate values of the mean number of trials can be achieved by finding $(k^*,m^*)$ such that $\Delta R(k,m;\alpha,\beta)$ is largest among those below $\Delta_{\mathrm{min}}$ and varying $q_{k^*,m^*}$ over $(0,1)$.
This idea is used to enforce an equal expected number of trials for trellises optimized with each method, thus allowing a fair comparison of their Bayes risks.
For a mean number of trials $\approx 95.36$, Algorithms~\ref{alg:dynamic-program} and~\ref{alg:greedy-offline} were found to give exactly the same trellis,
while online threshold-based termination gave a slightly different trellis with slightly higher mean Bayes risk.
Fig.~\ref{fig:dp-greedy-diff} illustrates the difference in $q_{k,m}$ values.
It is zero for the vast majority of nodes,
with 24 nodes at which the DP-designed trellis terminates but the threshold-based rule does not (red, $-1$),
and 54 nodes at which the threshold-based rule terminates but the DP-designed trellis does not (blue, $+1$).\footnote{The mean number of trials is equal. To be convinced that the blue and red nodes can balance, note that while there are more blue nodes, they are for larger values of $m$ and thus have lower probabilities of being reached.}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{dp-greedy-diff.pdf}
\caption{Dynamic programming solution minus online threshold-based termination result, both with mean number of trials $\approx 95.36$.
Online threshold-based termination achieves mean Bayes risk of $0.0016037$ whereas DP gives $0.0016036$.
At 24 nodes (red, `-1'), threshold-based rule performs additional trials and DP does not;
at 54 nodes (blue, `+1'), DP performs trials and threshold-based rule does not.
The greedily designed trellis coincides with the DP trellis, hence their difference plot is omitted.}
\label{fig:dp-greedy-diff}
\end{figure}
Illustrated in Fig.~\ref{fig:dynamic-vs-greedy} is a comparison of our three proposed implementable strategies,
applied for a uniform prior, over a range of trial budgets.
MSEs of DP
(Algorithm~\ref{alg:dynamic-program})
and the greedily optimized trellis
(Algorithm~\ref{alg:greedy-offline})
coincide for all trial budgets
because the trellises are identical~-- though we have not proven that this is guaranteed.
The online threshold-based stopping rule is only very slightly worse by a factor of at most 1.000195 (less than 0.001 dB).
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{dp-vs-greedy.pdf}
\caption{Comparison between proposed strategies.
DP and greedy algorithm solutions coincide for all trial budgets,
while threshold-based termination is only very slightly worse by a factor of at most 1.000195.
}
\label{fig:dynamic-vs-greedy}
\end{figure}
The phenomenon of more trials being merited when $p$ is near $\frac{1}{2}$ counteracts the MSE of $p(1-p)/n$ being largest for $p$ near $\frac{1}{2}$.
This is illustrated in Fig.~\ref{fig:bin_greedy_oracle-EN_ER-comp}(a), which shows mean numbers of trials allocated as a function of $p$.
We have optimized for MSE averaged over $p$ and, in so doing, obtained a modest improvement factor of $\approx 1.05$ in this average, comparing the online threshold-based termination to conventional binomial sampling.
A more significant reduction in the worst-case MSE is a by-product of the optimization (see Fig.~\ref{fig:bin_greedy_oracle-EN_ER-comp}(b)).
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{three-comparison-EN.pdf} \\
{\small (a)} \\
\includegraphics[width=0.8\linewidth]{three-comparison-ER.pdf} \\
{\small (b)}
\caption{Dependences on $p$ under a uniform prior.
(a) Conditional expectation of the number of trials, conditioned on $p$,
for binomial sampling, online threshold-based termination, and the oracle-aided binomial stopping rule.
Expected number of trials is 200 for all of the three methods.
(b) Dependence of conditional MSEs on the true Bernoulli parameter $p$ under a trial budget of $\trialbudget=200$.
}
\label{fig:bin_greedy_oracle-EN_ER-comp}
\end{figure}
\subsection{Asymptotic Comparison with Oracle-Aided Allocation}
\label{ssec:comparison_oracle}
Considering the non-degenerate cases $p \in (0,1)$, the threshold-based termination is asymptotically equivalent to oracle-aided optimal allocation.
For a large trial budget $\eta$, we will find an approximation for $m^\circ$,
the number of trials at which the online threshold-based rule terminates.
This will match the form of
\eqref{eq:oracleaided_solution}
or \eqref{eq:trial-allocation-random}.
Using \eqref{eq:DeltaR-km}, for an initial $\mathrm{Beta}(\alpha,\beta)$ prior,
the online rule continues at node $(k,m)$ if and only if
\begin{equation}
\frac{(\alpha+k)(\beta + m - k)}{(\alpha+\beta+m)^2 (\alpha+\beta+m+1)^2} \leq \Delta_{\mathrm{min}}.
\label{eq:BR_reduction_comparison}
\end{equation}
Since the trial budget is large and $p \in (0,1)$,
$k$, $m$, and $m-k$ are all large when nearing termination.
Hence, we approximate the expression in \eqref{eq:BR_reduction_comparison} as
\begin{align}
& \frac{(\alpha+k)(\beta + m - k)}
{(\alpha+\beta+m)^2 (\alpha+\beta+m+1)^2} \nonumber \\
& \quad
= \frac{m^2 (\alpha/m+k/m) (\beta/m + 1 - k/m)}
{m^4 (\alpha/m+\beta/m+1)^2 (\alpha/m+\beta/m+1+1/m)^2} \nonumber \\
& \quad
\approx \frac{(k/m) (1 - k/m)}
{m^2} \nonumber \\
& \quad
= \frac{ \widehat{p}_{\rm ML} (1 - \widehat{p}_{\rm ML}) }
{m^2},
\label{eq:asymptotic-equivalence-approx}
\end{align}
where $\widehat{p}_{\rm ML} = k/m$ is the ML estimate of $p$.
Substituting \eqref{eq:asymptotic-equivalence-approx}
into \eqref{eq:BR_reduction_comparison},
we obtain
\begin{equation}
m^\circ \approx \sqrt{
\frac{\widehat{p}_{\rm ML} (1 - \widehat{p}_{\rm ML})}{\Delta_{\mathrm{min}}}
}.
\label{eq:mcirc}
\end{equation}
By the law of large numbers, $\widehat{p}_{\rm ML} \rightarrow p$,
so \eqref{eq:mcirc} shows a match to
\eqref{eq:oracleaided_solution},
with $\Delta_{\mathrm{min}}$ determining the trial budget.
Furthermore, by comparison with \eqref{eq:trial-allocation-random},
we see an equivalence by choosing
$\Delta_{\mathrm{min}} = \big({\E\!\big[\sqrt{P(1-P)}\big]} / \trialbudget\big)^2$.
Fig.~\ref{fig:bin_greedy_oracle-EN_ER-comp}(a) illustrates an example of the approximate match
between threshold-based termination and oracle-aided sampling that is predicted by
the match among
\eqref{eq:oracleaided_solution},
\eqref{eq:trial-allocation-random}, and
\eqref{eq:mcirc}.
Note that convergence is not uniform in $p$;
a larger trial budget is needed to observe approximate equivalence in allocations
for $p$ near $0$ and near $1$.
Fig.~\ref{fig:coding-gain-Beta-1-x} shows the variation of the MSEs with mean number of trials budget constraint for conventional binomial sampling,
threshold-based termination,
and oracle-aided allocation.
The results are based on Monte Carlo simulations, with MATLAB, using the phantom image in Fig.~\ref{fig:Shepp-Logan}(a).
As expected the optimized rules consistently achieve MSE improvements over the conventional binomial sampling, for all simulated trial budgets.
In addition, when compared to the unrealizable oracle-aided method, the threshold-based approach only marginally under-performs at moderate mean number of trials budget constraints.
This observation is further underscored in Fig.~\ref{fig:bin_greedy_oracle-EN_ER-comp}, which shows significant overlap between threshold-based termination
and oracle-aided allocation, in terms of both trial allocations and the resulting MSEs. {Using {a} negative binomial sampling strategy yields significantly worse performance than binomial sampling and our proposed rules for estimating $p$; thus, we have omitted it from Figs.~\ref{fig:bin_greedy_oracle-EN_ER-comp} and~\ref{fig:coding-gain-Beta-1-x}, as well as other numerical simulations related to the estimation of $p$.}
\begin{figure}
\centering
\vspace{0.04cm}
\includegraphics[width=0.8\linewidth]{coding-gain-1-1.pdf} \\[0.03cm]
{\small (a)} \\[0.25cm]
\includegraphics[width=0.8\linewidth]{coding-gain-1-50.pdf} \\
{\small (b)} \vspace{0.05cm}
\caption{Results for the Shepp--Logan Phantom scaled to [0.001,\, 0.101], assuming: (a) $\mathrm{Beta}(1,1)$ and (b) $\mathrm{Beta}(1,50)$, for the online threshold-based termination.
The MSE has been computed from the average of 20 independent experiments for each mean number of trials.
Improvements are consistent with the trial allocation gain computed in
Example~\ref{ex:gains}(d).
}
\label{fig:coding-gain-Beta-1-x}
\end{figure}
In line with the earlier asymptotic analysis, the threshold-based termination and oracle-aided performances coincide for moderate
to high mean numbers of trials,
independent of the prior.
Under a highly skewed prior consistent with the true distribution of the phantom pixels,
Fig.~\ref{fig:coding-gain-Beta-1-x}(b) demonstrates that
it is possible for online threshold-based termination
to even outperform the oracle-aided binomial method, at low trial budgets.
This phenomenon is attributable to the
online method allocating more trials when the Bernoulli process realization has
a relatively high fraction of successes.
Put simply, it is allocating more trials for ``unlucky'' realizations where the MSE would be higher,
while the oracle-aided binomial method maintains a fixed number of trials.
\section{Estimating Functions of a Bernoulli Parameter}
\label{sec:estimating-functions}
When estimating an arbitrary function $f(p)$ of a Bernoulli parameter is of interest \cite{Hubert2000}, one can derive similar stopping strategies as before. In this section, we concern ourselves only with the estimation of $f(p) = \log p$ due to its prevalence in real-life scenarios. For instance, the subjective brightness perceived by the human vision system is a logarithmic function of the incident light intensity~\cite{Gonzalez2006,Milner2017}.
Also, the common \emph{log odds ratio} $\log( p/(1-p) )$, is approximately equal to $\log p$ when $p \ll 1$.
As before, we begin with a squared error loss
\begin{equation}
L(p,\cdot) = \big(f(p) - \hat{f}(\cdot)\big)^2,
\label{eq:loss-log}
\end{equation}
where $\hat{f}(\cdot)$ is the estimate of $f(p)$.
The expectation of this loss function over $p$ gives the Bayes risk
\begin{equation}
R(\cdot) = \E[L(p,\cdot)] = \E\!\big[\big(f(p) - \hat{f}(\cdot)\big)^2\big].
\label{eq:log-bayes-risk}
\end{equation}
For $f(p) = \log{p}$, suppose a sequence of trials leads to a node in the trellis corresponding to the posterior distribution $\mathrm{Beta}(a,b)$.
Under $P \sim \mathrm{Beta}(a,b)$, the MMSE estimator of $\log p$ is \cite{Karlis2005}
\begin{equation}
\hat{f}(a,b) := \E\!\left[\log P\right] = \psi^{(0)}(a) - \psi^{(0)}(a+b),
\end{equation}
where $\psi^{(m)}$ is the polygamma function of order $m$.
The Bayes risk \eqref{eq:log-bayes-risk} when no additional trial is performed, $R_{\mathrm{stop}}$, becomes the variance of $\log p$ \cite{Dette2017}:
\begin{equation}
R_{\mathrm{stop}} \stackrel{\text{\tiny{def}}}{=} R(a,b) = \var{\log P} = \psi^{(1)}(a) - \psi^{(1)}(a+b).
\end{equation}
If one additional trial is performed, the Bayes risk reduces to
\begin{align}
R_{\mathrm{cont}}(a,b) =&\ \E\!\left[ (1-P) \, R(a,b+1) + P \, R(a+1,b) \right] \nonumber \\
=&\ \frac{b}{a+b} \, R(a,b+1) + \frac{a}{a+b} \, R(a+1,b) .
\end{align}
Hence, the Bayes risk reduction from one additional trial is
\begin{equation}
R_{\mathrm{stop}}(a,b) - R_{\mathrm{cont}}(a,b)
= \frac{b}{a \, (a+b)^2}.
\label{eq:log-deltaBR}
\end{equation}
Starting with prior $\mathrm{Beta}(\alpha,\beta)$,
the counterpart to \eqref{eq:DeltaR-km} for estimation of $\log p$ is
\begin{equation}
\label{eq:log-DeltaR-km}
\Delta R(k,m;\alpha,\beta)
= \frac{\beta + m -k }
{(\alpha + k)(\alpha+\beta+m)^2} .
\end{equation}
As before, this can be used in
\eqref{eq:qkm-from-thresholding}
as an online threshold-based termination method.
The Bayes risk reductions for both $f(p)=p$
in \eqref{eq:DeltaR-km}
and $f(p) = \log p$
in \eqref{eq:log-DeltaR-km}, starting with a uniform prior,
are shown as heat maps in Fig.~\ref{fig:p-p-logp-trellises}.
When $f(p) = \log p$, the reduction from additional trials after observing sequences with low number of successes is significantly larger.
Thus, the online threshold-based termination of Section~\ref{ssec:greedy-online} is likely to assign more trials for the smaller underlying Bernoulli parameters.
Such a stopping rule is intuitive because a fixed amount of estimation error for $p$ would contribute more to the loss function defined in \eqref{eq:loss-log}, with $f(p) = \log p$, when $p$ is small.
In fact, one can choose a loss function that enforces different penalties for different $p$ values.
An example is the family of weighted mean squared errors,
$\E[w(p)(p-\hat{p})^2]$, where a weighting function $w(p)$ is designed according to the problem.
A special case is relative MSE $\E[(p-\hat{p})^2/p^2]$,
which is approximately the squared loss in \eqref{eq:log-bayes-risk} with $f(p) = \log p$ for estimates sufficiently close to the true value~\cite{Hubert2000}.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{p-trellis.pdf} \\
{\small (a)} \\
\includegraphics[width=0.7\linewidth]{logp-trellis.pdf} \\
{\small (b)}
\caption{Bayes risk reductions for estimation of (a) $f(p) = p$ and (b) $f(p) = \log{p}$, assuming $\mathrm{Beta}(1,1)$ assumed initially.
Low $p$ values are assigned significantly more trials when $\log{p}$ is estimated.}
\label{fig:p-p-logp-trellises}
\end{figure}
When smaller $p$ values are of more importance, it makes sense to use a strategy that allocates more trials to these instances.
Negative binomial sampling explained in Section~\ref{ssec:ExistingProtocols} achieves this type of trial allocation.
For the estimation of $f(p) = \log p$, we compare the performances of binomial sampling, negative binomial sampling, and online threshold-based termination in Fig.~\ref{fig:log-estimation}.
Threshold-based termination outperforms both binomial and negative binomial sampling, with improvement factors of $2.037$ and $1.491$, respectively.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{bin-vs-adaptive-log.pdf}
\caption{Estimation of $\log p$.
Conditional MSE of $\log p$, conditioned on $p$, is shown as a function of $p$ for binomial sampling ($\eta = 72$),
negative binomial sampling ($\ell = 10$, inducing $\eta = 71.66$), and
online threshold-based termination ($\eta = 71.77$).
The threshold-based termination uses a uniform prior,
and averaging over a uniform prior gives
MSE of 0.0668 for binomial sampling,
0.0489 for negative binomial sampling,
and 0.0328 for threshold-based termination\@.
{These values are computed directly from their corresponding trellises (hence exact), not through numerical integration.}
}
\label{fig:log-estimation}
\end{figure}
\section{Applications to Active Imaging}
\label{sec:applications}
Active imaging systems typically raster scan the scene by probing patch $(i,j)$, $i=1,\ldots,N_i$ and $j=1,\ldots,N_j$, using pulsed illumination.
The measured data~-- used to form an image of the scene~-- are arrays $[k_{i,j}]_{i,j}$ and $[m_{i,j}]_{i,j}$;
i.e., the number of detections (successes) and number of illumination pulses (trials) for each scene patch.
Note that the conventional approach of a fixed number of trials makes $m_{i,j}= \trialbudget$ for all $(i,j)$ and $\{k_{i,j}\}$ random, whereas both $\{k_{i,j}\}$ and $\{m_{i,j}\}$ are random when the proposed approach is applied.
The parameters of the Bernoulli processes generated by probing a scene patch and its neighbors are typically correlated.
This can be exploited in the image formation stage through mechanisms inspired by any of various image compression or denoising methods.
For this initial demonstration of adaptive acquisition, we apply total variation (TV) regularization \cite{louchet2008}.
We present simulation results using the Shepp--Logan phantom in Fig.~\ref{fig:Shepp-Logan}(a),
two lidar {datasets provided by the Alaska Department of Natural Resources~\cite{AlaskaLidar}, and scanning electron microscopy (SEM) images
{\emph{Foraminifera}\footnote{Quanta SEM image of Protozoan group secreting a calcareous shell
by Philippe Crassous,
\url{https://www.fei.com/image-gallery/Foraminifera-Protozoan/}}
and \emph{HairStyle}\footnote{Quanta SEM image of the upper part of the style and stigma from an \emph{Arabidopsis} flower by Guichuan Hou,
\url{https://www.fei.com/uploadedImages/FEISite/Content/Image_Gallery/Images/2013_Image_Contest/FEI/IM_20130718_Hou_18_HairStyle_lg.jpg}}}
taken from ThermoFisher {Scientific}.
All images have been rescaled to take on values in the range $[0.001,0.101]$.}
\subsection{Estimation of $f(p) = p$}
\label{sec:image_p}
We focus here on comparing conventional binomial sampling against online threshold-based termination
\eqref{eq:qkm-from-thresholding}
applied for each pixel.
For $f(p) = p$, the relevant Bayes risk reduction per trial is given by \eqref{eq:DeltaR-km}.
\subsubsection{MMSE Estimation Under I.I.D. Prior}
\label{sec:pixelwise-mmse-images}
When not exploiting any spatial correlations, each pixel estimation is performed separately using the methods of Section~\ref{ssec:greedy-online}.
Under a $\mathrm{Beta}(\alpha,\beta)$ prior, the MMSE estimate is
$\widehat{p}_{\rm MMSE}{[i,j]} = \Frac{(k_{i,j} + \alpha)}{(m_{i,j} + \alpha + \beta)}$.
Fig.~\ref{fig:phantom-mmse}(a) shows that with the choice of the $\mathrm{Beta}(2,152)$ prior,
MSE improvement of $2.42 \, \mathrm{dB}$ is attained for trial budget $\eta = 200$ for the Shepp--Logan phantom. In Figs.~\ref{fig:phantom-mmse}(b)--(e), MSE improvements were also demonstrated for the lidar and SEM images under various trial budgets and initial prior parameters. Improvements ranging from $0.92~\mathrm{dB}$ for lidar \#1 to
$2.02~\mathrm{dB}$ for \emph{Foraminifera} were obtained.
For the same corresponding choices of prior,
we also perform $100$ independent experiments
for each test image
at each of
two different trial budgets;
the results
indicated in Table~\ref{tbl:Average_MSE_p} show similar improvement factors.
In particular, the performance gains are close to the prediction from the trial allocation gain
(Example~\ref{ex:gains}(d) for the Shepp--Logan phantom image). The same is observed for other images.
Furthermore, we show in Fig.~\ref{fig:mse-2-beta} that significant MSE improvements can be attained for a large range of Beta priors,
using the Shepp--Logan phantom and \emph{HairStyle} datasets.
We also observe that the performance of binomial sampling
is degraded more by a mismatched prior than the performance of threshold-based termination.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth, trim={8mm 0mm 7mm 2mm},clip]{Phantom_MMSE.pdf} \\[-3mm] \vspace{-0.55cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 20.0 dB \hspace*{5.5mm} PSNR 22.4 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(a) Shepp--Logan phantom, $\mathrm{Beta}{(2,152)}$ and $\eta = 200$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{lidar1_MMSE.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 21.4 dB \hspace*{5.5mm} PSNR 22.3 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(b) lidar \#1, $\mathrm{Beta}{(2,162)}$ and $\eta = 800$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{lidar2_MMSE.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 20.6 dB \hspace*{5.5mm} PSNR 22.3 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(c) lidar \#2, $\mathrm{Beta}{(2,172)}$ and $\eta = 800$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{cell1_MMSE.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 20.4 dB \hspace*{5.5mm} PSNR 22.4 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(d) \emph{Foraminifera}, $\mathrm{Beta}{(2,162)}$ and $\eta = 500$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{cell3_MMSE.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 21.2 dB \hspace*{5.5mm} PSNR 22.7 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(e) \emph{HairStyle}, $\mathrm{Beta}{(2,162)}$ and $\eta = 400$.}
\caption{Images reconstructed through pixelwise MMSE estimation showing MSE improvements of up to $2.42\,\mathrm{dB}$ for online threshold-based termination in place of conventional binomial sampling. Assumed priors and trial budgets are indicated below each test image. All images are scaled to $[0.001,0.101]$.
}
\label{fig:phantom-mmse}
\end{figure}
\begin{table}
\centering
\caption{Average reconstruction PSNRs, averaged over 100 experiments,
for conventional binomial sampling and online threshold-based termination, for trial budgets $\eta$.}
\label{tbl:Average_MSE_p}
\begin{tabular}{@{}cccccc@{}}
\hline
\multirow{2}{*}{Image} & \multirow{2}{*}{$\eta$} & \multicolumn{2}{@{}c@{}}{\scriptsize Pixelwise MMSE estimation}
& \multicolumn{2}{@{}c@{}}{\scriptsize TV+ML estimation}
\\ \cline{3-6}
&
& {\scriptsize Binomial}
& {\scriptsize Threshold-based}
& {\scriptsize Binomial}
& {\scriptsize Threshold-based} \\ \hline
\multirow{2}{*}{\scriptsize Shepp--Logan} & $100$ & $17.5$ dB & $\mathbf{20.0}$ $\mathbf{dB}$ & $21.9$ dB & $\mathbf{26.3}$ $\mathbf{dB}$ \\
& $200$ & $19.9$ dB & $\mathbf{22.1}$ $\mathbf{dB}$ & $24.9$ dB & $\mathbf{28.5}$ $\mathbf{dB}$ \\ \hline
\multirow{2}{*}{\scriptsize lidar \#1} & $400$ & $18.1$ dB & $\mathbf{18.8}$ $\mathbf{dB}$ & $24.0$ dB & $\mathbf{25.9}$ $\mathbf{dB}$ \\
& $800$ & $21.4$ dB & $\mathbf{22.3}$ $\mathbf{dB}$ & $25.2$ dB & $\mathbf{27.7}$ $\mathbf{dB}$ \\ \hline
\multirow{2}{*}{\scriptsize lidar \#2} & $400$ & $16.9$ dB & $\mathbf{18.3}$ $\mathbf{dB}$ & $26.2$ dB & $\mathbf{26.8}$ $\mathbf{dB}$ \\
& $800$ & $20.7$ dB & $\mathbf{22.3}$ $\mathbf{dB}$ & $27.6$ dB & $\mathbf{28.8}$ $\mathbf{dB}$ \\ \hline
\multirow{2}{*}{\scriptsize \emph{Foraminifera}} & $400$ & $19.2$ dB & $\mathbf{21.2}$ $\mathbf{dB}$ & $25.5$ dB & $\mathbf{26.7}$ $\mathbf{dB}$ \\
& $500$ & $20.4$ dB & $\mathbf{22.4}$ $\mathbf{dB}$ & $23.5$ dB & $\mathbf{27.5}$ $\mathbf{dB}$ \\ \hline
\multirow{2}{*}{\scriptsize \emph{HairStyle}} & $400$ & $21.2$ dB & $\mathbf{22.7}$ $\mathbf{dB}$ & $26.8$ dB & $\mathbf{27.8}$ $\mathbf{dB}$ \\
& $500$ & $22.3$ dB & $\mathbf{23.8}$ $\mathbf{dB}$ & $27.8$ dB & $\mathbf{28.5}$ $\mathbf{dB}$ \\ \hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{mse-plot-2-beta.pdf}
\centerline{\footnotesize (a) Shepp--Logan phantom, $\eta=200$}
\medskip
\includegraphics[width=0.8\linewidth]{mse-plot-2-beta-cell3.pdf}
\centerline{\footnotesize (b) \emph{HairStyle}, $\eta=400$
}
\caption{Dependence of MSE on $\beta$
when $\mathrm{Beta}(2,\beta)$ prior is assumed and pixelwise MMSE estimation is performed; each plotted point is obtained by averaging over $100$ independent experiments.
}
\label{fig:mse-2-beta}
\end{figure}
\subsubsection{TV-Regularized ML Estimation}
\label{sec:tv-images}
Reconstruction quality can be improved through the use of TV-regularized ML estimation~\cite{louchet2008,Shin2015}.
In one typical experimental trial shown in Fig.~\ref{fig:phantom-mmse-TV}(a),
the TV-regularized reconstruction from data obtained with online threshold-based termination outperforms the conventional binomial sampling by $4.36\,\mathrm{dB}$ in MSE for the Shepp--Logan phantom;
the trial budget of $\eta = 200$ and prior of $\mathrm{Beta}(2,152)$ are the same as used previously. As anticipated, Figs.~\ref{fig:phantom-mmse-TV}(b)--(e) also demonstrate significant improvements in MSE, ranging from $1.15\,\mathrm{dB}$ to $4.17\,\mathrm{dB}$, for the remaining test images.
Furthermore, keeping the corresponding priors and trial budgets used for each test image in Figs.~\ref{fig:phantom-mmse-TV}, Table~\ref{tbl:Average_MSE_p} also provides results averaged over 100 experiments for statistical significance.
In many cases, imposing TV regularization increases the performance gained from the data-adaptive stopping rule. Most importantly, it does not completely diminish the gain of adapting the acquisition.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth, trim={8mm 0mm 7mm 2mm},clip]{Phantom_reg_ML.pdf} \\[-3mm] \vspace{-0.55cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 24.7 dB \hspace*{5.5mm} PSNR 29.1 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(a) $\mathrm{Beta}{(2,152)}$ and $\eta = 200$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{lidar1_TV.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 25.2 dB \hspace*{5.5mm} PSNR 27.7 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(b) $\mathrm{Beta}{(2,162)}$ and $\eta = 800$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{lidar2_TV.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 27.5 dB \hspace*{5.5mm} PSNR 28.9 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(c) $\mathrm{Beta}{(2,172)}$ and $\eta = 800$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{cell1_TV.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 23.4 dB \hspace*{5.5mm} PSNR 27.6 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(d) $\mathrm{Beta}{(2,162)}$ and $\eta = 500$.} \vspace{0.1cm}
\includegraphics[width=0.9\linewidth, trim={8mm 11mm 7mm 8mm},clip]{cell3_TV.pdf} \\[-3mm] \vspace{-0.05cm}
{\footnotesize\begin{flushleft} \hspace*{31.7mm} PSNR 26.7 dB \hspace*{5.5mm} PSNR 27.9 dB \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(e) $\mathrm{Beta}{(2,162)}$ and $\eta = 400$.}
\caption{Images reconstructed through TV-regularized ML estimation showing MSE improvements of up to $4.36\,\mathrm{dB}$ for online threshold-based termination in place of conventional binomial sampling. Assumed priors and trial budgets are indicated below each test image. All images are scaled to $[0.001,0.101]$.
}
\label{fig:phantom-mmse-TV}
\end{figure}
\subsection{Estimation of $f(p) = \log{p}$}
\label{sec:image_log_p}
Now we present simulation results for the estimation of the logarithm of the previous test images.
For $f(p) = \log p$, the Bayes risk reduction per trial to use in online threshold-based termination
\eqref{eq:qkm-from-thresholding}
is given by
\eqref{eq:log-DeltaR-km}.
Fig.~\ref{fig:phantom-mmse-log} shows simulation results wherein
improvement factors of
$1.48\,\mathrm{dB}$ to $1.86\,\mathrm{dB}$ are observed using threshold-based termination compared with binomial sampling,
and $2.56\,\mathrm{dB}$ to $3.78\,\mathrm{dB}$ when using threshold-based termination over negative binomial.
Note that since the contributions to the error from the lower pixel values are higher,
the comparison is provided at much higher trial budgets (i.e., $\eta = 3000$, {$1600$, $1700$}, $2800$, and $1800$ for Figs.~\ref{fig:phantom-mmse-log}(a)--(e), respectively) than in Section~\ref{sec:image_p} to obtain meaningful results.
Dark regions having values of $0.001$, for instance, require $1000$ trials on average to observe a success.
The effect of increasing number of trials is apparent in Table~\ref{tbl:Average_MSE_log}.
An increase in the improvement factor obtained by using the threshold-based rule, compared to both binomial and negative binomial stopping is observed, as the trial budget is increased.
For example, for {the} Shepp--Logan phantom, an increase in the improvement factor from $0.77\,\mathrm{dB}$ to $1.75\,\mathrm{dB}$ is observed when increasing $\eta$ from $1800$ to $3000$, if the threshold-based rule is used over conventional binomial stopping. An even larger increase is obtained when the threshold-based rule is compared against negative binomial stopping. The trend of increasing MSE improvements with increased trial budgets is persistent for other test images too.
\begin{figure}
\centering
\includegraphics[width=1\linewidth, trim={8mm 18mm 8mm 8mm},clip]{phantom_log.pdf} \\[-3mm] \vspace{-0.1cm}
{\footnotesize\begin{flushleft} \hspace*{22.5mm} MSE 0.2152 \hspace*{4.2mm} MSE 0.2526 \hspace*{4.2mm} MSE 0.1401 \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(a) $\eta = 3000$} \vspace{0.1cm}
\includegraphics[width=1\linewidth, trim={8mm 18mm 8mm 15mm},clip]{lidar1_log.pdf} \\[-3mm] \vspace{-0.1cm}
{\footnotesize\begin{flushleft} \hspace*{22.5mm} MSE 0.1567 \hspace*{4.2mm} MSE 0.2508 \hspace*{4.2mm} MSE 0.1115 \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(b) $\eta = 1600$} \vspace{0.1cm}
\includegraphics[width=1\linewidth, trim={8mm 18mm 8mm 15mm},clip]{lidar2_log.pdf} \\[-3mm] \vspace{-0.1cm}
{\footnotesize\begin{flushleft} \hspace*{22.5mm} MSE 0.1523 \hspace*{4.2mm} MSE 0.2370 \hspace*{4.2mm} MSE 0.0993 \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(c) $\eta = 1700$} \vspace{0.1cm}
\includegraphics[width=1\linewidth, trim={8mm 18mm 8mm 15mm},clip]{cell1_log.pdf} \\[-3mm] \vspace{-0.1cm}
{\footnotesize\begin{flushleft} \hspace*{22.5mm} MSE 0.1960 \hspace*{4.2mm} MSE 0.2439 \hspace*{4.2mm} MSE 0.1302 \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(d) $\eta = 2700$} \vspace{0.1cm}
\includegraphics[width=1\linewidth, trim={8mm 18mm 8mm 15mm},clip]{cell3_log.pdf} \\[-3mm] \vspace{-0.1cm}
{\footnotesize\begin{flushleft} \hspace*{22.5mm} MSE 0.1667 \hspace*{4.2mm} MSE 0.2501 \hspace*{4.2mm} MSE 0.1185 \end{flushleft}} \vspace{-0.3cm}
\centerline{\footnotesize(e) $\eta = 1800$}
\caption{Estimation of $\log p$.
Images reconstructed through pixelwise MMSE estimation.
All images are scaled to $[0.001,0.101]$ and uniform prior is assumed.
MSE improvements for online threshold-based termination in place of conventional binomial and negative binomial sampling respectively, are:
(a) $1.86\,\mathrm{dB}$ and $2.56\,\mathrm{dB}$.
(b) $1.48\,\mathrm{dB}$ and $3.52\,\mathrm{dB}$.
(c) $1.86\,\mathrm{dB}$ and $3.78\,\mathrm{dB}$.
(d) $1.78\,\mathrm{dB}$ and $2.73\,\mathrm{dB}$.
(e) $1.48\,\mathrm{dB}$ and $3.24\,\mathrm{dB}$.
The negative binomial results have been obtained with $\ell = 5$.}
\label{fig:phantom-mmse-log}
\end{figure}
\begin{table}
\centering
\caption{Estimation of $f(p) = \log(p)$.
Average reconstruction MSEs, averaged over 100 experiments,
for conventional binomial sampling, negative binomial sampling and threshold-based termination,
for trial budgets $\eta$.
}
\label{tbl:Average_MSE_log}
\begin{tabular}{ccccc}
\hline
\multirow{2}{*}{Image} & \multirow{2}{*}{$\eta$} & \multicolumn{3}{c}{\scriptsize Method} \\
\cline{3-5}
& & \multicolumn{1}{c}{\scriptsize Binomial} & \multicolumn{1}{c}{\scriptsize Neg. Binomial} & \multicolumn{1}{c}{\scriptsize Threshold-based} \\ \hline
\multirow{2}{*}{Shepp--Logan} & $1800$ & $0.271$ & $0.481$ & $\mathbf{0.227}$ \\& $3000$ & $0.208$ & $0.253$ & $\mathbf{0.139}$ \\
\hline
\multirow{2}{*}{lidar \#1} & $1600$ & $0.156$ & $0.246$ & $\mathbf{0.113}$ \\
& $2200$ & $0.133$ & $0.165$ & $\mathbf{0.080}$ \\
\hline
\multirow{2}{*}{lidar \#2} & $1700$ & $0.154$ & $0.237$ & $\mathbf{0.099}$ \\
& $2300$ & $0.135$ & $0.161$ & $\mathbf{0.073}$ \\
\hline
\multirow{2}{*}{\emph{Foraminifera}} & $1700$ & $0.253$ & $0.462$ & $\mathbf{0.208}$ \\
& $2700$ & $0.206$ & $0.245$ & $\mathbf{0.131}$ \\
\hline
\multirow{2}{*}{\emph{HairStyle}} & $1800$ & $0.166$ & $0.247$ & $\mathbf{0.118}$ \\
& $2500$ & $0.143$ & $0.166$ & $\mathbf{0.087}$ \\
\hline
\end{tabular}
\end{table}
\section{Conclusion}
We established a novel framework for estimating Bernoulli parameters where we represent each Bernoulli process with a simple trellis graph.
By exploiting the mathematical convenience that comes from assuming Beta priors, we propose and study three stopping strategies with varying complexities but yielding very nearly equal performances.
All strategies give significant performance improvements over the conventional binomial and negative binomial stopping rules in simulated active imaging applications.
The simple online threshold-based termination was shown to asymptotically allocate trials in the same manner as an
oracle-aided solution that assumes the Bernoulli parameters are \emph{a priori} known.
Whilst we only study herein oracle-aided binomial sampling, similar analyses are possible for functions of Bernoulli parameters or oracle-aided negative binomial sampling;
these are omitted here because they do not yield clean expressions like the binomial case.
Finally, the proposed online threshold-based termination is extended to the estimation of $\log{p}$.
Other functions of $p$, such as $f(p) = 1/p$, can prove useful for scenarios wherein distinguishing between small parameters is paramount.
In the formulation of optimizing a trellis to minimize MSE,
a beta prior is convenient but not at all fundamental.
The reduction of the design problem from a general tree to a trellis holds for any prior,
and one may in principle compute Bayes risk reduction per trial for any trellis node
and any prior.
Developing an analogous theory for minimax estimation is also of interest
but is less clear because of a lack of additivity of the cost function.
This is especially intriguing because the optimization for MSE incidentally reduces the maximum over $p$ of the risk (see Fig.~\ref{fig:bin_greedy_oracle-EN_ER-comp}(b)).
\section*{Acknowledgment}
The authors thank Charles Saunders for assistance with implementation of regularized
estimators and Joshua Rapp for discussions on influence in spatial neighborhoods.
\balance
\bibliographystyle{ieeetr}
|
\section{Introduction}
\label{intro}
It may be useful when dealing with a large set of curves differing slightly from one to other to provide a smart representation. Indeed, usually a reduction taking into account some prior knowledge on the curves may lead to a better understanding of the variability within the population. One way to perform such reduction consists in considering that the set of curves has been obtained by deforming the template. This point of view is usually called curve registration and has been widely studied in statistics (see for example \cite{gasser1995,golubev1988,hardle1990,kneip1995,kneip2000,lawton1972} and references therein). More recently, this topic has found a second wind pushed by applications in image and signal processing. We refer for example to \cite{grenander93,McGuire1998ParameterRecovery,
allassonniere2013,gamboa2007,bigot2010,bigot09,gassiat2006,vimond2006}
for models and techniques related to signal processing problems.
The set of transformations often consists in a parametric family of operator acting on curves \cite{bigot2012}. One very popular of such model is the so-called shape invariant model (SIM) introduced in \cite{lawton1972} and studied for example in \cite{kneip1995, lindstrom1995,hardle1990,ke2001,gamboa2007}.
Multidimensional extension of SIM has been studied for functions defined on the plane. In this case, the parametric transformations involve
rotation and scaling parameters. This model has been studied for image registration see for example \cite{bigot09} or \cite{McGuire1998ParameterRecovery}. For non parametric approaches, we refer to \cite{ramsay} and references therein.
Generally, the parametric transformation on the curve acts independently on the argument and on the value of the template function. More precisely, let $\tilde{f}$ denotes the template function and $\theta$ denotes the variable parameterizing the transformation. Then,
the transformed function evaluated at $x$ may be written as
$T_{2,\theta}(f(T_{1,\theta}(x))$ where $T_{2,\theta}$ is an application on
$\mathbb{R}$ and $T_{1,\theta}$ acts on the set where $\tilde{f}$ is defined.
The parametric estimation is then performed using some $M$-estimation method (see \cite{vdV00}). In this paper, we work on functions defined on $[0,1]$ and consider a somehow different class of parametric transformations. Indeed, we consider transformations acting jointly on the argument and on the value of the template function. We work with the plane similarities acting on the whole curve
$\tilde{C}:= (x,\tilde{f}(x))_{x\in[0,1]}$.\\
\noindent
In our paper, we consider first the case where the template is known on the whole design space. Then, we extend the results to the more realistic set up where curves and the template are only observed on a grid. The estimations of the unknown parameters are performed using a $M$-estimation procedure.\\
\noindent
Our paper is organized as follow: in Section \ref{sec:1}, we define our model, we develop $M$-estimation techniques to perform estimation and study the asymptotic behaviour of the estimators. Section \ref{sec:2} is devoted to examples. We first give a toy model example validating our procedure. Then, we apply the methodology to an aeronautic model: the prediction of aeronautic loads (see \cite{fournier}). \\
\noindent
All the proofs are postponed to the last section.
\section{Framework, model and analytic results}
\label{sec:1}
In this section, we describe the statistical model studied and give the asymptotic behavior of the $M$-estimators of the unknown parameters.
\subsection{The observations}
\label{sec:11}
\begin{notation} Let be $x \in [0,1]$, and $f: [0,1]\to \mathbb{R}^{+}$. We denote by $C$ the curve\\
$C:=\begin{pmatrix} x \\ f(x) \end{pmatrix}_{x\in [0,1]}$.
\end{notation}
In our framework, we have at hand $K+1$ curves. $\tilde{C}$ is the reference curve and we assume that the $K$ other curves $C_{j},\ j=1,...,K$ are the images of $\tilde{C}$ by the transformation model described bellow. The $K$ curves are observed on the same random grid $\mathcal{D}_{N}:=\{X_{1},...,X_{N}\}$ where $(X_{i})_{i=1,...,N}$ are iid random variables with uniform distribution. Hence we have at our disposal\\
\begin{center}
\begin{tabular}{l}
$C_{j}^{N}=(X_{i},f_{j}(X_{i}))_{i=1,...,N,\ j=1,...,K}$.\\
\end{tabular}
\end{center}
We will consider the following two cases\\
\begin{itemize}
\item[] i) $\tilde{C}$ is known everywhere: $\tilde{C}=(x,\tilde{f}(x))$, $\forall x \in [0,1]$,\\
\item[] ii) $\tilde{C}$ is only known on $\mathcal{D}_{N}$: $\tilde{C}=(x,\tilde{f}(x))$, $\forall x \in \mathcal{D}_{N}$.\\
\end{itemize}
\subsection{Transformation model}
\label{sec:12}
Before defining the transformation model linking the $K$ observed curves to the pattern $\tilde{C}$, one must ensure that the functions $f_{j}$, $j=1,...,K$ and $\tilde{f}$ are "admissible". This is the aim of the next definition:\\
\begin{definition}Let $\theta_{0} \in ]0,\frac{\pi}{2}[$, $\mathcal{F}_{\theta_{0}}$ is the set of applications defined by:
\begin{center}
$\mathcal{F}_{\theta_{0}}:=\{f:[0,1]\to \mathbb{R}^{+}, f\ \mathrm{is}\ \mathrm{differentiable}\ \mathrm{on}\ [0,1],$\\
$f(0)>0, f(1)=0$,\\
$f'(x)<0,f'(1)=0, f'(0)>-\cot{\theta_{0}}\}$.
\end{center}
\end{definition}
Let now define the parametric family of transformation that we will put in action.\\
\begin{notation}For any function $T_{\alpha}: \mathbb{R}^{2} \to \mathbb{R}^{2}$ depending on a parameter $\alpha\in\Theta$, we will denote by $T_{\alpha}^{1}$ and $T_{\alpha}^{2}$ its two coordinates.\\
\end{notation}
In the following, we define our transformation model.\\
Set $0<\theta_{0}<\theta_{1}<\frac{\pi}{2}$, and $0<\lambda_{min}<\lambda_{max}$. For any $\alpha:=(\theta,\lambda) \in \Theta= [-\theta_{0},\theta_{1}]\times [\lambda_{min},\lambda_{max}]$, set\\
\begin{center}
$T_{\alpha}:=H_{\lambda}\circ S_{\theta}\circ R_{\theta}$
\end{center}
the composition of a rotation $R_{\theta}$, a rescaling $S_{\theta}$ applied on the $x$-axis, and a homothetic transformation $H_{\lambda}$ on the $y$-axis. More precisely\\
\begin{itemize}
\item[$\bullet$] $R_{\theta}$ is the rotation centered in $\begin{pmatrix} 1 \\ 0 \end{pmatrix}$ and of angle $\theta$. That is\\
\begin{align*}
R_{\theta}: [0,1]\times\mathbb{R}^{+} &\to [-A_{min},1]\times\mathbb{R} \\ \begin{pmatrix} x \\ y \end{pmatrix} &\to \begin{pmatrix} (x-1)\cos{\theta}-y\sin{\theta}+1 \\ (x-1)\sin{\theta}+y\cos{\theta} \end{pmatrix},
\end{align*}
with $A_{min}=\sqrt{1+y(0)^{2}}-1$.\\
\item[$\bullet$] After rotating a curve, depending of the angle $\theta$, the resulting design space of the curve is generally no more [0,1]. Hence, we define the following transformation to re-position the curve on [0,1]:\\
\begin{align*}
S_{\theta}: [-A_{min},1]\times\mathbb{R} &\to [0,1]\times\mathbb{R} \\ \begin{pmatrix} x \\ y \end{pmatrix} &\to \begin{pmatrix} \frac{x-a}{1 -a} \\ y \end{pmatrix},
\end{align*}
with $a$ is the real minimum value of the design space after rotating the curve.\\
\item[$\bullet$] The scaling transformation of parameter $\lambda>0$ acting on the second coordinate of the curve:
\begin{align*}
H_{\theta}: [0,1]\times\mathbb{R^{+}} &\to [0,1]\times\mathbb{R}^{+} \\ \begin{pmatrix} x \\ y \end{pmatrix} &\to \begin{pmatrix} x \\ \lambda y \end{pmatrix}.
\end{align*}
\end{itemize}
We now introduced the assumption that will be used:\\
\begin{itemize}
\item[ ] \textbf{(A1)}: The functions $(f_{j})_{j=1,...,K}$, and $\tilde{f}$ belong to $\mathcal{F}_{\theta_{0}}$.\\
\item[ ] \textbf{(A2)}: For any $\theta \in [-\theta_{0},\theta_{1}]$, for any $x \in [0,1]$ and any $f \in \mathcal{F}_{\theta_{0}}$, the second coordinate of $R_{\theta}$ is positive.\\
\item[ ] \textbf{(A3)}: $\Theta=[-\theta_{0},\theta_{1}]\times [\lambda_{min},\lambda_{max}]$.\\
\item[ ] \textbf{(A4)}: $\tilde{C}$ is known on $[0,1]$.\\
\item[ ] \textbf{(A5)}: $\tilde{C}$ is known on the grid $\mathcal{D}_{N}$.\\
\end{itemize}
Thus, the transformation model considered is as follows:\\
\begin{align*}
T_{\alpha}: [0,1]\times\mathbb{R}^{+} &\to [0,1]\times\mathbb{R}^{+} \\ \begin{pmatrix} x \\ f(x) \end{pmatrix} &\to \begin{pmatrix} \frac{(x-1)\cos{\theta}-f(x)\sin{\theta}}{\cos{ \theta}+f(0)\sin{\theta}}+1 \\ \lambda((x-1)\sin{ \theta}+f(x)\cos{\theta}) \end{pmatrix}.
\end{align*}
\subsection{Regression model}
\label{sec:13}
Recall that we wish to adjust the reference curve $\tilde{C}$ on the other curves $C_{j}$ ($j=1,...,K$) by transformations defined in the previous subsection. Notice that these transformations act on both axis. For any $\alpha$, we want to compare the value of the transformed curve $(T_{\alpha}\tilde{C})(X_{i})$ with $f_{j}(X_{i})$. Since the abscissa points are affected by the transformation, we denote by $X_{i}(\alpha)$ the point such that $T_{\alpha}^{1}(X_{i}(\alpha))=X_{i}$. For that reason, we introduce the following definition:
\begin{definition}We denote by $x(\alpha,g)$ the solution of the equation:
\begin{equation}
T_{\alpha}^{1}(u,g(u))=x.
\label{eq:eq1}
\end{equation}
\end{definition}
To ease the notation, we finally set $x(\alpha):=x(\alpha,g)$, so $X_{i}(\alpha)=X_{i}(\alpha,\tilde{f})$. We consider the parametric regression model:
\begin{equation}
f_{j}(X_{i})=T_{\alpha_{j}^{*}}^{2}(X_{i}(\alpha_{j}^{*}),\tilde{f}(X_{i}(\alpha_{j}^{*})))+\epsilon_{j,i}, (j=1,...,K).
\label{eq:eq2}
\end{equation}
Where:
\begin{itemize}
\item[$\bullet$] $(X_{i})$ are iid with distribution $\mathbb{P}_{X}$. It is the design on which we observe the curves $C_{j}$;
\item[$\bullet$] $\alpha_{j}^{*}=(\theta_{j}^{*}$,$\lambda_{j}^{*})$ is the couple of true parameters for each curve ($j=1,...,K$);
\item[$\bullet$] $\epsilon_{j,i},\ \forall i=1,...,N, \ \forall j=1,...,K$ are iid $\mathcal{N}(0,\sigma^{2})$ random variables. These variables are assumed to be independent of $X_{i}$.\\
\end{itemize}
\subsection{Estimation}
\label{sec:14}
\subsubsection{Estimation when $\tilde{C}$ is known on $[0,1]$}
\label{sec:141}
For the sake of simplicity, let us fix $j$. Relying on a classical $M$-estimation procedure, we consider a semi-parametric method to estimate the parameters and define consequently the following empirical contrast function to fit the reference curve $\tilde{C}$ to $C_{j}$ ($j=1,...,K$):
\begin{equation}
\begin{split}
M^{j}_{N}(\alpha)&=\frac{1}{N}\sum_{i=1}^{N}(f_{j}(X_{i})-T_{\alpha}^{2}(X_{i}(\alpha),\tilde{f}(X_{i}(\alpha))))^{2}\\
&=\frac{1}{N}\sum_{i=1}^{N}m_{\alpha}^{j}(X_{i}).
\end{split}
\label{eq:eq3}
\end{equation}
The random function $M^{j}_{N}$ is non negative. Furthermore, intuitively, its minimum value
should be reached close to the true parameter $\alpha_{j}^{*}$. Indeed, the following theorem gives the consistency of the $M$-estimator, defined by :
\begin{equation}
\hat{\alpha}^{j}_{N}=\underset{\alpha \in \Theta}{\mathrm{arg min}}\ M^{j}_{N}(\alpha).\\
\label{eq:eq4}
\end{equation}
Recall that our empirical contrast function enters in the general theory of $M$-estimator. The Central Limit Theorem will be shown by using $M$-estimator arguments.\\
\begin{theorem}
Assume that A1, A2, A3 and A4 are satisfied. Then
\begin{align}
\begin{split}\label{eq:eq5}
i)\ {}& \hat{\alpha}_{N}^{j} \xrightarrow[N \to +\infty]{\mathbb{P}} \alpha_{j}^{*},\\
\end{split}\\
\begin{split}\label{eq:7}
ii)\ {}& \sqrt{N}(\hat{\alpha}_{N}^{j} -\alpha_{j}^{*}) \xrightarrow[N\to + \infty]{\mathcal{L}} \mathcal{N}(0,\Gamma_{\alpha_{j}^{*}}).\\
\end{split}
\end{align}
In particular, the covariance matrix has the following form
\begin{equation}
\Gamma_{\alpha_{j}^{*}} = V_{\alpha_{j}^{*}}^{-1}2\sigma^{2},
\label{eq:eq8}
\end{equation}
with $V_{\alpha_{j}^{*}}=2\mathbb{E}[\dot T_{\alpha_{j}^{*}}^{2}\dot T_{\alpha_{j}^{*}}^{2\textbf{T}}]$, and $\dot T_{\alpha_{j}^{*}}^{2}$ is the vector of partial derivatives of $T_{\alpha_{j}}^{2}$ w.r.t elements of $\alpha$ taken at $\alpha_{j}^{*}$.
\end{theorem}
\subsubsection{Estimation when $\tilde{C}$ is observed on $\mathcal{D}$}
\label{sec:142}
In this section, we consider the case where the reference curve $\tilde{C}$ is observed on the same grid $\mathcal{D}:=(X_{i})_{i=1,..,N}$ as the other curve $C_{j}$, i.e $\tilde{C}=\begin{pmatrix} x \\ \tilde{f}(x) \end{pmatrix}_{x \in D}$. By applying the transformation $T_{\alpha}$ to $\tilde{C}$, the transformed pattern $T_{\alpha}\tilde{C}$ is no longer observable on $\mathcal{D}$. As a consequence, one must make use of an approximation process over $\tilde{f}$. Let $\tilde{f}_{N}$ be the linear interpolate of $\tilde{f}$, defined by:\\
\begin{equation}
\tilde{f}_{N}(x)=\sum_{i=1}^{N}\Delta_{i}(x)\mathds{1}_{x\in[X_{(i)},X_{(i+1)})},
\label{eq:eq9}
\end{equation}
where
\begin{equation}
\Delta_{i}(x)=\frac{\tilde{f}(X_{(i+1)})-\tilde{f}(X_{(i)})}{X_{(i+1)}-X_{(i)}}\ x\ +\ \tilde{f}(X_{(i)})-\frac{\tilde{f}(X_{(i+1)})-\tilde{f}(X_{(i)})}{X_{(i+1)}-X_{(i)}}X_{(i)}.\\
\label{eq:eq10}
\end{equation}
It is easy to see that $\tilde{f}_{N}$ belongs also to $\mathcal{F}_{\theta_{0}}$. Replacing $\tilde{C}$ by $\hat{C}$ in (\ref{eq:eq3}) we obtain\\
\begin{center}
$\hat{M}^{j}_{N}(\alpha)=\frac{1}{N}\sum_{i=1}^{N}(f_{j}(X_{i})-T_{\alpha}^{2}(X_{i}(\alpha,N),\tilde{f}_{N}(X_{i}(\alpha,N))))^{2}$,\\
\end{center}
where $X_{i}(\alpha,N)$ is the solution to the equation $T_{\alpha}^{1}(u,\tilde{f}_{N}(u))=X_{i}$.\\
Using the linear interpolate defined by (\ref{eq:eq9}), we show the consistency and asymptotic normality of our $M$-estimator defined as follow:
\begin{equation}
\hat{\hat{\alpha}}^{j}_{N}=\underset{\alpha \in \Theta}{\mathrm{arg min}}\ \hat{M}^{j}_{N}(\alpha).\\
\label{eq:eq32}
\end{equation}
\begin{theorem}
Assume that A1, A2, A3 and A5 are satisfied. Let $\tilde{f}_{N}$ be defined by (\ref{eq:eq9}) and assume that $\exists C>0$ s.t $\forall x\in[0,1]$, $\tilde{f}_{N}'(x)\leq C$, and $\tilde{f}'(x)\leq C$, then
\begin{align}
\begin{split}\label{eq:eq33}
i)\ {}& \hat{\hat{\alpha}}_{N}^{j} \xrightarrow[N \to +\infty]{\mathbb{P}} \alpha_{j}^{*},\\
\end{split}\\
\begin{split}\label{eq:34}
ii)\ {}& \sqrt{N}(\hat{\hat{\alpha}}_{N}^{j} -\alpha_{j}^{*}) \xrightarrow[N\to + \infty]{\mathcal{L}} \mathcal{N}(0,\Gamma_{\alpha_{j}^{*}})\\
\end{split}
\end{align}
with $\Gamma_{\alpha_{j}^{*}}$ such defined in (\ref{eq:eq8}).
\end{theorem}
\section{Simulations and applications}
\label{sec:2}
In this section we illustrate the method on numerical applications. The first subsection is dedicated to some simulated toy example while the second to a real problem. The optimisation problems (\ref{eq:eq4}) and (\ref{eq:eq32}) will be numerically solved by using the BFGS algorithm \cite{NoceWrig06}.
\subsection{Simulated toy example}
\label{sec:21}
We consider the following model:\\
\begin{equation*}
f_{\lambda,\theta}(x)=\lambda[(x-1)\sin{\theta}+g(x)\cos{\theta}],
\end{equation*}
with $g(x)=2(\cos{\pi x}+1)$. We observe $f_{\lambda_{j},\theta_{j}}(x_{i})$ for $i=1,...,100$, for $J=25$ values of $(\lambda,\theta)$ with some iid errors $\epsilon_{ij}$.
\begin{itemize}
\item[$\bullet$] The observations points $x_{i},\ i=1,...,100$ are iid random variables with uniform distribution on $[0,1]$.
\item[$\bullet$] The parameters are chosen randomly with the following arbitrary distribution $(\lambda,\theta) \hookrightarrow \mathcal{U}([0,15])\times \mathcal{U}([-\frac{\pi}{3},\frac{\pi}{30}])$.
\item[$\bullet$] The errors are assumed to be $\mathcal{N}(0,0.01)$.
\end{itemize}
Results are given in Figure \ref{label-simu}. The simulated data are shown in Figure \ref{label-simu} (a). Each of these curves is rescaled to $[0,1]$ to avoid numerical issues. The curve in blue is the reference curve. After the estimation of the parameters $\lambda_{j}$ and $\theta_{j}$, all curves can be rescaled back to their original space as shown in Figure \ref{label-simu} (b). \\
Rescaling the curves allows to easily choose the initial point of parameters for the optimization algorithm taking 1 for $\lambda$ and 0 for $\theta$.\\
\begin{figure}[h!]
\centering
\subfigure[]{\includegraphics[width=6cm]{Init-4.png}}\subfigure[]{\includegraphics[width=6cm]{Fitted-4.png}}
\caption{Rotating and scaling results for simulated data: (a) Simulated data; (b) Fitted data}
\label{label-simu}
\end{figure}
\subsection{Aeronautic loads}
\label{sec:22}
An airframe structure is a complex system and its design is a complex task involving today many simulation activities generating massive amounts of data. This is, for example, the process of loads and stress computations of an aircraft. That is the computations of the forces and the mechanical strains suffered by the structure. The overall process exposed in Figure \ref{label-process} is run to identify load cases (i.e aircraft mission and configurations: maneuvers, speed, loading, stiffness...), that are critical in terms of stress endured by the structure and, of course, the parameters which make them critical. The final aim is to size and design the structure (and potentially to reduce loads in order to reduce the weight of the structure). Typically for an overall aircraft structure, millions of load cases can be generated and for each of these load cases millions of structural responses (i.e how structural elements react under such conditions) have to be computed. As a consequence, computational times can be significant.\\
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{ComputingChain.jpeg}
\caption{Flowchart for loads and stress analysis process}
\label{label-process}
\end{figure}
In an effort to continuously improve methods, tools and ways-of-working, Airbus has invested a lot in digital transformation and the development of infrastructures allowing to treat data (newly or already produced). The main industrial challenge for Airbus is to reduce lead time in the computation and preliminary sizing of an airframe as well as extracting value from already calculated loads. In this paper, we focus on the external loads of a wing: for each load case are calculated the shear forces (transverse forces near to vertical arising from aerodynamic pressure and inertia) and bending moments (resulting from the shear forces, they represent the flexion of the wing) such as shown in Figure \ref{label-BM}.
\begin{figure}[h!]
\centering
\includegraphics[width=5.5cm]{Examples.png}\includegraphics[width=5.5cm]{Bending.jpeg}\\
\caption{(a) Examples of bending moments along the wing for different load cases - (b) Finite element model of a generic aircraft representing the wing deformation \cite{ritter01}}
\label{label-BM}
\end{figure}
These external loads appeared to be extremely regular and one can legitimately suppose that it exists a link between all those curves. Indeed, it is natural to assume that it exists a reference bending moment (a reference curve) which can be morphed through a deformation model to give all the other curves.\\
In \cite{fournier}, the authors present an aeronautic model that computes the loads (forces and moments) on the wing of some aircraft model denoted by $ACM1$. They present several statistical methods in order to study these data. In this section, we will compare the method used in \cite{fournier} with the model presented in Section \ref{sec:1} for a new aircraft model called $ACM2$. The data at our disposal represents bending moments of a wing (representing its flexion) of an aircraft calculated for 1152 different configurations (load cases). Each configuration is defined by 28 features (speed of the aircraft, mass, altitude, quantity of fuel, etc.), leading to a bending moment calculated on 45 stations along the wing. In a more formal way, we observe the couple $(X_{j},Y_{j})_{j=1,...,1152}$, where $X_{j}=(X_{j}^{1},...,X_{j}^{28})$ are the features and $Y_{j}=(Y_{j}^{1},...,Y_{j}^{45})$ is the bending moment. The idea is to predict the bending moment for different configurations. The data are represented in Figure \ref{label-210}.\\
\begin{figure}[h!]
\centering
\includegraphics[width=6.5cm]{examples210.png}\\
\caption{Representation of all the bending moments of a wing of our data base: the wing root is located at the null coordinate, where the strains are maximum when the wing bends.}
\label{label-210}
\end{figure}
Due to the discontinuities at the $3^{rd}$ and $20^{th}$ stations, we apply our methodology to each section independently. Then, each section can be represented by its minimum and maximum values, and by its rotation and scaling coefficients $\lambda_{j}$ and $\theta_{j}$. Figure \ref{label-matched} assess the quality of the matching process (the reference curve used is the average bending moment).\\
\begin{figure}[h!]
\centering
\includegraphics[width=6.5cm]{unscaled.png}\\
\caption{Results of the matching process}
\label{label-matched}
\end{figure}
Thus the dimensional space of the outputs is reduce to 12 instead of 45. We compare our method to three other methods of \cite{fournier} applied on the outputs: no transformation (we call it raw - we build 45 models, one per station); a PCA (the three first principal components represent 99,9\% of the explained variance - 3 models instead of 45); a polynomial fitting per section (of degree 4 for the first section, of degree 2 for the second and of degree 1 for the third section) which leads to 10 models instead of 45. The Table \ref{table:tabout} sums up the number of outputs to predict depending on the method used.\\
\begin{table}[h!]
\centering
\caption{Number of outputs to be predicted depending on the method used on the raw outputs: Raw, Deformation Model, PCA, polynomial fitting}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& Number of outputs & Names of outputs \\
\hline
\hline
Raw & 45 & Bending moment value at \\
& &station 0 to 44 \\
\hline
Deformation & 12 & $\theta_{1}, \theta_{2}, \theta_{3}, \lambda_{1}, \lambda_{2}, \lambda_{3}, min_{1}, min_{2}, $\\
Model & &$min_{3}, max_{1}, max_{2}, max_{3}$ \\
\hline
PCA & 3 & Principal components 1 to 3 \\
\hline
Polynomial fitting & 10 & Coefficients of polynomials \\
\hline
\end{tabular}\\
\label{table:tabout}
\end{table}
The significant advantage of the reduction dimension techniques used is that the response of the model would have a physical form contrary to the simple linear models performed on the raw data. To build our models, we use the Orthogonal Greedy Algorithm (OGA) also known as the Matching Pursuit Algorithm. Detailed explanations can be found in \cite{barron2008}, \cite{sancetta2016} and \cite{mallat93}. Roughly speaking, we consider the problem of approximating a function by a sparse linear combination of inputs.\\
To assess the goodness of fit of our models, we defined for a curve of bending moment $j$ the error rate as follows:\\
\begin{center}
$error(j)=\sqrt{\frac{\sum_{i=1}^{45}(\hat{y}(x_{i})-y_{j}(x_{i}))^{2}}{\sum_{i=1}^{45}y_{j}^{2}(x_{i})}}$, $j=1,...,n_{test}$,\\
\end{center}
where $n_{test}$ is the size of the sample of test. We compute the error rates on (the sample of test is of 25\% the size of the total database). It gives an idea of how far our predictions are. For this standpoint, we can easily compute the empirical cumulative distribution function (CDF): $\forall\ j=1,...,n_{test}$, let $\alpha \in [0,1]$. The empirical CDF is defined as:\\
\begin{center}
$\alpha \to G(\alpha)=\frac{1}{n}\sum_{j=1}^{n_{test}}\mathds{1}_{(error(j)\leq\alpha)}$
\end{center}
In Table \ref{table:tab6}, we give the values of $G(\alpha)$ for $\alpha=1\%,2\%,5\%,10\%$ and the mean error. In Figure \ref{label-CDF} we give the plots of the function $G(\alpha)$ for the different methods.\\
\begin{table}[h!]
\caption{Average estimated $\mathbb{P}(error\leq 1\%)$, $\mathbb{P}(error\leq 2\%)$, $\mathbb{P}(error\leq 5\%)$ $\mathbb{P}(error\leq 10\%)$, $\mathbb{E}(error)$ calculated on several random test data set (25\% of the size of the total dataset)}
\makebox[\textwidth][c]{
\begin{tabular}{|c|c|c|c|c||}
\hline
&Deformation Model &Polynomial Fitting &PCA &Raw\\
\hline
\hline
$\mathbb{P}(error\leq 1\%)$ & 17\% & 14\% &16\% &15\% \\
\hline
$\mathbb{P}(error\leq 2\%)$ & 45\% & 45\% &43\% &51\% \\
\hline
$\mathbb{P}(error\leq 5\%)$ & 88\% & 88\% &86\% &88\% \\
\hline
$\mathbb{P}(error\leq 10\%)$ & 98\% &97\% &95\% &98\% \\
\hline
$\mathbb{E}(error)$ & 2.9\% & 2.9\% & 3\% & 2.8\%\\
\hline
\hline
\end{tabular}}
\label{table:tab6}
\end{table}
\begin{figure}[h!]
\centering
\includegraphics[width=6.7cm]{Res-3.png}\\
\caption{Empirical CDF of error rates ($\mathbb{P}(error\leq \alpha)$)}
\label{label-CDF}
\end{figure}
Concerning the approximation and prediction of loads, our model is equivalent in average to other tested methods, there are just slightly more observations with an error below 1\%. Nevertheless, in our case, the linear models built through the deformation model are sparser than the other. Indeed, in average, 11 variables are chosen as optimal parameter of the greedy algorithm by cross-validation for the deformation model, 13 for the polynomial fitting, 15 for the PCA and 14 for the raw outputs one.\\
Even though the prediction of loads with the deformation model is so likely equivalent to none transformation, it obtains better results than the polynomial fitting and the PCA. Besides, it is important to notice that it gives to engineers a physical interpretation and idea of how react the wing to new constraints. Besides, using this deformation model gives a physical response contrary to a simple linear model per station whose response could be irregular.\\
\section{Proofs and technical result}
\label{sec:3}
\subsection{Technical result}
\label{sec:31}
This section is dedicated to the technical result used in the proof of Theorem 4.\\
\begin{lemma}Let $X_{1},...,X_{N}$ be $N$ independent and identically distributed random variables with uniform distribution on $[0,1]$ and let $X_{(1)}\leq...\leq X_{(N)}$ be the reordered sample . Let $a_{N}=O(\sqrt{N})$, then:
\begin{equation*}
\mathbb{P}(a_{N}\underset{j}{\sup}|X_{(j+1)}-X_{(j)}|\geq \epsilon)\xrightarrow[N \to +\infty]{} 0.
\end{equation*}
\end{lemma}
\begin{proof}[Proof of Lemma 1] Let $Z_{1},...,Z_{N+1}$ be $N$ independent and identically distributed random variables with exponential distribution with parameter $1$. It is a well known fact that \\
$(\frac{Z_{1}}{\sum_{k=1}^{N+1}Z_{k}},\frac{Z_{1}+Z_{2}}{\sum_{k=1}^{N+1}Z_{k}},...,\frac{Z_{1}+...+Z_{N}}{\sum_{k=1}^{N+1}Z_{k}})\overset{(\mathcal{L})}{=}(X_{(1)},...,X_{(N)})$ and we have
\begin{equation}
X_{(j+1)}-X_{(j)}\overset{(\mathcal{L})}{=}\frac{Z_{j+1}}{\sum_{k}Z_{k}}.
\label{eq:ordered}
\end{equation}
Now, for $\epsilon >0$
\begin{equation*}
\begin{split}
\mathbb{P}(\underset{j}{\sup}|X_{(j+1)}-X_{(j)}|\geq \epsilon) &\leq \sum_{j} \mathbb{P}(X_{(j+1)}-X_{(j)}\geq \epsilon)\\
&\leq N \max_{j} \mathbb{P}(X_{(j+1)}-X_{(j)}\geq \epsilon).\\
\end{split}
\end{equation*}
By using (\ref{eq:ordered}), we have
\begin{equation*}
\begin{split}
\mathbb{P}(X_{(j+1)}-X_{(j)}\geq \epsilon)=(1-\epsilon)^{N-1}.\\
\end{split}
\end{equation*}
Then,
\begin{equation*}
\begin{split}
\mathbb{P}(\underset{j}{\sup}|X_{(j+1)}-X_{(j)}|\geq \epsilon) &\leq N(1-\epsilon)^{N-1}.\\
\end{split}
\end{equation*}
The result follows replacing $\epsilon$ by $\frac{\epsilon}{a_{N}}$ and letting $N \to +\infty$.
\end{proof}
\subsection{Proofs of Theorems 3 and 4}
\label{sec:32}
\begin{proof}[Proof of Theorem 3] To ease the notation, we do not display the dependency in $j$.\\
i) By (\ref{eq:eq3}) it is easy to see that $M_{N}(\alpha)$ is an empirical mean of iid bounded random variables. Thus, by the Strong Law of Large Number (SLLN)
\begin{center}
$M_{N}(\alpha) \xrightarrow[N \to +\infty]{p.s} M(\alpha)$,
\end{center}
with $M(\alpha)=\mathbb{E}[\epsilon^{2}]+ \mathbb{E}[(T_{\alpha}^{2}(X(\alpha),\tilde{f}(X(\alpha)))-T_{\alpha^{*}}^{2}(X(\alpha^{*}),\tilde{f}(X(\alpha^{*}))))^{2}]$.\\
$M(\alpha)$ is continuous and has an obvious unique minimum $\alpha^{*}$. Since $\Theta$ is compact, this implies that $\underset{\alpha:d(\alpha,\alpha^{*})\geq\epsilon}{\inf}\ M(\alpha) >M(\alpha^{*})$ is satisfied (see Problem 27 p. 84 in \cite{vdV00}).\\
It remains to prove that $\{m_{\alpha} : \alpha \in \Theta\}$ is a Glivenko-Cantelli class. Thanks to the remark following the proof of Theorem 5.9 in \cite{vdV00}, this is an easy consequence of the continuity of $\alpha \to m_{\alpha}$ and the fact that the function is bounded by a continuous and integrable function on $[0,1]$. Indeed, it exists at least a function $f^{*}$ in $\mathcal{F}_{\theta_{0}}$ which bounds every other functions, and two constants $K_{1} >0, K_{2}>0$ such that
\begin{center}
$m_{\alpha}(x) \leq K_{1}(f^{*}(x)+K_{2})^{2}$,
\end{center}
and
\begin{equation}
\underset{\alpha\in\Theta}{\sup}\ |M_{N}(\alpha)-M(\alpha)|\xrightarrow[N \to +\infty]{\mathbb{P}}\ 0.\\
\label{eq:eq6}
\end{equation}
The result follows from the Theorem 5.7 in \cite{vdV00}.\\
ii) The Central Limit Theorem will be a consequence of Theorem 5.23 in \cite{vdV00}. Recall that
\begin{center}
$m_{\alpha}(x)=[f(x)-\lambda((x(\alpha)-1)\sin \theta + \cos \theta \tilde{f}(x(\alpha)) )]^{2}$.
\end{center}
By the Implicit Function Theorem, that is easy to see that $\alpha \to x(\alpha)$ is $C^{1}$ on a compact set. This implies that the norm of the gradient of $m_{\alpha}$ is uniformly bounded in $\alpha$. Hence $\exists \dot\phi(x) \in L^{1}$ such that $||\nabla_{\alpha}m_{\alpha}(x)||\leq \dot\phi(x)$ hence\\
\begin{center}
$|m_{\alpha_{1}}(x)-m_{\alpha_{2}}(x)| \leq \dot\phi(x)\times ||\alpha_{1}-\alpha_{2}||.$
\end{center}
In order to give an explicit formula for the limit variance, we apply the results of Example 5.27 in \cite{vdV00} where $f_{\theta}$ becomes in our case $T_{\alpha}^{2}$ and hence, we have
\begin{equation*}
\sqrt{N}(\hat{\alpha}_{N}^{j} -\alpha_{j}^{*}) \xrightarrow[N\to + \infty]{\mathcal{L}} \mathcal{N}(0,\Gamma_{\alpha_{j}^{*}}),\\
\end{equation*}
with $\Gamma_{\alpha_{j}^{*}} = V_{\alpha_{j}^{*}}^{-1}2\sigma^{2}$ and $V_{\alpha_{j}^{*}}=2\mathbb{E}[\dot T_{\alpha_{j}^{*}}^{2}\dot T_{\alpha_{j}^{*}}^{2\textbf{T}}]$.
\end{proof}
\begin{proof}[Proof of Theorem 4] \ \\
i) To prove the consistency of $\hat{\hat{\alpha}}_{N}$ we have to show that\\
\begin{center}
$\underset{\alpha\in\Theta}{\sup}|\hat{M}_{N}(\alpha)-M(\alpha)|\xrightarrow[N \to +\infty]{\mathbb{P}}\ 0$.\\
\end{center}
We have,
\begin{center}
$\underset{\alpha\in\Theta}{\sup}|\hat{M}_{N}(\alpha)-M(\alpha)| \leq \underset{\alpha\in\Theta}{\sup}|\hat{M}_{N}(\alpha)-M_{N}(\alpha)| + \underset{\alpha\in\Theta}{\sup}|M_{N}(\alpha)-M(\alpha)|$.\\
\end{center}
It has been shown in the proof of Theorem 3 that
\begin{center}
$\underset{\alpha\in\Theta}{\sup}|M_{N}(\alpha)-M(\alpha)|\xrightarrow[N \to +\infty]{\mathbb{P}}\ 0$.\\
\end{center}
It remains to prove that
\begin{center}
$\underset{\alpha\in\Theta}{\sup}|\hat{M}_{N}(\alpha)-M_{N}(\alpha)|\xrightarrow[N \to +\infty]{\mathbb{P}}\ 0$.\\
\end{center}
To ease the notation, we write $T_{\alpha}^{2}(i,N)=T_{\alpha}^{2}(X_{i}(\alpha,N),\tilde{f}_{N}(X_{i}(\alpha,N)))$, and $T_{\alpha}^{2}(i)=T_{\alpha}^{2}(X_{i}(\alpha),\tilde{f}(X_{i}(\alpha)))$. Set
\begin{equation*}
\begin{split}
D_{N}(\alpha)&=M_{N}(\alpha)-\hat{M}_{N}(\alpha)\\
&=\frac{1}{N}\sum_{i=1}^{N}2y(X_{i})[T_{\alpha}^{2}(i,N)-T_{\alpha}^{2}(i)]-\frac{1}{N}\sum_{i=1}^{N}[T_{\alpha}^{2}(i,N)-T_{\alpha}^{2}(i)][T_{\alpha}^{2}(i,N)+T_{\alpha}^{2}(i)].\\
\end{split}
\end{equation*}
As $f$ and $T_{\alpha}^{2}$ are continuous and bounded on $\Theta\times [0,1]$, this implies that:\\
\begin{equation*}
\begin{split}
|D_{n}(\alpha)|&\leq |\frac{1}{N}\sum_{i=1}^{N}2f(X_{i})[T_{\alpha}^{2}(i,N)-T_{\alpha}^{2}(i)]|+|\frac{1}{N}\sum_{i=1}^{N}[T_{\alpha}^{2}(i,N)-T_{\alpha}^{2}(i)][T_{\alpha}^{2}(i,N)+T_{\alpha}^{2}(i)]|\\
&\leq K(\frac{1}{N}\sum_{i=1}^{N}[T_{\alpha}^{2}(i,N)-T_{\alpha}^{2}(i)]^{2})^{\frac{1}{2}}\\
&\leq K'(\frac{1}{N}\sum_{i=1}^{N}[(X_{i}(\alpha,N)-X_{i}(\alpha))(1+C)+(\tilde{f}_{N}(X_{i}(\alpha))-\tilde{f}(X_{i}(\alpha))]^{2})^{\frac{1}{2}}.\\
\end{split}
\end{equation*}
By construction, there exists $j$ such that $X_{(j)}\leq X_{i}(\alpha) \leq X_{(j+1)}$, and $X_{(j)}\leq X_{i}(\alpha,N) \leq X_{(j+1)}$ which leads to:
\begin{equation*}
X_{i}(\alpha,N)-X_{i}(\alpha)=\gamma(X_{(j+1)}-X_{(j)}).\\
\end{equation*}
Besides, since there exists $\gamma'>0$ such that\\
$\tilde{f}_{N}(X_{i}(\alpha))=\gamma'\tilde{f}(X_{(j+1)})+(1-\gamma')\tilde{f}(X_{(j)})$ we have
\begin{equation*}
\begin{split}
|\tilde{f}_{N}(X_{i}(\alpha))-\tilde{f}(X_{i}(\alpha))|&\leq \gamma'|\tilde{f}(X_{(j+1)})-\tilde{f}(X_{(j)})|\\
&\leq C\gamma'|X_{(j+1)}-X_{(j)}|, \\
\end{split}
\end{equation*}
and
\begin{align*}
|D_{n}(\alpha)|&\leq K'(\frac{1}{N}\sum_{j=1}^{N}[X_{(j+1)}-X_{(j)}]^{2})^{\frac{1}{2}} \\
\underset{\alpha\in\Theta}{\sup}|D_{n}(\alpha)|&\leq K'(\frac{1}{N}\sum_{j=1}^{N}\underset{\alpha\in\Theta}{\sup}[X_{(j+1)}-X_{(j)}]^{2})^{\frac{1}{2}}\\
&\leq K' \underset{j}{\sup}|X_{(j+1)}-X_{(j)}|.\\
\end{align*}
By Lemma 1 $\mathbb{P}(K'\underset{j}{\sup}|X_{(j+1)}-X_{(j)}|\geq \epsilon)\xrightarrow[N \to +\infty]{} 0$. Hence $D_{N}$ is bounded by an integrable and continuous function which goes to 0 in probability on $\Theta$\\
\begin{center}
$\underset{\alpha\in\Theta}{\sup}|\hat{M}_{N}(\alpha)-M(\alpha)|\xrightarrow[N \to +\infty]{\mathbb{P}}\ 0$.\\
\end{center}
So we may conclude.\\
ii) First, we use that $\sqrt{N}(\hat{\hat{\alpha}}_{N}-\alpha^{*})=\sqrt{N}(\hat{\hat{\alpha}}_{N}-\hat{\alpha}_{N})+\sqrt{N}(\hat{\alpha}_{N}-\alpha^{*})$. By Theorem 3, $\sqrt{N}(\hat{\alpha}_{N} -\alpha^{*}) \xrightarrow[N\to + \infty]{\mathcal{L}} \mathcal{N}(0,\Gamma_{\alpha^{*}})$ with $\Gamma_{\alpha^{*}}$ defined in (\ref{eq:7}). It remains to prove that $\sqrt{N}(\hat{\hat{\alpha}}_{N} -\hat{\alpha}) \xrightarrow[N\to + \infty]{\mathbb{P}} 0$.\\
Using the same arguments as in the proof i), we have\\
\begin{equation}
\mathbb{P}(\sqrt{N}\underset{\alpha\in\Theta}{\sup}|\hat{M}_{N}(\alpha)-M_{N}(\alpha)|\geq \epsilon)\leq \mathbb{P}(K\sqrt{N}\underset{\alpha\in\Theta}{\sup}|X_{(j+1)}-X_{(j)}|\geq \epsilon)
\label{eq:lemm}
\end{equation}
The right hand side of (\ref{eq:lemm}) converges to $0$ by Lemma 1. This implies that $\sqrt{N}(\hat{\hat{\alpha}}_{N} -\hat{\alpha}) \xrightarrow[N\to + \infty]{\mathbb{P}} 0$.\\
\end{proof}
\section{Perspectives and conclusion}
One of the main quality of our approach is that it is easy to implement and execute. The cost function being simple, we use a BFGS algorithm to find the optimal parameters, and because of the regularity of curves we deal with, the initial points for optimization can be easily defined.\\
Furthermore, the search of the coordinate of the reference curve which is sent to the coordinate of the curve to fit can be easily implemented with a simple value search.\\
Besides, the deformation parameters can be exploited through an explainable model such as the linear model used in the real world problem.\\
It seems that the deformation model is robust if the noise is controlled. An interesting extension of this work would be to study what is going on when the reference curve is noisy. A generalization of this work to less regular functions would be worthwhile. Finally, it would be interesting to include in the model a way to handle discontinuities in order to reduce the dimension and have a more global representation of the deformation.\\
\bibliographystyle{spmpsci}
|
\subsubsection*{\bibname}}
\bibliographystyle{apalike}
\usepackage{amsmath, amssymb, amsthm}
\usepackage{algorithm}
\usepackage{algorithmicx}
\usepackage{algpseudocode}
\usepackage{enumerate}
\usepackage{booktabs}
\usepackage{graphicx}
\usepackage{bm}
\newcommand{\operatorname{Normal}}{\operatorname{Normal}}
\newcommand{\mathbb R}{\mathbb R}
\newcommand{\bm 0}{\bm 0}
\newcommand{\mathrm I}{\mathrm I}
\newcommand{\stackrel{\text{iid}}{\sim}}{\stackrel{\text{iid}}{\sim}}
\newcommand{\stackrel{\text{ind}}{\sim}}{\stackrel{\text{ind}}{\sim}}
\newcommand{\mathcal T}{\mathcal T}
\newcommand{\mathcal M}{\mathcal M}
\newcommand{\mathcal L}{\mathcal L}
\newcommand{\operatorname{Uniform}}{\operatorname{Uniform}}
\newcommand{\operatorname{Exponential}}{\operatorname{Exponential}}
\newcommand{\operatorname{Cauchy}}{\operatorname{Cauchy}}
\newcommand{\operatorname{Dirichlet}}{\operatorname{Dirichlet}}
\newcommand{\supp}[1]{^{(#1)}}
\newcommand{\operatorname{DP}}{\operatorname{DP}}
\newcommand{\operatorname{Categorical}}{\operatorname{Categorical}}
\newcommand{\operatorname{Var}}{\operatorname{Var}}
\newcommand{\text{TP}}{\text{TP}}
\newcommand{\text{FP}}{\text{FP}}
\newcommand{\text{FN}}{\text{FN}}
\newcommand{\text{Prec}}{\text{Prec}}
\newcommand{\text{Rec}}{\text{Rec}}
\RequirePackage[pdfusetitle, colorlinks,citecolor=blue,urlcolor=blue]{hyperref}
\begin{document}
\twocolumn[
\aistatstitle{Interaction Detection with Bayesian Decision Tree Ensembles}
\aistatsauthor{ Junliang Du \And Antonio R. Linero}
\aistatsaddress{ Department of Statistics, Florida State University}
]
\begin{abstract}
Methods based on Bayesian decision tree ensembles have proven valuable in constructing high-quality predictions, and are particularly attractive in certain settings because they encourage low-order interaction effects. Despite adapting to the presence of low-order interactions for prediction purpose, we show that Bayesian decision tree ensembles are generally anti-conservative for the purpose of conducting interaction detection. We address this problem by introducing Dirichlet process forests (DP-Forests), which leverage the presence of low-order interactions by clustering the trees so that trees within the same cluster focus on detecting a specific interaction. We show on both simulated and benchmark data that DP-Forests perform well relative to existing interaction detection techniques for detecting low-order interactions, attaining very low false-positive and false-negative rates while maintaining the same performance for prediction using a comparable computational budget.
\end{abstract}
\section{INTRODUCTION}
In many scientific problems, a primary goal is to discover structures which allow the problem to be described parsimoniously. For example, one may wish to find a small subset of candidate variables that are predictive of a response of interest; this structure is referred to as \emph{sparsity}. Another structure is \emph{interaction} (or \emph{additive}) structure. An extreme case of additive structure is a generalized additive model (see, e.g., \citealp{hastie2017generalized}), where the effects of the predictors combine additively without any interactions. Teasing out additive structures can be valuable because it can substantially simplify the interpretation of a model. For example, if a given predictor does not interact with other predictors then it can be interpreted in isolation without reference to the values of other predictors. When predictors do interact, interpretation of the interactions is typically simplified whenever the interactions are of low-order. We consider the nonparametric regression problem $Y_i = f_0(X_i) + \epsilon_i$, $\epsilon_i \sim \operatorname{Normal}(0,\sigma^2)$,
where $Y_i$ is a response of interest and $X_i \in \mathbb R^P$ is a vector of predictors, however the methods we develop here can be easily extended to many other settings. The variables $x_j$ and $x_k$ are said to \emph{interact} if $f_0(x)$ cannot be written as $f_0(x) = f_{0\backslash j}(x) + f_{0\backslash k}(x)$ where $f_{0\backslash j}$ and $f_{0\backslash k}$ do not depend on $x_j$ and $x_k$ respectively. One can define higher order interactions similarly: a group of $K$ variables is said to have a $K$-way interaction if $f_0(x)$ cannot be decomposed as a sum of $K$ or fewer functions, each of which depends on fewer than $K$ of the variables.
Methods which estimate $f_0(x)$ using an ensemble of Bayesian decision trees have proven useful in a number of statistical problems. Beginning with the seminal work of \citet{chipman2010bart}, Bayesian additive regression trees (BART) have been successfully applied in a diverse range of settings including survival analysis \citep{sparapani2016nonparametric}, causal inference \citep{hahn2017bayesian}, variable selection in high dimensional settings \citep{linero2016bayesian, bleich2014variable}, loglinear models \citep{murray2017log}, and analysis of functional data \citep{starling2018functional}. A key motivating factor for the use of BART is precisely that it is designed to taking advantage of low-order interactions in the data. Indeed, \citet{linero2017abayesian} and \citet{rockova2017posterior} illustrate theoretically that the presence of low-order interactions is precisely the type of structure which BART excels at capturing. Hence BART appears to be an ideal tool for extracting low-order and potentially non-linear interactions.
Surprisingly, we show that, despite the ability of BART to capture low-order interactions for \emph{prediction} purposes, it is nonetheless not suitable for conducting fully-Bayesian inference for the \emph{selection} task of interaction detection. When taken at face value as a Bayesian model, we show empirically that BART generally leads to the detection of spurious interaction effects. This is not contradictory because optimal prediction accuracy is generally \emph{not} sufficient to guarantee consistency in variable selection (see, e.g., \citealp{wang2007tuning}).
We discuss the general problem which leads to the detection of spurious interactions; while this development is couched in the BART framework, we believe that the fundamental issues also occur for other decision tree ensembling methods. Specifically, the problem is that there is no penalty associated to including spurious interaction terms in the model.
We then introduce a suitable modification to the BART framework which addresses this problem and allows BART detect interactions in a fully-Bayesian fashion. We accomplish this by clustering the trees into non-overlapping groups. Intuitively, the shallow trees comprising each cluster work together to learn a single low-order interaction. To bypass the need to specify the number of clusters, we induce the clustering through a Dirichlet process prior \citep{ferguson1973}. We refer to the ensemble constructed in this fashion as a Dirichlet Process Forest (DP-Forest).
\subsection{A Simple Example}
\label{sec:a-simple-example}
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{./figure/Rplot06}
\caption{The interaction structure detected in the example from Section~\ref{sec:a-simple-example}. ``Truth'' denotes the true interaction structure in the example.}
\label{fig:interaction-graph}
\end{figure}
To motivate the problem, we consider a simulated data example of \citet{vo2016sparse}. This example takes $P = 100$, $N = 100$, $X_i \sim \operatorname{Normal}(\bm 0, 0.02\,\mathrm I)$, and
$f_0(x) = x_1 + x_2^2 + x_3 + x_4^2 + x_5 + x_1x_2 + x_2 x_3 + x_3x_4$.
We compare the DP-Forest we propose to a variant of BART referred to as SBART \citep{linero2017abayesian} which can accommodate sparsity in variable selection. We also consider the recently proposed iterative random forests algorithm of \citet{basu2018iterative}, selecting interactions whose stability score is higher than 0.5. In Figure~\ref{fig:interaction-graph} we display the interaction structure detected by each method on this data; while we considered only one iteration of this experiment here, these results are typical of replications of the experiment.
Here, SBART detects a spurious edge between $x_2$ and $x_4$. This occurs because BART, despite its fundamentally additive nature, does not include any penalization which discourages unnecessary interactions from being included. On the contrary, BART \emph{expect} interactions to occur between relevant predictions; considering a draw from a BART prior such that $x_2$ and $x_4$ are included in the model, an interaction between these variables is a-priori likely. Adapting Bayesian decision tree ensembles to interaction detection then requires a prior which discourages the inclusion of weak interactions. The iRF similarly detects two spurious interactions and misses a relevant interaction between $x_3$ and $x_4$.
\subsection{Related Work}
\label{sec:related-work}
Recent work has studied the theoretical properties of BART. \citet{linero2017abayesian} and \citet{rockova2017posterior} show that certain variants of BART are capable of adaptively attaining near-minimax-optimal rates of posterior concentration when $f_0$ can be expressed as a sum of low-order interaction terms $f_0(x) = \sum_{v = 1}^V f_{0v}(x)$ with each $f_{0v}(x)$ depending on a small subset of $\mathcal S_v$ of the predictors.
In view this, one might conclude that no modification to BART is needed. This is true if one cares only about the mean integrated squared error $\int (f_0(x) - f(x))^2 \, F_0(dx)$ where $X_i \stackrel{\text{iid}}{\sim} F_0$. Optimal prediction performance, however, does not imply that variable selection and interaction detection are being performed adequately. If $\mathcal S_0$ is the true interaction structure of the data $\mathcal S$ is an estimate of $\mathcal S_0$, then attaining the minimax estimation rate for $f_0$ in terms of prediction error typically only guarantees that $\mathcal S_0 \subseteq \mathcal S$ (not $\mathcal S \subseteq \mathcal S_0$).
Several other methods have been recently proposed in the literature specifically for the task of interaction detection. We offer a non-comprehensive review. For a recent review, see \citet{bien2013lasso}. \citet{lim2015learning} proposed a hierarchical group-lasso which enforces the constraint that the presence of a given interaction implies the presence of the associated main effects; a similar approach is given by \citet{bien2013lasso}. A potential shortcoming of these approaches is that they focus on linear models and allow only pairwise interactions. \citet{radchenko2010variable} propose the VANISH algorithm, which allows for nonlinear effects through the use of basis function expansions, but again limits to pairwise interactions. Several decision-tree based methods have also been proposed. The additive groves procedure of \citet{sorokina2008detecting} uses an adaptive boosting-type algorithm to sequentially test for the presence of interactions between variables after performing a variable screening step. \citet{basu2018iterative} propose the iterative random forest (iRF) algorithm which flags ``stable'' interaction effects as those which appear consistently in many trees in a certain random forest.
\section{BAYESIAN TREE ENSEMBLES}
\label{sec:review}
\subsection{The BART Prior}
Our starting point is the Bayesian additive regression trees (BART) framework of \citet{chipman2010bart}, which treats the function $f_0(\cdot)$ as the realization of a sum of random decision trees
\begin{align*}
f(x) = \sum_{t = 1}^T g(x ; \mathcal T_t, \mathcal M_t),
\end{align*}
where $\mathcal T_t$ denotes the tree structure (including the decision rules) of the $t^{\text{th}}$ tree and $\mathcal M_t = \{\mu_{t\ell} : \ell \in \mathcal L_t\}$ denotes the parameters associated to the leaf nodes; here, $\mathcal L_t$ denotes the collection of leaf nodes of $\mathcal T_t$. Let $[x \leadsto (t,\ell)]$ denote the event that the point $x$ is associated to leaf $\ell$ in tree $t$. The function $g(x; \mathcal T_t, \mathcal M_t)$ then returns $\mu_{t\ell}$ whenever $[x \leadsto (t,\ell)]$ occurs
We follow \citet{chipman2010bart} and specify a branching process prior for the tree structure $\mathcal T_t$.
A sample from the prior for $\mathcal T_t$ is generated iteratively, starting from a tree with a single node of depth $d = 0$; this is made a branch with two children with probability $q(d) = \gamma / (1 + \beta)^d$, and is made a leaf node otherwise. We repeat this process independently for all nodes of depth $d = 1, 2, \ldots$ until all nodes at depth $d$ are leaves. After the structure of the tree is generated, each branch $b$ is associated with a decision rule of the form $[x_j \le C_b]$. The coordinate $j$ used to construct the decision rule is sampled with probability $s_j$ where $s = (s_1, \ldots, s_P)$ is a probability vector. The splitting proportion $s$ will play a key role later as an avenue for inducing sparsity in the regression function. Finally, we generate $C_b \sim \operatorname{Uniform}(L_j, U_j)$ where $(L_1, U_1) \times \cdots \times (L_P, U_P)$ is the hyper-rectangle corresponding to the values of $x$ that lead to branch $b$. We remark that this choice for $C_b$ differs from the scheme used by other BART implementations; we adopt it to simplify the full conditionals we derive in Section~\ref{sec:dirichlet}.
For the prior on $\mathcal M_t$ we set $\mu_{t\ell} \stackrel{\text{iid}}{\sim} \operatorname{Normal}(0, \sigma^2_\mu / T)$ conditional on $\mathcal T_t$ and $\sigma^2_\mu$. By taking the variance to be $\sigma_\mu^2/T$ we ensure that the prior level of signal is constant as $T$ increases. The normal prior is selected for its conjugacy; we note, however, that any prior for $\mu_{t\ell}$ with mean $0$ and variance $\sigma_\mu/T$ leads to the approximation $f(x) \sim \operatorname{Normal}(0, \sigma^2_\mu)$ by the central limit theorem. We fix $\beta = 2$ and $\gamma = 0.95$; we refer readers to \citet{linero2017abayesian} for further details regarding prior specification, and to \citet{chipman2013bayesian} and \citet{linero2017review} for detailed reviews of Bayesian decision tree methods.
\subsection{Leveraging Structural Information}
Several recent developments have extended the BART methodology to take advantage of structural information. \citet{linero2016bayesian} noted that sparsity in $f_0(x)$ can be accommodated automatically by setting $s \sim \operatorname{Dirichlet}(\alpha / P, \ldots, \alpha / P)$. Recall here that $s_j$ denotes the prior probability that, for a fixed branch, coordinate $j$ will be used to construct a split at a that branch. Hence, if $s$ is nearly-sparse with $d$ non-sparse entries, the prior will encourage realizations from the prior to include only the $d$ predictors with non-sparse entries.
\citet{linero2017abayesian} showed that this prior for $s$ induces highly desirable posterior concentration properties; in particular, the posterior of $f(x)$ concentrates at close to the oracle minimax rate if we had known the relevant predictors beforehand.
\begin{figure*}[!t]
\centering
\includegraphics[width=.7\textwidth]{./figure/illustration-1}
\caption{Schematic showing the effect of clustering trees within the ensemble. When $Z_t = 1$ split are constructed with $X_1$, but when $Z_t = 2$ splits are constructed with $(X_2, X_3)$.}
\label{fig:illustration-1}
\end{figure*}
\citet{linero2017abayesian} also introduce the SBART model, which uses soft decision trees \citep{irsoy2012soft} which effectively replace the decision boundaries of BART with smooth sigmoid functions. This allows the SBART model to adapt to the smoothness level of $f(x)$; consequently, if $f_0(x)$ is assumed to be $\alpha$-H\"older, the posterior for the SBART model concentrates around $f_0(x)$ at close to the oracle minimax rate obtainable when the smoothness level is known a-priori. While the methodology we develop applies to the usual BART models, we will use the SBART model with the sparsity-inducing Dirichlet prior in all of our illustrations.
\section{DP-FORESTS}
\label{sec:dirichlet}
The distribution of $(\mathcal T_t, \mathcal M_t)$ in the BART model is parameterized by the splitting proportions $s$, leaf variance $\sigma^2_\mu$, and tree topology parameters $(\gamma, \beta)$. To encourage a small number of low-order interactions, we specify a prior which clusters the trees into non-overlapping groups such that each cluster constructs splits using different subsets of the predictors. A schematic is given in Figure~\ref{fig:illustration-1} with $T = 4$. In this figure we see that the first two trees are dedicated to learning a main effect for $x_1$ while the second two trees are dedicated to learning an interaction between $x_2$ and $x_3$.
We induce a clustering by using tree-specific splitting proportions $s\supp{t} \sim G$ and using a Dirichlet process prior on $G$ \citep{ferguson1973}. Specifically, we let $s\supp{t} \stackrel{\text{iid}}{\sim} G$ conditional on $G$ and let $G \sim \operatorname{DP}(\omega G_0)$ where $G_0$ is a $\operatorname{Dirichlet}(\alpha w_1, \ldots, \alpha w_P)$ distribution and $\omega$ denotes the precision parameter of the Dirichlet process. Using the latent-cluster interpretation of the Dirichlet process (see, .e.g, \citealp{teh2006hierarchical}) this can be approximated by the following generative model:
\begin{enumerate}
\item Draw $\pi \sim \operatorname{Dirichlet}(\omega / K, \ldots, \omega / K)$ for large $K$.
\item Draw $Z_1, \ldots, Z_T \stackrel{\text{ind}}{\sim} \operatorname{Categorical}(\pi)$.
\item Draw $s\supp{1}, \ldots s\supp{K} \stackrel{\text{ind}}{\sim} \operatorname{Dirichlet}(\alpha w_1, \ldots, \alpha w_P)$ where
$\sum_{p=1}^P w_p = 1, w_p \ge 0$.
\item For $t = 1, \ldots, T$, draw $(\mathcal T_t, \mathcal M_t)$ as described in Section~\ref{sec:review} with $s = s\supp{Z_t}$.
\end{enumerate}
The $Z_t$'s cluster trees such that the trees within each group capture a single low-order interaction. Note that the use of the the sparsity inducing prior in step 3 above ensures that each $s\supp{k}$ will be nearly-sparse, and hence the trees with $Z_t = k$ will split on only a small subset of the predictors. The role played by this weight vector $w$ is to encourage a subset of the predictors to appear in multiple \emph{different} interactions. For example, if there are interactions $(X_1, X_2)$ and $(X_2, X_3)$ we do not want to encourage an additional $(X_1, X_3)$ interaction. A large value of $w_2$ allows for this by encouraging $X_2$ to appear in several interactions.
\subsection{Properties of the Prior}
The degree of sparsity within each cluster of trees, as well as the overall number of clusters used, are determined by the hyperparameters $\alpha$ and $\omega$. These hyperparameters are key in determining the interaction structures that the prior favors.
To help anchor intuition we first consider several special cases of the DP-Forests model. First, we consider the behavior of the prior as $\alpha \to 0$ with $\omega$ fixed. In this case, with high probability each $s\supp{t}$ will have only one non-sparse entry. Consequently, each tree in the ensemble will split on at most one predictor. Because the trees are composed additively, this implies that none of the variables interact, and hence the prior concentrates on a sparse generalized additive model (SPAM, \citealp{ravikumar2007spam}). On the other hand, as $\alpha \to \infty$ we see that $s\supp{t} \to (w_1, \ldots, w_P)$ so that the prior reverts to original BART model with splitting proportions given by $(w_1, \ldots, w_P)$ described by \citet{bleich2014variable}.
We can conduct a similar analysis with $\alpha$ fixed and $\omega$ with $K \to \infty$. As $\omega \to \infty$, each tree will be associated to a unique $s\supp{t}$. As $\omega \to 0$, on the other hand, all of the trees share the same $s\supp{t}$ so that the model collapses to the Dirichlet additive regression trees model described by \citet{linero2016bayesian}.
The key difference between BART and a DP-Forest is that, once two variables are included, BART does not penalize interactions. Let $A_i$ and $A_j$ denote the event that variable $i$ and $j$ are included in the model, let $A_{ij}$ denote the event that variables $i$ and $j$ interact, and let $\Pi_{\alpha,\omega}$ denote the joint prior distribution for $\mathcal T_1, \ldots, \mathcal T_T$. We study the prior on the interaction structure by examining the probabilities
\begin{math}
\Lambda(\alpha, \omega) = \Pi_{\alpha, \omega}(A_{ij} \mid A_i \cap A_j),
\end{math}
and
\begin{math}
\Xi(\alpha,\omega) = \Pi_{\alpha,\omega}(A_{ik} \mid A_{ij} \cap A_{kj}).
\end{math}
In words, $\Lambda$ is the probability that $(i,j)$ interact given that both variables are relevant, while $\Xi$ represents the probability that $(i,k)$ interact given that $(i,j)$ and $(k,j)$ interact. Additionally, we examine the relationship between the average number of two-way interactions included in the model and the number of variables included.
\begin{figure}[!t]
\centering
\includegraphics[width=.45\textwidth]{./figure/testing-prior-1}
\caption{Plots of various quantities for $\omega = 0$ (solid, corresponding to SBART) and $\omega = 1$ (dashed) with $P = 5$ and $T = 50$. Left: plot of $\alpha$ against $\Lambda$. Middle: Plot of $\alpha$ against $\Xi$. Right: plot of the number of variables included in the model against the number of interactions.
}
\label{fig:testing-prior}
\end{figure}
Figure~\ref{fig:testing-prior} shows several relationships between these quantities as $\alpha$ varies for both SBART and DP-Forests. We see that $\Lambda$ is quite large for all values of $\alpha$ with SBART, implying that the prior expects any variables included in the model to interact; the trend is decreasing in $\alpha$ only because a larger number of predictors will be included in the model, causing variables to compete for branches in the ensemble. DP-Forests do not encourage the inclusion of interactions, particularly when $\alpha$ is small. Next, we see that $\Xi$ is also uniformly large for SBART. This implies that the prior does not encourage interaction structures like the truth from Figure~\ref{fig:interaction-graph}, while a DP-Forest with a small choice of $\alpha$ does.
\subsection{Default Prior Settings}
A benefit of the BART framework is the existence of default priors which require minimal tuning from users. Where applicable, we do not stray from the defaults recommended in Section~\ref{sec:review}. Specific to DP-Forests, the key parameter controlling the behavior of the model is $\alpha$. On the basis of Figure~\ref{fig:testing-prior} we recommend choosing $\alpha$ to be small; we have found setting $\alpha \sim \operatorname{Exponential}$ with mean $0.1$ to work well. Conversely, in our illustrations the results for the DP-Forest model do not depend strongly on $\omega$, and we set $\omega \sim \operatorname{Exponential}(1)$. This leaves the weight vector $w = (w_1, \ldots, w_P)$ to be specified. In our illustrations, we first run a screening step which removes irrelevant predictors. In principle any method can be used for screening; in our illustrations, we use SBART to screen variables which have posterior inclusion probability below $50\%$, and set $w_j \propto I(\text{$j$ is not screened})$. A more principled alternative is to use another sparsity-inducing prior on $w$ but we do not pursue this strategy here.
\subsection{Computation and Inference}
\label{sec:computation}
Inference for the DP-Forest model can be carried out using a Gibbs sampler with the Bayesian backfitting approach of \citet{chipman2010bart}. The Gibbs sampler operates on the state space $(\{\mathcal T_t, \mathcal M_t, Z_t\}_{t = 1}^T, \{s\supp{k}, \pi_k\}_{k=1}^K, \alpha, \omega, \sigma^2_\mu, \sigma^2)$. We use standard Metropolis-within-Gibbs proposals to update $\mathcal T_t$ and $\mathcal M_t$; see \citet{kapelner2014bartmachine} and \citet{pratola2016efficient} for details.
The parameters $\alpha$, $\omega$, $\sigma^2_\mu$, and $\sigma^2$ can all be updated easily using the slice sampling algorithm of \citet{slicesampling}. Finally, $Z_t, s\supp{k}$, and $\pi$ all have conjugate full-conditional distributions:
\textbf{Full conditional for $\pi$:}
Note that $\pi$ is conditionally independent of all parameters given $(\omega, Z)$. By conjugacy of the Dirichlet distribution to multinomial sampling we have the full conditional $\pi \sim \operatorname{Dirichlet}(\omega/K + m_1, \ldots, \omega/K + m_K)$ where $m_k = \sum_t I(Z_t = k)$.
\textbf{Full conditional for $s\supp{k}$:}
The conjugacy of the Dirichlet prior to multinomial sampling implies a Dirichlet full-conditional when a single $s$ is used. To account for the clustering, we only consider the branches associated to trees with $Z_t = k$, giving the full conditional
$s\supp{k} \sim \operatorname{Dirichlet}(\alpha w_1 + c_1\supp{k}, \ldots, \alpha w_P + c_P\supp{k})$ where $c_j\supp{k}$ is the number of branches associated to cluster $k$ which split on predictor $j$.
\textbf{Full conditional for $Z_t$:}
Let $p(k)$ denote the full conditional for $Z_t$. The term $[Z_t = k]$ comes in only through the factors $\pi_k$ (the prior probability of $Z_t = k$) and $\prod_{j = 1}^P s_j^{(k)c_{tj}}$ where $c_{tj}$ is the number of branches of tree $t$ which split on predictor $j$ (the likelihood of tree $t$ having split on the predictors that it has, give $Z_t = k$). Hence $p(k) \propto \pi_k \prod_{j = 1}^P s_j^{(k)c_{tj}}$.
Putting these pieces together, we arrive at Algorithm~\ref{alg:bayesian-backfitting}, which describes a single iteration of the Gibbs sampler.
\begin{algorithm}[h]
\caption{Bayesian backfitting algorithm} \label{alg:bayesian-backfitting}
\begin{algorithmic}[1]
\For{$t = 1, \ldots, T$}
\State Update $(\mathcal T_t, \mathcal M_t)$ via Metropolis-Hastings.
\State Sample $Z_t \sim p(k), k = 1, \ldots, K$ where
\begin{math}
p(k) \propto \pi_k \prod_{j = 1}^P s_j^{(k)c_{tj}}
\end{math}
and $c_{tj}$ is the number of branches associated to tree $t$ which split on predictor $j$.
\EndFor
\For{$k = 1,\ldots,K$}
\State Sample \(s\supp{k} \sim \operatorname{Dirichlet}(\alpha w_1 + c\supp{k}_1, \ldots, \alpha w_P + c\supp{k}_P)\) where $c\supp{k}_j$ is the number of branches associated to cluster $k$ which split on predictor $j$.
\EndFor
\State Sample $\pi \sim \operatorname{Dirichlet}(\omega / K + m_1, \ldots, \omega/K + m_K)$ where $m_k = \sum_{t = 1}^T I(Z_t = k)$.
\State Sample $(\sigma, \sigma_{\mu}, \alpha, \omega)$ using slice sampling.
\end{algorithmic}
\end{algorithm}
\section{EXPERIMENTS}
We now compare DP-Forests to existing methods on a number of synthetic datasets. We consider the following methods in addition to DP-Forests and SBART.
\textbf{Additive groves}: The additive groves procedure of \citet{sorokina2008detecting}. Because tuning of the additive groves algorithm is compute-intensive, we ran several pilot studies to choose appropriate tuning parameters which perform well for the given simulation settings.
\textbf{Hierarchical group lasso}: The hierarchical group lasso proposed by \citet{lim2015learning} for interaction detection;
we abbreviate this method by HL. This procedure was designed with linearity of $f_0(x)$ in mind. Tuning parameters are selected by cross-validation.
\textbf{Hierarchical group lasso, least squares}: HL is used to \emph{select} the interactions and main effects, while the coefficients are estimated by least squares; we abbreviate this method by HL-LS. Tuning parameters are selected by cross validation.
\textbf{Iterative random forests}: The iterative random forests (iRF) procedure proposed by \citet{basu2018iterative} as implemented in the \texttt{iRF} package on \texttt{CRAN}. We use the default $T = 500$ trees and 10 iterations of the iRF algorithm.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{figure/sim_results_1-1.pdf}
\caption{Barplot of results for interaction detection. The top row gives the average $F_1$ score for each method for detecting interactions. The second row gives the average number of false positive interactions detected. The bottom row gives the average number of false negatives detected. The average for each method is given on each bar.}
\label{fig:sim_results_1_1}
\end{figure}
Our simulation settings are borrowed from several existing works; we do not compare our methods to these other works due to a lack of publicly available software.
\begin{enumerate}
\item[(S1)] \citep{radchenko2010variable} We generate $X_i \sim \operatorname{Uniform}([0,1]^P)$ where $P = 50$, $N = 300$, and $\sigma^2 = 1$. We let $f_0(x)$ be
\begin{align*}
\sqrt{0.5} \bigg[\sum_{v = 1}^V f_v(x) + f_1(x) f_2(x) + f_1(x) f_3(x)\bigg]
\end{align*}
where $f_1(x) = x_1$, $f_2(x) = (1+x_2)^{-1}$, $f_3 = \sin(x_3)$, $f_4(x) = e^{x_4}$, and $f_5(x) = x_5^2$.
Each $f_v(x)$ is further centered and scaled so that $E(f_v(X_i)) = 0$ and $\operatorname{Var}(f_v(X_i)) = 1$.
\item[(S2)] \citep{vo2016sparse} We generate $X_i \sim \operatorname{Normal}(\bm 0, \mathrm I)$ with $N = 100$, $P = 100$, and $\sigma = 0.14$. We let
\begin{math}
f_0(x) = x_1 + x_2^2 + x_3 + x_4^2 + x_5 + x_1x_2 + x_2x_3 + x_3x_4.
\end{math}
\item[(S3)] Same as (S2), but without the interaction effects.
\item[(S4)] \citep{friedman1991multivariate} A common test case for BART, we generate $X_i \sim \operatorname{Uniform}([0,1]^P)$ with $P = 250, N = 250$, and $\sigma^2 = 1$. We set
\begin{math}
f_0(x) = 10 \sin(x_1x_2) + 20(x_3 - 0.5)^2 + 10 x_4 + 5 x_5.
\end{math}
\end{enumerate}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{figure/sim_results_1-2.pdf}
\caption{Barplot of results for detecting main effects.}
\label{fig:sim_results_1_2}
\end{figure}
Each of these scenarios was replicated $100$ times. We evaluate each method according to the average number of false positives (FPs), false negatives (FNs), $F_1$ score, and integrated root-mean squared error $\|f_0 - \widehat f\|_2$. The $F_1$ score is a commonly used measure of overall accuracy that balances false positives against false negatives in variable selection tasks; see, for example, \citet{zhang2015cross}.
Results for interaction detection are given in Figure~\ref{fig:sim_results_1_1}. We omit the results for HL because HL-LS performs uniformly better. Under all simulation settings, DP-Forests perform better than all other methods according to $F_1$ score. SBART is also competitive with other procedures on many of the datasets. As expected, the primary problem with SBART is that it has a relatively large number of false positives, i.e. it is susceptible to detecting spurious interactions. This issue is most pronounced on (S2) and (S3), with SBART detecting between 1.5 and 2 spurious interactions.
Additive groves and iterative random forests generally perform worse than SBART. In addition to having a larger false positives rate, these procedures are also prone to false negatives under simulation (S2). With the exception of (S1), the hierarchical group-lasso (HL-LS) performs worse than the other methods. Under (S1), HL-LS has reasonable performance as each component of $f_0(x)$ can be reasonably well-approximated by the assumed linear model. HL-LS also appears to perform well under (S3); this, however, is due to the fact that HL-LS typically misses several main effects, which is a substantially worse outcome than detecting a spurious interaction. The nonlinearities under (S2) and (S4) also create problems for HL-LS
All methods perform better for detecting the main effects. SBART and DP-Forests give identical results for the main effects due to the use of SBART in screening for DP-Forests. (S1) is the easiest setting, with all methods having very few false negatives and HL-LS the only method having non-negligible false-positives.
Under (S2), the non-Bayesian procedures all have non-negligible false negatives, and iRF and HL-LS are additionally prone to false positives; the story is similar under (S3), with HL-LS performing better in terms of false positives but worse in terms of false negatives. All methods perform well in terms of false positives under (S4), however iRF and HL-LS also suffer from many false negatives.
\begin{figure}[t]
\centering
\includegraphics[width=.5\textwidth]{figure/sim_results_1_rmse-1.pdf}
\caption{Boxplots given the distribution of integrated root mean-squared error for each method for each simulation setting.}
\label{fig:sim_results_1_rmse}
\end{figure}
Results for assessing prediction performance in terms of integrated root mean-squared error (RMSE) are given in Figure~\ref{fig:sim_results_1_rmse}. SBART and DP-Forests perform very similarly in terms of RMSE. All other methods perform substantially worse under all settings. This is likely due to a multitude of factors. First, any false negatives will contribute to poor predictive performance. Second, SBART and DP-Forests are able to take advantage of underlying smoothness in the response function which additive groves and iterative random forests cannot, while HL and HL-LS suffer from an incorrect model specification.
SBART and DP-Forests are competitive in terms of runtime. For example, on a single replicate of (S4), SBART and DP-Forests took 118 seconds and 241 seconds respectively to obtain 40,000 samples from the posterior. By comparison, iRF took 279 second, HL-LS took 91 seconds, and additive groves took 4966 seconds. Additive groves was by far the slowest procedure, due to the fact that recursive feature elimination is used. We conclude that, under these settings, DP-Forests outperform all competitors are a competitive computational budget.
We also consider the publicly available Boston housing dataset of \citet{harrison1978hedonic}. Analysis of the interaction structures present in this dataset was previously undertaken by \citet{radchenko2010variable} and \citet{vo2016sparse}. This dataset consists of $P = 13$ predictors and $N = 506$ neighborhoods, and a continuous response corresponding to the median house value in a given neighborhood.
\begin{table}
\centering
\begin{tabular}{lc}
\toprule
Method & RMSE \\
\midrule
DP-Forests & 1.00\\
iRF & 1.22\\
HL & 1.18\\
Additive Groves & 1.16\\
\bottomrule
\end{tabular}
\caption{Cross-validation estimate of root mean-squared prediction error on the Boston housing dataset normalized by the RMSE of the DP-Forest.}
\label{tab:boston-rmspe}
\end{table}
We compare the methods in terms of goodness-of-fit, which is evaluated using a 5-fold cross validated estimate of root mean squared prediction error. Results are given in Table~\ref{tab:boston-rmspe}. For prediction, the DP-Forest and SBART outperform the competing methods.
The DP-Forest includes most of the predictors in the model. This can be contrasted with the fit of a sparse additive model (SPAM) \citet{ravikumar2007spam} and the fit of the VANISH model reported by \citet{radchenko2010variable}, which include only a small number of predictors. Like the VANISH algorithm, the DP-Forest selects one interaction: there is strong evidence of an interaction between \texttt{DIS} (distance to an employment center in Boston) and \texttt{LSTAT} (the proportion of individuals in a neighborhood who are lower-status). This interaction was highly stable, and was selected by every fit to the data during cross-validation; additionally, this interaction was selected by additive groves in 4 out of 5 folds during cross-validation. Interestingly, this interaction was reportedly \emph{not} selected by VANISH, which instead selects an interaction between the variables \texttt{NOX} (nitrus-oxide concentration) and \texttt{LSTAT}.
\begin{figure}[t]
\centering
\includegraphics[width = .5\textwidth]{./figure/boston_results-1}
\caption{Graphical summary of the effect of distance \texttt{DIS} on \texttt{MEDV} for various values of \texttt{LSTAT}.}
\label{fig:boston_results}
\end{figure}
Figure~\ref{fig:boston_results} gives a visualization of the \texttt{LSTAT}-\texttt{DIS} interaction. To summarize the interaction we use a ``fit-the-fit'' strategy and fit a generalized additive model to the fitted-values of the DP-Forest with a thin plate spline term for the interaction \citep{wood2003thin}. The plot then displays the \texttt{LSTAT}-specific effect of \texttt{DIS} for the $10^{\text{th}}, 20^{\text{th}}, \ldots, 90^{\text{th}}$ quantiles of \texttt{LSTAT}. This GAM nearly reproduces the fitted values from the DP-Forest and is easier to visualize.
We see in Figure~\ref{fig:boston_results} a clear interaction between \texttt{DIS} and \texttt{LSTAT}. Intuitively, one expects that the closer a neighborhood is to an industry center the more expensive the housing will be. This is correct for areas with fewer lower-status individuals; however, this trend does not hold when there is a higher percentage of lower-status individuals. We remark also that the data is well supported near $0$ for all values of \texttt{LSTAT}, so that this behavior is unlikely to be due to extrapolation, though extrapolation may be an issue for large values of both \texttt{LSTAT} and \texttt{DIS}.
\section{DISCUSSION}
\label{sec:discussion}
We have introduced Dirichlet process forests (DP-Forests) and applied them to the problem of interaction detection. We demonstrated on both synthetic and real data that DP-Forests lead to improved interaction detection. Additionally, we demonstrated that DP-Forests are highly competitive with commonly used machine learning techniques for detecting low-order interactions.
There are a number of modifications one might make to improve performance further. One possibility is to allow $\sigma_{\mu}$ to also vary by mixture component. This would allow different mixture components to have different signal levels; for example, under simulation (S4), we would expect that a smaller value of $\sigma^2_{\mu}$ is appropriate for the mixture component responsible for $x_5$ relative to $x_4$. The proposed DP-Forests model captures this feature only indirectly through the number of trees assigned to each mixture component.
Additionally, it would be interesting to quantify the improvement in performance of DP-Forests over SBART theoretically. It is unknown whether SBART is variable-selection consistent, and establishing theoretically that DP-Forests are consistent for interaction detection while SBART is not remains an open problem.
|
\section{INTRODUCTION}
Despite the tremendous advances in modern computational methods and availability of efficient and powerful hardware, chemical applications of quantum mechanics are still heavily limited by the system size. The need to overcome this limitation becomes more pressing with the increasing interest in chemical properties of large systems, stimulated by research in chemical biology and materials science. One common approach to this problem involves breaking down the system into fragments for which properties can be obtained at a lower computational cost. The total properties of the system are then obtained from the properties of the fragments corrected for the inter-fragment interactions. \cite{L68,KSANU99,KIANU99} Among modern embedding methods that can be used for this purpose \cite{C91,WW93,cr2,HPC11,HC11,SC16}, those that use the electronic density as the main variable \cite{JN14,NW14,WSZ15} have the advantage of simplicity and can be directly connected to Kohn-Sham DFT \cite{KS65}; and among density-based embedding methods, Partition Density Functional Theory (PDFT) \cite{wasserman_2010, NW14} has the advantage of producing localized fragment densities that facilitate the connection to traditional chemical concepts. \cite{GPL03,jctc2018}\\
PDFT is conceptually analogous to Kohn-Sham DFT. In PDFT, a system of interacting fragments is uniquely mapped onto a fictitious system of non-interacting fragments in a global (\textit{i.e.} same for all fragments) external potential. This partition potential, $v_p(\mathbf{r})$, is unique for a given set of fragments. Although there are infinitely many ways to partition a system, chemically relevant fragments are usually the most natural choice of partitioning as they enable meaningful chemical properties to be calculated. \cite{pdt} In the case of the formic acid dimer\cite{pccp2016,jpca2017,jctc2018_2} analyzed here, we choose two monomers as fragments and label them \textit{left} ($L$) and \textit{right} ($R$). \\
As opposed to most other density embedding methods that minimize the total energy \cite{JN14,KSGP15}, in PDFT, we search for a minimum of the sum of fragment energies:
\begin{equation}
E_{f}[n_{L}(\mathbf{r}),n_{R}(\mathbf{r})]= E_{L}[n_{L}(\mathbf{r})]+E_{R}[n_{R}(\mathbf{r})]
\label{eq:fragE}
\end{equation}
subject to the constraint that fragment densities sum to the total density of the system at each point in space, $\mathbf{r}$:
\begin{equation}
n_{tot}(\mathbf{r})=n_{L}(\mathbf{r})+n_{R}(\mathbf{r})
\label{eq:totn}
\end{equation}
This constrained optimization can be replaced by the unconstrained optimization of the following functional:
\begin{equation}
G[n_{L}(\mathbf{r}),n_{R}(\mathbf{r}),v_p(\mathbf{r})]=E_{f}[n_{L}(\mathbf{r}),n_{R}(\mathbf{r})]+\int d\mathbf{r}v_p(\mathbf{r})(n_{tot}(\mathbf{r})-n_{L}(\mathbf{r})-n_{R}(\mathbf{r}))
\label{eq:G}
\end{equation}
In this formalism, the partition potential $v_p(\mathbf{r})$ appears as the Lagrange multiplier that controls the density constraint. More generally, PDFT can be formulated for varying non-integer fragment occupations. \cite{pdt, cr1, cr2} In this work, however, we fix occupations to the ones of isolated fragments. This simplification increases the efficiency of the method as optimization with respect to occupation numbers is not needed. Previous work on simpler systems suggests that occupations usually lock to integers when fragments have similar electronic structures. Since we work with neutral dimers with small dipole moments, we choose neutral fragments and focus attention on monomer density deformations:
\begin{equation}
\Delta n_i(\mathbf{r})=n_i(\mathbf{r})-n_i^0(\mathbf{r})
\label{e:dn}
\end{equation}
where $i$ stands for either $L$ or $R$.
Among the many types of intermolecular interactions, hydrogen-bonding is of particular interest because hydrogen bonds (HBs) are known to be responsible for stabilization of various chemical systems, with direct implications in a wide range of scenarios, from the life-supporting properties of liquid water \cite{w4, w5, w6, w7} to the tertiary structures of biomolecules in charge of storing and replicating genetic information. \cite{bio_book} Although the very nature of hydrogen bonding is not without controversy \cite{wein1,aw1,grabowski, buckingham}, several types of HBs are recognized in the literature. \cite{gilli} Of particular interest to this work are the conventional \textit{primary hydrogen bonds}, where a hydrogen atom sits between two electronegative atoms, and the non-conventional, \textit{secondary hydrogen bonds}, where a proton is donated from a non-polar C--H bond.\\
Although PDFT is particularly well suited to study molecular clusters, these systems are challenging because intermolecular interaction energies in clusters are significantly smaller than energies associated with formal bonds. Individual molecules in clusters retain their chemical identities to a large degree and require carefully constructed partition potentials to account for the comparatively weak interactions. PDFT was recently successfully applied to water dimers, \cite{w2-pdft} where it was shown that the partition potential and PDFT densities can be used to describe the mechanism of hydrogen-bond formation. \\
The question we address in this paper is whether the partition potential has transferable features corresponding to particular types of HBs. Finding transferability would imply that the partition potential around a hydrogen bond in one molecule could be used as a starting point to calculate approximate interaction energies in other molecules with similar HBs, an appealing prospect for computational chemistry. The formic acid dimers, (FA)$_2$, are ideal systems to investigate this question because their four lowest-energy conformations have two types of primary HBs (C=O$\cdots$H--O and H--O$\cdots$H--O ) and two types of secondary HBs (C=O$\cdots$H--C, H--O$\cdots$H--C). Is $v_p(\mathbf{r})$ in the vicinity of a primary HB in one of these four dimers a good approximation to $v_p(\mathbf{r})$ for a primary HB in a different dimer? What about the same question for secondary HBs? Previous work on one-dimensional model systems \cite{ZW10} taught us that the transferability of PDFT \textit{densities} was about an order of magnitude higher than that of real-space partitioning schemes, so it is reasonable to expect transferable features in the underlying partition potentials. However, we find that the answer is \textit{no} in both cases (primary and secondary), contrary to naive intuition. Conversely, monomer density deformations \textit{do} have specific features that can be used to distinguish between different types of HBs. \\
\section{LOW-ENERGY STRUCTURES OF FORMIC-ACID DIMERS}
Despite being the smallest carboxylic acid, the conformational space for the formic acid dimers is notoriously rich, with a considerable number of structures already experimentally detected. \cite{lundell} Farf\'an \textit{et al.} \textit{et al.} calculated a set of 21 well-defined minima in MP2/6-311++G$(d,p)$ potential energy surface (PES). \cite{fad} For our work, we selected four lowest energy motifs from this set (shown in Figure\ref{f:structures}). We reoptimized their geometries and confirmed the found stationary points are true minima by frequency calculations using the B3LYP XC functional and Dunning's aug-cc-pVTZ basis set. PDFT calculations were performed over the resulting geometries. B3LYP and PW91 were used as XC functionals in the construction of the effective potentials. The partition potential was expanded using aug-cc-pVTZ basis set. PW91 has been shown to be useful for the evaluation of intermolecular interactions in hydrogen-bonded systems. In particular, the dimers of water and formic acid for which PW91 computed interaction energies showed only slight changes with respect to CCSD(T). \cite{tsuzuki} It has also been concluded that large basis sets for $v_p\left(\mathbf{r}\right)$ lead to accurate total densities. \cite{w2-pdft} All calculations were carried out using the NWChem package. \cite{nwchem} \\
\begin{figure}[h]
\centering
\includegraphics[scale=0.8]{fig1.pdf}
\caption{Lowest energy dimers of formic acid from the MP2/6--311++G($d,p$) Potential Energy Surface. \cite{fad} Both monomers are in the $anti$ conformation. BEs are the CCSD(T)/6--311++G($d,p$) binding energies calculated as the difference between the given dimer and the isolated fragments. Relative energies with respect to the global minimum are shown. All energies in kcal/mol and corrected for zero-point vibrational energies. The right monomer ($R$), which simultaneously acts as a donor and acceptor of hydrogen bonds, is shown in the same perspective in all cases. The origin of coordinates is placed at the oxygen atom in the left ($L$) monomer. The two conformations of the water dimer discussed in the text are presented.}
\label{f:structures}
\end{figure}
\section{RESULTS AND DISCUSSION}
\subsection{Energy Analysis}
PDFT describes fragment interactions by means of two energy contributions: the preparation energy, $E_{prep}$, and the interaction energy among the "prepared" fragments (partition energy, $E_p$). The preparation energy is defined as the energy required to distort the density of isolated fragments to the density of fragments within the dimer, satisfying the condition of eq. \ref{eq:totn} and minimizing $G$ of eq. \ref{eq:G}:
\begin{equation}
E_{prep}=E_{f}[n_{L}(\mathbf{r}),n_{R}(\mathbf{r})]-(E_{L}[n_{L}^0(\mathbf{r})]+E_{R}[n_{R}^0(\mathbf{r})]),
\label{eq:eprep}
\end{equation}
where $n_i^0(\mathbf{r})$ is the density of an isolated $i$-fragment. We can also identify the preparation energy of a fragment $i$:
\begin{equation}
E_{prep}^{i} = E_{i}[n_{i}(\mathbf{r})]-E_{i}[n_{i}^0(\mathbf{r})].
\label{eq:eprepi}
\end{equation}
The partition energy describes the interaction of these distorted fragments and is defined as:
\begin{equation}
E_{p} = E_{Dimer}[n_{L}(\mathbf{r}),n_{R}(\mathbf{r})]-E_{f}[n_{L}(\mathbf{r}),n_{R}(\mathbf{r})].
\label{eq:epart}
\end{equation}
$v_{p}(\mathbf{r})$ is the functional derivative of $E_{p}$ with respect to the total density. In our calculations, however, $v_p(\mathbf{r})$ and $E_p$ are calculated separately: $v_p(\mathbf{r})$ through a numerical inversion and $E_{p}$ by eq. \ref{eq:epart} from the resulting energies. The inversion algorithm is based on computing the fragment density response, $\chi_f(\mathbf{r,r^\prime})$ at each step and updating $v_p(\mathbf{r})$ through: \cite{w2-pdft,WY03}
\begin{equation}
\delta v_p(\mathbf{r}) = \int d\mathbf{r^\prime} \chi_f(\mathbf{r,r^\prime})^{-1} \left( \sum_i n_i(\mathbf{r^\prime}) - n_f(\mathbf{r^\prime}) \right),
\label{eq:WY}
\end{equation}
where $n_f(\mathbf{r})=n_L(\mathbf{r})+n_R(\mathbf{r})$. The binding energy, $BE$, is simply the sum of the two terms:
\begin{equation}
BE=E_{p}+E_{prep} = E_{Dimer}[n_{L}(\mathbf{r}),n_{R}(\mathbf{r})]-(E_{L}[n_{L}^0(\mathbf{r})]+E_{R}[n_{R}^0(\mathbf{r})])
\label{eq:be}
\end{equation}
Since all four dimers have their atoms lie on a plane, it is most convenient to visualize the densities and potentials at values of $\mathbf{r}$ belonging to this plane. \\
Table \ref{energies} lists relevant quantities extracted from the PDFT calculations using both B3LYP and PW91 functionals. The experimental BE for D$_1$ is -59.5 $\pm$ 0.5 kJ/mol (2.27 $\times$ 10$^{-2}$ a.u.) \cite{be_fad_exp}. The BEs computed with eq. \ref{eq:be} yield 2.90 $\times$ 10$^{-2}$ a.u. (B3LYP) and 3.14 $\times$ 10$^{-2}$ a.u. (PW91). We note that these values do not include zero-point vibrational energies and thermal corrections. \\
\begin{sidewaystable}
\centering
\caption{Relevant energies (a.u.) from PDFT calculations (B3LYP, PW91, $v_p\left(\mathbf{r}\right)$ with the aug--cc--pVTZ basis set) on the lowest energy formic acid dimers (Figure \ref{f:structures}). The energy for the isolated $anti$ formic acid monomer is $E_{L}^0=E_{R}^0=$-189.846 a.u. Analogous results for the water dimer in two different symmetries\cite{w2-pdft} are also shown for comparison. All energies in kcal/mol.}
\label{energies}
\begin{tabular}{|lllllll|} \hline \hline
System & $BE$ & $E_{prep}^{D_i}$ (Eq. \ref{eq:eprep}) & $E_{prep}^{L}$ & $E_{prep}^{R}$ & $E_{p}$ & Intermolecular Interactions \\
\hline \hline
D$_1$ \tiny{(B3LYP)} & -18.19 & 7.14 & 3.57 (50\%) & 3.57 (50\%) & -25.33 & \tiny C=O $\cdots$ H--O $\qquad$ C=O $\cdots$ H--O\\
D$_1$ \tiny{(PW91)} & -19.70 & 6.99 & 3.49 (50\%) & 3.49 (50\%) & -26.69 & \tiny C=O $\cdots$ H--O $\qquad$ C=O $\cdots$ H--O\\
D$_2$ \tiny{(B3LYP)} & -9.61 & 4.67 & 2.56 (55\%) & 2.11 (45\%) & -14.28 & \tiny C=O $\cdots$ H--O $\qquad$ C=O $\cdots$ H--C \\
D$_2$ \tiny{(PW91)} & -10.55 & 4.76 & 2.76 (58\%) & 2.00 (42\%) & -15.31 & \tiny C=O $\cdots$ H--O $\qquad$ C=O $\cdots$ H--C \\
D$_3$ \tiny{(B3LYP)} & -6.87 & 2.08 & 1.14 (55\%) & 0.94 (45\%) & -8.95 & \tiny C=O $\cdots$ H--O $\qquad$ H--O $\cdots$ H--O \\
D$_3$ \tiny{(PW91)} & -7.84 & 2.08 & 1.11 (54\%) & 0.97 (46\%) & -9.92 & \tiny C=O $\cdots$ H--O $\qquad$ H--O $\cdots$ H--O \\
D$_4$ \tiny{(B3LYP)} & -5.01 & 1.71 & 1.06 (62\%) & 0.66 (38\%) & -6.72 & \tiny C=O $\cdots$ H--C $\qquad$ H--O $\cdots$ H--O\\
D$_4$ \tiny{(PW91)} & -5.62 & 1.75 & 1.13 (65\%) & 0.62 (35\%) & -7.37 & \tiny C=O $\cdots$ H--C $\qquad$ H--O $\cdots$ H--O\\
W$_2$ C$_s$ \tiny{(B3LYP)} & -4.45 & 1.86 & 0.88 (47\%) & 0.98 (53\%) & -6.31 & \tiny H--O $\cdots$ H--O \\
W$_2$ C$_{2h}$ \tiny{(B3LYP)} & -3.09 & 0.42 & 0.21 (50\%) & 0.21 (50\%) & -3.50 &\tiny H--O $\cdots$ H--O $\qquad$ H--O $\cdots$ H--O\\
\hline
\end{tabular}
\end{sidewaystable}
An inventory of intermolecular interactions in the four formic acid dimers studied here is provided in the rightmost column of Table \ref{energies}. D$_1$ and D$_3$ are stabilized by primary hydrogen bonds only while D$_2$ and D$_4$ include one secondary hydrogen bond each. D$_1$ exhibits two equivalent primary hydrogen bonds where the hydroxyl group in one monomer donates a proton to the carbonyl group of the other (see Figure {\ref{f:structures}). D$_2$ and D$_4$ have two types of contacts: a secondary C=O $\cdots$ H--C hydrogen bond and a C=O $\cdots$ H--O (D$_2$) and H--O $\cdots$ H--O (D$_4$) primary HB. D$_3$ has two non--equivalent primary hydrogen bonds, where $R$ simultaneously acts as donor and acceptor in two different functional groups, the O--H bond in $L$ acts as donor and acceptor of both HBs, freeing the carbonyl group in $L$ of intermolecular interactions. We note that although secondary hydrogen bonds are typically considered weaker than primary hydrogen bonds \cite{gilli}, the overall stability of the dimers is not correlated with the primary or secondary nature of the HBs. For example, D$_2$ is lower in energy than D$_3$, even though D$_2$ has one primary and one secondary HBs and D$_3$ has two primary HBs. This lack of correlation extends to the number of hydrogen bonds, as seen for example in the two water dimers listed in Table \ref{energies}. \\
Table \ref{energies} also lists preparation energies for each dimer and its components according to eqs. \ref{eq:eprep} and \ref{eq:eprepi}. It is clear from eqs. \ref{eq:fragE} and \ref{eq:eprep} that $E_{prep}$ is always positive. We also expect larger values of $E_{prep}$ for fragments that are more distorted relative to their isolated states. $E_{prep}$ decreases in going from D$_1$ to D$_4$. Fig. \ref{f:dens_diff_plots}, which shows the densities on the plane of two monomers, makes it obvious that this decrease corresponds to the decrease in the total density deformation. \\
There are characteristic deformation patterns for the primary and secondary bonds, as shown in Fig. \ref{f:dens_diff_plots}. The O atom of the H-donating O--H group has a significant density increase along the approximate direction of the HB in a dumbbell-like shape. The O atom of the acceptor has a density decrease of similar shape and direction. The H atom of the O--H group also has some density deficiency around it. The secondary bond pattern is very similar (note that the O--H donor is now replaced with C--H), but the deformation is smaller in magnitude and is more disperse. These observations suggest that the stronger intermolecular bonds require larger deformation of the original wavefunctions of the fragments, a result that may appear obvious to many chemists, but can not be quantified without a rigorous definition of fragments within a molecule. This is also consistent with the orbital interaction picture where it is generally thought that the gain of electron density in the $\sigma_{\rm O-H}^{\star}$ region of the donor and the simultaneous loss of charge in the region associated to the O atom of the acceptor is responsible for the formation of a hydrogen bond. \cite{wein1, aw1, fad} Although fragment occupation numbers remain constant in the present PDFT implementation, fragment densities are indeed distorted; these distortions are linked to the charge transfer within fragments, provided by the orbital interactions. \\
The fragment preparation energies can be analyzed further. Since in D$_1$ both monomers are the same, their preparation energies are identical. In D$_2$ and D$_4$ the left monomer acts as a donor of a secondary hydrogen bond. In those cases, $E_{prep}^{L}$ is significantly larger than $E_{prep}^{R}$, even though the density deformation reaches higher values in the $R$-monomer. The $L$-monomer has a more delocalized density deformation. This imbalance can also be attributed to the fact that weaker secondary HBs require smaller preparation energies. \cmmnt{We attribute such imbalance to the reaccommodation of charge transferred to $\sigma^{\star}_{\rm C-H}$ orbitals in $L$, which seems to require larger energies than the ones needed to distort the geometries of $R$ to incorporate the transferred charge to the $\sigma^{\star}_{\rm O-H}$ orbitals.}In D$_3$, the energy needed to prepare $L$ is larger than the energy needed to prepare $R$ because of the double donor/acceptor function of the O--H group in $L$. \\
Partition energies, $E_p$, also shown in Table \ref{energies}, are always negative and their magnitudes are correlated with the corresponding $E_{prep}$'s. That $E_p$ is negative can be proven from the variational principle, but the observed correlation with $E_{prep}$ (\textit{i.e.} that $E_{prep}$ decreases as $E_p$ decreases) cannot. As predicted by the analysis of Fig. \ref{f:dens_diff_plots}, larger preparation energies lead to larger partition energies, which is seen for all dimers in Table \ref{energies}. This trend is followed not only by FA dimers but by all other systems we have studied so far. Whereas this observation seems obvious, a hard proof is missing. \\
\begin{figure}
\centering
\includegraphics[scale=0.6]{figure_den_diff_monomers.pdf}
\caption{Density distortions, eq. \ref{e:dn}, On the molecular plane for various dimer configurations (in a.u.). For clarity, atom positions are indicated by hollow circles centered at atomic positions and bond lines are omitted. Upper panels correspond to the left monomer; lower panels correspond to the right monomers.}
\label{f:dens_diff_plots}
\end{figure}
It was noticed in previous work \cite{w2-pdft} that the character of chemical interactions (formal bonds, long range, van der Waals, etc.) appears to be related to the magnitude of the preparation energy. Thus, as expected, preparation energies in Table \ref{energies} suggest a direct correlation with binding energies. Table \ref{t:prep_energies} lists recalculated total $E_{prep}$ for a set of diatomic and polyatomic systems already available in the literature, as well as our results for the formic acid dimers. Nicely, the preparation energies for (FA)$_2$ fall in the same range as that of the water dimer. They lead to interaction energies that are stronger than van der Waals contacts but weaker than ionic and covalent bonds. In hydrogen bonding, the distance separating the two moieties dictates the strength of the interaction. This tendency can be seen in Table \ref{t:prep_energies}.
\begin{table}[h]
\caption{Total preparation energies for different systems. All PDFT calculations using B3LYP/aug--cc--pVTZ with an expansion of $v_p(\mathbf{r})$ in the same basis set. $R_{\rm O-O}$ and $R_{\rm C-O}$ are the distances between oxygen atoms in primary hydrogen bonds and between carbon and oxygen atoms in secondary hydrogen bonds, respectively.}
\label{t:prep_energies}
\begin{center}
\begin{tabular}{|c|cc|}
\hline
\hline
System & Distance (\AA{}) & $E_{prep}$ (kcal/mol)$^a$ \\
\hline
LiH (neutral fragments)& 1.59 & 34.76\\
LiH (ionic fragments) & 1.59 & 23.59 \\
H$_2$ & 0.74 & 12.76\\
D$_1$& $R_{\rm O-O} =$ 2.67 & 7.14 \\
D$_2$& $R_{\rm O-O} =$ 2.73, $R_{\rm C-O} =$ 3.13 & 4.67 \\
D$_3$& $R_{\rm O-O} =$ 2.73, 2.89 & 2.08 \\
$C_s$ Water Dimer& $R_{\rm O-O} =$ 2.86 & 1.86 \\
D$_4$& $R_{\rm O-O} =$ 2.90, $R_{\rm C-O} =$ 3.35 & 1.71 \\
$C_{2h}$ Water Dimer& $R_{\rm O-O} = $ 2.76 & 0.42 \\
He$_2$ & 1.60 & 0.53 \\
\hline
\end{tabular}\\
\end{center}
$^a$Our results for the diatomic molecules in this table differ slightly from those of the original work of Nafziger, Wu, and Wasserman \cite{nafziger_2011} because we recalculated all energies using the aug--cc--pVTZ basis set. \\
\end{table}
\subsection{Partition potentials}
Hydrogen bonding is a complex interaction with various degrees of contribution from electrostatic, inductive, and dispersive forces depending on the nature of the individual molecules. In this work, we use the PW91 functional, which fortuitously yields accurate interaction energies and molecular geometries in weakly-bonded clusters such as the benzene and methane dimers among others \cite{tsuzuki}; and we also use the very popular B3LYP hybrid functional.\\
Fig. \ref{f:rho_vp_plots} shows that all features of the partition potential are largely insensitive to the choice of XC functional (we only show results for D$_1$, but the same is also true for D$_2$ $\rightarrow$ D$_4$). Since $v_p(\mathbf{r})$ is obtained through the density-to-potential inversion of eq. \ref{eq:WY}, this is due to the \textit{densities} being insensitive to the choice of XC functional. The question of whether the approximate XC functionals can accurately capture the exact features of $v_p(\mathbf{r})$ remains open. \cite{OW18} For both functionals, we were able to achieve density convergence to the order of $10^{-8}$ a.u. in a reasonable number of iterations (on the order of $10^2$). \\
Fig. \ref{f:vp_plots} compares the B3LYP partition potentials for all four dimers. In contrast to the monomer density deformations, similar bonds are \textit{not} characterized by similar features in $v_p(\mathbf{r})$. This is further highlighted in Figure \ref{t:rho_vp_d2_d3_d4}, where the partition potentials are plotted along the following nearly-linear intramolecular bonds: C=O $\cdots$ O--H in D$_1$ and D$_2$; O--H $\cdots$ O--H in D$_3$; and O--H $\cdots$ O--H in D$_4$. Note that although the density deformations in the binding regions are qualitatively similar in all four dimers, $v_p(\mathbf{r})$ is qualitatively different for the global minimum (D$_1$), where it is highly negative. \\
The non-transferability of $v_p(\mathbf{r})$ or any of its features indicates its sensitivity to the density variations in regions that may be far from $\mathbf{r}$. In contrast, the density deformations are highly localized. Qualitatively, the 2D density deformations of Figure \ref{f:dens_diff_plots} show that, due to the formation of the hydrogen bond, charge is accumulated in the region occupied by the antibonding orbital in the $R$ monomer and simultaneously withdrawn from the region occupied by the lone pair in the $L$ monomer. Quantitatively, accumulation of charge in the antibonding region of $R$ and depletion of charge in the lone pair region of $L$ are larger for stronger bonds.
\begin{figure}
\centering
\includegraphics[scale=0.2]{figure_vp_1D.pdf}
\caption{1D plots of $v_p(\mathbf{r})$ along the approximate bond axis. The B3LYP functional in conjunction with the aug-cc-pVTZ basis set was used for all calculations. Vertical lines in the 1D plots enclose the intermolecular bonding region.}
\label{t:rho_vp_d2_d3_d4}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.60]{figure_vp_D1.pdf}
\caption{Partition potential, $v_p(\mathbf{r})$, for D$_1$ using B3LYP (top) and PW91 (bottom) functionals. The aug-cc-pVTZ basis set was used for all calculations. The left two plots show the $v_p(\mathbf{r})$ map on the molecular plane. The right two plots show $v_p(\mathbf{r})$ along the approximate HB line (through H atom of the donor and O atom of the acceptor).}
\label{f:rho_vp_plots}
\end{figure}
\subsection{PDFT and descriptors of weak chemical binding}
Farf\'an \textit{et al.} \cite{fad} established a hierarchy of strengths of intermolecular contacts based on a combined view that includes topological analyses of the electron densities and orbital interactions. Their results suggest that the strengths of hydrogen bonding in (FA)$_2$ decrease as C=O $\cdots$ H--O > H--O $\cdots$ H--O > C=O $\cdots$ H--C > H--O $\cdots$ H--C. Table \ref{t:quant} lists several chemical descriptors adapted from the work of Farf\'an \textit{et al.} \textit{et al.}\cite{fad} supplemented with quantities derived from our own PDFT results. \\
\begin{sidewaystable}
\centering
\caption{Descriptors or chemical bonding for hydrogen bonds in the formic acid dimers. WBI: Wiberg bond indices, $n$ is the electron density and $\nabla^2 n$ is its Laplacian, $\mathcal{H}$ is the total energy density, all evaluated at the bond critical points. All quantities listed in atomic units except for orbital interaction energies, $E^{(2)}_{ov}$, which are given in kcal/mol. Max ($R$) and Min ($L$) refer to the points of maximum and minimum density deformation, extracted from Figure \ref{f:dens_diff_plots}. }
\label{t:quant}
\begin{tabular}{|c|c|ccccccc|}
\hline
\hline
Interaction & Dimer & WBI & $n\left(\mathbf{r}_c\right)$ & $\mathcal{H}\left(\mathbf{r}_c\right)$ & $\nabla^2 n\left(\mathbf{r}_c\right)$ & $E^{(2)}_{ov}$ & Max ($R$) & Min ($L$) \\
\hline
C=O $\cdots$ H--O & D$_1$ & 0.0918 & 0.0398 & -0.0024 & 0.1288 & 21.36 & 0.048 & -0.170\\
C=O $\cdots$ H--O & D$_2$ & 0.0711 & 0.0329 & 0.0001 & 0.1143 & 17.62 & 0.047 & -0.120\\
C=O $\cdots$ H--O & D$_3$ & 0.0367 & 0.0226 & 0.0016 & 0.0816 & 6.86 & Non-Linear & Non-Linear\\
H--O $\cdots$ H--O & D$_4$ & 0.0282 & 0.0248 & 0.0021 & 0.0972 & 7.39 & 0.021 & -0.070\\
H--O $\cdots$ H--O & D$_3$ & 0.0187 & 0.0211 & 0.0021 & 0.0829 & 3.96 & 0.016 & -0.060\\
C=O $\cdots$ H--C & D$_2$ & 0.0069 & 0.0108 & 0.0011 & 0.0370 & 1.05& Non-Linear & Non-Linear\\
C=O $\cdots$ H--C & D$_4$ & 0.0044 & 0.0094 & 0.0008 & 0.0301 & 0.77& Non-Linear & Non-Linear\\
\hline
\end{tabular}
\end{sidewaystable}
The criteria used by Farf\'an \textit{et al.} \textit{et al.} \cite{fad}, listed in Table \ref{t:quant}, may be dissected as follows
\begin{enumerate}
\item Wiberg bond indices are directly linked to the strength of the chemical interactions
\item Larger electron densities at bond critical points result in stronger interactions
\item The sign and magnitude of the total energy density at bond critical points, $\mathcal{H}\left(\mathbf{r}_c\right) = \mathcal{G}\left(\mathbf{r}_c\right) + \mathcal{V}\left(\mathbf{r}_c\right)$, are indications of the strengths and nature of the interactions because if $\mathcal{H}\left(\mathbf{r}_c\right)<0$ the potential (attractive) energy dominates leading to a concentration of electron charge in the vicinities of the BCP, strengthening the interaction, with an opposite effect for $\mathcal{H}\left(\mathbf{r}_c\right)>0$ when the kinetic (repulsive) energy dominates.
\item The sign of the Laplacian of the electron density at bond critical points tells apart local maxima and minima. Thus, negative Laplacians (local maxima) indicate local concentration of charge around the critical points, which suggests stronger interactions. Positive Laplacians (local minima) indicate local depletion of charge, suggesting weaker interactions.
\item Orbital interaction energies calculated via second order perturbation theory on the Fock matrix represented by the NBO basis reveal the mechanism for the formation of hydrogen bonds and are directly correlated to the strength of the interaction. In all (FA)$_2$ cases, charge transfer from a lone pair in an oxygen atom in either a carbonyl or in a hydroxyl group to a neighboring antibonding orbital is identified as the culprit. The orbital charge transfer maybe represented in a general way as $n_{\rm O} \rightarrow \sigma^{\star}_{\rm X-H}$ with X= O, C.
\end{enumerate}
\begin{figure}
\centering
\includegraphics[scale=0.6]{figure_vp.pdf}
\caption{Partition potential, $v_p(\mathbf{r})$, for dimers for D$_1$ $\rightarrow$ D$_4$ (B3LYP/aug-cc-pVTZ).}
\label{f:vp_plots}
\end{figure}
As is clear from Table \ref{t:quant}, the low-density deformations are consistent with the hierarchy of strengths of hydrogen bonds in (FA)$_2$ established by Farf\'an \textit{et al.} \cite{fad}
\section{CONCLUDING REMARKS}
Contrary to our original expectation when we started this study, we showed that the partition potential is \textit{not} transferable between systems with similar types of hydrogen-bonding. The result highlights the nonlocal character of $v_p(\mathbf{r})$ in contrast to the local features of density deformations of the individual fragments, which \textit{are} largely transferable. In practical calculations, we should take advantage of the fact that fragment calculations can be done locally while still preserving the global features of the partition potential. We also highlight the intuitive yet nontrivial observation that large binding energies correspond to large preparation energies, and that the strength of the partition potential is correlated with the overall stability of the dimer, as made obvious by our graphical abstract. \\
\section*{acknowledgements}
We acknowledge support from the Universidad de Antioquia - Purdue grant No. PURDUE14-2-02. A.W. and Y.O. acknowledge support from the National Science Foundation CAREER program under
Grant No. CHE-1149968.
\bibliographystyle{spphys}
|
\section{Introduction}\label{sec:introduction}
In the vast realm of nano- and micro-scale materials, low dimensional systems, such as colloidal mono- and bi-layers, are relevant for a broad range of applications as, $e.g.$, for antireflection coatings, biosensors, data-storage, optical and photovoltaic devices, or catalysts~\cite{NanoparticleArrays2000,BiosensorVelev,Lithography2012,Coatings2012}. The ordering of the colloidal particles on a surface strongly depends on a delicate balance between the properties of the particles and those of the substrate. An efficient method for the fabrication of two-dimensional colloidal structures consists, $e.g.$, in tuning the surface properties of the colloidal particles~\cite{Kagomeexp_2011,Pine_2012}. The guided growth of planar architectures can be also achieved by taking advantage of chemically or physically patterned substrates~\cite{review-substrates,stripe-substrates1,stripe-substrates2,stripe-substrates3,bechinger2007}. Here we combine these two approaches by considering both particles and substrates with a charged surface pattern, $i.e.$, we investigate the behavior of non-homogeneously charged colloids close to a non-homogeneously charged substrate.
Heterogeneously charged colloids have recently emerged as promising self-assembling units for material architectures with target properties~\cite{Yigit15a,Yigit15b,vanOostrum2015,Yura2015-2,Cruz_2016,Blanco,Hiero_2016,Lund2017,Mani2017}. Particles with distinct charged regions on their surface can be generally regarded as charged patchy units; in order to distinguish them from conventional patchy particles~\cite{patchyrevexp,patchyrevtheo,newreview}, they are often referred to as inverse patchy colloids (IPCs)~\cite{bianchi:2011}. Similar to conventional patchy colloids, IPCs are characterized by anisotropic interaction patterns and low bonding valence, but the competition between the orientation-dependent attraction and repulsion -- induced by the interactions between like/oppositely charged areas on the particle surface -- leads to more complex assembly scenarios as compared to conventional patchy units~\cite{ipc-review,newreview}. A review about this class of systems, covering the synthesis of model particles, their self-assembly, their coarse-grained modeling and the related numerical/analytical treatments, has been presented in Ref.~\cite{ipc-review}.
Within this broad class of systems, most of the attention has been devoted so far to IPCs with two charged polar patches and an oppositely charged equatorial belt. The characteristic feature of the bonding between IPCs with two identical patches is the direct contact between the polar region of one particle with the equatorial area of the other. This preferred pattern favors the formation of planar aggregates either as monolayers close to a homogeneously charged substrate~\cite{bianchi:2d2013,bianchi:2d2014} or as bulk equilibrium phases~\cite{ismene,evaemanJPCM,silvanonanoscale}.
The assembly of heterogeneously charged colloids close to a homogeneously charged substrates depends on different parameters, such as the particle/substrate charge ratio, the extension of the charged regions on the particle surface, and the net particle charge~\cite{bianchi:2d2013,bianchi:2d2014}. The resulting surface layers can have different densities: sometimes particles assemble into close-packed, hexagonally ordered crystalline aggregates, sometimes they form open, square-like layers, sometimes they form a monomer-phase where particles are adsorbed on the substrate~\cite{bianchi:2d2013,bianchi:2d2014}. Moreover, upon subtle changes of external parameters, such as the pH of the solution or the electrical charge of the substrate, it is possible to reversibly switch the assembly process on and off as well as to induce a transformation from a one specific particle arrangement to another~\cite{bianchi:2d2014}. The same morphological features observed in simulations are found in experimental samples of IPCs sedimented on a glass substrate~\cite{vanOostrum2015}. A qualitative comparison between simulations and experiments has been reported in Ref.~\cite{ipc-review}: while in numerical samples all clusters have the same spatial and orientational order, in the experimental sample different particle arrangements coexist, probably because of the patch size polydispersity of the IPCs synthesized so far.
Here we investigate how the ordering of IPCs on a charged surface is affected by the substrate pattern. In particular, we consider negatively charged colloids with positively charged polar caps and substrates characterized by parallel stripes forming an alternating charge pattern: negative/neutral, positive/neutral and positive/negative. On varying the size of the stripes, we observe the emergence of different particle ordering, from adsorbed monomers to adsorbed crystalline layers with distinct particle arrangements. We also investigate the robustness of the emerging scenarios with respect to changes in the substrate surface charge and in the electrostatic screening conditions.
\section{Model and Methods}\label{sec:model&methods}
We consider IPCs with two positively charged polar caps and a negatively charged equatorial belt. The polar patches are identical, $i.e.$, they have the same size as well as the same charge. We mostly focus on neutral IPCs, $i.e.$, heterogeneously charged particles with a zero net charge; we also consider charged IPCs carrying a negative net charge. Particles are placed under confinement between two parallel walls at a distance such that two particles cannot sit on top of each other; the top wall is always neutral, while the bottom wall is patterned with parallel stripes, which can be either neutral or positively/negatively charged.
We note that the Debye-H\"uckel approximation behind our coarse-grained description is strictly valid when $ Z_i q_e \Phi \ll k_BT$, where $Z_i q_e$ is the charge of the ionic species (where $Z_i=1$ for monovalent ions and $q_e$ is the elementary charge), $k_B T$ at room temperature is $\approx 26$ meV, and $\Phi$ is the potential at the surface. For IPC systems, by varying the different parameters that characterize the potential (namely, the colloidal size, the effective charges of the different surface areas, the geometric parameters defining the charge distribution and the Debye screening length), it is possible to determine in which parameter windows the linear approximation is expected to be reliable. In our specific case, $q_e \beta \Phi \lesssim 1$ for $\sigma \ge 30$ nm. For particles with $\sigma \approx 30$ nm, the salt concentration corresponding to our $\kappa\sigma$ is $\approx 10$ mM.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{IMAGES/model-tot.png}
\end{center}
\caption{Panel (a): schematic representation of a coarse-grained IPC. The gray sphere corresponds to the hard colloid, while the yellow caps represent the polar patches. It is worth noting that IPCs are spherical particles: the yellow caps are the emerging parts of the patch interaction spheres, whose centers are located inside the colloid. The model geometric parameters are the radius of the colloid $\sigma_{\rm c}$, the particle interaction range $\delta$ and the patch opening angle $\gamma$ (as labeled); the latter two are determined once the position of the patch centers $a$ and the patch interaction range $\sigma_{\rm p}$ (as labeled) are chosen. Panels (b) and (c): schematic representation of the interaction between an IPC and a patterned substrate. The different colors correspond to differently charged stripes on the substrate. The intersection between the (core or patch) interaction sphere and the substrate is a spherical cap: the sub-volumes of the caps are highlighted for both the core-substrate (b) and the patch-substrate (c) interaction. Panel (d): particle slices corresponding to the different sub-volumes highlighted in panel (b).}
\label{fig:model}
\end{figure}
\subsection{Particle-particle interaction}
\label{sec:ppint}
For the particle-particle interaction, we consider the coarse-grained description introduced in Ref.~\cite{bianchi:2011}. The model features a spherical, impenetrable colloid, of radius $\sigma_{\rm c}$ and central charge $Z_{\rm c}$, carrying two polar patches, each of charge $Z_{\rm p}$. The particle diameter $\sigma=2\sigma_{\rm c}$ sets the unit of length. The model has three sets of independent parameters: the particle interaction range $\delta$, the patch surface extent, specified by the half opening angle $\gamma$, and the interaction strengths between the different surface regions (see panel (a) Fig.~\ref{fig:model}). The pair potential is a sum of three different contributions: the patch/patch, the patch/core and the core/core interactions. The specific form of the potential is based on the postulate that all the interactions can be factorized into a characteristic energy strength and a geometric weight factor, the latter one taking into account the distance and relative orientation between the interacting particles. More specifically, beyond the hard core repulsion, the pair potential between two IPCs at distance $r_{ij}$ with a mutual orientation $\Omega_{ij}$ is given by
\begin{equation}\label{eq:U}
U (r_{ij},\Omega_{ij}) =
\left\{
\begin{array}{rl}
\sum_{\alpha\beta} u_{\alpha\beta} w_{\alpha\beta}(r_{ij},\Omega_{ij}) &{\hspace{1em}\rm if\hspace{1em}} \sigma <r_{ij}<\sigma+\delta \\
0 &{\hspace{1em}\rm if\hspace{1em}} r_{ij} \ge \sigma+\delta
\end{array}
\right.
\end{equation}
where $i$ and $j$ specify the particles, $\alpha$ and $\beta$ specify either a patch (P) or the core (C) of the first and second IPC, respectively, $w_{\alpha\beta}$ and $u_{\alpha\beta}$ are the geometric weight factor and the characteristic energy strength of the $\alpha\beta$ interaction, respectively. We note that, while the $u_{\alpha\beta}$ are constant values, the $w_{\alpha\beta}$ depend on both the inter-particle distance and the relative orientation of the two IPCs. For sake of simplicity we omit in the following the explicit dependence of the $w_{\alpha\beta}$ on $r_{ij}$ and $\Omega_{ij}$. The weight functions for a pair of particles in a given configuration are calculated as overlap volumes between pairs of (core and/or patch) interaction spheres~\cite{bianchi:2011}. The energy strengths are defined $via$ a mapping between a suitably developed Debye-H{\"uckel} potential~\cite{bianchi:2011,debyehueckel} -- fully determined by the physical properties of the underlying microscopic system -- and the patchy pair potential -- characterized by the yet undetermined $u_{\alpha\beta}$~\cite{bianchi:2011}.
Here we consider two IPC types: they share the same patch size and interaction range, while having different net particle charge. As the selected systems have been studied before~\cite{bianchi:2d2014}, we use the names already given in Ref.~\cite{bianchi:2d2014} for the sake of consistency. In the contest of Ref.~\cite{bianchi:2d2014} it was important to distinguish between IPCs with different patch size and different overall particle charge, thus the systems were labelled as 60n and 60c, where 60 referred to $\gamma$, while ``n" and ``c" indicated neutral and charged particles, respectively. In the following we give all the parameters that characterize both systems.
The interaction range is fixed by the electrostatic screening conditions of the surrounding solvent. We choose $\kappa\sigma=10$ and we define $\delta$ as a function of the Debye screening length $\kappa^{-1}$. While usually $\kappa\delta=1$, we extend the relation between $\delta$ and $\kappa$ to allow for a more quantitative evaluation of the characteristic interaction distance~\cite{bianchi:2011}, $i. e.,$ we assume $\kappa\delta=n$, where $n$ is not necessarily an integer number. Here we choose $n=2$ corresponding to $\delta=0.2$ in units of the particle diameter.
The patch size is ideally determined by the corresponding feature of experimentally synthesized particles. Once $\delta$ is chosen, $\gamma$ is fixed either by the position of the patch center of charge, $a$, or by the size of the patch interaction sphere $\sigma_{\rm p}$, since the following relations must be satisfied (see panel (a) of Figure~\ref{fig:model}):
\begin{eqnarray}
\frac{\delta}{2}&=&a+\sigma_{\rm p}-\sigma_{\rm c} \nonumber \\
\cos\gamma&=&\frac{\sigma_{\rm c}^2+a^2-\sigma_{\rm p}^2}{2\sigma_{\rm c} a}.
\end{eqnarray}
\label{eq:params}
We choose $\gamma \approx 60 \degree$, corresponding to $a = 0.16$, with $\sigma_{\rm p}= 0.44$, both in units of the particle diameter.
Finally, the energy parameters of the model, $i.e.$, the $u_{\alpha\beta}$ in equation (\ref{eq:U}), are related to the effective charges of the different surface regions. These charges are responsible for the ratio between the attractive and repulsive contributions to the pair energy. We mostly focus on neutral IPCs, $i.e.$, particles with a net charge $Z_{\rm tot}=Z_{\rm c}+2Z_{\rm p} = 0$. Specifically, we set $Z_{\rm c}=-180$ and $Z_{\rm p}=90$, where we have set the elementary charge to one. We also consider overall charged colloids, such that the net particle charge is negative, namely $Z_{\rm tot}=-\frac{2}{9} Z_{\rm p}<0$ with $Z_{\rm c}=-200$ and $Z_{\rm p}=90$. To derive the energy strengths $u_{\alpha\beta}$ in equation (\ref{eq:U}) we perform the ``max" mapping between the model potential of equation (\ref{eq:U}) and the Yukawa-like pair potential appropriately derived within the Debye-H{\"u}ckel approach for IPCs in water at room temperature~\cite{bianchi:2011}. Namely, we match the value of the two potentials for a pair of IPCs at contact ($i.e.$, the distance between the two cores is $r = \sigma$) with a fixed mutual orientation (shown in Fig.~\ref{fig:configs}). For each $\alpha\beta$ interaction we choose a reference particle configuration~\cite{bianchi:2011}. For the 60n system, we obtain $u_{\rm CC}= 0.1349$ for the core/core interaction, $u_{\rm PC}= -0.8483$ for the patch/core interaction, and $u_{\rm PP}= 4.3228$ for the patch/patch interaction. For the 60c system, we obtain $u_{\rm CC}= 0.4330$, $u_{\rm PC}= -1.9467$, and $u_{\rm PP}= 4.3228$. The pair interaction energy is normalized by the value corresponding to the minimum of the attraction, $\varepsilon_{\rm min}$; thus $|\varepsilon_{\rm min}|$ sets the energy unit.
The characteristic pair configurations -- reported in Fig.~\ref{fig:configs} -- are referred to as equatorial-polar (EP), equatorial-equatorial (EE) and polar-polar (PP).
We observe that, by construction, the equatorial-polar attraction for two particles at contact is alway -1 in both systems. In contrast, the strengths of the equatorial-equatorial and polar-polar repulsive interactions for two particles at contact are different in the two systems. For system 60n the polar-polar repulsion is the strongest interaction, while the equatorial-equatorial repulsion is the weakest; for system 60c the attraction is stronger than both repulsions, which in turn have comparable strength. A representation of the inter-particle potentials is reported in the Supporting Information of Ref.~\cite{bianchi:2d2014}, while contact energy values for the particle-particle configurations reported in Fig.~\ref{fig:configs} are listed in Tab.~\ref{tab:contactenergies}.
\subsection{Particle-substrate interaction}
We consider a system of IPCs confined between two hard walls in a quasi-2D slab along the $xy$-plane, the two walls are positioned at $z =$ 0 and $z =$ 1.45$\sigma$. Such a tight confinement prevents particles from sitting on top of each other. The top wall is always neutral, while the bottom wall -- the substrate -- has a charged pattern: we consider stripes of the same width along the $x$ direction, spanning the plane along the $y$ axis. The stripes form alternating charged motifs: we alternate negative and neutral (-/0), positive and neutral (+/0) or positive and negative (+/0) stripes. We consider different numbers of stripes, $N_s$, spanning from $N_s =$ 2 to $N_s =$ 90, considering only even numbers of stripes, so that the total area covered by each of the two kinds is the same for any choice of $N_s$. We note that, since the stripes fill the whole surface, a low number of stripes implies that each stripe is quite large, the width of a stripe being defined as $L_s = L/N_s$. As $N_s$ increases, stripes become thinner and thinner, until their size becomes smaller than the particle size. At that point, each particle can interact with multiple stripes at the same time: for our choice of the system size, this scenario occurs at $N_s =$ 50.
For the particle-substrate interaction, we consider the coarse-grained description introduced in Ref.~\cite{bianchi:2d2013}. The interaction between a coarse-grained IPC and a neutral substrate is a steric interaction modeled as a hard repulsion taking place when the particle is located at distance ${\rm z}\le\sigma_{\rm c}$ from the substrate. In the presence of a charged substrate with surface charge $\sigma_{\rm w}=Z_{\rm w}/4\sigma_{\rm c}^2$, a screened electrostatic interaction must be added to the hard particle-substrate repulsion. We consider either positive and negative substrates with a very small value of $Z_{\rm w}$ as compared to $Z_{\rm p}$, $i.e.$, $Z_{\rm w}=\pm\frac{5}{90}Z_{\rm p}=\pm$5.
Following Ref.~\cite{bianchi:2d2013}, the particle-substrate interaction is modeled according to the same coarse-grained philosophy introduced in Sec.~\ref{sec:ppint}: beyond the hard core repulsion, the interaction energy of an IPC at distance $z_{i}$ and orientation $\Omega_{i}$ with respect to the substrate (S) is given by
%
\begin{equation}\label{eq:V}
V (z_{i},\Omega_{i}) =
\left\{
\begin{array}{rl}
\sum_{\alpha} u_{\alpha {\rm S}}w_{\alpha {\rm S}}(z_{i},\Omega_{i}) &{\hspace{1em}\rm if\hspace{1em}} \sigma_{\rm c} <z_{i}<\sigma_{\rm c}+\delta \\
0 &{\hspace{1em}\rm if\hspace{1em}} z_{i} \ge \sigma_{\rm c}+\delta
\end{array}
\right.
\end{equation}
where $i$ specifies the particle, $\alpha$ specifies either a patch (P) or the core (C), while $w_{\alpha {\rm S}}$ and $u_{\alpha {\rm S}}$ are the geometric weight factor and the characteristic energy strength of the ${\alpha {\rm S}}$ interaction, respectively. On physical grounds, the particle-substrate interaction range is the same as the particle-particle interaction range. The energy strengths $u_{\alpha {\rm S}}$ are defined $via$ a mapping between an appropriately developed Debye-H{\"uckel} potential and the model potential~\cite{bianchi:2011}; for system 60n and positive substrate we have $u_{\rm PS}= -0.6278$ and $u_{\rm CS}=2.2310$ (same magnitude but opposite signs for the interactions with the negative substrate), while $u_{\rm PS}=-1.48901$ and $u_{\rm CS}= 2.2310$ for system 60c and positive substrate (again, same magnitude but opposite signs for the interactions with the negative substrate). The weight functions $w_{\alpha {\rm S}}$ are calculated as overlap volumes between the (core or patch) interaction sphere $\alpha$ and the substrate~\cite{bianchi:2d2014}. For sake of simplicity we omit in the following the explicit dependence of the $w_{\alpha S}$ on $z_{i}$ and $\Omega_{i}$. When $\alpha$ interacts only with one substrate type then the overlap volume is the volume of the spherical cap normalized by the volume of the particle. When $\alpha$ overlaps with multiple stripes we proceed as it follows. First, we compute the overlap cap between $\alpha$ and the substrate as if the substrate were homogeneous: the whole overlap cap is the sum of the overlap sub-volumes represented with dark -- blue and red -- colors in panels (b) and (c) of Figure~\ref{fig:model}. Then, we calculate the volumes of the particle slices each overlap sub-volume belongs to, represented in panel (d) of Figure~\ref{fig:model}. Finally, we use the volumes of the slices to define an effective interaction energy associated to the whole spherical cap, namely $u_{\alpha S}^{*}=({\rm v}_{S1}u_{\alpha S1}+{\rm v}_{S2}u_{\alpha S2})/({\rm v_{S1}+v_{S2}})$, where $\rm v_{S1,S2}$ is the total volume of all particle slices above the substrate of type S1 and S2, respectively. The resulting contribution to the particle-substrate interaction is then $u_{\alpha S}^{*}w_{\alpha {\rm S}}$. We remark that this procedure accounts straightforwardly for an arbitrary large number of interacting stripes. Similar to the particle-particle interaction, also the particle-substrate potential is normalized by the value corresponding to the minimum of the particle-particle attraction $\varepsilon_{\rm min}$. Characteristic interaction configurations are reported in Fig.~\ref{fig:configs} and are referred to as equatorial-substrate (ES) and polar-substrate (PS). A representation of the particle-substrate potentials is reported in the Supporting Information of Ref.~\cite{bianchi:2d2014}, while contact energy values for the particle-substrate configurations reported in Fig.~\ref{fig:configs} are listed in Tab.~\ref{tab:contactenergies}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.4\textwidth]{IMAGES/configs.png}
\end{center}
\caption{Typical particle-particle (top row) and particle-wall (bottom row) configurations: (from left to right, from top to bottom) equatorial-polar (EP), equatorial-equatorial (EE), polar-polar (PP), equatorial-substrate (ES) and patch-substrate (PS), as labeled.}
\label{fig:configs}
\end{figure}
\begin{table}[h]
\centering
\begin{tabular}{lccccc}
Particle name & EP & EE & PP & ES$_{(+)}$ & PS$_{(+)}$ \\
\hline
60n & -1 & 0.159 & 3.683 & -2.162 & 4.687 \\
60c & -1 & 0.265 & 0.258 & -2.611 & 0.347 \\
\end{tabular}
\caption{Contact energies for the investigated IPC-types on a positively charged substrate: two particles, either 60n or 60c, at contact ($r=\sigma$) and fixed mutual orientations, as labeled in Fig.~\ref{fig:configs}. For a graphical representation of the full potentials we refer the reader to Fig.~S2 in the supplementary materials of Ref.~\cite{bianchi:2d2014}. We note that when the substrate is negatively charged, the contact energy values ES and PS just change sign.}
\label{tab:contactenergies}
\end{table}
\subsection{Numerical Methods}
We perform Monte Carlo (MC) simulations of IPC systems with $N=1000$ particles in the canonical ensemble at $T^*=0.10$ (temperature in reduced units, $i.e.$, $T^*=k_BT/|\varepsilon_{\rm min}|$) in a volume $V=hL^2$, where $L=50\sigma$ and $h=1.45\sigma$. Each MC step consists of $N$ trial particle moves, where the acceptance rule is given by the Metropolis criterion. A particle move is defined as both a displacement in each Cartesian direction of a random quantity distributed uniformly between $\pm \delta r$ as well as a rotation around a random axis of a random angle distributed uniformly between $\pm \delta \theta$. The chosen values for the trial changes are $\delta r = 0.05\sigma$ and $\delta \theta = \delta r/2\sigma$ rad. We perform for each state point 16 MC parallel runs, starting from different initial conditions. Each run lasts for a total of $\approx 10^8-10^9$ MC steps. All quantities are averages over the final $10^7$ MC steps, corresponding to at least 100 different configurations per independent run.
\subsection{Post-processing analysis}
To characterize the different aggregation scenarios that emerge on varying the substrate pattern, we (i) study the modulation of the system density (in section~\ref{sec:distr}), (ii) compute the average number of monomers and the average number of neighbors at each state point (in section~\ref{sec:bondorder}) and (iii) perform a local order analysis using the bond-order parameters $\phi_4$ and $\phi_6$ in two dimensions (in section~\ref{sec:bondorder}).
Monomers are defined as particles with no bonded interactions, $i.e.,$ the contribution to their energy related to the inter-particle interaction is zero or greater than zero. Neighboring particles are defined following an energy criterion: two particles are considered bonded if their interaction energy is negative. The average number of neighbors is computed accounting only for bonded particles ($i.e.,$ monomers are not counted).
Considering the projection of the system on the $xy$-plane, the bond-order parameters are defined as
\begin{equation}
\phi_4(i) = \left|\frac{1}{\tilde{N}_b(i)} \sum_{j = 0}^{\tilde{N}_b(i)} {e^{i 4 \theta_{ij}}}\right| \qquad \qquad \phi_6(i) = \left|\frac{1}{\tilde{N}_b(i)} \sum_{j = 0}^{\tilde{N}_b(i)} {e^{i 6 \theta_{ij}}}\right|
\label{eq:bondord}
\end{equation}
where $\theta_{ij} = \arctan(y_{ij}/x_{ij})$ is the phase of the relative distance $\boldmath{r}_{ij}$ in two dimensions and $\tilde{N}_b(i)$ is the number of neighbors of particle $i$.
We also report selected simulation snapshots; for the sake of clarity, we represent the particles with a slightly smaller size than their real one and we use a particle color code to highlight the number of bonds per particle. These two choices help the visualization of compact structures and of their local symmetry.
\section{Results}
We start our discussion with a brief summary of the results previously collected for systems 60n and 60c close to a homogeneous substrate~\cite{bianchi:2d2014} (see Figure~\ref{fig:old}).
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{IMAGES/summary-old.png}
\end{center}
\caption{Simulation snapshots of IPC systems labeled as 60n (left panel) and 60c (right panel) on a positive (blue), negative (red) and neutral (white) substrate (as labeled). The particle color code refers to the number of bonds per particle, $N_b$: a bond between two particles is formed when their interaction energy is negative.}
\label{fig:old}
\end{figure}
When confined between two parallel neutral walls (0) both systems form crystalline domains with a triangular particle arrangement; close to a charged substrate both IPC-types can adsorb on the surface, depending on the sign of the product $Z_{\rm p}Z_{\rm w}$. The competition between the particle-particle and the particle-substrate interactions gives rise to different morphologies. More specifically, a negative substrate (-) disfavors the assembly of system 60n, thus leading to a fluid-like phase where monomeric particles are adsorbed on the substrate; in contrast a positive substrate (+) favors the formation of adsorbed crystalline domains with square-like particle arrangement. For system 60c the triangular arrangement is more or less unaltered~\cite{bianchi:2d2014} with respect to the neutral case, the difference being the adsorption of the domains on the positive substrate, while no adsorption occurs on the negative substrate.
In the following we describe how the interplay between competing length scales and anisotropic interaction patterns affects the symmetry of the described crystalline domains.
\subsection{Spatial distribution}
\label{sec:distr}
When the substrate is no longer homogeneous, the particles may prefer a substrate-type with respect to the other and thus occupy specific regions of the sample. To quantify the fluctuations of the particle distribution within the sample, we consider the partial density $\rho_s$, defined as the number of particles observed on a given substrate-type divided by half of the total surface, as a function of $N_s$ for 2~$\le N_s \le $~ 90. We note that, due to the substrate specific design, significant fluctuations in the particle distribution only occur along the $x$ direction (corresponding to the axis of the charge modulation), while along the $y$ direction particles are always uniformly distributed; moreover, since particles are in most cases adsorbed to the charged parts of the substrate, fluctuations along the $z$ axis are in most cases negligible.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60n_n0.jpg}
\end{center}
\caption{System 60n on a -/0 substrate. Upper panels: a) partial densities $\rho_{\rm s}$ on the two substrate types (as labelled) as a function of the number of stripes $N_s$, b) average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as a function of $N_s$. Lower panels: MC simulations snapshots at selected state points: ($a_1$) $N_s =$ 2, ($a_2$) $N_s =$ 8, ($a_3$) $N_s =$ 26 and ($a_4$) $N_s =$ 60. The particle color code refers to the number of bonds per particle and is it reported in Figure~\ref{fig:old}.}
\label{fig:60n_n0}
\end{figure*}
We start our discussion with system 60n. On a -/0 substrate with $N_s=2$, particles form crystalline, non-adsorbed domains on the neutral stripe and adsorb on the negative stripe without forming any aggregate (see snapshot ($\rm a_1$) in Fig.~\ref{fig:60n_n0}). As soon as $N_s$ increases, particles prefer to absorb on the negatively charged regions, thus leaving the neutral stripes completely empty (see snapshots ($\rm a_2$), ($\rm a_3$) and ($\rm a_4$) in Fig.~\ref{fig:60n_n0}). The corresponding $\rho_s$ (see panel a) of Fig.~\ref{fig:60n_n0}) clearly confirms that for $N_s=2$ most particles are located on the negative stripes and for $N_s\ge8$ the neutral stripes are completely empty. All populated stripes are characterized by adsorbed monomers.
The neutral stripes completely deplete -- as soon as $N_s >$ 2 -- also for system 60n on a +/0 substrate (see Fig.~\ref{fig:60n_p0}). In contrast to the previous case, particles adsorbed on the charged stripes form large crystalline clusters. By visual inspection, we can anticipate that the crystalline clusters become more and more elongated with increasing $N_s$ (see snapshots ($\rm b_1$) and ($\rm b_2$) in Fig.~\ref{fig:60n_p0}); on further increasing $N_s$ particles organize in double-particle lanes (see snapshot ($\rm b_3$) in Fig.~\ref{fig:60n_p0}); when the size of the stripes becomes comparable to $\sigma$ ($i.e.$, for $N_s = 50$) particles organize in single-particle strings (see snapshot ($\rm b_4$) in Fig.~\ref{fig:60n_p0}); finally, these strings merge and form again a crystalline pattern (see snapshot ($\rm b_8$) in Fig.~\ref{fig:60n_p0}). The symmetry of the emerging aggregates depends on the competition between the different energies and length scales and it is discussed in section~\ref{sec:bondorder}.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60n_p0.jpg}
\end{center}
\caption{System 60n on a +/0 substrate. Upper panels: a) partial densities $\rho_{\rm s}$ on the two substrate types (as labelled) as a function of the number of stripes $N_s$, b) average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as a function of $N_s$, c) local bond order parameters $\phi_4$ and $\phi_6$ as function of $N_s$, shaded sectors highlight regions where the statistics is relevant. Lower panels: MC simulations snapshots at selected state points: ($b_1$) $N_s =$ 2, ($b_2$) $N_s =$ 10, ($b_3$) $N_s =$ 26, ($b_4$) $N_s =$ 50, ($b_5$) $N_s =$ 60, ($b_6$) $N_s =$ 66, ($b_7$) $N_s =$ 78 and ($b_8$) $N_s =$ 90. The particle color code refers to the number of bonds per particle and is it reported in Figure~\ref{fig:old}.}
\label{fig:60n_p0}
\end{figure*}
When particles adsorb on both substrate-types the scenario becomes more complex. Results for system 60n on a +/- substrate are shown in Fig.~\ref{fig:60n_opposite}. Even though the attraction between the patches and the negative substrate is stronger than the attraction between the equators and the positive substrates (see Tab.~\ref{tab:contactenergies}), as long as $N_s\le60$, particles slightly prefer to adsorb on the positive stripes ($i.e.$, $\rho_s$-values for the positive and negative stripes are comparable to the average system density $\rho=0.4$, see panel a) of Fig.~\ref{fig:60n_opposite}), where they can also aggregate. Visually, clusters on the positive stripes elongate and reshape on increasing $N_s$ (see snapshots ($c_1$)-($c_4$) in Fig.~\ref{fig:60n_opposite}) to the point that, when the width of the stripes is equal to $\sigma$ (and slightly smaller), particles form extended aggregates spanning across several adjacent stripes (see snapshot ($c_5$) and ($c_6$) in Fig.~\ref{fig:60n_opposite}). The transition from vertical aggregates to extended clusters is discussed in section~\ref{sec:bondorder} and~\ref{sec:visual}. Finally, when $N_s > 60$, particles prefer to adsorb on the negative stripes, as signaled by the abrupt increase of the partial density associated to the negative stripes towards $2\rho=0.8$ (see panel a) in Fig.~\ref{fig:60n_opposite}). Particles adsorbed on the negative stripes do not bond to each other (see snapshot ($c_7$) and ($c_8$) in Fig.~\ref{fig:60n_opposite}).
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60n_pn.jpg}
\end{center}
\caption{System 60n on a +/- substrate. Upper panels: a) partial densities $\rho_{\rm s}$ on the two substrate types (as labelled) as a function of the number of stripes $N_s$, b) average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as a function of $N_s$, c) local bond order parameters $\phi_4$ and $\phi_6$ as function of $N_s$, shaded sectors highlight regions where the statistics is relevant. Lower panels: MC simulations snapshots at selected state points: ($c_1$) $N_s =$ 2, ($c_2$) $N_s =$ 10, ($c_3$) $N_s =$ 26, ($c_4$) $N_s =$ 38, ($c_5$) $N_s =$ 50, ($c_6$) $N_s =$ 56, ($c_7$) $N_s =$ 66 and ($c_8$) $N_s =$ 90. The particle color code refers to the number of bonds per particle and is it reported in Figure~\ref{fig:old}.}
\label{fig:60n_opposite}
\end{figure*}
In summary, for system 60n, $\rho_s$ depicts the following scenario: for patterns -/0 and +/0, the neutral stripes are completely empty already at very low $N_s$-values; for pattern +/- the partial densities of the two substrate-types are very similar to each other, with a slight preference for the positive substrate as long as $N_s$ is small enough, while at high $N_s$-values, namely, when the width of the stripes is smaller than $\sigma$, the negative stripes become favorite. Anticipating a more detailed analysis proposed in section~\ref{sec:bondorder} and~\ref{sec:visual}, we also observe that, when particles are preferentially adsorbed to the negative stripes, no aggregation occurs at $N_s=90$ (see Fig.~\ref{fig:60n_n0} and Fig.~\ref{fig:60n_opposite}), while, when particles adsorb on the positive stripes (see Fig.~\ref{fig:60n_p0}) a crystalline pattern is formed at $N_s=90$.
We now look at the corresponding data for systems 60c. On a -/0 substrate, the system forms crystalline domains almost irrespective of the substrate charge, since particles do not adsorb on the neutral nor on the negative stripes~\cite{bianchi:2d2014}. The analysis of the partial densities shows a slight preference for the neutral substrate at small $N_s$-values, which soon disappears (see panel a) of Fig.~\ref{fig:60c_n0}): already for $N_s \approx 20$ the partial densities of both substrate-types reach the average density of the system $\rho=0.4$.
Similar to the corresponding case for the 60n system, on a +/0 substrate, particles prefer to assemble on the positive stripes where they also adsorb. As soon as $N_s > 2$ (see panel a) of Fig.~\ref{fig:60c_p0}), the neutral stripes become empty and all particles adsorb on the positive stripes forming crystalline domains ($i.e.$, $\rho_s=0.8$ on the positive stripes); these domains become more and more elongated and narrow on increasing $N_s$; as soon as the width of the stripes is smaller than $\sigma$, particles organize in single-particle strings that eventually merge into extended clusters, as discussed in section~\ref{sec:bondorder}. We observe that, for $N_s>60$, $\rho_s$ on the positive (neutral) substrate slightly decreases (increases): a visual inspection of the configurations suggests that the emergent crystalline domains have distinct symmetries, as discussed in section~\ref{sec:bondorder}.
On a +/- substrate, 60c particles prefer again to occupy the positive stripes (since 60c particles do not adsorb on the negative substrate~\cite{bianchi:2d2014}) and the assembly scenario is similar to the +/0 case (see panel a) of Fig.~\ref{fig:60c_opposite}). In this case, when $N_s>60$, the decrease (increase) of $\rho_s$ on the positive (negative) substrate is bigger than in the previous case, and this is related to the symmetry of the aggregates, as discussed in section~\ref{sec:bondorder}.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60c_n0.jpg}
\end{center}
\caption{System 60c on a -/0 substrate. Panel a) partial densities $\rho_{\rm s}$ on the two substrate types (as labelled) as a function of the number of stripes $N_s$. Panel b) average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as a function of $N_s$. Panel c) local bond order parameters $\phi_4$ and $\phi_6$ as function of $N_s$.}
\label{fig:60c_n0}
\end{figure*}
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60c_p0.jpg}
\end{center}
\caption{System 60c on a +/0 substrate. Upper panels: a) partial densities $\rho_{\rm s}$ on the two substrate types (as labelled) as a function of the number of stripes $N_s$, b) average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as a function of $N_s$, c) local bond order parameters $\phi_4$ and $\phi_6$ as function of $N_s$, shaded sectors highlight regions where the statistics is relevant. Lower panels: MC simulations snapshots at selected state points: ($d_1$) $N_s =$ 2, ($d_2$) $N_s =$ 10, ($d_3$) $N_s =$ 26, ($d_4$) $N_s =$ 50, ($d_5$) $N_s =$ 78 and ($d_6$) $N_s =$ 90. The particle color code refers to the number of bonds per particle and is it reported in Figure~\ref{fig:old}.}
\label{fig:60c_p0}
\end{figure*}
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60c_pn.jpg}
\end{center}
\caption{System 60c on a +/- substrate. Upper panels: a) partial densities $\rho_{\rm s}$ on the two substrate types (as labelled) as a function of the number of stripes $N_s$, b) average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as a function of $N_s$, c) local bond order parameters $\phi_4$ and $\phi_6$ as function of $N_s$, shaded sectors highlight regions where the statistics is relevant. Lower panels: MC simulations snapshots at selected state points: ($e_1$) $N_s =$ 2, ($e_2$) $N_s =$ 10, ($e_3$) $N_s =$ 26, ($e_4$) $N_s =$ 50, ($e_5$) $N_s =$ 78 and ($e_6$) $N_s =$ 90. The particle color code refers to the number of bonds per particle and is it reported in Figure~\ref{fig:old}.}
\label{fig:60c_opposite}
\end{figure*}
In summary, for system 60c, $\rho_s$ depicts the following scenario: for pattern -/0 the system is almost insensitive to the presence of the stripes on the substrate, while for patterns +/0 and +/-, the positive stripes are mostly preferred and the particle assembly ranges from extended aggregates at small $N_s$-values, to elongated clusters, double- and single-particle lanes and finally crystalline domains again.
It is worth stressing that, for both systems, as soon as the particles prefer to adsorb to the positive stripes, we observe the emergence of crystalline domains at $N_s=90$: the features of these domains differ from case to case since they depend on the delicate balance between different factors. In the following we perform a neighbor analysis in order to describe the different aggregation scenarios sketched above.
\subsection{Neighbor analysis}
\label{sec:bondorder}
To characterize the symmetry of the domains formed by IPCs in presence of a patterned substrate we consider the average number of monomers $\langle N_{mon} \rangle$, the average number of neighbors $\langle N_{neigh} \rangle$, and the local bond order parameters $\phi_4$ and $\phi_6$. We have highlighted in blue, in all panels c) of Figs.~\ref{fig:60n_p0},~\ref{fig:60n_opposite},~\ref{fig:60c_p0} and~\ref{fig:60c_opposite}, the state points where clusters with crystalline arrangements have been observed. In the other regions, non-crystalline arrangements occur and the values of the bond order parameters should not be used as an indicator of the assembly state of the system.
We start again our discussion with system 60n. As mentioned in section~\ref{sec:distr}, when system 60n is close to a -/0 pattern most of the particles are monomers adsorbed to the substrate (see panel b) of Fig.~\ref{fig:60n_n0}). More precisely, at $N_s=2$, slightly more than half of the sample is in a monomer state (adsorbed to the negative stripe), while the remaining particles belong to the crystalline domains (formed in correspondence to the neutral stripe). We know from the homogeneous case (see Fig.~\ref{fig:old}, left panel) that the crystalline domains observed on a neutral substrate have a triangular arrangement~\cite{bianchi:2d2014} and this is consistent with the $\langle N_{neigh} \rangle$-value observed at $N_s=2$. As soon as $N_s>2$, $\langle N_{mon} \rangle$ rapidly levels off to $N$, $i.e.,$ most of the particles become monomers. This is consistent with the corresponding drop of $\langle N_{neigh} \rangle$ from five to zero. In this case, a bond order analysis has little significance (and thus it is not reported), due to poor statistics.
When we consider a +/0 pattern, we observe that $\langle N_{mon} \rangle$ is negligible at all $N_s$-values (see panel b) of Fig.~\ref{fig:60n_p0}), since its peak for $40<N_s< 60$ corresponds to $\langle N_{mon} \rangle \approx 13$. In this case, particles form crystalline domains (see panel ($b_1$) of Fig.~\ref{fig:60n_p0}) that get elongated on increasing $N_s$, from compact clusters (see panel ($b_2$) of Fig.~\ref{fig:60n_p0}) to two-particles lanes (see panel ($b_3$) of Fig.~\ref{fig:60n_p0}) and single-particle strings (see panel ($b_4$) of Fig.~\ref{fig:60n_p0}), that eventually merge into crystalline domains again (see panel ($b_5$)-($b_8$) of Fig.~\ref{fig:60n_p0}), as anticipated in section~\ref{sec:distr}. The corresponding $\langle N_{neigh} \rangle$ (see panel b) of Fig.~\ref{fig:60n_p0}) decreases from $\langle N_{neigh} \rangle=4$ at $N_s=2$ (compact domains) to $\langle N_{neigh} \rangle=2$ at $N_s=50$ (single-particle stripes) and increases again to reach $\langle N_{neigh} \rangle=3$ at $N_s=90$ (compact domains). To better characterize the symmetries of the emerging compact domains we consider $\phi_4$ and $\phi_6$ as function of $N_s$. In panel c) of Fig.~\ref{fig:60n_p0} the two regimes where the order parameters are relevant are highlighted in blue. Both regimes correspond to a predominance of square particle arrangements: at small $N_s$ ($N_s<40$, panels ($b_1$)-($b_3$) of Fig.~\ref{fig:60n_p0}) this is the effect of the competition between different anisotropic interactions, while at large $N_s$ ($N_s> 72$, panels ($b_7$) and ($b_8$) of Fig.~\ref{fig:60n_p0}) an additional competition between different length scales sets in. More specifically, at small $N_s$, after the initial coexistence between square and triangular lattices (see snapshot ($b_1$) in Fig.~\ref{fig:60n_n0}), neutral stripes get depleted and the system forms clusters adsorbed on the positive stripes with a square symmetry, as signaled by the increase (decrease) of $\phi_4$ ($\phi_6$) for $N_s>2$. Despite becoming narrower and more elongated on increasing $N_s$, clusters maintain a square-like order until $N_s \simeq$ 32 (see snapshots ($b_2$) and ($b_3$) in Fig.~\ref{fig:60n_p0}), where particle organize in square-like, double-particle lanes. On further increasing $N_s$, we observe a coexistence between double-particle lanes with square symmetry and single-particle strings, which eventually prevail as soon as the particles are not able to pair in line on the same stripe anymore, $i. e.$, at $N_s =$ 50 (see snapshot ($b_4$) in Fig.~\ref{fig:60n_n0}). On further decreasing the size of the stripes below $\sigma$, these single-particle strings start merging (see snapshot ($b_5$) in Fig.~\ref{fig:60n_n0}): the effect of the stripes is to create \emph{tracks} to guide the assembly; over these tracks, particles are arranged in a configuration that greatly favors square domains. Thus, even tough there is a sweet spot where triangular, packed arrangements emerge (see snapshot ($b_6$) in Fig.~\ref{fig:60n_n0}), upon increasing $N_s$ single-particle stripes are just led to merge together (see snapshot ($b_7$) in Fig.~\ref{fig:60n_n0}), to finally form square domains (see snapshot ($b_8$) in Fig.~\ref{fig:60n_n0}), as signaled by the abrupt drop (growth) of $\phi_6$ ($\phi_4$). In such a regime, we see clearly the onset of a strong pattern-induced particle arrangement, which span over large areas of the substrate itself. We stress that at all state points the crystalline arrangements are realized with particles mostly adsorbed on the positive stripes, as discussed in section~\ref{sec:distr}.
Finally, for system 60n close to a +/- substrate, at small $N_s$ slightly less than half of the sample is in a monomer state ($i.e.$, $\langle N_{mon} \rangle \approx 500$), while the remaining particles belong to the crystalline domains formed in correspondence to the positive stripes. At low $N_s$-values, large $\phi_4$-values are observed, suggesting that the particles in crystalline clusters have a square-like symmetry (see snapshots ($c_1$) and ($c_2$) in Fig.~\ref{fig:60n_opposite}). When $N_s \gtrsim10$, square domains break down in favor of double-particle lanes with triangular symmetry -- again on the positive stripes -- and $\phi_4 < \phi_6$ (see snapshots ($c_3$) of Fig.~\ref{fig:60n_opposite}). When $N_s \approx 26$, one single stripe cannot accomodate two fully embedded particles anymore, so single-particle strings form on the positive stripes, while the monomers adsorbed on the negative stripes act as boundaries of these lanes (see snapshot ($c_4$) of Fig.~\ref{fig:60n_opposite}). When $40<N_s< 70$, $\langle N_{mon} \rangle$ drops to zero. Within this region, as long as $N_s \gtrsim50$, single-particle strings on the positive stripes becomes connected $via$ monomers on the negative stripes and a visual inspection of the particle configurations suggests a very rich scenario: networks emerge with properties of that are deeply affected by the substrate pattern. In this case, the behavior of $\langle N_{neigh} \rangle$ is less significative. Additionally, since clusters in this $N_s$-region are not compact anymore, the local bond order parameters are not reliable and a different analysis, which is beyond the scope of this paper, is needed. Finally, for $N_s\gtrsim70$ monomers prevail as shown by the abrupt increase of $\langle N_{mon} \rangle$ (see panel c) of Fig.~\ref{fig:60n_opposite}).
To summarize, system 60n shows a very rich assembly scenario, where according to the charge modulation it is possible to induce the formation of (square-like or triangular-like) double-particle lanes, open networks spanning throughout the sample and pattern-induced square domains.
We now consider system 60c. As anticipated in section~\ref{sec:distr}, for the -/0 pattern particles tend to cluster, irrespective of the substrate type.
The analysis of $\langle N_{mon} \rangle$ and $\langle N_{neigh} \rangle$ suggests that only one crystalline symmetry, the triangular one, is maintained over the whole $N_s$ range: at all $N_s$ values, $\langle N_{mon} \rangle = 0$ and $\langle N_{neigh} \rangle \simeq 5$, suggesting the presence of clusters with a triangular particle arrangement (see panel b) Fig.~\ref{fig:60c_n0}). The relative local bond oder parameters support this scenario, since $\phi_6 \simeq 1$ and $\phi_4 \simeq 0$ at any $N_s$-value (see panel c) of Fig.~\ref{fig:60c_n0}).
In contrast, the +/0 and the +/- patterns show a very similar trend, corresponding to a transition from a regime of purely triangular domains to a regime where triangular-like domains coexist with square-like aggregates. The similarity between the two cases is due to the fact that particles 60c strongly absorb on the positive substrate, that is thus preferred to both the negative and neutral stripes. The transition from one crystalline regime to another can be inferred already from the analysis of $\langle N_{mon} \rangle$ and $\langle N_{neigh} \rangle$ (see panels b) of Fig.s~\ref{fig:60c_p0} and~\ref{fig:60c_opposite}): at low $N_s$ no monomers are found (for both substrate patterns) and particles have on average five neighbors (suggesting a predominance of triangular domains); at large $N_s$ very few monomers are observed and particles have on average 3.5 (+/0 substrate) and 4.5 (+/- substrate) neighbors, suggesting the presence of clusters with mixed symmetry. Between these two regimes $\langle N_{neigh} \rangle$ drops to 2 (together with a mild increase of $\langle N_{mon} \rangle$ to 40), suggesting that the crystalline order is broken and restored. The transition between the two regimes is similar to the one observed for system 60n close to a +/0 pattern: as long as the stripes are sufficiently large, particles form extended crystalline clusters; on increasing $N_s$, particles start to organize in double-particle lanes on the same stripe (see snapshots ($\rm d_3$) of Fig.~\ref{fig:60c_p0} and ($\rm e_3$) of Fig.~\ref{fig:60c_opposite}), then in single-particle strings (see snapshots ($\rm d_4$) of Fig.~\ref{fig:60c_p0} and ($\rm e_4$) of Fig.~\ref{fig:60c_opposite}) and finally at very large $N_s$ these strings form big domains (see snapshots ($\rm d_6$) of Fig.~\ref{fig:60c_p0} and ($\rm e_6$) of Fig.~\ref{fig:60c_opposite}). From the behavior of the local bond order parameters as function of $N_s$, we observe that at $N_s$=90, square-like domains prevail ($\phi_4 \approx 0.4$ while $\phi_6 \approx 0.8$) in the +/0 case, while triangular domains prevail ($\phi_4 \approx 0.8$ while $\phi_6 \approx 0.4$) in the +/- case. The competition between square- and triangular-arrangements is directly related to the balance between energy gains and losses in the particle-substrate and particle-particle interaction: while square domains favor the particle adsorption, triangular domains favor inter-particle bonding. Here, particles do not pay any energetic price when oriented in-plane on a neutral substrate, while the would pay a price for such an orientation on a negative stripe; for this reason more triangular domains are observed on a +/- substrate with respect to the +/0 case.
We observe that, while for system 60n on a +/- substrate the emergence of square domains at high $N_s$-values is compatibile with the behavior of the system at low $N_s$-values (where square domains prevail), the behavior of system 60c at low $N_s$-values (where triangular domains prevail) is quite different from the behavior at high $N_s$-values (where pattern-induced square domains emerge).
Finally, we note that, with respect to system 60n, systems 60c does not form open networks nor square-like double-particle lanes.
\subsection{Robustness of the self-assembly scenario}
\label{sec:var}
We discuss here the robustness of the assembly scenarios presented in sections~\ref{sec:distr} and~\ref{sec:bondorder} by considering small variations of the system parameters. In particular, we investigate the effects of different (i) surface charges and (ii) interaction ranges, the latter being related to the electrostatic screening conditions.
Since our model is not a toy model but a coarse-grained description of the effective interactions, the interaction parameters must be computer anew $via$ the procedure described in section~\ref{sec:model&methods} and reported in more details in Ref.~\cite{bianchi:2011}. The list of parameters considered here is reported in Tab.~\ref{tab:addparams}: for all the systems, $\cos \gamma \sim 60 \degree$, the geometric parameters follows Eq.~(\ref{eq:params}) and the energy constants are calculated following the mapping procedure introduced in Ref.~\cite{bianchi:2011}.
\begin{table*}[htb]
\centering
\begin{tabular}{lccccccccc}
Particle name & $Z_w$ & $\delta$ & $a$ & $u_{CC}$ & $u_{PC}$ & $u_{PP}$ & $u_{CS_{(+)}}$ & $u_{PS_{(+)}}$ & $\varepsilon_{min}$ \\
\hline
60n & 1.0 & 0.20 & 0.1600 & 0.1349 & -0.8483 & 4.3228 & -0.1256 & 0.4462 & -0.0337 \\
60n & 1.5 & 0.20 & 0.1600 & 0.1349 & -0.8483 & 4.3228 & -0.1883 & 0.6693 & -0.0337 \\
60n & 3.0 & 0.20 & 0.1600 & 0.1349 & -0.8483 & 4.3228 & -0.3767 & 1.3385 & -0.0337 \\
60n & 5.0 & 0.20 & 0.1600 & 0.1349 & -0.8483 & 4.3228 & -0.6278 & 2.2310 & -0.0337 \\
60c & 1.0 & 0.20 & 0.1600 & 0.4330 & -1.9467 & 4.3228 & -0.2978 & 0.4462 & -0.0781 \\
60c & 1.5 & 0.20 & 0.1600 & 0.4330 & -1.9467 & 4.3228 & -0.4467 & 0.6693 & -0.0781 \\
60c & 3.0 & 0.20 & 0.1600 & 0.4330 & -1.9467 & 4.3228 & -0.8934 & 1.3386 & -0.0781 \\
60c & 5.0 & 0.20 & 0.1600 & 0.4330 & -1.9467 & 4.3228 & -1.4890 & 2.2310 & -0.0781 \\
\hline
SR & 5.0 & 0.15 & 0.1265 & 0.0798 & -0.5050 & 2.5470 & -0.6694 & 2.3663 & -0.0123 \\
LR & 5.0 & 0.25 & 0.1907 & 0.2160 & -1.3068 & 6.6472 & -0.6019 & 2.1261 & -0.0736 \\
\end{tabular}
\caption{List of the model parameters used to induce variations either in the particle-substrate interaction (upper part) or in the particle-particle interaction (lower part): substrate charge, $Z_w$, interaction range, $\delta$, eccentricity $a$, energy parameter for the core/core (CC), patch-core (PC), patch-patch (PP), core-substrate (CS) and patch-substrate (PS) interaction (in case of positive substrate) and minimum of the attraction, $\varepsilon_{min}$.}
\label{tab:addparams}
\end{table*}
\subsubsection{Effect of $Z_w$}
We first address the effect of a substrate carrying a weaker charge, namely we consider $Z_w = \pm \frac{1}{90} Z_p = \pm 1$, $Z_w = \pm \frac{1.5}{90} Z_p = \pm 1.5$ and $Z_w = \pm \frac{3}{90} Z_p = \pm 3$. Substrates with higher surface charges have not been considered, as both particles 60n and 60c are rather strongly adsorbed on substrates already for $Z_w = \pm 5$. We perform simulations at selected $N_s$-values and selected surface patterns for both systems: namely, we consider system 60n on a +/0 substrate with $N_s=2,38,66$ and 90, system 60n on a +/- substrate with $N_s=2,20,32,38,66$ and 90, and system 60c on a +/- substrate with $N_s=2,42$ and 90. Clearly, at low $Z_w$-values the assembly is mostly driven by the particle-particle interaction, leading to the formation of clusters with triangular arrangements, as this is the preferred configuration in case of a neutral ($i.e.,$ non-adsorbing) substrate for both 60n and 60c systems (see panels a) in Figs.~\ref{fig:surfcharge_n1},~\ref{fig:surfcharge_n2} and~\ref{fig:surfcharge_c}): in all investigated cases, $\phi_4 \approx 0$ and $\phi_6 \approx 1$ over the whole $N_s$-range as long as $Z_w < 3$. As soon as $Z_w \geq $ 3, all systems start to be affected by the substrate, as described in the following.
We start our discussion with system 60n on a +/0 substrate (see Fig.~\ref{fig:surfcharge_n1}). In this case, the bond order parameters at $N_s=2$ for $Z_w=3$ and $Z_w=5$ are exactly the same, meaning that when particles assemble over extended areas of the substrate no significative effect occurs as long as $ 3 \lesssim Z_w \lesssim 5$. This is also supported by a visual comparison of the snapshots (see panel ($b_1$) of Fig.~\ref{fig:60n_p0} and panel ($g_1$) of Fig.~\ref{fig:surfcharge_n1}). When $N_s$ is increased, though, we observe that a less charged substrate favors the formation of triangular domains: even though $\phi_4$ and $\phi_6$ are less significative in the intermediate $N_s$-range (due to the reduced number of large compact clusters), it is worth noting that $\phi_4 \ll \phi_6$ at $N_s$ = 38 and 66 for $Z_w=3$, while for $Z_w=5$ $\phi_4 \approx \phi_6$ at the same $N_s$-values; the relative abundance of triangular particle arrangements for $Z_w=3$ can be also inferred by a visual inspection of the system (see panel ($b_6$) of Fig.~\ref{fig:60n_p0} and panel ($g_2$) of Fig.~\ref{fig:surfcharge_n1}). This effect can be related to a reduced preference of the particles to sit on the positive stripes, where they arrange in square-like aggregates, thus bringing bonding and adsorption on a comparable energetic ground. The competition between triangular and square particle arrangements is also observed at $N_s=90$, where the formation of the pattern-induced square domains is hindered by the persistence of some triangular domains (see panel ($b_8$) of Fig.~\ref{fig:60n_p0} and panel ($g_3$) of Fig.~\ref{fig:surfcharge_n1}), leading to a decreased (increased) $\phi_4$ ($\phi_6$) for $Z_w=3$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_Z_60n_p0.jpg}
\end{center}
\caption{System 60n substrate +/0: local bond order parameters, $\phi_4$f and $\phi_6$, as function of $N_s$ for different $Z_w$-values (as labeled): symbols refer to $Z_w =$ 1, 1.5, 3, while lines refer to $Z_w =$ 5 (reported as a reference). Data are reported for selected $N_s$, namely $N_s=2,38,66$, and 90. Panels ($g_1$)-($g_3$) are snapshots of the system with $Z_w =$ 3 at different number of stripes: ($g_1$) $N_s =$ 2, ($g_2$) $N_s =$ 66, ($g_3$) $N_s =$ 90. }
\label{fig:surfcharge_n1}
\end{figure*}
If we now consider system 60n on a +/- substrate (see Fig.~\ref{fig:surfcharge_n2}) no significative difference can be found between $Z_w=3$ and $Z_w=5$ as long as $N_s \lesssim 38$. In contrast, at $N_s=90$, we observe a striking difference between the two cases: while for $Z_w=5$ no aggregation occurs, for $Z_w=3$ particles mostly assemble into triangular clusters. In the first case, adsorption to the negative stripes is preferred over bonding: as soon as the interaction with the substrate is weak, the particle-particle based assembly sets in. Here we do not discuss the effect of $Z_w$ on the network forming region at intermediate $N_s$-values as first a deeper analysis would be needed to elucidate the properties of the emerging networks, but this is out of the scope of the present paper.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_Z_60n_pn.jpg}
\end{center}
\caption{System 60n substrate +/-: local bond order parameters, $\phi_4$f and $\phi_6$, as function of $N_s$ for different $Z_w$-values (as labeled): symbols refer to $Z_w =$ 1, 1.5, 3, while lines refer to $Z_w =$ 5 (reported as a reference). Data are reported for selected $N_s$, namely $N_s=2,20,32,38,66$, and 90. Panels ($h_1$)-($h_3$) are snapshots of the system with $Z_w =$ 3 at different number of stripes: ($h_1$) $N_s =$ 2, ($h_2$) $N_s =$ 32, ($h_3$) $N_s =$ 90.}
\label{fig:surfcharge_n2}
\end{figure*}
Finally, we consider system 60c on a +/- substrate (see Fig.~\ref{fig:surfcharge_c}): similar to the 60n system on a +/- substrate, we observe that triangular domains are favored when $Z_w=3$. While for $Z_w=5$ we observe triangular domains at low $N_s$-values (see snapshot ($e_1$) in Fig.~\ref{fig:60c_opposite}) and a coexistence between triangular and square domains at large $N_s$-values (see snapshot ($e_8$) in Fig.~\ref{fig:60c_opposite}), for $Z_w=3$ we observe triangular domains at both low and large $N_s$-values (see the bond order parameters in panel a) of Fig.~\ref{fig:surfcharge_c} and related snapshots ($f_1$) and ($f_3$)), with no presence of square-like domains at $N_s=90$.
In summary, when the surface charge of the substrate is reduced, bonding is favored with respect to adsorption and thus, the investigated systems tend to assemble into those crystalline domains that emerge spontaneously on a homogeneous, neutral substrate. For systems 60n and 60c on a +/- substrate, the threshold between the two regimes is $Z_w \approx 3$: at such a value of $Z_w$ system 60n on a +/- substrate forms triangular clusters instead of monomers, while system 60c forms triangular domains instead of a mixed phase of triangles and squares. For system 60n on a +/0 substrate in order to observe the triangular domains in place of the pattern-induced square aggregates the surface charge must be very low, $i.e.$, $Z_w \ll 3$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_Z_60c_pn.jpg}
\end{center}
\caption{System 60c substrate +/-: local bond order parameters, $\phi_4$f and $\phi_6$, as function of $N_s$ for different $Z_w$-values (as labeled): symbols refer to $Z_w =$ 1, 1.5, 3, while lines refer to $Z_w =$ 5 (reported as a reference). Data are reported for selected $N_s$, namely $N_s=2,42$, and 90. Panels ($f_1$)-($f_3$) are snapshots of the system with $Z_w =$ 3 at different number of stripes: ($f_1$) $N_s =$ 2, ($f_2$) $N_s =$ 42, ($f_3$) $N_s =$ 90.}
\label{fig:surfcharge_c}
\end{figure*}
\subsubsection{Effect of interaction range}
We address here the effect of small variations in the interaction range of system 60n. To this aim, we introduce two new systems: they are both neutral IPCs with patch size defined by $\cos \gamma \approx 60 \degree$ and they differ in their interaction range. The system labelled as SR (short range) has $\delta = $ 0.15, the system labelled LR (long range) has $\delta$ = 0.25. We perform Monte Carlo simulations for system 60n on both +/0 and +/- substrates at selected state points, namely at $N_s =$ 2, 26, 38, 60, 90.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60n_delta_positive.jpg}
\end{center}
\caption{a) Bond order parameters $\phi_4$ and $\phi_6$ for systems 60n (reported as a reference), SR and LR (as labelled) on a +/0 substrate as function of the number of stripes $N_s$ . b) Average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as function of $N_s$ for systems SR and LR (as labelled). Panels ($m^1_1$)-($m^1_4$) are snapshots of SR systems, ($m^2_1$)-($m^2_4$) are snapshots of LR systems, both at different number of stripes: ($m^1_1$) ($m^2_1$) $N_s =$ 2, ($m^1_2$) ($m^2_2$) $N_s =$ 38, ($m^1_3$) ($m^2_3$) $N_s =$ 60, ($m^1_4$) ($m^2_4$) $N_s =$ 90.}
\label{fig:delta_p0}
\end{figure*}
In Figure~\ref{fig:delta_p0}, we report the behavior of both SR and LR systems on a +/0 substrate, together with the reference 60n system. We observe that system SR behaves very similar to system 60n: square domains emerge both at low and high $N_s$-values, $\langle N_{mon} \rangle$ is negligible over the whole $N_s$-range with a mild peak at $N_s \approx 60$ where particles are organized in single-particle strings, while $\langle N_{neigh} \rangle$ fluctuates between four and 3.5 (see panels a) and b) of Fig.~\ref{fig:delta_p0} and the snapshots ($m^1_1$)-($m^1_4$)). In contrast, system LR displays a different assembly scenario. In this case, for $N_s \lesssim 20$, $\phi_4 \approx \phi_6$: a visual inspection of the configurations shows the presence of square domains on the positive stripes and triangular domains on the neutral stripes, whereas at shorter interaction ranges the neutral stripes were completely depleted. This observation already suggests that the adoption to the positive stripes at large $N_s$-values is less effective, thus leading to a competition between (less favored) square-like particles (adsorbed on positive stripes) and triangular-like particles (emerging on neutral -- non adsorbing -- stripes). On further increasing $N_s$ the tendency of the particles to form domains with triangular symmetry persists and increases, as signaled by the increase of $\langle N_{neigh} \rangle$ above four at large $N_s$-values and by the visual analysis (see snapshots($m^2_2$)-($m^2_3$) in Fig.~\ref{fig:delta_pn}). At $N_s =$ 90, we observe a competition between triangular and square-induced arrangements (see snapshot ($m^2_4$) in Fig.~\ref{fig:delta_pn}), while for system 60n and SR the latter
dominates.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.75\textwidth]{IMAGES/panel_60n_delta_opposite.jpg}
\end{center}
\caption{a) Bond order parameters $\phi_4$ and $\phi_6$ for systems 60n (reported as a reference), SR and LR (as labelled) on a +/- substrate as function of the number of stripes $N_s$ . b) Average number of monomers $\langle N_{mon} \rangle$ (full symbols) and average number of neighbors $\langle N_{neigh} \rangle$ (empty symbols) as function of $N_s$ for systems SR and LR (as labelled). Panels ($n^1_1$)-($n^1_4$) are snapshots of SR systems, ($n^2_1$)-($n^2_4$) are snapshots of LR systems, both at different number of stripes: ($n^1_1$) ($n^2_1$) $N_s =$ 2, ($n^1_2$) ($n^2_2$) $N_s =$ 38, ($n^1_3$) ($n^2_3$) $N_s =$ 60, ($n^1_4$) ($n^2_4$) $N_s =$ 90.}
\label{fig:delta_pn}
\end{figure*}
In Figure \ref{fig:delta_pn}, we report the behavior of both system SR and LR on a +/- substrate (again the case 60n is reported for comparison). We observe here a scenario similar to the one described for the +/0 case: system SR and 60n behaves in the same way, whereas LR displays a different assembly scenario as soon as $N_s \gtrsim$ 30. In particular, system LR does not form a proper network phase at any $N_s$ investigated, but it tends to form more compact clusters that span over many consecutive stripes, as shown in the snapshot ($n^2_3$) of Fig.~\ref{fig:delta_pn}. Moreover, for very high $N_s$, LR particles assemble into clusters characterized by triangular domains instead of dispersing as monomers on the negative stripes, in contrast to what happens at this $N_s$ for both 60n and SR.\\
In general, longer interaction ranges favor a particle-particle interaction driven assembly and thus the formation of triangular domains, while shorter interaction ranges favor adsorption with respect to bonding.
\subsection{Visual inspection of the configurations}
\label{sec:visual}
So far we have discussed our results in terms of the spacial arrangement of the particles. Given the anisotropy of the particle surface, a closer look at the orientational particle arrangements can provide additional insight into the behavior of the investigated systems. Here we do not perform a quantitative analysis of the particle orientations -- since it is beyond the scope of this paper -- but rather highlight the trends that can be observed in the bonding patterns of the different aggregation regimes described in sections~\ref{sec:distr},~\ref{sec:bondorder} and~\ref{sec:var}.
As described in section~\ref{sec:model&methods}, we consider IPCs in a quasi two-dimensional slab: particles are thus free to rotate in three dimensions. If we define the particle orientation as the vector joining the particle center and one of the two patches, we have that isolated particles orient themselves (i) parallel to the $xy$-plane on a positive surface or (ii) perpendicular to the $xy$-plane on a negative surface; as soon as the particle interact with other IPCs its orientation is the result of a trade-off between the bonding energy with its neighbors and the bonding energy with its substrate.
A detailed analysis of the energy balance between these two contributions has been carried for systems 60n and 60c on homogeneous substrates in Ref.~\cite{bianchi:2d2014}. Here we briefly summarize the bonding patterns observed in Ref.~\cite{bianchi:2d2014} for the purpose of our discussion (see Fig.~\ref{fig:old} as reference): in the monomer phase particles are perpendicular to the $xy$-plane; in square domains particles bond in a T-shape fashion with polar and equatorial regions of neighboring particles at contact; in triangular domains two types of bonding patterns emerge, defined as flower-like and grain-like patterns. Flower-like patterns emerge in system 60n and are characterized by rings of six particles oriented in the $xy$-plane with one up-right particle in their center. Grain-like patterns characterize system 60c and consist of triangular domains where particles are predominantly oriented along the horizontal direction and form an angle of approximately $60^{\degree}$ with the symmetry axes of the neighboring particles.
In view of the previous analysis we define here three types of bonding between IPCs (see top three panels in Fig.~\ref{fig:orientations}): T-bonds, G-bonds and F-bonds.
T-bonds are in-plane equatorial-polar contacts, they occur when the adsorption on positive stripes drives the assembly of the systems: when particles belong to compact clusters, then the T-like configuration leads to four bonds per particle (see panels I and II of Fig.~\ref{fig:orientations}), while when they belong to double-particle lanes, particles in T-like configurations form on average three bonds (see panel III of Fig.~\ref{fig:orientations}). We observe that, in contrast to square domains formed on homogeneous positive substrates (panel I), pattern-induced square domains have a fixed orientation throughout the sample (panel II).
G-bonds are in-plane grain-like contacts, they occur in system 60c, while they do not emerge for system 60n. In Fig.~\ref{fig:orientations} we report a snapshot of double-particles lanes with a G-bonds pattern, while for extended compact domain we refer to Ref.~\cite{bianchi:2d2014}. We observe that, when triangular and square domains coexist for system 60c at large $N_s$-values, G-bonds and T-bonds patterns emerge in the compact clusters, where T-like lanes are adsorbed only on the positive stripes, while G-like lanes can be on both stripe-types (see panel V of Fig.~\ref{fig:orientations}).
Finally, F-bonds are out-of-plane equatorial-polar -- or flower-like -- contacts, they occur in system 60n and they are responsible for a large range of aggregation scenarios, as discussed in the following. This is due to the fact that 60n particles on a negative substrate tend to orient themselves up-right (see panel VI of Fig.~\ref{fig:orientations}), possibly acting as bridges between particles adsorbed in-plane. In particular, for system 60n on a +/- substrate, when the stripes can accomodate exactly two particles, the T-bonds pattern on the positive stripes disappears in favor of double-particle lanes in a G-like fashion, where F-like particles adsorbed on the negative stripes set the boundaries of the clusters: two bridge-particles in an up-right position can sit on the same stripe, thus effectively forcing the clusters to grow along the $y$ direction (see panel VIII of Fig.~\ref{fig:orientations}). When stripes are such that a single particle cannot be fully embedded, we observe either F-like extended domains (on a +/- substrate) or a coexistence between square and F-like triangular domains (on a +/0 substrate). In both cases up-right particles lie on the negative substrate (see panel IX of Fig.~\ref{fig:orientations}) and act as bridge between two G-like compact clusters, thus effectively allowing the growth of the clusters along the $x$ direction.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.4\textwidth]{IMAGES/orientations}
\end{center}
\caption{Particles in different bonding configurations. Upper row, from left to right: T-like bonds, F-like bonds and G-like bonds. I: Interaction-induced square domains (from snapshot ($\rm b_2$) of Fig.~\ref{fig:60n_p0}). II: Substrate-induced, square domains (from snapshot ($\rm b_8$) of Fig.~\ref{fig:60n_p0}). III: Double-particle lanes with T-like bonds (from snapshot ($\rm b_3$) of Fig.~\ref{fig:60n_opposite}). IV: Double-particle lanes with F-like bonds (from snapshot ($\rm c_3$) of Fig.~\ref{fig:60n_opposite}). IV: Double-particle lanes with G-like bonds (from snapshot ($\rm d_3$) of Fig.~\ref{fig:60n_p0}). V: Mixed square and G-like triangular domains (from snapshot ($\rm e_6$) of Fig.~\ref{fig:60n_opposite}). VI: Monomers (from snapshot ($\rm a_4$) of Fig.~\ref{fig:60n_n0}). VII: F-like triangular domains (from snapshot ($\rm g_2$) of Fig.~\ref{fig:delta_p0}), XI: Mixed square and F-like triangular domains (from snapshot ($\rm m_4^2$ in Fig.~\ref{fig:delta_p0}). X and XI: Networks for system 60n on substrate +/- (from snapshots ($c_5 $) and ($c_6$) of Fig.~\ref{fig:60n_opposite}).
}
\label{fig:orientations}
\end{figure}
Finally, the interplay between the different anisotropic interactions and the length scales in system 60n can give rise to even more complex scenarios: panels X and XI of Fig.~\ref{fig:orientations} show examples of open clusters, characterized by strings of particles with alternating up-right and planar orientation. The structure of these networks seems to be very sensitive to the substrate pattern and a more focused investigation is needed.
\section{Conclusions}\label{sec:conclusions}
Within the framework of materials design, we have investigated how to create colloidal monolayers on substrates by tuning the surface properties of both the particles and the substrate. In particular, we have considered colloids with a charged, non-homogeneous surface close to a planar wall characterized by parallel stripes with alternating surface charge. By tuning the competition between the particle-particle and the particle-substrate interaction as well as between the particle size and the width of the stripes, we have shown how to control the spacial ordering of the colloids and how to drive the systems from compact, to elongated or open aggregates.
We have considered negatively charged colloids with two positively charged polar regions, the size of these regions being relatively large, the net charge of the particles being zero or slightly negative, with a relatively short interaction range. These systems were shown to give rise to a rich assembly scenario already close to a homogeneous substrate: crystalline domains with different spatial (square $versus$ triangular) arrangements or different orientational bonding (flower-like $versus$ grain-like) patterns as well as monomers with a well-defined orientation were observed on changing the charge ratios between the differently charged surfaces. Here we have investigated the effect of properly designed substrate motifs. In particular, we have considered neutral, positive or negative stripes in an alternating pattern: negative/neutral, positive/neutral or positive/negative.
One striking effect of the substrate pattern is the robust formation of pattern-induced crystalline domains with square symmetry: these particle arrangements emerge for both overall neutral and charged particles as soon as (i) particles are preferentially adsorbed on the positive stripes and (ii) the width of the stripes becomes much smaller than the particle size. In some cases, these pattern-induced square domains compete with (flower- as well as grain-like) triangular domains but it is possibile to control the energy balance between adsorption (that favor square domains) and bonding (that favors triangular domains) by, $e.g.$, changing the substrate surface charge or the screening conditions. On reducing the substrate charge, both grain- and flower-like triangular domains are highly favored with respect to square domains (for system 60c and 60n on a +/0) or to monomers (for system 60n on a +/- substrate). Changing the electrostatic screening conditions implies variation in the interaction range: we observe that long range interactions favor particle-particle bonding, while short range interactions enhance particle adsorption. On reducing the (particle-particle and particle-substrate) interaction range, we thus observe that pattern-induced scenarios prevail, meaning that a monomer phase (if particles prefer to adsorb on negative stripes) or a square pattern (if particles tend to adsorb on positive stripes) can prevail over the triangular arrangement, while for longer interaction ranges triangular domains or clusters with mixed symmetry are observed.
In summary, double particle lanes with square or triangular symmetry, elongated (along the $y$ axis) versus extended (along the $x$ axis) clusters and pattern-induced crystals with a well-defined orientation can be induced by playing with the competition between different anisotropic interactions and length scales. The balance between adsorption and bonding is affected my the many parameters of the systems and the extended simulations carried here show that non-intuitive behaviors can emerge.
We have focused mostly on the spacial arrangement of the particles. Future work should address, $e.g.,$ the orientational features of the bonding patterns within the clusters, the compactness and the shape of the aggregates, their preferred direction with respect to the substrate motif as well as their percolating properties.
Moreover, it would be interesting to gradually release the confinement towards bulk systems: the aggregation of colloidal particles on patterned substrates could be used to guide the crystallization of bulk colloidal crystals as well as to tailor the orientation and size of the resulting lattices. This process is referred to as ``colloidal epitaxy"~\cite{colloidalepitaxy,Dias2013,Dias2017}. We note, for instance that, in contrast to the square arrangements assembled on a homogeneous positive substrate, pattern-induced square domains have a well-defined crystallographic axis, which may be used as templates to assemble perfect square crystals on large (possibly macroscopic) scales.
\section{Acknowledgements}
The authors wishes to thank Benedikt Vitecek for preliminary results. EB acknowledges support from the Austrian Science Fund (FWF) under Proj. Nos. V249-N27. Computer time at the Vienna Scientific Cluster (VSC) is also gratefully acknowledged.
\bibliographystyle{rsc}
|
\section{Introduction}
A perfect matching, or a dimer configuration on a graph is a subset of edges such that each vertex is incident to exactly one edge in the subset. Dimer configurations appear naturally in statistical physics to model the structure of matter, for example, the perfect matchings on the hexagon lattice is a mathematical model for the molecule structure of graphite. With explicit combinatorial correspondence, the dimer model is also closely related to other lattice models in statistical mechanics, including the Ising model (\cite{ZL12,ZLsp}), the 1-2 model (\cite{ZL14,ZL141,GL15,GL17}) and a general polygon model (\cite{GLPO}). By developing the technique of Kasteleyn, Temperley and Fisher (\cite{Ka61,TF61}), the partition function (weighted sum of configurations) of dimer configurations on a finite plane graph can be expressed explicitly as the determinant or pfaffian of a weighted adjacency matrix; the local statistics can be computed (\cite{Ken01}). By study the spectral curve of the periodic dimer model using algebraic geometry technique, the sharp phase transition result can be established (\cite{KOS,Ken06}). The asymptotics of the rescaled dimer height function on a graph approximating a simply-connected domain can also studied by a variational principle (\cite{CKP,KO}); and also by the asymptotics of certain symmetric functions (see \cite{OR01,AB07,AB11,LP14,LP142,GP15,bg,bk,BL17,Li18}).
The Gaussian free field (GFF) is a high-dimensional time analogue of Brownian motion. The main aim of this paper is to investigate the connection between the height fluctuations of the dimer model on a contracting square-hexagon lattice and the Gaussian free field. It was first shown in \cite{RK00,RK01} that the (non-rescaled) height function for the dimer model with uniform underlying measure on a simply-connected square grid with Temperley boundary condition converge to a GFF in distribution. The result was later proved for the whole-plane isoradial graph (\cite{BdT07}) and the simply-connected isoradial graph with Temperley boundary condition (\cite{Li13}). For boundary conditions other than the Temperley boundary condition, the convergence of height fluctuation for the dimer model with uniform underlying measure on a contracting hexagon lattice to GFF was proved in \cite{LP142}; the corresponding result on a Aztec diamond (contracting square grid) with uniform underlying measure was proved in \cite{bk}, by analyzing the asymptotics of the Schur function in a neighborhood of $(1,1,\ldots,1)$ (\cite{bg16}).
A related model is the dimer model on a contracting square-hexagon lattice whose underlying measure depends on periodically assigned edge weights with period $1\times n$. In \cite{BL17,Li18}, we studied this model by establishing an identity of the partition function of the dimer model on such a graph and the value of the Schur function depending on edge weights; and then analyze the asymptotics of the Schur function in a neighborhood of a generic point $(x_1,\ldots,x_N)$. The law of large numbers for the rescaled height function was proved for two specific boundary conditions on the bottom boundary (1) each remaining vertex on the boundary is followed by $(m-1)$ removed vertices, where $m\geq 2$ is a positive integer; (2) the bottom boundary is divided to alternate line segments with either all vertices removed or all the vertices preserved in each segment. We shall call the first boundary condition the \textbf{uniform boundary condition} and the second boundary condition the \textbf{piecewise boundary condition}. In this paper, we study the non-rescaled height fluctuations for the dimer model on the contracting square-hexagon lattice with the above two boundary conditions, and show that they converge to GFF in certain sense, building on the analysis of the Schur function at a generic point (see \cite{BL17,Li18}) and the techniques to relate fluctuations of particle systems determined by Schur generating functions and GFF developed in \cite{bg16}.
The organization of the paper is as follows. In Sect.~\ref{bkg}, we introduce the contracting square-hexagon lattice and the main technical tools used in this paper. In Sect.~\ref{unr}, we introduce the uniform boundary conditions and review the limit shape result for the dimer model on a contracting square hexagon lattice with $1\times n$ periodic edge weights and the uniform boundary conditions. In Sect.~\ref{unc}, we prove that certain statistics constructed from the dimer model on a contracting square hexagon lattice with $1\times n$ periodic edge weights and the uniform boundary conditions convergence to Gaussian distribution in the scaling limit. In Sect.~\ref{pbr}, we introduce the piecewise boundary conditions and review the limit shape result for the dimer model on a contracting square hexagon lattice with $1\times n$ periodic edge weights and the piecewise boundary conditions. In Sect.~\ref{pbc}, we prove that certain statistics contracted from the dimer model on a contracting square hexagon lattice with $1\times n$ periodic edge weights and the piece boundary conditions convergence to a sum of finitely many independent Gaussian random variables in the scaling limit; where the number of independent Gaussian random variables depends on the size of the period $n$. In Sect.~\ref{gffu}, we show that the statistics constructed from the dimer model on a contracting square hexagon lattice with $1\times n$ periodic edge weights and the uniform boundary conditions convergence to GFF in the upper half plane, under an homeomorphism from the liquid region to the upper half plane.
In Sect.~\ref{gffp}, we show that the statistics constructed from the dimer model on a contracting square hexagon lattice with $1\times n$ periodic edge weights and the piecewise boundary conditions convergence to a sum of $n$ independent GFFs in the upper half plane.
\section{Background}\label{bkg}
In this section, we define a general class graph (the contracting square-hexagon lattice) on which the height fluctuations of the dimer model is studied in this paper. Dimer model on such graphs has been studied in (\cite{BF15,bbccr,BL17,Li18}), and the limit shape result was explicitly established. We also review the main technical tools used in this paper, including the Schur function, the Young diagram, etc.
\subsection{Square-hexagon Lattices}Consider a doubly-infinite binary sequence indexed by integers $\ZZ=\{\ldots,-2,-1,0,1,2,\ldots\}$.
\begin{eqnarray}
\check{c}=(\ldots,c_{-2},c_{-1},c_0,c_1,c_2,\ldots)\in\{0,1\}^{\ZZ}.\label{ca}
\end{eqnarray}
The \textbf{whole-plane square-hexagon lattice} associated with the sequence $\check{c}$, is a bipartite plane graph $\mathrm{SH}(\check{c})$ defined as follows. Its vertex set is a subset of $\frac{\ZZ}{2}\times \frac{\ZZ}{2} $. Each vertex of $\mathrm{SH}(\check{c})$ is either black or white, and we identify the vertices with points on the plane.
For $m\in \ZZ $, the black vertices have $y$-coordinate $m$; while the white vertices have $y$-coordinate $m-\frac{1}{2}$. We will label all the vertices with coordinate $m$ as vertices in the $(2m)$th row, and all the vertices with coordinate $m-\frac{1}{2}$ as vertices in the $(2m-1)$th row. We further require that
\begin{itemize}
\item each black vertex in the $(2m)$th row is adjacent to two white vertices in the $(2m+1)$th row; and
\item if $c_m=1$, each white vertex on the $(2m-1)$th row is adjacent to exactly one black vertex in the $(2m)$th row;
if $c_m=0$, each white vertex on the $(2m-1)$th row is adjacent to two black vertices in the $(2m)$th row.
\end{itemize}
See Figure \ref{lcc}.
\begin{figure}
\subfloat[Structure of $\mathrm{SH}(\check{c})$ between the $(2m)$th row and the $(2m+1)$th row]{\includegraphics[width=.6\textwidth]{bw}}\\
\subfloat[Structure of $\mathrm{SH}(\check{c})$ between the $(2m-1)$th row and the $(2m)$th row when $c_m=0$]{\includegraphics[width = .6\textwidth]{wb0}}\\
\subfloat[Structure of $\mathrm{SH}(\check{c})$ between the $(2m-1)$th row and the $(2m)$th row when $c_m=1$]{\includegraphics[width = .55\textwidth]{wb1}}
\caption{Graph structures of the square-hexagon lattice on the $(2m-1)$th, $(2m)$th, and $(2m+1)$th rows depend on the values of $(c_m)$. Black vertices are along the $(2m)$th row, while white vertices are along the $(2m-1)$th and $(2m+1)$th row.}
\label{lcc}
\end{figure}
Note that for any $\check{c}\in \{0,1\}^{\ZZ}$, the faces of $\SH(\check{c})$ is either a square or a hexagon. if $c_i=0$ for all $i\in\ZZ$, $\SH(\check{c})$ is a square grid; while if $c_i=1$ for all $\SH(\check{c})$ is a hexagonal lattice. See Figure \ref{fig:SH} for an example of a square-hexagon lattice.
\begin{figure}
\includegraphics{SH}
\caption{Contracting square-hexagon lattice with $N=3$, $m=3$, $\Omega=(1,3,6), (c_1,c_2,c_3)=(1,0,1)$.}
\label{fig:SH}
\end{figure}
We shall assign edge weights to the whole-plane square-hexagon lattice $\SH(\check{c})$ satisfying the following assumption; see Figure \ref{lcc}.
\begin{assumption}\label{apew} For $m\geq 1$, we assign weight $x_m>0$ to each NE-SW edge joining the $(2m)$th row to the $(2m+1)$th row of $\mathrm{SH}(\check{c})$. We assign weight $y_m>0$ to each NE-SW edge joining the $(2m-1)$th row to the $(2m)$th row of $\mathrm{SH}(\check{c})$, if such an edge exists. We assign weight $1$ to all the other edges.
\end{assumption}
A \textbf{contracting square-hexagon lattice} is built from a whole-plane square-hexagon lattice as follows:
\begin{definition}\label{dfr}Let $N\in \NN$. Let $\Omega=(\Omega_1,\ldots,\Omega_N)$ be an $N$-tuple of positive integers, such that $1=\Omega_1<\Omega_2<\cdots<\Omega_{N}$. Set $m=\Omega_N-N$.
The contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$ is a subgraph of $\mathrm{SH}(\check{c})$ built of $2N$ or $2N+1$ rows.
The rows of $\mathcal{R}(\Omega,\check{c})$ inductively, starting from the bottom, can be enumerated as follows:
\begin{itemize}
\item The first row consists of vertices $(i,j)$ with $i=\Omega_1-\frac{1}{2},\ldots,\Omega_N-\frac{1}{2}$ and $j=\frac{1}{2}$. We call this row the boundary row of $\mathcal{R}(\Omega,\check{c})$.
\item When $k=2s$, for $s=1,\ldots N$, the $k$th row consists of vertices $(i,j)$ with $j=\frac{k}{2}$ and incident to at least one vertex in the $(2s-1)th$ row of the whole-plane square-hexagon lattice $\mathrm{SH}(\check{c})$ lying between the leftmost vertex and rightmost vertex of the $(2s-1)$th row of $\mathcal{R}(\Omega,\check{c})$
\item When $k=2s+1$, for $s=1,\ldots N$, the $k$th row consists of vertices $(i,j)$ with $j=\frac{k}{2}$ and incident to two vertices in the $(2s)$th row of of $\mathcal{R}(\Omega,\check{c})$.
\end{itemize}
\end{definition}
\subsection{Partitions, Young diagrams and Schur functions}
We denote by $\GT_N$ the set of $N$-tuples $\lambda$ of integers satisfying $\lambda_1\geq \lambda_2\ldots\geq \lambda_N$, and let $\GT_N^+$ be a subset of $\GT_N$ consisting of all the $\lambda$'s in $\GT_N$ such that $\lambda_N\geq 0$. For $\lambda\in \GT_N^+$, Let
\begin{eqnarray*}
|\lambda|:=\sum_{i=1}^{N}\lambda_i.
\end{eqnarray*}
A graphic way to represent a non-negative signature $\mu$ is through its
\emph{Young diagram} $Y_\lambda$, a collection of $|\lambda|$ boxes arranged on
non-increasing rows aligned on the left: with
$\lambda_1$ boxes on the first row, $\lambda_2$ boxes on the second row,\dots $\lambda_N$
boxes on the $N$th row. Note that elements in $\GT_N^+$ are in bijection with all the Young diagrams with $N$ rows (rows are allowed to have zero length).
\begin{definition}
Let $Y,W$ be two Young diagrams. We say that $Y\subset W$ \emph{differ by a
horizontal strip} if the
collection of boxes in $Z=W\setminus Y$ contains at most one box in every
column. We say that they \emph{differ by a vertical strip} if $Z$ contains at
most one box in every row.
We say that two non-negative signatures $\lambda$ and $\mu$ \emph{interlace}, and
write $\lambda \prec \mu$ if $Y_\lambda\subset Y_\mu$ differ by a horizontal
strip. We say they \emph{cointerlace} and write $\lambda\prec'\mu$ if
$Y_\lambda\subset Y_\mu$ differ by a vertical strip.
\end{definition}
\begin{definition}
Let $\lambda\in\GT_N^+$ be a partition of length $N$. We define the
counting measure $m(\lambda)$ corresponding to $\lambda$ as follows.
\begin{equation}
m(\lambda)=\frac{1}{N}\sum_{i=1}^{N}\delta\left(\frac{\lambda_i+N-i}{N}\right).\label{ml}
\end{equation}
\end{definition}
\begin{definition}Let $\lambda\in \GT_N$. The rational Schur function is
\begin{eqnarray*}
s_{\lambda}(u_1,\ldots,u_N)=\frac{\det_{i,j=1,\ldots,N}(u_i^{\lambda_j+N-j})}{\prod_{1\leq i<j\leq N}(u_i-u_j)}
\end{eqnarray*}
\end{definition}
\subsection{Dimer model}
\begin{definition}\label{dfvl}A dimer configuration, or a perfect matching $M$ of a contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$ is a set of edges $((i_1,j_1),(i_2,j_2))$, such that each vertex of $\mathcal{R}(\Omega,\check{c})$ belongs to an unique edge in $M$.
The set of perfect matchings of $\mathcal{R}(\Omega,\check{c})$ is denoted by
$\mathcal{M}(\Omega,\check{a})$.
\end{definition}
\begin{definition}
Let $M\in \mathcal{M}(\Omega,\check{c})$ be a perfect matching of
$\mathcal{R}(\Omega,\check{c})$. We call an edge $e=((i_1,j_1),(i_2,j_2))\in
M$ a \emph{$V$-edge} if $\max\{j_1,j_2\}\in\NN$ (i.e.\@ if its higher
extremity is black) and we call it a \emph{$\Lambda$-edge}
otherwise. In other words, the edges going upwards starting from an odd row
are $V$-edges and those ones starting from an even row are $\Lambda$-edges. We
also call the corresponding vertices-$(i_1,j_1)$ and $(i_2,j_2)$ $V$-vertices
and $\Lambda$-vertices accordingly.
\end{definition}
\begin{definition} The partition function of the dimer model of a finite graph
$G$ with edge weights $(w_e)_{e\in E(G)}$ is given by
\begin{equation*}
Z=\sum_{M\in \mathcal{M}}\prod_{e\in M}w_e,
\end{equation*}
where $\mathcal{M}$ is the set of all perfect matchings of $G$. The
Boltzmann dimer probability measure on $M$ induced by the weights $w$ is
thus defined by declaring that probability of a perfect matching is equal to
\begin{equation*}
\frac{1}{Z}\prod_{e\in M} w_e.
\end{equation*}
\end{definition}
We shall associate to each perfect matching in $\mathcal{M}(\Omega,\check{a})$
a sequence of non-negative signatures, one for each row of the graph.
\begin{construction}\label{ct}To the boundary row $\Omega=(\Omega_1<\cdots<\Omega_N)$ of a contracting
square-hexagon lattice is naturally associated a non-negative signature $\omega$
of length $N$ by:
\begin{equation*}
\omega=(\Omega_N-N,\dotsc,\Omega_1-1).
\end{equation*}
Let $j\in\{2,\dots,2N+1\}$. Assume that the $j$th row of
$\mathcal{R}(\Omega,\check{a})$ has $n_j$ V-vertices and $m_j$
$\Lambda$-vertices. The a dimer configuration at the $j$th row of $\mathcal{R}(\Omega,\check{a})$
corresponds to a signature $\mu\in \GT_{n_j}^+$, such that
\begin{itemize}
\item $\mu=(\mu_1,\ldots,\mu_{n_j})$;
\item We label all the $V$-vertices on the $j$th row by the 1st $V$-vertex, the 2nd $V$-vertex, \ldots, the $n_j$th $V$-vertex,
such that the $1$st $V$-vertex is the rightmost $V$-vertex on the $j$th row.
for $1\leq k\leq n_j$, $\mu_k$ is the number of $\Lambda$-vertices to the left of the $k$th $V$-vertex.
\end{itemize}
\end{construction}
Then we have
\begin{theorem}[\cite{BL17} Theorem 2.13]
For given $\Omega$, $\check{c}$, let $\omega$ be the signature associated to
$\Omega$. Then the construction~\ref{ct} defines a
bijection between the set of perfect matchings
$\mathcal{M}(\Omega,\check{c})$ and the set $S(\omega,\check{c})$ of
sequences of non-negative signatures
\begin{equation*}
\{(\mu^{(N)},\nu^{(N)},\dots,\mu^{(1)}, \nu^{(1)}, \mu^{(0)}\}
\end{equation*}
where the signatures satisfy the following properties:
\begin{itemize}
\item All the parts of $\mu^{(0)}$ are equal to 0;
\item The signature $\mu^{(N)}$ is equal to $\omega$;
\item For $0\leq i\leq N$, $\mu^{(i)}\in \GT_i^{+}$.
\item The signatures satisfy the following (co)interlacement relations:
\begin{equation*}
\mu^{(N)} \prec' \nu^{(N)} \succ \mu^{(N-1)} \prec' \cdots
\mu^{(1)} \prec' \nu^{(1)} \succ \mu^{(0)}.
\end{equation*}
\end{itemize}
Moreover, if $a_m=1$, then $\mu^{(N+1-k)}=\nu^{(N+1-k)}$.
\label{myb}
\end{theorem}
The following proposition, proved in \cite{BL17}, shows that the partition function of dimer configurations on a contracting square-hexagon lattice can be computed by a Schur function depending on the boundary condition and the edge weights. Therefore it opens the door for investigating the asymptotics of periodic dimer model on a contracting square-hexagon lattice by studying the corresponding Schur functions.
\begin{proposition}\label{p16}Let $\mathcal{R}(\Omega,\check{c})$ be a contracting square-hexagon lattice built from a whole-plane square-hexagon lattice $\SH(\check{c})$ with edge weights $\{x_i,y_i,1\}_{1\leq i\leq N}$ assigned as in Assumption \ref{apew}.
Let
\begin{eqnarray*}
I_2=\{i|i\in\{1,2,\ldots,N\}, c_i=0\}
\end{eqnarray*}
Then
the partition function for perfect matchings on $\mathcal{R}(\Omega,\check{c})$ is given by
\begin{eqnarray*}
Z=\left[\prod_{i\in I_2}\Gamma_i\right] s_{\omega}(x_{1},\ldots,x_{N})
\end{eqnarray*}
where $\omega$ is the $N$-tuple corresponding to the boundary row of $\mathcal{R}(\Omega,\check{c})$, and $\Gamma_i$ is defined by
\begin{eqnarray}
\Gamma_i=\prod_{t=i+1}^{N}\left(1+y_{i}x_{t}\right).\label{gi}
\end{eqnarray}
\end{proposition}
\begin{proof}See Proposition 2.18 of \cite{BL17}.
\end{proof}
\section{Uniform Boundary Conditions}\label{unr}
In this section, we introduce the uniform boundary conditions on the bottom boundary of a contracting square hexagon lattice, and review the limit shape result of the dimer model on such a lattice.
Consider a contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$ with edge weights assigned as in Assumption \ref{apew}. Suppose that the configuration on the bottom row corresponds to the following signature
\begin{eqnarray}
\lambda(N)=((m-1)(N-1),(m-1)(N-2),\ldots,(m-1),0),\label{ln}
\end{eqnarray}
where $m\geq 1$ is a positive integer. More precisely, each remaining vertex on the boundary row is followed by $(m-1)$ removed vertices in the boundary row; the left most vertex and the rightmost vertex on the boundary row are remaining vertices. In Example 1.3.7 of \cite{IGM15}, the Schur function of such a signature is computed explicitly as follows
\begin{eqnarray}
s_{\lambda(N)}(x_1,\ldots,x_N)=\prod_{1\leq i<j\leq N}\frac{x_i^m-x_j^m}{x_i-x_j}.\label{sm}
\end{eqnarray}
\begin{definition}\label{df33}Let
\begin{eqnarray}
X=(x_1,x_2,\ldots,x_N)\in\RR^N\label{xn}
\end{eqnarray}
Let $\rho_N$ be a probability measure on $\GT_N$. The the \textbf{Schur generating function} with respect to $\rho_N$, $X$ is given by
\begin{eqnarray*}
\mathcal{S}_{\rho_N,X}(u_1,\ldots,u_N)=\sum_{\lambda\in \GT_N}\rho_N(\lambda)\frac{s_{\lambda}(u_1,\ldots,u_N)}{s_{\lambda}(x_1,\ldots,x_N)}
\end{eqnarray*}
\end{definition}
For a positive integer $s$, let $\ol{s}=s\mod n$. We make the following assumption on edge weights.
\begin{assumption}\label{pw}Assume that the edge weights $x_i$ ($1\leq i\leq N$), $y_j$ ($j\in I_2$) changes periodically with period $n$; i.e.
\begin{eqnarray}
x_{\ol{i}}=x_i\label{px};\\
y_{\ol{j}}=y_j\label{py}
\end{eqnarray}
for $1\leq i\leq n$.
\end{assumption}
\begin{lemma}\label{lmm212}Let $x_i>0 (1\leq i\leq N)$, $y_j (j\in I_2)$ be edge weights of a contracting square- hexagon lattice $\mathcal{R}(\Omega,\check{c})$ satisfying Assumptions \ref{apew} and \ref{pw}.
Let
\begin{eqnarray*}
X^{(N-t)}&=&(x_{\ol{t+1}},\ldots,x_{\ol{N}}),\\
Y^{(t)}&=&(x_{\ol{1}},\ldots,x_{\ol{t}})
\end{eqnarray*}
for each integer $t$ satisfying $0\leq t\leq N-1$, where $x_{i}>0\ (1\leq i\leq n)$ are weights of NE-SW edges joining the $(2i)$th row to the $(2i+1)$th row of the contracting square-hexagon lattice, see Figure \ref{fig:SH}. Let $\lambda(N)$ be the partition corresponding to the configuration on the boundary row; and let $\rho^k$ be the probability measure on $\GT_{N-t}^+$ which is the distribution of partitions corresponding to the dimer configuration on the $k$th row of vertices of $\mathcal{R}(\Omega,\check{c})$, counting from the bottom. Then
we have
\begin{eqnarray*}
\mathcal{S}_{\rho^k,X^{(N-t)}}(u_1,\ldots,u_{N-t})&=&\frac{s_{\lambda(N)}\left(u_1,\ldots,u_{N-t},Y^{(t)}\right)}{s_{\lambda(N)}(X^{(N)})}\prod_{i\in\{1,\ldots,t\}\cap I_2}\prod_{j=1}^{N-t}\left(\frac{1+y_{\ol{i}}u_j}{1+y_{\ol{i}}x_{\ol{t+j}}}\right),
\end{eqnarray*}
if $k=2t+1$, for $t=0,1,\ldots,N-1$
Moreover,
\begin{eqnarray*}
\mathcal{S}_{\rho_{N-t},X^{(N-t)}}(u_1,\ldots,u_{N-t})&=&\frac{s_{\lambda(N)}\left(u_1,\ldots,u_{N-t},Y^{(t)}\right)}{s_{\lambda(N)}(X^{(N)})}\prod_{i\in\{1,\ldots,t+1\}\cap I_2}\prod_{j=1}^{N-t}\left(\frac{1+y_{\ol{i}}u_j}{1+y_{\ol{i}}x_{\ol{t+j}}}\right),
\end{eqnarray*}
for $k=2t+2,\ t=0,1,\ldots,N-1$.
\end{lemma}
\begin{proof}See Lemma 3.17 of \cite{BL17}.
\end{proof}
\begin{lemma}\label{l33}Let $\kappa\in(0,1)$. Let $\mathcal{R}(\Omega,\check{c})$ be a contracting square-hexagon lattice.
Let $\{\lambda(\lfloor1-\kappa)N\rfloor)\}_{N\in \NN}$ be a sequence of partitions corresponding to dimer configurations on the $[2(N-\lfloor (1-\kappa )N\rfloor)+1]$th row of $\mathcal{R}(\Omega,\check{c})$, counting from the bottom. Let $X$ be an $N$-tuple of integers given by (\ref{xn}), which are also edge weights of $\mathcal{R}(\Omega,\check{c})$ satisfying (\ref{px}). Let $\rho_{\lfloor (1-\kappa )N\rfloor}:=\rho^{2(N-\lfloor (1-\kappa )N)\rfloor)+1}$ be a probability measure on $\GT_{\lfloor (1-\kappa )N\rfloor}^+$. Note that $\rho_{N}:=\delta_{\lambda(N)}$ is the distribution of partitions corresponding to the dimer configurations on the bottom row, in which $\lambda(N)$ has probability 1 to occur, while any other configuration has probability 0 to occur. Let $\mathcal{S}_{\rho_{\lfloor(1-\kappa)N\rfloor},X}(u_1,\ldots,u_{\lfloor(1-\kappa)N\rfloor})$ be the Schur generating function corresponding to $\rho_{\lfloor(1-\kappa)N\rfloor}$ and $X$.
Then we have
\begin{enumerate}
\item Assume $1\leq i\leq n$, then
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\left.\frac{1}{(1-\kappa)N}\frac{\partial\log \mathcal{S}_{\rho_{(1-\kappa)N},X}(u_1,\ldots,u_{\lfloor(1-\kappa)N\rfloor})}{\partial u_i}\right|_{(u_1,\ldots,u_{\lfloor(1-\kappa)N\rfloor})=(x_{1+N-\lfloor(1-\kappa) N\rfloor},\ldots,x_{N})}\\
&=& H_i(X,Y,\kappa);
\end{eqnarray*}
where
\begin{eqnarray*}
H_i(X,Y,\kappa)&=&\frac{1}{(1-\kappa)n}\left\{\left[\sum_{j\in\{1,2,\ldots,n\},j\neq i}\left(\frac{mx_{i}^{m-1}}{x_{i}^m-x_j^m}-\frac{1}{x_{i}-x_j}\right)\right]+\frac{m-1}{2x_{i}}\right\}\\&&+\frac{\kappa}{(1-\kappa)n}\sum_{j\in\{1,2,\ldots,n\}\cap I_2}\frac{y_j}{1+y_j x_{i}}
\end{eqnarray*}
\item Assume $1\leq i,j\leq \lfloor(1-\kappa)N\rfloor$ and $i\neq j$. Then
\begin{eqnarray*}
\left.\lim_{N\rightarrow\infty}\frac{\partial^2\log \mathcal{S}_{\lfloor\rho_{(1-\kappa)N}\rfloor,X}(u_1,\ldots,u_{\lfloor(1-\kappa)N\rfloor})}{\partial u_i\partial u_j}\right|_{(u_1,\ldots,u_{\lfloor(1-\kappa)N\rfloor})=(x_{1+N-\lfloor(1-\kappa) N\rfloor},\ldots,x_{N})}=G(x_i,x_j);
\end{eqnarray*}
where
\begin{eqnarray*}
G(x_i,x_j)=
&=&\left\{\begin{array}{cc}\frac{m^2x_i^{m-1}x_j^{m-1}}{(x_i^m-x_j^m)^2}-\frac{1}{(x_i-x_j)^2}&\mathrm{if}\ x_i\neq x_j\\0&\mathrm{otherwise}\end{array}\right.
\end{eqnarray*}
\item Assume $i,j,k\in\{1,2,\ldots,n\}$ are three distinct integers, then
\begin{eqnarray*}
\left.\lim_{N\rightarrow\infty}\frac{\partial^3\log \mathcal{S}_{\rho_{\lfloor(1-\kappa)N\rfloor},X}(u_1,\ldots,u_{\lfloor(1-\kappa)N\rfloor})}{\partial u_i\partial u_j\partial u_k}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}=0.
\end{eqnarray*}
\end{enumerate}
\end{lemma}
\begin{proof}Applying Lemma \ref{lmm212}, (\ref{sm}) and Definition \ref{df33}, and by explicit computations.
\end{proof}
\begin{remark}\label{rm25}Lemma \ref{l33} still holds if we define $\rho_{\lfloor (1-\kappa )N\rfloor}:=\rho^{2(N-\lfloor (1-\kappa )N)\rfloor)+2}$ and $\{\lambda(\lfloor1-\kappa)N\rfloor)\}_{N\in \NN}$ be a sequence of signatures corresponding to dimer configurations on the $[2(N-\lfloor (1-\kappa )N)\rfloor)+2]$th row of $\mathcal{R}(\Omega,\check{c})$.
\end{remark}
\begin{proposition}\label{plm}Let $\mathcal{R}(\Omega(N),\check{c})$ be a contracting square hexagon lattice with the configuration at the bottom boundary given by
\begin{eqnarray*}
\Omega(N)=(1,m+1,2m+1,\ldots,(N-1)m+1)
\end{eqnarray*}
Assume also that the edge weights are assigned as in Assumption \ref{apew} (see Figure \ref{fig:SH} for an example) and periodically with period $n$; i.e. the edge weights satisfy (\ref{px}) and (\ref{py}).
Let
\begin{eqnarray*}
F_{\kappa,m}(z)=\frac{\kappa z}{n(1-\kappa)}\sum_{i\in\{1,2,\ldots,n\}\cap I_2}\frac{y_i}{1+y_iz}+\sum_{j=1}^{n}\frac{z}{n(z-x_{j})}+\frac{z}{n(1-\kappa)}\sum_{j=1}^{n}\left(\frac{mz^{m-1}}{z^m-x_j^m}-\frac{1}{z-x_j}\right)
\end{eqnarray*}
Let $\rho_N^k$ be the measure on the configurations of the $k$th row, and let $\kappa\in (0,1)$, such that $k=[2\kappa N]$, Then the corresponding counting measure $m(\rho_N^k)$ converges to $\mathbf{m}^{\kappa}$ in probability as $N\rightarrow\infty$, and the moments of $\mathbf{m}^{\kappa}$ is given by
\begin{eqnarray}
\int_{\RR}y^{p}\textbf{m}^{\kappa}(dy)=\sum_{i=1}^{n}\frac{1}{2(p+1)\pi \mathbf{i}}\oint_{x_{t+i}}\frac{dz}{z}\left[ F_{\kappa,m}(z)\right]^{p+1}\label{mtl}
\end{eqnarray}
\end{proposition}
\begin{proof}See Section 8 of \cite{BL17}.
\end{proof}
We can compute the Stieltjes transform of the limit measure $\mathbf{m}^{\kappa}$ when $x$ is in neighborhood of infinity by
\begin{eqnarray*}
\sum_{j=0}^{\infty}\frac{\int_{\RR}y^{j}\textbf{m}^{\kappa}(dy)}{x^{j+1}}=-\sum_{i=1}^{n}\frac{1}{2\pi\mathbf{i}}\oint_{x_{t+i}}\frac{dz}{z}\mathrm{log}\left(1-\frac{F_{\kappa,m}(z)}{x}\right)
\end{eqnarray*}
Integrating by parts we have
\begin{eqnarray*}
\sum_{j=0}^{\infty}\frac{\int_{\RR}y^{j}\textbf{m}^{\kappa}(dy)}{x^{j+1}}=\sum_{i=1}^{n}\frac{1}{2\pi\mathbf{i}}\oint_{x_{t+i}}\log(z)\frac{\partial_z\left(1-\frac{F_{\kappa,m}(z)}{x}\right)}{\left(1-\frac{F_{\kappa,m}(z)}{x}\right)}dz
\end{eqnarray*}
The integrand has poles at roots of
\begin{eqnarray}
F_{\kappa,m}(z)=x.\label{fmz}
\end{eqnarray}
\begin{definition}
\label{df41}Let $\mathcal{R}$ be the rescaled square-hexagon lattice, i.e. $\mathcal{R}=\frac{1}{N}\mathcal{R}(\Omega,\check{c})$, with coordinates $(\chi,\kappa)$.
Let $\mathcal{L}$ be the set of $(\chi,\kappa)$ inside $\mathcal{R}$ such that
the density $d\mathbf{m}^{\kappa}\left(\frac{\chi}{1-\kappa}\right)$ is not
equal to 0 or 1. Then $\mathcal{L}$ is called the \emph{liquid region}. Its boundary
$\partial \mathcal{L}$ is called the \emph{frozen boundary}.
\end{definition}
\section{Central Limit Theorem for Uniform Boundary Conditions} \label{unc}
In this section, we construct certain statistics from the (random) dimer configuration on a contracting square hexagon lattice with uniform boundary conditions, and show that the converge in distribution to Gaussian random variables in the scaling limit. The main theorem proved in this section is Theorem \ref{gff1}.
Let $X$ be given by (\ref{xn}), and let $V_N(X)$ be the Vandermonde determinant, i.e.
\begin{eqnarray*}
V_N(X)=\prod_{1\leq i<j\leq N}(x_j-x_i)
\end{eqnarray*}
\begin{proposition}\label{pn41}Let $\rho_N$ be a probability measure on $\GT_N$, and let $\lambda\in\GT_N$. Let
\begin{eqnarray*}
U=(u_1,u_2,\ldots,u_N)\in \CC^N.
\end{eqnarray*}
Then
\begin{eqnarray}
\mathbf{E}\sum_{i=1}^{N}(\lambda_i+N-i)^k:&=&\sum_{\lambda\in\GT_N}\rho_N(\lambda)\sum_{i=1}^N(\lambda_i+N-i)^k\notag\\
&=&\left.\frac{1}{V_N(U)}\sum_{i=1}^{N}(u_i\partial_i)^k V_N(U) \mathcal{S}_{\rho_N,X}(U)\right|_{U=X};\label{e1}
\end{eqnarray}
and
\begin{eqnarray}
&&\mathbf{E}\left(\sum_{i=1}^{N}(\lambda_i+N-i)^k\sum_{j=1}^{N}(\lambda_j+N-j)^l\right)\notag\\
&=&\left.\frac{1}{V_N(U)}\sum_{i=1}^N(u_i\partial_i)^k\sum_{j=1}^{N}(u_j\partial_j)^l V_N(U) \mathcal{S}_{\rho_N,X}(U)\right|_{U=X}\label{e2}
\end{eqnarray}
\end{proposition}
\begin{proof}Let $\lambda\in \GT_N$, and let $s_{\lambda}$ be the Schur function with respect to $\lambda$. By Proposition 4.3 of \cite{bg}, we have
\begin{eqnarray}
\frac{1}{V_N(U)}\sum_{i=1}^{N}(u_i\partial_i)^k V_N(U) s_{\lambda}(U)=\sum_{i=1}^{N}(\lambda_i+N-1)^k s_{\lambda}(U).\label{ds}
\end{eqnarray}
Dividing by $s_{\lambda}(X)$ to both sides of (\ref{ds}), then taking expectations for $\lambda$ with respect to the distribution $\rho_N$; then evaluate at $U=X$, we obtain (\ref{e1}). The expression (\ref{e2}) can be obtained similarly by performing the above process twice.
\end{proof}
Let $f(x_1,\ldots,x_r)$ be a function of $r$ variables. Define
\begin{eqnarray*}
\mathrm{Sym}_{x_1,\ldots,x_r}f(x_1,\ldots,x_r)=\frac{1}{r!}\sum_{\sigma\in \Si_r} f(x_{\sigma(1)},x_{\sigma(2)},\ldots,x_{\sigma(r)}),
\end{eqnarray*}
where $\Si_r$ is the symmetric group of $r$ elements.
For an integer $l>0$, and $t\in(0,1]$, let
\begin{eqnarray}
U_t&=&(u_1,u_2,\ldots,u_{\lfloor tN\rfloor});\label{ut}\\
X_{t}&=&(x_{N-\lfloor tN\rfloor+1},\ldots, x_N).\label{xt}
\end{eqnarray}
and
\begin{eqnarray}
\mathcal{F}_{(l,t)}(U_t)=\frac{1}{\mathcal{S}_{\rho_{\lfloor tN\rfloor},X_t}(U_t)V_{\lfloor tN\rfloor}(U_t)}\sum_{i=1}^{\lfloor tN\rfloor}(u_i\partial_i)^l V_{\lfloor tN\rfloor}(U_t)\mathcal{S}_{\rho_{\lfloor tN\rfloor},X_t}(U_t);\label{flu}
\end{eqnarray}
where $\rho_{\lfloor tN\rfloor}$ is a probability measure on $\GT_{\lfloor tN\rfloor}$.
In order to analyze the asymptotics, we first introduce the following technical lemma.
\begin{lemma}\label{l552}Let $f(z)$ be a complex analytic function in a neighborhood of $1$ and let $r$ be a positive integer. Then
\begin{eqnarray*}
\mathrm{Sym}_{z_1,\ldots,z_{r+1}}\left.\left(\frac{f(z_1)}{(z_1-z_2)\ldots(z_1-z_{r+1})}\right)\right|_{(z_1,\ldots,z_{r+1})=(1,\ldots,1)}=\left.\frac{1}{(r+1)!}\frac{\partial^r f(z)}{\partial z^r}\right|_{z=1}.
\end{eqnarray*}
\end{lemma}
\begin{proof}See Lemma 5.5 of \cite{bg}.
\end{proof}
\begin{proposition}\label{p36}Assume the assumption of Lemma \ref{l33} holds. We use the notation $\partial_i$ to denote $\frac{\partial}{\partial u_i}$. Then we have
\begin{enumerate}
\item the functions $\mathcal{F}_{(l,t)}(U_t)|_{U_t=X_t}$ have $N$-degree at most $l+1$;
\item for $1\leq i\leq N$, the functions $\partial_i\mathcal{F}_{(l,t)}(U_t)|_{U_t=X_t}$ have $N$-degree at most $l$; moreover
\begin{eqnarray*}
&&\partial_i\mathcal{F}_{(l,t)}(U_t)|_{U_t=X_t}=\partial_i\left[\sum_{r=0}^{l}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!\right.\\
&&\times\sum_{\{a_1,\ldots,a_{r+1}\}\subset\{1,2,\ldots,\lfloor{tN}\rfloor\}}\left.\left.\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left(\frac{u_{a_1}^l(\partial_{a_1}[\log S_{\rho_{\lfloor tN\rfloor},X_t}])^{l-r}}{(u_{a_1}-u_{a_2})\ldots(u_{a_1}-u_{a_{r+1}})}\right)\right]\right|_{U_t=X_t}+T_{(l,t)}(U_t)|_{U_t=X_t}
\end{eqnarray*}
where $T_{(l,t)}(X_t)$ has $N$-degree less than $l$.
\item for any $1\leq i, j\leq N$ and $i\neq j$, the functions $\partial_i\partial_j \mathcal{F}_{(l,t)}(U_t)|_{U_t=X_t}$ have $N$-degree at most $l-1$.
\end{enumerate}
\end{proposition}
\begin{proof}When $t=1$ and $X_1=(1,\ldots,1)$, the proposition is proved in Lemma 5.5 of \cite{bg16}. Consider a general $S_{\rho_{\lfloor tN\rfloor},X_t}$ with $X_t$ given by (\ref{xt}). Since $S_{\rho_{\lfloor tN\rfloor},X_t}(X_t)=1$, the function $\log S_{\rho_{\lfloor tN\rfloor},X_t}$ is well defined in a neighborhood of $X_t$. Note that
\begin{eqnarray*}
\frac{\partial_i S_{\rho_{\lfloor tN\rfloor},X_t}}{S_{\rho_{\lfloor tN\rfloor},X_t}}=\partial_i(\log S_{\rho_{\lfloor tN\rfloor},X_t}).
\end{eqnarray*}
This way we can write $\mathcal{F}_{(l,t)}(U_t)$ as a large sum of factors of the form
\begin{eqnarray*}
\frac{c_0u_i^{l-s_0}(\partial_i^{s_1}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_1}\ldots (\partial_i^{s_t}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_t}}{(u_i-u_{a_1})\ldots (u_i-u_{a_r})}
\end{eqnarray*}
where $i,\ a_1,\ \ldots,\ a_r$ are distinct indices, $s_j,d_j\in \NN\cup\{0\}$ for $j=1,\ldots t$, and
\begin{eqnarray}
s_1<s_2<\ldots<s_t;\notag\\
r+s_0+\sum_{j=1}^t s_jd_j=l.\label{fsc}
\end{eqnarray}
Moreover, $c_0$ depends on $r$, $s_j$, $d_j$, but is independent of $N$, $a_1,\ldots,a_r$. By symmetry we can write
\begin{eqnarray}
&&\mathcal{F}_{(l,t)}(U_t)=\sum_{r,\{s_j\},\{d_j\}}(r+1)!\times\sum_{\{a_1,\ldots,a_{r+1}\}\subset \{1,2,\ldots,\lfloor tN\rfloor\}}\label{exf}\\
&& \mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left(\frac{c_0u_{a_1}^{l-s_0}(\partial_{a_1}^{s_1}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_1}\ldots (\partial_{a_1}^{s_t}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_t}}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{r+1}})}\right),\notag
\end{eqnarray}
where the first sum are over $r,\{s_j\},\{d_j\}$ satisfying (\ref{fsc}), and $c_0$ depends on $r,\{s_j\},\{d_j\}$.
By Lemma \ref{l33}, for each $1\leq w\leq t$ the degree of $N$ in each factor $(\partial_{a_1}^{s_w}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_w}$ is at most $d_w$. For each given choice of $\{a_1,\ldots,a_{r+1}\}$, we define an equivalence relation on the set $\{a_1,\ldots,a_{r+1}\}$: for $1\leq i,j\leq r+1$, we say $a_i$ and $a_j$ are equivalent if and only if $[a_i\mod n]=[a_j\mod n]$. Let $A_1,\ldots, A_{w}$ be all the distinct equivalence classes under this equivalence relation, where $w$ is a positive integer satisfying $w\leq r+1$. For $1\leq i\leq w$, let $C_i=\{a_1,\ldots,a_{r+1}\}\setminus A_i$
For $1\leq i\leq r+1$, let $a_1,\ldots, \hat{a}_i,\ldots,a_{r+1}$ be $r$ distinct integers obtained from $a_1,\ldots, a_{r+1}$ by removing $a_i$. Then
\begin{eqnarray}
&&\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left(\frac{c_0u_{a_1}^{l-s_0}(\partial_{a_1}^{s_1}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_1}\ldots (\partial_{a_1}^{s_t}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_t}}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{r+1}})}\right)\label{sm1}\\
&=&\frac{1}{(r+1)!}\sum_{i=1}^{w}{{r}\choose{|C_i|}}|A_i|!|C_i|!\mathrm{Sym}_{A_i}\left[\mathrm{Sym}_{C_i}\right.\notag\\&&\left(\frac{c_0u_{a_i}^{l-s_0}(\partial_{a_i}^{s_1}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_1}\ldots (\partial_{a_i}^{s_t}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_t}}{\prod_{j\in C_i}(u_{a_i}-u_{a_j})}\right)\notag\\
&&\times\left.\left(\frac{1}{\prod_{a_j\in A_i}(u_{a_i}-u_{a_j})}\right)\right]\notag\\
&=&\sum_{i=1}^{w}\frac{|A_i|}{r+1}\mathrm{Sym}_{A_i}\left[\left(\frac{c_0u_{a_i}^{l-s_0}(\partial_{a_i}^{s_1}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_1}\ldots (\partial_{a_i}^{s_t}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_t}}{\prod_{j\in C_i}(u_{a_i}-u_{a_j})}\right)\right.\notag\\
&&\times\left.\left(\frac{1}{\prod_{a_j\in A_i}(u_{a_i}-u_{a_j})}\right)\right]\notag
\end{eqnarray}
By Lemma \ref{l33} and (\ref{fsc}), the degree of $N$ in
\begin{eqnarray*}
\left(\frac{c_0u_{a_i}^{l-s_0}(\partial_{a_i}^{s_1}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_1}\ldots (\partial_{a_i}^{s_t}[\log S_{\rho_{\lfloor tN\rfloor}, X_t}])^{d_t}}{\prod_{j\in C_i}(u_{a_i}-u_{a_j})}\right)
\end{eqnarray*}
is at most $l-r$. By Lemma \ref{l552}, the degree of $N$ in (\ref{sm1}) is at most $l-r$. Summing over all the choices $\{a_1,\ldots,a_{r+1}\}\subset\{1,2,\ldots,\lfloor tN\rfloor\}$ (there are $O(N^{r+1})$ such choices), we obtain that the degree of $N$ in $\mathcal{F}_{(l,t)}(U_t)$ is at most $l+1$; then Part (1) of the proposition follows.
\end{proof}
For positive integers $l_1,l_2$, we define
\begin{eqnarray*}
\mathcal{G}_{l_1,l_2,t}(U_t)&=&l_1\sum_{r=0}^{l_1-1}\left(\begin{array}{c}l_1-1\\r\end{array}\right)\sum_{\{a_1,\ldots,a_{r+1}\}\subset\{1,2,\ldots,\lfloor tN\rfloor\}}(r+1)!\\&&\times\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\frac{u_{a_1}^{l_1}\partial_{a_1}[\mathcal{F}_{(l_2,t)}](\partial_{a_1}[\log S_{\rho_{\lfloor tN\rfloor},X_t}])^{l_1-1-r}}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{r+1}})}
\end{eqnarray*}
\begin{lemma}Assume the assumption of Lemma \ref{l33} holds.
Let $l_1,l_2$ be arbitrary positive integers, and $t\in(0,1]$, then
\begin{eqnarray}
&&\frac{1}{V_{\lfloor tN\rfloor}S_{\rho_{\lfloor tN\rfloor},X_t}}\sum_{i_1=1}^{\lfloor tN\rfloor}(u_{i_1}\partial_{i_1})^{l_1}\sum_{i_2=1}^{\lfloor tN\rfloor} (u_{i_2}\partial_{i_2})^{l_2}[V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor},X_t}]\notag\\
&=&\mathcal{F}_{(l_1,t)}(U_t)\mathcal{F}_{(l_2,t)}(U_t)+\mathcal{G}_{(l_1,l_2,t)}(U_t)+T(U_t)\label{l5i}
\end{eqnarray}
where $\mathcal{G}_{(l_1,l_2,t)}(U_t)|_{U_t=X_t}$ has $N$-degree at most $l_1+l_2$ and $T(U_t)|_{U_t=X_t}$ has $N$ degree less than $l_1+l_2$. Moreover, for any index $i$ the function $\partial_i\mathcal{G}_{(l_1,l_2,t)}(U_t)|_{U_t=X_t}$ has $N$-degree less than $l_1+l_2$.
\end{lemma}
\begin{proof}The proof follows from similar arguments as in the proof of Lemma 5.7 in \cite{bg16}, in which the case $X=1^N$ and $t=1$ is proved. We sketch the idea here. Note that the left hand side of (\ref{l5i}) is exactly
\begin{eqnarray*}
\frac{1}{V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor},X_t}}\sum_{i_1=1}^{N}(u_{i_1}\partial_{i_1})^{l_1}[V_{\lfloor tN\rfloor} \mathcal{S}_{\rho_{\lfloor tN\rfloor},X_t} \mathcal{F}_{(l_2)}(U_t)]
\end{eqnarray*}
It can be rewritten as the sum of terms of the form
\begin{eqnarray*}
\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\frac{c_0 u_{a_1}^{l_1-s_0}\partial_{a_1}^{s_1}[\mathcal{F}_{(l_2,t)}](\partial_{a_1}^{s_2}[\log S_{\rho_{\lfloor tN\rfloor},X_t}])^{d_2}\ldots(\partial_{a_1}^{s_p}[\log S_{\rho_{\lfloor tN\rfloor},X_t}])^{d_p}}{(u_{a_1}-u_{a_2})(u_{a_1}-u_{a_3})\ldots(u_{a_1}-u_{a_{r+1}})},
\end{eqnarray*}
where $r,s_0,s_1,\ldots,s_p,d_2,\ldots,d_p$ are nonnegative integers and
\begin{eqnarray*}
&&s_2<s_3<\ldots< s_p;\\
&&s_0+s_1+s_2d_2+\ldots+s_pd_p+r=l_1.
\end{eqnarray*}
Then $\mathcal{F}_{(l_1,t)}(U_t)\mathcal{F}_{(l_2,t)}(U_t)$ comes from the terms with $s_1=0$; $\mathcal{G}_{(l_1,l_2,t)}(U_t)$ comes from the terms with $s_0=0$, $s_1=1$, $s_2=1$, $d_2=l_1-1-r$. The $N$-degrees of these terms can be obtained by applying Lemma \ref{l33}.
\end{proof}
Let $s$ be a positive integer. For a subset $\{j_1,\ldots,j_p\}\subset \{1,2,\ldots s\}$, let $\mathcal{P}^s_{j_1,\ldots,j_p}$ be the set of all pairings of the set $\{1,2,\ldots,s\}\setminus\{j_1,\ldots,j_p\}$. The set $\mathcal{P}^s_{j_1,\ldots,j_p}$ is non-empty only when $s-p$ is even. For a pairing $P$ let $\prod_{(a,b)\in P}$ denote the product over all pairs $(a,b)$ from this pairing.
\begin{proposition}Assume that the assumption of Lemma \ref{l33} holds.
Let $s,l_1,\ldots,l_s$ be arbitrary positive integers, and let $t\in (0,1]$. Then
\begin{eqnarray*}
&&\frac{1}{V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor},X_t}}\sum_{i_1=1}^{\lfloor tN\rfloor}(u_{i_1}\partial_{i_1})^{l_1}\sum_{i_2=1}^{\lfloor tN\rfloor}(u_{i_2}\partial_{i_2})^{l_2}\ldots \sum_{i_s=1}^{\lfloor tN\rfloor}(u_{i_s}\partial_{i_s})^{l_s}[V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor},X_t}]\\
&=&\sum_{p=0}^{s}\sum_{\{j_1,\ldots,j_p\}\subset\{1,2,\ldots,s\}}\mathcal{F}_{(l_{j_1},t)}(U_t)\ldots \mathcal{F}_{(l_{j_p},t)}(U_t)\left(\sum_{P\in \mathcal{P}^s_{j_1,\ldots,j_p}}\prod_{(a,b)\in P}\mathcal{G}_{(l_a,l_b,t)}(U_t)+T_{j_1,\ldots,j_p}^{1;s}(U_t)\right),
\end{eqnarray*}
where $T_{j_1,\ldots,j_p}^{1;s}(U_t)|_{U_t=X_t}$ has $N$-degree less than $\sum_{i=1}^{s}l_i-\sum_{i=1}^p l_{j_i}$.
\end{proposition}
\begin{proof}The proposition can be proved by induction on $s$, similar to the proof of Proposition 5.10 of \cite{bg}, where the case when $X=1^N$ is proved.
\end{proof}
Let $l$ be a positive integer and $t\in(0,1]$. Let
\begin{eqnarray*}
E_{l,t}:=\mathcal{F}_{(l,t)}(X_t)=\frac{1}{V_{\lfloor tN\rfloor}S_{\rho_{\lfloor tN\rfloor}, X_t}}\sum_{i=1}^{\lfloor tN\rfloor}(u_i\partial_i)^l V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor}, X_t}(U_t)|_{U_t=X_t}
\end{eqnarray*}
\begin{lemma}\label{lc}Assume the assumption of Lemma \ref{l33} holds. Let $s, l_1,\ldots,l_s$ be arbitrary positive integers, and let $t\in(0,1]$. Then
\begin{eqnarray*}
&&\frac{1}{V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor}, X_t}}\left(\sum_{i_1=1}^{\lfloor t N\rfloor}(u_{i_1}\partial_{i_1})^{l_1}-E_{l_1,t}\right)\left(\sum_{i_2=1}^{\lfloor tN\rfloor}(u_{i_2}\partial_{i_2})^{l_2}-E_{l_2,t}\right)\\
&&\left.\times\cdots\left(\sum_{i_2=1}^{\lfloor tN\rfloor}(u_{i_s}\partial_{i_s})^{l_s}-E_{l_s,t}\right)V_{\lfloor tN\rfloor} S_{\rho_{\lfloor tN\rfloor},X_t}\right|_{U_t=X_t}=\left.\sum_{P\in \mathcal{P}_{\emptyset}^s}\prod_{(a,b)\in P}\mathcal{G}_{(l_a,l_b,t)}(U)\right|_{U_t=X_t}+T_{\emptyset}(U_t)|_{U_t=X_t},
\end{eqnarray*}
where $T_{\emptyset}(U_t)|_{U_t=X_t}$ has $N$-degree less than $\sum_{i=1}^s l_i$.
\end{lemma}
Let $\kappa\in[0,1)$ and $\lambda\in\GT_{N-\lfloor \kappa N\rfloor}$.
\begin{eqnarray*}
p_j^{(N-\lfloor \kappa N\rfloor)}=\sum_{i=1}^{N-\lfloor\kappa N\rfloor}(\lambda_i+(N-\lfloor\kappa N\rfloor)-i)^j;\qquad \mathrm{for}\ j=1,2,\ldots.
\end{eqnarray*}
Assume the distribution of $\lambda$ is $\rho_{N-\lfloor\kappa N\rfloor}$. Let $\mathbf{E}$ be the expectation under the probability measure $\rho_{N-\lfloor \kappa N\rfloor}$, and let
\begin{eqnarray*}
\mathrm{cov}\left(p_{k}^{(N-\lfloor\kappa N\rfloor)},p_{l}^{(N-\lfloor\kappa N\rfloor)}\right)=\mathbf{E}\left(p_{k}^{(N-\lfloor \kappa N \rfloor)},p_{l}^{(N-\lfloor \kappa N \rfloor)}\right)-\mathbf{E}p_{k}^{(N-\lfloor\kappa N\rfloor)}\mathbf{E}p_{l}^{(N-\lfloor \kappa N\rfloor)}.
\end{eqnarray*}
then by Lemma \ref{lc}, we have
\begin{eqnarray}
\lim_{N\rightarrow\infty}\frac{\mathrm{cov}\left(p_{k}^{(N-\lfloor \kappa N\rfloor)},p_{l}^{(N-\lfloor \kappa N\rfloor)}\right)}{N^{k+l}}=\lim_{N\rightarrow\infty}\frac{\mathcal{G}_{(k,l)}(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}{N^{k+l}}.\label{cov}
\end{eqnarray}
We have
\begin{eqnarray*}
&&\mathcal{G}_{(k,l)}(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)\\
&=&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}k-1\\q\end{array}\right)(q+1)!\\&&\left.\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\frac{u_{a_1}^k\partial_{a_1}[\mathcal{F}_{(l)}](\partial_{a_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}]^{k-1-q})}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&\approx&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}k-1\\q\end{array}\right)(q+1)!\\
&&\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left(\frac{u_{a_1}^k(\partial_{a_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}]^{k-1-q})}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right.\\
&\times&\partial_{a_1}\left[\sum_{r=0}^{l}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!\right.\\
&\times &\left.\left.\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(\partial_{b_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}])^{l-r}}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right]\right)\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}
\end{eqnarray*}
The approximate equality above contains only leading terms of $\partial_{a_1}[\mathcal{F}_{(l,1-\kappa)}]$; see Proposition \ref{p36} (2).
We first consider the case that
\begin{eqnarray*}
\{a_1,\ldots,a_{q+1}\}\cap\{b_1,\ldots,b_{r+1}\}=\emptyset.
\end{eqnarray*}
By Lemma \ref{l33}, we have
\begin{eqnarray*}
&&\partial_{a_1}\left[\sum_{r=0}^{l}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!\right.\\
&\times &\left.\left.\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(\partial_{b_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},X_{\kappa}}])^{l-r}}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right]\right)\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&=&\sum_{r=0}^{l}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!(l-r)\\
&\times &\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(\partial_{b_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},X_{\kappa}}])^{l-r-1}\partial_{a_1}\partial_{b_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},X_{\kappa}}]}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&\approx&\sum_{r=0}^{l}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!(l-r)\\
&\times &\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(H_{b_1}(X,Y,\kappa))^{l-r-1}N^{l-r-1}G(x_{a_1},x_{b_1})}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
Moreover,
\begin{eqnarray*}
&&\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left(\frac{u_{a_1}^k(\partial_{a_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}]^{k-1-q})}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right.\\
&\times&\partial_{a_1}\left[\sum_{r=0}^{l}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!\right.\\
&\times &\left.\left.\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(\partial_{b_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}])^{l-r}}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right]\right)\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}\\
&\approx&\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left(\frac{u_{a_1}^k([H_{a_1}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right.\\
&&\sum_{r=0}^{l}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!(l-r)\\
&\times &\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(H_{b_1}(X,Y,\kappa))^{l-r-1}N^{l-r-1}G(x_{a_1},x_{b_1})}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
\begin{lemma}
Let
\begin{eqnarray*}
H(z)&=&\frac{1}{n}\left[\sum_{j\in\{1,2,\ldots,n\}}\left(\frac{mz^{m-1}}{z^m-x_j^m}-\frac{1}{z-x_j}\right)\right]
+\frac{\kappa}{n}\sum_{j\in\{1,2,\ldots,n\}\cap I_2}\frac{y_j}{1+y_j z}.
\end{eqnarray*}
The contribution of the terms when $\{a_1,\ldots,a_{q+1}\}\cap\{b_1,\ldots,b_{r+1}\}=\emptyset$ to $\mathcal{G}_{(k,l,1-\kappa)}(X_{1-\kappa})$, as $N\rightarrow\infty$, is asymptotically
\begin{eqnarray}
\frac{(1-\kappa)^{k+l}N^{k+l}}{(2\pi\mathbf{i})^2}\sum_{i,j=1}^{n}\oint_{|z-x_i|=\epsilon }\oint_{|w-x_j|=\epsilon}\left(\sum_{i=1}^n\frac{z}{n(z-x_i)}+\frac{z H(z)}{1-\kappa}\right)^k\\
\times\left(\sum_{i=1}^n\frac{w}{n(w-x_i)}+\frac{w H(w)}{1-\kappa}\right)^l
G(z,w)dz dw\label{ct0}
\end{eqnarray}
where $\epsilon>0$ is sufficiently small such that for each $1\leq i\leq n$ the disk centered at $x_i$ with radius $\epsilon$ contains exactly one singularity $x_i$ of the integrand.
\end{lemma}
\begin{proof}Note that
\begin{eqnarray*}
\sum_{r=0}^{l}{{l}\choose{r}}(l-r)=l\sum_{r=0}^{l-1}{{l-1}\choose {r}}
\end{eqnarray*}
By the computations above, The contribution of the terms when $\{a_1,\ldots,a_{q+1}\}\cap\{b_1,\ldots,b_{r+1}\}=\emptyset$ to $\mathcal{G}_{(k,l,1-\kappa)}(X_{1-\kappa})$, as $N\rightarrow\infty$, is asymptotically
\begin{eqnarray*}
I:&=&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}{{k-1}\choose{q}}(q+1)!\\
&&\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left[\frac{u_{a_1}^k([H_{a_1}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right.\\
&&l\sum_{r=0}^{l-1}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset\}}{l-1\choose r}(r+1)!\\
&\times &\left.\left.\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^l(H_{b_1}(X,Y,\kappa))^{l-r-1}N^{l-r-1}G(x_{a_1},x_{b_1})}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
We consider the equivalence relation on $\{a_1,\ldots,a_{q+1}\}$ (resp.\ $\{b_1,\ldots,b_{r+1}\}$) such that for $1\leq i\leq j\leq q+1$, $a_i$ and $a_j$ (resp.\ for $1\leq i\leq j\leq r+1$, $b_i$ and $b_j$) are equivalent if and only if $(i\mod n)=(j\mod n)$. Let $A_1,\ldots, A_h$ (resp.\ $B_1,\ldots,B_g$) be all the equivalence classes in $\{a_1,\ldots,a_{q+1}\}$ (resp.\ $\{b_1,\ldots,b_{r+1}\}$) under such an equivalence relation, where $h,g$ are positive integers satisfying $1\leq h\leq q+1$, $1\leq g\leq r+1$. For $1\leq i\leq h$ and $1\leq j\leq g$, let
\begin{eqnarray*}
C_i=\{a_1,\ldots,a_{q+1}\}\setminus A_i;\qquad D_j=\{b_1,\ldots,b_{r+1}\}\setminus B_j.
\end{eqnarray*}
Then we have
\begin{eqnarray*}
I&=&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}{{k-1}\choose{q}}(q+1)!\sum_{i=1}^{g}\frac{|A_i|}{(q+1)}\\
&&\mathrm{Sym}_{A_i}\left[\frac{u_{a_i}^k([H_{a_i}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\frac{1}{\prod_{a_t\in A_i\setminus\{a_i\}}(u_{a_i}-u_{a_t})}\\
&&l\sum_{r=0}^{l-1}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset\}}{l-1\choose r}(r+1)!\sum_{j=1}^{h}\frac{|B_j|}{(r+1)}\\
&\times &\left.\left.\mathrm{Sym}_{B_j}\left(\frac{u_{b_j}^l(H_{b_j}(X,Y,\kappa))^{l-r-1}N^{l-r-1}G(x_{a_i},x_{b_j})}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\frac{1}{\prod_{b_w\in B_j\setminus\{b_j\}}(u_{b_j}-u_{b_w})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
By Lemma \ref{l552}, we have
\begin{eqnarray*}
I&=&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}{{k-1}\choose{q}}(q+1)!\sum_{i=1}^{g}\frac{|A_i|}{(q+1)}\\
&&\frac{1}{|A_i|!}\frac{\partial^{|A_i|-1}}{\partial u_{a_i}^{|A_i|-1}}\left[\frac{u_{a_i}^k([H_{a_i}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\\
&&l\sum_{r=0}^{l-1}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset\}}{l-1\choose r}(r+1)!\sum_{j=1}^{h}\frac{|B_j|}{(r+1)}\\
&\times &\frac{1}{|B_j|!}\frac{\partial^{|B_j|-1}}{\partial u_{b_j}^{|B_j|-1}}\left.\left.\left(\frac{u_{b_j}^l(H_{b_j}(X,Y,\kappa))^{l-r-1}N^{l-r-1}G(x_{a_i},x_{b_j})}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\sum_{i=1}^{g}\frac{k!}{(k-1-q)!}\\
&&\frac{1}{(|A_i|-1)!}\frac{\partial^{|A_i|-1}}{\partial u_{a_i}^{|A_i|-1}}\left[\frac{u_{a_i}^k([H_{a_i}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\\
&&\sum_{r=0}^{l-1}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset\}}\frac{l!}{(l-1-r)!}\sum_{j=1}^{h}\\
&\times &\frac{1}{(|B_j|-1)!}\frac{\partial^{|B_j|-1}}{\partial u_{b_j}^{|B_j|-1}}\left.\left.\left(\frac{u_{b_j}^l(H_{b_j}(X,Y,\kappa))^{l-r-1}N^{l-r-1}G(x_{a_i},x_{b_j})}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}
\end{eqnarray*}
Using the residue theorem, we obtain
\begin{eqnarray*}
I&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\sum_{i=1}^{g}\frac{k!}{(k-1-q)!}\\
&&\mathrm{Res}_{z=u_{a_i}}\left[\frac{z^k([H(z)]^{k-1-q} N^{k-1-q})}{(z-u_{a_i})^{|A_i|}\prod_{a_s\in C_i}(z-u_{a_s})}\right.\\
&&\sum_{r=0}^{l-1}\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset\}}\frac{l!}{(l-1-r)!}\sum_{j=1}^{h}\\
&\times &\mathrm{Res}_{w=u_{b_j}}\left.\left.\left(\frac{w^l(H(w))^{l-r-1}N^{l-r-1}G(z,w)}{(w-u_{b_j})^{|B_j|}\prod_{b_v\in D_j}(w-u_{b_v})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&\approx&N^{k+l}\sum_{s=1}^{n}
\mathrm{Res}_{z=x_s}\left[\left(\frac{1}{N}\sum_{i=\lfloor \kappa N\rfloor+1}^{N}\frac{z}{z-x_i}+zH(z)\right)^k\right.\\
&&\times\left.\left[\sum_{j=1}^{n}\mathrm{Res}_{w=x_j}\left(\left(\frac{1}{N}\sum_{i=\lfloor \kappa N\rfloor+1}^{N}\frac{w}{w-x_i}+wH(w)\right)^lG(z,w)\right)\right]\right]\\
&\approx&\frac{N^{k+l}(1-\kappa)^{k+l}}{(2\pi \mathbf{i})^2}\sum_{i=1}^{n}\sum_{j=1}^{n}
\oint_{\left|z-x_i\right|=\epsilon}\left(\frac{1}{n}\sum_{p=1}^{n}\frac{z}{z-x_p}+\frac{zH(z)}{1-\kappa}\right)\\
&\times &\oint_{\left|w-x_j\right|=\epsilon}\left(\frac{1}{n}\sum_{q=1}^{n}\frac{w}{w-x_q}+\frac{wH(w)}{1-\kappa}\right)G(z,w)dwdz
\end{eqnarray*}
Then the lemma follows.
\end{proof}
Now we consider the case when
\begin{eqnarray}
|\{a_1,\ldots,a_{q+1}\}\cap\{b_1,\ldots,b_{r+1}\}|=1.\label{io}
\end{eqnarray}
Without loss of generality we suppose that $a_1=b_1$, and all the other indices are distinct. Then we have
\begin{eqnarray*}
&&\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left(\frac{u_{a_1}^k(\partial_{a_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}]^{k-1-q})}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right.\\
&\times&\partial_{a_1}\left[\sum_{r=0}^{l}\sum_{\{b_2,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!\right.\\
&\times &\left.\left.\left.\mathrm{Sym}_{a_1,b_2,\ldots,b_{r+1}}\frac{u_{a_1}^l(\partial_{a_1}[\log S_{\rho_{N-\lfloor\kappa N\rfloor},(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}])^{l-r}}{(u_{a_1}-u_{b_2})\ldots(u_{a_1}-u_{b_{r+1}})}\right]\right)\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}\\
&\approx&\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left(\frac{u_{a_1}^k N^{k-1-q} H(u_{a_1})^{k-1-q}}{(1-\kappa)^{k-1-q}(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{q+1}})}\right.\\&&\times\partial_{a_1}\left[\sum_{r=0}^{l}\sum_{\{b_2,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\left(\begin{array}{c}l\\r\end{array}\right)(r+1)!\right.\\
&&\times \left.\left.\left.\mathrm{Sym}_{a_1,b_2,\ldots,b_{r+1}}\frac{u_{a_1}^l N^{l-r}[H(u_{a_1})]^{l-r}}{(1-\kappa)^{l-r}(u_{a_1}-u_{b_2})\ldots (u_{a_1}-u_{b_{q+1}})}\right]\right)\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=(x_{\lfloor\kappa N\rfloor+1},\ldots,x_N)}
\end{eqnarray*}
The summation over indices when (\ref{io}) holds gives terms of order $N^{r+q+1}$. We can see that the contribution of the terms when (\ref{io}) holds to $\mathcal{G}_{(l,k,1-\kappa)}(X_{1-\kappa})$ as $N\rightarrow\infty$, is
\begin{eqnarray*}
\tilde{I}:&=&\tilde{I}_1+\tilde{I}_2;
\end{eqnarray*}
where
\begin{eqnarray*}
\tilde{I}_1&=&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}{{k-1}\choose{q}}(q+1)!\sum_{i=1}^{g}\frac{|A_i|}{(q+1)}\\
&&\mathrm{Sym}_{A_i}\left\{\frac{u_{a_i}^k([H_{a_i}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\frac{1}{\prod_{a_t\in A_i\setminus\{a_i\}}(u_{a_i}-u_{a_t})}\\
&&\frac{\partial}{\partial u_{a_i}}\left[\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_j=a_i\}}\sum_{j=1}^{h}\frac{|B_j|}{(r+1)}\mathbf{1}_{b_j\in B_j}{l\choose r}(r+1)!\right.\\
&\times &\left.\left.\left.\mathrm{Sym}_{B_j}\left(\frac{u_{b_j}^l(H_{b_j}(X,Y,\kappa))^{l-r}N^{l-r}}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\frac{1}{\prod_{b_w\in B_j\setminus\{b_j\}}(u_{b_j}-u_{b_w})}\right)\right]\right\}\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
\end{eqnarray*}
and
\begin{eqnarray*}
\tilde{I}_2&=&k\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}{{k-1}\choose{q}}(q+1)!\sum_{i=1}^{g}\frac{|A_i|}{(q+1)}\\
&&\mathrm{Sym}_{A_i}\left\{\frac{u_{a_i}^k([H_{a_i}(X,Y,\kappa)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\frac{1}{\prod_{a_t\in A_i\setminus\{a_i\}}(u_{a_i}-u_{a_t})}\\
&&\frac{\partial}{\partial u_{a_i}}\left[\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_s,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_s,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_s=a_i\}}\sum_{j=1}^{h}\frac{|B_j|}{(r+1)}\mathbf{1}_{b_s\notin B_j}\right.\\&&{l\choose r}(r+1)!
\left.\left.\left.\mathrm{Sym}_{B_j}\left(\frac{u_{b_j}^l(H_{b_j}(X,Y,\kappa))^{l-r}N^{l-r}}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\frac{1}{\prod_{b_w\in B_j\setminus\{b_j\}}(u_{b_j}-u_{b_w})}\right)\right]\right\}\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
\end{eqnarray*}
Here we use $\mathbf{1}$ to denote the indicator function.
By Lemma \ref{l552}, we infer
\begin{eqnarray*}
\tilde{I}_1&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\frac{k!}{(k-1-q)!}\sum_{i=1}^{g}|A_i|\\
&&\frac{1}{|A_i|!}\frac{\partial^{|A_i|-1}}{\partial u_{a_i}^{|A_i|-1}}\left[\frac{u_{a_i}^k([H(u_{a_i})]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\\
&&\frac{\partial}{\partial u_{a_i}}\left(\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_j=a_i\}}\frac{l!}{(l-r)!}\right.\\
&\times &\sum_{j=1}^{h}|B_j|\mathbf{1}_{b_j\in B_j}\left.\left.\left.\frac{1}{|B_j|!}\frac{\partial^{|B_j|-1}}{\partial u_{b_j}^{|B_j|-1}}\frac{u_{b_j}^l(H(u_{b_j}))^{l-r}N^{l-r}}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
and
\begin{eqnarray*}
\tilde{I}_2&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\frac{k!}{(k-1-q)!}\sum_{i=1}^{g}|A_i|\\
&&\frac{1}{|A_i|!}\frac{\partial^{|A_i|-1}}{\partial u_{a_i}^{|A_i|-1}}\left[\frac{u_{a_i}^k([H(u_{a_i})]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\\
&&\frac{\partial}{\partial u_{a_i}}\left(\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_s,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_s,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_s=a_i\}}\frac{l!}{(l-r)!}\right.\\
&\times &\sum_{j=1}^{h}|B_j|\mathbf{1}_{b_s\notin B_j}\left.\left.\left.\frac{1}{|B_j|!}\frac{\partial^{|B_j|-1}}{\partial u_{b_j}^{|B_j|-1}}\frac{u_{b_j}^l(H(u_{b_j}))^{l-r}N^{l-r}}{\prod_{b_v\in D_j}(u_{b_j}-u_{b_v})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
By the residue theorem, we have
\begin{eqnarray*}
\tilde{I}_1&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\frac{k!}{(k-1-q)!}\sum_{i=1}^{g}|A_i|\\
&&\frac{1}{|A_i|!}\frac{\partial^{|A_i|-1}}{\partial u_{a_i}^{|A_i|-1}}\left[\frac{u_{a_i}^k([H(u_{a_i})]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\\
&&\left(\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_j=a_i\}}\frac{l!}{(l-r)!}\sum_{j=1}^{h}|B_j|\mathbf{1}_{b_j\in B_j}\right.\\
&\times &\left.\left.\left.\frac{1}{|B_j|!}\frac{\partial^{|B_j|}}{\partial u_{a_i}^{|B_j|}}\frac{u_{a_i}^l(H(u_{a_i}))^{l-r}N^{l-r}}{\prod_{b_v\in D_j}(u_{a_i}-u_{b_v})}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\frac{k!}{(k-1-q)!}\sum_{i=1}^{g}\\
&&\mathrm{Res}_{z=u_{a_i}} \left[\frac{z^k([H(z)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(z-u_{a_s})}\frac{1}{(z-u_{a_i})^{|A_i|}}\right.\\
&&\left(\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_j,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_j=a_i\}}\sum_{j=1}^{h}|B_j|\mathbf{1}_{b_j\in B_j}\frac{l!}{(l-r)!}\right.\\
&\times &\left.\left.\left.\mathrm{Res}_{w=u_{a_i}}\left(\frac{w^l(H(w))^{l-r}N^{l-r}}{\prod_{b_v\in D_j}(w-u_{b_v})}\frac{1}{(w-u_{a_i})^{|B_j|-1}}\frac{1}{(w-z)^2}\right)\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&=&\frac{1}{(2\pi \mathbf{i})^2}\sum_{j=1}^{n}
\oint _{|z-x_j|=\epsilon}\left(\sum_{i=1}^{N-\lfloor \kappa N\rfloor}\frac{z}{z-u_i}+z N H(z)\right)^k
\oint_{|w-x_j|=\epsilon}\left(\sum_{i=1}^{N-\lfloor \kappa N\rfloor}\frac{w}{w-u_i}+w N H(w)\right)^l\\
&&\frac{1}{(w-z)^2}dwdz
\end{eqnarray*}
We also have
\begin{eqnarray*}
\tilde{I}_2&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\frac{k!}{(k-1-q)!}\sum_{i=1}^{g}|A_i|\\
&&\frac{1}{|A_i|!}\frac{\partial^{|A_i|-1}}{\partial u_{a_i}^{|A_i|-1}}\left[\frac{u_{a_i}^k([H(u_{a_i})]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(u_{a_i}-u_{a_s})}\right.\\
&&\left(\sum_{r=0}^{l}\sum_{\{b_1,\ldots,\hat{b}_s,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_1,\ldots,\hat{b}_s,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_s=a_i\}}\frac{l!}{(l-r)!}\right.\\
&\times &\sum_{j=1}^{h}|B_j|\mathbf{1}_{b_s\notin B_j}\left.\left.\left.\frac{1}{|B_j|!}\frac{\partial^{|B_j|-1}}{\partial u_{b_j}^{|B_j|-1}}\frac{u_{b_j}^l(H(u_{b_j}))^{l-r}N^{l-r}}{\prod_{b_v\in D_j\setminus\{b_s\}}(u_{b_j}-u_{b_v})}\frac{1}{(u_{b_j}-u_{b_s})^2}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}.
\end{eqnarray*}
where $b_j\in B_j$.
Using the residue theorem, we can also infer
\begin{eqnarray*}
\tilde{I}_2&=&\sum_{q=0}^{k-1}\sum_{\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor\}}\frac{k!}{(k-1-q)!}\sum_{i=1}^{g}\\
&&\mathrm{Res}_{z=u_{a_i}}\left[\frac{z^k([H(z)]^{k-1-q} N^{k-1-q})}{\prod_{a_s\in C_i}(z-u_{a_s})}\frac{1}{(z-u_{a_i})^{|A_i|}}\right.\\
&&\sum_{r=0}^{l}\sum_{\{b_2,\ldots,b_{r+1}\}\subset\{1,2,\ldots,N-\lfloor\kappa N \rfloor,\ \{b_2,\ldots,b_{r+1}\}\cap \{a_1,\ldots,a_{q+1}\}=\emptyset,\ b_1=a_i\}}\frac{l!}{(l-r)!}\sum_{j=1}^{h}\mathbf{1}_{b_1\notin B_j}\\
&\times &\left.\left.\mathrm{Res}_{w=u_{b_j}}\left(\frac{w^l(H(w))^{l-r}N^{l-r}}{\prod_{b_v\in D_j\setminus\{b_1\}}(w-u_{b_v})}\frac{1}{(w-u_{b_j})^{|B_j|}}\frac{1}{(w-z)^2}\right)\right]\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}\\
&=&\frac{1}{(2\pi\mathbf{i})^2}\sum_{j\in\{1,2,\ldots,n\}}\sum_{p\in\{1,2,\ldots,n\},p\neq j}
\oint _{|z-x_j|=\epsilon}\left(\sum_{i=1}^{N-\lfloor \kappa N\rfloor}\frac{z}{z-u_i}+z N H(z)\right)^k\\
&&\oint_{|w-x_p|=\epsilon}\left(\sum_{i=1}^{N-\lfloor \kappa N\rfloor}\frac{w}{w-u_i}+w N H(w)\right)^l
\left.\frac{1}{(w-z)^2}dwdz\right|_{(u_1,\ldots,u_{N-\lfloor \kappa N\rfloor})=X_{1-\kappa}}
\end{eqnarray*}
Therefore we have
\begin{eqnarray}
&&\lim_{N\rightarrow\infty}\frac{\tilde{I}}{N^{k+l}}\label{ct1}\\
&=&\frac{(1-\kappa)^{k+l}}{(2\pi\mathbf{i})^2}\sum_{i=1}^{n}\sum_{j=1}^{n}\oint_{|z-x_i|=\epsilon}\oint_{|w-x_j|=\epsilon}\left(\sum_{i=1}^n\frac{z}{n(z-x_i)}+ \frac{zH(z)}{1-\kappa}\right)^k\notag\\
&&\times\left(\sum_{i=1}^n\frac{w}{n(w-x_i)}+ \frac{wH(w)}{1-\kappa}\right)^l\frac{1}{(z-w)^2}dz dw\notag
\end{eqnarray}
Finally let us consider the case when
\begin{eqnarray*}
|\{a_1,\ldots,a_{q+1}\}\cap\{b_1,\ldots,b_{r+1}\}|\geq 2.
\end{eqnarray*}
By Lemma \ref{l33}, we can see that the contribution of these terms to $\mathcal{G}_{(l,k)}(X_{1-\kappa})$ has $N$-degree strictly less than $k+l$.
Therefore we have the following proposition.
\begin{proposition}Assume the assumption of Lemma \ref{l33} holds. Then
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\mathrm{cov}(p_k^{((1-\kappa)N)},p_{l}^{((1-\kappa)N)})}{N^{k+l}}\\
&=&\frac{(1-\kappa)^{k+l}}{(2\pi\mathbf{i})^2}\sum_{i=1}^{n}\sum_{j=1}^{n}\oint_{|z-x_i|=\epsilon}\oint_{|w-x_j|=\epsilon}\left(\sum_{i=1}^{n}\frac{z}{n(z-x_i)}+ \frac{zH(z)}{1-\kappa}\right)^k\\
&&\times\left(\sum_{j=1}^{n}\frac{z}{n(z-x_j)}+ \frac{zH(z)}{1-\kappa}\right)^l Q(z,w)dzdw
\end{eqnarray*}
where
\begin{eqnarray*}
Q(z,w)=G(z,w)+\frac{1}{(z-w)^2}.
\end{eqnarray*}
\end{proposition}
\begin{proof}By (\ref{cov}), $\lim_{N\rightarrow\infty}\frac{\mathrm{cov}(p_k^{((1-\kappa)N)},p_{l}^{((1-\kappa)N)})}{N^{k+l}}$ should be the sum of (\ref{ct0}) and (\ref{ct1}), divided by $N^{k+l}$. Then the proposition follows.
\end{proof}
\subsection{Multi-Level Correlations}
Define a mapping $\phi: \{1,\ldots,2N+1\}\rightarrow \{\mu^{(i)},\nu^{(j)}:i,j\in \{1,2,\ldots,N\}\}$ as follows
\begin{eqnarray*}
\phi(n)=\left\{\begin{array}{cc}\mu^{\left(\frac{n-1}{2}\right)}&\mathrm{if}\ n\ \mathrm{is\ odd}\\\nu^{\left(\frac{n}{2}\right)}&\mathrm{if}\ n\ \mathrm{is\ even}\end{array}\right.
\end{eqnarray*}
For $i\in \{1,2,\ldots,N\}$, define
\begin{eqnarray*}
C_i=(x_i,x_{i+1},\ldots,x_N)\in \RR^{N-i+1};
\end{eqnarray*}
and for $i\in I_2$, let
\begin{eqnarray*}
B_i=y_i C_i=(y_ix_i,y_ix_{i+1},\ldots,y_ix_N)\in \RR^{N-i+1}.
\end{eqnarray*}
Let $1\leq n_1\leq n_2\leq\ldots \leq n_s=2N+1$ be positive row numbers of the square-hexagon lattice, counting from the top.
For $1\leq i\leq s$, let $\rho_{n_i}$ be the induced probability measure on dimer configurations of the $n_i$th row. Then the induced probability measure on the state space
\begin{eqnarray*}
\GT_{\lfloor\frac{n_1}{2}\rfloor}\times\GT_{\lfloor\frac {n_2}{2}\rfloor}\times \ldots\times \GT_{\lfloor \frac{n_s}{2}\rfloor}
\end{eqnarray*}
by the measure proportional to the product of weights of present edges of dimer configurations on the square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$ can be expressed as follows:
\begin{eqnarray}
\mathrm{Prob}\left(\phi(n_s),\ldots,\phi(n_1)\right)=\rho_{n_s}\left((\phi(n_s)\right)\prod_{i=2}^{k}\mathrm{Prob}\left[\phi(n_{i-1})|\phi(n_i)\right]\label{prb}
\end{eqnarray}
Here $\mathrm{Prob}\left[\phi(n_{i-1})|\phi(n_i)\right]$ is the probability of $\phi(n_{i-1})$ conditional on $\phi(n_i)$. In particular, we have
\begin{eqnarray}
\mathrm{Prob}[\mu^{(t-1)}|\nu^{(t)}]&=&\mathrm{pr}_{C_{N-t+1}}(\nu^{(t)}\rightarrow\mu^{(t-1)})\label{cc1}\\
\mathrm{Prob}[\nu^{(t)}|\mu^{(t)}]&=&\begin{cases}\mathrm{st}_{B_{N-t+1}}(\mu^{(t)}\rightarrow\nu^{(t)}),\ \mathrm{If}\ N-t+1\in I_2\\ \mathbf{1}_{\nu^{(t)}=\mu^{(t)}},\ \mathrm{otherwise}\end{cases}\label{cc2}
\end{eqnarray}
where $\mathbf{1}_{\nu^{(t)}=\mu^{(t)}}$ is the indicator of $\nu^{(t)}=\mu^{(t)}$; $x_i,y_j$ are edge weights; and
\begin{eqnarray*}
\mathrm{pr}_{C_{N-t+1}}(\nu^{(t)}\rightarrow\mu^{(t-1)})=\begin{cases}x_{N-t+1}^{|\nu^{(t)}|-|\mu^{(t-1)}|}\frac{s_{\mu^{(t-1)}}(x_{N-t+2},\ldots,x_{N})}{s_{\nu^{(t)}}(x_{N-t+1},\ldots,x_N)},\ \mathrm{If}\ \mu^{(t-1)}\prec \nu^{(t)}\\0,\ \mathrm{otherwise}\end{cases}\\
\mathrm{st}_{B_{N-t+1}}(\mu^{(t)}\rightarrow\nu^{(t)})=\begin{cases}\frac{y_{N-t+1}^{|\nu(t)|-|\mu(t)|}}{\prod_{j=N-t+1}^{N}(1+y_{N-t+1}x_j)}\frac{s_{\nu^{(t)}}(x_{N-t+1},\ldots,x_N)}{s_{\mu^{(t)}}(x_{N-t+1},\ldots,x_N)},\ \mathrm{If}\ \mu^{(t)}\subset \nu^{(t)}\\0,\ \mathrm{otherwise}\end{cases};
\end{eqnarray*}
see Section 2.4 of \cite{BL17}.
\begin{definition}\label{df59}(Multi-dimensional Schur generating function) Let $1\leq N_1\leq N_2\leq\ldots\leq N_s$ be positive integers. For a probability measure $\rho$ on $\prod_{t=1}^{s}\GT_{N_t}$, we define the $s$-dimensional Schur generating function with respect to $X=(x_1,\ldots,x_N)$ by
\begin{eqnarray*}
&&\mathcal{S}_{\rho,X}(u_{1,1},\ldots,u_{N_1,1};\ldots;u_{1,s},\ldots,u_{N_s,s})\\
&=&\sum_{\lambda^1\in \GT_{N_1},\ldots,\lambda^s\in\GT_{N_s}}\rho(\lambda^1,\ldots,\lambda^s)\prod_{t=1}^{s}\frac{s_{\lambda^t}(u_{1,t},\ldots,u_{\lfloor\frac{n_t}{2} \rfloor,t})}{s_{\lambda^t}(x_{N-\lfloor\frac{n_s}{2}\rfloor+1},\ldots,x_N)}
\end{eqnarray*}
\end{definition}
The multi-dimensional Schur generating function with respect to $(1,\ldots,1)$ was defined in \cite{BG17}.
\begin{lemma}\label{l59}Let $m_1,\ldots,m_k$ be positive integers. Let $1\leq n_1\leq n_2\leq\ldots \leq n_k\leq 2N+1$ be positive row numbers of the square-hexagon lattice, counting from the top. Assume that $(\phi(n_s),\ldots,\phi(n_1))$ has distribution $\rho$ defined by (\ref{prb}).
For $1\leq s\leq k$, let $\mathcal{D}_{l}^{(s)}$ be the $l$-th order differential operator defined by
\begin{eqnarray}
\mathcal{D}_{l}^{(n_s)}=\frac{1}{V_{\lfloor \frac{n_s}{2}\rfloor}}\left(\sum_{i=1}^{\lfloor \frac{n_s}{2}\rfloor}\left(u_{i,s}\frac{\partial}{\partial u_{i,s}}\right)^l\right)V_{\lfloor \frac{n_s}{2}\rfloor}\label{dd}
\end{eqnarray}
where $V_{\lfloor \frac{n_s}{2}\rfloor}$ is the Vandermonde determinant on $\lfloor \frac{n_s}{2}\rfloor$ variables $u_{1,s},\ldots,u_{\lfloor \frac{m}{2}\rfloor,s}$.
Let
\begin{eqnarray*}
\ol{X}_{k}=\left(x_{N-\lfloor \frac{n_k}{2}\rfloor+1},\ldots,x_{N}\right).
\end{eqnarray*}
Then
\begin{eqnarray*}
&&\left.\mathcal{D}_{m_1}^{(n_1)}\mathcal{D}_{m_2}^{(n_2)}\ldots\mathcal{D}_{m_k}^{(n_k)}\mathcal{S}_{\rho,X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};\ldots;u_{1,k},\ldots,u_{\lfloor \frac{n_k}{2}\rfloor,k})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \forall 1\leq s\leq k}\\
&=&\mathbf{E}\left(\sum_{i=1}^{\lfloor \frac{n_1}{2}\rfloor}(\phi(n_1)_{i_1}+\lfloor \frac{n_1}{2}\rfloor-i_1)^{m_1}\sum_{i_2=1}^{\lfloor\frac {n_2}{2}\rfloor}(\phi(n_2)_{i_2}+\lfloor\frac{n_2}{2}\rfloor-i_2)^{m_2}\cdot\ldots\cdot\sum_{i_k=1}^{\lfloor\frac{n_k}{2}\rfloor}(\phi(n_k)_{i_k}+\lfloor \frac{n_k}{2}\rfloor-i_k)^{m_k}\right).
\end{eqnarray*}
where $\mathbf{E}$ is the expectation with respect to the probability measure defined by (\ref{prb}), and $\mathcal{S}_{\rho,X}$ is the multi-dimensional Schur generating function as defined in Definition \ref{df59}.
\end{lemma}
\begin{proof}The theorem follows from explicit computations.
\end{proof}
\begin{lemma}\label{ll511}Suppose the assumptions in Lemma \ref{l59} hold. For $1\leq s\leq k$, let
\begin{eqnarray*}
t_s=N-\lfloor\frac{n_s}{2} \rfloor.
\end{eqnarray*}
For $1\leq s\leq k-1$, let
\begin{eqnarray*}
\tilde{\mathcal{D}}_{m_s,k}^{(n_s)}:=\frac{1}{\hat{V}_{\lfloor \frac{n_s}{2}\rfloor}}\left(\sum_{i=t_s+1}^{N}\left(u_{i}\frac{\partial}{\partial u_{i}}\right)^{m_s}\right)\hat{V}_{\lfloor \frac{n_s}{2}\rfloor}\prod_{i\in \{t_{s+1}+1,\ldots,t_s\}\cap I_2}\prod_{j= t_s+1}^{N}\left(\frac{1+y_{\ol{i}}u_j}{1+y_{\ol{i}}x_{\ol{j}}}\right),
\end{eqnarray*}
and let
\begin{eqnarray*}
\tilde{\mathcal{D}}_{m_k,k}^{(n_k)}:=\frac{1}{\hat{V}_{\lfloor \frac{n_k}{2}\rfloor}}\left(\sum_{i=t_k+1}^{N}\left(u_{i}\frac{\partial}{\partial u_{i}}\right)^{m_k}\right)\hat{V}_{\lfloor \frac{n_k}{2}\rfloor};
\end{eqnarray*}
where $\hat{V}_{\lfloor \frac{n_s}{2}\rfloor}$ is the Vandermonde determinant on $\lfloor \frac{n_s}{2}\rfloor$ variables $u_{t_s+1},\ldots,u_{N}$. Then
\begin{eqnarray*}
&&\left.\mathcal{D}_{m_1}^{(n_1)}\mathcal{D}_{m_2}^{(n_2)}\ldots\mathcal{D}_{m_k}^{(n_k)}\mathcal{S}_{\rho,X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};\ldots;u_{1,k},\ldots,u_{\lfloor \frac{n_k}{2}\rfloor,k})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \forall 1\leq s\leq k}\\
&=&\tilde{\mathcal{D}}_{m_1,k}^{(n_1)}\tilde{\mathcal{D}}_{m_2,k}^{(n_2)}\ldots\tilde{\mathcal{D}}_{m_k,k}^{(n_k)}\left\{
\left.\mathcal{S}_{\rho_{\lfloor\frac{n_k}{2} \rfloor},X}(u_{N-\lfloor\frac{n_k}{2} \rfloor+1},\ldots,u_N)\right\}\right|_{(u_{1},\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
where $\mathcal{S}_{\rho_{\lfloor\frac{n_k}{2} \rfloor},X}(u_1,\ldots,u_N)$ is the one-dimensional Schur generating function defined as in Definition \ref{df33}, and $\rho_{\frac{n_k}{2}}$ is a probability measure on $\GT_{\lfloor \frac{n_k}{2}\rfloor}^+$ defined as in Lemma \ref{l33}.
\end{lemma}
\begin{proof}We shall prove the lemma by induction on $k$. First of all, when $k=1$, the lemma obviously holds. Assume that the lemma holds when $k=l-1$, where $\geq 2$ is a positive integer. Then when $k=l$, we have
\begin{eqnarray*}
&&\left.\mathcal{D}_{m_1}^{(n_1)}\mathcal{D}_{m_2}^{(n_2)}\ldots\mathcal{D}_{m_l}^{(n_l)}\mathcal{S}_{\rho,X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};\ldots;u_{1,l},\ldots,u_{\lfloor \frac{n_l}{2}\rfloor,l})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \forall 1\leq s\leq l}\\
&=&\sum_{\lambda^{l}\in \GT_{\lfloor \frac{n_l}{2}\rfloor}}\rho_{\lfloor\frac{n_l}{2} \rfloor}(\lambda^l)\mathcal{D}_{m_l}^{(n_l)}\left(\frac{s_{\lambda^l}(u_{1,l},\ldots,u_{\lfloor \frac{n_l}{2}\rfloor,l})}{s_{\lambda^l}(\ol{X}_l)}\right)\\
&&\left.\mathcal{D}_{m_1,l-1}^{(n_1)}\ldots\mathcal{D}_{m_{l-1},l-1}^{(n_{l-1})}\mathcal{S}_{\rho(\cdot|\lambda^{l}),X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};\ldots;u_{1,{l-1}},\ldots,u_{\lfloor \frac{n_{l-1}}{2}\rfloor,{l-1}})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \forall 1\leq s\leq l}
\end{eqnarray*}
where $\rho(\cdot|\lambda^l)$ is a probability on $\prod_{s=1}^{l-1}\GT_{\lfloor\frac{n_s}{2} \rfloor}$ obtained from $\rho$ by conditional on the configuration $\lambda^l$ on $\GT_{\lfloor\frac{n_l}{2} \rfloor}$. By the induction hypothesis and Lemma \ref{lmm212}, we have
\begin{eqnarray*}
&&\left.\mathcal{D}_{m_1}^{(n_1)}\ldots\mathcal{D}_{m_{l-1}}^{(n_{l-1})}\mathcal{S}_{\rho(\cdot|\lambda^{l}),X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};\ldots;u_{1,{l-1}},\ldots,u_{\lfloor \frac{n_{l-1}}{2}\rfloor,{l-1}})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \forall 1\leq s\leq l-1}\\
&=&\tilde{\mathcal{D}}_{m_1,l-1}^{(n_1)}\tilde{\mathcal{D}}_{m_2,l-1}^{(n_2)}\ldots\tilde{\mathcal{D}}_{m_{l-1},l-1}^{(n_{l-1})}\left\{\left.\mathcal{S}_{\rho_{\lfloor\frac{n_{l-1}}{2} \rfloor}(\cdot|\lambda^l),X}(u_{N-\lfloor\frac{n_{l-1}}{2} \rfloor+1},\ldots,u_N)\right\}\right|_{(u_{1},\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&\tilde{\mathcal{D}}_{m_1,l}^{(n_1)}\tilde{\mathcal{D}}_{m_2,l}^{(n_2)}\ldots\tilde{\mathcal{D}}_{m_{l-1},l-1}^{(n_{l-1})}\left\{\left[\prod_{i\in \{t_{l},t_{l}+1,\ldots,t_{l-1}\}\cap I_2}\prod_{j\in t_{l-1}+1}^{N}\left(\frac{1+y_{\ol{i}}u_j}{1+y_{\ol{i}}x_{\ol{j}}}\right)\right]\right.\\
&&\left.\left.\frac{s_{\lambda^l}(x_{N-\lfloor\frac{n_l}{2} \rfloor+1},\ldots,x_{N-\lfloor\frac{n_{l-1}}{2}\rfloor},u_{N-\lfloor\frac{n_{l-1}}{2} \rfloor+1},\ldots,u_N)}{s_{\lambda^l}(x_{N-\lfloor\frac{n_l}{2} \rfloor+1},\ldots,x_N)}\right\}\right|_{(u_{1},\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&\tilde{\mathcal{D}}_{m_1,l}^{(n_1)}\tilde{\mathcal{D}}_{m_2,l}^{(n_2)}\ldots\tilde{\mathcal{D}}_{m_{l-1},l}^{(n_{l-1})}\left\{\left.\frac{s_{\lambda^l}(x_{N-\lfloor\frac{n_l}{2} \rfloor+1},\ldots,x_{N-\lfloor\frac{n_{l-1}}{2}\rfloor},u_{N-\lfloor\frac{n_{l-1}}{2} \rfloor+1},\ldots,u_N)}{s_{\lambda^l}(x_{N-\lfloor\frac{n_l}{2} \rfloor+1},\ldots,x_N)}\right\}\right|_{(u_{1},\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
Note also that
\begin{eqnarray*}
&&\left.\mathcal{D}_{m_l}^{(n_l)}\left(\frac{s_{\lambda^l}(u_{1,l},\ldots,u_{\lfloor \frac{n_l}{2}\rfloor,l})}{s_{\lambda^l}(\ol{X}_l)}\right)\right|_{(u_{1,l},\ldots,u_{\lfloor \frac{n_l}{2}\rfloor,l})=\ol{X}_l}=\sum_{j=1}^{\lfloor\frac{n_l}{2} \rfloor}\left(\lambda_j^{l}+\lfloor \frac{n_l}{2}\rfloor-j\right)^{m_l}\\
&=&\tilde{\mathcal{D}}_{m_l,l}^{(n_l)}\left.\left\{\frac{s_{\lambda^l}(u_{N-\lfloor\frac{n_l}{2} \rfloor+1},\ldots,\ldots,u_N)}{s_{\lambda^l}(x_{N-\lfloor\frac{n_l}{2} \rfloor+1},\ldots,x_N)}\right\}\right|_{(u_{1},\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
Then the lemma follows.
\end{proof}
Let $m_1,\ldots,m_k$ and $n_1,\ldots,n_k$ be as in Lemma \ref{l59}.
For $1\leq s\leq k$, we introduce the notation
\begin{eqnarray*}
\left.E_{l,s}=\mathcal{F}_{(l,\frac{n_s}{2N}) }(u_{N-\lfloor\frac{n_s}{2} \rfloor}+1,\ldots,u_N)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}.\\
\end{eqnarray*}
where $\mathcal{F}$ is defined by (\ref{flu}) and can be expressed as in (\ref{exf}).
Let $s_1<s_2$ be positive integers between $1$ and $k$. Define
\begin{eqnarray}
&&G_{s_1,s_2}(u_{N-\lfloor\frac{n_{s_2}}{2}\rfloor+1},\ldots,u_N)\label{g12}\\
&=&m_{s_1}\sum_{r=0}^{m_{s_1}-1}\left(\begin{array}{c}m_{s_1}-1\\r\end{array}\right)\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{N-\lfloor\frac{n_{s_1}}{2} \rfloor+1,\ldots,N\}}(r+1)!\notag\\
&&\times\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^{m_{s_1}}\frac{\partial}{\partial u_{b_1}}\left[\mathcal{F}_{(m_{s_2},\frac{n_{s_2}}{2N})}\right]\left(\frac{\partial}{\partial u_{b_1}}\left[\log \mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2} \rfloor},X}\right]\right)^{m_{s_1}-1-r}}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}.\notag
\end{eqnarray}
\begin{lemma}\label{ll66}
\begin{enumerate}
\item the degree of $N$ in $\left.G_{s_1,s_2}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}$ is at most $m_{s_1}+m_{s_2}$. Moreover, for any index $i$ the degree of $N$ in $\left.\frac{\partial}{\partial u_i}G_{s_1,s_2}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}$ is less than $m_{s_1}+m_{s_2}$.
\item
\begin{eqnarray*}
&&\frac{1}{\hat{V}_{\lfloor\frac{n_{s_1}}{2}\rfloor} \mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2}\rfloor},X} }m_{s_1}\sum_{i\in\{N-\lfloor \frac{n_{s_1}}{2}\rfloor+1,\ldots,N\}}\left(u_i\frac{\partial}{\partial u_i}\left[\mathcal{F}_{(m_{s_2},\frac{n_{s_2}}{2N})}\right]\right)\\
&&\left.\left(u_i\frac{\partial}{\partial u_i}\right)^{m_{s_1}-1}\left[\hat{V}_{\lfloor\frac{n_{s_1}}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2}\rfloor},X} \right]\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&\left.G_{s_1,s_2}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}+R
\end{eqnarray*}
where the degree of $N$ in $R$ is less than $l+k$.
\end{enumerate}
\end{lemma}
\begin{proof}We first prove Part (1). By Lemma \ref{l33}, the degree of $N$ in $\left(\frac{\partial}{\partial u_{b_1}}\left[\log \mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2} \rfloor},X}\right]\right)^{m_{s_1}-1-r}$ is at most $m_{s_1}-1-r$. By Lemma \ref{p36}, the degree of $N$ in $\frac{\partial}{\partial u_{b_1}}\left[\mathcal{F}_{(m_{s_2},\frac{n_{s_2}}{2N})}\right]$ is at most $m_{s_2}$. The summation gives $O(N^{r+1})$ terms. Therefore, the degree of $N$ in $\left.G_{s_1,s_2}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}$ is at most $m_{s_1}+m_{s_2}$. The fact that the degree of $N$ in $\frac{\partial}{\partial u_i}\left.G_{s_1,s_2}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}$ is at most $m_{s_1}+m_{s_2}$ also follows from Lemmas \ref{l33} and \ref{p36}, and by discussing the cases when $i\in\{b_1,\ldots,b_{r+1}\}$ and $i\notin\{b_1,\ldots,b_{r+1}\}$ separately.
Now we prove Part (2). We have
\begin{eqnarray*}
&&\frac{1}{\hat{V}_{\lfloor\frac{n_{s_1}}{2}\rfloor} \mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2}\rfloor},X} }m_{s_1}\sum_{i\in\{N-\lfloor \frac{n_{s_1}}{2}\rfloor+1,\ldots,N\}}\left(u_i\frac{\partial}{\partial u_i}\left[\mathcal{F}_{(m_{s_2},\frac{n_{s_2}}{2N})}\right]\right)\\
&&\left.\left(u_i\frac{\partial}{\partial u_i}\right)^{m_{s_1}-1}\left[\hat{V}_{\lfloor\frac{n_{s_1}}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2}\rfloor},X} \right]\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&m_{s_1}\sum_{t_0+t_1d_1+\ldots t_q d_q+r=m_{s_1}-1, t_1<t_2<\ldots<t_q}\left(\begin{array}{c}m_{s_1}-1\\r\end{array}\right)\sum_{\{b_1,\ldots,b_{r+1}\}\subset\{N-\lfloor\frac{n_{s_1}}{2} \rfloor+1,\ldots,N\}}(r+1)!\notag\\
&&\times\mathrm{Sym}_{b_1,\ldots,b_{r+1}}\frac{u_{b_1}^{m_{s_1}-t_0}\frac{\partial}{\partial u_{b_1}}\left[\mathcal{F}_{(m_{s_2},\frac{n_{s_2}}{2N})}\right]\left(\frac{\partial^{t_1}}{\partial u_{b_1}}\left[\log \mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2} \rfloor},X}\right]\right)^{d_1}\ldots\left(\frac{\partial^{t_q}}{\partial u_{b_1}}\left[\log \mathcal{S}_{\rho_{\lfloor \frac{n_{s_1}}{2} \rfloor},X}\right]\right)^{d_q}}{(u_{b_1}-u_{b_2})\ldots(u_{b_1}-u_{b_{r+1}})}
\end{eqnarray*}
By Lemmas \ref{l33} and \ref{p36}, the degree of $N$ in the expression above is at most
\begin{eqnarray}
d_1+\ldots+d_q+m_{s_2}+r+1.\label{dne}
\end{eqnarray}
Given that $t_0+t_1d_1+\ldots t_q d_q+r=m_{s_1}-1, t_1<t_2<\ldots<t_q$, the maximal of (\ref{dne}) achieves when $t_0=d_2=\ldots=d_q=0$, $t_1=1$, $d_1=m_{s_1}-r-1$; with maximal value $m_{s_1}+m_{s_2}$. Then Part (2) follows.
\end{proof}
\begin{lemma}\label{p510}Let $m_1,\ldots,m_k$ and $n_1,\ldots,n_k$ be as in Lemma \ref{l59}. Then
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N^{m_1+m_2+\ldots+m_k}}\left(\mathcal{D}_{m_1}^{(n_1)}-E_{m_1,s_1}\right)
\left(\mathcal{D}_{m_2}^{(n_2)}-E_{m_2,s_{2}}\right)\ldots\left(\mathcal{D}_{m_k}^{(n_k)}-E_{m_k,s_k}\right)\\
&& \left.\mathcal{S}_{\rho,X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};\ldots;u_{1,k},\ldots,u_{\lfloor \frac{n_k}{2}\rfloor,k})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \forall 1\leq s\leq m}\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{m_1+m_2+\ldots+m_k}}\sum_{P\in \mathcal{P}_{\emptyset}^{k}}\left.\prod_{(s_1,s_2)\in P}G_{s_1,s_2}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}.
\end{eqnarray*}
\end{lemma}
\begin{proof}The proposition follows from explicit computations and by analyzing the degree of $N$ of each term in the expansion. We sketch the proof here.
The lemma obviously holds when $k=1$, for which both the left hand side and the right hand side are 0. When $k=2$, by Lemma \ref{ll511}, we have
\begin{eqnarray*}
\mathcal{E}_2:&=&\mathcal{D}_{m_1}^{(n_1)}\mathcal{D}_{m_2}^{(n_2)} \left.\mathcal{S}_{\rho,X}(u_{1,1},\ldots,u_{\lfloor \frac{n_1}{2}\rfloor,1};u_{1,2},\ldots,u_{\lfloor \frac{n_2}{2}\rfloor,2})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=\ol{X}_s,\ \mathrm{for}\ s=1,2}\\
&=&\tilde{\mathcal{D}}_{m_{1,2}}^{(n_1)}\tilde{\mathcal{D}}_{m_2,2}^{(n_2)} \left.\mathcal{S}_{\rho_{\lfloor\frac{n_2}{2}\rfloor},X}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_N)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&\frac{1}{\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_N)}\tilde{\mathcal{D}}_{m_1}^{(n_1)}\tilde{\mathcal{D}}_{m_2}^{(n_2)}\\
&&\left.\left[ \exp\left(\log\left[\mathcal{S}_{\rho_{\lfloor\frac{n_2}{2} \rfloor},X}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_N)\right]\right)\right]\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&\frac{1}{\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_N)V_{\lfloor\frac{n_1}{2} \rfloor}(u_{N-\lfloor\frac{n_1}{2} \rfloor+1},\ldots,u_N)}\sum_{i\in\{N-\lfloor\frac{n_1}{2} \rfloor+1,\ldots,N\}}\left(u_{i}\frac{\partial}{\partial u_{i}}\right)^{m_1}\\
&&\frac{V_{\lfloor\frac{n_1}{2} \rfloor}(u_{N-\lfloor\frac{n_1}{2} \rfloor+1},\ldots,u_N)}{V_{\lfloor\frac{n_2}{2} \rfloor}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_N)}\prod_{l\in\{N-\lfloor\frac{n_2}{2} \rfloor+1,\ldots,N-\lfloor\frac{n_1}{2} \rfloor\}\cap I_2}\prod_{s=N-\lfloor\frac{n_1}{2}\rfloor+1}^{N}\left(\frac{1+y_{\ol{l}}u_s}{1+y_{\ol{l}}x_{\ol{s}}}\right)\\
&&\sum_{j\in\{N-\lfloor\frac{n_2}{2} \rfloor+1,\ldots,N\}}\left(u_{j}\frac{\partial}{\partial u_{j}}\right)^{m_2}V_{\lfloor\frac{n_2}{2} \rfloor}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_{N})\\
&&\left.\left[ \exp\left(\log\left[\mathcal{S}_{\rho_{\lfloor \frac{n_2}{2}\rfloor},X}(u_{N-\lfloor\frac{n_2}{2}\rfloor+1},\ldots,u_N)\right]\right)\right]\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
By Lemma \ref{lmm212}, we have
\begin{eqnarray*}
\prod_{l\in\{N-\lfloor\frac{n_2}{2} \rfloor+1,\ldots,N-\lfloor\frac{n_1}{2} \rfloor\}\cap I_2}\prod_{s=N-\lfloor\frac{n_1}{2}\rfloor+1}^{N}\left(\frac{1+y_{\ol{l}}u_s}{1+y_{\ol{l}}x_{\ol{s}}}\right)=\frac{\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}(u_{N-\lfloor \frac{n_1}{2}\rfloor}+1,\ldots,u_N)}{\mathcal{S}_{\rho_{\lfloor \frac{n_2}{2}\rfloor},X}(u_{N-\lfloor \frac{n_2}{2}\rfloor}+1,\ldots,u_N)}
\end{eqnarray*}
Then
\begin{eqnarray*}
\mathcal{E}_2&=&\frac{1}{\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}(u_{N-\lfloor\frac{n_2}{2} \rfloor+1},\ldots,u_N)V_{\lfloor\frac{n_1}{2} \rfloor}(u_{N-\lfloor\frac{n_1}{2} \rfloor+1},\ldots,u_N)}\\
&&\sum_{i\in\{N-\lfloor\frac{n_1}{2}\rfloor+1,\ldots,N\}}\left(u_{i}\frac{\partial}{\partial u_{i}}\right)^{m_1}V_{\lfloor\frac{n_1}{2} \rfloor}(u_{N-\lfloor\frac{n_1}{2} \rfloor+1},\ldots,u_N)\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}(u_{N-\lfloor\frac{n_2}{2}\rfloor+1},\ldots,u_N)\\
&&\left.\mathcal{F}_{m_2,\lfloor\frac{n_2}{2N} \rfloor}(u_{N-\lfloor\frac{n_2}{2}\rfloor+1},\ldots,u_N)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
Hence $\mathcal{E}_2$ is a sum of terms of the following form
\begin{eqnarray*}
&&\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\\
&&\left.\left[\frac{c_0u_{a_1}^{m_1-q_0}\frac{\partial^{q_1}}{\partial u_{a_1}^{q_1}}[\mathcal{F}_{m_2,\frac{n_2}{2N}}]\left[\frac{\partial^{q_2}}{\partial u_{a_1}^{q_2}}(\log \mathcal{S}_{\rho_{\lfloor\frac{n_1}{2}\rfloor},X})\right]^{d_2}\ldots \left[\frac{\partial^{q_t}}{\partial u_{a_1}^{q_t}}(\log \mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X})\right]^{d_t}}{(u_{a_1}-u_{a_2})\ldots (u_{a_1}-u_{a_{r+1}})}\right]\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}.
\end{eqnarray*}
such that
\begin{itemize}
\item $r$, $q_0,q_1,\ldots,q_t$, $d_2,\ldots,d_t$ are nonnegative integers; and
\item $q_2<q_3<\ldots<q_t$; and
\item
\begin{eqnarray}
q_0+q_1+q_2d_2+\ldots+q_td_t+r=m_1; \label{ssdk}
\end{eqnarray}
and
\item $\{a_1,\ldots,a_{r+1}\}\subset\{N-\lfloor \frac{n_1}{2}\rfloor+1,\ldots,N\}$
\end{itemize}
When $q_1=0$, we obtain $E_{m_1,s_1}E_{m_2,s_2}$.
Now we consider the terms corresponding to $q_1\geq 1$. By Lemma \ref{p36}, the degree of $N$ in $\partial_{a_1}^{q_1}\left[\mathcal{F}_{m_2,\frac{n_2}{2N}}\right]$ is at most $m_2$. By Lemma \ref{l33}, the total degree of $N$ in these terms is at most $m_2+d_2+\ldots+d_t+r+1$. By (\ref{ssdk}) and the assumption that $s_1\geq 1$, we have
\begin{eqnarray*}
m_2+d_2+\ldots+d_t+r+1\leq m_2+m_1；
\end{eqnarray*}
and the equality holds when $q_0=d_3=\ldots=d_t=0$, $q_1=q_2=1$; $d_2=m_1-1-r$; this corresponds to $G_{1,2}$, in which the degree of $N$ is at most $m_1+m_2$ . The degree of $N$ is less than $m_1+m_2$ in all the other terms. This completes the proof when $k=2$.
We shall finish the rest of the proof by induction. For $1\leq l\leq k-1$, let
\begin{eqnarray*}
A_l=\prod_{[i\in\{t_{l+1}+1,\ldots,t_l\}]}\prod_{j=t_l+1}^{N}\left(\frac{1+y_{\ol{i}}u_j}{1+y_{\ol{i}}x_{\ol{j}}}\right)
\end{eqnarray*}
By Lemma \ref{lmm212}, we have
\begin{eqnarray*}
A_l=\frac{\mathcal{S}_{\rho_{\lfloor\frac{n_l}{2} \rfloor},X}(u_{N-\lfloor\frac{n_l}{2} \rfloor+1,},\ldots,u_N)}{\mathcal{S}_{\rho_{\lfloor\frac{n_{l+1}}{2} \rfloor},X}(u_{N-\lfloor\frac{n_{l+1}}{2} \rfloor+1,},\ldots,u_N)}
\end{eqnarray*}
Assume that the lemma holds for $k=r-1$, where $r\geq 2$ is a positive integer. When $k=r$, by induction hypothesis, we have
\begin{eqnarray*}
&&\frac{1}{\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}}\sum_{[i_1\in\{N-\lfloor\frac{n_1}{2}\rfloor+1,\ldots,N\}]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{m_1}A_1\frac{\hat{V}_{\lfloor\frac{n_1}{2} \rfloor}}{\hat{V}_{\lfloor\frac{n_2}{2} \rfloor}}\sum_{[i_2\in\{N-\lfloor\frac{n_2}{2} \rfloor\rfloor+1,\ldots,N\}]}\left(u_{i_2}\frac{\partial}{\partial u_{i_2}}\right)^{m_2}
\\&&\frac{\hat{V}_{\lfloor\frac{n_2}{2}\rfloor}}{\hat{V}_{\lfloor\frac{n_3}{2}\rfloor}}A_{2}\cdots\left.\sum_{[i_r\in\{N-\lfloor\frac{n_r}{2} \rfloor+1,\ldots,N\}}\left(u_{i_r}\frac{\partial}{\partial u_{i_r}}\right)^{m_r}\left[\hat{V}_{\lfloor\frac{n_r}{2} \rfloor}\mathcal{S}_{\rho_{\lfloor \frac{n_t}{2}\rfloor},X}\right]\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}\\
&=&\frac{1}{\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}}\sum_{[i_1\in\{N-\lfloor\frac{n_1}{2}\rfloor+1,\ldots,N\}]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{m_1}\left[\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}\right]\\
&&\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,r\}]}\mathcal{F}_{(m_{w_1},\frac{n_{w_1}}{2N})}\mathcal{F}_{(m_{w_2},\frac{n_{w_2}}{2N})}\ldots \mathcal{F}_{(m_{w_p},\frac{n_{w_p}}{2N})}\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}.\\
&=& S_1+S_2+S_3;
\end{eqnarray*}
where by induction hypothesis the degree of $N$ in $R_{1,w_1,\ldots,w_p}$ is less than $\sum_{i=2}^{s}l_i-\sum_{i=1}^{p}l_{w_i}$; and
\begin{eqnarray*}
S_1&=&\left\{\frac{1}{\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}}\sum_{[i_1\in\{N-\lfloor\frac{n_1}{2} \rfloor+1,\ldots,N\}]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}\left[\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}\right]\right\}\\
&&\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,r\}]}E_{m_{w_1},w_1}E_{m_{w_2},w_2}\ldots E_{m_{w_p},w_p}\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}.
\end{eqnarray*}
and
\begin{eqnarray*}
S_2&=&\frac{\ell_1}{\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}}\sum_{[i_1\in\{N-\lfloor\frac{n_1}{2} \rfloor+1,\ldots,N\}]}\left\{\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1-1}\left[\hat{V}_{\frac{n_1}{2}}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}\right]\right\}\\
&&\times\left\{\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}\mathcal{F}_{(m_{w_1},\frac{n_{w_1}}{2N})}\mathcal{F}_{(m_{w_2},\frac{n_{w_2}}{2N})}\ldots \mathcal{F}_{(m_{w_p},\frac{w_p}{2N})}\right.\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)};
\end{eqnarray*}
and $S_3$ consists of all the other terms.
By the definition of $E_{m_i,i}$ we have
\begin{eqnarray*}
S_1&=&E_{m_1,1}\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,r\}]}E_{m_{w_1},w_1}E_{m_{w_2},w_2}\ldots E_{m_{w_p},w_p}\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}.
\end{eqnarray*}
By Lemma \ref{ll66}, we have
\begin{eqnarray*}
S_2&=&\left\{\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,r\}]}\sum_{x=1}^{p}E_{m_{w_1,w_1}}\ldots E_{m_{w_{x-1}},w_{x-1}} E_{m_{w_{x+1}},w_{x+1}}\ldots E_{m_{w_p},w_p}\right.\right.\\
&&\left.\left.\left.\left[\left(G_{1,x}+R_{1,x}\right)\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{1,w_1,\ldots,w_p}\right)\right]\right)\right\}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)},
\end{eqnarray*}
where the degree of $N$ in $R_{1,x}$ is less than $l_1+l_x$.
Hence we have
\begin{eqnarray*}
S_2&=&\left\{\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,r\}]}\sum_{x=1}^{p}E_{m_{w_1,w_1}}\ldots E_{m_{w_{x-1}},w_{x-1}} E_{m_{w_{x+1}},w_{x+1}}\ldots E_{m_{w_p},w_p}\right.\right.\\
&&\left.\left.\left.\left[G_{1,x}\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{w_1,\ldots,\hat{w}_x,\ldots,w_p}\right)\right]\right\}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
where the degree of $N$ in $R_{w_1,\ldots,\hat{w}_x,\ldots,w_p}$ is less than $\sum_{i=2}^{t}l_i-\sum_{i=1}^{p}l_{w_i}$.
Note that
\begin{eqnarray*}
S_1+S_2&=&\sum_{p=0}^{r}\sum_{[w_1,\ldots,w_p\in\{1,2,\ldots,r\}]}E_{m_{w_1},w_1}E_{m_{w_2},w_2}\ldots E_{m_{w_p},w_p}\\
&&\left.\left(\sum_{P\in \mathcal{P}_{w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P} G_{a,b}+R_{w_1,\ldots,w_p}\right)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
where the degree of $N$ in $R_{w_1,\ldots,w_p}$ is less than $\sum_{i=1}^{t}l_i-\sum_{i=1}^{p}l_{w_i}$.
It remains to show that $S_3$ does not contribute to the leading terms. Define
\begin{eqnarray*}
\mathcal{H}_{j_1,\ldots,j_p}=\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{r}}\prod_{(a,b)\in P}G_{a,b}+R_{1,w_1,\ldots,w_p}\right);
\end{eqnarray*}
By Lemma \ref{ll66}, the degree of $N$ in $\left.\mathcal{H}_{j_1,\ldots,j_p}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is at most $\sum_{i=2}^r l_i-\sum_{j=1}^p l_{w_j}$. Moreover, by Lemma \ref{ll66}, for any index $i$, the degree of $N$ in $\left.\frac{\partial}{\partial u_i}\mathcal{H}_{j_1,\ldots,j_p}\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}$ is less than $\sum_{i=2}^t l_i-\sum_{j=1}^p l_{w_j}$.
We write
\begin{eqnarray*}
&&\frac{1}{\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}}\sum_{[i_1\in\{N-\lfloor\frac{n_1}{2}\rfloor+1,\ldots,N\}]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{m_1}\left[\hat{V}_{\lfloor \frac{n_1}{2}\rfloor}\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}\right]\\
&&\left.\left(\sum_{p=0}^{r-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,r\}]}\mathcal{F}_{(m_{w_1},\frac{n_{w_1}}{2N})}\mathcal{F}_{(m_{w_2},\frac{n_{w_2}}{2N})}\ldots \mathcal{F}_{(m_{w_p},\frac{n_{w_p}}{2N})}\mathcal{H}_{j_1,\ldots,j_p}\right)\right|_{(u_1,\ldots,u_N)=(x_1,\ldots,x_N)}
\end{eqnarray*}
as a sum of terms of the following form
\begin{eqnarray}
\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left[\frac{u_{a_1}^{m_1-s_0}(\partial_{a_1}^{s_1}[\log\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}])^{d_1}\ldots (\partial_{a_t}^{s_t}[\log\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}])^{d_t} }{(u_{a_1}-u_{a_2})\ldots(u_{a_1}-u_{a_{q+1}})}\right.\label{ssa1}\\
\left.\frac{\partial_{a_1}^{f_1}\mathcal{F}_{(m_{w_1},\frac{n_{w_1}}{2N})} \ldots\partial_{a_1}^{f_p}\mathcal{F}_{(m_{w_p},\frac{n_{w_p}}{2N})}\partial_{a_1}^{h_0}\mathcal{H}_{j_1,\ldots,j_p} }{}\right]\notag
\end{eqnarray}
where
\begin{itemize}
\item $\{a_1,\ldots,a_{q+1}\}\subset\{N-\lfloor\frac{n_1}{2} \rfloor+1,\ldots,N\}$;
\item $s_1<s_2<\ldots<s_t$ are positive integers;
\item $f_1,\ldots,f_p,h_0$ are nonnegative integers;
\item
\begin{eqnarray}
q+s_0+s_1d_1+\ldots+s_td_t+f_1+\ldots+f_p+h_0=m_1\label{rsdm}
\end{eqnarray}
\end{itemize}
By Lemma \ref{l33}, the degree of $N$ in $[\log\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}])^{d_1}\ldots (\partial_{a_t}^{s_t}[\log\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}])^{d_t}$ is at most $d_1+\ldots+d_t$; therefore, the terms in (\ref{ssa1}) with highest degree of $N$ has the form
\begin{eqnarray}
\mathrm{Sym}_{a_1,\ldots,a_{q+1}}\left[\frac{u_{a_1}^{m_1}(\partial_{a_1}[\log\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}])^{d_1} }{(u_{a_1}-u_{a_2})\ldots(u_{a_1}-u_{a_{q+1}})}\right.\\
\left.\frac{\partial_{a_1}^{f_1}\mathcal{F}_{(m_{w_1},\frac{n_{w_1}}{2N})} \ldots\partial_{a_1}^{f_p}\mathcal{F}_{(m_{w_p},\frac{n_{w_p}}{2N})}\partial_{a_1}^{h_0}\mathcal{H}_{j_1,\ldots,j_p} }{}\right]
\label{ssa2}
\end{eqnarray}
where
\begin{eqnarray}
s_0=d_2=\ldots=d_t=0;\ s_1=1.\label{sd0}
\end{eqnarray}
Let
\begin{eqnarray*}
B=\{i\in\{1,2,\ldots,p\}:f_i=0\}.
\end{eqnarray*}
Then
\begin{eqnarray*}
(\ref{ssa2})=\left[\prod_{i\in B}\mathcal{F}_{(m_{w_i},\frac{n_{w_i}}{2N})}\right] S(u_1,\ldots,u_{N})
\end{eqnarray*}
It suffices to show that the degree of $N$ in $S$, except for $S_1$ and $S_2$, is less than $\sum_{i=1}^{s}l_i-\sum_{i\in B}l_i$. Note that the degree of $N$ in $\partial_{a_1}[\log\mathcal{S}_{\rho_{\lfloor \frac{n_1}{2}\rfloor},X}])^{d_1}$ is at most $d_1$ by Lemma \ref{l33}. The summation over $\{a_1,\ldots,a_{q+1}\}\subset\{N-\lfloor\frac{n_1}{2} \rfloor+1,\ldots,N\}$ gives $O(N^{q+1})$ terms. By Lemma \ref{p36}, when $i\notin B$, the degree of $N$ in $\partial_{a_1}^{f_i}\mathcal{F}_{(m_{w_i},\frac{n_{w_i}}{2N})} $ is at most $m_{w_i}$. Therefore the degree of $N$ in $S(u_1,\ldots,u_N)$ is at most
\begin{eqnarray*}
\sum_{i=2}^{r}m_i-\sum_{i=1}^{p}m_{w_i}+d_1+\sum_{i\in \{1,2,\ldots,p\}\setminus B}m_{w_i}+q+1
\end{eqnarray*}
By (\ref{rsd}) and (\ref{sd0}), if $|B|\leq p-2$, $q+d_1+1\leq m_1-1$, then the degree of $N$ in $S(u_1,\ldots,u_N)$ is at most
\begin{eqnarray*}
\sum_{i=1}^{r}m_i-\sum_{i\in B}m_{w_i}-1
\end{eqnarray*}
Therefore only the terms where at most one $f_i$ is nonzero contribute to the leading order. In these terms if $h_0>0$, then by Lemma \ref{l67}, the degree of $N$ is less than $\sum_{i=1}^{r}m_i-\sum_{i\in B}m_{w_i}$. So only the terms where $h_0=0$ and at most one $f_i$ is nonzero contribute to the leading order. These terms are in $S_1$ and $S_2$. Then the proof is complete.
\end{proof}
\begin{theorem}\label{gff1}The collection of random variables
\begin{eqnarray*}
\{N^{-l}\left[p_{l}^{((1-\kappa) N)}-\mathbb{E}p_{l}^{((1-\kappa) N)}\right]\}_{l\in \NN;\kappa=a_1,\ldots,a_m}
\end{eqnarray*}
converges, as $N\rightarrow\infty$, in the sense of moments, to the Gaussian vector with zero mean and covariance
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\mathrm{cov}\left(p_{l_1}^{\lfloor(1-\kappa_1) N\rfloor},p_{l_2}^{\lfloor(1-\kappa_2)N\rfloor}\right)}{N^{l_1+l_2}}\\
&=&\frac{(1-\kappa_1)^{l_1}(1-\kappa_2)^{l_2}}{(2\pi\mathbf{i})^2}\sum_{i=1}^{n}\sum_{j=1}^{n}\oint_{|z-x_i|=\epsilon}\oint_{|w-x_j|=\epsilon}\left(\sum_{i=1}^{n}\frac{z}{n(z-x_i)}+\frac{z H(z)}{1-\kappa_1}\right)^{l_1}\\
&&\times\left(\sum_{i=1}^{n}\frac{w}{n(w-x_j)}+\frac{w H(w)}{1-\kappa_2}\right)^{l_2}Q(z,w)dzdw,
\end{eqnarray*}
where
\begin{itemize}
\item $\epsilon>0$ is sufficiently small such that the disk centered at $x_i$ with radius $\epsilon$ contains exactly one singularity $x_i$ of the integrand.
\item the $z$- and $w$-contours of integration are counter-clockwise.
\end{itemize}
\end{theorem}
\begin{proof}The theorem follows from Lemma \ref{p510} similar arguments as in the single-level case. The only difference is in the expansion $\{b_1,\ldots,b_{r+1}\}\subset\{1,2,\ldots,\lfloor (1-t_1)N\rfloor\}$, while $\{a_1,\ldots,a_{q+1}\}\subset\{1,2,\ldots,\lfloor(1-t_2)N \rfloor\}$.
\end{proof}
\section{Piecewise Boundary Conditions}\label{pbr}
In this section, we introduce the piecewise boundary conditions on the bottom boundary of a contracting square-hexagon lattice, and review the limit shape results for perfect matchings on such a graph.
For $N\geq 1$, let $\lambda(N)\in \GT_N^+$. We consider the following special asymptotical case of $\lambda(N)$ as $N\rightarrow\infty$. Let
\begin{eqnarray*}
\Omega=(\Omega_1<\Omega_2<\ldots<\Omega_N)=(\lambda_N(N)+1,\lambda_{N-1}(N)+2,\ldots,\lambda_1(N)+N)
\end{eqnarray*}
Indeed, $\Omega_1,\ldots,\Omega_N$ are the locations of the $N$ remaining vertices on the bottom boundary of the contracting square-hexagon lattice.
Assume
\begin{eqnarray}
\Omega&=&(A_1,A_1+1,\ldots.B_1-1,B_1,\label{abt}\\
&&A_2,A_2+1,\ldots,B_2-1,B_2,\ldots,A_s,A_s+1,\ldots,B_s-1,B_s).\notag
\end{eqnarray}
where
\begin{eqnarray*}
\sum_{i=1}^{s}(B_i-A_i+1)=N.
\end{eqnarray*}
and $s$ is a fixed positive integer independent of $N$. Suppose as $N\rightarrow\infty$,
\begin{eqnarray*}
A_i(N)=a_iN+o(N),\qquad B_i(N)=b_iN+o(N),\qquad \mathrm{for}\ 1\leq i\leq s,
\end{eqnarray*}
and $a_1<b_1<\ldots<a_s<b_s$ are fixed parameters independent of $N$ and satisfy $\sum_{i=1}^{s}(b_i-a_i)=1$. Assume the edge weights $\{x_i\}_{i=1}^{N}$ $\{y_j\}_{j\in I_2\cap\{1,2,\ldots,N\}}$ satisfy (\ref{px}) and (\ref{py}).
Let $\Sigma_N$ be the permutation group of $N$ elements and let $\sigma\in \Sigma_N$. Let
\begin{eqnarray*}
X=(x_1,\ldots,x_N).
\end{eqnarray*}
Let $x_1,\ldots, x_n$ be all the distinct elements in $\{x_1,\ldots,x_N\}$. Let $\Sigma_N^{X}$ be the subgroup of $\Sigma_N$ that preserves the value of $X$; more precisely
\begin{eqnarray*}
\Sigma_N^{X}=\{\sigma\in \Sigma_N: x_{\sigma(i)}=x_i,\ \mathrm{for}\ 1\leq i\leq N\}
\end{eqnarray*}
Let $[\Sigma_N/\Sigma_N^X]^r$ be the collection of all the right cosets of $\Sigma_N^X$ in $\Sigma_N$. More precisely,
\begin{eqnarray*}
[\Sigma_N/\Sigma_N^X]^r=\{\Sigma_N^X\sigma:\sigma\in \Sigma_N\},
\end{eqnarray*}
where for each $\sigma\in \Sigma_N$
\begin{eqnarray*}
\Sigma_N^X\sigma=\{\xi\sigma:\xi\in \Sigma_N^X\}
\end{eqnarray*}
and $\xi\sigma\in \Sigma_N$ is defined by
\begin{eqnarray*}
\xi\sigma(k)=\xi(\sigma(k)),\ \mathrm{for}\ 1\leq k\leq N.
\end{eqnarray*}
For $1\leq j\leq N$, let
\begin{eqnarray}
\eta_j^{\sigma}(N)=|\{k:k>j,x_{\sigma(k)}\neq x_{\sigma(j)}\}|.\label{et}
\end{eqnarray}
For $1\leq i\leq n$, let
\begin{eqnarray}
\Phi^{(i,\sigma)}(N)=\{\lambda_j(N)+\eta_j^{\sigma}(N):x_{\sigma(j)}=x_i\}\label{pis}
\end{eqnarray}
and let $\phi^{(i,\sigma)}(N)$ be the partition obtained by decreasingly ordering all the elements in $\Phi^{(i,\sigma)}(N)$. Let $\Sigma_N^{X}$ be the subgroup $\Sigma_N$ that preserves the value of $X$; more precisely
\begin{eqnarray*}
\Sigma_N^{X}=\{\sigma\in \Sigma_N: x_{\sigma(i)}=x_i,\ \mathrm{for}\ 1\leq i\leq N\}
\end{eqnarray*}
Let $[\Sigma/\Sigma_N^X]^r$ be the collection of all the right cosets of $\Sigma_N^X$ in $\Sigma_N$. More precisely,
\begin{eqnarray*}
[\Sigma/\Sigma_N^X]^r=\{\Sigma_N^X\sigma:\sigma\in \Sigma_N\},
\end{eqnarray*}
where for each $\sigma\in \Sigma_N$
\begin{eqnarray*}
\Sigma_N^X\sigma=\{\xi\sigma:\xi\in \Sigma_N^X\}
\end{eqnarray*}
and $\xi\sigma\in \Sigma_N$ is defined by
\begin{eqnarray*}
\xi\sigma(k)=\xi(\sigma(k)),\ \mathrm{for}\ 1\leq k\leq N.
\end{eqnarray*}
For simplicity, we make the following assumptions.
\begin{assumption}\label{ap423}Let $(x_1,\ldots,x_N)$ be an $N$-tuple of real numbers at which we evaluate the Schur polynomial.
\begin{itemize}
\item $N$ is an integral multiple of $n$; and.
\item $\{x_i\}_{i=1}^{N}$ are periodic with period $n$, i.e., $x_{i}=x_{j}$ for $1\leq i,j\leq N$ and $[i\mod n]=[j\mod n]$;
\item $x_1>x_2>\ldots>x_n>0$.
\end{itemize}
\end{assumption}
We may further make the assumptions below
\begin{assumption}\label{ap32}Assume $x_{1,N}=x_1>0$ and $(x_{2,N},\ldots,x_{n,N})$ changes with $N$. Assume that for each fixed $N$, $(x_{1,N},\ldots,x_{n,N})$ satisfies Assumption \ref{ap423}. Moreover, assume that
\begin{eqnarray*}
\liminf_{N\rightarrow\infty} \frac{\log\left(\min_{1\leq i<j\leq n}\frac{x_{i,N}}{x_{j,N}}\right)}{\log N}\geq \alpha>0,
\end{eqnarray*}
where $\alpha$ is a sufficiently large positive constant independent of $N$.
\end{assumption}
Let $\ol{\sigma}_0\in [\Si_N/\Si_N^X]^r$ be the unique element in $[\Si_N/\Si_N^X]^r$ satisfying the condition that for any representative $\sigma_0\in\ol{\sigma}_0$, we have
\begin{eqnarray}
x_{\si_0(1)}\geq x_{\si_0(2)}\geq\ldots\geq x_{\si_0(N)}.\label{sz}
\end{eqnarray}
Let $\bm_{i}$ be the limit of the counting measures for $\phi^{(i,\si_0)}(N)$ as $N\rightarrow\infty$.
\begin{assumption}\label{ap428}Assume $x_1,\ldots,x_N$ satisfy Assumption \ref{ap423}.
Let $A_i$, $B_i$ be given as in (\ref{abt}). For $1\leq i\leq s$, let
\begin{eqnarray*}
B_i-A_i+1=K_i.
\end{eqnarray*}
By (\ref{abt}), we may assume
\begin{eqnarray*}
&&\lambda_1=\lambda_2=\ldots=\lambda_{K_s}=\mu_1;\\
&&\lambda_{K_s+1}=\lambda_{K_s+2}=\ldots=\lambda_{K_s+K_{s-1}}=\mu_2;\\
&&\ldots\\
&&\lambda_{\sum_{t=2}^{s}K_t}=\lambda_{1+\sum_{t=2}^{s}K_t}=\ldots=\lambda_{\sum_{t=1}^{s}K_t}=\mu_s;
\end{eqnarray*}
and note that
\begin{eqnarray}
\mu_1>\ldots>\mu_s\label{mi}
\end{eqnarray}
are all the distinct elements in $\{\lambda_1,\lambda_2,\ldots,\lambda_N\}$. Let
\begin{eqnarray}
J_i=\{t:1\leq p\leq N, 1\leq t\leq s, [\si_0(p)\mod n]=i,\lambda_p=\mu_t\}\label{ji}.
\end{eqnarray}
Suppose that all the following conditions hold
\begin{itemize}
\item If $1\leq i<j\leq n$, $\ell\in J_i$, and $t\in J_j$, then $\ell<t$; and
\item For any $p,q$ satisfying $1\leq p\leq s$ and $1\leq q\leq s$, and $q>p$.
\begin{eqnarray*}
C_1N \leq \mu_p-\mu_q\leq C_2N
\end{eqnarray*}
where $C_1$, $C_2$ are constants independent of $N$.
\end{itemize}
\end{assumption}
Let
\begin{eqnarray}
H_{\mathbf{m}_i}(u)=\int_{0}^{\ln(u)}R_{\mathbf{m}_i}(t)dt+\ln\left(\frac{\ln(u)}{u-1}\right)\label{hmi}
\end{eqnarray}
and $\mathbf{R}_{\mathbf{m}_i}$ is the Voiculescu R-transform of $\mathbf{m}_i$ given by
\begin{eqnarray*}
R_{\bm_i}=\frac{1}{S_{\bm_i}^{(-1)}(z)}-\frac{1}{z};
\end{eqnarray*}
Where $S_{\bm_i}$ is the moment generating function for $\bm_i$ given by
\begin{eqnarray*}
S_{\bm_i}(z)=z+M_1(\bm_i)z^2+M_2(\bm_i)z^3+\ldots;
\end{eqnarray*}
$M_k(\bm_i)=\int_{\RR}x^k\bm_i(dx)$; and $S_{\bm_i}^{-1}(z)$ is the inverse series of $S_{\bm_i}(z)$.
See also Section 2.2 of \cite{bg} for details.
\begin{proposition}\label{tm1}Suppose Assumptions \ref{ap32} and \ref{ap428} hold. Let $\kappa\in(0,1)$ be a positive number. Let $\rho_{\lfloor (1-\kappa)N\rfloor}$ be a probability measure on $\GT_{\lfloor (1-\kappa)N\rfloor}^+$ as defined in Lemma \ref{l33} or Remark \ref{rm25}. Let $\bm[\rho_{\lfloor (1-\kappa)N\rfloor}]$ be the corresponding random counting measure. Then as $N\rightarrow\infty$, $\bm[\rho_{\lfloor (1-\kappa)N\rfloor}]$ converge in probability, in the sense of moments to a deterministic measure $\bm^{\kappa}$, whose moments are given by
\begin{eqnarray*}
\int_{\RR}x^{p}\textbf{m}^{\kappa}(dx)=
\frac{1}{2n(p+1)\pi \mathbf{i}}\sum_{i=1}^{n}\oint_{C_{1}}\frac{dz}{z}\left(zQ_{i,\kappa}'(z)+\frac{n-i}{n}+\frac{z}{n(z-1)}\right)^{p+1}
\end{eqnarray*}
where for $1\leq i\leq n$
\begin{eqnarray*}
Q_{i,\kappa}(z)=\left\{\begin{array}{cc}\frac{1}{(1-\kappa)n}\left[
H_{\bm_i}(z)-(n-i)\log z+\kappa\sum_{r\in\{1,2,\ldots,n\}\cap I_2}\log\frac{1+y_rzx_1}{1+y_rx_1}\right]&\mathrm{if}\ i=1\\\frac{1}{(1-\kappa)n}\left[H_{\bm_i}(z)-(n-i)\log z\right]
&\mathrm{otherwise}\end{array}\right.
\end{eqnarray*}
and for $i\geq n+1$,
\begin{eqnarray*}
Q_{i,\kappa}(z)=\left\{\begin{array}{cc}Q_{(i\mod n),\kappa}(z),&\mathrm{if}\ (i\mod n)\neq 0\\Q_{n,\kappa}(z),&\mathrm{if}\ (i\mod n)= 0 \end{array}\right\}.
\end{eqnarray*}
\end{proposition}
\begin{proof}See Theorem 2.18 of \cite{Li18}.
\end{proof}
\begin{lemma}\label{p437}Let $k$ be a positive integer such that $1\leq k\leq N$. Let
\begin{eqnarray*}
w_{i}=\left\{\begin{array}{cc}u_i&\mathrm{if}\ 1\leq i\leq k\\x_i&\mathrm{if}\ k+1\leq i\leq N\end{array}\right.
\end{eqnarray*}
Assume
\begin{eqnarray*}
k=qn+r,\qquad \mathrm{where}\ r<n,
\end{eqnarray*}
and $q,r$ are positive integers.
Then the Schur function can be computed by the following formula
\begin{eqnarray}
&&\label{s0s}s_{\lambda}(w_1,\ldots,w_N)\\
&=&\sum_{\ol{\sigma}\in[\Sigma_N/\Sigma_N^X]^r} \left(\prod_{i=1}^{n}x_i^{|\phi^{(i,\sigma)}(N)|}\right)\left(\prod_{i=1}^{r}s_{\phi^{(i,\sigma)}(N)}\left(\frac{u_i}{x_i},\frac{u_{n+i}}{x_i}\ldots,\frac{u_{qn+i}}{x_i},1,\ldots,1\right)\right)\notag\\
&&\times\left(\prod_{i=r+1}^{n}s_{\phi^{(i,\sigma)}(N)}\left(\frac{u_i}{x_i},\frac{u_{n+i}}{x_i}\ldots,\frac{u_{(q-1)n+i}}{x_i},1,\ldots,1\right)\right)\notag\\
&&\times\left(\prod_{i<j,x_{\sigma(i)}\neq x_{\sigma(j)}}\frac{1}{w_{\sigma(i)}-w_{\sigma(j)}}\right)\notag
\end{eqnarray}
where $\sigma\in \ol{\sigma}\cap \Sigma_N$ is a representative.
\end{lemma}
\begin{proof}See Corollary 3.4 of \cite{Li18}.
\end{proof}
\begin{theorem}\label{tm2}Under Assumptions \ref{ap32} and \ref{ap428}, for each given $\{a_i,b_i\}_{i=1}^{n}$, when $\alpha$ in Assumption \ref{ap32} is sufficiently large and $k\leq n$ we have
\begin{eqnarray}
&&\lim_{N\rightarrow\infty}\frac{1}{N}\log \frac{s_{\lambda(N)}(u_1x_{1,N},\ldots,u_kx_{k,N},x_{k+1,N},\ldots,x_{N,N})}{s_{\lambda(N)}(x_{1,N},\ldots,x_{N,N})}=\sum_{i=1}^{k}[Q_i(u_i)]\label{fc}
\end{eqnarray}
where for $1\leq i\leq k$,
\begin{enumerate}
\item if $[i\mod n]\neq 0$,
\begin{eqnarray*}
Q_i(u)=\frac{H_{\mathbf{m}_{i\mod n}}(u)}{n}-\frac{(n-[i\mod n])\log(u)}{n}.
\end{eqnarray*}
and the convergence of (\ref{fc}) is uniform when $u_1,\ldots,u_k$ are in an open complex neighborhood of $1$.
\item if $[i\mod n]=0$,
\begin{eqnarray*}
Q_i(u)=\frac{H_{\mathbf{m}_n}(u)}{n}.
\end{eqnarray*}
\end{enumerate}
\end{theorem}
\begin{proof}See Theorem 2.8 of \cite{Li18}.
\end{proof}
\begin{lemma}\label{l440}Let $\si_0$ satisfy (\ref{sz}), and let $\ol{\si}_0\in [\Si_N/ \Si_N^X]^r$. For $1\leq i\leq k$, assume $\frac{u_i}{x_i}$ is in an open complex neighborhood of $1$. For any $\si\in \Si_N$, let
\begin{eqnarray*}
&&L_{\si}\left(\frac{u_1}{x_1},\ldots,\frac{u_k}{x_k}\right)\\
&=& \left(\prod_{i=1}^{n}x_i^{|\phi^{(i,\sigma)}(N)|}\right)\left(\prod_{i=1}^{r}s_{\phi^{(i,\sigma)}(N)}\left(\frac{u_i}{x_i},\frac{u_{n+i}}{x_i}\ldots,\frac{u_{qn+i}}{x_i},1,\ldots,1\right)\right)\notag\\
&&\times\left(\prod_{i=r+1}^{n}s_{\phi^{(i,\sigma)}(N)}\left(\frac{u_i}{x_i},\frac{u_{n+i}}{x_i}\ldots,\frac{u_{(q-1)n+i}}{x_i},1,\ldots,1\right)\right)\notag\\
&&\times\left(\prod_{i<j,x_{\sigma(i)}\neq x_{\sigma(j)}}\frac{1}{w_{\sigma(i)}-w_{\sigma(j)}}\right)
\end{eqnarray*}
Suppose that Assumption \ref{ap32} holds. When $\alpha$ in Assumption \ref{ap32} is sufficiently large, we have
\begin{eqnarray*}
\left|\frac{L_{\si_0}}{L_{\si}}\right|\geq e^{CN}
\end{eqnarray*}
where $C>0$ is a constant independent of $\si$, $N$ and $(u_1,\ldots,u_k)$. Moreover,
\begin{eqnarray*}
\lim_{\alpha\rightarrow+\infty} C=+\infty.
\end{eqnarray*}
\end{lemma}
\begin{proof}See Lemma 4.5 of \cite{Li18}.
\end{proof}
\section{Central Limit Theorem for Piecewise Boundary Conditions}\label{pbc}
In this section, we construct certain statistics from the (random) dimer configuration on a contracting square hexagon lattice with piecewise boundary conditions, and show that they converge in distribution to sum of $n$ independent Gaussian random variables in the scaling limit, where $1\times n$ is the size of a fundamental domain. The main theorem proved in this section is Theorem \ref{clt2}.
\subsection{First order moments}
For the piecewise boundary conditions, the proof of Proposition \ref{pn41} still holds. For $\kappa\in(0,1)$, let
\begin{eqnarray*}
X_N&=&(x_{1,N},\ldots,x_{N,N});\\
X_{N,\kappa}&=&(x_{1+N-\lfloor(1-\kappa) N\rfloor,N},\ldots,x_{N,N})\\
U_{N,\kappa}&=&(u_1,u_2,\ldots,u_{\lfloor (1-\kappa)N\rfloor})\\
U_{N,\kappa,X}&=&(u_1x_{1+N-\lfloor(1-\kappa) N\rfloor,N},u_2x_{2+N-\lfloor(1-\kappa) N\rfloor,N},\ldots,u_{\lfloor (1-\kappa)N\rfloor}x_{N,N})
\end{eqnarray*}
Let $\lambda\in\GT_{\lfloor (1-\kappa)N \rfloor}$, and $\rho_{\lfloor(1-\kappa)N \rfloor}$ be a probability measure on $\GT_{\lfloor(1-\kappa)N}\rfloor$ as defined in Proposition \ref{tm1}. Then we have
\begin{eqnarray*}
&&E_{k,\kappa,N}:=\mathbf{E}\sum_{i=1}^{\lfloor (1-\kappa)N\rfloor}(\lambda_i+\lfloor(1-\kappa)N\rfloor-i)^k\\
&=&\sum_{\lambda\in\GT_{\lfloor(1-\kappa) N\rfloor}}\rho_{\lfloor(1-\kappa)N \rfloor}(\lambda)\sum_{i=1}^{\lfloor(1-\kappa)N \rfloor}(\lambda+\lfloor(1-\kappa)N\rfloor-i)^k\\
&=&\left.\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(U_{N,\kappa,X})}\sum_{i=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_i\frac{\partial}{\partial u_i}\right)^kV_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\mathcal{S}_{\rho_{\lfloor(1-\kappa)N \rfloor},X_{N,\kappa}}(U_{N,\kappa,X})\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
\end{eqnarray*}
For $1\leq i\leq n$, let
\begin{eqnarray*}
v_i=\begin{cases}x_{i,N},\ \mathrm{if}\ 1\leq i\leq N-\lfloor (1-\kappa)N\rfloor \\ x_{i,N} u_{i-N+\lfloor (1-\kappa)N\rfloor},\ \mathrm{if}\ N-\lfloor (1-\kappa)N\rfloor+1\leq i\leq N \end{cases}
\end{eqnarray*}
Let $\lambda(N)$ be the partition corresponding to the boundary condition. For $1\leq i\leq n$, let
\begin{eqnarray}
R(i)=\{1\leq j\leq \lfloor(1-\kappa)N \rfloor: [(j+N-\lfloor(1-\kappa)N \rfloor)\mod n]=[i\mod n]\}
\end{eqnarray}
and for $1\leq i\leq \lfloor (1-\kappa)N\rfloor$ let
\begin{eqnarray*}
j(i)=\begin{cases}[i+N-\lfloor (1-\kappa)N \rfloor]\mod n,\ \mathrm{if}\ \left([i+N-\lfloor (1-\kappa)N \rfloor]\mod n\right)\neq 0\\
n,\ \mathrm{if}\ \left([i+N-\lfloor (1-\kappa)N \rfloor]\mod n\right)=0
\end{cases}
\end{eqnarray*}
Assume
\begin{eqnarray*}
\lfloor(1-\kappa)N\rfloor=q_{N,\kappa}n+r_{N,\kappa}
\end{eqnarray*}
where $q_{N,\kappa}$ and $r_{N,\kappa}$ are nonnegative integers satisfying $r_{N,\kappa}<n$.
By Lemmas \ref{lmm212}, \ref{p437}, \ref{l440}, we obtain
\begin{eqnarray*}
&&E_{k,\kappa,N}\\
&=&\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(X_{N,\kappa})}\sum_{i=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_i\frac{\partial}{\partial u_i}\right)^kV_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})
\left[\frac{s_{\lambda(N)}\left(U_{N,\kappa,X},x_{1,N},\ldots, x_{N-\lfloor (1-\kappa)N \rfloor,N}\right)}{s_{\lambda(N)}(X_N)}\right.\\
&&\left.\left.\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\prod_{j=1}^{\lfloor(1-\kappa)N \rfloor}\left(\frac{1+y_{\ol{l}}u_jx_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(X_{N,\kappa})}\sum_{i=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_i\frac{\partial}{\partial u_i}\right)^kV_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\frac{T_N}{P_N}\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
where
\begin{eqnarray*}
&&T_N=\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\prod_{j=1}^{\lfloor(1-\kappa)N \rfloor}\left(\frac{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}u_j}{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}\right)\\
&\times&\left(\prod_{i=1}^{r}s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},u_{n+i}\ldots,u_{q_{N,\kappa}n+i},1,\ldots,1\right)\right)\\
&&\times\left(\prod_{i=r+1}^{n}s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},u_{n+i}\ldots,u_{(q_{N,\kappa}-1)n+i},1,\ldots,1\right)\right)\\
&&\times\left(\prod_{i<j,x_{\sigma_0(i)}\neq x_{\sigma_0(j)}}\frac{1}{v_{\sigma_0(i)}-v_{\sigma_0(j)}}\right)\left(1+o(1)\right)
\end{eqnarray*}
and
\begin{eqnarray*}
P_N&=&\left(\prod_{i=1}^{n}s_{\phi^{(i,\sigma_0)}(N)}\left(1,\ldots,1\right)\right)\\
&&\left(\prod_{i<j,x_{\sigma_0(i)}\neq x_{\sigma_0(j)}}\frac{1}{x_{\sigma_0(i)}-x_{\sigma_0(j)}}\right)\left(1+o(1)\right)
\end{eqnarray*}
Then we have
\begin{eqnarray*}
E_{k,\kappa,N}=\sum_{j=1}^{n}E_{k,\kappa,N}^{(j)}.
\end{eqnarray*}
where
\begin{eqnarray*}
E_{k,\kappa,N}^{(j)}:&=&\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}
\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(X_{N,\kappa})}\\
&&\left.\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\right)^kV_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\frac{T_N}{P_N}\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}\left.\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\right)^k\frac{T_{N,i}}{P_{N,i}}\right|_{U_{N,\kappa}=(1,\ldots,1)};
\end{eqnarray*}
\begin{eqnarray*}
T_{N,i}&=&\left[\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\left(\frac{1+y_{\ol{l}}x_{j(i)}u_i}{1+y_{\ol{l}}x_{j(i)}}\right)\right]s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},1,\ldots,1\right)\\
&&\times\left(\prod_{j(i)<k\leq n }\left(\frac{1}{u_ix_{j(i)}-x_k}\right)^{\frac{\kappa N}{n}}\right)\left(\prod_{k<j(i)\leq n}\left(\frac{1}{x_k-u_ix_{j(i)}}\right)^{\frac{\kappa N}{n}}\right)\\
&&\left(\prod_{N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ k\mod n=j(i),\ k-N+\lfloor(1-\kappa)N \rfloor\neq i}\left[u_ix_{j(i)}-x_{k}\right]\right)e^{N o(1)}
\end{eqnarray*}
and
\begin{eqnarray*}
P_{N,i}&=&s_{\phi^{(j(i),\sigma_0)}(N)}\left(1,\ldots,1\right)\\
&&\times\left(\prod_{j(i)<k\leq n }\left(\frac{1}{x_{j(i)}-x_k}\right)^{\frac{\kappa N}{n}}\right)\left(\prod_{k<j(i)\leq n}\left(\frac{1}{x_k-x_{j(i)}}\right)^{\frac{\kappa N}{n}}\right)\\
&&\left(\prod_{N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ k\mod n=j(i),\ k-N+\lfloor(1-\kappa)N \rfloor\neq i}\left[x_{j(i)}-x_{k}\right]\right)e^{No(1)}.
\end{eqnarray*}
In the expressions above, the $o(1)$ terms converge to 0 uniformly when $u_i$ is in a neighborhood of $1$.
If the edge weights satisfy Assumption \ref{ap32}, we obtain
\begin{eqnarray*}
E_{k,\kappa,N}^{(j)}:=\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}\left.\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\right)^k\frac{\tilde{V}_{N,i}\tilde{T}_{N,i}}{\tilde{W}_{N,i}\tilde{P}_{N,i}}\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
where
\begin{eqnarray*}
\tilde{T}_{N,i}&=&\left[\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\left(\frac{1+y_{\ol{l}}x_{j(i)}u_i}{1+y_{\ol{l}}x_{j(i)}}\right)\right]s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},1,\ldots,1\right)\\
&&\times\left(\prod_{j(i)<k\leq n }\left(\frac{1}{u_ix_{j(i)}-x_k}\right)^{\frac{\kappa N}{n}}\right)e^{No(1)};
\end{eqnarray*}
\begin{eqnarray*}
\tilde{P}_{N,i}&=&s_{\phi^{(j(i),\sigma_0)}(N)}\left(1,\ldots,1\right)\left(\prod_{j(i)<k \leq n}\left(\frac{1}{x_{j(i)}-x_k}\right)^{\frac{\kappa N}{n}}\right)e^{No(1)};
\end{eqnarray*}
\begin{eqnarray*}
\tilde{V}_{N,i}&=&\prod_{[N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ (k\mod n)=(j(i)\mod n)],\ k-N+\lfloor(1-\kappa)N \rfloor\neq i}\left[u_ix_{j(i)}-x_{k}\right]\\
\tilde{W}_{N,i}&=&\prod_{[N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ (k\mod n)=(j(i)\mod n)],\ k-N+\lfloor(1-\kappa)N \rfloor\neq i}\left[x_{j(i)}-x_{k}\right]
\end{eqnarray*}
Note that
\begin{eqnarray}
&&E_{k,\kappa,N}^{(j)}\label{ekj}\\
&=&\frac{1}{\tilde{W}_{N,i}\tilde{P}_{N,i}}\left.\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\right)^k\tilde{V}_{N,i}\exp\left[\log\left(\tilde{T}_{N,i}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\notag
\end{eqnarray}
and
\begin{eqnarray*}
\frac{\partial \tilde{T}_{N,i}}{\partial u_i}=\frac{\partial}{\partial u_i}\exp\left[\log\left(\tilde{T}_{N,i}\right)\right]=\exp\left[\log\left(\tilde{T}_{N,i}\right)\right]\frac{\partial}{\partial u_i}\left[\log\left(\tilde{T}_{N,i}\right)\right]
\end{eqnarray*}
\begin{lemma}\label{l51}Assume $\kappa \in (0,1)$ and the edge weights satisfy Assumption \ref{ap32}, then
\item For $1\leq i\leq \lfloor (1-\kappa)N\rfloor$ and $i\in R(j)$
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N}\frac{\partial [\log\tilde{T}_{N,i}]}{\partial u_i}=\frac{\kappa}{n}\sum_{l\in I_2\cap\{1,2,\ldots,n\}}\frac{y_lx_{j(i)}}{1+y_l x_{j(i)}u_i}-\frac{\kappa}{n}\frac{n-j(i)}{u_i}+\frac{1}{n} H_{\bm_{j(i)}}'(u_i);
\end{eqnarray*}
and the convergence is uniform when $u_i$ is in a neighborhood of $1$.
\end{lemma}
\begin{proof}The lemma follows from explicit computations and Theorem 3.6 of \cite{bk}; see also \cite{bg,GP15,GM05}.
\end{proof}
\subsection{Second order moments}
Let
\begin{eqnarray*}
&&E_{k,\ell,\kappa,N}:=\mathbf{E}\left(\sum_{i=1}^{\lfloor (1-\kappa)N\rfloor}(\lambda_i+\lfloor(1-\kappa)N\rfloor-i)^k\sum_{j=1}^{\lfloor (1-\kappa)N\rfloor}(\lambda_j+\lfloor(1-\kappa)N\rfloor-j)^l\right)\\
&=&\sum_{\lambda\in\GT_{\lfloor(1-\kappa) N\rfloor}}\rho_{\lfloor(1-\kappa)N \rfloor}(\lambda)\sum_{i=1}^{\lfloor(1-\kappa)N \rfloor}(\lambda_i+\lfloor(1-\kappa)N\rfloor-i)^k\sum_{j=1}^{\lfloor (1-\kappa)N\rfloor}(\lambda_j+\lfloor(1-\kappa)N\rfloor-j)^l\\
&=&\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(U_{N,\kappa,X})}\sum_{i=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_i\frac{\partial}{\partial u_i}\right)^k\sum_{j=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_j\frac{\partial}{\partial u_j}\right)^l\\
&&\left.V_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\mathcal{S}_{\rho_{\lfloor(1-\kappa)N \rfloor},X_{N,\kappa}}(U_{N,\kappa,X})\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
\end{eqnarray*}
Again by Lemmas \ref{lmm212}, \ref{p437}, \ref{l440}, we obtain
\begin{eqnarray*}
&&E_{k,\ell,\kappa,N}\\
&=&\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(X_{N,\kappa})}\sum_{i=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_i\frac{\partial}{\partial u_i}\right)^k\sum_{j=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_j\frac{\partial}{\partial u_j}\right)^l V_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\\
&&\left[\frac{s_{\lambda(N)}\left(U_{N,\kappa,X},x_{1,N},\ldots, x_{N-\lfloor (1-\kappa)N \rfloor,N}\right)}{s_{\lambda(N)}(X_N)}\right.\\
&&\left.\left.\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\prod_{j=1}^{\lfloor(1-\kappa)N \rfloor}\left(\frac{1+y_{\ol{l}}u_jx_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(X_{N,\kappa})}\sum_{i=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_i\frac{\partial}{\partial u_i}\right)^k\sum_{j=1}^{\lfloor(1-\kappa)N\rfloor}\left(u_j\frac{\partial}{\partial u_j}\right)^lV_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\frac{T_N}{P_N}\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
Then we have
\begin{eqnarray*}
E_{k,l,\kappa,N}=\sum_{s=1}^{n}\sum_{t=1}^{n}E_{k,l,\kappa,N}^{(s,t)};
\end{eqnarray*}
where
\begin{eqnarray}
E_{k,l,\kappa,N}^{(s,t)}:&=&\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}
\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(X_{N,\kappa})}\label{e2}\\
&&\sum_{[i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(s)]}\sum_{[r\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(t)]}\left(u_i\frac{\partial}{\partial u_i}\right)^k\left(u_r\frac{\partial}{\partial u_r}\right)^l
\notag\\&&\left.V_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\frac{T_N}{P_N}\right|_{U_{N,\kappa}=(1,\ldots,1)}\notag\\
&=&\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(s)}\left(u_i\frac{\partial}{\partial u_i}\right)^k\notag\\
&&\sum_{[r\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(t)]}\left(u_r\frac{\partial}{\partial u_r}\right)^l\left.\frac{V_{N,(i,r)}^{(s,t)}T_{N,(i,r)}^{(s,t)}}{W_{N,(i,r)}^{(s,t)}P_{N,(i,r)}^{(s,t)}}\right|_{U_{N,\kappa}=(1,\ldots,1)}\notag
\end{eqnarray}
and
where (assume the edge weights satisfy Assumption \ref{ap32})
\begin{enumerate}
\item if $s=t$,
\begin{eqnarray*}
T_{N,(i,r)}^{(s,s)}&=&
\left[\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\left(\frac{1+y_{\ol{l}}x_{s}u_i}{1+y_{\ol{l}}x_{s}}\right)\right]
\left[\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\left(\frac{1+y_{\ol{l}}x_{s}u_r}{1+y_{\ol{l}}x_{s}}\right)\right]\\
&&s_{\phi^{(s,\sigma_0)}(N)}\left(u_{i},u_j,1,\ldots,1\right)\\
&&\times\left(\prod_{s<k\leq n }\left(\frac{1}{u_ix_{s}-x_k}\right)^{\frac{\kappa N}{n}}\right)\left(\prod_{s<k\leq n }\left(\frac{1}{u_rx_{s}-x_k}\right)^{\frac{\kappa N}{n}}\right)e^{No(1)};\\
P_{N,(i,r)}^{(s,s)}&=&s_{\phi^{(s,\sigma_0)}(N)}\left(1,\ldots,1\right)\left(\prod_{s<k \leq n}\left(\frac{1}{x_{s}-x_k}\right)^{\frac{2\kappa N}{n}}\right)e^{No(1)}\\
V_{N,(i,r)}^{(s,s)}&=&\left(\prod_{[N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ (k\mod n)=(s\mod n)],\ k-N+\lfloor(1-\kappa)N \rfloor\notin \{i,r\}}\left[u_ix_{s}-x_{k}\right]\right)\\
&&\times \left(\prod_{[N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ (k\mod n)=(s\mod n)],\ k-N+\lfloor(1-\kappa)N \rfloor\notin \{i,r\}}\left[u_rx_{s}-x_{k}\right]\right)\\
&&\times (u_ix_s-u_rx_s)\\
W_{N,(i,r)}^{(s,s)}&=&\prod_{[N- \lfloor(1-\kappa)N\rfloor+1\leq k\leq N,\ (k\mod n)=(s\mod n)],\ k-N+\lfloor(1-\kappa)N \rfloor\notin \{i,r\}}\left[x_{s}-x_{k}\right]^2\\
&&\times \left(x_{i+N-\lfloor(1-\kappa)N\rfloor}-x_{r+N-\lfloor (1-\kappa)N\rfloor}\right)
\end{eqnarray*}
In the expressions above, the $o(1)$ terms converge to 0 uniformly when $u_i$, $u_r$ are in a neighborhood of $1$ as $N\rightarrow\infty$; moreover, the operator $\frac{\partial^2}{\partial u_r\partial u_s}$ acting on $\log o(1)$ is identically 0 (instead of converging to $0$ as $N\rightarrow\infty$).
\item if $s\neq t$,
\begin{eqnarray*}
T_{N,(i,r)}^{(s,t)}&=&
\left[\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\left(\frac{1+y_{\ol{l}}x_{s}u_i}{1+y_{\ol{l}}x_{s}}\right)\right]
\left[\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\left(\frac{1+y_{\ol{l}}x_{s}u_r}{1+y_{\ol{l}}x_{s}}\right)\right]\\
&&s_{\phi^{(s,\sigma_0)}(N)}\left(u_{i},1,\ldots,1\right)s_{\phi^{(t,\sigma_0)}(N)}\left(u_r,1,\ldots,1\right)\\
&&\times\left(\prod_{s<k\leq n}\left(\frac{1}{u_ix_{s}-x_k}\right)^{\frac{\kappa N}{n}}\right)\left(\prod_{t<k\leq n}\left(\frac{1}{u_rx_{t}-x_k}\right)^{\frac{\kappa N}{n}}\right) \left(\frac{1}{u_ix_{s}-u_rx_t}\right)e^{No(1)};\\
P_{N,(i,r)}^{(s,t)}&=&\tilde{P}_{N,i}\tilde{P}_{N,r}e^{No(1)}\\
V_{N,(i,r)}^{(s,t)}&=&\tilde{V}_{N,i}\tilde{V}_{N,r}\\
W_{N,(i,r)}^{(s,t)}&=&\tilde{W}_{N,i}\tilde{W}_{N,r}
\end{eqnarray*}
In the expressions above, the $o(1)$ terms converge to 0 uniformly when $u_i$, $u_r$ are in a neighborhood of $1$ as $N\rightarrow\infty$; moreover, the operator $\frac{\partial^2}{\partial u_r\partial u_s}$ acting on $\log o(1)$ is identically 0 (instead of converging to $0$ as $N\rightarrow\infty$).
\end{enumerate}
\begin{lemma}\label{l52}Assume $\kappa \in (0,1)$ and the edge weights satisfy Assumption \ref{ap32}, then
\begin{enumerate}
\item For $1\leq i<j\leq \lfloor (1-\kappa)N\rfloor$ and $i,j\in R(s)$
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\partial^2 [\log T^{(s,s)}_{N,(i,r)}]}{\partial u_i\partial u_r}\\
&=&\frac{\partial^2}{\partial u_i u_r}\left[\log\left(1-(u_i-1)(u_r-1)\frac{u_i H_{\bm_s}'(u_i)-u_r H'_{\bm_s}(u_r)}{u_i-u_r}\right)\right]
\end{eqnarray*}
and the convergence is uniform when $u_i$ is in a neighborhood of $1$.
\item For $1\leq i<j\leq \lfloor (1-\kappa)N\rfloor$ and $i\in R(s)$, $j\in R(t)$ with $s\neq t$
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\partial^2 [\log T^{(s,t)}_{N,(i,r)}]}{\partial u_i\partial u_s}=0.
\end{eqnarray*}
and the convergence is uniform when $u_i$ is in a neighborhood of $1$.
\end{enumerate}
\end{lemma}
\begin{proof}First we consider the case that $i,j\in R(s)$, we have
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\partial^2 [\log T^{(s,s)}_{N,(i,r)}]}{\partial u_i\partial u_r}
=\lim_{N\rightarrow\infty}\frac{\partial^2 [\log s_{\phi^{(s,\sigma_0)}(N)}\left(u_{i},u_j,1,\ldots,1\right)]}{\partial u_i\partial u_r}
\end{eqnarray*}
Then Part (1) of the lemma follows from Theorem 6.8 of \cite{bk}; see also \cite{GP15,bg16}.
Now we consider the case that $i\in R(s)$, $j\in R(t)$ and $s\neq t$. In this case
\begin{eqnarray*}
\lim_{N\rightarrow\infty}\frac{\partial^2 [\log T^{(s,s)}_{N,(i,r)}]}{\partial u_i\partial u_r}
&=&\lim_{N\rightarrow\infty}\frac{\partial^2 [-\log(u_ix_{s,N}-u_rx_{t,N})]}{\partial u_i\partial u_r}\\
&=&\lim_{N\rightarrow\infty}\frac{x_{s,N}x_{t,N}}{(u_ix_{s,N}-u_rx_{t,N})^2};
\end{eqnarray*}
and the limit is 0 by Assumption \ref{ap32}.
\end{proof}
\subsection{Asymptotical analysis}
Let
\begin{eqnarray}
F_{k,\kappa,N}^{(s)}&=&\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}
\frac{1}{V_{\lfloor (1-\kappa)N\rfloor}(U_{N,\kappa,X})T_N}\label{fkn}\\
&&\sum_{[i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(s)]}\left(u_i\frac{\partial}{\partial u_i}\right)^kV_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})T_N\notag
\end{eqnarray}
For simplicity, we use the notation $\partial_i$ to denote $\frac{\partial}{\partial u_i}$.
Expanding the right hand side of (\ref{ekj}), we can write $E_{k,\kappa,N}^{(j)}$ as a sum of terms with the following form
\begin{eqnarray}
\frac{c_0 x_{j(i)}^r (\partial_i^{s_1}[\log\tilde{T}_{N,i}])^{d_1}\ldots (\partial_i^{s_t}[\log\tilde{T}_{N,i}])^{d_t}}{(x_{j(i)}-x_{a_1})\ldots(x_{j(i)}-x_{a_r})}\label{rt}
\end{eqnarray}
Similarly, we can write the right hand side of (\ref{fkn}) as a large sum of the following form
\begin{eqnarray*}
\frac{c_0 x_{j(i)}^r u_i^{k-s_0}(\partial_i^{s_1}[\log T_N])^{d_1}\ldots (\partial_i^{s_t}[\log T_N])^{d_t}}{(x_{j(i)}u_i-x_{a_1}u_{a_1-N+\lfloor (1-\kappa)N\rfloor})\ldots(x_{j(i)}u_i-x_{a_r}u_{a_r-N+\lfloor (1-\kappa)N\rfloor})}
\end{eqnarray*}
such that
\begin{itemize}
\item for $1\leq s\leq r$, $a_s$ is a positive integer satisfying $N-\lfloor (1-\kappa)N\rfloor+1\leq a_s\leq N$; and
\item In (\ref{rt}), we have $(a_s\mod n)=(j(i)\mod n)$ for all $1\leq s\leq r$; and
\item $N-\lfloor (1-\kappa)N\rfloor+i, a_1,\ldots,a_r$ are distinct; and
\item $\{s_j\}_{j=0}^{t}$ and $\{d_j\}_{j=1}^{t}$ are nonnegative integers satisfying $s_1<s_2<\ldots <s_t$; and
\item
\begin{eqnarray}
r+s_0+s_1d_1+\ldots+s_td_t=k; \label{rsd}
\end{eqnarray}
and
\item $c_0$ is a constant independent of $N$ and $a_1,\ldots,a_r$.
\end{itemize}
From the expression (\ref{ekj}) we see that any term obtained by permuting $i,a_1-N+\lfloor(1-\kappa)N\rfloor,\ldots,a_r-N+\lfloor (1-\kappa)N\rfloor$ of (\ref{rt}) with in $\{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)$ are still present in the sum. Let
\begin{eqnarray*}
\tilde{a}_1=a_1-N+\lfloor (1-\kappa)N\rfloor
\end{eqnarray*}
Hence we have
\begin{eqnarray*}
E_{k,\kappa,N}^{(j)}&=&\sum_{r,\{s_j\}_{j=0}^{t},\{d_j\}_{j=1}^{t}\ \mathrm{satisfy}\ (\ref{rsd})}(r+1)!
\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\\
&&\left.\left[\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}(\partial_{\tilde{a}_1}^{s_1}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_1}\ldots (\partial_{\tilde{a}_1}^{s_t}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_t}}{(x_{a_1}-x_{a_2})\ldots(x_{a_1}-x_{a_{r+1}})}\right]\right|_{U_{N,\kappa}=\{1,\ldots,1\}}.
\end{eqnarray*}
where the constant $c_0$ may depend on $r$, $\{s_j\}_{j=0}^t$ and $\{d_j\}_{j=1}^{t}$.
\begin{lemma}\label{le53} Let
\begin{eqnarray*}
&&\tilde{T}_N=\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa)N \rfloor\}\cap I_2}\prod_{j=1}^{\lfloor(1-\kappa)N \rfloor}\left(\frac{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}u_j}{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}\right)\\
&\times&\left(\prod_{i=1}^{r}s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},u_{n+i}\ldots,u_{q_{N,\kappa}n+i},1,\ldots,1\right)\right)\\
&&\times\left(\prod_{i=r+1}^{n}s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},u_{n+i}\ldots,u_{(q_{N,\kappa}-1)n+i},1,\ldots,1\right)\right)\\
&&\times\left(\prod_{N-\lfloor(1-\kappa)N\rfloor+1\leq i\leq N,\ 1\leq j\leq N-\lfloor(1-\kappa)N\rfloor,\ \tilde{i}\in R(p),\tilde{j}\in R(q),p<q}\frac{1}{x_iu_{\tilde{i}}-x_j}\right)\left(1+o(1)\right)
\end{eqnarray*}
and
\begin{eqnarray*}
\tilde{V}_{\lfloor (1-\kappa)N\rfloor}=\prod_{k=1}^{n}\prod_{[1+N-\lfloor(1-\kappa)N \rfloor\leq i<j\leq N, (i\mod n)=(j\mod n)=(k\mod n)]}(x_iu_{\tilde{i}}-x_ju_{\tilde{j}})
\end{eqnarray*}
Assume $a,b,c\in R(j)$, and $a,b,c$ are distinct positive integers. Assume that the $o(1)$ in the definition of $\tilde{T}_N$ converges to $0$ uniformly when $u_a,u_b,u_c$ are in a neighborhood of 1. Then
\begin{eqnarray*}
\left.\lim_{N\rightarrow\infty}\frac{\partial^3\log[\tilde{T}_N]}{\partial u_a\partial u_b\partial u_c}\right|_{U_{N,\kappa}=(1,\ldots,1)}=0.
\end{eqnarray*}
\end{lemma}
\begin{proof}We have
\begin{eqnarray*}
\left.\lim_{N\rightarrow\infty}\frac{\partial^3\log[\tilde{T}_N]}{\partial u_a\partial u_b\partial u_c}\right|_{U_{N,\kappa}=(1,\ldots,1)}=\lim_{N\rightarrow\infty}\frac{\partial^3\log[s_{\phi^{(j,\sigma_0)}(N)}\left(u_a,u_b,u_c,1,\ldots,1\right)]}{\partial u_a\partial u_b\partial u_c}
\end{eqnarray*}
To compute the derivative on the right hand side, note that
\begin{eqnarray*}
\lim_{N\rightarrow \infty}\bm[\phi^{(j,\sigma_0)}(N)]=\bm_j
\end{eqnarray*}
By Theorem 6.8 of \cite{bk} (see also \cite{GP15}, \cite{bg16})，
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\partial^2\log[s_{\phi^{(j,\sigma_0)}(N)}\left(u_a,u_b,u_c,1,\ldots,1\right)]}{\partial u_a\partial u_b}\\
&=&\frac{\partial^2}{\partial u_a\partial u_b}\log\left(1-(u_a-1)(u_b-1)\frac{u_a H_{\bm_j}'(u_a)-u_bH_{\bm_j}'(u_b)}{u_a-u_b}\right);
\end{eqnarray*}
and the convergence is uniform when $(u_a,u_b,u_c)$ is in a neighborhood of $(1,1,1)$. Therefore we can take the derivative with respect to $u_c$ on both sides. The right hand side is independent of $u_c$, and the derivative with respect to $u_c$ is 0. Then the lemma follows.
\end{proof}
\begin{lemma}\label{l53}Let $\kappa\in (0,1)$ and the edge weights satisfy Assumption \ref{ap32}. Then
\begin{enumerate}
\item the degree of $N$ in $E_{k,\kappa,N}^{(j)}$ is at most $k+1$.
\item For any integer $i$ satisfying $1\leq i\leq \lfloor (1-\kappa)N\rfloor$ and $i\in R(j)$, the degree of $N$ in $\left.\frac{\partial}{\partial u_i} F_{k,\kappa,N}^{(j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is at most $k$;
\item For any integer $i$ satisfying $1\leq i\leq \lfloor (1-\kappa)N\rfloor$ and $i\notin R(j)$, the degree of $N$ in $\left.\frac{\partial}{\partial u_i} F_{k,\kappa,N}^{(j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is less than $k$;
\item For any integers $i_1,i_2$ satisfying $1\leq i_1<i_2\leq \lfloor (1-\kappa)N\rfloor$ and $i_1,i_2\in R(j)$, the degree of $N$ in $\left.\frac{\partial^2}{\partial u_{i_1}\partial u_{i_2}} F_{k,\kappa,N}^{(j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is less than $k$.
\end{enumerate}
\end{lemma}
\begin{proof}We first consider the asymptotics of
\begin{eqnarray}
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\label{sls}\\
&& \left.\left[\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}(\partial_{\tilde{a}_1}^{s_1}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_1}\ldots (\partial_{\tilde{a}_1}^{s_t}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_t}}{(x_{a_1}-x_{a_2})\ldots(x_{a_1}-x_{a_{r+1}})}\right]\right|_{U_{N,\kappa}=\{1,\ldots,1\}}\notag
\end{eqnarray}
By Lemma \ref{l51}, the degree of $N$ in each factor $(\partial_{\tilde{a}_1}^{s_l}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_1}$ is at most $d_1$. By Lemma \ref{l52} and identity (\ref{rsd}), the degree of $N$ in (\ref{sls}) is at most $k-r$. Summing over the permutations, we obtain that
\begin{eqnarray*}
&&\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}(\partial_{\tilde{a}_1}^{s_1}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_1}\ldots (\partial_{\tilde{a}_1}^{s_t}[\log\tilde{T}_{N,\tilde{a}_1}])^{d_t}}{(x_{a_1}-x_{a_2})\ldots(x_{a_1}-x_{a_{r+1}})}
\end{eqnarray*}
is the sum of $O(N^{r+1})$ terms, the degree of $N$ in each of which is at most $k-r$. Therefore, the degree of $N$ in $E_{k,\kappa,N}^{(j)}$ is at most $k+1$, and we complete the proof of Part (1).
Now we prove Parts (2) and (3).
We consider the following two cases.
\begin{itemize}
\item Consider
\begin{eqnarray*}
\mathcal{D}_i:&=&\left.\frac{\partial F_{k,\kappa,N}^{(j)}}{\partial u_i}\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\frac{\partial}{\partial u_i}\left[\sum_{[r\geq 0,\{s_i\geq0\}_{i=0}^{t},\{d_i\geq 0\}_{i=1}^{t}:s_0+s_1d_1+\ldots s_tdt+r=k]}\right.\\&&\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_w}\longrightarrow x_{j}],\ 1\leq w\leq r+1}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\\
&&\left.\left.\left(\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}(\partial_{\tilde{a}_1}^{s_1}[\log T_{N,(\tilde{a}_1,i)}^{(j,r)}])^{d_1}\ldots (\partial_{\tilde{a}_1}^{s_t}[\log T_{N,(\tilde{a}_1,i)}^{(j,r)}])^{d_t}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}})}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
with
\begin{eqnarray*}
i\notin \{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}
\end{eqnarray*}
When $\kappa\in(0,1)$, there are $O(N^{r+1})$ such terms. Notice that if $i\in R(r)$
\begin{eqnarray*}
&&\left.\frac{\partial}{\partial u_i}\left[\frac{\partial^{s_q}(\log T_{N,(\tilde{a}_1,i)}^{(j,r)})}{\partial \tilde{a}_1^{s_q}}\right]^{d_q}\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.d_q\left[\frac{\partial^{s_q}(\log T_{N,(\tilde{a}_1,i)}^{(j,r)}}{\partial \tilde{a}_1^{s_q}}\right]^{d_q-1}\frac{\partial}{\partial u_i}\left[\frac{\partial^{s_q}(\log T_{N,(\tilde{a}_1,i)}^{(j,r)})}{\partial \tilde{a}_1^{s_q}}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.d_q\left[\frac{\partial^{s_q}\left[\log \tilde{T}_{N,\tilde{a}_1}\right]}{\partial \tilde{a}_1^{s_q}}\right]^{d_q-1}\frac{\partial}{\partial u_i}\left[\frac{\partial^{s_q}\left[\log T_{N,(\tilde{a}_1,i)}^{(j,r)}\right]}{\partial \tilde{a}_1^{s_q}}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
By Lemmas \ref{l51} and \ref{l52}, the degree of $N$ in the expressions above is at most $d_q-1$ if $r=j$, and is strictly less than $d_q-1$ when $r\neq j$. Note that $\mathcal{D}_i$ can be written as a sum of terms of the form
\begin{eqnarray*}
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\\
&&\left.\left[\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}\left(\partial_i\left[\left(\partial_{\tilde{a}_1}^{s_1}\left[\log T_{N,(\tilde{a}_1,i)}^{(j,r)}\right]\right)^{d_1}\right]\right)\ldots \left(\partial_{\tilde{a}_1}^{s_t}\left[\log \tilde{T}_{N,\tilde{a}_1}\right]\right)^{d_t}}{(x_{a_1}-x_{a_2})\ldots(x_{a_1}-x_{a_{r+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
By Lemmas \ref{l51} and \ref{l52}, the degree of $N$ in the expressions above is at most $d_1+\ldots+d_t-1$ when $r=j$; and is less than $d_1+\ldots+d_t-1$ when $r\neq j$. Since $r+s_0+s_1d_1+\ldots+s_td_t=k$, we have
\begin{eqnarray*}
d_1+\ldots+d_t-1\leq k-r-1;
\end{eqnarray*}
and in the sum over $\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}$, it is the sum of $O(N^{r+1})$ of such terms, we obtain that the degree of $N$ in this sum is at most $k$ when $r=j$; and is less than $k$ when $r\neq j$. This completes the proof of Part (3).
\item Consider the case when
\begin{eqnarray*}
i\in \{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}
\end{eqnarray*}
Note that there are $O(N^r)$ such terms in total, since $i$ is fixed. By Lemmas \ref{l51} and \ref{l52}, the degree of $N$ in
\begin{eqnarray*}
\left.\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}\left(\partial_i\left[\left(\partial_{\tilde{a}_1}^{s_1}\left[\log T_{N,(\tilde{a}_1,i)}^{(j,r)}\right]\right)^{d_1}\right]\right)\ldots \left(\partial_{\tilde{a}_1}^{s_t}\left[\log \tilde{T}_{N,\tilde{a}_1}\right]\right)^{d_t}}{(x_{a_1}-x_{a_2})\ldots(x_{a_1}-x_{a_{r+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
is at most $l-r$. This completes the proof of Part (2).
\end{itemize}
Now we prove Part (4).
Note that
\begin{eqnarray*}
&&\left.\frac{\partial^2 F_{k,\kappa,N}^{(j)}}{\partial u_{i_1}\partial u_{i_2}}\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\frac{\partial^2 }{\partial u_{i_1}\partial u_{i_2}}\left[\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}
\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}(U_{N,\kappa,X})\tilde{T}_N}\right.\\
&&\left.\left.\sum_{[i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)]}\left(u_i\frac{\partial}{\partial u_i}\right)^k\tilde{V}_{\lfloor(1-\kappa)N \rfloor}(U_{N,\kappa,X})\tilde{T}_N\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\frac{\partial^2}{\partial u_{i_1}\partial u_{i_2}}\left[\sum_{[r\geq 0,\{s_i\geq0\}_{i=0}^{t},\{d_i\geq 0\}_{i=1}^{t}:s_0+s_1d_1+\ldots s_td_t+r=k]}\right.\\
&&\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\left.\left.\lim_{[x_{a_w}\longrightarrow x_{j}],\ 1\leq w\leq r+1}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\frac{c_0 x_{a_1}^r u_{\tilde{a}_1}^{k-s_0}(\partial_{\tilde{a}_1}^{s_1}[\log \tilde{T}_{N}])^{d_1}\ldots (\partial_{\tilde{a}_1}^{s_t}[\log \tilde{T}_{N}])^{d_t}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
The following cases might occur
\begin{itemize}
\item $\{i_1,i_2\}\cap\{\tilde{a}_1,\ldots,\tilde{a}_{r+1}\}=\emptyset$, and both $\partial_{i_1}$ and $\partial_{i_2}$ are applied to the same $\log\tilde{T}_N$. We have
\begin{eqnarray*}
&&\frac{\partial^2}{\partial u_{i_1}\partial u_{i_2}}\left[\frac{\partial^{s_w}[\log \tilde{T}_N]}{[\partial u_{\tilde{a}_1}]^{s_w}}\right]^{d_w}\\
&=&d_w(d_w-1)\left[\frac{\partial^{s_w}[\log \tilde{T}_N]}{\partial u_{\tilde{a}_1}^{s_w}}\right]^{d_w-2}\frac{\partial^{s_w+1}[\log \tilde{T}_N]}{\partial u_{i_1}\partial u_{\tilde{a}_1}^{s_w}}\frac{\partial^{s_w+1}[\log \tilde{T}_N]}{\partial u_{i_2}\partial u_{\tilde{a}_1}^{s_w}}\\
&&+d_w \left[\frac{\partial^{s_w}[\log \tilde{T}_N]}{\partial u_{\tilde{a}_1}^{s_w}}\right]^{d_w-1}\frac{\partial^{s_w+2}[\log \tilde{T}_N]}{\partial u_{i_1}\partial u_{i_2}\partial u_{\tilde{a}_1}^{s_w}}
\end{eqnarray*}
By Lemma \ref{le53}, the degree of $N$ in the expression above is less than $d_{w}-1$. Taking into account all the other factors, as well as the sum of $O(N^{r+1})$ terms, in this case the degree of $N$ in $\left.\frac{\partial^2 F_{k,\kappa,N}^{(j)}}{\partial u_{i_1}\partial u_{i_2}}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is less than
\begin{eqnarray*}
d_1+d_2+\ldots+d_t-1+r+1\leq k.
\end{eqnarray*}
\item $\{i_1,i_2\}\cap\{\tilde{a}_1,\ldots,\tilde{a}_{r+1}\}=\emptyset$, and $\partial_{i_1}$ and $\partial_{i_2}$ are applied to different $\log\tilde{T}_N$. In this case, the degree of $N$ is at most
\begin{eqnarray*}
d_1+d_2+\ldots+d_t-2+r+1\leq k-1.
\end{eqnarray*}
\item $i_1\in\{a_1,\ldots,a_{r+1}\}$ and $i_2\notin\{a_1,\ldots,a_{r+1}\}$. In this case, we take the sum over $O(N^r)$ terms, since one element in $\{a_1,\ldots,a_{r+1}\}$ is fixed to be $i_1$. Then the degree of $N$ in $\left.\frac{\partial^2 F_{k,\kappa,N}^{(j)}}{\partial u_{i_1}\partial u_{i_2}}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is at most
\begin{eqnarray*}
d_1+d_2+\ldots+d_t-1+r\leq k-1.
\end{eqnarray*}
\item $\{i_1,i_2\}\subset \{a_1,\ldots,a_{r+1}\}$. In this case, we take the sum over $O(N^{r-1})$ terms, since two elements in $\{a_1,\ldots,a_{r+1}\}$ are fixed to be $i_1$ and $i_2$. Then the degree of $N$ in $\left.\frac{\partial^2 F_{k,\kappa,N}^{(j)}}{\partial u_{i_1}\partial u_{i_2}}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is at most
\begin{eqnarray*}
d_1+d_2+\ldots+d_t+r-1\leq k-1.
\end{eqnarray*}
\end{itemize}
\end{proof}
\subsection{Covariance}
Let
\begin{eqnarray*}
\tilde{G}_{\kappa,N,(k,l)}^{(j,s)}&=&k\sum_{r=0}^{k-1}{{k-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\\
&&\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{k-s_0} \partial_{\tilde{a}_1}[F_{l,\kappa,N}^{(s)}](\partial_{\tilde{a}_1}[\log \tilde{T}_{N}])^{k-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]
\end{eqnarray*}
\begin{lemma}\label{l55}Let $l,k$ be arbitrary positive integers. Then
\begin{eqnarray*}
E_{k,l,\kappa,N}^{(j,s)}:=E_{k,\kappa,N}^{(j)}E_{l,\kappa,N}^{(s)}+\left.\tilde{G}_{\kappa,N,(l,k)}^{(j,s)}\right|_{U_{N,\kappa}=(1,\ldots,1)}+R
\end{eqnarray*}
\end{lemma}
where
\begin{itemize}
\item if $j=s$, the degree of $N$ in $\left.\tilde{G}_{\kappa,N,(l,k)}^{(j,s)}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is at most $l+k$;
\item if $j\neq s$, the degree of $N$ in $\left.\tilde{G}_{\kappa,N,(l,k)}^{(j,s)}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is less than $l+k$;
\item the degree of $N$ in $R$ is less than $l+k$.
\end{itemize}
\begin{proof}By (\ref{e2}) and (\ref{fkn}), we obtain
\begin{eqnarray*}
E_{k,l,\kappa,N}^{(j,s)}&=&\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\right)^k\\
&&\sum_{[r\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(s)]}\left(u_r\frac{\partial}{\partial u_r}\right)^l\left.\frac{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}{W_{N,(i,r)}^{(j,s)}P_{N,(i,r)}^{(j,s)}}\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\frac{1}{{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}}\lim_{[x_k\rightarrow (x_{k\mod n}),\mathrm{for}\ 1\leq k\leq N]}\sum_{i\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\right)^k{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}\\
&&\frac{1}{{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}}\sum_{[r\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(s)]}\left(u_r\frac{\partial}{\partial u_r}\right)^l\left.{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}\right|_{U_{N,\kappa}=(1,\ldots,1)}.
\end{eqnarray*}
Note that for any integer $w$, we have
\begin{eqnarray*}
&&\left.\frac{\partial^w}{\partial u_i^w}\left[\frac{1}{{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}}\sum_{[r\in\{1,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(s)]}\left(u_r\frac{\partial}{\partial u_r}\right)^l{V_{N,(i,r)}^{(j,s)}T_{N,(i,r)}^{(j,s)}}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\frac{\partial^w}{\partial u_i^w} F_{l,\kappa,N}^{(s)}\right|_{U_{N,\kappa}=(1,\ldots,1)}.\\
\end{eqnarray*}
Then $E_{k,l,\kappa,N}^{(j,s)}$ can be written as a sum of following terms
\begin{eqnarray*}
&&\left.\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{c_0x_{a_1}^ru_{\tilde{a}_1}^{k-s_0}\partial_{\tilde{a}_1}^{s_1}[F_{l,\kappa,N}^{(s)}][\partial_{\tilde{a}_1}^{s_2}(\log T_{N,(\tilde{a}_1,r)}^{(j,s)})]^{d_2}\ldots [\partial_{\tilde{a}_1}^{s_t}(\log T_{N,(\tilde{a}_1,r)}^{(j,s)})]^{d_t}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots (x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{c_0x_{a_1}^ru_{\tilde{a}_1}^{k-s_0}\partial_{\tilde{a}_1}^{s_1}[F_{l,\kappa,N}^{(s)}][\partial_{\tilde{a}_1}^{s_2}(\log \tilde{T}_N]^{d_2}\ldots [\partial_{\tilde{a}_1}^{s_t}(\log \tilde{T}_N)^{(j,s)})]^{d_t}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots (x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}.
\end{eqnarray*}
such that
\begin{itemize}
\item $r$, $s_0,s_1,\ldots,s_t$, $d_2,\ldots,d_t$ are nonnegative integers; and
\item $s_2<s_3<\ldots<s_t$; and
\item
\begin{eqnarray}
s_0+s_1+s_2d_2+\ldots+s_td_t+r=k; \label{sdk}
\end{eqnarray}
and
\item $\{a_1,\ldots,a_{r+1}\}\subset\{N-\lfloor(1-\kappa)N \rfloor+1,N-\lfloor(1-\kappa)N \rfloor+2,\ldots,N\}\cap R(j)$
\end{itemize}
When $s_1=0$, we obtain $\left.F_{l,\kappa,N}^{(s)}F_{k,\kappa,N}^{(j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}$.
Now we consider the terms corresponding to $s_1\geq 1$. By Lemma \ref{l53}, the degree of $N$ in $\partial_{\tilde{a}_1}^{s_1}[F_{l,\kappa,N}^{(s)}]$ is at most $l$ when $j=s$; and is less than $l$ when $j\neq s$. Therefore, the total degree of $N$ in these terms is at most $l+d_2+\ldots+d_t+r+1$ when $j=s$; and is less than $l+d_2+\ldots+d_t+r+1$ when $j\neq s$. By (\ref{sdk}) and the assumption that $s_1\geq 1$, we have
\begin{eqnarray*}
l+d_2+\ldots+d_t+r+1\leq l+k；
\end{eqnarray*}
and the equality holds when $s_0=d_3=\ldots=d_t=0$, $s_1=s_2=1$; $d_2=k-1-r$; this corresponds to $G_{\kappa,N,(k,l)}^{(j,s)}$, in which the degree of $N$ is $l+k$ when $j=s$, and the degree of $N$ is less than $l+k$ when $j\neq s$. The degree of $N$ is less than $l+k$ in all the other terms. This completes the proof.
\end{proof}
\begin{lemma}\label{l56}
\begin{eqnarray*}
&&\left.\frac{1}{\tilde{V}_{\lfloor(1-\kappa)N\rfloor} \tilde{T}_N }k\sum_{i\in\{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\cap R(j)}\left(u_i\frac{\partial}{\partial u_i}\left[F_{l,\kappa,N}^{(s)}\right]\right)\left(u_i\frac{\partial}{\partial u_i}\right)^{k-1}\left[\tilde{V}_{\lfloor(1-\kappa)N\rfloor} \tilde{T}_N \right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\tilde{G}_{\kappa,N,(k,l)}^{(j,s)}\right|_{U_{N,\kappa}=(1,\ldots,1)}+R
\end{eqnarray*}
where the degree of $N$ in $R$ is less than $l+k$.
\end{lemma}
\begin{proof}This follows from the proof of Lemma \ref{l55}.
\end{proof}
\begin{lemma}\label{l67}Let $i\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}$. Then the degree of $N$ in
\begin{eqnarray*}
\left.\frac{\partial}{\partial u_i}\tilde{G}_{\kappa,N,(k,l)}^{(j,s)}\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
is less than $k+l$.
\end{lemma}
\begin{proof}Note that $\tilde{G}_{\kappa,N,(k,l)}^{(j,s)}$ is the sum of terms
\begin{eqnarray*}
\mathrm{Sym}_{a_1,\ldots,a_{r+1}}
\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{k-s_0} \partial_{\tilde{a}_1}[F_{l,\kappa,N}^{(s)}](\partial_{\tilde{a}_1}[\log \tilde{T}_{N}])^{k-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]
\end{eqnarray*}
If we take derivatives $\frac{\partial}{\partial u_i}$, the following cases might occur
\begin{enumerate}
\item $i\in \{a_1,\ldots,a_{r+1}\}$. Since one element in $\{a_1,\ldots,a_{r+1}\}$ is fixed to be $i$, we take the sum over $O(N^r)$ terms. By Lemma \ref{le53}, and Lemma \ref{l53}, the degree of $N$ is at most
\begin{eqnarray*}
l+(k-1-r)+r=l+k-1
\end{eqnarray*}
\item $i\notin\{a_1,\ldots,a_{r+1}\}$. In this case, we take the sum over $O(N^r)$ terms. Again by Lemmas \ref{le53} and \ref{l53}, the degree of $N$ is less than
\begin{eqnarray*}
l+k-1-r+r+1=l+k.
\end{eqnarray*}
Then the lemma follows.
\end{enumerate}
\end{proof}
\subsection{Products of Moments}
Recall that $\mathcal{P}_{w_1,\ldots,w_p}^{s}$ is the set of all pairings of the set $\{1,2,\ldots,s\}\setminus \{w_1,\ldots,w_p\}$.
We have the following lemma concerning the products of moments.
\begin{lemma}\label{l58}Let $s,l_1,\ldots,l_s$ be positive integers, and let $j_1,\ldots,j_s\in\{1,2,\ldots,n\}$. Then
\begin{eqnarray*}
&&\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}\sum_{[i_2\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_2)]}\left(u_{i_2}\frac{\partial}{\partial u_{i_2}}\right)^{l_2}\cdots\\&&\left.\sum_{[i_s\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_s)]}\left(u_{i_s}\frac{\partial}{\partial u_{i_s}}\right)^{l_s}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\sum_{p=0}^{s}\sum_{[w_1,\ldots,w_p\in\{1,2,\ldots,s\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_s},\kappa,N}^{(j_{w_p})}\left(\sum_{P\in \mathcal{P}_{w_1,\ldots,w_p}^{s}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R\right)\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
where the degree of $N$ in $R$ is less than $\sum_{i=1}^{s}l_i-\sum_{i=1}^{p}l_{w_i}$.
\end{lemma}
\begin{proof} The lemma can be proved by induction on $s$, similar to the proof of proposition 5.10 in \cite{bg16}. We shall now sketch the proof. When $s=1$ the lemma follows from the definition of $F_{l,\kappa,N}^{(j)}$. When $s=2$, the lemma follows from Lemma \ref{l55}. Assume that the lemma holds for $s=t-1$, where $t\geq 2$ is a positive integer. When $s=t$, by induction hypothesis, we have
\begin{eqnarray*}
&&\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}\sum_{[i_2\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_2)]}\left(u_{i_2}\frac{\partial}{\partial u_{i_2}}\right)^{l_2}\cdots\\&&\left.\sum_{[i_t\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_t)]}\left(u_{i_t}\frac{\partial}{\partial u_{i_t}}\right)^{l_t}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]\\
&&\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_p},\kappa,N}^{(j_{w_p})}\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{t}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{U_{N,\kappa}=(1,\ldots,1)}.\\
&=& S_1+S_2+S_3;
\end{eqnarray*}
where by induction hypothesis the degree of $N$ in $R_{1,w_1,\ldots,w_p}$ is less than $\sum_{i=2}^{s}l_i-\sum_{i=1}^{p}l_{w_i}$.
\begin{eqnarray*}
S_1&=&\left\{\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]\right\}\\
&&\times\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_p},\kappa,N}^{(j_{w_p})}\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{t}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{U_{N,\kappa}=(1,\ldots,1)}.
\end{eqnarray*}
and
\begin{eqnarray*}
S_2&=&\frac{\ell_1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left\{\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1-1}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]\right\}\\
&&\times\left\{\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_p},\kappa,N}^{(j_{w_p})}\right.\right.\\
&&\left.\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{t}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{1,w_1,\ldots,w_p}\right)\right)\right\}\right|_{U_{N,\kappa}=(1,\ldots,1)}.
\end{eqnarray*}
Indeed, $S_1$ corresponds to the terms where all the differentiations $\frac{\partial}{\partial u_{i_1}}$ are applied to $\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N$ or $u_{i_1}$; $S_2$ corresponds to the terms where all the differentiations $\frac{\partial}{\partial u_{i_1}}$ except one are applied to $\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N$ or $u_{i_1}$, and $S_3$ are all the other terms.
By the definition of $F_{l,\kappa,N}^{(j)}$ we have
\begin{eqnarray*}
S_1&=&F_{l_1,\kappa,N}^{(j_1)}\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_s},\kappa,N}^{(j_{w_s})}\right.\\
&&\left.\left.\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{s}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{1,w_1,\ldots,w_p}\right)\right)\right|_{U_{N,\kappa}=(1,\ldots,1)}.
\end{eqnarray*}
By Lemma \ref{l56}, we have
\begin{eqnarray*}
S_2&=&\left\{\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}\sum_{x=1}^{p}F_{l_{w_1},\kappa,N}^{(j_{w_1})}\ldots F_{l_{w_{x-1}},\kappa,N}^{(j_{w_{x-1}})} F_{l_{w_{x+1}},\kappa,N}^{(j_{w_{x+1}})}\ldots F_{l_{w_s},\kappa,N}^{(j_{w_s})}\right.\right.\\
&&\left.\left.\left.\left[\left(\tilde{G}_{\kappa,N,(l_1,l_x)}^{(j_1,j_x)}+R_{1,x}\right)\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{s}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{1,w_1,\ldots,w_p}\right)\right)\right]\right\}\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
where the degree of $N$ in $R_{1,x}$ is less than $l_1+l_x$.
Hence we have
\begin{eqnarray*}
S_2&=&\left\{\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}\sum_{x=1}^{p}F_{l_{w_1},\kappa,N}^{(j_{w_1})}\ldots F_{l_{w_{x-1}},\kappa,N}^{(j_{w_{x-1}})} F_{l_{w_{x+1}},\kappa,N}^{(j_{w_{x+1}})}\ldots F_{l_{w_s},\kappa,N}^{(j_{w_s})}\right.\right.\\
&&\left.\left.\left.\left[\tilde{G}_{\kappa,N,(l_1,l_x)}^{(j_1,j_x)}\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{t}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{w_1,\ldots,\hat{w}_x,\ldots,w_p}\right)\right]\right\}\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
where the degree of $N$ in $R_{w_1,\ldots,\hat{w}_x,\ldots,w_p}$ is less than $\sum_{i=2}^{t}l_i-\sum_{i=1}^{p}l_{w_i}$.
Note that
\begin{eqnarray*}
S_1+S_2&=&\sum_{p=0}^{s}\sum_{[w_1,\ldots,w_p\in\{1,2,\ldots,s\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_t},\kappa,N}^{(j_{w_t})}\\
&&\left.\left(\sum_{P\in \mathcal{P}_{w_1,\ldots,w_p}^{t}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{w_1,\ldots,w_p}\right)\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
where the degree of $N$ in $R_{w_1,\ldots,w_p}$ is less than $\sum_{i=1}^{t}l_i-\sum_{i=1}^{p}l_{w_i}$.
It remains to show that $S_3$ does not contribute to the leading terms. Define
\begin{eqnarray*}
\tilde{H}_{j_1,\ldots,j_p}=\left(\sum_{P\in \mathcal{P}_{1,w_1,\ldots,w_p}^{t}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R_{1,w_1,\ldots,w_p}\right);
\end{eqnarray*}
By Lemma \ref{l55}, the degree of $N$ in $\left.\tilde{H}_{j_1,\ldots,j_p}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is at most $\sum_{i=2}^t l_i-\sum_{j=1}^p l_{w_j}$. Moreover, by Lemma \ref{l67}, for any index $i$, the degree of $N$ in $\left.\frac{\partial}{\partial u_i}\tilde{H}_{j_1,\ldots,j_p}\right|_{U_{N,\kappa}=(1,\ldots,1)}$ is less than $\sum_{i=2}^t l_i-\sum_{j=1}^p l_{w_j}$.
We write
\begin{eqnarray*}
&&\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]\\
&&\left.\left(\sum_{p=0}^{t-1}\sum_{[w_1,\ldots,w_p\in\{2,\ldots,t\}]}F_{l_{w_1},\kappa,N}^{(j_{w_1})}F_{l_{w_2},\kappa,N}^{(j_{w_2})}\ldots F_{l_{w_p},\kappa,N}^{(j_{w_p})}\tilde{H}_{j_1,\ldots,j_p}\right)\right|_{U_{N,\kappa}=(1,\ldots,1)}.
\end{eqnarray*}
as a sum of terms of the following form
\begin{eqnarray}
&&\lim_{x_{a_1},\ldots,x_{a_{r+1}}\rightarrow\ x_{j_1}}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\notag\\
&&\left[\frac{x_{a_1}^ru_{\tilde{a}_1}^{l_1-s_0}(\partial_{\tilde{a}_1}^{s_1}[\log\tilde{T}_N])^{d_1}\ldots (\partial_{\tilde{a}_t}^{s_t}[\log\tilde{T}_N])^{d_t}\partial_{\tilde{a}_1}^{f_1}F_{l_{w_1},\kappa,N}^{(j_{w_1})} \ldots \partial_{\tilde{a}_1}^{f_p}F_{l_{w_p},\kappa,N}^{(j_{w_p})}\partial_{\tilde{a}_1}^{h_0}\tilde{H}_{j_1,\ldots,j_p} }{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}})}\right]\label{sa1}
\end{eqnarray}
where
\begin{itemize}
\item $\{\tilde{a}_1,\ldots,\tilde{a}_{r+1}\}\subset\{1,2,\ldots,\lfloor(1-\kappa)N\}\cap R(j_1)$;
\item $s_1<s_2<\ldots<s_t$ are positive integers;
\item $f_1,\ldots,f_p,h_0$ are nonnegative integers;
\item
\begin{eqnarray}
r+s_0+s_1d_1+\ldots+s_td_t+f_1+\ldots+f_p+h_0=l_1\label{rsdl}
\end{eqnarray}
\end{itemize}
By Lemma \ref{l51}, the degree of $N$ in $(\partial_{\tilde{a}_1}^{s_1}[\log\tilde{T}_N])^{d_1}\ldots (\partial_{\tilde{a}_t}^{s_t}[\log\tilde{T}_N])^{d_t}$ is at most $d_1+\ldots+d_t$; therefore, the terms in (\ref{sa1}) with highest degree of $N$ has the form
\begin{eqnarray}
\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{x_{a_1}^ru_{\tilde{a}_1}^{l_1}(\partial_{\tilde{a}_1}[\log\tilde{T}_N])^{d_1}\partial_{\tilde{a}_1}^{f_1}F_{l_{w_1},\kappa,N}^{(j_{w_1})} \ldots \partial_{\tilde{a}_1}^{f_p}F_{l_{w_p},\kappa,N}^{(j_{w_p})}\partial_{\tilde{a}_1}^{h_0}\tilde{H}_{j_1,\ldots,j_p} }{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}})}\right]\label{sa2}
\end{eqnarray}
where
\begin{eqnarray}
s_0=d_2=\ldots=d_t=0;\ s_1=1.\label{sd0}
\end{eqnarray}
Let
\begin{eqnarray*}
B=\{i\in\{1,2,\ldots,p\}:f_i=0\}.
\end{eqnarray*}
Then
\begin{eqnarray*}
(\ref{sa2})=\left[\prod_{i\in B}F_{l_{w_i},\kappa,N}^{(j_{w_i})}\right] S(u_1,\ldots,u_{\lfloor(1-\kappa) N\rfloor})
\end{eqnarray*}
where $S(u_1,\ldots,u_{\lfloor1-\kappa N\rfloor})$ is a symmetric function. It suffices to show that the degree of $N$ in $S$, except for $S_1$ and $S_2$, is less than $\sum_{i=1}^{s}l_i-\sum_{i\in B}l_i$. Note that the degree of $N$ in $\partial_{\tilde{a}_1}[\log\tilde{T}_N])^{d_1}$ is at most $d_1$ by Lemma \ref{l51}. The summation over $\{\tilde{a}_1,\ldots,\tilde{a}_{r+1}\}\subset\{1,2,\ldots,\lfloor(1-\kappa)N\}\cap R(j_1)$ gives $O(N^{r+1})$ terms. By Lemma \ref{l53}, when $i\notin B$, the degree of $N$ in $\partial_{\tilde{a}_1}^{f_i}F_{l_{w_i},\kappa,N}^{(j_{w_i})} $ is at most $l_{w_i}$. Therefore the degree of $N$ in $S(u_1,\ldots,u_{\lfloor(1-\kappa) N\rfloor})$ is at most
\begin{eqnarray*}
\sum_{i=2}^{s}l_i-\sum_{i=1}^{p}l_{w_i}+d_1+\sum_{i\in \{1,2,\ldots,p\}\setminus B}l_{w_i}+r+1
\end{eqnarray*}
By (\ref{rsd}) and (\ref{sd0}), if $|B|\leq p-2$, $r_1+d_1+1\leq l_1-1$, then the degree of $N$ in $S(u_1,\ldots,u_{\lfloor(1-\kappa) N\rfloor})$ is at most
\begin{eqnarray*}
\sum_{i=1}^{s}l_i-\sum_{i\in B}l_i-1
\end{eqnarray*}
Therefore only the terms where at most one $f_i$ is nonzero contribute to the leading order. In these terms if $h_0>0$, then by Lemma \ref{l67}, the degree of $N$ is less than $\sum_{i=1}^{s}l_i-\sum_{i\in B}l_i$. So only the terms where $h_0=0$ and at most one $f_i$ is nonzero contribute to the leading order. These terms are in $S_1$ and $S_2$. Then the proof is complete.
\end{proof}
\begin{lemma}\label{le59}Let $s,l_1,\ldots,l_s$ be positive integers, and let $j_1,\ldots,j_s\in\{1,2,\ldots,n\}$. Then
\begin{eqnarray*}
&&\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\left[\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}-E_{l_1,\kappa,N}^{(j_1)}\right]\cdots\\&&\left.\left[\sum_{[i_s\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_s)]}\left(u_{i_s}\frac{\partial}{\partial u_{i_s}}\right)^{l_s}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]-E_{l_s,\kappa,N}^{(j_s)}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\left.\sum_{P\in \mathcal{P}_{\emptyset}^{s}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}+R\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
where the degree of $N$ in $R$ is less than $\sum_{i=1}^{s}l_i$.
\end{lemma}
\begin{proof}The lemma follows from Lemma \ref{l58} by explicit computations. See also the proof of Lemma 5.11 in \cite{bg16}.
\end{proof}
\begin{lemma}\label{le510}Let $s,l_1,\ldots,l_s$ be positive integers, and let $j_1,\ldots,j_s\in\{1,2,\ldots,n\}$. Let $\tilde{\mathcal{P}}_{\emptyset}^{s}\subset \mathcal{P}_{\emptyset}^{s}$ consisting of all the pairings of $\{1,2,\ldots,s\}$ such that in each pair $(a,b)$ in the pairing, $j_a=j_b$.
Then
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+\ldots+l_s}}\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\left[\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}-E_{l_1,\kappa,N}^{(j_1)}\right]\cdots\\&&\left.\left[\sum_{[i_s\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_s)]}\left(u_{i_s}\frac{\partial}{\partial u_{i_s}}\right)^{l_s}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]-E_{l_s,\kappa,N}^{(j_s)}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+\ldots+l_s}}\left.\sum_{P\in \tilde{\mathcal{P}}_{\emptyset}^{s}}\prod_{(a,b)\in P}\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_a)}\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
where the degree of $N$ in $R$ is less than $\sum_{i=1}^{s}l_i$.
\end{lemma}
\begin{proof}The lemma follows from Lemma \ref{le59} and the fact that the degree of $N$ in $R$, in Lemma \ref{le59}, is less than $l_1+\ldots+l_s$, and that the degree of $N$ in $\tilde{G}_{\kappa,N,(l_a,l_b)}^{(j_a,j_b)}$ is less than $l_a+l_b$ if $j_a\neq j_b$.
\end{proof}
\subsection{Integral formula for covariance}
Assume that $\kappa\in (0,1)$ and $k$ is a positive integer. Let
\begin{eqnarray*} p_k^{(\lfloor (1-\kappa)N \rfloor)}=\sum_{i=1}^{\lfloor (1-\kappa)N\rfloor}(\lambda_i+\lfloor (1-\kappa)N\rfloor-i)^k
\end{eqnarray*}
where $\lambda=(\lambda_1,\ldots,\lambda_{\lfloor (1-\kappa)N \rfloor})\in \GT_{\lfloor (1-\kappa)N \rfloor}$ has the distribution $\rho_{\lfloor (1-\kappa)N\rfloor }$ as defined in Lemma \ref{l33}. Explicit computations show that
\begin{eqnarray}
&&\mathbf{E}\left(p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa)N}\right)\left(p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}\right)\label{mts}\\
&&\cdots \left(p_{l_s}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_s}^{(\lfloor (1-\kappa)N\rfloor)}\right)\notag\\
&=&\sum_{j_1,\ldots,j_s\in \{1,2,\ldots,n\}}\frac{1}{\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N}\left[\sum_{[i_1\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_1)]}\left(u_{i_1}\frac{\partial}{\partial u_{i_1}}\right)^{l_1}-E_{l_1,\kappa,N}^{(j_1)}\right]\cdots
\notag
\\&&\left.\left[\sum_{[i_s\in\{1,2,\ldots,\lfloor(1-\kappa)N\rfloor\}\cap R(j_s)]}\left(u_{i_s}\frac{\partial}{\partial u_{i_s}}\right)^{l_s}\left[\tilde{V}_{\lfloor (1-\kappa)N\rfloor}\tilde{T}_N\right]-E_{l_s,\kappa,N}^{(j_s)}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\notag
\end{eqnarray}
\begin{lemma}
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\mathbf{E}\left(p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}\right) \left(p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}\right)\\
&=&\left.\sum_{j=1}^{n}\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\tilde{G}_{\kappa,N,(l_1,l_2)}^{(j,j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
\end{lemma}
\begin{proof}The lemma follows from (\ref{mts}), Lemma \ref{le510}, Lemma \ref{l55}, and the fact that
\begin{eqnarray*}
\left.\tilde{G}_{\kappa,N,(l_1,l_2)}^{(j,j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}=G_{\kappa,N,(l_1,l_2)}^{(j,j)}
\end{eqnarray*}
\end{proof}
Therefore, in order to obtain an explicit integral formula for the covariance
\begin{eqnarray*}
\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\mathbf{E}\left(p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}\right) \left(p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}\right);
\end{eqnarray*}
It suffices to obtain an explicit integral formula for
\begin{eqnarray*}
\left.\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\tilde{G}_{\kappa,N,(l_1,l_2)}^{(j,j)}\right|_{U_{N,\kappa}=(1,\ldots,1)},
\end{eqnarray*}
where $1\leq j\leq N$.
We have
\begin{eqnarray*}
&&\left.\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\tilde{G}_{\kappa,N,(l_1,l_2)}^{(j,j)}\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{l_1} (\partial_{\tilde{a}_1}[\log \tilde{T}_{N}])^{l_1-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]\\
&&\partial_{\tilde{a}_1}\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\right.(s+1)!\\
&&\left.\left.\lim_{[x_{b_w}\longrightarrow x_{j}],\ 1\leq w\leq s+1}\mathrm{Sym}_{b_1,\ldots,b_{s+1}}\frac{c_0 x_{b_1}^s u_{\tilde{b}_1}^{l_2}(\partial_{\tilde{b}_1}[\log \tilde{T}_N])^{l_2-s}}{(x_{b_1}u_{\tilde{b}_1}-x_{b_2}u_{\tilde{b}_2})\ldots(x_{b_1}u_{\tilde{b}_1}-x_{b_{s+1}}u_{\tilde{b}_{s+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
We consider the following cases
\begin{itemize}
\item If $\{a_1,\ldots,a_{r+1}\}\cap \{b_1,\ldots,b_{s+1}\}=\emptyset$, we have
\begin{eqnarray*}
\partial_{\tilde{a}_1}(\partial_{\tilde{b}_1}[\log \tilde{T}_N])^{l_2-s}=(l_2-s)(\partial_{\tilde{b}_1}[\log \tilde{T}_N])^{l_2-s-1}\partial_{\tilde{a}_1}(\partial_{\tilde{b}_1}[\log \tilde{T}_N])
\end{eqnarray*}
where the degree of $N$, when $U_{N,\kappa}=(1,\ldots,1)$, is at most $l_2-s-1$. By Lemma \ref{l51} and Lemma \ref{l52}, and note that
\begin{eqnarray*}
\end{eqnarray*}
we have
\begin{eqnarray*}
I_1:&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{l_1} (\partial_{\tilde{a}_1}[\log \tilde{T}_{N}])^{l_1-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]\\
&&\partial_{\tilde{a}_1}\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\emptyset]}\right.\\
&&(s+1)!\left.\left.\lim_{[x_{b_w}\longrightarrow x_{j}],\ 1\leq w\leq s+1}\mathrm{Sym}_{b_1,\ldots,b_{s+1}}\frac{c_0 x_{b_1}^s u_{\tilde{b}_1}^{l_2}(\partial_{\tilde{b}_1}[\log \tilde{T}_N])^{l_2-s}}{(x_{b_1}u_{\tilde{b}_1}-x_{b_2}u_{\tilde{b}_2})\ldots(x_{b_1}u_{\tilde{b}_1}-x_{b_{s+1}}u_{\tilde{b}_{s+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{l_1} [A_j(u_{\tilde{a}_1})N]^{l_1-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]\\
&&\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}(l_2-s)\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\emptyset]}\right.\\
&&(s+1)!\left.\left.\lim_{[x_{b_w}\longrightarrow x_{j}],\ 1\leq w\leq s+1}\mathrm{Sym}_{b_1,\ldots,b_{s+1}}\frac{ x_{b_1}^s u_{\tilde{b}_1}^{l_2}([A_j(u_{\tilde{b}_1})N])^{l_2-s-1}B_j(u_{\tilde{a}_1},u_{\tilde{b}_1})}{(x_{b_1}u_{\tilde{b}_1}-x_{b_2}u_{\tilde{b}_2})\ldots(x_{b_1}u_{\tilde{b}_1}-x_{b_{s+1}}u_{\tilde{b}_{s+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
where
\begin{eqnarray}
A_j(z)&=&\begin{cases}\frac{\kappa}{n}\sum_{l\in I_2\cap\{1,2,\ldots,n\}}\frac{y_l x_1}{1+y_lx_1 z}-\frac{\kappa}{n}\frac{n-1}{z}+\frac{1}{n}H'_{\bm_i}(z),\ \mathrm{if}\ j=1;\\-\frac{\kappa}{n}\frac{n-j}{z}+\frac{1}{n}H'_{\bm_j}(z),\ \mathrm{if}\ 2\leq j\leq n \end{cases}\label{aj}
\end{eqnarray}
and
\begin{eqnarray}
B_j(z,w)&=&\frac{\partial^2}{\partial z\partial w}\left[\log\left(1-(z-1)(w-1)\frac{z H_{\bm_j}'(z)-wH'_{\bm_j}(w)}{z-w}\right)\right]\label{bj}
\end{eqnarray}
By Lemma \ref{l552}, we obtain
\begin{eqnarray*}
I_1&\approx&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\sum_{r=0}^{l_1-1}\frac{l_1!}{(l_1-1-r)! r!}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\frac{\partial^r}{\partial z^r}\left[z^{l_1} [NA_j(z)]^{l_1-1-r}\right]\left[\sum_{s=0}^{l_2}\frac{l_2!}{(l_2-s-1)! s!}\right.\\
&&\left.\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\emptyset]}\right.\\
&&\frac{\partial^s}{\partial w^s}\left.\left[w^{l_2}[NA_j(w)]^{l_2-s-1}B_j(z,w)\right]\right|_{(z,w)=(1,1)}
\end{eqnarray*}
By the residue theorem, we deduce that
\begin{eqnarray*}
I_1&\approx&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\sum_{r=0}^{l_1-1}\frac{l_1!}{(l_1-1-r)!}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\mathrm{Res}_{z=1}\left(\frac{z^{l_1} [NA_j(z)]^{l_1-1-r}}{(z-1)^{r+1}}\right.\\
&&\left[\sum_{s=0}^{l_2-1}\frac{l_2!}{(l_2-s-1)!}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\emptyset]}\right.\\
&&\mathrm{Res}_{w=1}\left.\left[\frac{w^{l_2}[NA_j(w)]^{l_2-s-1}B_j(z,w)}{(w-1)^{s+1}}\right]\right)\\
&\approx&\frac{1}{(2\pi\mathbf{i})^2}\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\oint_{|z-1|=\epsilon}
\left(\frac{\lfloor (1-\kappa)N\rfloor}{n}\frac{z}{z-1}+zNA_j(z)\right)^{l_1}\\
&&\oint_{|w-1|=\epsilon}\left(\frac{\lfloor (1-\kappa)N\rfloor}{n}\frac{w}{w-1}+wNA_j(w)\right)B_j(z,w)dwdz\\
&=&\frac{1}{(2\pi\mathbf{i})^2}\oint_{|z-1|=\epsilon}
\left(\frac{1-\kappa}{n}\frac{z}{z-1}+zA_j(z)\right)^{l_1}\oint_{|w-1|=\epsilon}\left(\frac{1-\kappa}{n}\frac{w}{w-1}+wA_j(w)\right)^{l_2}B_j(z,w)dwdz
\end{eqnarray*}
\item If $|\{a_1,\ldots,a_{r+1}\}\cap \{b_1,\ldots,b_{s+1}\}|\geq 2$, then the degree of $N$ in these terms is at most
\begin{eqnarray*}
l_2-s+l_1-1-r+r+1+s+1-2=l_1+l_2-1<l_1+l_2,
\end{eqnarray*}
therefore the contribution of these terms to the limit is 0.
\item If $|\{a_1,\ldots,a_{r+1}\}\cap \{b_1,\ldots,b_{s+1}\}|=1$. Then
\begin{eqnarray*}
I_2:&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{l_1} (\partial_{\tilde{a}_1}[\log \tilde{T}_{N}])^{l_1-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]\\
&&\partial_{\tilde{a}_1}\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),|\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}|=1]}\right.\\
&&(s+1)!\left.\left.\lim_{[x_{b_w}\longrightarrow x_{j}],\ 1\leq w\leq s+1}\mathrm{Sym}_{b_1,\ldots,b_{s+1}}\frac{c_0 x_{b_1}^s u_{\tilde{b}_1}^{l_2}(\partial_{\tilde{b}_1}[\log \tilde{T}_N])^{l_2-s}}{(x_{b_1}u_{\tilde{b}_1}-x_{b_2}u_{\tilde{b}_2})\ldots(x_{b_1}u_{\tilde{b}_1}-x_{b_{s+1}}u_{\tilde{b}_{s+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{l_1} [A_j(u_{\tilde{a}_1})N]^{l_1-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]\\
&&\partial_{\tilde{a}_1}\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),|\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}|=1]}\right.\\
&&(s+1)!\left.\left.\lim_{[x_{b_w}\longrightarrow x_{j}],\ 1\leq w\leq s+1}\mathrm{Sym}_{b_1,\ldots,b_{s+1}}\frac{ x_{b_1}^s u_{\tilde{b}_1}^{l_2}([A_j(u_{\tilde{b}_1})N])^{l_2-s}}{(x_{b_1}u_{\tilde{b}_1}-x_{b_2}u_{\tilde{b}_2})\ldots(x_{b_1}u_{\tilde{b}_1}-x_{b_{s+1}}u_{\tilde{b}_{s+1}})}\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
By Lemma \ref{l552}, we deduce
\begin{eqnarray*}
I_2:&=&I_3+I_4
\end{eqnarray*}
where
\begin{eqnarray*}
I_3:&=&
\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\frac{\partial^r}{\partial u_{\tilde{a}_1}^r}\left[ u_{\tilde{a}_1}^{l_1} [A_j(u_{\tilde{a}_1})N]^{l_1-1-r}\right]\\
&&\partial_{\tilde{a}_1}\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\{b_1\}]}\right.\\
&&\left.\left.\frac{\partial^s}{\partial u_{\tilde{b}_1}^s}\left( u_{\tilde{b}_1}^{l_2}[A_j(u_{\tilde{b}_1})N]^{l_2-s}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)};\\
&=&
\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\frac{\partial^r}{\partial u_{\tilde{a}_1}^r}\left[u_{\tilde{a}_1}^{l_1} [A_j(u_{\tilde{a}_1})N]^{l_1-1-r}\right]\\
&&\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\{b_1\}]}\right.\\
&&\left.\left.\frac{\partial^{s+1}}{\partial u_{\tilde{b}_1}^{s+1}}\left( u_{\tilde{b}_1}^{l_2}[A_j(u_{\tilde{b}_1})N]^{l_2-s}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)};
\end{eqnarray*}
and
\begin{eqnarray*}
I_4&=&
\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}l_1\sum_{r=0}^{l_1-1}{{l_1-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\frac{\partial^r}{\partial u_{\tilde{a}_1}^r}\left[u_{\tilde{a}_1}^{l_1}[A_j(u_{\tilde{a}_1})N]^{l_1-1-r}\right]\\
&&\partial_{\tilde{a}_1}\left[\sum_{s=0}^{l_2}{{l_2}\choose{s}}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}=\{b_j\},j\neq 1]}\right.\\
&&\left.\left.\frac{\partial^s}{\partial u_{\tilde{b}_1}^s}\left( u_{\tilde{b}_1}^{l_2}[A_j(u_{\tilde{b}_1})N]^{l_2-s}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)};\\
&=&0.
\end{eqnarray*}
By the residue theorem, we infer
\begin{eqnarray*}
I_2&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\sum_{r=0}^{l_1-1}\frac{l_1!}{(l_1-1-r)!}\sum_{[\{a_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j)]}\\
&&\mathrm{Res}_{z=u_{\tilde{a}_1}}\left[ \frac{z^{l_1} [A_j(z)N]^{l_1-1-r}}{(z-1)^{r+1}}\right]\\
&&\left[\sum_{s=0}^{l_2}\frac{l_2!}{(l_2-s)!}\sum_{[\{b_1-N+\lfloor(1-\kappa)N \rfloor,\ldots,b_{s+1}-N+\lfloor(1-\kappa)N\rfloor\}\subset \{1,2,\ldots,\lfloor(1-\kappa)N \rfloor\}\cap R(j),|\{a_1,\ldots,a_{r+1}\}\cap\{b_1,\ldots,b_{s+1}\}|=1]}\right.\\
&&\left.\left.(s+1)\mathrm{Res}_{w=u_{\tilde{b}_1}}\left( \frac{w^{l_2}[A_j(w)N]^{l_2-s}}{(w-1)^{s+2}}\right)\right]\right|_{U_{N,\kappa}=(1,\ldots,1)}\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\frac{1}{(2\pi\mathbf{i})^2}\oint_{|z-1|=\epsilon}\left(\frac{\lfloor (1-\kappa)N\rfloor}{n}\frac{z}{z-1}+NzA_j(z)\right)^{l_1}\\
&&\oint_{|w-1|=\epsilon}\left(\frac{\lfloor (1-\kappa)N\rfloor}{n}\frac{w}{w-1}+NzA_j(w)\right)^{l_2}\frac{1}{(z-w)^2}dwdz\\
&=&\frac{1}{(2\pi\mathbf{i})^2}\oint_{|z-1|=\epsilon}\left(\frac{ (1-\kappa)}{n}\frac{z}{z-1}+zA_j(z)\right)^{l_1}\\
&&\oint_{|w-1|=\epsilon}\left(\frac{ (1-\kappa)}{n}\frac{w}{w-1}+wA_j(w)\right)^{l_2}\frac{1}{(z-w)^2}dwdz
\end{eqnarray*}
\end{itemize}
Then we have the following proposition
\begin{proposition}\label{p612}
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\mathbf{E}\left(p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa)N\rfloor)}\right) \left(p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}-\mathbf{E}p_{l_2}^{(\lfloor (1-\kappa)N\rfloor)}\right)\\
&=&\sum_{j=1}^n\frac{1}{(2\pi\mathbf{i})^2}\oint_{|z-1|=\epsilon}\left(\frac{ (1-\kappa)}{n}\frac{z}{z-1}+zA_j(z)\right)^{l_1}\\
&&\oint_{|w-1|=\epsilon}\left(\frac{ (1-\kappa)}{n}\frac{w}{w-1}+wA_j(w)\right)^{l_2}\left[B_j(z,w)+\frac{1}{(z-w)^2}\right]dwdz
\end{eqnarray*}
where for $1\leq j\leq n$, $A_j(z)$ and $B_j(z,w)$ are given by (\ref{aj}), (\ref{bj}).
\end{proposition}
\subsection{Central limit theorem in multiple levels}
Let
\begin{eqnarray*}
&&1\geq\kappa_1\geq \kappa_2\geq\ldots\geq\kappa_k>0\\
&&1\leq n_1\leq n_2\ldots\leq n_k\leq 2N+1;
\end{eqnarray*}
such that for $1\leq i\leq k$,
\begin{eqnarray*}
\lfloor\frac{n_i}{2}\rfloor =\lfloor (1-\kappa_i)N\rfloor.
\end{eqnarray*}
and
\begin{eqnarray*}
U_{N,\kappa_1,\ldots,\kappa_k,X}=&&\left(u_{1,1}x_{1+N-\lfloor(1-\kappa_1) N\rfloor,N},u_{2,1}x_{2+N-\lfloor(1-\kappa_1) N\rfloor,N},\ldots,u_{\lfloor (1-\kappa_1)N\rfloor,1}x_{N,N}\right.;\\
&&u_{1,2}x_{1+N-\lfloor(1-\kappa_2) N\rfloor,N},u_{2,2}x_{2+N-\lfloor(1-\kappa_2) N\rfloor,N},\ldots,u_{\lfloor (1-\kappa_2)N\rfloor,2}x_{N,N}\\
&&\ldots\\
&&\left.u_{1,k}x_{1+N-\lfloor(1-\kappa_k) N\rfloor,N},u_{2,k}x_{2+N-\lfloor(1-\kappa_k) N\rfloor,N},\ldots,u_{\lfloor (1-\kappa_k)N\rfloor,k}x_{N,N}\right)
\end{eqnarray*}
Let $\mathcal{S}_{\rho,X}(U_{N,\kappa_1,\ldots,\kappa_s,X})$ be the multi-dimensional Schur generating function as defined in \ref{df59}, where $\rho$ is the joint distribution of partitions on the $n_1$th, $n_2$th, \ldots, $n_k$th row of the square-hexagon lattice, counting from the top. Then explicit computations show that
\begin{eqnarray*}
&&\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa_1)N\rfloor)} p_{l_2}^{(\lfloor (1-\kappa_2)N\rfloor)}\cdots p_{l_k}^{(\lfloor (1-\kappa_k)N\rfloor)}\\
&=&\left.\mathcal{D}_{l_1}^{(n_1)}\mathcal{D}_{l_2}^{(n_2)}\ldots\mathcal{D}_{l_k}^{(n_k)}\mathcal{S}_{\rho,X}(U_{N,\kappa_1,\ldots,\kappa_k,X})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=(1,\ldots,1),\ \forall 1\leq s\leq k}
\end{eqnarray*}
where $\mathcal{D}_{l_i}^{(n_i)}$ is defined in (\ref{dd}).
\begin{lemma}\label{ll613}Suppose the assumptions in Lemma \ref{l59} hold. For $1\leq s\leq k$, let
\begin{eqnarray*}
t_s=N-\lfloor\frac{n_s}{2} \rfloor.
\end{eqnarray*}
Let
\begin{eqnarray*}
\mathbf{D}_{l_s}^{(n_s)}:=\frac{1}{\hat{V}_{\lfloor \frac{n_s}{2}\rfloor}}\left(\sum_{i=t_s+1}^{N}\left(u_{i}\frac{\partial}{\partial u_{i}}\right)^{l_s}\right)\hat{V}_{\lfloor \frac{n_s}{2}\rfloor},
\end{eqnarray*}
where $\hat{V}_{\lfloor \frac{n_s}{2}\rfloor}$ is the Vandermonde determinant on $\lfloor \frac{n_s}{2}\rfloor$ variables $u_1x_{t_s+1}, u_2x_{t_s+2},\ldots,u_{\lfloor\frac{n_s}{2} \rfloor}x_{N}$. Then
\begin{eqnarray*}
&&\left.\mathcal{D}_{l_1}^{(n_1)}\mathcal{D}_{l_2}^{(n_2)}\ldots\mathcal{D}_{l_k}^{(n_k)}\mathcal{S}_{\rho,X}(U_{N,\kappa_1,\ldots,\kappa_k,X})\right|_{(u_{1,s},\ldots,u_{\lfloor \frac{n_s}{2}\rfloor,s})=(1,\ldots,1),\ \forall 1\leq s\leq k}\\
&=&\frac{1}{\mathcal{S}{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}(U_{N,\kappa_1,X})}\mathbf{D}_{l_1}^{(n_1)}\frac{\mathcal{S}_{\rho_{\lfloor\frac{n_1}{2} \rfloor},X}(U_{N,\kappa_1,X})}{\mathcal{S}_{\rho_{\lfloor\frac{n_2}{2} \rfloor},X}(U_{N,\kappa_2,X})}\mathbf{D}_{l_2}^{(n_2)}\ldots\\
&&\frac{\mathcal{S}_{\rho_{\lfloor\frac{n_{k-1}}{2} \rfloor},X}(U_{N,\kappa_{k-1},X})}{\mathcal{S}_{\rho_{\lfloor\frac{n_k}{2} \rfloor},X}(U_{N,\kappa_k,X})} \mathbf{D}_{l_k}^{(n_k)}\left\{
\left.\mathcal{S}_{\rho_{\lfloor\frac{n_k}{2} \rfloor},X}(U_{N,\kappa_k,X})\right\}\right|_{(u_{1},\ldots,u_N)=(1,\ldots,1)}
\end{eqnarray*}
where $\mathcal{S}_{\rho_{\lfloor\frac{n_k}{2} \rfloor},X}(U_{N,\kappa_k,X})$ is the one-dimensional Schur generating function defined as in Definition \ref{df33}, and $\rho_{\frac{n_k}{2}}$ is a probability measure on $\GT_{\lfloor \frac{n_k}{2}\rfloor}^+$ defined as in Lemma \ref{l33}.
\end{lemma}
\begin{proof}The lemma follows from same arguments as the proof of Lemma \ref{ll511}.
\end{proof}
For $1\leq s\leq N$, let $F_{k,\kappa,N}^{(s)}$ be defined as in (\ref{fkn}). Let
\begin{eqnarray*}
\tilde{G}_{\kappa_1,\kappa_2,N,(k,l)}^{(j,s)}&=&k\sum_{r=0}^{k-1}{{k-1}\choose{r}}\sum_{[\{a_1-N+\lfloor(1-\kappa_1)N \rfloor,\ldots,a_{r+1}-N+\lfloor(1-\kappa_1)N\rfloor\}\in \{1,2,\ldots,\lfloor(1-\kappa_1)N \rfloor\}\cap R(j)]}(r+1)!\\
&&\lim_{[x_{a_1},\ldots,x_{a_{r+1}}\longrightarrow x_j]}\mathrm{Sym}_{a_1,\ldots,a_{r+1}}\\
&&\left[\frac{ x_{a_1}^r u_{\tilde{a}_1}^{k-s_0} \partial_{\tilde{a}_1}[F_{l,\kappa_2,N}^{(s)}](\partial_{\tilde{a}_1}[\log \tilde{T}_{N,\kappa_1}])^{k-1-r}}{(x_{a_1}u_{\tilde{a}_1}-x_{a_2}u_{\tilde{a}_2})\ldots(x_{a_1}u_{\tilde{a}_1}-x_{a_{r+1}}u_{\tilde{a}_{r+1}}))}\right]
\end{eqnarray*}
where
\begin{eqnarray*}
&&\tilde{T}_N=\prod_{l\in\{1,\ldots,N-\lfloor(1-\kappa_1)N \rfloor\}\cap I_2}\prod_{j=1}^{\lfloor(1-\kappa_1)N \rfloor}\left(\frac{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}u_j}{1+y_{\ol{l}}x_{\ol{N-\lfloor (1-\kappa)N\rfloor+j}}}\right)\\
&\times&\left(\prod_{i=1}^{r}s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},u_{n+i}\ldots,u_{q_{N,\kappa}n+i},1,\ldots,1\right)\right)\\
&&\times\left(\prod_{i=r+1}^{n}s_{\phi^{(j(i),\sigma_0)}(N)}\left(u_{i},u_{n+i}\ldots,u_{(q_{N,\kappa}-1)n+i},1,\ldots,1\right)\right)\\
&&\times\left(\prod_{N-\lfloor(1-\kappa)N\rfloor+1\leq i\leq N,\ 1\leq j\leq N-\lfloor(1-\kappa)N\rfloor,\ \tilde{i}\in R(p),\tilde{j}\in R(q),p<q}\frac{1}{x_iu_{\tilde{i}}-x_j}\right)\left(1+o(1)\right)
\end{eqnarray*}
\begin{lemma}
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+\ldots+l_s}}\mathbf{E}\left(p_{l_1}^{(\lfloor (1-\kappa_1)N\rfloor)}-\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa_1)N}\right)\left(p_{l_2}^{(\lfloor (1-\kappa_2)N\rfloor)}-\mathbf{E}p_{l_2}^{(\lfloor (1-\kappa_2)N\rfloor)}\right)\\
&&\cdots \left(p_{l_s}^{(\lfloor (1-\kappa_k)N\rfloor)}-\mathbf{E}p_{l_s}^{(\lfloor (1-\kappa_k)N\rfloor)}\right)\notag\\
&=&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+\ldots+l_s}}\left.\sum_{P\in \tilde{\mathcal{P}}_{\emptyset}^{s}}\prod_{(a,b)\in P}\tilde{G}_{\kappa_a,\kappa_b,N,(l_a,l_b)}^{(j_a,j_a)}\right|_{U_{N,\kappa}=(1,\ldots,1)}
\end{eqnarray*}
\end{lemma}
\begin{proof}The lemma follows from similar arguments as the proof of Lemma \ref{le510}.
\end{proof}
\begin{proposition}\label{p615} Assume $\kappa_1,\kappa_2\in (0,1)$
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{1}{N^{l_1+l_2}}\mathbf{E}\left(p_{l_1}^{(\lfloor (1-\kappa_1)N\rfloor)}-\mathbf{E}p_{l_1}^{(\lfloor (1-\kappa_1)N\rfloor)}\right) \left(p_{l_2}^{(\lfloor (1-\kappa_2)N\rfloor)}-\mathbf{E}p_{l_2}^{(\lfloor (1-\kappa_2)N\rfloor)}\right)\\
&=&\sum_{j=1}^n\frac{1}{(2\pi\mathbf{i})^2}\oint_{|z-1|=\epsilon}\left(\frac{ (1-\kappa_1)}{n}\frac{z}{z-1}+zA_j(z)\right)^{l_1}\\
&&\oint_{|w-1|=\epsilon}\left(\frac{ (1-\kappa_2)}{n}\frac{w}{w-1}+zA_j(w)\right)^{l_2}\left[B_j(z,w)+\frac{1}{(z-w)^2}\right]dwdz
\end{eqnarray*}
where for $1\leq j\leq n$, $A_j(z)$ and $B_j(z,w)$ are given by (\ref{aj}), (\ref{bj}).
\end{proposition}
\begin{proof}The proposition follows from similar arguments as the proof of Proposition \ref{p612}.
\end{proof}
Then we have the following theorem:
\begin{theorem}\label{clt2}Assume $\kappa_1,\kappa_2\in (0,1)$. Then the random variables $\left\{\frac{1}{N^{l}}[p_{l}^{\lfloor(1-\kappa)N \rfloor}-\mathbf{E}p_{l}^{\lfloor(1-\kappa)N \rfloor}]\right\}_{l,\kappa}$ converge in distribution to a mean 0 Gaussian vector with covariance given by Proposition \ref{p615}. Moreover, each $\frac{1}{N^{l}}[p_{l}^{\lfloor(1-\kappa)N\rfloor}-\mathbf{E}p_{l}^{\lfloor(1-\kappa)N \rfloor}]$ converge in distribution to the sum of $n$ independent mean 0 Gaussian random variables.
\end{theorem}
\section{Gaussian free field in uniform boundary condition}\label{gffu}
\subsection{Height function}
Let $\mathcal{R}(\Omega(N),\check{c})$ be a contracting square-hexagon lattice. Assume that the edge weights of $\SH(\check{c})$ satisfy Assumption \ref{apew}.
The planar dual graph $\mathrm{SH}^*(\check{c})$ of $\mathrm{SH}(\check{c})$ is obtained by placing a vertex of $\mathrm{SH}^*(\check{c})$ inside each face of $\mathrm{SH}(\check{c})$; two vertices of $\SH^*(\check{c})$ are adjacent, or joined by an edge in $\SH^*(\check{c})$, if and only if the two corresponding faces of $\SH(\check{c})$ share an edge of $\SH(\check{c})$.
We place a vertex of $\SH^*(\check{c})$ at the center of each face of $\SH(\check{c})$, and obtain an embedding of $\SH^*(\check{c})$ into the plane. Each face of $\SH^*(\check{c})$ is either a triangle or a square, depending on whether the corresponding vertex of $\SH(\check{c})$ inside the dual face in $\SH^*(\check{c})$ is degree-3 or degree-4.
For a contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$, let $\mathcal{R}^*(\Omega,\check{c})$ be a finite triangle-square lattice such that
\begin{itemize}
\item $\mathcal{R}^*(\Omega,\check{c})$ is a finite subgraph of $\SH^*(\check{c})$ as constructed above; and
\item $\mathcal{R}(\Omega,\check{c})$ is the interior dual graph of $\mathcal{R}^*(\Omega,\check{c})$.
\end{itemize}
In other words, $\mathcal{R}^*(\Omega,\check{c})$ is the subgraph of $\SH^*(\check{c})$ consisting of all the faces of $\SH^*(\check{c})$ corresponding to vertices of $\mathcal{R}(\Omega,\check{c})$; see Figure \ref{fig:SHht}.
\begin{definition}\label{dfn35}Let $M\in\mathcal{M}(\Omega,\check{c})$ be a perfect matching of a contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$. We color the vertices of $\mathcal{R}(\Omega,\check{c})$ by black and white such that vertices of the same color cannot be adjacent and the boundary row of $\mathcal{R}(\Omega,\check{c})$ on the bottom consists of white vertices. A height function $h_M$ is an integer-valued function on vertices of $\mathcal{R}^*(\Omega,\check{c})$ that satisfies the following property.
Let $f_1,f_2$ be a pair of adjacent vertices of $\mathcal{R}^*(\Omega,\check{c})$. Let $(f_1,f_2)$ denote the non-oriented edge of $\mathcal{R}^*(\Omega,\check{c})$ with endpoints $f_1$ and $f_2$; and let $[f_1,f_2\rangle$ (resp. $[f_2,f_1\rangle$) denote the oriented edge starting from $f_1$ (resp.\ $f_2$) and ending in $f_2$ (resp.\ $f_1$).
\begin{itemize}
\item if $(f_1,f_2)$ is a dual edge crossing a NW-SE edge or a NE-SW edge of $\SH(\check{c})$,
\begin{itemize}
\item if the oriented dual edge $[f_1,f_2\rangle$ crosses an absent edge $e$ of $\SH(\check{c})$ in $M$ then $h_M(f_2)=h_M(f_1)+1$ if $[f_1,f_2\rangle$ has the white vertex or $e$ on the left, and $h_M(f_2)=h_M(f_1)-1$ otherwise.
\item if an oriented dual edge $[f_1,f_2\rangle$ crosses a present edge $e$ of $\SH(\check{c})$ in $M$ then $h_M(f_2)=h_M(f_1)-3$ if $[f_1,f_2\rangle$ has the white vertex of $e$ on the left, and $h_M(f_2)=h_M(f_1)+3$ otherwise.
\end{itemize}
\item if $(f_1,f_2)$ is a dual edge crossing a vertical edge of $\SH(\check{a})$.
\begin{itemize}
\item If an oriented dual edge $[f_1,f_2\rangle$ crosses an absent edge $e$ of $\SH(\check{c})$ in $M$, then $h_M(f_2)=h_M(f_1)+2$ if $[f_1,f_2\rangle$ has the white vertex of $e$ on the left, and $h_M(f_2)=h_M(f_1)-2$ otherwise.
\item If an oriented dual edge $[f_1,f_2\rangle$ crosses a present edge $e$ of $\SH(\check{c})$ in $M$ then $h_M(f_2)=h_M(f_1)-2$ if $[f_1,f_2\rangle$ has the white vertex of $e$ on the left, and $h_M(f_2)=h_M(f_1)+2$ otherwise.
\end{itemize}
\item $h_{M}(f_0)=0$, where $f_0$ is the lexicographic smallest vertex of $\mathcal{R}^*(\Omega,\check{c})$.
\end{itemize}
\end{definition}
It is straightforward to verify that the height function above is well-defined, by checking that around either a degree-3 vertex or a degree-4 vertex, the total height change is 0.
Moreover, since none of the boundary edges of $\mathcal{R}(\Omega,\check{c})$ (by boundary edges we mean edges of $\SH(\check{c})$ joining exactly one vertex of $\mathcal{R}(\Omega,\check{c})$ and one vertex outside $\mathcal{R}(\Omega,\check{c})$) are present in any perfect matching of $\mathcal{R}(\Omega,\check{c})$, the height function restricted on the boundary vertices of $\mathcal{R}^*(\Omega,\check{c})$ is fixed and independent of the random perfect matching; see Figure \ref{fig:SHht}.
\begin{figure}
\includegraphics{SHht}
\caption{Contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$, dual graph $\mathcal{R}^*(\Omega,\check{c})$ and height function on the boundary. The black lines represent the graph $\mathcal{R}(\Omega,\check{c})$, the gray lines represent boundary edges of $\mathcal{R}(\Omega,\check{c})$, the red lines represent the dual graph $\mathcal{R}^*(\Omega,\check{c})$, and the height function is defined on vertices of the dual graph. The values of the height function on the boundary vertices of $\mathcal{R}^*(\Omega,\check{c})$ are also shown in the figure.}
\label{fig:SHht}
\end{figure}
\begin{theorem}(Law of large numbers for the height function.) Assume that the assumptions of Proposition \ref{plm} hold.
Let $\rho_N^k$ be the measure on the configurations of the $k$th row, and let $\kappa\in (0,1)$, such that $k=[2\kappa N]$. Let $\mathbf{m}^{\kappa}$ be the limit of the counting measures $m(\rho_N^k)$ in probability as $N\rightarrow\infty$ with moments given by (\ref{mtl})
Define
\begin{eqnarray}
\mathbf{h}(\chi,\kappa):=2\left(2(1-\kappa)\int_0^{\frac{\chi-\frac{\kappa r}{2n}}{1-\kappa}}d\mathbf{m}^{\kappa}-2\chi+2\kappa\right)\label{lh}
\end{eqnarray}
Then the random height function $h_M$ associate to a random perfect matching $M$, as defined by Definition \ref{dfn35}, has the following law of large numbers
\begin{eqnarray*}
\frac{h_{M}([\chi N],[\kappa N])}{N}\rightarrow\mathbf{h}(\chi,\kappa), \mathrm{when}\ N\rightarrow\infty
\end{eqnarray*}
where $\chi, \kappa$ are new continuous parameters of the domain.
\end{theorem}
\begin{proof}Same arguments as in the proof of Theorem 3.24 in \cite{BL17}.
\end{proof}
\subsection{Variational principle and complex Burger's equation}
Let $\mathrm{SH}(\check{c})$ be the whole plane square hexagon lattice with edge weights assigned as in Assumption \ref{apew} and periodically with period $n$ such that (\ref{px}) and (\ref{py}) hold. Then $\mathbb{Z}^2$ acts on $\mathrm{SH}(\check{c})$ by vertex-color-preserving and edge-weight-preserving isomorphisms of $\mathrm{SH}(\check{c})$. Let $\mathrm{SH}_1(\check{c})$ be the quotient graph of $\mathrm{SH}(\check{c})$ under the action of $\ZZ^2$. The graph $\mathrm{SH}_1(\check{c})$ is called a \textbf{fundamental domain} of $\mathrm{SH}(\check{c})$, which a finite graph that can be embedded into a torus.
Let $\gamma_x$ and $\gamma_y$ be two directed simple cycles winding once around the two homology generators of the torus where $\mathrm{SH}_1(\check{c})$ is embedded. Assume the edge weights of the square-hexagon lattice satisfy Assumption \ref{apew}. We shall modify the edge weights of the graph and construct a modified weighted adjacency matrix (Kasteleyn matrix) for $\mathrm{SH}_1(\check{c})$, which plays an essential role in the analysis of periodic dimer models, see \cite{Ka61,TF61,RK01}.
\begin{itemize}
\item Multiply all the edge weights $x_i$ by $-1$. This way around each face of degree 4, there are an odd number of ``$-$'' signs multiplied by edge weights; while around each face of degree 6, there are an even number of ``$-$'' signs multiplied by edge weights.
\item Multiply the weight of each edge crossed by $\gamma_x$ with $w$ (resp.\ $w^{-1}$) if the black vertex of the edge is on the left (resp.\ right) of the path; then multiply the weight of each edge crossed by $\gamma_y$ with $z$ (resp.\ $z^{-1}$) if the black vertex of the edge is on the left (resp.\ right) of the path.
\end{itemize}
Let $K(z,w)$ be the weighted adjacency matrix of $\mathrm{SH}_1(\check{c})$ with respect to the modified edge weights after the multiplication above. More precisely, the rows of $K(z,w)$ are labeled by white vertices of $\mathrm{SH}_1(\check{c})$, while the columns of $K(z,w)$ are labeled by black vertices of $\mathrm{SH}_1(\check{c})$. For a black vertex $B$ and a white vertex $W$ of $G_1$; the entry $K_{BW}(z,w)=0$ if $B$ and $W$ are not adjacent; if $B$ and $W$ are joined by an edge $e_{BW}$ in $\mathrm{SH}_1(\check{c})$, then the entry $K_{BW}(z,w)$ is the modified weight of the edge $e_{BW}$. Let $P(z,w)=\det K(z,w)$ be the \textbf{characteristic polynomial}. See \cite{KOS} for more results about the characteristic polynomial and the phase transitions of the dimer model on a bipartite, periodic graph.
\begin{example}Consider a fundamental domain of a square-hexagon lattice as illustrated in Figure \ref{fig:SHfd}.
\begin{figure}
\includegraphics{SHfd1}
\caption{A fundamental domain in a periodic square-hexagon lattice. The subgraph bounded by the dashed lines is a fundamental domain.}
\label{fig:SHfd}
\end{figure}
We have
\begin{eqnarray*}
K(z,w)=\left(\begin{array}{cc}z-x_2&w\\1+y_2z& z-x_1\end{array}\right)
\end{eqnarray*}
and
\begin{eqnarray*}
P(z,w)=\det K(z,w)=(z-x_1)(z-x_2)-w(1+y_2z).
\end{eqnarray*}
\end{example}
\begin{proposition}Let $\mathbf{h}$ be the limit height function as given by (\ref{lh}).
In the liquid region, we have
\begin{eqnarray*}
\left(\frac{1}{4}\frac{\partial \mathbf{h}}{\partial x}+\frac{1}{2},\frac{1}{4}\frac{\partial \mathbf{h}}{\partial y}+\frac{n}{2}\right)=\frac{1}{\pi}(\mathrm{arg}z,-\mathrm{arg}w),
\end{eqnarray*}
where the functions $z$ and $w$ solve the differential equation
\begin{eqnarray}
\frac{z_y}{z}+\frac{w_x}{w}=0;\label{bge}
\end{eqnarray}
and the algebraic equation $P(z,w)=0$.
\end{proposition}
\begin{proof}Same arguments as the proof of Theorem 1 in \cite{KO}; see also \cite{CKP}.
\end{proof}
When the edge weights satisfy Assumptions \ref{apew} and \ref{pw}, we can choose a fundamental domain such that $P(z,w)$ is linear in $w$. More precisely, when the period of the graph is $1\times n$, each row of the weighted adjacency matrix $K(z,w)$ has exactly two non-vanishing entries. Choose a fundamental domain consisting of $2n$ rows and $2$ columns such that the topmost row is a row of white vertices, and the rightmost column is a column of white vertices. Assume $\gamma_x$ is oriented from the left to the right and $\gamma_y$ is oriented from the top to the bottom. Let $\mathrm{SH}_1(\check{c})$ be the toroidal graph constructed from the fundamental domain above by identifying the left and right boundary as well as the top and bottom boundary. Let $v_b$ be a white vertex of $\mathrm{SH}_1(\check{c})$. The following cases might occur
\begin{itemize}
\item If $v_b$ has degree 3,
\begin{itemize}
\item when the vertex is not incident to an edge crossed by $\gamma_x$, the two non-vanishing entries of $K(z,w)$ on the row corresponding to $v_b$ are $z-x_i$ and $1$;
\item when the vertex is incident to an edge crossed by $\gamma_x$;, the two non-vanishing entries of $K(z,w)$ on the row corresponding to $v_b$ are $z-x_n$ and $w$;
\end{itemize}
\item If $v_b$ has degree 4
\begin{itemize}
\item when the vertex is not incident to an edge crossed by $\gamma_x$, the two non-vanishing entries of $K(z,w)$ on the row corresponding to $v_b$ are $z-x_i$ and $1+y_{i+1}z$;
\item when the vertex is incident to an edge crossed by $\gamma_x$;, the two non-vanishing entries of $K(z,w)$ on the row corresponding to $v_b$ are $z-x_{n-1}$ and $w(1+y_{n}z)$;
\end{itemize}
\end{itemize}
In the toroidal graph $\mathrm{SH}_1(\check{c})$, each vertex is adjacent to exactly two vertices, with possible multiple edges joining two adjacent vertices. We may consider $\mathrm{SH}_1(\check{c})$ with vertices located on a circle, then $P(z,w)=\det K(z,w)$ counts the (signed) partition function of dimer configurations on the circle.
Solving the equations $P(z,w)=0$ for $w$, we have
\begin{eqnarray*}
w=R(z),
\end{eqnarray*}
where $R(z)$ is a rational function of $z$ (quotient of two polynomials in $w$). By the explanations above, $R(z)$ can be written down explicitly as
\begin{eqnarray*}
R(z)=\frac{\prod_{i=1}^{n}(z-x_i)}{\prod_{j=\{1,2,\ldots,n\}\cap I_2}(1+y_jz)}.
\end{eqnarray*}
Therefore given $P(z,w)=0$, (\ref{bge}) becomes
\begin{eqnarray}
z_x+\frac{R(z)}{zR'(z)}z_y=0.\label{ws}
\end{eqnarray}
\begin{lemma}\label{l43}Let $z$ be a solution of the following equation
\begin{eqnarray}
F_{\kappa,m}(z)=\frac{\chi}{1-\kappa},\label{fk}
\end{eqnarray}
in the upper half plane, then $z$ also satisfies the differential equation (\ref{ws}).
\end{lemma}
\begin{proof}Differentiate (\ref{fk}) with respect to $\chi$ and $\kappa$, we have
\begin{eqnarray}
\frac{\frac{\partial z}{\partial \chi}}{\frac{\partial z}{\partial \kappa}}=\frac{1}{-\frac{z}{n}\sum_{i\in I_2\cap\{1,2,\ldots,n\}}\frac{y_i}{1+y_iz}+\sum_{j=1}^n\frac{z}{n(z-x_j)}}\label{rd}
\end{eqnarray}
Given the different scalings of $(x,y)$ and $(\chi,\kappa)$, (more precisely, in the $(x,y)$-system, we assume each fundamental domain has height 1 and width 1; while in the $(\chi,\kappa)$ system, we assume each fundamental domain has width 1 and height $n$). Then we have
\begin{eqnarray*}
\frac{\partial}{\partial x}&=&\frac{\partial }{\partial \chi}\ \mathrm{and}\ \frac{\partial}{\partial y}=n\frac{\partial}{\partial \kappa }
\end{eqnarray*}
The right hand side of (\ref{rd}) divided by $n$ is exactly $\frac{R(z)}{zR'(z)}$.
\end{proof}
\begin{lemma}\label{l44}Assume all the edge weights $x_j$'s are distinct. Let $m\geq 1$ be a positive integer. Then
\begin{enumerate}
\item For $\kappa=0$ and any $\chi\in(0,m)$, the equation
\begin{eqnarray}
F_{0,m}(z)=\chi\label{f0m}
\end{eqnarray}
has a unique root satisfying
\begin{eqnarray*}
\mathrm{Arg}(z)=\frac{\pi}{m}
\end{eqnarray*}
\item For each $z$ satisfying $\mathrm{Arg}(z)=\frac{\pi}{m}$, there exists $\chi\in(0,m)$ such that equation (\ref{f0m}) holds.
\end{enumerate}
\end{lemma}
\begin{proof}We first prove Part (1). The equation (\ref{f0m}) has the following form
\begin{eqnarray*}
\frac{1}{n}\sum_{j=0}^{n}\frac{mz^m}{z^m-x_j^m}=\chi.
\end{eqnarray*}
Let $\sqrt[m]{a}$ be the nonnegative $m$th root of a nonnegative number $a$.
Assume $z=\sqrt[m]{R}e^{\frac{\mathbf{i}\pi}{m}}$, where $R=|z^m|\geq 0$, then we have
\begin{eqnarray}
\frac{1}{n}\sum_{j=1}^{n}\frac{-mR}{-R-x_j^m}=\chi.\label{vve}
\end{eqnarray}
It suffices to show that (\ref{vve}) has a unique solution in $R\in[0,\infty)$. Explicit computations show that (\ref{vve}) is equivalent to the following equation
\begin{eqnarray*}
f(-R):=\sum_{j=1}^{n}(-mR)\prod_{i\in\{1,2,\ldots,n\},i\neq j}(-R-x_i^m)-\chi n\prod_{i=1}^{n}(-R-x_i^m)=0.
\end{eqnarray*}
Without loss of generality, assume that
\begin{eqnarray*}
0<x_1<x_2<\ldots<x_n.
\end{eqnarray*}
We have
\begin{eqnarray*}
\mathrm{sgn}[f(x_i^m)]=(-1)^{n-i}.
\end{eqnarray*}
Moreover, when $\chi>0$,
\begin{eqnarray*}
\mathrm{sgn}[f(0)]=(-1)^{n+1}.
\end{eqnarray*}
Given $\chi<m$, we have
\begin{eqnarray*}
\mathrm{sgn}[f(-\infty)]=(-1)^n,
\end{eqnarray*}
Therefore, the equation $f(z)=0$ has a solution in each one of the following intervals
\begin{eqnarray*}
(-\infty,0),(x_1^m,x_2^m),(x_2^m,x_3^m)\ldots,(x_{n-1}^m,x_n^m).
\end{eqnarray*}
Given $\chi<m$, $f(z)$ is a degree-$n$ polynomial and has at most $n$ distinct roots in $\CC$. Therefore, we have $f(z)$ has exactly one root each one of the above intervals, and Part (1) of the lemma follows.
Now we prove Part (2) of the lemma. Let
\begin{eqnarray*}
g(t)=\frac{1}{n}\sum_{j=1}^{n}\frac{mt}{t+x_j^m}
\end{eqnarray*}
Then
\begin{eqnarray*}
g'(t)=\frac{1}{n}\sum_{j=1}^{n}\frac{m x_j^m}{(t+x_j^m)^2}>0
\end{eqnarray*}
for any $t\in \RR$. Moreover
\begin{eqnarray*}
g(0)=0, \qquad \lim_{t\rightarrow\infty}g(t)=m.
\end{eqnarray*}
Then Part (2) of the Lemma follows.
\end{proof}
\begin{lemma}\label{l45}Assume all the edge weights $x_j$'s are distinct. Let $m\geq 1$ be a positive integer. Assume
\begin{eqnarray*}
|I_2\cap\{1,2,\ldots,n\}|=r.
\end{eqnarray*}
For each $\kappa\in[0,1]$ and each $\chi\in[0,m]$, the equation
\begin{eqnarray}
F_{\kappa,m}(z)=\frac{\chi}{1-\kappa}\label{fkm}
\end{eqnarray}
has a unique root $z_0(\chi,\kappa)$ satisfying
\begin{enumerate}
\item
$\mathrm{Arg}z_0(\chi,0)=\frac{\pi}{m}$.
\item $z_0(0,\kappa)=0.$
\item $\lim_{\chi\rightarrow m+\left(\frac{r}{n}-1\right)\kappa}z_0\left(\chi,\kappa\right)=\infty$.
\item $z_0(\chi,1)\in[0,+\infty)$.
\item $z_0(\chi,\kappa)$ is continuous in $(\chi,\kappa)$.
\end{enumerate}
\end{lemma}
\begin{proof}Let
\begin{eqnarray*}
g_{m-1,j}(z)=\sum_{k=0}^{m-1}z^kx_j^{m-1-k}.
\end{eqnarray*}
Then when $\kappa=1$, the equation $(1-\kappa)F_{\kappa,m}=\chi$ has the following form
\begin{eqnarray*}
p(z):=\frac{z}{n}\sum_{i\in I_2\cap\{1,2,\ldots,n\}}\frac{y_i}{1+y_iz}+\frac{z}{n}\sum_{j=1}^n\frac{\frac{\partial g_{m-1,j}(z)}{\partial z}}{g_{m-1,j}(z)}-\chi=0.
\end{eqnarray*}
Note that $p(z)$ is well-defined on $(0,+\infty)$, since $g_{m-1,j}(z)>0$ whenever $z>0$.
When $\chi\in (0,m-1+\frac{r}{n})$, we have
\begin{eqnarray*}
p(0)<0;\ \mathrm{and}\ p(+\infty)>0.
\end{eqnarray*}
Moreover, we have
\begin{lemma}
\begin{eqnarray*}
p'(z)>0
\end{eqnarray*}
when $z\in (0,+\infty)$ and $x_i,y_j>0$.
\end{lemma}
\begin{proof}Note that
\begin{eqnarray*}
p'(z)=\frac{1}{n}\sum_{i\in I_2\cap\{1,2,\ldots,n\}}\frac{y_i}{(1+y_iz)^2}+\frac{1}{n}\sum_{j=1}^{n}\frac{(zg_{m-1,j}')'g_{m-1,j}-z(g_{m-1,j}')^2}{g_{m-1,j}^2}
\end{eqnarray*}
It suffices to show that for $z>0$, $x_i>0$, we have
\begin{eqnarray}
T_m(z):=(zg_{m-1,j}')'g_{m-1,j}-z(g_{m-1,j}')^2\geq 0.\label{iie}
\end{eqnarray}
We prove (\ref{iie}) by induction on $m$. When $m=1$, we have $g_{m-1,j}=1$ and $T_m(z)=0$. Assume (\ref{iie}) holds when $m=l-1$, $l\geq 2$. When $m=l$, we have
\begin{eqnarray*}
g_{l-1,j}=z g_{l-2,j}+x_j^{m-1};
\end{eqnarray*}
and
\begin{eqnarray*}
T_{l}(z)&=&[zg_{l-2,j}+z^2g'_{l-2,j}]'(zg_{l-2,j}+x_j^{l-1})-z(g_{l-2,j}+zg_{l-2,j}')^2\\
&=&z^2 T_{l-1}(z)+x_j^{l-1}g_{l-2,j}+z x_j^{l-1}(3g_{l-2,j}'+g_{l-2,j}'')>0
\end{eqnarray*}
Since $T_{l-1}(z)>0$ by induction hypothesis, $g_{l-2,j}>0$, $g_{l-2,j}'>0$ and $g_{l-2,j}''>0$. Then the lemma follows.
\end{proof}
Hence $p(z)=0$ has exactly one root in $(0,\infty)$ when $\chi\in (0,m-1+\frac{r}{n})$. The root converges to 0 when $\chi$ goes to 0, and the root approaches $+\infty$ when $\chi$ goes to $m-1+\frac{r}{n}$.
The fact that there is a unique root of (\ref{fkm}) satisfying Condition (1) follows from Lemma \ref{l44}. The root when $\kappa=0$ satisfying condition (1) and the root when $\kappa=1$ satisfying condition (4) can be considered as boundary conditions for the Burgers equation. More precisely, the slope of height is $\frac{1}{m}$ on the bottom boundary $\kappa=0$ of the rescaled square-hexagon lattice $\mathcal{R}$, while the slope of height is $0$ on the top boundary $\kappa=1$. Since the surface tension function is strictly convex in the liquid region, the solution of the Burgers equations satisfying the given boundary conditions is unique; see \cite{CKP,KO}. By Lemma \ref{l43}, a root satisfying (1)-(5) is also a solution of the Burgers equation satisfying given boundary conditions, then the lemma follows.
\end{proof}
\begin{lemma}\label{l46}Let $\mathcal{L}$ be liquid region of the limit shape, defined by\begin{enumerate}
\item the region is a subset of
\begin{eqnarray*}
\mathcal{R}:=\left\{(\chi,\kappa): 0\leq \kappa <1;0<\chi<m+\left(\frac{r}{n}-1\right)\kappa\right\}
\end{eqnarray*}
\item the solution $z_0(\chi,\kappa)$ of $F_{\kappa,m}(z)=\frac{\chi}{1-\kappa}$ given as in Lemma \ref{l45} remains non-real in the region;
\item the region includes the bottom boundary $\kappa=0$, $0<\chi<m$ when $m\geq 2$.
\end{enumerate}
Let $\mathcal{L}_o$ be the interior of $\mathcal{L}$ Let $T_{\mathcal{L}}$ be a mapping on $\mathcal{L}$ which maps each point $(\chi,\kappa)\in \mathcal{L}$ to $z_0(\chi,\kappa)$. Then restricted on $\mathcal{L}_o$, $T_{\mathcal{L}}$ is a homeomorphism from the interior of $\mathcal{L}_o$ to $T_{\mathcal{L}}(\mathcal{L}_o)$, where
\begin{eqnarray*}
T_{\mathcal{L}}(\mathcal{L}_o)=\left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}
\end{eqnarray*}
and $\mathrm{Arg}z$ is the principal argument of $z$.
\end{lemma}
\begin{proof}First we show that $T_{\mathcal{L}}$ is one-to-one from $\mathcal{L}$ to $T_{\mathcal{L}}(\mathcal{L})$. Let
\begin{eqnarray*}
U(z)&=&\frac{z}{n}\sum_{i\in\{1,2,\ldots,n\}\cap I_2}\frac{y_i}{1+y_iz};\\
V(z)&=&\frac{z}{n}\sum_{j=1}^n\frac{1}{z-x_j}\\
W(z)&=&\frac{z}{n}\sum_{j=1}^n\left(\frac{mz^{m-1}}{z^m-x_j^m}-\frac{1}{z-x_j}\right)
\end{eqnarray*}
Then if $z$ is a solution of $F_{\kappa,m}(z)=\frac{\chi}{1-\kappa}$, we have $F_{\kappa,m}(\ol{z})=\frac{\chi}{1-\kappa}$. Hence we can solve for $\chi$ and $\kappa$ in terms of $z$ as follows
\begin{eqnarray}
&&\chi_{\mathcal{L}}(z)\label{clm}\\
&=&\frac{W(\ol{z})U(z)+V(\ol{z})U(z)-U(\ol{z})V(z)-W(\ol{z})V(z)-W(z)U(\ol{z})+W(z)V(\ol{z})}{U(z)-U(\ol{z})-V(z)+V(\ol{z})}\notag\\
&&\kappa_{\mathcal{L}}(z)=\frac{W(\ol{z})-W(z)+V(\ol{z})-V(z)}{U(z)-U(\ol{z})-V(z)+V(\ol{z})}\label{klm}
\end{eqnarray}
Therefore $T_{\mathcal{L}}$ is one-to-one from $\mathcal{L}_o$ to $T_{\mathcal{L}}(\mathcal{L}_o)$. From the expression of $T_{\mathcal{L}}$, $\chi_{\mathcal{L}}$ and $\kappa_{\mathcal{L}}$ it is straightforward to see that both $T_{\mathcal{L}}$ and $\chi_{\mathcal{L}}$, $\kappa_{\mathcal{L}}$ are continuous.
Then we claim that
\begin{eqnarray}
\mathcal{L}_o=\mathcal{L}\setminus \{\kappa=0\};\label{lo}
\end{eqnarray}
and $\mathcal{L}_o$ is connected. To see why that is true,
let $(\chi_1,\kappa_1)\in\mathcal{L}\setminus\{\kappa=0\}$ and $z_1= T_{\mathcal{L}}(\chi_1,\kappa_1)$. From the definition of $\mathcal{L}$, we have $z_1\notin \RR$. Note that $z_1$ be the root of $F_{\kappa_1}(z)=\frac{\chi_1}{1-\kappa_1}$ as given by Lemma \ref{l45}. Given $0<\kappa_1<1$, we obtain
\begin{eqnarray*}
z_1\notin\{x_je^{\frac{2t\pi}{m}}:1\leq j\leq n,1\leq t\leq m\}.
\end{eqnarray*}
We shall show that $(\chi_2,\kappa_2)\in \mathcal{L}\setminus \{\kappa=0\}$ whenever $|\chi_1-\chi_2|$ and $|\kappa_1-\kappa_2|$ are sufficiently small. Fix $\epsilon>0$ such that
\begin{itemize}
\item \begin{eqnarray}
B(z_1,\epsilon)\cap [\RR\cup\{x_je^{\frac{2t\pi}{m}}:1\leq j\leq n,1\leq t\leq m\}]=\emptyset \label{zeo}
\end{eqnarray}
\item
\begin{eqnarray*}
\inf_{z\in\partial B(z_1,\epsilon)}\left|F_{\kappa_1}(z)-\frac{\chi_1}{1-\kappa_1}\right|>0;
\end{eqnarray*}
\item $F_{\kappa_1}(z)=\frac{\chi_1}{1-\kappa_1}$ has a unique zero in $B(z_1,\epsilon)$.
\item $B(z_1,\epsilon)\subset \mathcal{R}$.
\end{itemize}
By (\ref{zeo})
\begin{eqnarray*}
\left|F_{\kappa_1}(z)-\frac{\chi_1}{1-\kappa_1}-F_{\kappa_2}(z)+\frac{\chi_2}{1-\kappa_2}\right|<\epsilon,
\end{eqnarray*}
for any $\epsilon>0$, whenever $|\chi_1-\chi_2|$ and $|\kappa_1-\kappa_2|$ are sufficiently small, and $z\in B(z_1,\epsilon)$.
Therefore when $|\chi_1-\chi_2|$ and $|\kappa_1-\kappa_2|$ are sufficiently small, we have
\begin{eqnarray*}
\left|F_{\kappa_1}(z)-\frac{\chi_1}{1-\kappa_1}\right|>\left|F_{\kappa_1}(z)-\frac{\chi_1}{1-\kappa_1}-F_{\kappa_2}(z)+\frac{\chi_2}{1-\kappa_2}\right|.
\end{eqnarray*}
for any $z\in \partial B(z_1,\epsilon)$.
By Rouch\'e's theorem, $F_{\kappa_2}(z)-\frac{\chi_2}{1-\kappa_2}$ has a root in $B(z_1,\epsilon)$, which is as described in Lemma \ref{l45}. Hence $(\chi_2,\kappa_2)\in\mathcal{L}\setminus \{\kappa=0\}$, and we obtain (\ref{lo}). The statement that $\mathcal{L}_o$ is connected follows from the fact that $\mathcal{L}$ is connected and that for any point $(\chi,0)$ with $0<\chi<\mathcal{L}$ there is a neighborhood $B_{\delta}(\chi,0)$ such that
$B_{\delta}(\chi,0)\cap R\in \mathcal{L}$.
Now we claim that
\begin{eqnarray}
T_{\mathcal{L}}(\mathcal{L}_o)\subseteq \left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}\label{tls}
\end{eqnarray}
To see why that is true, assume there exists a point $c\in T_{\mathcal{L}}(\mathcal{L}_o)\setminus \left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}$. Since $T_{\mathcal{L}}(\mathcal{L}_o)\cap \left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}\neq \emptyset$, we can find $c_1\in T_{\mathcal{L}}(\mathcal{L}_o)\cap \left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}$. Let $d,d_1\in \mathcal{L}_o$ such that
\begin{eqnarray*}
T_{\mathcal{L}}(d)=c;\qquad T_{\mathcal{L}}(d_1)=c_1
\end{eqnarray*}
Since $\mathcal{R}$ is connected, we can find a path $p_{dd_1}$ in $\mathcal{R}$ joining $d$ and $d_1$ such that $p_{dd_1}\cap \{\kappa=0\}=\emptyset$. By continuity $T_{\mathcal{L}}(p_{dd_1})$ is a continuous curve in $\CC$ joining the point $c$ satisfying $\mathrm{Arg}c<\frac{\pi}{m}$ and the point $c_1$ satisfying $\mathrm{Arg}c_1>\frac{\pi}{m}$. Then there exists $c_2\in T_{\mathcal{L}}(d_2)$ such that $d_1\in p_{dd_1}$ and $\mathrm{Arg} c_2=\frac{\pi}{m}$. By (\ref{klm}), $d_2$ on the line $\kappa=0$, but this is a contraction to the fact that $p_{dd_1}\cap \{\kappa=0\}=\emptyset$. The contradiction implies (\ref{tls}).
We finally show that
\begin{eqnarray*}
\left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}\subseteq T_{\mathcal{L}}(\mathcal{L}_o).
\end{eqnarray*}
Assume that there exists $t\in \partial T_{\mathcal{L}}(\mathcal{L}_o)$, such that $t\in \left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}\setminus T_{\mathcal{L}}(\mathcal{L}_o)$. Then there exists a sequence $\{t_n\}_{n\in \NN}\subset T_{\mathcal{L}}(\mathcal{L}_o)$ such that
\begin{eqnarray*}
\lim_{n\rightarrow\infty}t_n=t.
\end{eqnarray*}
By continuity of (\ref{clm}), (\ref{klm}) we have
\begin{eqnarray*}
\lim_{n\rightarrow\infty}\chi(t_n)&=&\chi(t)\\
\lim_{n\rightarrow\infty}\kappa(t_n)&=&\kappa(t).
\end{eqnarray*}
Hence $(\chi(t)$, $\kappa(t))\in\mathcal{L}_o\cup \partial \mathcal{L}_o$ is such that $F_{\kappa(t),m}(z)=\frac{\chi(t)}{1-\kappa(t)}$ has a root $t$ in $\left\{z\in \mathbb{C}:0< \mathrm{Arg}z<\frac{\pi}{m}\right\}$ as described in Lemma \ref{l45}.
Note that $\kappa(t)\neq 1$, because if $\kappa(t)=1$, then $\mathrm{Arg} t=0$. Also $\chi(t)\neq 0$, because if $\chi(t)=0$, then $t=0$. Moreover, $\chi(t)\neq m+\left(\frac{r}{n}-1\right)\kappa(t)$, because otherwise $t=\infty$.
Then $\chi(t),\kappa(t)\in \mathcal{L}$. Since $\mathrm{Arg} t\neq \frac{\pi}{m}$, $\kappa(t)\neq 0$, we have $(\chi(t),\kappa(t))\in \mathcal{L}_o$ and $t\in T_{\mathcal{L}}(\mathcal{L}_o)$. Then the proof is complete.
\end{proof}
\subsection{Gaussian Free Field}
Let $C_0^{\infty}$ be the space of smooth real-valued functions with compact support in the upper half plane $\HH$. The \textbf{Gaussian free field} (GFF) $\Xi$ on $\HH$ with the zero boundary condition is a collection of Gaussian random variables $\{\xi_{f}\}_{f\in C_0^{\infty}}$ indexed by functions in $C_0^{\infty}$, such that the covariance of two Gaussian random variables $\xi_{f_1}$, $\xi_{f_2}$ is given by
\begin{eqnarray*}
\mathrm{Cov}(\xi_{f_1},\xi_{f_2})=\int_{\HH}\int_{\HH}f_1(z)f_2(w)G_{\HH}(z,w)dzd\ol{z}dwd\ol{w},
\end{eqnarray*}
where
\begin{eqnarray*}
G_{\HH}(z,w):=-\frac{1}{2\pi}\ln\left|\frac{z-w}{z-\ol{w}}\right|,\qquad z,w\in \HH
\end{eqnarray*}
is the Green's function of the Dirichlet Laplacian operator on $\HH$. The Gaussian free field $\Xi$ can also be considered as a random distribution on $C_0^{\infty}$ of $\HH$, such that for any $f\in C_0^{\infty}$, we have
\begin{eqnarray*}
\Xi(f)=\int_{\HH}f(z)\Xi(z)dz:=\xi_f.
\end{eqnarray*}
See \cite{SS07} for more about GFF.
Consider a contracting square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$. Let $\omega$ be a signature corresponding to the boundary row.
Let
\begin{eqnarray*}
\mathcal{S}^N=(\mu^{(N)},\nu^{(N)},\ldots,\mu^{(1)},\nu^{(1)}).
\end{eqnarray*}
be the sequence of (random) partitions corresponding to the (random) dimer configuration on the contraction square-hexagon lattice $\mathcal{R}(\Omega,\check{c})$, as in Proposition \ref{p16}.
Define a function $\Delta^N$ on $\RR_{\geq 0}\times \RR_{\geq 0}\times \mathcal{S}\rightarrow \NN$ as follows
\begin{eqnarray*}
&& \Delta^{N}: (x,y,(\mu^{(N)},\nu^{(N)},\ldots,\mu^{(1)},\nu^{(1)}))\longrightarrow\\
&&\sqrt{\pi}|\{1\leq s\leq N-\lfloor y\rfloor\}:\mu_s^{N-\lfloor y\rfloor}+(N-\lfloor y\rfloor)-s\geq x\}|
\end{eqnarray*}
Let $\Delta_{\mathcal{M}}^N(x,y)$ be the pushforward of the measure $P_{\omega}^N$ on $\mathcal{S}^N$ with respect to $\Delta^N$. For $z\in \HH$, define
\begin{eqnarray*}
\mathbf{\Delta}_{\mathcal{M}}^N(z):=\Delta_{M}^N(N\chi_{\mathcal{L}}(z),N\kappa_{\mathcal{L}}(z)),
\end{eqnarray*}
where $\chi_{\mathcal{L}}(z)$, $\kappa_{\mathcal{L}}(z)$ are defined by (\ref{clm}), (\ref{klm}), respectively.
Here is the main theorem we shall prove in the section.
\begin{theorem}\label{t710}Let $\mathbf{\Delta}_{\mathcal{M}}^{N}(z)$ be a random function corresponding to the random perfect matching of the contracting square-hexagon lattice, as explained above. Then
\begin{eqnarray*}
\mathbf{\Delta}_{\mathcal{M}}^{N}(z)-\EE\mathbf{\Delta}_{\mathcal{M}}^{N}(z)\rightarrow \Xi(z),\ \mathrm{as}\ N\rightarrow\infty.
\end{eqnarray*}
Here $\Xi(z)$ is the Gaussian free field in $\mathbb{S}$ with zero boundary conditions as defined above. The convergence is in the sense that for $0<\kappa\leq 1$, $j\in \NN$,
\begin{eqnarray}
M_j^{\kappa}\rightarrow \mathbf{M}_j^{\kappa},\ \mathrm{as}\ N\rightarrow\infty,\label{mjk}
\end{eqnarray}
where
\begin{eqnarray*}
M_j^{\kappa}=\int_{-\infty}^{+\infty}\chi^j(\mathbf{\Delta}_{\mathcal{M}}^{N}(N\chi,N\kappa)-\mathbb{E}\mathbf{\Delta}_{\mathcal{M}}^{N}(N\chi,N\kappa))d\chi,
\end{eqnarray*}
and
\begin{eqnarray*}
\mathbf{M}_j^{\kappa}=\int_{z\in\HH;\kappa_{\mathcal{L}}(z)=\kappa}\chi_{\mathcal{L}}(z)^j\Xi(z)d\chi_{\mathcal{L}}(z).
\end{eqnarray*}
\end{theorem}
\begin{proof}By Theorem \ref{gff1}, we have
\begin{eqnarray}
&&\lim_{N\rightarrow\infty}\frac{\mathrm{cov}\left(p_{l_1}^{\left(1-\kappa_1) N\right)},p_{l_2}^{\left((1-\kappa_2)N\right)}\right)}{N^{l_1+l_2}}\label{cor}\\
&=&\frac{(1-\kappa_1)^{l_1}(1-\kappa_2)^{l_2}}{(2\pi\mathbf{i})^2}\sum_{i=1}^{n}\sum_{j=1}^{n}\oint_{|z-x_i|=\epsilon}\oint_{|w-x_j|=\epsilon}\left[F_{\kappa_1,m}(z)\right]^{l_1}\left[F_{\kappa_2,m}(w)\right]^{l_2}Q(z,w)dzdw,\notag
\end{eqnarray}
The poles of $F_{\kappa,m}(z)$ are of 3 types
\begin{enumerate}
\item $x_1,\ldots,x_n$ lying on the positive real axis;
\item $-\frac{1}{y_j}$ for $j\in I_2\cap\{1,2,\ldots,n\}$ lying on the negative real axis;
\item roots of $z^m=x_j^m$ except $x_j$ for $j=1,\ldots,n$, lying on the circle centered at $0$ with radius $x_j$.
\end{enumerate}
We may change the sum of contour integrals in the RHS of (\ref{cor}) into an integral over a contour enclosing all the poles of $F_{\kappa,m}(z)$ of type (1), yet enclosing no poles of types (2) and (3), with respect to both $z$ and $w$.
For $\kappa\in(0,1)$, let
\begin{eqnarray*}
\mathcal{Z}_1(\kappa)=\left\{z: F_{\kappa,m}(z)=\frac{\chi}{1-\kappa}; \chi\in \left[0, m+\kappa\left(\frac{r}{n}-1\right)\right]; z\ \mathrm{is\ a\ root\ as\ given\ by\ Lemma\ \ref{l45}}\right\}.
\end{eqnarray*}
Let
\begin{eqnarray*}
\mathcal{Z}_2(\kappa)=\left\{z:\ol{z}\in \mathcal{Z}_1(\kappa)\right\}.
\end{eqnarray*}
and
\begin{eqnarray*}
\mathcal{Z}(\kappa)=\mathcal{Z}_1(\kappa)\cup \mathcal{Z}_2(\kappa).
\end{eqnarray*}
We claim that for $\kappa\in(0,1)$, $\mathcal{Z}(\kappa)$ is a contour in the complex plane $\CC$ enclosing all the poles of $F_{\kappa,m}(z)$ of type (1), yet enclosing no poles of type (2) and (3). By Lemma \ref{l46}, we have
\begin{eqnarray*}
\mathcal{Z}_1(\kappa)\in \{0,+\infty\}\cup \left\{z:0<\mathrm{Arg}z<\frac{\pi}{m}\right\}.
\end{eqnarray*}
Since $\mathcal{Z}_2(\kappa)$ is the complex conjugate of $\mathcal{Z}_1(\kappa)$, then the claim follows.
For $\kappa_1,\kappa_2\in(0,1)$, (\ref{cor}) becomes
\begin{eqnarray*}
&&\lim_{N\rightarrow\infty}\frac{\mathrm{cov}\left(p_{l_1}^{\left(1-\kappa_1) N\right)},p_{l_2}^{\left((1-\kappa_2)N\right)}\right)}{N^{l_1+l_2}}\\
&=&\frac{(1-\kappa_1)^{l_1}(1-\kappa_2)^{l_2}}{(2\pi\mathbf{i})^2}\oint_{z\in \mathcal{Z}(\kappa_1)}\oint_{w\in\mathcal{Z}(\kappa_2)}\left[F_{\kappa_1,m}(z)\right]^{l_1}\left[F_{\kappa_2,m}(w)\right]^{l_2}Q(z,w)dzdw,\\
&=&\frac{1}{(2\pi\mathbf{i})^2}\oint_{z\in \mathcal{Z}(\kappa_1)}\oint_{w\in\mathcal{Z}(\kappa_2)}\left[\chi_{\mathcal{L}}(z)\right]^{l_1}\left[\chi_{\mathcal{L}}(z)\right]^{l_2}Q(z,w)dzdw
\end{eqnarray*}
where
\begin{eqnarray*}
Q(z,w)=\frac{m^2z^{m-1}w^{m-1}}{(z^m-w^m)^2}
\end{eqnarray*}
Integrate (\ref{mjk}) by parts we obtain
\begin{eqnarray*}
M_j^{\kappa}=\frac{N^{-(j+1)}\sqrt{\pi}}{j+1}(p_{j+1}^{\kappa}-\mathbf{E}p_{j+1}^{\kappa}).
\end{eqnarray*}
Hence the random variables $\{M_j^{\kappa}\}_{\kappa\in(0,1),j\in \NN}$ converge to the Gaussian distribution with mean 0 and limit covariance
\begin{eqnarray}
&&\lim_{N\rightarrow\infty}\mathrm{cov}(M_{j_1}^{\kappa_1},M_{j_2}^{\kappa_2})\label{ra}\\
&=&\frac{-1}{4\pi(j_1+1)(j_2+1)}\oint_{z\in \mathcal{Z}(\kappa_1)}\oint_{2\in \mathcal{Z}(\kappa_2)}\chi_{\mathcal{L}}(z)^{j_1+1}\chi_{\mathcal{L}}(w)^{j_2+1}\frac{m^2z^{m-1}w^{m-1}}{(z^m-w^m)^2}dzdw.\notag
\end{eqnarray}
Let
\begin{eqnarray*}
\mathbb{S}=\left\{z:0<\mathrm{Arg} z<\frac{\pi}{m}\right\}.
\end{eqnarray*}
By definition, the random variables $\{\mathcal{M}_j^{\kappa}\}_{\kappa\in(0,1),j\in \NN}$ are Gaussian with mean 0 and covaraince
\begin{eqnarray}
&&\mathrm{cov}(\mathcal{M}_{j_1}^{\kappa_1},\mathcal{M}_{j_2}^{\kappa_2})\label{rb}\\
&=&\oint_{z\in\mathbb{S}:\kappa_{\mathcal{L}}(z)=\kappa_1}\oint_{w\in \mathbb{S}:\kappa_{\mathcal{L}}(w)=\kappa_2}\chi_{\mathcal{L}}(z)^{j_1}\chi_{\mathcal{L}}(w)^{j_2}\frac{d\chi_{\mathcal{L}}(z)}{dz}\frac{d\chi_{\mathcal{L}}(w)}{dw}\mathcal{G}_{\mathbb{S}}(z,w)dzdw,\notag
\end{eqnarray}
where $\mathcal{G}_{\mathbb{S}}(z,w)$ is the Green's function on $\mathbb{S}$ given by
\begin{eqnarray*}
-\frac{m^2z^{m-1}w^{m-1}}{2\pi}\mathrm{ln}\left|\frac{z^m-w^m}{z^m-\ol{w}^m}\right|
\end{eqnarray*}
Then integration by parts shows that the right hand sides of (\ref{ra}) and (\ref{rb}) are equal.
\end{proof}
\section{Gaussian Free Field in Piecewise Boundary Condition}\label{gffp}
For $1\leq i\leq n$, let
\begin{eqnarray*}
F_{i,\kappa}(z)=\frac{1}{n}\frac{z}{z-1}+\frac{1}{1-\kappa}zA_i(z);
\end{eqnarray*}
where $A_i(z)$ is defined by (\ref{aj}).
\begin{lemma}For any $\chi>0$, $\kappa\in (0,1)$ and $1\leq i\leq n$ the equation
\begin{eqnarray}
F_{i,\kappa}(z)=\frac{\chi}{1-\kappa}.\label{fik}
\end{eqnarray}
has at most one pair of complex conjugate roots.
\end{lemma}
\begin{proof}See Proposition 7.2 of \cite{Li18}.
\end{proof}
Let $J_i$ be defined as in (\ref{ji}). Under Assumption \ref{ap32}, we may assume that
\begin{eqnarray*}
J_i=\left\{\begin{array}{cc}\{d_i,d_i+1,\ldots,d_{i+1}-1\}& \mathrm{if}\ 1\leq i\leq n-1\\ \{d_n, d_n+1,\ldots,s\}& \mathrm{if}\ i=n \end{array}\right.
\end{eqnarray*}
where for $1\leq i\leq n$,
\begin{eqnarray*}
1=d_1<d_2<\ldots <d_n\leq s
\end{eqnarray*}
\begin{lemma}Let $1\leq i\leq n$. Let $\mathcal{S}_i$ consisting of all the $(\chi,\kappa)$ in $\mathcal{R}$ (the rescaled square-hexagon lattice $\frac{1}{N}\mathcal{R}(\Omega,\check{c})$ in the limit as $N\rightarrow\infty$) such that equation (\ref{fik}) has exactly one pair of complex conjugate roots. Let $\HH$ be the upper half plane defined by
\begin{eqnarray*}
\HH=\{z: \mathrm{Im}z>0\}.
\end{eqnarray*}
The mapping
\begin{eqnarray*}
T_{\mathcal{S}_i}:\mathcal{S}_i\rightarrow \mathbb{H}
\end{eqnarray*}
maps $(\chi,\kappa)\in \mathcal{S}_i$ to $t:=\mathrm{St}_{\mathbf{m}_i}^{(-1)}(\log z)$, where $z$ is the unique root of (\ref{fik}) in $\HH$. Let
\begin{eqnarray*}
p(t)&=&\sum_{l\in I_2\cap\{1,2,\ldots,n\}}\frac{y_lx_1 \exp[\mathrm{St}_{\mathbf{m}_i}(t)]}{1+y_lx_1 \exp[\mathrm{St}_{\mathbf{m}_i}(t)]}\\
q(t)&=&\frac{\exp(\mathrm{St}_{\mathbf{m}_i}(t))}{\exp(\mathrm{St}_{\mathbf{m}_i}(t))-1}
\end{eqnarray*}
Then $T_{\mathcal{S}_1}$ is a homeomorphism with inverse $t\rightarrow (\chi_{\mathcal{S}_1}(t),\kappa_{\mathcal{S}_1}(t))$ for all $t\in \HH$, given by
\begin{eqnarray}
\chi_{\mathcal{S}_1}(t)&=&\frac{\ol{t}(p(t)-q(t))-t(p(\ol{t})-q(\ol{t}))-(n-1)(\ol{t}-t)}{n[p(t)-p(\ol{t})-q(t)+q(\ol{t})]}\label{cst1}\\
\kappa_{\mathcal{S}_1}(t)&=&\frac{\ol{t}-t}{p(t)-p(\ol{t})-q(t)+q(\ol{t})}\label{kst1}
\end{eqnarray}
and for $2\leq i\leq n$
\begin{eqnarray}
\chi_{\mathcal{S}_i}(t)&=&\frac{\ol{t}q(t)-tq(\ol{t})+(\ol{t}-t)(n-i)}{n[q(t)-q(\ol{t})]}\label{csti}\\
\kappa_{\mathcal{S}_i}(t)&=&\frac{\ol{t}-t}{-q(t)+q(\ol{t})}\label{ksti}
\end{eqnarray}
\end{lemma}
\begin{proof}The proof is an adaptation of Proposition 6.2 of \cite{bk}; see also Theorem 2.1 of \cite{DM15}.
It suffices to show all the following statements for each $1\leq i\leq n$:
\begin{enumerate}
\item $\mathcal{S}_i$ is nonempty.
\item $\mathcal{S}_i$ is open.
\item $T_{\mathcal{S}_i}: \mathcal{S}_i\rightarrow \HH$ is continuous.
\item $T_{\mathcal{S}_i}: \mathcal{S}_i\rightarrow \HH$ is injective.
\item $T_{\mathcal{S}_i}: \mathcal{S}_i\rightarrow T_{\mathcal{S}_i}(\mathcal{S}_i)$ has continuous inverse for all $t\in T_{\mathcal{S}_i}(\mathcal{S}_i)$.
\item $T_{\mathcal{S}_i}(\mathcal{S}_i)=\HH$.
\end{enumerate}
We first prove (1). Explicit computations show that $(\chi_{\mathcal{S}_i}(t),\kappa_{\mathcal{S}_i}(t))$ satisfies (\ref{cst1}), (\ref{kst1}) when $i=1$ and (\ref{csti}), (\ref{ksti}) when $2\leq i\leq n$. Since $\mathbf{m}_{i}$ is a measure on $\RR$ with compact support, assume that $\mathrm{Support}(\mathbf{m}_i)\subset[a,b]$ where $a,b\in\RR$. The Stieltjes transform satisfies
\begin{eqnarray*}
\mathrm{St}_{\mathbf{m}_i}(t)=\frac{1}{t}+\frac{\alpha}{t^2}+\frac{\beta}{t^3}+O(|t|^{-4}),
\end{eqnarray*}
where
\begin{eqnarray*}
\alpha&=&\int_a^b x\mathbf{m}_{i}(dx)\\
\beta&=&\int_a^b x^2\mathbf{m}_{i}(dx).
\end{eqnarray*}
Let
\begin{eqnarray*}
c_i=\frac{1}{y_ix_1}
\end{eqnarray*}
After computations we have
\begin{eqnarray*}
\chi_{\mathcal{S}_1}&=&\frac{\alpha-1}{n}-\frac{n-1}{n}+O(|t|^{-1}).\\
\kappa_{\mathcal{S}_1}&=&1+\left(\alpha^2-\beta-\sum_{i\in I_2\cap\{1,2,\ldots,n\}}\frac{c_i}{(1+c_i)^2}\right)\frac{1}{|t|^2}+O(|t|^{-3}).
\end{eqnarray*}
and for $2\leq i\leq n$,
\begin{eqnarray*}
\chi_{\mathcal{S}_i}&=&\frac{\alpha-1}{n}-\frac{n-i}{n}+O(|t|^{-1}).\\
\kappa_{\mathcal{S}_i}&=&1+\left(\alpha^2-\beta\right)\frac{1}{|t|^2}+O(|t|^{-3}).
\end{eqnarray*}
Let $\lambda$ be the Lebesgue measure on $\RR$. Recall that $\mathbf{m}_{i}$ is the limit counting measure for $\phi^{(i,\si_0)}$ as $N\rightarrow\infty$. By Assumption \ref{ap428}, we have
\begin{eqnarray*}
\alpha\geq n-i+1
\end{eqnarray*}
Similarly,
\begin{eqnarray*}
\beta-\alpha^2\geq \frac{1}{2}\int_{0}^1\int_{0}^1(x-y)^2dxdy=\frac{1}{12}.
\end{eqnarray*}
As a result, $(\chi,\kappa)\in\left(0,\infty\right)\times(0,1)$ whenever $|t|$ is sufficiently large. Then (1) follows.
The facts (2) and (3) follow from Rouch\'e's theorem by the same arguments as in the proof of Proposition 6.2 in \cite{bk}. The facts (4)-(6) can also be obtained by the same arguments as in the proof of Proposition 6.2 in \cite{bk}.
\end{proof}
\begin{theorem}Let $\mathbf{\Delta}_{\mathcal{M}}^{N}(z)$ be a random function corresponding to the random perfect matching of the contracting square-hexagon lattice, as in Theorem \ref{t710}, but with piecewise boundary conditions satisfying Assumptions \ref{ap423}, \ref{ap32} and \ref{ap428}. Then
\begin{eqnarray*}
\mathbf{\Delta}_{\mathcal{M}}^{N}(z)-\EE\mathbf{\Delta}_{\mathcal{M}}^{N}(z)\rightarrow \Xi_1(z)+\Xi_2(z)+\ldots+\Xi_n(z),\ \mathrm{as}\ N\rightarrow\infty.
\end{eqnarray*}
Here for $1\leq i\leq n$ $\Xi_i(z)$'s are $n$ independent Gaussian free fields in $\mathbb{H}$ with zero boundary conditions. The convergence is in the sense that for $0<\kappa\leq 1$, $j\in \NN$,
\begin{eqnarray}
M_j^{\kappa}\rightarrow \mathbf{M}_j^{\kappa},\ \mathrm{as}\ N\rightarrow\infty,\label{mjk}
\end{eqnarray}
where
\begin{eqnarray*}
M_j^{\kappa}=\int_{-\infty}^{+\infty}\chi^j(\mathbf{\Delta}_{\mathcal{M}}^{N}(N\chi,N\kappa)-\mathbb{E}\mathbf{\Delta}_{\mathcal{M}}^{N}(N\chi,N\kappa))d\chi,
\end{eqnarray*}
and
\begin{eqnarray*}
\mathbf{M}_j^{\kappa}=\int_{z\in\HH;\kappa_{\mathcal{L}}(z)=\kappa}\chi_{\mathcal{L}}(z)^j\sum_{i=1}^n\Xi_i(z)d\chi_{\mathcal{L}}(z).
\end{eqnarray*}
\end{theorem}
\begin{proof}Similar arguments as the proof of Theorem 6.3 in \cite{bk}.
\end{proof}
\bigskip
\bigskip
\noindent\textbf{Acknowledgements.} ZL's research is supported by National Science Foundation grant DMS 1608896. ZL thanks Vadim Gorin for helpful discussions.
|
\section{Introduction} \label{sec:intro}
The hypercube $Q_k$ is a $k$-regular graph on $2^k$ nodes
that can be constructed by
labeling the nodes by the $2^k$ subsets of the set $\{1,2,\dots,k\}$
and placing an edge between two nodes whenever
the two node labels (i.e., the two subsets) differ in
a single element.
{
Hypercubes are ubiquitous, both in mathematics and in the ``real world.''
It can be seen that $Q_k$ is the ``covering graph'' of the family of
all subsets of a $k$-element set,
see \cite{Bol},
and hence, properties of hypercubes have universal appeal.
(Recall that the covering graph of
the poset/powerset of a set $U$ has a node for each subset $A$ of
$U$, and the nodes of subsets $A,B$ are adjacent iff $|A \bigtriangleup B|=1$.)
}
Hypercubes (and their variants) are useful in
computer communication networks, VLSI design, etc., and
there is extensive literature in this area, see \cite{CD90,FKL02,H95,L92,SS88}.
An \textit{orientation} of an (undirected) graph $G=(V,E)$ is a directed~graph
$D=(V,A)$ such that each edge $\{v,w\}\in E$ is replaced
by exactly one of the arcs $(v,w)$ or $(w,v)$.
Orientations of hypercubes have applications in practical domains
such as broadcasting in computer communication networks
and the design of parallel computer architectures.
The connectivity properties of hypercubes and orientations of hypercubes
have been studied, see \cite{CD90,H95,SS88}, and
orientations of hypercubes that achieve the maximum possible node
connectivity are of interest, see \cite[Proposition~9]{H95}.
Our key result states that the optimal node connectivity
among orientations of $Q_{2k}$
can be achieved in a trivial way:
pick any orientation such that
the indegree is equal to the outdegree at every node.
\subsection{Smooth orientations and Eulerian orientations} \label{sec:smooth}
For a node $v$ of a directed graph, we use $d^{in}(v)$
to denote the number of arcs with head~$v$;
similarly, $d^{out}(v)$ denotes the number of arcs with tail~$v$.
An orientation of an (undirected) graph $G$ is called \textit{smooth}
if the absolute value of the difference between the
indegree and the outdegree of every node is at most one, that is,
$\big| d^{in}(v)-d^{out}(v) \big| \le1,\;\forall{v}\in{V(G)}$.
A smooth orientation of an Eulerian graph $G$ is
called an \textit{Eulerian} orientation;
such an orientation satisfies
$d^{in}(v) = d^{out}(v),\;\forall{v}\in{V(G)}$.
Moreover, it can be seen that
for every {Eulerian} orientation,
for every subset of the nodes $W$,
the number of arcs leaving $W$ is equal to
the number of arcs entering $W$, see \cite[Ch.6.1]{Di}.
Therefore, every Eulerian orientation of a
$2k$-edge connected Eulerian graph results in a directed graph
that is $k$-edge connected.
An Eulerian orientation of an Eulerian graph
can be found by orienting the edges of
each connected component according to an Euler tour.
\subsection{Nash-Williams' results and possible extensions} \label{sec:NWtheorems}
A well-known result of Nash-Williams says that
the edges of a $k$-edge connected graph
can be oriented such that the resulting
directed~graph is $\myfloor{\frac{k}{2}}$-edge connected
\cite{N60}, \cite[Ch.9]{BM}.
A long-standing goal in the area is to
extend Nash-Williams' result to obtain analogous results for
other types of connectivity,
such as node connectivity and element connectivity,
see \cite{F-HC95,KS06,KL08,T89,T2015}.
\subsection{Our results} \label{sec:results}
We show that {every} Eulerian orientation
of the hypercube $Q_{2k}$
is strongly $k$-node connected;
recall that a directed~graph is called \textit{strongly $k$-node connected} if
it has $\ge k+1$ nodes and
the deletion of any set of $\leq (k-1)$ nodes results in a
strongly-connected directed~graph.
Let us mention that there are easy inductive constructions that prove
that there exists a ``good orientation'' for a hypercube of even degree;
we describe one such construction in Fact~\ref{f:exists-goodorient}.
For hypercubes of odd degree, the smoothness condition does not
guarantee ``good orientations;''
for example, there exist smooth orientations of $Q_3$ that are not
strongly connected.
\section{Preliminaries} \label{sec:prelims}
This section has some definitions and preliminary results.
Also, see \cite{Di} for standard definitions and notation.
The hypercube $Q_k$ is the Cartesian product of $k$ copies of $K_2$,
see \cite{wiki-HG}.
There are other constructions of $Q_k$, and we describe three of them.
\begin{enumerate}[(i)]
\item
Label $2^k$ nodes by $k$-bit binary strings, and
place an edge between two nodes whenever
their labels differ in exactly one bit
(i.e., the Hamming distance between the two strings is one).
\item
Label $2^k$ nodes by the $2^k$ subsets of a set with $k$ elements,
and place an edge between two nodes whenever
the two node labels (i.e., the two subsets) differ in a single element.
\item
Take two disjoint hypercubes $Q_{k-1}$, and place an edge
between corresponding pairs of nodes in the
two copies of $Q_{k-1}$; thus, the edges between the
two copies of $Q_{k-1}$ form a perfect matching.
\end{enumerate}
By a $d$-hypercube we mean a hypercube of degree~$d$.
For a node~set $S$ of a graph $G$, we use $N_G(S)$ to denote the
set of neighbors of $S$, thus,
$N_G(S) = \{ w\in V(G)-S \;:\; \exists v\in S ~\tn{such that}~ \{v,w\}\in E(G) \}$.
\begin{fact} \label{f:exists-goodorient}
For each integer $k\ge1$, there exists an Eulerian orientation of $Q_{2k}$
that is strongly $k$-node connected.
\end{fact}
\begin{proof}
Let $k\ge1$ be an integer.
We sketch an inductive construction that gives a strongly $(k+1)$-node~connected
Eulerian orientation for the hypercube $Q_{2k+2}$.
Observe that any Eulerian orientation of $Q_2$ (the 4-cycle) is
strongly 1-connected. Assume (by induction) that $Q_{2k}$ has a
strongly $k$-node~connected Eulerian orientation.
View the $(2k+2)$-hypercube as four $2k$-hypercubes (i.e., four copies of $Q_{2k}$)
together with $2^{2k}$ 4-cycles,
where each of these 4-cycles $C_i$
contains a distinct node $i$ of
the first copy of $Q_{2k}$
as well as the image of $i$ in each of the other three copies of $Q_{2k}$.
By the induction hypothesis,
there exists a strongly $k$-node~connected Eulerian orientation for $Q_{2k}$.
Fix such an orientation for each of the four copies of $Q_{2k}$.
Moreover, for each of the 4-cycles $C_i$,
fix any Eulerian orientation of $C_i$.
Let $D$ be the resulting directed~graph (i.e., orientation of $Q_{2k+2}$).
We claim that $D$ is strongly $(k+1)$-node~connected.
To see this, consider any set of nodes $Z$ of size $\le k$.
Suppose that one of the four copies of $Q_{2k}$ contains $Z$;
then it is clear that each of the other three copies of $Q_{2k}$
is strongly connected in $D-Z$, and hence,
(using the $2^{2k}$ oriented 4-cycles of $D$)
it can be seen that $D-Z$ is strongly connected.
Otherwise, each of the four copies of $Q_{2k}$ has $\le k-1$ nodes of $Z$,
hence, the removal of $Z$ from any one of the four copies of $Q_{2k}$
results in a strongly connected directed~graph;
again (using the $2^{2k}$ oriented 4-cycles of $D$),
it can be seen that $D-Z$ is strongly connected.
\end{proof}
\section{Eulerian orientations of $2k$-hypercubes} \label{sec:cube}
This section has our results and proofs.
In this section, we assume that $k$ is a positive integer.
{
\begin{theorem} \label{thm:oriG}
Let $G$ be a $2k$-regular $2k$-node~connected graph
such that for every set of nodes $S$ with $1\leq |S| \leq |V(G)|/2$
we have
$\displaystyle |N_G(S)| ~>~ \min\{ k^2-1,\; (k-1)(|S|+1) \}$.
Then every Eulerian orientation of $G$ is strongly $k$-node connected.
\end{theorem}
\begin{proof}
Let $D$ denote an arbitrary Eulerian orientation of $G$.
(In what follows, when we refer to \textit{the orientation} of an edge
of $G$ we mean the corresponding directed edge of $D$.)
By way of contradiction, suppose that $D$ is not strongly $k$-node
connected. Then there is a node~set $Z$ of size $\leq k-1$ whose
deletion from $D$ results in a directed~graph that has a partition
$(S, \bar{S})$ of its node~set $V(G)-Z$ such that
both $S,\bar{S}$ are nonempty and
the edges of $G-Z$ in this cut either
are all oriented from $S$ to $\bar{S}$ or
are all oriented from $\bar{S}$ to ${S}$.
We fix the notation such that $|S| \leq |\bar{S}|$.
(Now, observe that $|S|$ satisfies the condition stated in the hypothesis.)
Moreover, without loss of generality,
we assume that the edges are oriented from $S$ to $\bar{S}$
(the arguments are similar for the other case).
Observe that $G-Z$ has $\ge |N_G(S)| - |Z|$ edges in the cut $(S, \bar{S})$.
Thus, $D$ has $\ge |N_G(S)| - |Z|$ edges oriented out from $S$
(and into $\bar{S}$).
Consider the cut $(S, \bar{S} \cup{Z})$ of $G$, and
observe that it has $\leq \min\{ k|Z|,\; |S|\,|Z| \}$ edges oriented
into $S$ (and out of $Z$), because
(i)~all such edges are incident to nodes of $Z$ and only $k$
of the $2k$~edges incident to a node $w\in Z$ are oriented out of $w$;
(ii)~each such edge is incident to a node $s\in S$ and a node $w\in Z$
(and each pair $s,w$ contributes at most one such edge).
Thus, the cut $(S, \bar{S} \cup{Z})$ of $G$ has
$\ge |N_G(S)| - |Z| \ge |N_G(S)| - (k-1)$ edges oriented out of $S$ and
$\leq \min\{ k|Z|,\; |S|\,|Z| \} \leq \min\{ k(k-1),\; |S|(k-1) \}$
edges oriented into $S$;
the hypothesis (in the theorem) implies that the former quantity
is greater than the latter quantity.
This is a contradiction:
in an Eulerian orientation of an Eulerian graph,
every cut has the same number of outgoing edges and incoming edges.
\end{proof}
}
In the next subsection we show that hypercubes of even degree
satisfy all the conditions stated in Theorem~\ref{thm:oriG};
this gives our main result.
\subsection{Bounds for the ${2k}$-hypercube} \label{sec:iso}
The main goal of this subsection is to show that the hypercube $Q_{2k}$
satisfies the inequalities stated in Theorem~\ref{thm:oriG}.
Our analysis has two parts depending on the size $m$ of the set $S\subseteq V(Q_{2k})$
(in the statement of Theorem~\ref{thm:oriG});
the first part (Fact~\ref{f:cube-2}) applies for $1\le m\le 2k+1$ and it follows easily;
the second part (Fact~\ref{f:big-m}) applies for $2k+2\le m\le 2^{2k-1}$ and it follows by
exploiting properties of the hypercube.
In more detail, in the second part, we show that the
minimum of $|N_{Q_{2k}}(S)|$ over all sets $S\subseteq V(Q_{2k})$ of size $m$
(where $2k+2\le m\le 2^{2k-1}$) is $>k^2-1$;
our proof avoids elaborate computations by exploiting structural
properties of hypercubes;
a key point is to focus on a subgraph of the hypercube induced by
the set of binary strings of Hamming weight~$i$ and
the set of binary strings of Hamming weight~$i-1$
(see Claim~\ref{cl:lower-shadow} in the proof of Fact~\ref{f:big-m}).
We follow the notation of \cite{Bol} and use $b_v(m,Q_{2k})$ to denote
$\min \{ |N_{Q_{2k}}(S)| \;:\; S\subseteq V(Q_{2k}),\; |S|=m \}$;
thus, $b_v(m,Q_{2k})$ denotes the minimum
over all node~sets $S\subseteq V(Q_{2k})$ of size~$m$
of the number of neighbors of $S$.
For the sake of exposition,
we mention that
the node~sets $S$ with $|N_{Q_{2k}}(S)| = b_v(m,Q_{2k})$
(i.e., the minimizers of $b_v(m,Q_{2k})$)
are Hamming balls (see \cite[page~126]{Bol}), and
the formula for $b_v(m,Q_{2k})$ (stated in Theorem~\ref{thm:harper} below)
is obtained by computing the minimum number of neighbors of such sets.
Harper, see \cite{H66} and also see \cite{FF81}, proved the following result:
\begin{theorem}[Theorem~4, Ch.~16, \cite{Bol}] \label{thm:harper}
Every integer $m$, $1\le m \le 2^{2k}-1$, has a unique representation
in the form
\begin{align*}
m &~=~ \sum_{i=r+1}^{2k} {2k \choose i} + m', \quad 0<m' \leq {2k \choose r}, \\
m' &~=~ \sum_{j=s}^r {m_j \choose j}, \quad 1\le s\le m_s<m_{s+1}<\dots<m_r. \\
\tn{Moreover,~} & \\
b_v(m,Q_{2k}) &~=~ {2k \choose r} - m' + \sum_{j=s}^r {m_j \choose {j-1}}.
\end{align*}
\end{theorem}
\smallskip
\begin{remark}
To find the unique representation of $m$ stated in the above theorem,
we start by taking $r$ to be the largest integer $x\in \{1,\dots,2k\}$
such that $m \leq \sum_{i=x}^{2k} {{2k} \choose i}$, and then we fix
$m'=m-\sum_{i=r+1}^{2k} {{2k} \choose i}$; clearly, $m' \leq {2k \choose r}$.
Then we write $m'$ (uniquely) in the form $\sum_{j=s}^r {m_j \choose j}$;
for this, we take $m_r$ to be the largest integer $y$ such that
${y \choose r} \leq m'$; if $m'= {m_r \choose r}$, then we are done,
otherwise, we iterate by replacing $m'$ and $r$ by $m'-{m_r \choose r}$ and $r-1$, respectively, and then
applying the previous step.
For example, if $k=3$ and $m=17$,
then $r=4$, and $m={6\choose 6}+{6\choose 5}+m'$,
where $m'=10$ and $m'={5 \choose 4}+{4\choose 3}+{2 \choose 2}$.
\end{remark}
In what follows, we use the abbreviated notation $\phi(m)$ for $b_v(m,Q_{2k})$.
Now, our goal is to show that for $m=1,\dots,2^{2k-1}$,
we have $\phi(m) > \min\{ k^2-1,\; (k-1)(m+1) \}$.
This will imply that the hypercube $Q_{2k}$
satisfies the inequalities stated in Theorem~\ref{thm:oriG}.
We first consider the case $m=1,\dots,2k+1$.
We claim that $\phi(m) ~=~ 1 + (m/2) (4k-m-1)$.
This can be easily verified for $m=1$ and $m=2k+1$
(by applying Theorem~\ref{thm:harper}).
Now, suppose that $m=2,\dots,2k$;
then, observe that the unique representation of
$m$ (see Theorem~\ref{thm:harper}) is $1 + m'$,
where $m'=m-1$ and $r=2k-1$, and moreover,
$m' = {{2k-1} \choose {2k-1}} + {{2k-2} \choose {2k-2}} +\dots+
{{2k-m'} \choose {2k-m'}}$,
hence, $\phi(m) ~=~ (2k) - m' + \Big((2k-1)+(2k-2) +\dots+ (2k-m')\Big)
~=~ 1 + (m/2) (4k-m-1)$.
\begin{fact} \label{f:small-m}
\label{f:cube-2}
For each $m=1,\dots,2k+1$, we have
\[ \phi(m) ~>~ \min\{ (k-1)(m+1),\; (k-1)(k+1) \}.
\]
\end{fact}
\begin{proof}
{
We have $\phi(m) ~=~ 1 + (m/2) (4k-m-1)$, for $m=1,\dots,2k+1$.
Our goal is to show that
\[ \Delta ~=~ 1 + (m/2) (4k-m-1) - \min\{ (k-1)(m+1),\; (k-1)(k+1) \}
\]
is positive.
First, suppose that $m \leq k$.
Then, we have
\[
2\Delta ~=~ 2 + m(4k-m-1) - 2(k-1)(m+1) ~=~ m(k-m) + (k+1)(m-2) + 6.
\]
It can be seen that this quantity is $\ge4$ for $1\leq m\leq k$.
\big(For $2\leq m\leq k$, note that $m(k-m)\ge0$ and $(k+1)(m-2)\ge0$,
hence, $2\Delta\ge 6$; moreover, for $m=1$, we have $2\Delta = 4$.\big)
Next, suppose that $k \leq m$.
Then, we have
\[
2\Delta ~=~ 2 + m(4k-m-1) - 2(k-1)(k+1) ~=~ (2k+1-m)(m-k+1) + (m-1)(k-1) + 2.
\]
Clearly, this quantity is $\ge2$ for $1\leq k\leq m\leq 2k+1$.
}
\end{proof}
\begin{fact} \label{f:big-m}
For each $m=2k+2,\dots,2^{2k-1}$, we have
\[ \phi(m) ~>~ (k-1)(k+1).
\]
\end{fact}
\begin{proof}
{
Let $\alpha$ denote
$\sum_{i=0}^{k-1} {{2k} \choose i} = \sum_{i=k+1}^{2k} {{2k} \choose i}$;
observe that
$2^{2k} = \sum_{i=0}^{2k} {{2k} \choose i} = 2\alpha + {{2k} \choose k}$,
hence, $\alpha = \frac12 2^{2k} - \frac12 {{2k} \choose k}$.
Suppose that $m=2^{2k-1}$.
Then $m = \frac12 2^{2k} = \alpha + \frac12 {{2k} \choose k}$, hence,
$\sum_{i=k+1}^{2k} {{2k} \choose i} < m \leq
\sum_{i=k}^{2k} {{2k} \choose i}$.
Hence, for each $m=2k+2,\dots,2^{2k-1}$,
we have $k\leq r \leq 2k-2$
in the unique representation of $m$ given by Theorem~\ref{thm:harper},
i.e., we have
$\displaystyle m = \sum_{i=r+1}^{2k} {2k \choose i} + m', ~\tn{where}~
0 < m' \leq {2k \choose r}, ~\tn{and}~ k\leq r \leq 2k-2$;
moreover, we have
$m' = \sum_{j=s}^r {m_j \choose j}, \quad 1\le s\le m_s<m_{s+1}<\dots<m_r$.
We will use this notation in the rest of the proof.
To complete the proof, we examine two cases, namely,
(1)~$r=k$, and
(2)~$k+1 \le r \le 2k-2$.
\begin{description}
{
\item[Case 1: $r=k$.]
Since $m = \alpha + m' \le 2^{2k-1}$, we have
$1 \le m' \le 2^{2k-1} - \alpha = \frac12 {2k \choose k}$.
Hence, $\displaystyle \phi(m) =
{2k \choose r} - m' + \sum_{j=s}^r {m_j \choose {j-1}} \geq
{2k \choose r} - m' \geq
{2k \choose r} - \frac12 {2k \choose k} = \frac12 {2k \choose k}$.
Clearly, for $k=3$, we have $\frac12 {2k \choose k} > k^2-1$,
and for $k\ge3$, we have $\frac12 {2k \choose k} \ge \frac12 {2k \choose 3} > k^2-1$.
Moreover, for $k=1$, Fact~\ref{f:big-m} holds vacuously, and
for $k=2$, by the 4-node~connectivity of $Q_4$, we have
$\phi(m) \ge 4 > k^2-1=3,\; \forall m\in\{4,\dots,8\}$.
\item[Case 2: $k+1 \le r \le 2k-2$.]
Claim~\ref{cl:lower-shadow}, see below, states the key inequality
\[ m' < \sum_{j=s}^r {m_j \choose {j-1}}.
\]
This immediately implies that
\[ \phi(m) = {2k \choose r} - m' + \sum_{j=s}^r {m_j \choose {j-1}} >
{2k \choose r} \ge {2k \choose 2} =
k(2k-1) > k^2-1 ~~(\tn{for~} k\ge1),
\]
as required;
observe that the second inequality uses the upper~bound on $r$
(as well as the lower~bound $r\ge k+1\ge 2$).
}
\end{description}
\medskip
\begin{claim} \label{cl:lower-shadow}
For $r\ge k+1$,
we have
$\displaystyle \sum_{j=s}^r {m_j \choose {j-1}} > m'$.
\end{claim}
To prove this claim, it is convenient to view the $2^{2k}$ nodes
of $Q_{2k}$ as the $2^{2k}$ subsets of the set $\{1,2,\dots,2k\}$
(recall the second construction in Section~\ref{sec:prelims}).
Let $L_i \subset V(Q_{2k})$ denote the set of nodes corresponding to
$i$-element subsets of $\{1,2,\dots,2k\}$.
For $A \subseteq L_i$, let $\Gamma(A)$ denote $N_{Q_{2k}}(A)\cap{L_{i-1}}
~=~ \{ v \in L_{i-1} : \exists w \in A \mbox { such that } \{v,w\}
\in E(Q_{2k}) \}$; $\Gamma(A)$ is called the lower shadow of $A$.
(We mention that the lower shadow of $A$ is denoted by $\partial A$
in \cite{Bol}.)
Following \cite[Ch.5]{Bol}, let $\partial^{(r)}(m')$ denote
$\sum_{j=s}^r {m_j \choose {j-1}}$.
Let $M'\subseteq L_r$ consist of the first $m'$ nodes
(in colex order) of $L_r$, and let
$S' \subseteq L_{r-1}$ consist of the first $\partial^{(r)}(m')$ nodes
(in colex order) of $L_{r-1}$.
It is well known that the lower~shadow of the first $m'$ nodes (in
colex order) of $L_r$ consists of precisely the first
$\partial^{(r)}(m')$ nodes (in colex order) of $L_{r-1}$;
see \cite[pp.~28--32]{Bol}.
Thus, we have $\Gamma(M') = S'$.
Our key inequality can be restated as $\displaystyle m' = |M'| < |S'|$.
We will derive it by examining the subgraph $H$ of $Q_{2k}$ induced by
$M' \cup S'$.
Note that $H$ is a bipartite graph with the node bipartition $M'$, $S'$.
Observe that for each node of $M'$
(which corresponds to an $r$-element set),
there are exactly $r$ neighbors in $\Gamma(M') = S'$.
On the other hand, a node in $S'$
(which corresponds to an $(r-1)$-element set)
has $\leq 2k - r + 1 < r$ neighbors in $M'$
(the strict inequality follows from $k+1 \leq r$).
It follows that $|M'| < |S'|$.
This proves the inequality
$\displaystyle \sum_{j=s}^r {m_j \choose {j-1}} > m'$
of our claim.
}
\end{proof}
\medskip
Our main result follows from Theorem~\ref{thm:oriG}, Theorem~\ref{thm:harper},
the fact that $Q_{2k}$ is $2k$-regular and $2k$-connected,
and the inequalities stated above
(see Facts~\ref{f:cube-2},~\ref{f:big-m}).
\begin{theorem} \label{thm:main}
{Every} Eulerian orientation
of a hypercube of degree $2k$
is strongly $k$-node connected.
\end{theorem}
\bigskip
\bigskip
\noindent
{\bf Acknowledgments}:
We thank Zoltan Szigeti for several suggestions that
improved the paper, and we thank
Andre Linhares for his comments on a preliminary draft.
We are grateful to other colleagues and reviewers for their comments.
|
\section{The Farey determinants matrix}
The Farey sequence $F_N$ of order $N$ is an ascending sequence of irreducible
fractions between 0 and 1 whose denominators do not exceed $N$~\cite{hardy}.
Let $h_i/k_i\ <\ h_{i+1}/k_{i+1}$ be two
Farey neighbors then, $h_{i+1}k_i - h_ik_{i+1}=1$.
The next Farey fraction to appear between two Farey neighbors is given by
the mediant as
\begin{equation}\nonumber
\frac{h_i}{k_i} < \frac{h_i+h_j}{k_i+k_j} < \frac{h_j}{k_j}\ .
\end{equation}
We define the determinant of any two Farey fractions $h_i/k_i$ and $h_j/k_j$, both in $F_N$, as
\begin{equation}\nonumber
d_{ij}(N)=\begin{vmatrix}h_j&h_i\\k_j&k_i\end{vmatrix}= h_jk_i - h_ik_j\ .
\end{equation}
$d_{ij}(N)$ is also the numerator of the difference $h_j/k_j-h_i/k_i$,
so that $d_{ii}=0$ and $d_{i(i\pm1)}=\pm1$.
Note that $(N)$ is omitted from $d(N)$ when it is not necessary.
In the following a matrix operation is introduced
resembling the mediant that allows to iteratively compute $d(N)$
for increasing $N$.
Starting from $F_1=\{0/1,\ 1/1\}$, the skew-symmetric unitary matrix $d(1)$
is given by
\begin{equation}\nonumber
\begin{matrix}0&-1\\1 & 0\end{matrix}\ ,
\end{equation}
we insert one row and one column in the middle with values given
by the sum of the neighboring horizontal or vertical entries.
This is illustrated as follows,
\begin{center}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
0 & \w & -1 \\
\w & \w & \w \\
1 & \w & 0\\
};
\draw [BarreStyle=blue] (A-2-1.center) to (A-2-1.center) ;
\draw [BarreStyle=blue] (A-2-2.center) to (A-2-2.center) ;
\draw [BarreStyle=blue] (A-1-2.center) to (A-1-2.center) ;
\draw [BarreStyle=blue] (A-3-2.center) to (A-3-2.center) ;
\draw [BarreStyle=blue] (A-2-3.center) to (A-2-3.center) ;
\draw (A-1-1.north) edge[pil,bend right=-30,color=blue, line width=0.5mm, opacity=0.6] (A-1-2.north);
\draw (A-1-3.north) edge[pil,bend right=30,color=blue, line width=0.5mm, opacity=0.6] (A-1-2.north);
\draw (A-3-1.south) edge[pil,bend right=30,color=blue, line width=0.5mm, opacity=0.6] (A-3-2.south);
\draw (A-3-3.south) edge[pil,bend right=-30,color=blue, line width=0.5mm, opacity=0.6] (A-3-2.south);
\draw (A-1-1.west) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=-40,in=210] (A-2-1.west);
\draw (A-3-1.west) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=40,in=-210] (A-2-1.west);
\draw (A-1-3.east) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=40,in=-210] (A-2-3.east);
\draw (A-3-3.east) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=-40,in=210] (A-2-3.east);
\end{tikzpicture}
$\ \ \rightarrow \ \ $
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
\ww{0} & -1 & -1 \\
1 & & -1 \\
1 & 1 & 0\\
};
\draw [BarreStyle=blue] (A-2-2.center) to (A-2-2.center) ;
\draw (A-2-1.north) edge[pil,bend right=-30,color=blue, line width=0.5mm, opacity=0.6] (A-2-2.north);
\draw (A-2-3.north) edge[pil,bend right=30,color=blue, line width=0.5mm, opacity=0.6] (A-2-2.north);
\end{tikzpicture}
$\ \ \rightarrow \ \ $
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
\ww{0} & -1 & -1 \\
1 & 0 & -1 \\
1 & 1 & 0\\
};
\end{tikzpicture}
\end{center}
\begin{center}
$ \ \ \rightarrow \ \ $
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
0 & \w & -1 & \w & -1 \\
\w & \w &\w & \w & \w \\
1&\w & 0 & \w & -1 \\
\w& \w & \w & \w &\w \\
1& \w & 1 & \w & 0\\
};
\draw [BarreStyle=blue] (A-1-2.center) to (A-1-2.center) ;
\draw [BarreStyle=blue] (A-2-2.center) to (A-2-2.center) ;
\draw [BarreStyle=blue] (A-2-1.center) to (A-2-1.center) ;
\draw [BarreStyle=blue] (A-4-1.center) to (A-4-1.center) ;
\draw [BarreStyle=blue] (A-3-2.center) to (A-3-2.center) ;
\draw [BarreStyle=blue] (A-4-2.center) to (A-4-2.center) ;
\draw [BarreStyle=blue] (A-5-2.center) to (A-5-2.center) ;
\draw [BarreStyle=blue] (A-1-4.center) to (A-1-4.center) ;
\draw [BarreStyle=blue] (A-2-4.center) to (A-2-4.center) ;
\draw [BarreStyle=blue] (A-3-4.center) to (A-3-4.center) ;
\draw [BarreStyle=blue] (A-4-4.center) to (A-4-4.center) ;
\draw [BarreStyle=blue] (A-5-4.center) to (A-5-4.center) ;
\draw [BarreStyle=blue] (A-2-3.center) to (A-2-3.center) ;
\draw [BarreStyle=blue] (A-4-3.center) to (A-4-3.center) ;
\draw [BarreStyle=blue] (A-2-5.center) to (A-2-5.center) ;
\draw [BarreStyle=blue] (A-4-5.center) to (A-4-5.center) ;
\end{tikzpicture}
$\ \ \rightarrow \ \ $
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
0 & -1 & -1 & -2 & -1 \\
1 & 0 & -1 & -3 & -2\\
1 & 1 & 0 & -1 & -1 \\
2 & 3 & 1 & 0 & -1\\
1 & 2 & 1 & 1 & 0\\
};
\end{tikzpicture}
\end{center}
where $d(2)$ and $d(3)$ have been generated as the reader can verify. Between $d(N)$ and $d(N+1)$ new rows and columns should be inserted at the same positions as the new
fractions appearing between $F_N$ and $F_{N+1}$.
By construction the top row consists
of the numerators of Farey fractions with opposite sign.
The bottom row consists of numerators of Farey fractions
in reverse order. Similarly happens for the
first and last columns.
The denominators corresponding to the numeratos in the bottom row
can be obtained by subtracting the top row to the bottom row.
This is illustrated with $d(4)$ together with the Farey fractions corresponding
to the first and last elements in rows and columns,
\begin{equation}\nonumber
\begin{matrix}
& \frac{0}{1} & \frac{1}{4} &\frac{1}{3} &\frac{1}{2} & \frac{2}{3}& \frac{3}{4} &\frac{1}{1} & \\[0.3cm]
\scriptstyle0/1 & 0 &-1&-1 & -1 & -2 &-3 &-1 & \scriptstyle1/1 \\
\scriptstyle1/4 & 1 & 0 &-1 & -2 & -5 &-8 &-3& \scriptstyle3/4 \\
\scriptstyle1/3 & 1 & 1 & 0 & -1 & -3 &-5 &-2 & \scriptstyle2/3 \\
\scriptstyle1/2 & 1 & 2 & 1 & 0 & -1 & -2 &-1 & \scriptstyle1/2 \\
\scriptstyle2/3 & 2 & 5 & 3 & 1 & 0 &-1 &-1& \scriptstyle1/3 \\
\scriptstyle3/4 & 3 & 8 & 5 & 2 & 1 & 0&-1& \scriptstyle1/4 \\
\scriptstyle1/1 & 1 & 3 & 2 & 1 & 1 & 1& 0& \scriptstyle0/1 \\[0.3cm]
& \frac{1}{1} & \frac{3}{4} &\frac{2}{3} &\frac{1}{2} & \frac{1}{3}& \frac{1}{4} &\frac{0}{1}
\end{matrix}
\end{equation}
However it is not yet demonstrated that this process actually generates $d(N)$.
Let $\delta$ be a $n\times n$ matrix built following the above
procedure.
The element $\delta_{ij}$ is a linear combination
of the corresponding top and bottom elements $\delta_{i0}$
and $\delta_{in}$. This linear combination is the same
for all the elements in row $j$, so we can use the elements
$\delta_{0j}$ and $\delta_{nj}$ to reconstruct the linear
combination as follows
\begin{equation}\nonumber
\delta_{ij} = \delta_{i0}\frac{\delta_{nj}}{\delta_{n0}} +
\delta_{in} \frac{\delta_{0j}}{\delta_{0n}}
= -\delta_{i0}\delta_{nj} + \delta_{in} \delta_{0j}\ ,
\end{equation}
where we have used $\delta_{n0}=-1$ and $\delta_{0n}=1$, which are true
by construction.
The first row and column correspond
to the numerators of Farey fractions as $\delta_{i0}=-h_i$ and $\delta_{0j}=h_j$, respectively.
The last row and column correspond to the numerators of Farey fractions
in reverse order as $\delta_{in}=h_{n-i+1}=k_i-h_i$ and $\delta_{nj}=-h_{n-j+1}=-(k_j-h_j)$, respectively.
Therefore,
\begin{equation}
\delta_{ij} = -h_i(k_j-h_j) + (k_i-h_i)h_j
= -h_i k_j + k_ih_j =
\begin{vmatrix}h_j&h_i\\k_j&k_i\end{vmatrix}\ ,
\nonumber
\end{equation}
and $\delta_{ij}=d_{ij}$.
By construction $\delta$, or $d$, is
a square, skew-symmetric matrix with rank equal 2
as the new rows are a linear combination
of the neighboring rows for $N>2$.
$d$ is not only skew-symmetric but also
symmetric around the secondary diagonal and therefore
one could keep only one quarter of the matrix still being able to
generate higher order $d$'s.
Starting from $d(3)$,
\begin{center}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=0.6,line width=0.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{0 & -1 & -1 \\
1 & 0 & -1\\
1 & 1 & 0 \\
};
\draw [BarreStyle=blue] (A-1-1.north west) to (A-2-2.east) ;
\draw [BarreStyle=blue] (A-2-2.east) to (A-3-1.south west);
\draw [BarreStyle=blue] (A-3-1.south west) to (A-1-1.north west);
\end{tikzpicture}
\end{center}
we can proceed as before but only taking the quarter of the matrix
highlited with the blue triangle,
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
\w & 0 & \w \\
0 & \ww{1} & 1 \\
};
\end{tikzpicture}
$\ \ \rightarrow \ \ $
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
\w & \w & 0 & \w & \w \\
& 0 & \w & \w & \w\\
0 & \w &\ww{1} &\w & 1 \\
};
\draw [BarreStyle=blue] (A-2-2.center) to (A-2-2.center) ;
\draw [BarreStyle=blue] (A-3-2.center) to (A-3-2.center) ;
\draw [BarreStyle=blue] (A-2-3.center) to (A-2-3.center) ;
\draw [BarreStyle=blue] (A-2-4.center) to (A-2-4.center) ;
\draw [BarreStyle=blue] (A-3-4.center) to (A-3-4.center) ;
\draw (A-3-1.south) edge[pil,bend right=30,color=blue, line width=0.5mm, opacity=0.6] (A-3-2.south);
\draw (A-3-3.south) edge[pil,bend right=-30,color=blue, line width=0.5mm, opacity=0.6] (A-3-2.south);
\draw (A-3-3.south) edge[pil,bend right=30,color=blue, line width=0.5mm, opacity=0.6] (A-3-4.south);
\draw (A-3-5.south) edge[pil,bend right=-30,color=blue, line width=0.5mm, opacity=0.6] (A-3-4.south);
\draw (A-1-3.west) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=-40,in=210] (A-2-3.west);
\draw (A-3-3.west) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=40,in=-210] (A-2-3.west);
\end{tikzpicture}
$\ \ \rightarrow \ \ $
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
\w & \w & 0 & \w & \w \\
& 0 & 1 & \w & \w\\
0 & 1 &\ww{1} &2 & 1 \\
};
\draw [BarreStyle=blue] (A-2-4.center) to (A-2-4.center) ;
\draw (A-3-4.west) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=40,in=-210] (A-2-4.west);
\draw (A-2-3.north) edge[pil,bend right=0,color=blue, line width=0.5mm, opacity=0.6, out=40,in=-210] (A-2-4.north);
\end{tikzpicture}
\begin{equation}
\ \ \rightarrow \ \ \nonumber
\begin{matrix}
& & 0 & \\
& 0 & 1 & 3 \\
0& 1 & 1 & 2 & 1
\end{matrix}
\end{equation}
Note that to obtain the 3 we recall
the symmetries of the original matrix and we add the 2 and the 1
below and left of the 3, respectively.
In Section~\ref{index} $d$ is related to the index of
Farey fractions as defined in~\cite{index,generalizedindex}.
Section~\ref{gcd} illustrates an equality among the greatest common divisors
between elements in $d$ as presented in~\cite{partfransums}.
Section~\ref{maps} shows how some $d(N)$ contain other smaller $d(i)$
thanks to maps in~\cite{arxiv} that preserve the determinant of
two Farey fractions.
\section{$d_{ij}$ and $k$-indexes}\label{index}
The index $\nu(x_i)$ and the $k$-index $\nu_k(x_i)$ of the $i^{th}$ Farey
fraction in $F_N$
are introduced in~\cite{index} and~\cite{generalizedindex}, respectively.
We can relate them to the determinant matrix as
\begin{eqnarray}
\nu(x_i)&=&d_{(i-1)(i+1)}\ , \ {\rm for}\ 2 \leq i \leq |F_N|-1,\nonumber\\
\nu_k(x_i)&=&d_{(i-1)(i+k-1)}\ ,\ {\rm for}\ 2 \leq i \leq |F_N|-k+1.\nonumber
\end{eqnarray}
Being $\nu_k(x_i)$ a generalized definition of the index: $\nu_2(x_i)=\nu(x_i)$.
The highlighted diagonals in the following $d(3)$ matrix contain part of the $k$-indexes of fractions in $F_3$.
\begin{center}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{0 & -1 & -1 & -2 & -1 \\
1 & 0 & -1 & -3 & -2\\
1 & 1 & 0 & -1 & -1 \\
2 & 3 & 1 & 0 & -1\\
1 & 2 & 1 & 1 & 0\\
};
\draw [BarreStyle=blue] (A-2-1.center) to (A-5-4.center) ;
\draw [BarreStyle=green] (A-4-1.center) to (A-5-2.center) ;
\draw [BarreStyle=red] (A-3-1.center) to (A-5-3.center) ;
\draw [BarreStyle=yellow] (A-5-1.center) to (A-5-1.center) ;
\draw [] (A-2-1.west) node[left] {\color{blue}$\nu_1$} ;
\draw [] (A-3-1.west) node[left] {\color{red}$\nu=\nu_2$} ;
\draw [] (A-4-1.west) node[left] {\color{green}$\nu_3$} ;
\draw [] (A-5-1.west) node[left] {\color{yellow}$\nu_4$} ;
\end{tikzpicture}
\end{center}
The sum of the numbers in the red diagonal for $d(N)$ is easily obtained from Theorem~1 in~\cite{index} as
\[
\sum_{i=2}^{|F_N|-1} d_{(i-1)(i+1)} = 3(|F_N|-1)-2N-1\ .
\]
Any two adjacent rows (or columns) of $d$ correspond to
ordered lists of numerators and denominators of
Farey neighbours, illustrated as follows.
\begin{center}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=.4,line width=4.6 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{0 & -1 & -1 & -2 & -1 \\
1 & 0 & -1 & -3 & -2\\
1 & 1 & 0 & -1 & -1 \\
2 & 3 & 1 & 0 & -1\\
1 & 2 & 1 & 1 & 0\\
};
\draw [BarreStyle=green] (A-2-1.center) to (A-2-5.center) ;
\draw [BarreStyle=green] (A-3-1.center) to (A-3-5.center) ;
\end{tikzpicture}
$\displaystyle \ \ \rightarrow \ \ \left\{ \frac{1}{1},\frac{0}{1},\frac{-1}{0},\frac{-3}{-1},\frac{-2}{-1} \right\}$
\end{center}
This is easy to demonstrate by realizing that the first and
last elements of the rows fulfill
\[d_{0i}d_{n(i+1)} - d_{0(i+1)}d_{ni}=1\ ,
\]
as $d_{0i}=h_i$ and $d_{ni}=-(k_i-h_i)$.
Hence $d_{0i}/d_{0(i+1)}$ and $d_{ni}/d_{n(i+1)}$ are Farey neighbors. Therefore, the numbers within the rows
constitute Farey fractions as they are computed as mediants. Adjacent fractions are also Farey neighbors.
This is equivalent to Lemma~1 in~\cite{generalizedindex} expressed here as
\begin{equation}\nonumber
\begin{vmatrix}d_{ij}&d_{(i+1)j}\\ d_{i(j+1)}&d_{(i+1)(j+1)}\end{vmatrix}=1\ , {\rm\ for\ }i\leq|F_N|-1\ , \
j\leq|F_N|-1\ .
\end{equation}
\section{Great common divisors among $d_{ij}$}\label{gcd}
According to Theorem 1 in~\cite{partfransums}
the following equality holds between the great common divisors
of elements $d_{pq}$ with $p$ and $q$ in $\{i > j > k\}$,
\[ {\rm gcd}\left(d_{kj}, d_{ki}\right) = {\rm gcd}(d_{kj}, d_{ji}) =
{\rm gcd}(d_{ki}, d_{ji})\ .
\]
This property is illustrated using $d(6)$ as follows,
\begin{center}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=0.5,line width=0.5 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
0& -1& -1& -1& -1& -2& -1& -3& -2& -3& -4& -5& -1\\
1& 0& -1& -2& -3& -7& -4& -13& -9& -14& -19& -24& -5\\
1& 1& 0& -1& -2& -5& -3& -10& -7& -11& -15& -19& -4\\
1& 2& 1& 0& -1& -3& -2& -7& -5& -8& -11& -14& -3\\
1& 3& 2& 1& 0& -1& -1& -4& -3& -5& -7& -9& -2\\
2& 7& 5& 3& 1& 0& -1& -5& -4& -7& -10& -13& -3\\
1& 4& 3& 2& 1& 1& 0& -1& -1& -2& -3& -4& -1\\
3& 13& 10& 7& 4& 5& 1& 0& -1& -3& -5& -7& -2\\
2& 9& 7& 5& 3& 4& 1& 1& 0& -1& -2& -3& -1\\
3& 14& 11& 8& 5& 7& 2& 3& 1& 0& -1& -2& -1\\
4& 19& 15& 11& 7& 10& 3& 5& 2& 1& 0& -1& -1\\
5& 24& 19& 14& 9& 13& 4& 7& 3& 2& 1& 0& -1\\
1& 5& 4& 3& 2& 3& 1& 2& 1& 1& 1& 1& 0\\
};
\draw [blue,line width=0.5mm,opacity=0.6] (A-10-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-9-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-8-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-7-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-6-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-5-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-4-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-3-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-3) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-4) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-5) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-6) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-7) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-8) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-9) circle (1.5ex);
\draw [BarreStyle=blue] (A-9-2.east) to (A-10-9.north) ;
\draw [BarreStyle=blue] (A-8-2.east) to (A-10-8.north) ;
\draw [BarreStyle=blue] (A-7-2.east) to (A-10-7.north) ;
\draw [BarreStyle=blue] (A-6-2.east) to (A-10-6.north) ;
\draw [BarreStyle=blue] (A-5-2.east) to (A-10-5.north) ;
\draw [BarreStyle=blue] (A-4-2.east) to (A-10-4.north) ;
\draw [BarreStyle=blue] (A-3-2.east) to (A-10-3.north) ;
\end{tikzpicture}
\end{center}
where the number in the blue circle together with another 2 numbers
connected by any blue line define a triplet of numbers for which
the gcd's computed for all possible combinations within the triplet are equal,
e.g.,
\[
{\rm gcd}(14,2)={\rm gcd}(2,8)={\rm gcd}(14,8)=2.
\]
Let the blue circled number be on the antidiagonal $k=n-j+1$, assuming $d$ is
a $n\times n$ matrix. This is illustrated for $d(5)$ as follows,
\begin{center}
\renewcommand\ww[1]{\makebox[0.7em]{$#1$}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=0.5,line width=0.5 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
0 \\
1 & 0 \\
1 & 1 & 0\\
1 & 2 & 1 & 0\\
2 & 5 & 3 & 1 & 0 \\
1 & 3 & 2 & 1 & 1 & 0 \\
3 & 10 & 7 & 4 & 5 & 1 & 0\\
2 & 7 & 5 & 3 & 4 & 1 & 1 & 0\\
3 & 11 & 8 & 5 & 7 & 2 & 3 & 1 & 0\\
4 & 15 & 11 & 7 & 10 & 3 & 5 & 2 & 1 & 0 \\
\ww{1} & \ww4 & \ww3 & \ww2 & \ww3 & \ww{1} & \ww2 & \ww{1} & \ww1 & \ww1 & \ww0\\
};
\draw [blue,line width=0.5mm,opacity=0.6] (A-10-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-9-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-8-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-7-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-6-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-5-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-4-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-3-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-3) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-4) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-5) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-6) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-7) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-8) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-10-9) circle (1.5ex);
\draw [BarreStyle=blue] (A-9-2.east) to (A-10-9.north) ;
\draw [BarreStyle=blue] (A-8-2.east) to (A-10-8.north) ;
\draw [BarreStyle=blue] (A-7-2.east) to (A-10-7.north) ;
\draw [BarreStyle=blue] (A-6-2.east) to (A-10-6.north) ;
\draw [BarreStyle=blue] (A-5-2.east) to (A-10-5.north) ;
\draw [BarreStyle=blue] (A-4-2.east) to (A-10-4.north) ;
\draw [BarreStyle=blue] (A-3-2.east) to (A-10-3.north) ;
\end{tikzpicture}
\end{center}
The symmetry of $d$ around
the antidiagonal $d_{(n-j+)i}=d_{j(n-i+1)}$ implies
\[ {\rm gcd}(d_{(n-j+1)j}, d_{(n-j)i}) = {\rm gcd}(d_{(n-j)j}, d_{(n-j)(n-i+1)}) =
{\rm gcd}(d_{(n-j)i}, d_{(n-j)(n-i)})\ ,
\]
so the previous property can be seen as applied to a single column (or row),
as shown in the following illustration.
\begin{center}
\renewcommand\ww[1]{\makebox[0.5em]{$#1$}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=0.5,line width=0.5 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
0 \\
1 & 0 \\
1 & 1 & 0\\
1 & 2 & 1 & 0\\
2 & 5 & 3 & 1 & 0 \\
1 & 3 & 2 & 1 & 1 & 0 \\
3 & 10 & 7 & 4 & 5 & 1 & 0\\
2 & 7 & 5 & 3 & 4 & 1 & 1 & 0\\
3 & 11 & 8 & 5 & 7 & 2 & 3 & 1 & 0\\
4 & 15 & 11 & 7 & 10 & 3 & 5 & 2 & 1 & 0 \\
\ww{1} & \ww4 & \ww3 & \ww2 & \ww3 & \ww{1} & \ww2 & \ww{1} & \ww1 & \ww1 & \ww0\\
};
\draw [blue,line width=0.5mm,opacity=0.6] (A-10-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-9-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-8-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-7-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-6-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-5-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-4-2) circle (1.5ex);
\draw [red,line width=0.5mm,opacity=0.6] (A-3-2) circle (1.5ex);
\draw [blue,line width=0.5mm, opacity=0.6] (A-9-2.east) to [bend right=80] (A-3-2.east) ;
\draw [blue,line width=0.5mm, opacity=0.6] (A-8-2.east) to [bend right=80] (A-4-2.east) ;
\draw [blue,line width=0.5mm, opacity=0.6] (A-7-2.east) to [bend right=50] (A-5-2.east) ;
\path
(A-6-2.east) edge [ -,color=blue, line width=0.5mm, opacity=0.6] node {} (A-6-3.west) node {} (A-6-2.north);
\end{tikzpicture}
\end{center}
Note that the 3 is paired to himself, and the property above reads:
\[
{\rm gcd}(3,3)={\rm gcd}(15,3)=3\ .
\]
In general, assuming $k=n-j+1$ and $i=(n+1)/2$ implies $(n-i+1) = (n+1)/2$ and
therefore
\[ {\rm gcd}\left(d_{k\, (n-k+1)}, d_{k\, \frac{n+1}{2}}\right) = d_{k\, \frac{n+1}{2}} \ ,
\]
which implies that column-by-column the element in the antidiagonal is a multiple
of the element in the middle row. This last property is also easily shown from the properties of the Farey sequence,
as $d_{k\, (n-k+1)}=b_k(b_k-2a_k)$, with $a_k/b_k$ being the $k^{\rm th}$ element in $F_N$
and $d_{k\, \frac{n+1}{2}}=b_k-2a_k$, since $a_{\frac{n+1}{2}}/b_{\frac{n+1}{2}}=1/2$.
\section{Maps preserving the determinants matrix}\label{maps}
Let $F_{N}^{1/a,\, 1/b}$ be the subsequence of $F_N$ defined
as all the fractions of $F_N$ in $[1/a,\ 1/b]$ with $1\leq b \leq a \leq N$.
In~\cite{arxiv} it is demonstrated that
the map
\begin{equation}
F_i \rightarrow F_{N}^{1/q,\, 1/(q-1)}\ , \ \ \ \ \ \ \frac{h}{k} \mapsto \frac{k}{kq-h }\label{map}\ ,
\end{equation}
is bijective between $F_i$ and $F_{N}^{1/q,\, 1/(q-1)}$ when $N$ is a multiple of $i(i+1)$ and $N/(i+1)< q \leq N/i$. It is straight forward to show that this map preserves
the determinant, meaning that
\begin{equation}\nonumber
\begin{vmatrix}h& h'\\k &k'\end{vmatrix} =
\begin{vmatrix}k &k'\\ kq-h & k'q-h'\end{vmatrix}\ ,
\end{equation}
for $h/k$ and $h'/k'$ belonging to $F_i$.
Therefore $d(N)$
contains $d(i)$ as a matrix block. In other words,
$d(i)$ is contained
$p$ times in $d(i(i+1)p)$.
As an illustration a portion of $d(30)$ is shown containing $d(5)$,
\begin{center}
\begin{tikzpicture}[baseline=(A.center)]
\tikzset{BarreStyle/.style = {opacity=0.5,line width=0.5 mm,line cap=round,color=#1}}
\matrix (A) [matrix of math nodes, nodes = {node style ge},,column sep=0 mm]
{
F_{5} & F_{30} &d & & & & & \\
&\scriptstyle4/25 & 0 \\
\scriptstyle0/1 &\scriptstyle1/6 &1 & 0 \\
\scriptstyle1/5 &\scriptstyle5/29 & 9 & 1& 0 \\
\scriptstyle1/4 &\scriptstyle4/23 & 8 & 1 & 1& 0 \\
\scriptstyle1/3 &\scriptstyle3/17 & 7 & 1 & 2 & 1& 0 \\
\scriptstyle2/5 &\scriptstyle5/28 &13 & 2 & 5 & 3 & 1& 0 \\
\scriptstyle1/2 &\scriptstyle2/11 & 6 & 1 & 3 & 2 & 1 & 1& 0 \\
\scriptstyle3/5 &\scriptstyle5/27 &17 & 3 & 10 & 7 & 4 & 5 & 1& 0 \\
\scriptstyle2/3 &\scriptstyle3/16 &11 & 2 & 7 & 5 & 3 & 4 & 1 & 1& 0 \\
\scriptstyle3/4 &\scriptstyle4/21 &16 & 3 & 11 & 8 & 5 & 7 & 2 & 3 & 1& 0 \\
\scriptstyle4/5 &\scriptstyle5/26 &21 & 4 & 15 & 11 & 7 & 10 & 3 & 5 & 2 & 1& 0 \\
\scriptstyle1/1 &\scriptstyle1/5 & 5 & 1 & 4 & 3 & 2 & 3 & 1 & 2 & 1 & 1 & 1& 0 \\
&\scriptstyle6/29 &34 & 7 & 29 & 22 & 15 & 23 & 8 & 17 & 9 & 10 & 11 & 1& 0 \\
};
\draw [BarreStyle=blue] (A-2-3.north east) to (A-13-4.south west) ;
\draw [BarreStyle=blue] (A-13-4.south west) to (A-14-15.north east);
\draw [BarreStyle=blue] (A-2-3.north east) to (A-14-15.north east);
\end{tikzpicture}
\end{center}
the left columns show the corresponding Farey fraction in $F_{5}$ and $F_{30}$ according
to the map in Eq.~(\ref{map}) with $q=6$. The blue triangle highlights $d(5)$.
|
\section*{Background}
One of the most efficient way to analyze a complex network is by partitioning the nodes into blocks, shedding light on the internal structure of the network. For example, community detection seeks to decompose a network into blocks of nodes with many internal edges, and few edges falling between the blocks. This approach allows analyzing large networks as a sum of dense but weakly interconnected subnetworks. In the last couple of decades, the wider and wider availability of various large network-shaped data has pushed the need for new formalisations to the task of community detection, and more efficient algorithms~\cite{fortunato2010}.
Often, one same situation can be modelled by several networks of interests, where nodes represent entities at different levels of abstractions. For example, at one level a node may represent an individual person, and at another level it may represent the aggregation of several persons sharing an attribute, for example belonging to the same age class or living in the same municipality. In this case, the edge between two aggregation classes is typically weighted as the sum of the weight of all edges linking the individuals across the two classes. The reasons for considering an aggregated network rather than a disaggregated one are many. For instance only the aggregated network may be available to the researcher due to privacy reasons, or due to limited resources (e.g. only aggregate flows may be accessible to the measurement, or the disaggregated network may be too large to handle for a given community detection algorithm). The aggregated network may also be more relevant for a given analysis, because the aggregation removes possible noise present at the individual level, and creates statistically robust entities.
In all these situations, it is natural to wonder whether the communities computed on different levels of aggregation will be comparable in any way. This is the question that we explore in this article.
That some statistical patterns, for instance correlations, may differ starkly when computed either on a dataset or on an aggregated version of the same dataset is well known in statistical sciences. Extrapolating observations on categories of individuals to the individuals themselves is generically called an \emph{ecological fallacy}, with Simpson's paradox \cite{simpson1951interpretation,simpson_first_explanation} or Robinson's paradox~\cite{robinson_original_work} as well-known examples. In geographical sciences, a particular form of such fallacy is called the Modifiable Areal Unit Problem (MAUP). In the earliest detected occurrence of MAUP, Gelhke and Biel \cite{gehlke1934certain} showed that the value of the correlation coefficient of geolocalised features was influenced by the size of the spatial units used in their analyses. Openshaw further showed that the results of quantitative spatial models and statistics may depend highly on the size and shape of the basic spatial units used \cite{openshaw1984modifiable}. This problem has been broadly studied and is the object of extensive literature, see \cite{wong} for a review. The \emph{atomic fallacy} can be seen as the bias generated by extrapolating patterns present at the individual level to the level of the group to which those individuals or their geographical entities belong.
To the best of our knowledge, however, the impact of atomic and ecological fallacies on community detection has not been considered in the literature, despite its high relevance in practical applications. This is a gap that we aim to fill in the present paper, by measuring quantitatively the impact of node aggregation on the community structure in networks. We first bring a theoretical argument showing that some community detection methods are more robust than others to node aggregation, in that whenever the communities found optimal by the method on the finer network happen to be unambiguously aggregated, the aggregated communities are also found optimal by the community detection method.
Then, we introduce the aggregability index, a quantitative proxy for the robustness of the community structure of a given network with respect to given node aggregation classes
We illustrate our considerations on two real-life examples. Both compare networks of places, where the nodes are geographical areas, and the edges represent interactions between areas. In these examples it is easy to generate a series of aggregated networks by merging the places into larger and larger areas, either according to an administrative hierarchy (districts, municipalities, counties, etc.) or according to coarser and coarser square grids. In the first real-life example, each node is the mean position of a Twitter user in Belgium, and edges count the reply-to tweets between two such users. Aggregated versions of this network are produced by merging the positions into larger and larger administrative units or grid cells, and merging the edges accordingly. In the second example, the nodes are mobile phone towers in and around Brussels (the capital city of Belgium), and the edges count the number of phones calls between two towers. Aggregated versions of this network are also produced by successive merging of the nodes and edges similarly to the Twitter case. The quantitative tools we introduce allow to observe that the Twitter networks exhibit significantly different community structures at different levels of aggregation, while the mobile phone networks' communities are relatively insensitive to aggregation.
\section*{Edge-counting objective functions for optimal partitioning}
Partitioning the nodes of a network is often performed through optimising an objective function, and assigning a real number to each partition. We characterise a class of objective functions that preserve optimality of the community partition under aggregation whenever possible, as we now define.
Assume we want to detect communities in a weighted, undirected graph $G$, understood as a (non-overlapping) partition $\mathcal{C}$ of the nodes of $G$. Let us assume that we are also interested in optimising a certain criterion, capturing structural patterns of interest, typically high density of edges inside the communities and low density across communities. Some other criteria are also possible as, for instance one may want to detect core-periphery structure or general stochastic block models
\cite{cucuringu2014detection, newman2015generalized, goldenberg2010survey}. We want to underline here that there is a variety of possible criteria whose relevance is strongly dependent on the network and the application. For instance, some methods integrate a resolution parameter that imposes a preference for small or large communities \cite{reichardt2006,delvenne2013stability}. Some methods based on comparison with a generative model for the graph are highly dependent on the choice of the model \cite{peel2016ground}. Even more broadly, different goals for community detection may lead to entirely different objective functions~\cite{schaub2016many,chakraborty2017metrics}. As many of those methods proceed by optimising a `goodness' criterion, we talk of \textquotedblleft the optimal partition\textquotedblright to denote the communities found to be optimal for the criterion of interest --- we suppose for simplicity that the partition is unique and can be discovered effectively, although in practice most algorithms are only heuristics.
Assume moreover that a graph $G'$ is obtained from the aggregation of the nodes and edges of $G$, following a partition $\mathcal{P}$ of the nodes of $G$. In other words, if $\mathcal{P}$ partitions nodes of $G$ into $k$ ``aggregation classes'', then $G'$ has $k$ nodes. The weight of the edge (if any) between node $i$ and node $j$ of $G'$ is the sum of the weights of all edges of $G$, between nodes in the corresponding aggregation class $I$ and aggregation class $J$ of the partition $\mathcal{P}$. In particular, node $i$ of $G'$ has a self-loop aggregating the weight of all the edges inside the corresponding aggregation class $I$, representing the interactions between different nodes of the same class. In summary, the weights in $G'$ are given by
\begin{equation}
w_{ij}=\sum_{u \in I, v \in J, u \neq v} w_{uv},
\end{equation}
where $w_{uv}$ is the weight of the edge between nodes $u$ and $v$, in the initial disaggregated network $G$. In all cases, we insist that node aggregation as considered in this paper leads to weighted aggregated graphs, even when the original graphs are unweighted.
In general, we want to understand the relationship between the communities of $G$ and $G'$. Those communities takes place on different graphs, thus a direct comparison is not possible. We can nevertheless ``lift'' the communities of $G'$ back to $G$, by replacing each node in $C'$ with its aggregation class in $G$. Indeed a community of $G'$ is a set of nodes of $G'$, each of which represents an aggregation class of $G$. To state it more formally, if $f_{\mathcal{P}}:\mathrm{Nodes}(G) \to \mathrm{Nodes}(G')$ is the aggregation function relating every node of the original graph $G$ to its corresponding node in the aggregated graph $G'$, then a community $C' \subset \mathrm{Nodes}(G')$ is lifted back to $f^{-1}(C')$, which is a notation for the set $\{x \in \mathrm{Nodes}(G) : f(x) \in C'\}$. Doing so for each community of $G'$, we obtain a partition of the nodes of $G$, which we denote $f^{-1}(\mathcal{C'})$, and which we call the ``lifting'' of $\mathcal{C'}$. This partition can now be compared to $\mathcal{C}$.
A specific case of interest is when we want to know the communities of $G$ while we only have access to $G'$. Clearly the best scenario is when the aggregation classes are subsets of the optimal communities in $G$, i.e. when each community is a union of aggregation classes. In this case, the aggregation transforms unambiguously the optimal community partition $\mathcal{C}$ of $G$ into a (possibly non-optimal) community partition $\mathcal{C}'$ of $G'$. From the knowledge of the community partition $\mathcal{C}'$ of $G'$, it is then possible to recover $\mathcal{C}$ as $f^{-1}(\mathcal{C'})$, the lifting of $\mathcal{C'}$.
If, moreover the community structure $\mathcal{C}'$ is also optimal in $G'$ then we have a natural way to recover the community structure of the original $G$: first compute $\mathcal{C'}$ as the optimal community structure of $G'$ then lift it to $\mathcal{C}$. However, whether $\mathcal{C'}$ is indeed optimal for $G'$ depends on the criterion used to define `(optimal) communities'.
This can be guaranteed if the objective function, evaluated on a given graph $G$ and a proposed community partition $\mathcal{C}$, only depends on the graph $G''$, defined as the graph obtained by aggregating $G$ with respect to the partition $\mathcal{C}$ of the nodes. In other words, we require that the objective function depends only on the total weight of all edges between any pair of communities (including from a community to itself), but not on the way those edges are distributed inside a community or between communities. We call such a function an \emph{edge-counting function}.
This natural result is proved simply. Since we assume that $G'$ is obtained from $G$ by aggregation with respect to a partition $\mathcal{P}$, and that the partition $\mathcal{C}$ is coarser than $\mathcal{P}$, then the aggregation of $G'$ with respect to $\mathcal{C}'$ coincides with the aggregation of $G$ with respect to $\mathcal{C}$. Therefore, the
edge-counting objective function takes the same value for $(G,\mathcal{C})$ and $(G',\mathcal{C}')$. Thus if $\mathcal{C}$ is optimal for $G$ then $\mathcal{C}'$ is also optimal for $G'$.
Despite its simplicity, this first result suggests that some methods of the literature are more appropriate than others in presence of node aggregation. Such edge-counting
criteria include modularity~\cite{newman2004}, the Hamiltonian given by Potts models~\cite{reichardt2007partitioning},
linearised partition stability~\cite{delvenne2013stability}, Infomap's description length~\cite{rosvall2008}, conductance\cite{Kannan07},
Normalised Cuts~\cite{Shi2000}, and their natural extension to weighted graphs.
On the other hand, methods based on counting paths rather than edges depend on the way edges are distributed inside a community and not only the
number of edges or total weights. Such methods include Markov clustering\cite{vanDongen00}, Walktrap~\cite{LatapyPons08}, partition stability~\cite{delvenne2013stability}, etc., and should be used with the greatest caution in case of aggregated data.
\section*{Different aggregations lead to different community structures}
Even an edge-counting objective function cannot preserve the community structure in the context of arbitrary aggregation classes. Assume for instance, that the aggregation classes are chosen randomly, every node being attributed uniformly randomly to one of the classes. Then, it is reasonable to assume that the aggregated graph will behave like a complete graph with all edges of similar weight. Such a graph is expected to exhibit either no community structure, or communities created only by the small random fluctuations in the weights, retaining no information from the optimal communities of $G$.
One can also generate examples where well chosen classes generate a graph with entirely different, yet relevant, community structure. See Fig.~\ref{fig:toy} for an illustration on a toy 4-node network, and two aggregated 2-node networks, whose communities lift back to different community structures on the fine-scale network. These partitions may or may not coincide with the community structure computed directly on the fine-scale 4-node network---depending on the criterion for detecting communities. Here we do not specify an explicit community detection criterion, but it is reasonable to assume that if the self-loop (omitted for clarity in Fig.~\ref{fig:toy}) on each node of a 2-node aggregated network is heavy enough compared to the internode link, then the criterion will find the two 1-node-community partition as optimal.
Suppose for example that on each aggregating partition (either by colour or by shape) of the figure, the community detection criterion is such that the 2-community partition is optimal. Suppose as well that the community partition criterion finds the same-colour communities to be optimal on the 4-node network. We see that this partition is `orthogonal' to the partition that would be lifted from the communities on the (bottom) same-shape aggregated network. Yet all community partitions are `correct' and relevant for their respective networks: one should refrain from thinking that the aggregation leads to the `wrong' communities.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6 \textwidth]{fig1f.pdf
\end{center}
\caption{Community detection over two examples of aggregations of a same 4-node network. Self-loops in aggregated networks are omitted for clarity. We assume that the community
detection criterion is such that each aggregated network admits the trivial two-community partition as optimal. The community structure on each aggregated network lifts to two
possible partitions on the 4-node network. The community structure could be coincide with either of the two, or with that 4-community partition, according to the respective weight
of the edges. On the depicted example, it may coincide with the same-colour communities.}
\label{fig:toy}
\end{figure}
A more general example is built with the Kronecker product of an $n_1$-node graph $G_1$ and an $n_2$-node graph $G_2$. In the product graph $G_1 \otimes G_2$, whose node set is the Cartesian product of the two individual node sets, a node $(i,j)$ is connected to the node $(i',j')$ if $i$ and $i'$ are neighbours in $G_1$, as well as $j$ and $j'$ in $G_2$. If the graphs are weighted, then the weight on an edge in the product graph is simply the product of the weights in the corresponding edges in $G_1$ and $G_2$. The product graph can be aggregated in two natural ways, in one that retrieves $G_1$ as aggregated graph, and another one that retrieves $G_2$. Assume that the fine-grained network is $G_1 \otimes G_2$. Both aggregated graphs $G_1$ and $G_2$ may have a significant community structure, thus the community detection on both aggregations will provide interesting, distinct insights on the underlying fine-grained network.
A real-life analogy would involve, for instance, aggregating a social network according either to geographical location (e.g., counties), or to age class: both may exhibit relevant community structures explaining on the one hand which counties interact together, and the other hand which age classes interact together. Both community structures can be lifted back to the social network. If all age groups are equally present in each location, then those two partitions of the social network, although both interesting in their own rights, are `orthogonal' to each other as in the examples above. Thus at least one of them will differ sharply from communities found directly on the social network.
In summary, different aggregations of the original network may induce community structures on the original network that are completely disaligned with one another, without necessarily being `wrong', and that are either similar or dissimilar to the community structure computed directly on the original graph.
\section*{The aggregability index}
Between the two extremes situations where the aggregating partition is completely aligned with the community structure in $G$, or completely orthogonal to it, one finds intermediate situations where node aggregation is expected to perturb more or less the community detection.
We propose a metric capable of capturing to what extent node aggregation will preserve community detection by introducing the \emph{aggregability index}, $\eta$, as the fraction of information required to identify the community of a randomly chosen node, that is provided by the knowledge of its aggregation class:
\begin{equation}
\eta = \frac{I(\mathcal{C} ; \mathcal{P})}{H(\mathcal{C})}.
\label{eq1}
\end{equation}
Here $H(\mathcal{C})$ is the Shannon entropy of the community partition, defined in the following way. As a thought experiment, pick a node uniformly at random in $G$. The community of the node is a random variable with Shannon entropy $H(\mathcal{C}) \triangleq - \sum_{C\in \mathcal{C}} P(C) \log P(C)$, with probability $P(C)$ of a community $C$ being proportional to its number of nodes. Similarly, $I(\mathcal{C} ; \mathcal{P})$ is the Shannon mutual information between the community in the partition $\mathcal{C}$ and the random aggregation class in $\mathcal{P}$ of a randomly picked node of $G$.
Our newly-defined aggregability index, $\eta$, ranges from $0$ to $1$. In the $\eta=0$ limit, the aggregation classes are independent from the communities in $\mathcal{C}$, which implies that each node is aggregated with nodes from other communities. In particular, the community structure $\mathcal{C'}$ that we may compute in the aggregated network, once lifted back to the initial graph $G$, form communities which are unions of aggregation classes, thus also independent from the communities in $\mathcal{C}$. In short, using the notations above, we can write $I(\mathcal{C} ; f^{-1}(\mathcal{C'}))=0$.
In the $\eta=1$ limit, the aggregation classes are subset of the communities, thus any edge-counting criterion will preserve the community structure. In short, we write $\mathcal{C} = f^{-1}(\mathcal{C'})$.
Between these extreme situations, the lifted communities $f^{-1}(\mathcal{C'})$ are neither independent nor fully aligned with $\mathcal{C'}$. In this case, we observe, due to the fact that $f^{-1}(\mathcal{C'})$ is a coarser partition than the aggregation partition $f^{-1}(\mathcal{C'})$, that $I(\mathcal{C} ; f^{-1}(\mathcal{C'})) \leq I(\mathcal{C} ; \mathcal{P})$ (in application of the so-called data-processing inequality in information theory). In summary, we have in all cases:
\begin{equation}\label{eq:eta-vs-I}
\eta \geq \frac{I(\mathcal{C} ; f^{-1}(\mathcal{C'})) }{H(\mathcal{C})},
\end{equation}
which confirms $\eta$ as a `best-case' estimate of the closeness between the community structure on the original graph $G$ and its aggregation $G'$.
We may relax (\ref{eq:eta-vs-I}) to make it more symmetric in $\mathcal{C}$ and $\mathcal{C'}$, by increasing the denominator:
\begin{equation}\label{eq:eta-vs-I-2}
\eta \geq \frac{I(\mathcal{C} ; f^{-1}(\mathcal{C'})) }{H(\mathcal{C})+H(f^{-1}(\mathcal{C'}))},
\end{equation}
which can be written equivalently as
\begin{equation}\label{eq:eta-vs-NMI}
\eta \geq \frac{1}{2} \text{NMI}(\mathcal{C},f^{-1}(\mathcal{C'})),
\end{equation}
where NMI denotes a popular way to measure the similarity between two partitions, explained in the Methods (see Eq. \ref{eq:NMI}).
Note that if the aggregating partition is very coarse, with a few large aggregation classes, then we expect that $H(f^{-1}(\mathcal{C'})) \ll H(\mathcal{C})$, and
Eq. \ref{eq:eta-vs-I-2} is not more conservative than Eq. \ref{eq:eta-vs-I}. If, on the other hand, we only have a few nodes in each aggregation class and $H(\mathcal{C})$ is relatively large then we may heuristically expect $H(\mathcal{C}) \approx H(\mathcal{C'})$, and it is more relevant to write
\begin{equation}\label{eq:eta-vs-NMI-2}
\eta \gtrsim \text{NMI}(\mathcal{C},f^{-1}(\mathcal{C'})).
\end{equation}
There is no reason that these inequalities should be always tight. Assume for instance that exactly one aggregation class overlaps over two communities $C_1$ and $C_2$ (so that $\eta <1$, if only by a little). Then in the aggregated network, the node resulting from this aggregation class will create edges whose weight typically depends on the \emph{density} of the two communities $C_1$ and $C_2$. Thus if $C_1$ and $C_2$ are sparse enough, the links so created in the aggregated network may be negligible so that the optimal community structure will not be modified, and the ideal situation
$\mathcal{C} = f^{-1}(\mathcal{C'})$ that holds for $\eta=1$ and edge-counting criteria still holds. If on the other hand, the aggregation class cuts into dense communities, this will result in heavy weights in the aggregated that might disrupt significantly the overall community structure. We expect therefore that a network that is heterogeneous in terms of density of links may be potentially more fragile to aggregation, in terms of community structure. In section SA.2. of the Supplementary Information, we investigate the behaviour of the aggregability index $\eta$ in synthetic graphs with planted communities of heterogeneous densities.
In the next sections we show empirically how the aggregability index $\eta$ correlates with the NMI distance between the optimal partitions found for the original and aggregated networks on two datasets. Albeit embedded in the same geographical area ---Belgium--- these two case studies will reveal different behaviours with respect to aggregation. In both cases, we
know a network $G$, aggregate it according to administrative units or regular squares, compute the aggregability index and observe the distorsion of the communities found to be optimal in the new (aggregated) networks.
\section*{Methods}
We now describe the datasets, the definition of community and the way to compare partitions in an empirical approach. Both datasets are localised on parts of Belgium. See section SA.1. of the Supplementary Information for a visualisation and description of the territory.
\subsection*{Twitter networks}
Our first dataset is composed of 291,552 tweets geolocalised on the Belgian territory between 18,327 Twitter users, obtained as described in Supplementary Information SA.2. From this dataset we build a network $N_0$ as follows. The nodes are the users, and the weighted edges count the number of reply-to tweets between the two users (without taking the directionality into account, in order to keep the graph undirected). Each node is associated to a position, obtained as the barycentre of positions of the user recorded in each sent tweet. In this way we see $N_0$ as a network linking positions together. By the means of how the dataset was collected, those positions are spread over the Belgian territory.
A list of aggregated networks was created from $N_0$. The territory of Belgium is divided into 589 municipalities, and used to be divided into 2,675 smaller municipalities until a merge took place in 1979. We first build two aggregated versions, where nodes represent former ($N_{fm}$) and current ($N_{m}$) municipalities, respectively, by merging all nodes of $N_0$ positioned in the same (former or current) municipality. Edges are merged accordingly, receiving a weight that aggregates the weights of all corresponding edges of $N_0$.
We also applied a regular grid of 125 m square cells onto the Belgian territory, and merged into a single node all nodes of $N_0$ positioned in the same cell, creating the aggregating network $N_{125}$. Increasingly coarser square grids of cell size 250 m to 32 km, were used in the same way to create the aggregated networks $N_{250}$ to $N_{32k}$ respectively. The number of nodes and edges are described in Table S1 of the Supplementary Information (SA.3.).
\subsection*{Phone networks}
Our second dataset counts the numbers of phone calls between towers in the territory of Brabant, a former administrative unit (province) of 111 municipalities including and surrounding
Brussels, the capital of Belgium. The derived undirected network, called $M_0$, is composed of 1,168 nodes (towers). A weighted edge between two towers counts the number of communications between the towers in either direction, for a total of 13M communications over the network. As each tower is associated with a precise position, one may again consider $M_0$ as a network between places. We may aggregate those places into municipalities, thus forming the network $M_m$, or into cells of regular size 125 m to 32 km, creating the networks $M_{125}$ to $M_{32k}$, as for the Twitter dataset. See Table S2 of the Supplementary Information, section SA.3, for the number of nodes and edges of each network.
\subsection*{Linearised stability maximisation}
Communities are intuitively meant here as sets of strongly interconnected nodes with comparatively few connections between the communities. Among the many formalisations of this concept, one of the most popular is modularity \cite{newman2004detecting}, quantifying the goodness of a given partition $\mathcal{C}$ of nodes as
\begin{equation}
Q_\mathcal{C}=\frac{1}{2m} \sum_{C \in \mathcal{C}} \sum_{i,j \in C} (A_{ij} - \frac{k_i k_j}{2m}),
\end{equation}
where $m$ is the sum of all weights of the networks' edges, and $k_i$ represents the (weighted) degree of node $i${.} $A_{ij}$ is the weighted adjacency matrix of the network, and
$C (\in \mathcal{C}$) represents a community of the partition.
We use a generalisation, called linearised partition stability \cite{delvenne2013stability}, or equivalently Potts model ~\cite{reichardt2007partitioning}, which introduces a resolution parameter $\rho$ varying from $0$ to $\infty$ as follows:
\begin{equation}
r_{lin}(\rho,\mathcal{C}) = (1-\rho) + \rho \frac{1}{2m} \sum_{C \in \mathcal{C}} \sum_{i,j \in C} ( A_{ij} - \frac{1}{\rho} \frac{k_i k_j}{2m}),
\end{equation}
At $\rho=0$, single nodes are optimal as communities, while partitions with larger communities emerge for increasing values of $\rho$, until a single community is
optimal at $\rho \to \infty$. For $\rho=1$, the linearised stability is the modularity, $r_{lin}(1,\mathcal{C})=Q_\mathcal{C}$.
The resolution parameter $\rho$ is hereafter
called timescale, because linearised stability is formally derived in \cite{delvenne2013stability} as capturing the ability of incumbent communities to retain the flow of a
diffusion of random walkers across the network for a timescale of the order of $\rho$. The original Potts model~\cite{reichardt2007partitioning} uses the parameter $\gamma = 1/\rho$.
As most community detection criteria, linearised stability is NP-hard to optimise except
for extreme values of $\rho$, and we use the Louvain method \cite{blondel2008fuc,lambiotte2014random} as a heuristic.
Whenever appropriate, we will use the linearised stability method to detect communities, because it is an edge-counting criterion, because it includes an extremely popular criterion (modularity, for $\rho=1$) as a special case, and because it allows adapting the timescale parameter $\rho$ in order to create partitionings on different networks with the same or similar number of communities. There are certainly many methods of merits sharing the same properties. Our goal in the Results section is not to find the most sociologically relevant Twitter or phone call communities in Belgium, but illustrate how partitions found with an edge-counting criterion are modified in presence of aggregation. Therefore, the various arguments in favor or against the practical significance of the communities delivered by one or another method are not relevant here.
\subsection*{Normalised mutual information for comparing partitions}
We compute the normalised mutual information \cite{ana2003robust}, between the two partitions $\mathcal{C}$ and $\mathcal{D}$ of the same set of nodes, to evaluate how similar they are, as
\begin{equation} \label{eq:NMI}
\mathrm{NMI}(\mathcal{C},\mathcal{D})= \frac{I(\mathcal{C};\mathcal{D})}{( H(\mathcal{C})+H(\mathcal{D}) )/2},
\end{equation}
where $I(\mathcal{C};\mathcal{D})$ denotes the mutual information between the two partitions, i.e. between the set in $\mathcal{C}$ and the set in $\mathcal{D}$ containing a randomly picked node of the graph. Note that in this article, the sets of nodes belonging to a partition are either called `communities' (if found by community detection algorithm) or `aggregation classes' (if defining a way to aggregate the network).
Similarly, $H(\mathcal{{C}})$ or $H(\mathcal{{D}})$ denotes the Shannon entropy of each partition, i.e., the Shannon entropy of the set of a randomly picked node
of the graph. The NMI takes values between $0$, for independent (thus maximally dissimilar) partitions, and $1$, for identical partitions.
In our case, we also want to be able to compare community partitions at different levels of aggregation, let us say for example the optimal partition $\mathcal{C}$ and $\mathcal{D}$ of networks
$N_0$ and $N_{125}$, respectively. In this case, we lift the communities of $N_{125}$ into communities of $N_0$, replacing each node of $N_{125}$ by its aggregation classes in
$N_0$. We call $\mathcal{D}'$ this partition of the nodes of $N_0$.
We now compare the two partitions $\mathcal{C}$ and $\mathcal{D}'$ with the quantity $\mathrm{NMI}(\mathcal{C},\mathcal{D}')$, which we will also sometimes denote $\mathrm{NMI}(\mathcal{C},\mathcal{D})$ by abuse of notations.
\section*{Results}
In the following, we illustrate on the two real-life datasets the concepts explained above on toy networks. Specifically, we show how the aggregation process over the Twitter and phone call networks strongly affects the community partition in the former case, and mildly so in the latter.
We also show how the magnitude of this distorsion, as the aggregation grid becomes coarser and coarser, correlates with the proposed \emph{aggregability index}.
\subsection*{Twitter networks}
Figure \ref{fig:comgen}-a shows the communities extracted from the network $N_m$ of municipalities, using a timescale $\rho=1$. Each figure from \ref{fig:comgen}-b to
\ref{fig:comgen}-f shows the spatial footprint of one community of individual Twitter users. We have used a timescale $\rho=10$, in order to illustrate the case with a number of communities (namely 5) comparable to the 7 communities of the $N_m$ network. The colour intensity in each municipality represents the proportion of users positioned in this municipality who belong to the community being represented.
\begin{figure}[h!]
\begin{center}
\subfloat[Communities detected in the network of municipalities $N_m$]{\label{fig:comuni}\includegraphics[width=0.47\textwidth]{repto_com_0426.png}}
\hspace{5pt}
\subfloat[Community 1, network $N_0$]{\label{fig:comu1}\includegraphics[width=0.47\textwidth]{repto_ind_com1_0427.png}}
\hspace{5pt}
\subfloat[Community 2, network $N_0$]{\label{fig:comu2}\includegraphics[width=0.47\textwidth]{repto_ind_com2_0427.png}}
\hspace{5pt}
\subfloat[Community 3, network $N_0$]{\label{fig:comu3}\includegraphics[width=0.47\textwidth]{repto_ind_com3_0427.png}}
\hspace{5pt}
\subfloat[Community 4, network $N_0$]{\label{fig:comu4}\includegraphics[width=0.47\textwidth]{repto_ind_com4_0427.png}}
\hspace{5pt}
\subfloat[Community 5, network $N_0$]{\label{fig:comu5}\includegraphics[width=0.47\textwidth]{repto_ind_com5_0427.png}}
\end{center}
\caption{Spatial footprint of communities detected in Twitter networks $N_m$ (municipalities, timescale $\rho=1$) and $N_0$ (individual users, timescale $\rho=10$).}
\label{fig:comgen}
\end{figure}
Some communities of $N_0$ (for example those represented on Figures \ref{fig:comgen}-b and \ref{fig:comgen}-c) show a remarkable geographical dispersion, and in particular do not seem to match any community of $N_m$ (only community $4$ in Fig.\ref{fig:comgen}-e seems to match a community in the network of municipalities in Fig.\ref{fig:comgen}-a, namely the dark blue one).
In order to analyse quantitatively the effect of aggregating data, we systematically test different levels of spatial aggregation, all at the same timescale parameter $\rho=1$. In other words, for the next analysis, we look at the maximum modularity communities, as approximated by the Louvain method.
Figure \ref{fig:formermunigrids} shows communities at different levels of aggregation: municipalities, former (smaller) municipalities and square cells of size 1km, 2km, 4km, 8km. As the aggregation classes become larger and larger they step over several communities forcing their re-arrangement and giving rise to different partitionings.
We can see that as the nodes are increasingly aggregated, some communities gathering distant places, such as the light green community in Fig. \ref{fig:formermunigrids}-a) to \ref{fig:formermunigrids}-c), are re-arranged into geographically localised communities (light green in Fig. \ref{fig:formermunigrids}-f).
White areas represent the physical space where no event has been recorded. At the finest level ($N_0$), nodes are represented as a single point (the average position of a user), thus almost all space is white. As the aggregation scale increases, the white space is progressively removed, being merged with neighbouring space with non-zero activity. We observe that this
effect is more visible in areas with low levels of activity, as the southern part of the country.
The normalised mutual information (NMI) between the disaggregated network $N_0$ and several aggregated networks is depicted on Figure
\ref{fig:nmi_evol}. Starting with the first level of aggregation (125 m), we observe that the NMI already drops rather steeply, even though
there is some fit (NMI $\approx 0.7$) between the communities displayed by aggregated units of 125 m and the non-aggregated ones. Values of NMI continue to decrease with the size of the aggregation.
\begin{figure}[h!]
\begin{center}
\subfloat[Network $N_{1km}$, cells of side 1km]{\label{fig:grid1}\includegraphics[width=0.47\textwidth]{grid1km.png}}
\hspace{5pt}
\subfloat[Network $N_{2km}$, cells of side 2km]{\label{fig:grid25}\includegraphics[width=0.47\textwidth]{grid2km.png}}
\hspace{5pt}
\subfloat[Network $N_{fm}$]{\label{fig:formermuni}\includegraphics[width=0.47\textwidth]{repto_ancom_ts1_rescal0_lcc1_ntrials1000.png}}
\hspace{5pt}
\subfloat[Network $N_{4km}$, cells of side 4km]{\label{fig:grid5}\includegraphics[width=0.47\textwidth]{grid4km.png}}
\hspace{5pt}
\subfloat[Network $N_{m}$]{\label{fig:grid5}\includegraphics[width=0.47\textwidth]{repto_com_0426.png}}
\hspace{5pt}
\subfloat[Network $N_{8km}$, cells of side 8km]{\label{fig:grid10}\includegraphics[width=0.47\textwidth]{grid8km.png}}
\end{center}
\caption{Communities detected in the Twitter network aggregated at the level of former municipalities, $N_{fm}$, (c), at the level of current municipalities in Belgium, $N_{m}$, (e),
and aggregated into grids of square cells of different sizes (a-b, d, f) (timescale parameter set to 1).}
\label{fig:formermunigrids}
\end{figure}
\subsection*{Mobile phone networks}
Fig. \ref{fig:mpcomgen} shows the communities found at the disaggregated level of towers $M_0$ (note that although towers are characterised by a single point, for the visual depiction we represent them by the Voronoi polygone associated to it), and the aggregated level of municipalities $M_m$. The normalised mutual information, NMI, between community partitions found on networks $M_0$ and $M_m$ is 0.64. Thus, the similarity between the communities found on the two levels of aggregation is higher than the similarity observed in the Twitter network between the disaggregated network of users ($N_0$), and the aggregated versions (see Fig.~\ref{fig:nmi_evol}). On Fig.~\ref{fig:nmi_evol} we also notice that the NMI between the communities found on $M_0$ and versions aggregated with larger and larger cells is consistently higher than in the case of the Twitter dataset.
\begin{figure}[htp]
\begin{center}
\subfloat[Network of cell towers $M_t$]{\label{fig:mpcomuni}\includegraphics[width=0.47\textwidth]{phonescell_bra_075.png}}
\hspace{5pt}
\subfloat[Network of communes $M_m$]{\label{fig:mpcomcell}\includegraphics[width=0.47\textwidth]{phonescom_bra_075.png}}
\end{center}
\caption{Communities detected in the mobile phone networks, at the disaggregated level of towers $M_0$ (note that although towers are characterised by a single point, for the visual depiction we represent them by the Voronoi polygone associated to it), and the aggregated level of municipalities $M_m$. The normalised mutual information, NMI, between partitions of network $M_0$ and $M_m$ is 0.64. Thus, the similarity between the communities found on the two levels of aggregation is higher than the similarity observed in the Twitter network between the disaggregated network of users ($N_0$), and the aggregated versions (see Fig.~\ref{fig:nmi_evol}). The timescale parameter $\rho$ is set to $0.75$, as suggested by another study on the same dataset \cite{thomas2017}. }
\label{fig:mpcomgen}
\end{figure}
\subsection*{Aggregability index and NMI}
In Fig. \ref{fig:nmi_evol}, we compare for both datasets, the results of community detection on the original network ($N_0$ or $M_0$) with communities found on the networks of square cells of sides 125 m, 250 m, 500 m, 1 km, 2 km, 4 km, 8 km, 16 km and 32 km. We also plot the aggregability indices, comparing the community structure found on the original networks ($N_0$ or $M_0$) with the aggregating partitions into square cells of sides 125 m, 250 m, 500 m, 1 km, 2 km, 4 km, 8 km, 16 km and 32 km.
The aggregability index, $\eta$, requires the knowledge of the partition into communities and of the partition into aggregation classes at the finest level, but not of the communities that are deemed to be relevant to the aggregated graph. It measures to what extent every aggregation class is a subset of a single community, which is a sufficient condition for the community structure to be left invariant by the (edge-counting) community detection method, as argued in this paper.
The shape of the $\eta$ and NMI curves in Fig. \ref{fig:nmi_evol} is in line with the following facts:
\begin{itemize}
\item If $\eta=1$ then $\text{NMI}=1$ (because we use an edge-counting criterion for detecting communities),
\item For a small aggregation scale we expect from Eq. \ref{eq:eta-vs-NMI-2} that $\eta \gtrsim \text{NMI}$,
\item For further aggregation scales, we know from Eq. \ref{eq:eta-vs-NMI} that $\eta \geq \text{NMI}/2$.
\end{itemize}
Low values of $\eta$ can be seen as a warning signal that communities on the aggregated network (once lifted back to the original network) will necessarily be significantly different than the original communities. In Fig. \ref{fig:nmi_evol} we observe that the value of $\eta$ for mobile phone calls stays remarkably steady until the aggregation scale of 1 or 2 km, while the $\eta$ value for the Twitter dataset dips comparably faster --- and so does the NMI between the community partitions at different scales, as expected.
The fact that the NMI curve of the Twitter dataset drops significantly faster than the $\eta$ curve shows that Eqs \ref{eq:eta-vs-NMI} and \ref{eq:eta-vs-NMI-2} need not be tight. In line with the arguments we discuss above (below Eq. \ref{eq:eta-vs-NMI-2}), one reason for this discrepancy may be a strong heterogeneity of the data in terms of density, as Figs \ref{fig:comgen} and \ref{fig:formermunigrids} suggest.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.7\textwidth]{yer_new_fig.pdf}
\caption{In circles is shown the evolution of normalized mutual information, NMI, between communities found in the network prior aggregation, and communities found in aggregated
networks at several square sizes. In squares, the evolution of the aggregability index, $\eta$, between communities and aggregability at the finest level compared with the
same sizes as before. For Twitter data (in blue) the initial level corresponds to users centroids and time scale kept to $\rho=1$. For mobile phone data (in pink), the initial level corresponds to cell towers and the timescale was kept
constant with a value of 0.75. }
\label{fig:nmi_evol}
\end{center}
\end{figure}
\section*{Discussion}
In this paper, we have studied the impact of data aggregation on community detection in networks. We have shown on theoretical and empirical examples
that data aggregation can preserve the community structure, destroy it, or highlight another relevant community structure. We have identified a class of methods able to preserve the community structure whenever it is aligned with the aggregation classes. We have defined an aggregability index that measures how aligned the community structure is with the aggregation classes.
The article has been structured as a proof of concept. The examples have focused on the most standard notion of communities, as highly interconnected set of nodes. Communities were computed with one of the most popular quality functions, namely modularity and its multiscale extension. We focused on aggregating geographical coordinates into spatial units of increasing size, in line with the well-known Modified Areal Unit Problem in geography.
Nevertheless, from the theoretical considerations, we see that the conclusions may be potentially relevant for different notions of partitioning (e.g. stochastic block modelling) with various aggregation criteria, according to any node metadata such as age, school, etc.
Therefore, broadly speaking, we see our investigation as a warning to data scientists grappling with networks on several levels of aggregation. Our message being that the results of their analyses may depend starkly on the level and nature of the aggregation.
We chose two datasets behaving differently with respect to aggregation, as an illustration for our proposed parameter, the aggregability index. The fact that these two datasets are geographic in nature is incidental in our study, whose scope includes in principle any kind of network and their aggregations. Nonetheless, this might indicate a potential privileged applicability to space-embedded networks. Example of networks embedded in space abound, and the interaction between their structure and the way they unfold in space has triggered some interesting developments, see for example \cite{mascolo_measuring_urban_social_diversity, rosvall_networks_and_cities}. As to explaining why the two datasets behave differently with respect to a same aggregation strategy, one can only formulate hypotheses, whose investigation is beyond the scope of this paper, and may involve the analysis of other datasets with other community detection methods on the same geographical area \cite{arnaudadamJGS}. While the mobile phone calls dataset is shaped by the condition of previous social interaction, this constraint is not present, or to a lesser extent, in the Twitter dataset. Further differences between the datasets include the heterogenous density of events in the Twitter network,
and the different geographic area (Belgium or surroundings of Brussels). Even more importantly, the mobile phone network's nodes at the finest scale are towers, which already aggregate a large number of users.
The present study is certainly not without caveats. For instance in many cases it may be that the full network is inaccessible to the measurement (such as in the human brain connectomes, only available under aggregated form), or too large for most community detection algorithms. In this case, a computation of the aggregation index $\eta$ may not be available. Also, in many cases the aggregated network is available with weights on the edges that do not represent the sum of all interactions between all nodes of the aggregation classes, but only a tresholded version of it, for instance. Ways to cope with this may be a focus of further research.
\section*{List of abbreviations}
$m$ : Meters \\
MAUP : Modifiable Areal Unit Problem \\
NMI : Normalised mutual information \\
$N_0$ : Disaggregated network of Twitter users\\
$N_{fm}$ : Network of former municipalities \\
$N_m$ : Network of municipalities \\
$N_p$ : Networks of Twitter users aggregated into cells of size $p$ meters \\
$M_0$ : Disaggregated network of towers\\
$M_p$ : Networks of towers aggregated into cells of size $p$ meters \\
\section*{Competing interests}
The authors declare no competing interests.
\section*{Data availability statements}
The anonymised phone call datasets used in this paper cannot be made publicly available due to a privacy contract signed between the authors and the phone company in order to avoid privacy issues. The twitter dataset analysed during the current study are available from the corresponding author on reasonable request.
\section*{Author's contributions}
All authors contributed to designe the research, perform the research, write the manuscript.
\section*{Acknowledgements}
This work was supported by Innoviris (project Anticipate - Prospective Research 88 BRU-NET), Federation Wallonia-Brussels (Concerted Research Action ARC 14/19-060), and Flagship European Research Area Network (FLAG-ERA) Joint Transnational Call FuturICT 2.0.
\footnotesize {
|
\section{Introduction}\label{section:intro}
The fundamental interest of operator algebra theory is on separable operator algebras.
Nevertheless, non-separable operator algebras
naturally arise even if one is only interested in countable objects.
For instance, they arise as ultraproducts, central sequence algebras, injective envelopes \cite{Ham79}, \cite{Ham85}, (generalized) Calkin algebras, uniform Roe algebras \cite{Roe93}, and so on.
Also, some set theorists and logicians care about problems on non-separable operator algebras as they often involve statements independent of ZFC and thus provide interesting sources to study axioms beyond ZFC (see e.g., \cite{Far}, \cite{Wea}).
The purpose of the present paper is to clarify
that even in the tamest class of simple operator algebras, that is, the approximately finite dimensional case,
non-separable operator algebras can be surprisingly rigid,
contrary to softness of simple separable amenable operator algebras (see e.g., \cite{Con}, \cite{Win}).
Throughout the paper, we say that a von Neumann algebra is {\it AFD}
(approximately finite dimensional)
if it admits an increasing net of finite dimensional $\ast$-subalgebras
whose union is strongly dense.
This corresponds to Murray and von Neumann's approximate finiteness (C)
introduced in \cite{MvN}.
Similarly, we say a \Cs-algebra is {\it AF} (approximately finite dimensional)
if there is an increasing net of finite dimensional $\ast$-subalgebras
with the norm dense union.
In this paper, we exclude finite dimensional algebras from these classes.
In \cite{FHKT}, Farah--Hathaway--Katsura--Tikuisis studied non-separable approximately finite dimensional operator algebras.
{\it Under the continuum hypothesis}, they have constructed
an AFD type II$_1$ factor and a simple unital AF-algebra
whose central sequences are all trivial.
This concludes that even for approximately finite dimensional algebras,
in the non-separable category, the tensor product structure
of approximately finite dimensional operator algebras is not necessary flexible unlike the separable case.
(We recall that in the separable case, any AFD factor tensorially absorbs the AFD type II$_1$ factor,
and any simple AF-algebra tensorially absorbs the Jiang--Su algebra $\mathcal{Z}$ \cite{JS}.)
The problem if these results are provable in ZFC was left open
(\cite{FHKT}, Question 4.3).
In this paper, we settle these problems.
In fact, we solve their questions in the following stronger form.
\begin{Thmint}[Corollaries \ref{Cor:primeAFD} and \ref{Cor:primeAF}]\label{Thmint:prime}
The following statements hold true in ZFC.
\begin{enumerate}\upshape
\item There is a prime AFD type II$_1$ factor without infinite dimensional regular abelian $\ast$-subalgebras.
\item There is a tensorially prime simple unital AF-algebra
without infinite dimensional regular abelian $\ast$-subalgebras.
\end{enumerate}
\end{Thmint}
We notice that the obstruction on regular abelian $\ast$-subalgebras
yields that these operator algebras cannot be realized
as a (twisted) groupoid algebra of an \'{e}tale groupoid $G$ with $G^{(0)}$ an infinite set.
See Section \ref{subsection:regular} and references given there for more details.
We also note that R{\o}rdam \cite{Ror} has constructed examples of simple unital nuclear
tensorially prime \Cs-algebras.
More precisely, he has constructed simple unital (separable) nuclear \Cs-algebras
which are neither stably finite nor purely infinite.
Tensorial primeness of such \Cs-algebras then follows from Kirchberg's dichotomy (see Theorem 4.1.10 of \cite{Ror02}).
To the author's knowledge, R{\o}rdam's examples
are the only previously known examples of tensorially prime simple nuclear \Cs-algebras.
To construct such algebras, we introduce a new class of groups
which we call {\it iterated wreath product groups} (along an ordinal).
We then apply Popa's orthogonality method \cite{Pop83} to von Neumann algebras of iterated wreath product groups to confirm the primeness and the absence of regular abelian $\ast$-subalgebras.
Then, to obtain the AFD condition of these group von Neumann algebras,
we extend Murray and von Neumann's theorem on the equivalence of
three approximate finiteness conditions for separable type II$_1$ factors
to type II$_1$ factors of density character $\aleph_1$ (Proposition \ref{Prop:AFD}). This is a von Neumann algebra analogue of Proposition 5.6 in \cite{FK}. By combining this result and Connes's classification theorem \cite{Con}, we obtain the desired AFD type II$_1$ factor and AF-algebra.
We also directly determine central sequences for certain crossed products of iterated wreath product groups.
This allows to answer Question 4.3 of \cite{FHKT} even in the class of purely infinite algebras.
In particular we conclude the failure of the Kirchberg $\mathcal{O}_\infty$-absorption theorem (\cite{KP}, Theorem 3.15) in the non-separable case.
Here we recall that the Kirchberg $\mathcal{O}_\infty$-absorption theorem
states that any simple separable nuclear purely infinite \Cs-algebra (called a Kirchberg algebra)
tensorially absorbs the Cuntz algebra $\mathcal{O}_\infty$.
Flexibility of the tensor product structure of amenable operator algebras, e.g., $\mathcal{O}_\infty$-stability, $\mathcal{Z}$-stability, the McDuff property, play a central role in the classification theory of operator algebras (see e.g., \cite{Con}, \cite{Win}).
\begin{Thmint}[Corollary \ref{Cor:NSKT}, Remark \ref{Rem:Cartan}]\label{Thmint:central}
The following statements hold true in ZFC.
\begin{enumerate}\upshape
\item There is a simple unital AF-algebra
with no non-trivial central sequences.
\item There is an AFD type II$_1$ factor
with no non-trivial central sequences.
\item There is an AFD type III factor with no non-trivial central sequences.\item There is a simple unital nuclear purely infinite \Cs-algebra
with no non-trivial central sequences.
\end{enumerate}
Moreover, one can construct these algebras as the reduced crossed product
of a group action on an abelian operator algebra.
\end{Thmint}
By the last statement, algebras in Theorem \ref{Thmint:central} can be non-isomorphic
to algebras in Theorem \ref{Thmint:prime}.
Furthermore, finite dimensional approximations
and AF-algebras in Theorem \ref{Thmint:central}
are directly given. In particular we avoid Connes's classification theorem \cite{Con}
and Proposition \ref{Prop:AFD}.
We remark that a unital \Cs-algebra with no non-trivial central sequences
cannot have non-trivial separable tensor product components.
See Section \ref{subsection:central} for details.
We also construct tensorially prime simple unital \Cs-algebras of decomposition rank one. Here we do not review the decomposition rank of \Cs-algebras. We refer the reader to \cite{KW} for the definition of decomposition rank
and \cite{Win} for its important role in the classification theory of \Cs-algebras.
\begin{Thmint}[Theorem \ref{Thm:dr1}]\label{Thmint:dr1}
There is a tesorially prime simple unital \Cs-algebra of decomposition rank one with no non-trivial central sequences.
\end{Thmint}
We note that AF-algebras are of decomposition rank zero (see Examples 6.1 (i) in \cite{KW}).
Thus these (nuclear) \Cs-algebras have a different feature to those in Theorems \ref{Thmint:prime} and \ref{Thmint:central}.
To show tensorial primeness, we again apply Popa's orthogonality method.
\subsection*{Organization of the paper}
In Section \ref{section:pre}, we summarize basic facts and terminologies.
Besides from known facts, we include a new general statement on
finite GNS-completions of regular abelian $\ast$-subalgebras (Proposition \ref{Prop:regular}).
This in particular strengthens and unifies Corollaries 4.7 and 4.11 of \cite{LR} (see Corollary \ref{Cor:ssolid}). In Section \ref{section:iterated}, we introduce and study iterated wreath product groups. In Section \ref{section:AFD}, we establish the equivalence
of approximate finiteness (A), (B), (C) of Murray--von Neumann for type II$_1$ factors of density character $\aleph_1$. In Section \ref{section:prime},
we prove Theorem \ref{Thmint:prime}. In Section \ref{section:central},
we study central sequences of certein crossed products.
As a consequence, we obtain Theorem \ref{Thmint:central}.
In Section \ref{section:dr1}, by modifying the constructions and proofs in
Sections \ref{section:prime} and \ref{section:central}, we prove Theorem \ref{Thmint:dr1}.
\section{Preliminries}\label{section:pre}
Here we recall a few basic facts.
Our basic references on operator algebras are \cite{BO}, \cite{KR}.
We first fix notations which we frequently use.
\subsection*{Notations}
\begin{itemize}
\item The symbols `$\votimes$' and `$\otimes$' stand for the von Neumann algebra tensor product and the minimal tensor product respectively.
\item The symbols `$\vnc$' and `$\rc$' stand for the von Neumann algebra crossed product and the reduced \Cs-crossed product respectively.
\item For a subset $S$ of a von Neumann algebra $M$,
denote by $\Wso(S)$ the von Neumann subalgebra of $M$ generated by $S$.
\item For a von Neumann algebra $M$, $U(M)$ denotes the unitary group of $M$.
\item We denote the unit element of a group by $e$.
(The ambient group is always clear from the context.)
\end{itemize}
\subsection{Ordinals}
For the definition and basic facts on (von Neumann) ordinals, see Chapter 1 of \cite{Kun} for instance.
Here we only recall a few basic terminologies which we frequently use.
As usual, denote by $\omega_1$ the smallest uncountable ordinal.
Denote by $\aleph_0$ and $\aleph_1(= \omega_1)$ the countable infinite cardinal and the smallest uncountable cardinal respectively.
For each ordinal $\alpha$, there is at most one ordinal $\beta$
with $\alpha= \beta+1$.
If such $\beta$ exists, $\alpha$ is said to be a successor ordinal,
and $\beta$ is said to be the predecessor of $\alpha$.
Otherwise $\alpha$ is said to be a limit ordinal.
A limit ordinal $\alpha$ is said to be of {\it uncountable cofinality}
if there is no countable subset $S$ in $\{ \beta : \beta < \alpha\}$
satisfying $\sup(S) =\alpha$.
For instance, $\omega_1$
is of uncountable cofinality, as the union of countably many countable sets is again countable.
\subsection{Tensorial primeness}
A von Neumann algebra $M$ is
said to be {\it prime} if for any isomorphism $M \cong M_1 \votimes M_2$,
one of $M_i$ is type $I$.
Similarly, a \Cs-algebra $A$ is said to be {\it tensorially prime}
if for any isomorphism $A \cong A_1 \otimes A_2$,
one of $A_i$ is a type $I$ \Cs-algebra.
(Since the term `primeness' is reserved for another property of \Cs-algebras,
we use this terminology.)
It is clear that any strongly dense \Cs-subalgebra of a prime type II$_1$ factor is tensorially prime.
\subsection{Normalizers and regularity for subalgebras}\label{subsection:regular}
Throughout the paper, we only consider nonzero subalgebras.
For a $\ast$-subalgebra $N\subset M$ of a von Neumann algebra $M$,
define the {\it normalizer} of $N$ in $M$ to be
\[\mathcal{N}_M(N):= \{ u\in U(M): u N u^\ast =N\}.\]
A $\ast$-subalgebra $N$ of $M$ is said to be {\it regular}
if it satisfies $\mathcal{N}_M(N)''=M$.
A typical regular abelian $\ast$-subalgebra arises from the crossed product construction.
Indeed, if $\Gamma$ is a group acting on an abelian von Neumann algebra $A$,
then the inclusion $A \subset A \vnc \Gamma$ is regular.
For a $\ast$-subalgebra $ A \subset B$ of a \Cs-algebra $B$,
define the {\it normalizer} of $A$ in $B$ to be
\[N_B(A) := \{ n\in B: n A n^\ast \cup n^\ast A n \subset A\}.\]
A \Cs-subalgebra $A$ of $B$ is said to be {\it regular} if $N_B(A)$ generates
$B$ as a \Cs-algebra.
We note that, in this case, $N_B(A)$ spans a dense subspace of $B$.
Regular abelian $\ast$-subalgebras naturally arise from the groupoid constructions.
More precisely, for an \'etale locally compact Hausdorff groupoid $G$ with a twist $\Sigma$,
the inclusions $C_0(G^{(0)}) \subset {\rm C}^\ast_{\rm r}(G, \Sigma)$ and
$C_0(G^{(0)}) \subset {\rm C}^\ast(G, \Sigma)$ are regular (for their definitions, see Section 4 in \cite{Ren08}).
For the proof of this fact, see Corollary 4.9 in \cite{Ren08}.
(Note that the second countability assumption is not used for this conclusion, and the same proof works for
the full twisted groupoid \Cs-algebras.)
Thus the absence of regular abelian $\ast$-subalgebras
implies the absence of (twisted) groupoid \Cs-algebra realizations.
Obviously, the definitions of regularity as a \Cs- and von Neumann algebra
are very different. For subalgebras of von Neumann algebras, we always consider regularity in the von Neumann algebra sense.
We next record a few observations on regular abelian $\ast$-subalgebras.
The next lemma is standard and well-known.
Since the author does not know a reference for this result, we include a proof.
\begin{Lem}\label{Lem:diffusevn}
Let $A \subset M$ be a regular abelian $\ast$-subalgebra of a type II$_1$ factor $M$.
Then the strong closure of $A$ contains $1_M$ and is either diffuse or finite dimensional.
\end{Lem}
\begin{proof}
By taking the strong closure if necessary, we may assume that $A$ is strongly closed in $M$.
We first show that $1_M \in A$.
Let $1_A$ denote the unit of $A$.
Then for any $u \in \mathcal{N}_M(A)$, $1_A= u1_A u^\ast$.
By the factoriality of $\mathcal{N}_M(A)''$($=M$), we have $1_A=1_M$.
Now assume we have a minimal projection $p$ in $A$.
Then for any $u \in \mathcal{N}_M(A)$,
$u p u^\ast$ is again a minimal projection in $A$.
Since $A$ is abelian, for any $u, v \in \mathcal{N}_M(A)$,
we have one of the relations $upu^\ast = vpv^\ast$ or $upu^\ast vpv^\ast=0$.
Then, since $M$ is finite,
the set
\[P_p:=\{ u p u^\ast : u \in \mathcal{N}_M(A)\}\]
is finite.
Define $r := \sum_{q \in P_p} q$.
Then for any $u\in\mathcal{N}_M(A)$, we have $uru^\ast =r$.
Hence $r \in A \cap \mathcal{N}_M(A)' = \mathbb{C}$.
Therefore $A$ is finite dimensional.
\end{proof}
We next prove the following general proposition.
This implies a \Cs-algebra analogue of Lemma \ref{Lem:diffusevn}.
We note that the last statement is shown in Lemma 4.10 of \cite{LR} (see also Corollary 4.11 in \cite{LR}) for Cantan subalgebras in separable \Cs-algebras. We point out that our proof does not rely on the groupoid realization theorem \cite{Ren08} unlike their proof.
\begin{Prop}\label{Prop:regular}
Let $C\subset A$ be a regular abelian $\ast$-subalgebra of a \Cs-algebra $A$.
Let $\pi \colon A \rightarrow \mathbb{B}(H)$
be a $\ast$-representation such that $\pi(A)''$ is finite.
Then $\pi(C)''$ is regular in $\pi(A)''$.
In particular, if $\pi(A)''$ is a type II$_1$ factor,
then $\pi(C)''$ is either diffuse or finite dimensional.
\end{Prop}
\begin{proof}
Note that the last statement follows from
the first statement and Lemma \ref{Lem:diffusevn}.
Put $\mathcal{C}:= \stcl(\pi(C))$, $\mathcal{A}:= \pi(A)''$.
We first show that the unit $p$ of $\mathcal{C}$ commutes with $\mathcal{A}$.
For any $n\in \pi( N_A(C))$, as $n \mathcal{C} n^\ast \cup n^\ast \mathcal{C} n \subset \mathcal{C}$, direct computations imply
\[(pn -pnp)^\ast(pn -pnp) =0,~ (np -pnp)(np -pnp)^\ast=0.\]
This yields $p\in \mathcal{A} \cap \pi( N_A(C))'= \mathcal{A} \cap \mathcal{A}'$.
Thanks to the centrality of $p$ in $\mathcal{A}$,
the regularity of the inclusion $\pi(C)'' \subset \mathcal{A}$
is equivalent to the regularity of $\mathcal{C} \subset \mathcal{A}p$.
Hence, by restricting $\pi$ to $pH$ if necessary, we may assume
$1_H \in \mathcal{C}$.
We show that, for any $n\in \pi(N_A(C))$,
there is $u\in \mathcal{N}_{\mathcal{A}}(\mathcal{C})$
satisfying $n=u|n|$. Since $|n| =(n^\ast 1_H n)^{1/2} \in \mathcal{C}$ for all $n \in \pi(N_A(C))$, this implies the regularity of $\mathcal{C} \subset \mathcal{A}$.
To show the claim, we first show that for any $n \in \pi(N_A(C))$,
denoting by $n= v|n|$ the polar decomposition of $n$,
we have $v \mathcal{C} v^\ast \cup v^\ast \mathcal{C} v\subset \mathcal{C}$.
To see this, for $k\in \mathbb{N}$, define a Borel function $f_k \colon \mathbb{R} \rightarrow \mathbb{R}$ to be $f_k(t) = t^{-1}$ for $t\in [1/k, k]$ and $f_k(t)=0$ otherwise. Set $a_k:=f_k(|n|)$.
Then, since $a_k \in \mathcal{C}$, we have
\[na_k \mathcal{C} a_k n^\ast \subset \mathcal{C}.\]
As the bounded sequence $(n a_k )_{k=1}^\infty$ $\ast$-strongly converges to $v$, we have $v \mathcal{C} v^\ast \subset \mathcal{C}$.
By applying the same argument to $n^\ast$, we obtain the other inclusion.
Define the set
\[\mathcal{V}:= \{ w\in \mathcal{A}: w{\rm~is~a~partial~isometry}, w^\ast \mathcal{C} w \cup w \mathcal{C} w^\ast \subset \mathcal{C}, wv^\ast v=v\}.\]
Note that $v\in \mathcal{V}$ hence $\mathcal{V}$ is not empty.
We equip the set $\mathcal{V}$ with the partial order $\preceq$ defined as follows.
For two elements $w_1, w_2 \in \mathcal{V}$,
we declare $w_1 \preceq w_2$ if $w_2 w_1^\ast w_1 =w_1$.
It is not hard to see that $(\mathcal{V}, \preceq)$ satisfies the assumption of Zorn's lemma.
Choose a maximal element $u$ of $\mathcal{V}$.
We show that $u$ is a unitary element.
To lead to a contradiction, assume that $u$ is not a unitary element.
Since $\mathcal{A}$ is finite,
the projections $p:=1-u^\ast u$ and $q:= 1-uu^\ast$ are Murray--von Neumann equivalent in $\mathcal{A}$.
Since $\pi(N_A(C))$ spans a strongly dense subspace of $\mathcal{A}$,
one can choose $m\in \pi(N_A(C))$ satisfying
$x:=qmp \neq 0$.
Since both $p, q$ are in $\mathcal{C}$,
we have $x \mathcal{C} x^\ast \cup x^\ast \mathcal{C} x \subset \mathcal{C}$.
Let $x= w|x|$ be the polar decomposition of $x$.
By the argument in the second previous paragraph, we have $w \mathcal{C} w^\ast \cup w^\ast \mathcal{C} w \subset \mathcal{C}$.
Also, since $x= q xp$, we have $w=qwp$.
This implies $w u^\ast = w^\ast u =0$.
Hence $s:=u+w$ is a partial isometry in $\mathcal{A}$.
It is clear from the definition that $s u^\ast u= u$
and $s\neq u$.
Moreover, we have $w \mathcal{C}u^\ast = w u^\ast u \mathcal{C} u^\ast =0$
and similarly $w^\ast \mathcal{C} u= 0$.
Hence
\[s \mathcal{C} s^\ast \cup s^\ast \mathcal{C} s \subset \mathcal{C}.\]
This contradicts the maximality of $u$.
We thus conclude $u\in \mathcal{N}_{\mathcal{A}}(\mathcal{C})$.
Since $uv^\ast v= v$,
we obtain the desired presentation $n=u|n|$.
\end{proof}
As an immediate consequence of Proposition \ref{Prop:regular}, we obtain the following corollary.
This strengthens and unifies Corollaries 4.7 and 4.11 in \cite{LR} (cf.~Corollary B of \cite{OP2}, Theorem B of \cite{CS}).
We again emphasize that our proof does not involve groupoids.
We recall from \cite{OP}, \cite{OP2} that
a von Neumann algebra $M$ is said to be {\it strongly solid}
if for any diffuse amenable von Neumann subalgebra $A$ of $M$,
$\mathcal{N}_M(A)''$ is amenable.
\begin{Cor}\label{Cor:ssolid}
For any non-amenable strongly solid type II$_1$ factor,
all its strongly dense \Cs-subalgebras have no infinite dimensional regular abelian $\ast$-subalgebras.
\end{Cor}
\subsection{Popa's orthogonality method}
Recall that a von Neumann algebra $M$ is said to be {\it countably decomposable}
if there is no uncountable family consisting of pairwise orthogonal nonzero projections in $M$.
Throughout the paper, for a countably decomposable finite von Neumann algebra $M$, we fix a faithful normal tracial state $\tau$ (otherwise specified).
In most cases, $M$ is a type II$_1$ factor and thus the choice of $\tau$ is canonical.
Our analysis in Section \ref{section:prime} (thus Theorem \ref{Thmint:prime}) is based on {\it Popa's orthogonality method} \cite{Pop83}.
Here we briefly review this method and related terminologies.
In the seminal paper \cite{Pop83}, Popa has developed a powerful strategy to determine
the normalizers of operator subalgebras.
Among other things, his method provides the following striking results;
the first examples of prime type II$_1$ factors (e.g., uncountable free group factors),
the first examples of type II$_1$ factors with no infinite dimensional regular abelian $\ast$-subalgebras (e.g., uncountable free group factors, ultraproducts of type II$_1$ factors). For more applications, see the original article \cite{Pop83}.
His analysis of the normalizer is based on the orthogonality relation between
two von Neumann subalgebras.
Here recall from Definition 2.2 of \cite{Pop83} that
two $\ast$-subalgebras $A, B$ in $M$ are said to be {\it orthogonal},
denoted by $A \perp B$,
if they satisfy the following condition:
$\tau(a b) = \tau(a) \tau(b)$ for all $a\in A$ and $b\in B$.
This condition has a number of useful characterizations; see Lemma 2.1 in \cite{Pop83}.
\subsection{Density character}
The {\it density character} of a topological space $X$
is
the cardinal \[\min\{\sharp S: S \subset X {\rm ~dense}\}.\]
For von Neumann algebras, we always consider
the density character with respect to the $\sigma$-strong topology.
Note that the weak topology, strong topology, strong $\ast$-topology,
$\sigma$-weak topology, $\sigma$-strong $\ast$-topology
define the same density character as the $\sigma$-strong topology. This follows from
the proof of statement (1) below.
The following proposition clarifies aspects of the density character of von Neumann algebras.
For future reference, we record the proof.
\begin{Prop}\label{Prop:dc}
\begin{enumerate}[\upshape(1)]
\item The density character of $M$ is equal to
\[\min\{\sharp S:S\subset M {\rm~infinite~ subset}, \Wso(S)=M\}.\]
\item When $M$ is countably generated, the density character of $M$ coincides with that of the predual $M_\ast$ (with respect to the norm topology).
\item When $M$ is infinite dimensional, for any faithful normal state $\varphi$ on $M$,
the density character of $M$ is equal to the dimension
of the GNS Hilbert space $L^2(M, \varphi)$.
\end{enumerate}
\end{Prop}
\begin{proof}[Proof $($outline$)$]
(1):
We observe that, for any infinite subset $S$ of $M$,
the $\ast$-subalgebra of $M$ generated by $S$ over the field $\mathbb{Q}(i)$
has the same cardinal as $S$.
This proves the claim.
(2): The duality $(M_\ast)^\ast =M$ shows that
the density character of $M$ is at most that of $M_\ast$.
To see the converse, take a faithful normal state $\varphi$ on $M$.
Then the Hahn--Banach theorem shows that
the set $\{\varphi(x \cdot): x\in M\}$ is dense in $M_\ast$.
(3): Since $M \subset \mathbb{B}(L^2(M, \varphi))$,
the density character of $M$ is at most the dimension of $L^2(M, \varphi)$.
The converse follows from the following observation.
For any strongly dense subset $S$ of $M$,
its canonical image in $L^2(M, \varphi)$ is dense.
\end{proof}
For \Cs-algebras, we consider the density character with respect to the norm topology.
Obviously the analogue of Proposition \ref{Prop:dc} (1) holds true in the \Cs-algebra context.
\subsection{Central sequences}\label{subsection:central}
For a unital \Cs-algebra $A$,
a bounded sequence $(x_n)_{n=1}^\infty$ of $A$ is said to be {\it central}
if it satisfies $\lim_{n \rightarrow \infty} \| x_ny -y x_n\| =0$ for all $y\in A$.
A central sequence $(x_n)_{n=1}^\infty$ in $A$ is said to be {\it trivial}
if there is a sequence $(\lambda_n)_{n=1}^\infty$ of scalars
satisfying $\lim_{n \rightarrow \infty} \| x_n -\lambda_n\|=0$.
We say that $A$ has {\it no non-trivial central sequences}
if all central sequences of $A$ are trivial.
We note that if a unital \Cs-algebra $A$ has no non-trivial central sequences, then $A$ has no separable tensor product factors other than full matrix algebras.
To see this, it suffices to show that any unital separable \Cs-algebra $B$
except full matrix algebras has a non-trivial central sequence.
The case $B$ is simple is shown in Proposition 2.10 of \cite{Kir}.
When $B$ is not simple, for a proper ideal $J$ of $B$,
a quasi-central approximate unit $(e_n)_{n=1}^\infty$ of $J$ in $B$
defines a non-trivial central sequence in $B$.
Similarly, for a von Neumann algebra $M$,
an (operator norm) bounded sequence $(x_n)_{n=1}^\infty$ of $M$
is said to be {\it central} if for any $y\in M$,
the sequence $(x_ny - y x_n)_{n=1}^\infty$ $\ast$-strongly converges to $0$.
A central sequence $(x_n)_{n=1}^\infty$ in $M$ is said to be {\it trivial}
if there is a sequence $(\lambda_n)_{n=1}^\infty$ of scalars
satisfying $x_n - \lambda_n \rightarrow 0$ $\ast$-strongly as $n$ tends to infinity.
We say that $M$ has {\it no non-trivial central sequences}
if all central sequences of $M$ are trivial.
We remark that, for von Neumann algebras, the definitions of centrality and triviality in the von Neumann and \Cs-algebra sense are very different.
In this paper, for von Neumann algebras, we always understand
these terminologies in the von Neumann algebra sense.
\section{Iterated wreath product groups and their basic properties}\label{section:iterated}
Throughout the paper, we assume that groups are discrete.
In this section, we introduce a new class of groups which we call {\it iterated wreath product groups}.
We also study their basic properties which we use.
We first recall the wreath product construction.
For two groups $\Gamma$ and $\Lambda$,
the wreath product group $\Gamma \wr \Lambda$
is given by
$(\bigoplus_{\Lambda} \Gamma) \rtimes_\sigma \Lambda$.
Here $\sigma \colon \Lambda \curvearrowright \bigoplus_{\Lambda} \Gamma$
denotes the left shift action.
We identify $\Lambda$ with the subgroup of $\Gamma \wr \Lambda$
via the canonical embedding $ \Lambda \ni g \mapsto (e, g) \in \Gamma \wr \Lambda$.
The main idea of the construction is to iterate
the wreath product construction along an ordinal.
More precisely, we construct a new group as follows.
Given an ordinal $\alpha$ and
a family $(\Gamma_\beta)_{\beta < \alpha}$ of groups.
We then define an increasing family $(G_\beta)_{\beta \leq \alpha}$
of groups as follows.
We first set $G_0 := \{ e \}$ (the trivial group).
Assume that the increasing family $(G_\gamma)_{\gamma<\beta}$
has been defined for an ordinal $\beta \leq \alpha$.
We then define $G_\beta$ as follows.
When $\beta$ is a limit ordinal,
we define $G_\beta := \bigcup_{\gamma < \beta} G_\gamma$.
When $\beta$ has the predecessor $\eta$,
we define
$G_\beta := \Gamma_\eta \wr G_\eta$.
This process extends the family
$(G_\gamma)_{\gamma<\beta}$ to $(G_\gamma)_{\gamma<\beta+1}$.
By the transfinite induction, we obtain the increasing family
$(G_\beta)_{\beta \leq \alpha}$ of groups.
We refer to the group $G:=G_\alpha$ as the {\it iterated wreath product of
$(\Gamma_\beta)_{\beta<\alpha}$}.
Throughout the paper, for $\beta<\alpha$, we denote by $G_\beta$ the subgroup of $G$ obtained in the construction of $G$ at the $\beta$-th step.
We next define the degree for elements in $G$.
For $g\in G$,
we define \[\deg(g):= \min \{ \beta \leq \alpha: g\in G_\beta\}.\]
By the definition of $(G_\beta)_{\beta \leq \alpha}$,
$\deg(g)$ must be a successor ordinal.
We now summarize basic properties of iterated wreath product groups.
The following lemma is useful to study properties of iterated wreath product groups.
Recall that a group is said to be {\it locally finite}
if it is a directed union of finite subgroups.
\begin{Lem}\label{Lem:lf}
Let $\alpha$ be an ordinal. Let $\mathcal{C}$ be a class of groups
closed under taking wreath products, isomorphisms, and directed unions
$($e.g., the class of amenable groups, the class of locally finite groups, the class of torsion-free groups$)$.
Let $(\Gamma_\beta)_{\beta < \alpha}$ be a family of groups in $\mathcal{C}$.
Then the iterated wreath product $G$ of $(\Gamma_\beta)_{\beta < \alpha}$ is contained in $\mathcal{C}$.
\end{Lem}
\begin{proof}
This follows by transfinite induction on $\alpha$.
\end{proof}
\begin{Lem}\label{Lem:ICC}
Let $\alpha$ be a limit ordinal.
Let $(\Gamma_\beta)_{\beta <\alpha}$ be a family of non-trivial groups.
Then the iterated wreath product $G$ of $(\Gamma_\beta)_{\beta <\alpha}$ is ICC.
\end{Lem}
\begin{proof}
Let $g\in G \setminus \{ e\}$ be given.
Put $\beta:=\deg(g)$.
Since $\alpha$ is a limit ordinal, for any $n\in \mathbb{N}$,
we have $\beta+n <\alpha$.
For each $n\in \mathbb{N}$, choose
$g_n\in \Gamma_{\beta + n}\setminus \{e\}$.
We identify $\Gamma_{\beta+n}$ with the $e$-th direct summand
of $\bigoplus_{G_{\beta+n}} \Gamma_{\beta+n} \subset G_{\beta+n+1}$.
Then $g_n g g_n^{-1} = g_n \sigma_g (g_n^{-1}) g \in G_{\beta+n +1}\setminus G_{\beta+n}$ for each $n\in \mathbb{N}$.
Hence the conjugacy class of $g$ in $G$ is infinite.
\end{proof}
The following disjointness property is the key of our results. Cf.~Proposition 4.1 in \cite{Pop83}.
\begin{Lem}\label{Lem:disjoint}
Let $\alpha$ be an ordinal.
Let $(\Gamma_\beta)_{\beta < \alpha}$ be a family of torsion-free groups.
Let $G$ be the iterated wreath product of $(\Gamma_\beta)_{\beta < \alpha}$.
Then for any $\beta <\alpha$ and for any $g\in G \setminus G_\beta$,
we have
\[G_\beta \cap g G_\beta g^{-1} = \{e \}.\]
\end{Lem}
\begin{proof}
Let $\gamma$ denote the predecessor of $\deg(g)$.
Note that by assumption, $\gamma \geq \beta$.
Since $G_\beta \subset G_\gamma$,
by replacing $\beta$ by $\gamma$ if necessary,
we may assume $g\in G_{\beta +1} \setminus G_\beta$.
Recall the definition $G_{\beta +1 } = \Gamma_\beta \wr G_\beta$.
By replacing $g$ by $gh$ for an appropriate $h \in G_\beta$,
we may assume
$g \in (\bigoplus_{G_\beta} \Gamma_\beta) \setminus \{ e \}$.
Then, since $G_\beta$ is torsion-free,
the (restricted) left shift action $\sigma \colon G_\beta \curvearrowright (\bigoplus_{G_\beta} \Gamma_\beta) \setminus \{ e \}$ is free.
Hence, for any $k\in G_\beta \setminus \{e \}$, we have
$g k g^{-1} = g \sigma_k(g^{-1}) k\in G \setminus G_\beta$.
\end{proof}
\section{Equivalence of Murray--von Neumann's approximate finiteness (A), (B), (C) for type II$_1$ factors of density character $\aleph_1$}\label{section:AFD}
In their fundamental work \cite{MvN}, Murray and von Neumann
developed the study of approximately finite dimensional von Neumann algebras.
They introduced three notions of approximate finiteness (A), (B), and (C), and show that all three conditions are equivalent for separable type II$_1$ factors.
In this section, following the strategy of Farah--Katsura \cite{FK},
we extend this result to type II$_1$ factors of density character $\aleph_1$.
Since the proof is basically the same as \cite{FK}, we only discuss the points where we need modifications.
Though in \cite{MvN} the definitions of approximate finiteness are stated in the separable case, we have obvious translations of these properties in general case as follows.
Before stating them, we introduce the following notation.
Let $M$ be a type II$_1$ factor.
For $S, T \subset M$ and for $\delta>0$,
denote by $S\subset^\delta T$ if they satisfy the following condition:
for any $x\in S$, there is $y\in T$ with
$\|x - y\|_2 <\delta$.
Here $\| a \|_2:= \tau(a^\ast a)^{1/2}$ for $a \in M$.
\begin{Def}[Definitions 4.3.1, 4.5.2, and 4.6.1 in \cite{MvN}]
Let $M$ be a type II$_1$ factor.
\begin{enumerate}[(A)]
\item
We say that $M$ has {\it approximate finiteness (A)}
if it satisfies the following condition:
for any finite subset $F \subset M$ and any $\epsilon>0$,
there is an $n\in \mathbb{N}$ with the following property:
for any $q\geq n$, there is a type I$_q$ subfactor $N$ of $M$
with $ F \subset^\epsilon N$.
\item
We say that $M$ has {\it approximate finiteness (B)}
if it satisfies the following condition:
for any finite subset $F \subset M$ and any $\epsilon>0$,
there is a finite dimensional $\ast$-subalgebra $N$ of $M$
with $ F \subset^\epsilon N$.
\item
We say that $M$ has {\it approximate finiteness (C)}
if it admits a strongly dense $\ast$-subalgebra obtained as a directed union of
finite dimensional $\ast$-subalgebras.
\end{enumerate}
\end{Def}
In \cite{Ell}, it is shown that injectivity is equivalent to approximate finiteness (B)
(certainly the proof depends on Connes's theorem \cite{Con}).
In a similar method to \cite{Ell}, one can show the equivalence of injectivity and approximate finiteness (A).
It is clear from the definition that (C) implies (B).
Hence to show the equivalence of (A), (B), (C), it suffices to show that injectivity implies (C).
Indeed, the following proposition holds true.
\begin{Prop}\label{Prop:AFD}
Let $M$ be an injective type II$_1$ factor of density character $\aleph_1$.
Then there is an increasing net $(M_i)_{i \in I}$ of
type I subfactors of $M$ whose union is strongly dense.
\end{Prop}
We point out that one can show the equivalence of three approximate finiteness conditions
for type II$_1$ factors of density character without using Connes's classification theorem.
See Remark \ref{Rem:Con} for details.
However we need Proposition 4.2 rather than just the equivalence of three approximate finiteness conditions.
We therefore give a direct proof of Proposition \ref{Prop:AFD}
(certainly with the aid of Connes's classification theorem \cite{Con}).
To prove Proposition \ref{Prop:AFD}, we need a few lemmas.
The next lemma plays the role of Lemma 5.1 in \cite{FK} in the von Neumann algebra setting. For a \Cs-algebra $A$,
we set $\mathcal{B}_A:= \{ x\in A: \| x \|\leq 1\}$.
\begin{Lem}\label{Lem:AFD}
For any $\epsilon>0$ and any $n\in \mathbb{N}$,
there exists $\delta>0$ with the following property.
For any type II$_1$ factor $M$,
for any type I$_n$ subfactor $D$ of $M$ and any type I$_{m}$ subfactor $B$ of $M$
satisfying $\mathcal{B}_D \subset ^\delta \mathcal{B}_B$ and $n|m$,
and for any subfactor $E\subset B \cap D$,
there is a unitary element $u \in M \cap E'$ satisfying
$u D u^\ast \subset B$, $\|u -1 \|_2 < \epsilon$.
\end{Lem}
\begin{proof}
Since $E$ is a type I subfactor contained in $B \cap D$, we have the tensor product decompositions
$M= E \votimes N$, $B= E \votimes P$, $D= E \votimes Q$,
where $N:= M \cap E'$, $P := B \cap E'$, $Q:= D \cap E'$.
Observe that $\mathcal{B}_D \subset ^\delta \mathcal{B}_B$
implies $\mathcal{B}_{P} \subset^\delta \mathcal{B}_{Q}$.
This reduces the proof of the statement to the case $E =\mathbb{C}$.
In this case, the statement is an immediate consequence of Lemmas 12.2.4 and 12.2.5 in \cite{KR}.
\end{proof}
We further introduce a few terminologies.
Let $\Lambda$ be a directed set. Let $M$ be a von Neumann algebra.
Recall from \cite{FK} that $\Lambda$ is said to be {\it $\sigma$-complete}
if any countable directed subset of $\Lambda$ has a supremum in $\Lambda$.
A family $(M_\lambda)_{\lambda \in \Lambda}$ of von Neumann subalgebras
of $M$
is said to be {\it directed} if
$\lambda \leq \mu$ if and only if $M_\lambda \subset M_\mu$.
Following \cite{FK}, we say that
a directed family $(M_\lambda)_{\lambda \in \Lambda}$
of von Neumann subalgebras of $M$
is {\it $\sigma$-complete} if it satisfies the following conditions:
\begin{itemize}
\item $\Lambda$ is $\sigma$-complete.
\item For any countable directed subset $S\subset \Lambda$,
we have $M_{\sup(S)} = \Wso(\bigcup_{s\in S}M_s)$.
\end{itemize}
We say that
a directed family $(M_\lambda)_{\lambda \in \Lambda}$
of von Neumann subalgebras of $M$ has {\it dense union}
if the union $\bigcup_{\lambda \in \Lambda} M_\lambda$ is strongly dense in $M$.
We next show the following basic lemma, which plays the role of Lemma 5.5 in \cite{FK}.
\begin{Lem}\label{Lem:directed}
Every type II$_1$ factor $M$ of density character $\aleph_1$ admits a directed $\sigma$-complete family of
separable subfactors with dense union indexed by $\omega_1$.
\end{Lem}
\begin{proof}
We first show that
any separable von Neumann subalgebra $N$ of $M$ is contained in a separable subfactor of $M$. To see this, take a countable strongly dense subset $S$ of $N$.
By Theorem 8.3.6 in \cite{KR},
for any $x\in M$, we have
\[\tau(x)\in \overline{\rm conv}\{ u x u^\ast: u \in U(M)\}.\]
Here and below, the symbol `$\overline{\rm conv}$'
stands for the norm closure of the convex hull.
Since $S$ is countable, one can choose a countable subset $T$ of $U(M)$
satisfying
\[\tau(x)\in \overline{\rm conv}\{ u x u^\ast: u \in T\} {\rm~for~ all~} x\in S.\]
Set $N_1 := \Wso(N \cup T)$.
Since $S$ is strongly dense in $N$,
it follows from the choice of $T$ that
any normal tracial state $\tau_1$ on $N_1$ satisfies
$\tau_1|_N = \tau|_N$.
By continuing this process, we obtain
an increasing sequence $(N_n)_{n=1}^\infty$
of separable von Neumann subalgebras in $M$
with the following property:
for any $n\in \mathbb{N}$ and for any normal tracial state $\tau_{n+1}$ on $N_{n+1}$,
we have $\tau_{n+1}|_{N_n}= \tau|_{N_n}$.
Now define $L:= \Wso(\bigcup_{n \in \mathbb{N}}N_n)$.
Then by the choice of $N_n$'s,
$L$ has a unique (faithful) normal tracial state.
Hence $L$ is a factor. Clearly $L$ is separable.
Now choose a strongly dense net $(x_\alpha)_{\alpha < \omega_1}$ in $M$.
We define a directed family $(M_\alpha)_{\alpha <\omega_1}$ of separable subfactors of $M$ as follows.
We first choose $M_0$ to be a separable subfactor of $M$ containing $x_0$.
For an ordinal $\alpha<\omega_1$, assume we have defined the family $(M_\beta)_{\beta<\alpha}$.
We then define $M_\alpha$ as follows.
When $\alpha$ is a limit ordinal,
set $M_\alpha := \Wso(\bigcup_{\beta<\alpha} M_\beta)$.
When $\alpha$ has the predecessor $\eta$,
fix $y\in M \setminus M_\eta$
and
define $M_\alpha$ to be a separable subfactor of $M$ containing $M_\eta \cup \{x_\gamma:\gamma\leq \alpha\} \cup \{y\}$.
By the transfinite induction,
we obtain a family $(M_\alpha)_{\alpha < \omega_1}$
of separable subfactors of $M$.
It is clear from the construction that the family is directed and $\sigma$-complete.
Moreover, by the construction,
$x_\alpha \in M_{\alpha+1}$ for all $\alpha<\omega_1$.
This shows the density of $(M_\alpha)_{\alpha < \omega_1}$ in $M$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{Prop:AFD}]
By Connes's classification theorem \cite{Con},
each separable type II$_1$ subfactor $N$ of $M$
admits an increasing sequence of subfactors of type $I_{2^n}$, $n\in \mathbb{N}$, whose union is strongly dense in $N$.
The rest of the proof is the same as the proof of Proposition 5.6 in \cite{FK}
modulo the following modifications.
\begin{enumerate}
\item
Replace the operator norm by the norm $\| \cdot \|_2$.
\item Use Lemmas \ref{Lem:AFD}, \ref{Lem:directed} in place of Lemmas 5.1, 5.5 in \cite{FK} respectively.
\end{enumerate}
(For item (2), since our approximating sequence consists of full matrix algebras, Lemma \ref{Lem:AFD}, a slightly weaker analogue of Lemma 5.1 in \cite{FK},
is sufficient for our purpose.)
\end{proof}
\begin{Rem}\label{Rem:Con}
We point out that the equivalence of approximate finiteness (A) to (C) for type II$_1$ factors of density character $\aleph_1$ can be shown without using Connes's classification theorem (but using the classification theorem of separable AFD type II$_1$ factors due to Murray and von Neumann \cite{MvN}).
Indeed, the following version of Lemma \ref{Lem:directed} holds true:
every type II$_1$ factor $M$ of density character $\aleph_1$ with approximate finiteness (B) admits a directed $\sigma$-complete family of
separable AFD subfactors with dense union indexed by $\omega_1$.
To see this, it suffices to show that any separable von Neumann subalgebra $N$ of $M$ is contained in a separable AFD subfactor of $M$.
(The rest of the proof is similar to that of Proposition \ref{Prop:AFD}.)
Choose a strongly dense sequence $(x_{n, 1})_{n=1}^\infty$ of $N$.
By the assumption, one can choose a finite dimensional $\ast$-subalgebra $F_1$ of $M$ with $\{ x_{1, 1}\} \subset^{1} F_1$.
Set $N_1 := \Wso(N \cup F_1)$.
Choose a strongly dense sequence $(x_{n, 2})_{n=1}^\infty$ of $N_1$.
Again by the assumption, one can choose a finite dimensional $\ast$-subalgebra $F_2$ of $M$ with $\{ x_{i, j}: 1\leq i, j \leq 2 \} \subset^{1/ 2} F_2$.
Set $N_2:= \Wso(N_1 \cup F_2)$.
By iterating this construction, we obtain an increasing sequence $(N_n)_{n=1}^\infty$
of separable von Neumann subalgebras of $M$ and a strongly dense sequence $(x_{n, m+1})_{n=1}^\infty$ in $N_m$ for each $m\in \mathbb{N}$ with the following property:
for any $m\in \mathbb{N}$, there is a finite dimensional $\ast$-subalgebra $F$ of $N_{m}$ with $\{ x_{i, j} :1\leq i, j \leq m \} \subset ^{1/m} F$.
This implies approximate finiteness (B) of $L_1 := \Wso(\bigcup_{n=1}^\infty N_n)$.
We have seen in the proof of Lemma \ref{Lem:directed}
that there is a separable subfactor $P_1$ of $M$ containing $L_1$.
By iterating the constructions, we obtain an increasing sequence
\[N \subset L_1 \subset P_1 \subset L_2 \subset P_2 \subset \cdots \subset L_i \subset P_i \subset \cdots\]
of separable von Neumann subalgebras of $M$ such
that each $L_i$ satisfies approximate finiteness (B) and that each $P_i$
is a factor.
Now set $Q:= \Wso(\bigcup_{i=1}^\infty L_i)= \Wso(\bigcup_{i=1}^\infty P_i)$.
Obviously $N \subset Q$.
By definition, $Q$ is separable and satisfies approximate finiteness (B),
while by the second equality, $Q$ is a subfactor of $M$.
Now by Murray and von Neumann's theorem \cite{MvN},
$Q$ is a (separable) AFD subfactor of $M$.
\end{Rem}
\section{Constructions of tensorially prime approximately finite dimensional operator algebras}\label{section:prime}
In this section, under a mild condition, we show primeness
and absence of regular abelian $\ast$-subalgebras of
the group von Neumann algebras of (uncountable) iterated wreath product groups.
As an application, we prove Theorem \ref{Thmint:prime}.
\begin{Prop}\label{Prop:groupvn}
Let $\alpha$ be a limit ordinal of uncountable cofinality.
Let $(\Gamma_\beta)_{\beta <\alpha}$ be a family
of non-trivial torsion-free groups
and let $G$ denote its iterated wreath product.
Then the group von Neumann algebra $L(G)$ is prime and has
no infinite dimensional regular abelian $\ast$-subalgebras.
\end{Prop}
\begin{proof}
The proof is similar to Popa's proof of the same results for uncountable free group factors \cite{Pop83}.
We first observe that $L(G)$ is a type II$_1$ factor by Lemma \ref{Lem:ICC}.
Observe that
for any countable subset $S$ of $L(G)$,
the set
\[T:=\{g\in G: \tau(s u_g^\ast)\neq 0 {\rm~for~some~}s\in S\}\]
is countable.
Since $\alpha$ is of uncountable cofinality, we have
\[\beta:=\sup \left\{ \deg(g): g\in T \right\} < \alpha.\]
Note that $S \subset L(G_\beta)$.
This observation together with Proposition 4.1 in \cite{Pop83} and Lemma \ref{Lem:disjoint} shows
that for any diffuse, countably generated von Neumann subalgebra
$B$ of $L(G)$,
there is $\beta <\alpha$ with $\mathcal{N}_{L(G)}(B)'' \subset L(G_\beta)$.
We now show that $L(G)$ has no infinite dimensional regular abelian $\ast$-subalgebras.
To see this, assume we have an infinite dimensional regular abelian $\ast$-subalgebra
$A \subset L(G)$. By taking the strong closure, we may assume that $A$ is strongly closed.
Then, by Lemma \ref{Lem:diffusevn}, $A$ is diffuse and contains the unit of $L(G)$.
Choose a countably generated, diffuse
von Neumann subalgebra $A_0 \subset A$.
By the observation in the previous paragraph,
one can choose $\beta <\alpha$ satisfying $A \subset \mathcal{N}_{L(G)}(A_0)'' \subset L(G_\beta)$.
By applying Proposition 4.1 in \cite{Pop83} and Lemma \ref{Lem:disjoint} to $A$ itself,
we conclude $\mathcal{N}_{L(G)}(A)'' \subset L(G_\beta) \subsetneq L(G)$.
This contradicts the regularity of $A$ in $L(G)$.
We now prove the primeness of $L(G)$.
To lead to a contradiction, assume we have a tensor product decomposition
$L(G) =M_1 \votimes M_2$ where both $M_i$ are infinite dimensional.
Then as $L(G)$ is a type II$_1$ factor, so are both $M_i$.
Choose a countably generated, diffuse von Neumann subalgebra
$B_i$ of $M_i$ for $i=1, 2$.
By the argument in the second previous paragraph,
one can choose $\beta <\alpha$ satisfying $\mathcal{N}_{L(G)}(B_i)'' \subset L(G_\beta)$ for $i=1, 2$.
This implies the relations $M_{1} \subset \mathcal{N}_{L(G)}(B_2)'' \subset L(G_\beta)$ and $M_2 \subset \mathcal{N}_{L(G)}(B_1)''\subset L(G_\beta)$.
Since $G_\beta \subsetneq G$, this is a contradiction.
\end{proof}
\begin{Cor}\label{Cor:primeAFD}
There is a prime AFD type II$_1$ factor without infinite dimensional regular abelian $\ast$-subalgebras.
\end{Cor}
\begin{proof}
Let $(\Gamma_{\alpha})_{\alpha < \omega_1}$
be a family of non-trivial torsion-free countable amenable groups.
Let $G$ denote the iterated wreath product of $(\Gamma_{\alpha})_{\alpha < \omega_1}$. By Lemmas \ref{Lem:lf} and \ref{Lem:ICC}, $G$ is amenable and ICC.
Hence $L(G)$ is an injective type II$_1$ factor.
Since $\sharp G= \aleph_1$,
the density character of $L(G)$ is $\aleph_1$.
Propositions \ref{Prop:AFD} and \ref{Prop:groupvn}
now imply that $L(G)$ possesses the desired properties.
\end{proof}
A unital \Cs-algebra is said to be {\it monotracial} if there is a unique tracial state on it.
\begin{Cor}\label{Cor:primeAF}
There is a tensorially prime simple unital monotracial AF-algebra without infinite dimensional regular abelian $\ast$-subalgebras.
\end{Cor}
\begin{proof}
This is a consequence of Propositions \ref{Prop:regular}, \ref{Prop:AFD}, and Corollary \ref{Cor:primeAFD}.
\end{proof}
\begin{Rem}
\begin{enumerate}
\item By Proposition \ref{Prop:AFD}, \Cs-algebras constructed in Corollary \ref{Cor:primeAF} are in fact {\it AM-algebras} in the sense of \cite{FK}, \cite{FHKT}. That is, they are obtained as the inductive limits of full matrix algebras.
\item
By slightly modifying the proof of Proposition \ref{Prop:AFD}, for any UHF-algebra $A$,
one can arrange a \Cs-algebra obtained in Corollary \ref{Cor:primeAF}
to have the same ordered K-theory as $A$.
To see this, take a sequence $(k_n)_{n=1}^\infty$ of natural numbers
satisfying $A \cong \bigotimes_{n=1}^\infty \mathbb{M}_{k_n}(\mathbb{C})$.
Set $K_n := \prod_{i=1}^n k_i$ for each $n\in \mathbb{N}$.
Then, for any separable injective type II$_1$ factor,
by Connes's classification theorem \cite{Con},
it admits an increasing sequence of subfactors of type $I_{K_n}$, $n\in \mathbb{N}$,
with dense union.
Use these approximation sequences instead of the one consisting of type $I_{2^n}$ subfactors to proceed the proof of Proposition \ref{Prop:AFD}.
Then the resulting inductive system consists of full matrix algebras of size $K_n$, $n\in \mathbb{N}$.
Thus the inductive limit \Cs-algebra satisfies the required properties.
\item It is not hard to see from the proof of Proposition \ref{Prop:AFD} (cf.~\cite{FK})
that the AF-algebras we obtained in Corollary \ref{Cor:primeAF}
are of density character $\aleph_1$.
\end{enumerate}
\end{Rem}
We will see in the next section that the group von Neumann algebras
in Proposition \ref{Prop:groupvn} have no non-trivial central sequences (see Theorem \ref{Thm:nocentral}).
\section{Central sequences in certain crossed products}\label{section:central}
In this section, we study central sequences of the reduced crossed products
of non-commutative Bernoulli shifts of iterated wreath product groups.
We first introduce a few notations.
Let $A$ be a unital \Cs-algebra.
We set $\bigotimes_{\emptyset} A := \mathbb{C}$ for convenience.
For a group $G$ and for a $G$-set $S$ (possibly empty),
we equip $\bigotimes_S A$ with the
tensor shift $G$-action induced from the $G$-action on $S$.
We set \[A \wrr{S} G := \left(\bigotimes_S A \right) \rc G.\]
When $H$ is a subgroup of $G$ and
$T$ is a (possibly empty) $H$-invariant subset of $S$,
we identify
$A \wrr{T} H$
with a \Cs-subalgebra of $A\wrr{S} G$ in the canonical way.
Here we regard $T$ as an $H$-set by restricting the original action.
For a group $G$, we equip $G$ with the left translation $G$-action.
In the most important case $S=G$,
we simply denote $A \wrr{G} G$ by $A \wr G$ for short.
We use similar notations for von Neumann algebras.
For instance, for a von Neumann algebra $M$ with a faithful normal state $\varphi$
and for a group $G$, we set $(M, \varphi) \vwr G := \left[\bigvotimes_G (M, \varphi) \right]\vnc G$.
\begin{Thm}\label{Thm:nocentral}
Let $\alpha$ be a limit ordinal of uncountable cofinality.
Let $(\Gamma_\beta)_{\beta <\alpha}$ be a family of non-trivial groups
and let $G$ denote its iterated wreath product. Then the following statements hold true.
\begin{enumerate}\upshape
\item
Let $A$ be a unital \Cs-algebra.
Then $A \wr G$
has no non-trivial central sequences.
\item
Let $M$ be a von Neumann algebra with a faithful normal state $\varphi$.
Then $(M, \varphi)\vwr G$
has no non-trivial central sequences.
\end{enumerate}
\end{Thm}
\begin{proof}
We only prove (1).
The statement (2) can be shown by a similar argument.
Let $(x_n)_{n=1}^\infty$ be a central sequence in $A \wr G$.
Since $\alpha$ is of uncountable cofinality, one can find
$\beta < \alpha$ satisfying $\{ x_n : n \in \mathbb{N} \} \subset A \wr G_\beta$.
Fix a state $\varphi$ on $A$.
For any subgroup $H$ of $G$
and any left $H$-invariant subset $S$ of $G$,
by Exercise 4.1.4 in \cite{BO},
one can find a conditional expectation
$E_{H, S}^{\varphi} \colon A \wr G \rightarrow A \wrr{S} H$ satisfying
\[E_{H, S}^{\varphi} (x u_g) = \Phi_S(x)\chi_H(g)u_g {\rm~for~} x\in \bigotimes_G A, g\in G,\]
where $\Phi_S:= \left(\bigotimes_S \id_{A}\right) \otimes (\bigotimes_{G\setminus S}\varphi)$ and $\chi_H$ denotes the characteristic function of $H$ on $G$.
Observe that for two subgroups $H_1, H_2 \subset G$
and left $H_i$-invariant subsets $S_i \subset G$; $i=1, 2$,
we have
\[E_{H_1, S_1}^{\varphi} (A \wrr{S_2} H_2) = A \wrr{S_1 \cap S_2} (H_1 \cap H_2).\]
Choose $g\in \Gamma_{\beta} \setminus \{e\}$.
We identify $\Gamma_{\beta}$ with the $e$-th direct summand
of $\bigoplus_{ G_\beta} \Gamma_{\beta} \subset G_{\beta +1}$.
Obviously $g G_\beta g^{-1} \cap G_\beta = \{ e\}$,
$g G_\beta \cap G_\beta = \emptyset$.
Since $u_g x_n u_g^\ast \in A \wrr{g G_\beta}(g G_\beta g^{-1})$,
we have $c_n:=E_{G_\beta, G_\beta}^{\varphi} (u_g x_n u_g^\ast) \in \mathbb{C}$.
Since $E_{G_\beta, G_\beta}^{\varphi} (x_n)=x_n$ for all $n$
and $\lim_{n \rightarrow \infty} \| x_n - u_g x_n u_g^\ast\| =0$,
we conclude
$\lim_{n \rightarrow \infty} \| x_n - c_n\| =0.$
\end{proof}
\begin{Rem}
The proof of Theorem \ref{Thm:nocentral} in fact shows the following slightly stronger statement:
for any iterated wreath product groups $G_1, \ldots, G_n$ as in Theorem \ref{Thm:nocentral}
and for any unital \Cs-algebras $A_1, \ldots, A_n$,
the tensor product
$\bigotimes_{i=1}^n ( A_i \wr G_i)$ has no non-trivial central sequences.
(The tensor products of conditional expectations used in the proof
of Theorem \ref{Thm:nocentral} play their role.)
A similar generalization is valid for von Neumann algebras.
\end{Rem}
We next give the following useful lemma on pure infiniteness of the reduced crossed product.
This is an immediate consequence of Lemma 3.2 of \cite{Kis}.
For the definition and consequences of pure infiniteness, we refer the reader to Chapter 4 of \cite{Ror02}.
Recall that an automorphism $\alpha$ of a \Cs-algebra $A$ is said to be {\it outer}
if there is no unitary element $u$ in the multiplier algebra $M(A)$
satisfying $\alpha(a)= uau^\ast$ for all $a\in A$.
An action $\alpha \colon \Gamma \curvearrowright A$ of a discrete group $\Gamma$ on a \Cs-algebra $A$ is said to be {\it outer} if
for any $s\in \Gamma \setminus \{e\}$, the automorphism $\alpha_s$ is outer.
\begin{Lem}\label{Lem:pi}
Let $\Gamma$ be a group.
Let $A$ be a simple purely infinite \Cs-algebra.
Let $\alpha \colon \Gamma \curvearrowright A$ be an outer action.
Then $A \rc \Gamma$ is simple and purely infinite.
\end{Lem}
\begin{proof}
It suffices to show the following claim.
For any $x = \sum_{g\in F} a_g u_g \in A \rc \Gamma$; $a_g\in A$, $e\in F \subset \Gamma$ finite subset,
with $a_e \geq 0$, $\| a_e\| >1$, for any positive element $b\in A$ of norm one, and for any $\epsilon>0$,
there is a contractive element $c\in A$ with
$\| c^\ast x c -b\| < \epsilon$.
By Lemma 3.2 of \cite{Kis},
one can choose a positive element $d_1 \in A$
satisfying $\|d_1 \| =1$,
$\| d_1 a_e d_1 \| \geq 1$,
and $ \sum _{g \in F \setminus \{e \}}\| d_1 a_g \alpha_g(d_1)\| < \epsilon/2$.
This implies
$\| d_1xd_1 - d_1 a_e d_1 \| <\epsilon/2$.
Since $A$ is simple and purely infinite,
there is $d_2\in A$
satisfying $\| d_2^\ast d_1 a_e d_1 d_2 - b \| <\epsilon/2$
and $\| d_2 \| \leq 1$.
Combining these inequalities,
we obtain
$\| d_2^\ast d_1 x d_1 d_2 -b \| < \epsilon$.
Thus $c:= d_1 d_2$ gives the desired element.
\end{proof}
The following corollary in particular shows that Kirchberg's $\mathcal{O}_\infty$-absorption theorem (\cite{KP}, Theorem 3.15)
fails in the non-separable setting.
\begin{Cor}\label{Cor:NSKT}
For any cardinal $\kappa > \aleph_0$,
there are operator algebras of density character $\kappa$ with no non-trivial central sequences in the following classes.
\begin{enumerate}\upshape
\item
Simple unital monotracial AF-algebras,
\item
AFD type II$_1$ factors,
\item
AFD type III factors,
\item
simple unital nuclear purely infinite \Cs-algebras.
\end{enumerate}
\end{Cor}
\begin{proof}
(1):
Choose a family $( \Gamma_\alpha)_{\alpha< \omega_1}$
of locally finite non-trivial groups
satisfying $\sharp \Gamma_0= \kappa$ (e.g., $\Gamma_0 := \bigoplus_\kappa \mathbb{Z}_2$) and $\sharp \Gamma_\alpha \leq \aleph_0$ for $0<\alpha <\omega_1$.
Let $G$ be the iterated wreath product group of the family $(\Gamma_\alpha)_{\alpha< \omega_1}$.
By Lemma \ref{Lem:lf}, $G$ is locally finite.
Set $A:=\mathbb{M}_2(\mathbb{C}) \wr G$.
Clearly $A$ is unital.
By Theorem 3.1 of \cite{Kis}, $A$ is simple.
It follows from the construction of $G$ that $\sharp G= \kappa$.
Hence the density character of $A$ is $\kappa$.
Take an increasing net $(H_\lambda)_{\lambda \in \Lambda}$
of finite subgroups of $G$ with $\bigcup_{\lambda \in \Lambda} H_\lambda =G$.
Then the net $(\mathbb{M}_2(\mathbb{C}) \wr H_{\lambda})_{\lambda \in \Lambda}$ shows that $A$ is an AF-algebra.
We also observe that $A$ is realized as an increasing union
of the net $(\mathbb{M}_2(\mathbb{C}) \wr_G H_{\lambda})_{\lambda \in \Lambda}$.
Hence Proposition 2.1 in \cite{ORS} yields
the uniqueness of tracial states on $A$.
By Theorem \ref{Thm:nocentral}, $A$ has no non-trivial central sequences.
(2): Let $G$ be as in the proof of (1).
Put $M:=(\mathbb{M}_2(\mathbb{C}), \tau)\vwr G$.
It follows from Theorem \ref{Thm:nocentral}
that $M$ has no non-trivial central sequences.
In particular $M$ is a factor.
By the same reason as in (1),
$M$ possesses the desired properties.
(3): Again let $G$ be as in the proof of (1). Take a faithful non-tracial state $\varphi$ on $\mathbb{M}_2(\mathbb{C})$.
Then, by Exercise 13.4.2 in \cite{KR},
$(M, \varphi) \vwr G$ is of type III.
The other required conditions are confirmed by the same way as (2).
(4):
Choose a family $( \Gamma_\alpha)_{\alpha< \omega_1}$
of amenable groups
satisfying $\sharp \Gamma_0= \kappa$ and $\sharp \Gamma_\alpha = \aleph_0$ for $0<\alpha <\omega_1$.
Let $G$ be the iterated wreath product group of the family $(\Gamma_\alpha)_{\alpha< \omega_1}$.
Fix a Kirchberg algebra $B$.
Set $A:=B \wr G$.
Clearly $A$ is unital.
By Theorem 4.2.6 in \cite{BO}, $A$ is nuclear.
By Lemma \ref{Lem:pi}, $A$ is simple and purely infinite.
Since $\sharp G= \kappa$, the density character of $A$ is $\kappa$.
Again by Theorem \ref{Thm:nocentral}, $A$ has no non-trivial central sequences.
\end{proof}
\begin{Rem}\label{Rem:Cartan}
In contrast to the absence of regular abelian $\ast$-subalgebras in Theorem \ref{Thmint:prime}, one can arrange \Cs-algebras stated in Corollary \ref{Cor:NSKT}
to be a crossed product of a group acting on an abelian operator algebra.
Indeed, for (4), choose $B$ in the proof to be a Kirchberg algebra of the form
$C(X) \rc \Gamma$ for some amenable minimal topologically free dynamical system $\Gamma \curvearrowright X$
(see e.g., \cite{Suz13}).
Let $G$ denote the iterated wreath product group as in the proof of Corollary \ref{Cor:NSKT}.
We equip $X^G$ with the action of $\Gamma \wr G$ defined as follows.
The group $\bigoplus_G \Gamma$ acts on $ X^G$ coordinate-wise
and $G$ acts on $X^G$ by the left shifts.
We then obtain an isomorphism
\[A:=B \wr G \cong C(X^G) \rc (\Gamma \wr G).\]
We point out that the action $\Gamma \wr G \curvearrowright X^G$ is minimal
and topologically free. Thus
$C(X^G)$ in fact gives rise a {\it Cartan subalgebra} \cite{Ren08} of $A$.
A similar remark applies to operator algebras constructed in the proofs
of (1) to (3) of Theorem \ref{Thm:nocentral}.
(To see this, note that $\mathbb{M}_2(\mathbb{C}) \cong C(\mathbb{Z}_2) \rc \mathbb{Z}_2$ where
$\mathbb{Z}_2$ acts on itself by the left translations.)
\end{Rem}
\begin{Rem}
Recall that the cofinality of a limit ordinal $\alpha$ is
the cardinal \[{\rm cf}(\alpha):=\min\{\sharp S:S \subset \alpha: \sup(S) =\alpha\}.\]
By slightly modifying the proofs of Theorem \ref{Thm:nocentral} and Corollary \ref{Cor:NSKT} (replacing $\omega_1$ by $\alpha$),
one can show the following generalized statement:
for any limit ordinal $\alpha$,
one can construct operator algebras of density character $\sharp \alpha$ from each classes (1) to (4) of Corollary \ref{Cor:NSKT} with the following property.
There is no non-trivial central net $(x_i)_{i \in I}$ with $\sharp I< {\rm cf}(\alpha)$. (Here centrality and triviality of a bounded net is defined analogously to the sequence case.)
Note that for \Cs-algebras, this condition implies the non-existence of non-trivial tensor product components
of density character less than ${\rm cf}(\alpha)$.
We point out that when $\kappa$ is a successor (infinite) cardinal,
then ${\rm cf}(\kappa)$ is equal to $\kappa$.
In general, for any cardinals $\lambda < \kappa$,
one can find an ordinal $\alpha$ with
$\sharp\alpha=\kappa$, ${\rm cf}(\alpha)>\lambda$.
To see this, choose a successor cardinal $\mu$
with $\lambda < \mu \leq \kappa$.
Then the ordinal $\alpha:= \kappa + \mu$ possesses the desired property.
\end{Rem}
\section{Rigid sides of simple \Cs-algebras of decomposition rank one}\label{section:dr1}
We close this article by giving examples of simple unital monotracial \Cs-algebras which are tensorially prime and have decomposition rank one. This is another application of Popa's orthogonality method \cite{Pop83}.
\begin{Thm}\label{Thm:dr1}
For any cardinal $\kappa > \aleph_0$,
there is a tesorially prime simple unital monotracial \Cs-algebra of decomposition rank one, of density character $\kappa$, with no non-trivial central sequences.
\end{Thm}
\begin{proof}
Define $\Gamma_0 := \mathbb{Z}$, $\Gamma_1 := \bigoplus_\kappa \mathbb{Z}$,
and
$\Gamma_\alpha:=\mathbb{Z}$ for $2\leq \alpha<\omega_1$.
Let $G$ be the iterated wreath product group of the family $(\Gamma_\alpha)_{\alpha<\omega_1}$.
We show that $\mathbb{M}_2(\mathbb{C}) \wr G$ possesses the required properties.
We have already seen in the proof of Corollary \ref{Cor:NSKT} that $\mathbb{M}_2(\mathbb{C}) \wr G$ is simple, unital, monotracial, of density character $\kappa$, with no non-trivial central sequences.
To show that the decomposition rank of $\mathbb{M}_2(\mathbb{C}) \wr G$ is one,
as explained in Examples 6.1 (i) of \cite{KW}, it suffices to show that
$\mathbb{M}_2(\mathbb{C}) \wr G$ has nonzero K$_1$-group and has decomposition rank at most one.
To show the first claim, we define subgroups $H_\alpha \subset G_\alpha$, $\alpha \leq\omega_1$
as follows. Set $H_0:= \{e\}$, $H_1 := \{e \}$.
Assume we have defined $(H_\beta)_{\beta<\alpha}$ for some $1 < \alpha\leq\omega_1$.
We then define $H_\alpha$ as follows.
When $\alpha$ is a limit ordinal,
set $H_\alpha:= \bigcup_{\beta<\alpha} H_\beta$.
When $\alpha$ has the predecessor $\gamma$,
set $H_\alpha$ to be the subgroup of $G_\alpha$
generated by $\bigoplus_{G_\gamma}\Gamma_\gamma$ and $H_\gamma$.
By transfinite induction, we obtain the increasing family $(H_\alpha)_{\alpha\leq\omega_1}$ of subgroups in $G$.
We set $H:= H_{\omega_1}$.
It is not hard to show by transfinite induction that for each $1 \leq \alpha<\omega_1$, $H_\alpha$ is normal in $G_\alpha$,
$H_\alpha \cap G_1 = \{e\}$, and $H_\alpha \cup G_1$ generates $G_\alpha$.
This yields that $H$ is normal in $G$
and $G/H \cong G_1$.
Since $G_1$ is isomorphic to $\mathbb{Z}$,
we obtain the semidirect product decomposition
$G= H \rtimes G_1$.
This induces the crossed product decomposition
\[\mathbb{M}_2(\mathbb{C}) \wr G =A\rc G_1,\]
where $A:=\mathbb{M}_2(\mathbb{C}) \wrr{G} H$.
Since $A$ is unital and admits a tracial state,
the unit element $[1_A]_0$ is nonzero in $K_0(A)$.
The Pimsner--Voiculescu exact sequence \cite{PV}
then implies that $K_1(\mathbb{M}_2(\mathbb{C}) \wr G)$ is nonzero.
We next show the second claim.
By Theorem 2.1 of \cite{ORS} and Theorem 1.1 in \cite{Sat}, for each countable subgroups $L \subset K$ of $G$ with
$[K: L]=\infty$,
$\mathbb{M}_2(\mathbb{C}) \wr_K L$ is monotracial and $\mathcal{Z}$-stable.
By Theorem F of \cite{STW} and Theorem 3.8 of \cite{ORS}, each $\mathbb{M}_2(\mathbb{C}) \wr_K L$
has decomposition rank at most one.
This proves the second claim.
We next show the tensorial primeness of $\mathbb{M}_2(\mathbb{C}) \wr G$.
We identify $\mathbb{M}_2(\mathbb{C}) \wr G$
with a strongly dense \Cs-subalgebra of $(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G$ in the canonical way.
To lead to a contradiction, assume we have a tensor product decomposition
\[\mathbb{M}_2(\mathbb{C}) \wr G = A_1 \otimes A_2\]
with both $A_i$ infinite dimensional.
By taking the strong closure, we obtain a tensor product decomposition
\[(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G =M_1 \votimes M_2,\]
where both $M_1, M_2$ are type II$_1$ factors.
We show that there is $\alpha< \omega_1$
satisfying $M_i \subset (\mathbb{M}_2(\mathbb{C}), \tau) \vwrr{G} G_\alpha$ for $i=1, 2$, which is obviously a contradiction.
To find such $\alpha$, choose a separable type II$_1$ subfactor
$N_i$ of $M_i$ for $i=1, 2$.
Take $\alpha<\omega_1$
satisfying $N_i \subset (\mathbb{M}_2(\mathbb{C}), \tau) \vwr G_\alpha$ for $i=1, 2$.
Let $U_1$, $U_2$ denote the unitary group
of the \Cs-subalgebras $\bigotimes _{G_\alpha} \mathbb{M}_2(\mathbb{C})$,
$\bigotimes _{G\setminus G_\alpha} \mathbb{M}_2(\mathbb{C})$ of $\bigvotimes _G(\mathbb{M}_2(\mathbb{C}), \tau)$ respectively.
For any $g\in G\setminus G_\alpha$, $v_1 \in U_1$, $v_2 \in U_2$, and $i=1, 2$,
we will show the relation
\[v_1 v_2 u_g N_i u_g^\ast v_2^\ast v_1^\ast \perp N_i.\]
Since the product $U_1 \cdot U_2$ spans a strongly dense subspace of $\bigvotimes _G(\mathbb{M}_2(\mathbb{C}), \tau)$,
this together with Corollary 2.6 of \cite{Pop83} concludes the desired inclusions
\[M_1 \subset \mathcal{N}_{(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G}(N_2)''
\subset (\mathbb{M}_2(\mathbb{C}), \tau) \vwrr{G} G_\alpha,\]
\[M_2 \subset \mathcal{N}_{(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G}(N_1)''
\subset (\mathbb{M}_2(\mathbb{C}), \tau) \vwrr{G} G_\alpha.\]
To prove the claim, as $U_1, N_1, N_2 \subset (\mathbb{M}_2(\mathbb{C}), \tau) \vwr G_\alpha$, it is enough to show
\[v_2 u_g [(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G_\alpha] u_g^\ast v_2^\ast \perp [(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G_\alpha]\]
for $g\in G \setminus G_\alpha$ and $v_2 \in U_2$.
To see this, it suffices to show the equation
\[\tau(v_2u_g xu_s u_g^\ast v_2^\ast y u_t) = \tau(v_2u_g x u_s u_g^\ast v_2^\ast)\tau(yu_t)\]
for $g\in G \setminus G_\alpha$, $v_2 \in U_2$, $x, y \in \bigotimes_{G_\alpha}\mathbb{M}_2(\mathbb{C})$, $s, t \in G_\alpha$.
When $\{s, t\} \neq \{e\}$,
by Lemma \ref{Lem:disjoint}, we have
\[\tau(v_2u_g xu_s u_g^\ast v_2^\ast y u_t) = 0=\tau(v_2u_g x u_s u_g^\ast v_2^\ast)\tau(yu_t).\]
We next consider the case $s= t=e$. In this case, observe that $v_2 u_g x u_g^\ast v_2^\ast \in \bigotimes_{G \setminus G_\alpha}\mathbb{M}_2(\mathbb{C})$,
while $y \in \bigotimes_{G_\alpha}\mathbb{M}_2(\mathbb{C})$.
We thus obtain
\[\tau(v_2u_g x u_g^\ast v_2^\ast y ) = \tau(v_2u_g x u_g^\ast v_2^\ast)\tau(y).\]
\end{proof}
\begin{Rem}
Let $G$ be the iterated wreath product of a family $(\Gamma_\alpha)_{\alpha<\omega_1}$,
where each $\Gamma_\alpha$ is countable infinite, amenable, and torsion-free.
Then, by the proof of Theorem \ref{Thm:dr1}, Proposition \ref{Prop:AFD}, and Remark \ref{Rem:Cartan}, $(\mathbb{M}_2(\mathbb{C}), \tau) \vwr G$ is
a prime AFD type II$_1$ factor with a regular maximal abelian $\ast$-subalgebra.
This construction also provides prime amenable type II$_1$ factors of
prescribed uncountable density character with a regular maximal abelian $\ast$-subalgebra. We do not know if these von Neumann algebras are AFD.
\end{Rem}
\subsection*{Acknowledgements}
Some parts of this work were carried out while the author was visiting Hokkaido University. The author is grateful to Reiji Tomatsu for his kind hospitality.
He also thanks Shuhei Masumoto for helpful comments on ordinals.
Finally, the author would like to thank the referees for their careful reading and valuable comments.
This work was supported by JSPS KAKENHI Grant-in-Aid for Young Scientists
(Start-up, No.~17H06737) and tenure track funds of Nagoya University.
|
\section{Introduction}
The discovery of the Higgs boson ($H$) by the ATLAS and CMS Collaborations at the Large Hadron Collider (LHC) in
2012~\cite{Aad:2012tfa,Chatrchyan:2012xdj} has confirmed the Brout-Englert-Higgs (BEH) mechanism of electroweak
symmetry breaking and mass generation~\cite{Englert:1964et,Higgs:1964pj}. The Higgs boson mass is about
125~GeV~\cite{Aad:2015zhl} and its measured properties are in agreement with the Standard Model (SM)
predictions~\cite{Khachatryan:2016vau}. However, the BEH mechanism does not only predict the existence
of a massive scalar particle, but it also requires the Higgs boson to couple to itself. In order to measure the
Higgs potential and thereby have a complete description of the SM, it is necessary to observe the production
of Higgs boson pairs and measure the Higgs self-coupling $\lambda_{HHH}$. Also, any deviation from the
SM predictions would open a window on new physics.\\
In the SM, Higgs boson pairs ($HH$) can be produced either in a heavy-quark loop or via Higgs self-coupling,
see the left-hand and central diagrams of Figure~\ref{fig:hh-diagrams}, respectively. However, due to the destructive
interference between the two diagrams, the SM Higgs boson pair production is only 33.41~fb at 13~TeV, as computed
at the next-to-leading-order (NLO) in QCD and fully accounting for top-quark mass effects~\cite{deFlorian:2016spz,Grazzini:2018bsd}.
Still, the $HH$ cross-section can be significantly enhanced in several Beyond-the-Standard-Model (BSM) scenarios, e.g.\ through
anomalous couplings, new contact interactions, or resonant production of $HH$ pairs, as illustrated in the right-hand diagram of
Figure~\ref{fig:hh-diagrams}.
Several BSM theories indeed predict the existence of heavy particles decaying into a pair of Higgs bosons. For instance,
two-Higgs-doublet models (2HDMs)~\cite{Djouadi:2005gj,Branco:2011iw} have a second CP-even Higgs boson, which may be
heavy enough to decay into two SM-like lighter Higgs bosons. Alternatively, $HH$ pairs can be produced in the decay of
a spin-2 graviton, as predicted in the Randall-Sundrum model of warped extra dimensions~\cite{Randall:1999ee}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.32\textwidth]{CIPANP2018-Ferrari-figure01a}
\includegraphics[width=0.32\textwidth]{CIPANP2018-Ferrari-figure01b}
\includegraphics[width=0.32\textwidth]{CIPANP2018-Ferrari-figure01c}
\caption{\label{fig:hh-diagrams} Leading-order Feynman diagrams for $HH$ production in the gluon-gluon fusion mode, via
(left) a heavy-quark loop and (centre) Higgs self-coupling in the SM, or through (right) an intermediate resonance ($X$).}
\end{figure}
Searches for non-resonant and resonant $HH$ production were conducted by the ATLAS collaboration, based on
proton-proton ($pp$) collision data corresponding to an integrated luminosity of up to 36.1~fb$^{-1}$. The ATLAS
experiment~\cite{PERF-2007-01,Capeans:2010jnh,ATLAS:2012rar} is a multi-purpose detector
with a forward-backward symmetric cylindrical geometry and nearly $4\pi$ coverage in solid angle.\footnote{ATLAS uses a right-handed
coordinate system with its origin at the nominal interaction point (IP) in the centre of the detector and the $z$-axis along the beam
pipe. The $x$-axis points from the IP to the centre of the LHC ring, and the $y$-axis points upwards. Cylindrical coordinates
$(r,\phi)$ are used in the transverse plane, $\phi$ being the azimuthal angle around the $z$-axis. The pseudo-rapidity is
defined in terms of the polar angle $\theta$ as $\eta = -\ln \tan(\theta/2)$, and the angular distance is measured in units of
$\Delta R \equiv \sqrt{(\Delta\eta)^{2} + (\Delta\phi)^{2}}$.} It consists of an inner tracking detector
surrounded by a thin superconducting solenoid providing a 2~T axial magnetic field, electromagnetic and hadron calorimeters,
and a muon spectrometer. A two-level trigger~\cite{Aaboud:2016leb,ATL-DAQ-PUB-2017-001} reduces the event rate to a
maximum of 1~kHz prior to offline data storage. Three search results were recently published by the ATLAS Collaboration, in the
$HH \to bbbb$~\cite{Aaboud:2018knk}, $HH \to bb\gamma\gamma$~\cite{Aaboud:2018ftw} and
$HH \to WW\gamma\gamma$~\cite{Aaboud:2018ewm} channels. They are summarised in the following sections of these proceedings.
\section{Search for $HH \to bbbb$ in ATLAS}
With a branching fraction of 33\%, $HH \to bbbb$ has the largest event rate among all Higgs boson pair production channels.
However, it also suffers from a large multi-jet background, which requires the use of novel data-driven estimation techniques.
Depending on the $HH$ production mode and the probed mass range, two event topologies are considered:
\begin{itemize}
\item For both the non-resonant $HH$ production mode and the search for a resonance decaying via $HH \to bbbb$ with a mass in the
range 260--1400~GeV, a resolved topology is considered, where at least four anti-$k_T$ jets~\cite{Cacciari:2008gp} with a parameter
radius $R=0.4$ are reconstructed with a transverse momentum $p_{\mathrm{T}} > 40~\mbox{GeV}$, and the number of $b$-tagged
jets (i.e.\ jets compatible with a $b$-hadron decay) is required to be exactly four (the efficiency of the working point is 70\% in simulated
$t\bar{t}$ events). Such events are recorded based on a combination of $b$-tagged jet triggers. Due to an inefficiency of vertex
reconstruction and thereby $b$-tagging at the trigger level, a fraction of the data collected in 2016 was not retained, hence the
integrated luminosity for this event topology is only 27.5~fb$^{-1}$.
\item For the search for a resonance decaying via $HH \to bbbb$ with a mass in the range 800--3000~GeV, a boosted topology
is considered, where at least two anti-$k_T$ jets with a larger parameter radius $R=1.0$ are reconstructed (including one firing the
corresponding trigger). The leading (sub-leading) large-$R$ jet is required to have $p_{\mathrm{T}} > 450~(250)~\mbox{GeV}$.
Then, $b$-tagging is performed on track-jets with a parameter radius $R=0.2$ and at least one $b$-tag per large-$R$ jet is
required, which yields three event categories with two, three or four $b$-tags.
\end{itemize}
\subsection{Resolved topology}
For events with a resolved topology, the four jets with highest $b$-tagging scores are used. The selection and pairing of jets
into Higgs boson candidates is performed using angular distances between jets ($\Delta R_{jj}$), the four-jet invariant mass
$m_{4j}$ and differences in $m_{2j}$. Then, $m_{4j}$- and $m_{2j}$-dependent requirements on the $p_{\mathrm{T}}$ and mass
of the Higgs boson candidates are applied. Events in which a three-jet combination is compatible with a top-quark decay are
vetoed to reduce the $t\bar{t}$ background contamination. As a result, the signal region is defined by a small area in the
($m_{2j}^{\mathrm{lead}}$;\,$m_{2j}^{\mathrm{sub-lead}}$) phase-space centered at (120~GeV;\,110~GeV), corresponding to
the reconstructed masses of the two (leading and sub-leading) $H \to bb$ candidates, as illustrated by the red dashed line in
the left-hand plot of Figure~\ref{fig:bbbb-first}. The efficiency of the various event selections is shown in the central (right-hand)
plot of Figure~\ref{fig:bbbb-first} for various masses of a spin-0 resonance decaying to a Higgs boson pair (non-resonant $HH$
production).
\begin{figure}[htbp]
\centering
\includegraphics[height=0.2\textheight]{CIPANP2018-Ferrari-figure02a}
\includegraphics[height=0.2\textheight]{CIPANP2018-Ferrari-figure02b}
\includegraphics[height=0.2\textheight]{CIPANP2018-Ferrari-figure02c}
\caption{\label{fig:bbbb-first} Analysis regions in the ($m_{2j}^{\mathrm{lead}}$;\,$m_{2j}^{\mathrm{sub-lead}}$) phase-space
(left) where the area within the red dashed line is the signal region, while the area between the orange and yellow full lines is the
sideband region used for data-driven background estimation, and event selection efficiencies for a spin-0 resonance decaying to
a Higgs boson pair (centre) and non-resonant $HH$ production (right)~\cite{Aaboud:2018knk}.}
\end{figure}
The dominant multi-jet background is estimated with data. A data sample is built with the nominal event selection,
but the number of $b$-jets is requested to be exactly two: one $H$ candidate is reconstructed from the two $b$-tagged jets, and
the other one from two non-$b$-tagged jets. Re-weighting is then applied to this $2b$+$2j$ sample. The weights are derived
by comparing $2b$+$2j$ and $4b$ samples in a sideband region, located between the orange and yellow full lines in the
($m_{2j}^{\mathrm{lead}}$;\,$m_{2j}^{\mathrm{sub-lead}}$) phase-space: a per-non-$b$-tagged-jet factor is obtained by
comparing jet multiplicities and a global event weight is obtained from the ratio of $4b$ and $2b$+$2j$ templates (after
subtracting $t\bar{t}$ events) for five variables sensitive to differences in $b$-tagging. As for the $t\bar{t}$ background,
its shape is taken from simulation, but the normalisations of the fully-hadronic and semi-leptonic $t\bar{t}$ backgrounds are
determined together with the multi-jet event yield by a simultaneous fit in three background-enriched regions of the sideband
region. A validation of the background estimate is performed in a dedicated region, which is between the orange full
line and the red dashed line in the left-hand plot of Figure~\ref{fig:bbbb-first}, and a good agreement between the
predicted and measured $m_{4j}$ distribution is found.\\
Figure~\ref{fig:bbbb-second} shows the invariant mass distribution the two Higgs boson candidates in the signal
region, split between the 2015 and 2016 datasets, as different trigger configurations were used. The largest local
deviation has a statistical significance of 3.6$\sigma$ at 280~GeV, while the global significance is 2.3$\sigma$.
\begin{figure}[htbp]
\centering
\includegraphics[height=0.2\textheight]{CIPANP2018-Ferrari-figure03a}
\includegraphics[height=0.2\textheight]{CIPANP2018-Ferrari-figure03b}
\caption{\label{fig:bbbb-second} Reconstructed invariant mass of the two $H \to bb$ candidates in the signal region of the resolved
topology, split by data-taking year~\cite{Aaboud:2018knk}.}
\end{figure}
\subsection{Boosted topology}
In the boosted topology, the two large-$R$ jets with highest $p_{\mathrm{T}}$ are used in order to reconstruct the two
$H \to bb$ candidates, with an angular distance $|\Delta \eta_{JJ}| < 1.7$. Following further requirements on the jet
masses, a signal region is defined in the ($m_{J}^{\mathrm{lead}}$;\,$m_{J}^{\mathrm{sub-lead}}$) phase-space centered at
(124~GeV;\,115~GeV), corresponding to the reconstructed masses of the two (leading and sub-leading) $H \to bb$ candidates,
each reconstructed as a single large-$R$ jet. Similarly to the data-driven estimation method employed in the case of resolved
topologies, multi-jet templates are built from "lower-tagged" event selections (i.e. in which one of the large-$R$ jet has no
$b$-tagged track-jet and at least one failing $b$-tagging) and the kinematic distributions of the non-$b$-tagged $J$ is
re-weighted in order to mimic a $H$ candidate. The shape of the $t\bar{t}$ background is taken from simulation. Finally,
the normalisation of the backgrounds is obtained from binned likelihood fits of the leading large-$R$ jet mass distribution
in a sideband region. After successful validation of the background model in a dedicated region between the sideband and
signal regions of the ($m_{J}^{\mathrm{lead}}$;\,$m_{J}^{\mathrm{sub-lead}}$) phase-space, the $m_{JJ}$ distribution
in the signal region is used as a discriminant, after correction of the large-$R$ jet momenta by $m_H/m_J$, separately
for events with two, three and four $b$-tags, see Figure~\ref{fig:bbbb-third}.
\begin{figure}[htbp]
\centering
\includegraphics[height=0.16\textheight]{CIPANP2018-Ferrari-figure04a}
\includegraphics[height=0.16\textheight]{CIPANP2018-Ferrari-figure04b}
\includegraphics[height=0.16\textheight]{CIPANP2018-Ferrari-figure04c}
\caption{\label{fig:bbbb-third} Reconstructed invariant mass of the two $H \to bb$ candidates in the signal region of the boosted
topology, split by number of $b$-tags~\cite{Aaboud:2018knk}.}
\end{figure}
\subsection{Results}
The 95\% confidence-level (CL) exclusion limits on the non-resonant $HH$ production are shown in Table~\ref{tab:bbbb}, in
units of the SM prediction. In the case of $HH \to bbbb$ production through a spin-0 resonance, a statistical combination of
the resolved and boosted topologies is performed in the mass range where they overlap, i.e.\ 800--1400~GeV. The 95\% CL
upper limits on the resonant production cross-section are shown in Figure~\ref{fig:bbbb-fourth}.
\begin{table}[htbp]
\begin{center}
\caption{95\% CL limits on non-resonant Higgs boson pair production, from the search for $HH \to bbbb$ in ATLAS.}
\vspace*{2mm}
\begin{tabular}{cccccc}
\hline
Observed & $-2\sigma$ & $-1\sigma$ & Expected & $+1\sigma$ & $+2\sigma$ \\
\hline
13.0 & 11.1 & 14.9 & 20.7 & 30.0 & 43.5 \\
\hline
\end{tabular}
\label{tab:bbbb}
\end{center}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[height=0.24\textheight]{CIPANP2018-Ferrari-figure05}
\caption{\label{fig:bbbb-fourth} 95\% CL limits on Higgs boson pair production from a spin-0 resonance, as a function of the hypothetised
mass, from the search for $HH \to bbbb$ in ATLAS~\cite{Aaboud:2018knk}.}
\end{figure}
\section{Search for $HH \to bb\gamma\gamma$ in ATLAS}
The search for $HH \to bb\gamma\gamma$ suffers from a low signal branching fraction (0.26\%), but the presence of a di-photon system
with a good mass resolution provides a very clean signature. The event selection starts with a di-photon trigger, with transverse energy
($E_{\mathrm{T}}$) thresholds at 25 and 35~GeV. Two photons with an invariant mass $m_{\gamma\gamma}$ between 105 and
160~GeV are selected, with $E_{\mathrm{T}}/m_{\gamma\gamma}$ above 0.35 (0.25) for the leading (sub-leading) photon.
The event selection proceeds with requesting at least two central jets with $p_{\mathrm{T}} > 25~\mbox{GeV}$, vetoing
events with more than two $b$-tags. The signal region consists of events with exactly two $b$-tags (with the same working
point as in the search for $HH \to bbbb$,) and events with one $b$-tag, however with a tighter working point corresponding
to a 60\% efficiency. In that latter case, a Boosted Decision Tree (BDT) is used to assign a second jet to the $H \to bb$ candidate.
In addition to the requirements on the photons and $b$-jets, two sets of event selections are applied:
\begin{itemize}
\item loose selection, used in the search for non-resonant production with a Higgs self-coupling different from the SM prediction and
in the search for a spin-0 resonance of mass 260--500~GeV decaying via $HH \to bb\gamma\gamma$: the $p_{\mathrm{T}}$ of the
leading (sub-leading) jet is greater than 40~(25)~GeV, the invariant mass of the two jets is $80~\mbox{GeV} < m_{jj} < 140~\mbox{GeV}$
and, in the case of the search for resonant $HH \to bb\gamma\gamma$ production, $m_{\gamma\gamma}$ must be within 4.7~GeV of
$m_H$;
\item tight selection, used in the search for SM non-resonant production and in the search for a spin-0 resonance heavier than
500~GeV decaying via $HH \to bb\gamma\gamma$: the $p_{\mathrm{T}}$ of the leading (sub-leading) jet is greater than
100~(30)~GeV, the invariant mass of the two jets is $90~\mbox{GeV} < m_{jj} < 140~\mbox{GeV}$ and, in the case of the search
for resonant $HH \to bb\gamma\gamma$ production, $m_{\gamma\gamma}$ must be within 4.3~GeV of $m_H$.
\end{itemize}
\subsection{Search for non-resonant $HH \to bb\gamma\gamma$ production}
The analysis strategy used in the search for non-resonant $HH \to bb\gamma\gamma$ production is to extract the signal from
the $m_{\gamma\gamma}$ distribution. Both the $HH$ signal and the single-$H$ backgrounds are taken from simulation, and
$m_{\gamma\gamma}$ is parameterised with a double-sided Crystal-Ball function. On the other hand, the continuum background
of multi-jet and multi-photon events is modelled by a fit to the data, with a first-order exponential function. This fit function was
found to minimise the spurious signal, i.e.\ the bias measured by fitting a signal+background model to a background-only sample.
Figure~\ref{fig:bbyy-first} shows the $m_{\gamma\gamma}$ distribution in the signal region with one and two $b$-tags, obtained
with the tight selection. In the absence of a statistically significant excess with respect to the SM prediction, 95\% CL exclusion
limits are set of the Higgs boson pair production cross-section, in units of the SM prediction, see Table~\ref{tab:bbyy}.
\begin{figure}[htbp]
\centering
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure06a}
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure06b}
\caption{\label{fig:bbyy-first} Di-photon mass spectrum in signal regions of the search for $HH \to bb\gamma\gamma$ in ATLAS,
with one (left) or two (right) $b$-tags, after applying a tight event selection~\cite{Aaboud:2018ftw}.}
\end{figure}
\begin{table}[htbp]
\begin{center}
\caption{95\% CL limits on non-resonant Higgs boson pair production, from the search for $HH \to bb\gamma\gamma$ in ATLAS.}
\vspace*{2mm}
\begin{tabular}{cccc}
\hline
Observed & $-1\sigma$ & Expected & $+1\sigma$ \\
\hline
22 & 20 & 28 & 40 \\
\hline
\end{tabular}
\label{tab:bbyy}
\end{center}
\end{table}
Higgs self-coupling values that differ from the SM prediction affect the signal acceptance. The variations of the exclusion
limits with $\kappa_{\lambda}$, defined as the ratio of the Higgs self-coupling $\lambda$ to its predicted value in the SM,
$\lambda_{\mathrm{SM}}$, are computed with the loose event selection and are shown in Figure~\ref{fig:bbyy-second}.
As a result, $\kappa_{\lambda}$ is constrained at the 95\% CL to be between $-8.2$ and 13.2.
\begin{figure}[htbp]
\centering
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure07}
\caption{\label{fig:bbyy-second} 95\% CL limits on non-resonant Higgs boson pair production as a function of $\kappa_{\lambda}$,
from the search for $HH \to bb\gamma\gamma$ in ATLAS. The red line indicates the theoretical prediction with its
uncertainty~\cite{Aaboud:2018ftw}.}
\end{figure}
\subsection{Search for resonant $HH \to bb\gamma\gamma$ production}
The analysis strategy used in the search for resonant $HH \to bb\gamma\gamma$ production is to extract the signal from
the $m_{\gamma\gamma jj}$ distribution, after scaling the di-jet four-momentum by $m_{H}/m_{jj}$. The resonant $HH$ signal
is modelled with a Gaussian distribution with exponential tails. The single-$H$ and SM non-resonant $HH$ backgrounds are taken
from simulation. On the other hand, the continuum background of multi-jet and multi-photon events is modelled by a fit to the data,
with a fit function chosen to minimise the spurious signal. Figure~\ref{fig:bbyy-third} shows the $m_{\gamma\gamma jj}$ distribution
in the signal region with one and two $b$-tags, obtained with the loose selection. In the absence of a statistically significant excess
with respect to the SM prediction, 95\% CL exclusion limits are set of the resonant Higgs boson pair production cross-section, see
Figure~\ref{fig:bbyy-fourth}.
\begin{figure}[htbp]
\centering
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure08a}
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure08b}
\caption{\label{fig:bbyy-third} Distribution of $m_{\gamma\gamma jj}$ in signal regions of the search for
$HH \to bb\gamma\gamma$ in ATLAS, with one (left) or two (right) $b$-tags, after applying a loose event selection~\cite{Aaboud:2018ftw}.}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[height=0.25\textheight]{CIPANP2018-Ferrari-figure09}
\caption{\label{fig:bbyy-fourth} 95\% CL limits on Higgs boson pair production from a spin-0 resonance, as a function of the hypothetised
mass, from the search for $HH \to bb\gamma\gamma$ in ATLAS. The vertical blue dashed line indicates a change in event
selections~\cite{Aaboud:2018ftw}.}
\end{figure}
\section{Search for $HH \to WW\gamma\gamma$ in ATLAS}
The good mass resolution of the di-photon system arising from a $H \to \gamma\gamma$ decay has also been exploited to
search for $HH \to WW\gamma\gamma$ in ATLAS, where one $W$-boson decays hadronically and the other one into a charged
lepton (electron or muon) and a neutrino. The trigger, as well as the selection and reconstruction of the two photons coming from
the $H \to \gamma\gamma$ decay are similar to those employed in the $HH \to bb\gamma\gamma$ search, except that
$m_{\gamma\gamma}$ is required to be within 3.4~GeV of $m_H$ when searching for both non-resonant and resonant
Higgs boson pair productions. In addition, a requirement $p_{\mathrm{T}}^{\gamma\gamma} > 100~\mbox{GeV}$ is
used in the search for SM non-resonant production and in the search for a spin-0 resonance heavier than 400~GeV.
In order to select $H \to WW$ decays, events are also requested to have at least one electron or muon with
$p_{\mathrm{T}} > 10~\mbox{GeV}$, and at least two central jets which are not $b$-tagged. Both the $HH$ signal
and the single-$H$ backgrounds are taken from simulation, and their $m_{\gamma\gamma}$ spectrum is parameterised
with a double-sided Crystal-Ball function. On the other hand, the continuum background of multi-jet and multi-photon events
is modelled by a fit to the data, with a second-order exponential function, chosen to minimise the spurious signal. \\
Figure~\ref{fig:WWyy-first} shows the predicted and measured $m_{\gamma\gamma}$ distribution when using or not the cut
$p_{\mathrm{T}}^{\gamma\gamma} > 100~\mbox{GeV}$. In the absence of a statistically significant excess with respect to the
SM prediction, 95\% CL exclusion limits are set of the non-resonant Higgs boson pair production cross-section, in units of
the SM prediction (see Table~\ref{tab:WWyy}), and on the cross-section for a spin-0 resonance decaying to $HH$ (see
Figure~\ref{fig:WWyy-second}).
\begin{figure}[htbp]
\centering
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure10a}
\includegraphics[height=0.21\textheight]{CIPANP2018-Ferrari-figure10b}
\caption{\label{fig:WWyy-first} Di-photon mass spectrum in signal regions of the search for $HH \to WW\gamma\gamma$ in ATLAS,
without (left) or with (right) a cut on $p_{\mathrm{T}}^{\gamma\gamma}$~\cite{Aaboud:2018ewm}.}
\end{figure}
\begin{table}[htbp]
\begin{center}
\caption{95\% CL limits on non-resonant Higgs boson pair production, from the search for $HH \to WW\gamma\gamma$ in ATLAS.}
\vspace*{2mm}
\begin{tabular}{cccccc}
\hline
Observed & $-2\sigma$ & $-1\sigma$ & Expected & $+1\sigma$ & $+2\sigma$ \\
\hline
230 & 90 & 120 & 160 & 240 & 340 \\
\hline
\end{tabular}
\label{tab:WWyy}
\end{center}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[height=0.24\textheight]{CIPANP2018-Ferrari-figure11}
\caption{\label{fig:WWyy-second} 95\% CL limits on Higgs boson pair production from a spin-0 resonance, as a function of the
hypothetised mass, from the search for $HH \to WW\gamma\gamma$ in ATLAS. The vertical dashed line indicates a change
in event selections~\cite{Aaboud:2018ewm}.}
\end{figure}
\section{Conclusion}
These proceedings report on three searches for Higgs boson pair production in up to 36.1~fb$^{-1}$ of LHC $pp$ collision
data, which have been recently published by the ATLAS Collaboration: $HH \to bbbb$, $HH \to bb\gamma\gamma$ and
$HH \to WW\gamma\gamma$. The data were analysed to search for non-resonant production of $HH$ pairs as well as
for heavy resonances decaying to two SM-like Higgs bosons, with various analysis strategies and event selections. The result
of a fourth search, based on the $HH \to bb\tau\tau$ final state, was published after the conference~\cite{Aaboud:2018sfw},
reaching 95\% CL exclusion limits of 12.7 times the SM expectation for the non-resonant $HH$ production mode. A statistical
combination of the three most sensitive searches was also performed by ATLAS recently~\cite{ATLAS-comb}. The combined
observed limit on the non-resonant Higgs boson pair cross-section is 0.223~pb at 95\% CL, which is equivalent to 6.7 times the
predicted SM cross-section. The ratio $\kappa_{\lambda}$ of the Higgs boson self-coupling to its SM expectation is constrained
at 95\% CL to $-5.0 < \kappa_{\lambda} < 12.1$. The search for Higgs boson pair production will remain at the core of the ATLAS
research program towards the end of the LHC Run-2 and beyond, as it allows to probe directly the Higgs potential as well as
search for new physics in the Higgs sector.\\
Copyright [2018] CERN for the benefit of the ATLAS Collaboration. Reproduction of this article or parts of it is
allowed as specified in the CC-BY-4.0 license.
|
\section{Introduction}
\label{sec:intro}
The topic or population polytope is a fundamental geometric object that underlies the presence of latent topic variables in topic and admixture models ~\citep{blei2003latent,pritchard2000inference,tang2014understanding}.
The geometry of topic models provides the theoretical basis for posterior contraction analysis of latent topics, in addition to helping to develop fast and quite accurate inference algorithms in parametric and nonparametric settings~\citep{nguyen2015posterior,yurochkin2016geometric,yurochkin2017conic,yurochkin2019dirichlet}.
When data and the associated topics are indexed by time dimension, it is of interest to study the temporal dynamics of such latent geometric structures. In this paper, we will study the modeling and algorithms for learning temporal dynamics of topic polytope that arises in the analysis of text corpora.
Several authors have extended the basic topic modeling framework to analyze how topics evolve over time. The Dynamic Topic Models (DTM) \citep{blei2006dynamic} demonstrated the importance of accounting for non-exchangeability between document groups, particularly when time index is provided. Another approach is to keep topics
fixed and consider only evolving topic popularity \citep{wang2006topics}. \citet{hong2011time} extended such an approach to multiple corpora.
\citet{ahmed2012timeline} proposed a nonparametric construction extending DTM where topics can appear or eventually die out. Although the evolution of the latent geometric structure (i.e., the topic polytope) is implicitly present in these works, it was not explicitly addressed nor is the geometry exploited. A related limitation shared by these modeling frameworks is the lack of scalability, due to inefficient joint modeling and learning of topics at each time point and topic evolution over time. To improve scalability, a natural solution is decoupling the two phases of inference.
To this end, we seek to develop a series of topic \emph{meta}-models, i.e. models for temporal dynamics of topic polytopes, assuming that the topic estimates from each time point have already been obtained via some efficient static topic inference technique.
The focus on inference of topic evolution offers novel opportunities and challenges. To start, what is the suitable ambient space in which the topic polytope is represented? As topics evolve, so are the number of topics that may become active or dormant, raising distinct modeling choices. Interesting issues arise in the inference, too. For instance, what is the principled way of \emph{matching} vertices of a collection of polytopes to their next reincarnations? Such question arises because we consider modeling of topics learned independently across timestamps and text corpora, which entails the need for preserving the topic structure's permutation invariance of the vertex labels.
We consider an isometric embedding of the unit sphere in the word simplex, so that the evolution of topic polytopes may be represented by a collection of (random) trajectories of points residing on the unit sphere. Instead of attempting to mix-match vertices in an ad hoc fashion, we appeal to a Bayesian nonparametric modeling framework that allows the number of topic vertices to be random and vary across time. The mix-matching between topics shall be guided by the assumption on the smoothness of the collection of global trajectories on the sphere using von Mises-Fisher dynamics \citep{mardia2009directional}. The selection of active topics at each time point will be enabled by a nonparametric prior on the random binary matrices via the (hierarchical) Beta-Bernoulli process~\citep{thibaux2007hierarchical}.
Our contribution includes a sequence of Bayesian nonparametric models in increasing levels of complexity: the simpler model describes a topic polytope evolving over time, while the full model describes the temporal dynamics of a collection of topic polytopes as they arise from multiple corpora. The semantics of topics can be summarized as follows: there is a collection of latent global topics of unknown cardinality evolving over time (e.g. topics in science or social topics in Twitter). Each year (or day) a subset of the global topics is elucidated by the community (some topics may be dormant at a given time point). The nature of each global topic may change smoothly (via varying word frequencies). Additionally, different subsets of global topics are associated with different groups (e.g. journals or Twitter location stamps), some becoming active/inactive over time.
Another key contribution includes a suite of scalable approximate inference algorithms suitable for online and distributed settings. In particular, we focus mainly on MAP updates rather than a full Bayesian integration. This is appropriate in an online learning setting, moreover such updates of the latent topic polytope can be viewed as solving an optimal matching problem for which a fast Hungarian matching algorithm can be applied. Our approach is able to perform dynamic nonparametric topic inference on 3 million documents in 20 minutes, which is significantly faster than prior static online and/or distributed topic modeling algorithms \citep{newman2008distributed,hoffman2010online,wang2011online,bryant2012truly,broderick2013streaming}.
The remainder of the paper is organized as follows. In Section \ref{sec:sphere_map} we define a Markov process over the space of topic polytopes (simplices). In Section \ref{sec:models} we present a series of models for polytope dynamics and describe our algorithms for online dynamic and/or distributed inference. Section \ref{sec:experiments} demonstrates experimental results. We conclude with a discussion in Section \ref{sec:discuss}.
\section{Temporal dynamics of a topic polytope}
\label{sec:sphere_map}
The fundamental object of inference in this work is the topic polytope arising in topic modeling which we shall now define~\citep{blei2003latent,nguyen2015posterior}. Given a vocabulary of $V$ words, a topic is defined as a probability distribution on the vocabulary. Thus a topic is taken to be a point in the vocabulary simplex, namely, $\Delta^{V-1}$, and a topic polytope for a corpus of documents is defined as a convex hull of topics associated with the documents. Geometrically, the topics correspond to the vertices (extreme points) of the (latent) topic polytope to be inferred from data.
In order to infer about the temporal dynamics of a topic polytope, one might consider the evolution of each topic variable, say $\theta^{(t)}$, which represents a vertex of the polytope at time $t$. A standard approach is due to~\citet{blei2006dynamic}, who proposed to use a Gaussian Markov chain $\theta^{(t)}|\theta^{(t-1)} \thicksim \mathcal{N}(\theta^{(t-1)},\sigma I)$ in $\mathbb{R}^V$ for modeling temporal dynamics and a logistic normal transformation $\pi(\theta^{(t)})_i := \frac{\exp(\theta^{(t)}_{i})}{\sum_i \exp(\theta^{(t)}_{i})}$, which sends elements
of $\mathbb{R}^V$ into $\Delta^{V-1}$. In our meta-modeling approach, we consider topics, i.e. points in $\Delta^{V-1}$, learned independently across time and corpora. Logistic normal map is many-to-one, hence it is undesirably ambiguous in mapping a collection of topic polytopes to $\mathbb{R}^V$.
We propose to represent each topic variable as a point in a unit sphere $\mathbb{S}^{V-2}$, which possesses a natural isometric embedding (i.e. one-to-one) in the vocabulary simplex $\Delta^{V-1}$, so that the temporal dynamics of a topic variable can be identified as a (random) trajectory on $\mathbb{S}^{V-2}$. This trajectory shall be modeled as a Markovian process on $\mathbb{S}^{V-2}$: $\theta^{(t)}|\theta^{(t-1)} \thicksim \vmf(\theta^{(t-1)}, \tau_0)$. Von Mises-Fisher (vMF) distribution is commonly used in the field of directional statistics \citep{mardia2009directional} to model points on a unit sphere and was previously utilized for text modeling \citep{banerjee2005clustering, reisinger2010spherical}.
\paragraph{Isometric embedding of $\mathbb{S}^{V-2}$ into the vocabulary simplex} We start with the directional representation of topic polytope~\citep{yurochkin2017conic}: let $B=\{\beta_1,\ldots,\beta_K\}$ be a collection of vertices of a topic polytope. Each vertex is represented as $\beta_k := C + R_k\tilde{\beta}_k$, where $C \in \conv(B)$ is a reference point in a convex hull of $B$, $\tilde{\beta}_k \in \mathbb{R}^V$ is a topic direction and $R_k \in \mathbb{R}_{+}$. Moreover, $R_k \in [0,1]$ is determined so that the tip of direction vector $\tilde{\beta}_k$ resides on the boundary of $\Delta^{V-1}$. Since the effective dimensionality of $\tilde{\beta}_k$ is $V-2$, we can now define an one-to-one and isometric map sending $\tilde{\beta}_k$ onto $\mathbb{S}^{V-2}$ as follows: map of the vocabulary simplex $\Delta^{V-1} \in \mathbb{R}^V$ where it is first translated so that $C$ becomes the origin and then rotated into $\mathbb{R}^{V-1}$, where resulting topics, say $\theta_1,\ldots,\theta_K \in \mathbb{S}^{V-2}$, are normalized to the unit length. Observe that this geometric map is an isometry and hence invertible. It preserves angles between vectors, therefore we can evaluate vMF density without performing the map explicitly, by simply setting $\theta_k := \frac{\beta_k - C}{\|\beta_k - C\|}$. The following lemma formalizes this idea.
\begin{lem}
\label{lem:simplex_to_sphere}
$\Gamma: \{\beta=(\beta_1,\ldots,\beta_V)\in \Delta^{V-1}: \beta_i=0 \text{ for some } i\} \rightarrow \{\theta \in \mathbb{S}^{V-1} : \mathbbm{1}_V^{T}\theta=0 \}$ is a homeomorphism, where $\Gamma(\beta)=\left(\beta-C\right)/\|\beta-C\|_2$, and $\Gamma^{-1}(\theta)=-\frac{\theta}{\max_i \theta_i/c_i}+C$, for any $C=(c_1,\ldots,c_V) \in \Delta^{V-1}$.
\end{lem}
Proofs of this Lemma is given in Supplement \ref{supp:lemma1}. The intuition behind the construction is provided via Figure \ref{fig:simplex_to_sphere} which gives a geometric illustration for $V=3$, vocabulary simplex $\Delta^{V-1}$ shown as red triangle. Two topics on the boundary (face) of the vocabulary simplex are $\beta_1 = C + \tilde{\beta}_1$ and $\beta_2 = C + \tilde{\beta}_2$. Green dot $C$ is the reference point and $\alpha = \angle (\tilde{\beta}_1, \tilde{\beta}_2)$. In Fig. \ref{fig:simplex_to_sphere} (left) we move $C$ by translation to the origin and rotate $\Delta^{V-1}$ from $xyz$ to $xy$ plane. In Fig. \ref{fig:simplex_to_sphere} (center left) we show the resulting image of $\Delta^{V-1}$ and add a unit sphere (blue) in $\mathbb{R}^2$. Corresponding to $\beta_1, \beta_2$ topics are the points $\theta_1, \theta_2$ on the sphere with $\angle(\theta_1,\theta_2) = \alpha$. Now, apply the inverse translation and rotation to \emph{both} $\Delta^{V-1}$ and
$\mathbb{S}^{V-2}$, the result is shown in Fig. \ref{fig:simplex_to_sphere} (center right) --- we are back to $\mathbb{R}^3$ and $\angle(\theta_1,\theta_2)=\angle(\tilde{\beta}_1,\tilde{\beta}_2)=\alpha$, where $\theta_k = \frac{\beta_k - C}{\|\beta_k - C\|_2}$. In Fig. \ref{fig:simplex_to_sphere} (right) we give a geometric illustration of the temporal dynamics.
As described above, each topic evolves in a random trajectory residing in a unit sphere, so the evolution of a collection of topics can be modeled by a collection of corresponding trajectories on the sphere. Note that the number of "active" topics may be unknown and vary over time. Moreover, a topic may be activated, become dormant, and then resurface after some time. New modeling elements are introduced in the next section to account for these phenomena.
\begin{figure*}[t]
\begin{center}
\centerline{\includegraphics[width=\textwidth]{sphere_simplex_small.pdf}}
\caption{Invertible transformation between unit sphere and a standard simplex; dynamics example}
\vskip -0.4in
\label{fig:simplex_to_sphere}
\end{center}
\end{figure*}
\section{Hierarchical Bayesian modeling for single or multiple topic polytopes}
\label{sec:models}
We shall present a sequence of models with increasing levels of complexity: we start by introducing a hierarchical model for online learning of the temporal dynamics of a single topic polytope, allowing for varying number of vertices over time. Next, a static model for \emph{multiple} topic polytopes learned on different corpora drawing on a common pool of global topics. Finally, we present a "full" model for modeling evolution of global topic trajectories over time and across groups of corpora.
\subsection{Dynamic model for single topic polytope}
\label{subsec:dynamic}
At a high level, our model maintains a collection of global trajectories taking values on a unit sphere. Each trajectory shall be endowed with a von Mises-Fisher dynamic described in the previous section. At each time point, a random topic polytope is constructed by selecting a (random) subset of points on the trajectory evaluated at time $t$. The random selection is guided by a Beta-Bernoulli process prior \citep{thibaux2007hierarchical}. This construction is motivated by a modeling technique of \citet{nguyen2010inference}, who studied a Bayesian hierarchical model for inference of smooth trajectories on an Euclidean domain using Dirichlet process priors. Our generative model, using Beta-Bernoulli process as a building block, is more appropriate for the purpose of topic discovery. Due to the isometric embedding of $\mathbb{S}^{V-2}$ in $\Delta^{V-1}$ described in the previous section, from here on we shall refer to topics as points on $\mathbb{S}^{V-2}$.
First, generate a collection of global topic trajectories using Beta Process prior (cf. \citet{thibaux2007hierarchical})
\footnote{\citet{thibaux2007hierarchical} write BP$(c, H)$, $H(\Omega)=\gamma_0$; we set $c=1$, $H=H/\gamma_0$ and write BP$(\gamma_0, H)$.}
with a base measure $H$ on the space of trajectories on $\mathbb{S}^{V-2}$ and mass parameter $\gamma_0$:
\begin{equation}
\label{eq:top_bp}
Q|\gamma_0, H \thicksim \text{BP}(\gamma_0,H).
\end{equation}
It follows that $Q = \sum_i q_i \delta_{\theta_i}$,
where $\{q_i\}_{i=1}^\infty$ follows a stick-breaking construction \citep{teh2007stick}: $\mu_i \thicksim \text{Beta}(\gamma_0,1),\, q_i = \prod_{j=1}^i \mu_j$, and each $\theta_i \thicksim H$ is a sequence of $T$ random elements on the unit sphere $\theta_i := \{\theta_i^{(t)}\}_{t=1}^T$, which are generated as follows:
\begin{equation}
\label{eq:vmf_time}
\begin{split}
& \theta_i^{(t)}| \theta_i^{(t-1)} \thicksim \vmf(\theta_i^{(t-1)}, \tau_0)\text{ for }t=1,\ldots, T,\\
& \theta_i^{(0)} \thicksim \vmf(\cdot,0) \; \text{-- uniform on} \; \mathbb{S}^{V-2}.
\end{split}
\end{equation}
At any given time $t=1,\ldots,T$, the process $Q$ induces a marginal measure $Q_t$, whose support is given by the atoms of $Q$ as they are evaluated at time $t$. Now, select a subset of the global topics that are active at $t$ via the Bernoulli process
$\ensuremath{\mathcal{T}}^{(t)}|Q_t \thicksim \text{BeP}(Q_t).$ Then
$\ensuremath{\mathcal{T}}^{(t)} := {\textstyle\sum}_{i=1} b_{i}^{(t)} \delta_{\theta^{(t)}_i}\text{, where } b_{i}^{(t)}|q_{i} \thicksim \text{Bern}(q_{i}),\,\forall i$.
$\ensuremath{\mathcal{T}}^{(t)}$ are supported by atoms $\{\theta^{(t)}_i : b_{i}^{(t)} = 1, i=1,2,\ldots \}$ representing topics active at time $t$. Finally, assume that noisy measurements of each of these topic variables are generated via:
\begin{equation}
\label{eq:t_vmf_kalman}
\begin{split}
& v^{(t)}_{k}|\ensuremath{\mathcal{T}}^{(t)} \thicksim \vmf(\ensuremath{\mathcal{T}}_{k}^{(t)}, \tau_1),\, k = 1,\ldots,K^{(t)}\text{, where}\\
& \,K^{(t)} := \card(\ensuremath{\mathcal{T}}^{(t)}); \ensuremath{\mathcal{T}}_k^{(t)}\text{ is $k$-th atom of $\ensuremath{\mathcal{T}}^{(t)}$}.
\end{split}
\end{equation}
Noisy estimates for the topics at any particular time point may come from either the global topics observed until the previous time point or a topic yet unexplored. We emphasize that topics $\{v^{(t)}_{k}\}_{k=1}^{K^{(t)}}$ for $t=1,\ldots,T$ are the quantities we aim to model, hence we refer to our approach as the \emph{meta}-model. These topics may be learned, for each time point independently, by any stationary topic modeling algorithms, and then transformed to sphere by applying Lemma \ref{lem:simplex_to_sphere}.
Let $B^{(t)}$ denote the binary matrix representing the assignment of observed topic estimates to global topics at time point $t$, i.e, $B_{ik}^{(t)}=1$ if the vector $v_k^{(t)}$ is a noisy estimate for $\theta_i^{(t)}$. In words, these random variables ``link up'' the noisy estimates at any time point to the global topics observed thus far.
By conditional independence, the joint posterior of the hidden $\theta^{(t)}$ given observed noisy $v^{(t)}$ is:
\begin{equation*}
\label{eq:posterior_dynamic_BBP}
\mathbb P\left( \theta^{(0)},\{\theta^{(t)},B^{(t)}\}_{t=1}^T | \{v^{(t)}\}_{t=1}^T \right) \propto
\mathbb P(\theta^{(0)})\prod_{t=1}^T \mathbb P( \theta^{(t)}, B^{(t)} | \theta^{(t-1)},\{ B^{(a)}\}_{a=1}^{t-1})\mathbb P(v^{(t)}| \theta^{(t)}, B^{(t)}).
\end{equation*}
At $t$, $\mathbb P( \theta^{(t)}, B^{(t)} | \theta^{(t-1)}, \{ B^{(a)}\}_{a=1}^{t-1}) \mathbb P(v^{(t)}| \theta^{(t)}, B^{(t)}) \propto \mathbb P(\theta^{(t)},B^{(t)}| \theta^{(t-1)},v^{(t)},\{ B^{(a)}\}_{a=1}^{t-1}) \propto$
\begin{equation}
\label{eq:time_bp_1}
\begin{split}
& \prod_{i=1}^{L_{t-1}} \left( \left(m_i^{(t-1)}/(t-m_i^{(t-1)})\right)^{\sum_{k=1}^ {K^{(t)}} B_{ik}^{(t)}} \exp(\tau_0 \langle \theta_i^{(t-1)}, \theta_i^{(t)}\rangle)\right) \\
& \cdot \frac{\exp(-\frac{\gamma_0}{t}) (\gamma_0/t)^{L_t-L_{t-1}}}{(L_t-L_{t-1})!} \exp(\tau_1 {\textstyle\sum}_{i=1}^{L_t} {\textstyle\sum}_{k=1}^{K^{(t)}} B_{ik}^{(t)}\langle \theta_i^{(t)},v_k^{(t)}\rangle).
\end{split}
\end{equation}
The equation above represents a product of four quantities: (1) probability of $B^{(t)}$s, where
$m_i^{(t)}$ denotes the number of occurrences of topic $i$ up to time point $t$ (cf. popularity of a dish in the Indian Buffet Process (IBP) metaphor \citep{ghahramani2005infinite}),
(2) vMF conditional of $\theta_i^{(t)}$ given $\theta_i^{(t-1)}$ (cf. Eq.~\eqref{eq:vmf_time}), (3) number of new global topics at time $t$, $L_t - L_{t-1} \sim \textrm{Pois}(\gamma_0/t)$, and
(4) emission probability $\mathbb P(v^{(t)} | \theta^{(t)}, B^{(t)})$ (cf. Eq.~\eqref{eq:t_vmf_kalman}). Derivation details are given in Supplement \ref{supp:sdm_posterior}.
\paragraph{Streaming Dynamic Matching (SDM)}
To perform MAP estimation in the streaming setting, we
highlight the connection of the maximization of the posterior \eqref{eq:time_bp_1} to the objective of an optimal \emph{matching} problem: given a cost matrix, workers should be assigned to tasks, at most one worker per task and one task per worker. The solution of this problem is obtained by employing the well-known Hungarian algorithm \citep{kuhn1955hungarian}. In the context of dynamic topic modeling, our goal is to match topics learned on the new timestamp to the trajectories of topics learned over the previous timestamps, where the cost is governed by our model. This connection is formalized by the following.
\begin{prop}
\label{prop:sdm}
Given the cost
$
C_{ik}^{(t)}=
\begin{cases} \| \tau_1 v_k^{(t)} +\tau_0 \theta_i^{(t-1)} \|_2 - \tau_0 + \log\frac{m_i^{(t-1)}}{t-m_i^{(t-1)}}, i\leq L_{t-1}\\
\tau_1 + \log\frac{\gamma_0}{t} - \log(i-L_{t-1}), L_{t-1} < i \leq L_{t-1} + K^{(t)}
\end{cases}
$
consider the optimization problem $\max_{B^{(t)}} \sum_{i,k}B_{ik}^{(t)}C_{ik}^{(t)}$ subject to the constraints that
(a) for each fixed $i$, at most one of $B_{ik}^{(t)}$ is $1$ and the rest are $0$, and
(b) for each fixed $k$, exactly one of $B_{ik}^{(t)}$ is $1$ and the rest are $0$.
Then, the MAP estimate for Eq. \eqref{eq:time_bp_1} can be obtained by the Hungarian algorithm, which solves for $((B_{ik}^{(t)}))$ to obtain $\theta_i^{(t)}$ as
\begin{equation}
\label{eq:sdm_theta}
\begin{cases} \frac{\tau_1 v_{k}^{(t)} +\tau_0 \theta_i^{(t-1)}}{\|\tau_1 v_{k}^{(t)} +\tau_0 \theta_i^{(t-1)}\|_2}, & \text{if }\exists \ k\ \text{s.t. }B_{ik}^{(t)}=1 \text{ and } i\leq L_{t-1}\\
v_k^{(t)}, & \text{if }\exists \ k\ \text{s.t. }B_{ik}^{(t)}=1 \text{ and } i>L_{t-1} \text{ (new topic)}\\
\theta_i^{(t-1)} & \text{otherwise (topic is dormant at $t$).}
\end{cases}
\end{equation}
\end{prop}
We defer proof to Supplement \ref{supp:sdm}. To complete description of the inference we shall discuss how noisy estimates are obtained from the bag-of-words representation of the documents observed at time point $t$. We choose to use CoSAC \citep{yurochkin2017conic} algorithm to obtain topics $\{\beta_k^{(t)} \in \Delta^{V-1}\}_{k=1}^{K^{(t)}}$ from $\{x_m^{(t)} \in \mathbb{N}^V\}_{m=1}^{M_t}$, collection of $M_t$ documents at time point $t$. CoSAC is a stationary topic modeling algorithm which can infer number of topics from the data and is computationally efficient for moderately sized corpora. We note that other topic modeling algorithms, e.g., variational inference \citep{blei2003latent} or Gibbs sampling \citep{griffiths2004finding, teh2006hierarchical}, can be used in place of CoSAC. Estimated topics are then transformed to $\{v^{(t)}_{k} \in \mathbb{S}^{V-2}\}_{k=1}^{K^{(t)}}$ using Lemma \ref{lem:simplex_to_sphere} and reference point $C_t = \sum_{a=1}^t\sum_{m=1}^{M_a} \frac{x_m^{(a)}}{N_m^{(a)}}/\sum_{a=1}^t M_a$,
where $N_m^{(a)}$ is the number of words in the corresponding document. Our reference point is simply an average (computed dynamically) of the normalized documents observed thus far.
Finally we update MAP estimates of global topics dynamics based on Proposition \ref{prop:sdm}. Streaming Dynamic Matching (SDM) is summarized in Algorithm \ref{algo:SDM}.
\paragraph{Additional related literature} utilizing similar technical building blocks in different contexts. \citet{fox2009sharing} utilized Beta-Bernoulli process in time series modeling to capture switching regimes of an autoregressive process, where the corresponding Indian Buffet Process was used to select subsets of the latent states of the Hidden Markov Model. \citet{williamson2010ibp} used Indian Buffet Process in topic models to sparsify document topic proportions. \citet{campbell2015streaming} utilized Hungarian algorithm for streaming mean-field variational inference of the Dirichlet Process mixture model.
\begin{algorithm}[ht]
\caption{Streaming Dynamic Matching (SDM)}
\label{algo:SDM}
\begin{algorithmic}[1]
\FOR{$t=1,\ldots,T$}
\STATE Observe documents $\{x_m^{(t)}\}_{m=1}^{M_t}$
\STATE Estimate topics $\{\beta_k^{(t)}\}_{k=1}^{K^{(t)}}$ = CoSAC($\{x_m^{(t)}\}_{m=1}^{M_t}$)
\STATE Map topics to sphere $\{v^{(t)}_{k}\}_{k=1}^{K^{(t)}}$ (Lemma \ref{lem:simplex_to_sphere})
\STATE Given $\{\theta^{(t-1)}\}_{i=1}^{L_{t-1}}$ and $\{v^{(t)}_{k}\}_{k=1}^{K^{(t)}}$ compute cost matrix as in Proposition \ref{prop:sdm}
\STATE Using Hungarian algorithm solve the corresponding matching problem to obtain $B^{(t)}$
\STATE Compute $\{\theta^{(t)}\}_{i=1}^{L_{t}}$ as in eq. \eqref{eq:sdm_theta}
\ENDFOR
\end{algorithmic}
\end{algorithm}
\subsection{Beta-Bernoulli Process for multiple topic polytopes}
\label{subsec:distributed}
We now consider meta-modeling in the presence of multiple corpora, each of which maintains its own topic polytope. Large text corpora often can be partitioned based on some grouping criteria, e.g. scientific papers by journals, news by different media agencies or tweets by location stamps. In this subsection we model the collection of topic polytopes observed at a single time point by employing the Beta-Bernoulli Process prior \citep{thibaux2007hierarchical}. The modeling of a collection of polytopes evolving over time will be described in the following subsection.
First, generate global topic measure $Q$ as in Eq. \eqref{eq:top_bp}. Here, we are interested only in a single time point, the base measure $H$ is simply a $\vmf(\cdot,0)$, the uniform distribution over $\mathbb{S}^{V-2}$. Next, for each group $j=1,\ldots,J$, select a subset of the global topics:
\begin{equation}
\label{tj_bern}
\ensuremath{\mathcal{T}}_j|Q \thicksim \text{BeP}(Q)\text{, then }
\ensuremath{\mathcal{T}}_j := {\textstyle\sum}_{i=1} b_{ji} \delta_{\theta_i}\text{, where }b_{ji}|q_{i} \thicksim \text{Bern}(q_{i}),\,\forall i.
\end{equation}
Notice that each group $\ensuremath{\mathcal{T}}_j := \{\theta_i : b_{ji} = 1, i=1,2,\ldots \}$ selects only a subset from the collection of global topics, which is consistent with the idea of partitioning by journals: some topics of ICML are not represented in SIGGRAPH and vice versa. The next step is analogous to Eq. \eqref{eq:t_vmf_kalman}:
\begin{equation}
\label{eq:j_vmf_kalman}
v_{jk}|\ensuremath{\mathcal{T}}_j \thicksim \vmf(\ensuremath{\mathcal{T}}_{jk}, \tau_1)\text{ for }k = 1,\ldots,K_j\text{, where }K_j := \card(\ensuremath{\mathcal{T}}_j).
\end{equation}
We again use $B$ to denote the binary matrix representing the assignment of global topics to the noisy topic estimates, i.e., $B_{jik}=1$ if the $k^{th}$ topic estimate for group $j$ arises as a noisy estimate of global topic $\theta_i$.
However, the \emph{matching} problem is now different from before: we don't have any information about the global topics as there is no history, instead we should match a \emph{collection} of topic polytopes to a global topic polytope.
The matrix of topic assignments is distributed a priori by an Indian Buffet Process (IBP) with parameter $\gamma_0$. The conditional probability for global topics $\theta_i$ and assignment matrix $B$ given topic estimates $v_{jk}$ has the following form:
\begin{equation}
\label{eq:post_hbp_1}
\mathbb P(B,\theta|v) \propto \exp(\tau_1 {\textstyle\sum}_{j,i,k} B_{jik} \langle\theta_i,v_{jk}\rangle)\text{IBP}(\{m_i\})\text{, where }m_i={\textstyle\sum}_{j,k} B_{jik}
\end{equation}
and IBP is the prior (see Eq. (15) in \citep{griffiths2011indian}) with $m_i$ denoting the popularity of global topic $i$.
\paragraph{Distributed Matching (DM)}
Similar to Section \ref{subsec:dynamic}, we look for point estimates for the topic directions $\theta$ and for the topic assignment matrix $B$. Direct computation of the global MAP estimate for Eq. \eqref{eq:post_hbp_1} is not straight-forward. The problem of matching across groups and topics is not amenable to a closed form Hungarian algorithm. However we show that for a fixed group the assignment optimization reduces to a case of the Hungarian algorithm.
This motivates the use of Hungarian algorithm iteratively, which guarantees convergence to a local optimum.
\begin{prop}
\label{distributed}
Given the cost
\begin{equation*}
C_{jik}=
\begin{cases} \tau_1 \|v_{jk} + \sum_{-j,i,k}B_{-jik}v_{-jk}\|_2 -
\tau_1\|\sum_{-j,i,k}B_{-jik}v_{-jk}\|_2 + \log \frac{m_{-ji}}{J-m_{-ji}} , \text{ if } i \leq L_{-j}\\
\tau_1 + \log\frac{\gamma_0}{J} - \log(i-L_{-j}), \text{ if }L_{-j}< i \leq L_{-j} + K_j,
\end{cases}
\end{equation*}
where $-j$ denotes groups excluding group $j$ and $L_{-j}$ is the number of global topics before group $j$ (due to exchangeability of the IBP, group $j$ can always be considered last). Then, a locally optimum MAP estimate for Eq. \eqref{eq:post_hbp_1} can be obtained by iteratively employing the Hungarian algorithm to solve: for each group $j$, $(((B_{jik})))$ which maximizes $\sum_{j,i,k}B_{jik}C_{jik}$, subject to constraints:
(a) for each fixed $i$ and $j$, at most one of $B_{jik}$ is $1$, rest are $0$ and
(b) for each fixed $k$ and $j$, exactly one of $B_{jik}$ is $1$, rest are $0$.
After solving for $(((B_{jik})))$, $\theta_i$ is obtained as
$\theta_{i} = \frac{\sum_{j,k}B_{jik}v_{jk}}{\|\sum_{j,k}B_{jik}v_{jk}\|_2}.$
\end{prop}
The noisy topics for each of the groups can be obtained by applying CoSAC to corresponding documents, which is trivially parallel. Proof of Proposition \ref{distributed} is in Supplement \ref{supp:dm} and Distributed Matching algorithm is in Supplement \ref{supp:algos}.
\subsection{Dynamic Hierarchical Beta Process}
\label{subsec:combined}
Our ``full'' model, the Dynamic Hierarchical Beta Process model (dHBP), builds on the constructions described in subsections \ref{subsec:dynamic} and \ref{subsec:distributed} to enable the inference of temporal dynamics of collections of topic polytopes. We start by specifying the upper level Beta Process given by Eq. \eqref{eq:top_bp} and base measure $H$ given by Eq. \eqref{eq:vmf_time}. Next, for each group $j=1,\ldots,J$, we introduce an additional level of hierarchy to model group specific distributions over topics
\begin{equation}
\label{eq:j_top__bp}
Q_j|Q \thicksim \text{BP}(\gamma_j, Q)\text{, then }
Q_j := \sum_i p_{ji} \delta_{\theta_i},
\end{equation}
where $p_{ji}$s vary around corresponding $q_i$. The distributional properties of $p_{ji}$ are described in \citep{thibaux2007hierarchical}.
At any given time $t$, each group $j$ selects a subset from the common pool of global topics:
\begin{equation}
\label{eq:tj_bern}
\ensuremath{\mathcal{T}}_j^{(t)}|Q_{jt} \thicksim \text{BeP}(Q_{jt})\text{, then }
\ensuremath{\mathcal{T}}_j^{(t)} := {\textstyle\sum}_{i=1} b_{ji}^{(t)} \delta_{\theta^{(t)}_i}\text{, where }b_{ji}^{(t)}|p_{ji} \thicksim \text{Bern}(p_{ji}),\,\forall i.
\end{equation}
Let $\ensuremath{\mathcal{T}}_j^{(t)} := \{\theta^{(t)}_i : b_{ji}^{(t)} = 1, i=1,2,\ldots \}$ be the corresponding collection of atoms -- topics active at time $t$ in group $j$.
Noisy measurements of these topics are generated by:
\begin{equation}
\label{eq:tj_vmf_kalman}
v^{(t)}_{jk}|\ensuremath{\mathcal{T}}_j^{(t)} \thicksim \vmf(\ensuremath{\mathcal{T}}_{jk}^{(t)}, \tau_1)\text{ for }k = 1,\ldots,K_j^{(t)},\text{ where }K_j^{(t)} := \card(\ensuremath{\mathcal{T}}_j^{(t)}).
\end{equation}
The conditional distribution of global topics at $t$ given the state of the global topics at $t-1$ is
\begin{equation}
\label{eq:posterior_dHBP}
\begin{split}
& \mathbb P(\theta^{(t)},B^{(t)}| \theta^{(t-1)},v^{(t)},\{ B^{(a)}\}_{a=1}^{t-1}) \propto \\
&
\exp\left(\tau_0 {\textstyle\sum}_{i=1} \langle \theta_i^{(t)}, \theta_i^{(t-1)} \rangle)\right) F(\{m_{ji}^{(t-1)}\},\{m_{ji}^{(t)}\})
\cdot \exp\left({\textstyle\sum}_{j,i,k} \tau_1 B_{jik}^{(t)} \langle \theta_i^{(t)}, v_{jk}^{(t)}\rangle\right),
\end{split}
\end{equation}
where $F(\{m_{ji}^{(t-1)}\},\{m_{ji}^{(t)}\})$ is the prior term dependent on the popularity counts history from current and previous time points. Analogous to the Chinese Restaurant Franchise \citep{teh2006hierarchical}, one can think of an Indian Buffet Franchise in the case of HBP. A headquarter buffet provides some dishes each day and the local branches serve a subset of those dishes. Although this analogy seems intuitive, we are not aware of a corresponding Gibbs sampler and it remains to be a question of future studies. Therefore, unfortunately, we are unable to handle this prior term directly and instead propose a heuristic replacement --- stripping away popularity of topics across groups and only considering group specific topic popularity (groups still remain dependent through the atom locations).
\paragraph{Streaming Dynamic Distributed Matching (SDDM)} We combine our results to perform approximate inference of the model in Section \ref{subsec:combined}. Using Hungarian algorithm, iterating over groups at time $t$ obtain estimates for $(((B_{jik}^{(t)})))$ based on the following cost $C_{jik}^{(t)} =$
\begin{equation*}
\begin{cases} \|\tau_1 v_{jk}^{(t)} + \tau_1 \sum_{-j,i,k}B_{-jik}^{(t)}v_{-jk}^{(t)} + \tau_0\theta_i^{(t-1)}\|_2 - \|\sum_{-j,i,k}B_{-jik}^{(t)}v_{-jk}^{(t)} + \tau_0\theta_i^{(t-1)}\|_2 + \log \frac{1+m_{ji}^{(t)}}{t - m_{ji}^{(t)}},\\
\tau_1 + \log\frac{\gamma_0}{J}- \log(i-L_{-j}^{(t)}),\text{ if } L_{-j}^{(t)} < i \leq L_{-j}^{(t)} + K_j^{(t)},
\end{cases}
\end{equation*}
where first case is if $i \leq L_{-j}^{(t)}$; $m_{ji}^{(t)}$ denotes the popularity of topic $i$ in group $j$ up to time $t$ (plus one is used to indicate that global topic $i$ exists even when $m_{ji}^{(t)}=0$). Then compute global topic estimates
$\theta_i^{(t)}= \frac{\tau_1\sum_{j,k}B_{jik}^{(t)}v_{jk}^{(t)} +\tau_0 \theta_i^{(t-1)}}{\|\tau_1\sum_{j,k}B_{jik}^{(t)}v_{jk}^{(t)} +\tau_0 \theta_i^{(t-1)}\|_2}.$
At time point $t$, the noisy topics for each of the groups can be obtained by applying CoSAC to corresponding documents in parallel. SDDM algorithm is described in Supplement \ref{supp:algos}.
\section{Experiments}
\label{sec:experiments}
We study ability of our models to learn the latent temporal dynamics and discover new topics that change over time. Next we show that our models scale well by utilizing temporal and group inherent data structures. We also study hyperparameters choices. We analyze two datasets: the Early Journal Content (\url{http://www.jstor.org/dfr/about/sample-datasets}), and a collection of Wikipedia articles partitioned by categories and in time according to their popularity.
\subsection{Temporal Dynamics and Topic Discovery}
\paragraph{Early Journal Content.}
The Early Journal Content dataset spans years from $1665$ up to $1922$. Years before $1882$ contain very few articles, and we aggregated them into a single timepoint. After preprocessing, dataset has $400k$ scientific articles from over $400$ unique journals. The vocabulary was truncated to $4516$ words. We set all articles from the last available year ($1922$) aside for the testing purposes.
\paragraph{Case study: epidemics.} The beginning of the $20$th century is known to have a vast history of disease epidemics of various kinds, such as smallpox, typhoid, yellow fever to name a few. Vaccines or effective treatments for the majority of them were developed shortly after. One of the journals represented in the EJC dataset is the "Public Health Report"; however, publications from it are only available starting $1896$. Primary objective of the journal was to reflect epidemic disease infections. As one of the goals of our modeling approach is topic discovery, we verify that the model can discover an epidemics-related topic around $1896$. Figure \ref{fig:dbp_epidemics} shows that SDM correctly discovered a new topic is $1896$ semantically related to epidemics. We plot the evolution of probabilities of the top $15$ words in this topic across time. We observe that word "typhoid" increases in probability towards $1910$ in the "epidemics" topic, which aligns with historical events such as Typhoid Mary in $1907$ and chlorination of public drinking water in the US in $1908$ for controlling the typhoid fever. The probability of "tuberculosis" also increases, aligning with foundation of the National Association for the Study and Prevention of Tuberculosis in $1904$.
\paragraph{Case study: law.} Some of the EJC journals are related to the topic of law. Our DM algorithm identified a global topic semantically similar to law by matching similar topics present in 32 out of the 417 journals. In Figure \ref{fig:ejc_dm_law} we present the learned global topic and 4 examples of the matched local topics with the corresponding journal names. Our algorithm correctly identified that these journals have a shared law topic.
\subsection{Scalability}
\begin{figure}
\centering
\begin{subfigure}{0.52\textwidth}
\centering
\includegraphics[scale=0.25]{dbp_epidemics.pdf}
\caption{SDM \emph{Epidemics}: evolution of top 15 words}
\label{fig:dbp_epidemics}
\end{subfigure}
\begin{subfigure}{0.46\textwidth}
\centering
\vspace{0.07in}
\includegraphics[scale=0.2]{interpretability_global_group_topics.pdf}
\vspace{0.07in}
\caption{DM \emph{Law}: matched topics from journals}
\label{fig:ejc_dm_law}
\end{subfigure}
\caption{Qualitative examples of topics learned by SDM and DM algorithms on the EJC data}
\label{fig:ejc_interpretability}
\vskip -0.1in
\end{figure}
\begin{wrapfigure}[5]{r}{0.5\linewidth}
\vspace{-.7in}
\centering
\begin{subfigure}{0.23\textwidth}
\centering
\includegraphics[width=\linewidth]{imgs/Wiki_CoSAC_VS_DM.pdf}
\vskip -0.1in
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\centering
\includegraphics[width=\linewidth]{imgs/Wiki_SDM_VS_SDDM.pdf}
\vskip -0.1in
\end{subfigure}
\vskip -0.1in
\caption{Comparison on Wiki Data (20 cores)}
\label{fig:model_comp_wiki}
\end{wrapfigure}
\paragraph{Wiki Corpus.} We collected articles from Wikipedia and their page view counts for the $12$ months of $2017$ and category information (e.g., Arts, History). We used categories as groups and partitioned the data across time according to the page view counts. Dataset construction details are given in Supplement \ref{supp:data_wiki}. The total number of documents is about $3$ million, and we reduced vocabulary to $7359$ words similarly to \citep{hoffman2010online}. For testing we set aside documents from category Art from December $2017$.
\paragraph{Modeling Grouping.}
In Fig. \ref{fig:model_comp_wiki} we present comparisons
on Wiki data: CoSAC \citep{yurochkin2017conic} v.s DM under the static \textit{distributed} setting and SDM v.s SDDM under the dynamic \textit{streaming} setting.
Fig. \ref{fig:model_comp_wiki} (left) shows that for data accessible in groups, DM outperforms CoSAC by$~\sim25X$, as DM runs CoSAC on different data groups in parallel and then matches the outputs. Matching time adds only a small overhead compared to the runtime of CoSAC. Similarly, in Fig. \ref{fig:model_comp_wiki} (right), SDDM is $\sim 6X$ faster than SDM, since SDDM can process documents of different groups in parallel and interleaves CoSAC with matching: while matching is being performed on data groups with timestamp $t$, CoSAC can process the data that arrives with timestamp $t+1$ in parallel.
\begin{table*}[t]
\vskip -0.1in
\caption{Modeling topics of EJC \hspace{0.5cm} || Modeling Wikipedia articles}
\vskip -0.1in
\centering
\begin{tabular}{lrrrr || rrrr}
\toprule
{} & Perplexity & Time & Topics & Cores & Perplexity & Time & Topics & Cores\\
\midrule
SDM & \textbf{1179} & 22min & 125 & 1 & 1254 & 2.4hours & 182 & 1\\
DM & 1361 & 5min & 125 & 20 & 1260 & \textbf{15min} & 182 & 20\\
SDDM & 1241 & \textbf{2.3min} & 103 & 20 & \textbf{1201} & 20min & 238 & 20\\
DTM & 1194 & 56hours & 100 & 1 & NA & >72hours & 100 & 1 \\
SVB & 1840 & 3hours & 100 & 20 & 1219 & 29.5hours & 100 & 20\\
CoSAC & 1191 & 51min & 132 & 1 & 1227 & 4.4hours& 173 & 1\\
\bottomrule
\end{tabular}
\label{table:results}
\vskip -0.2in
\end{table*}
\paragraph{Modeling temporality}
also benefits scalability. We compare our methods with other topic models on both Wiki and EJC datasets: Streaming Variational Bayes (SVB) \citep{broderick2013streaming} and Dynamic Topic Models (DTM) \citep{blei2006dynamic} trained with $100$ topics. Perplexity scores on the held out data, training times, computing resources and number of topics are reported in Table \ref{table:results}.
On the wiki dataset, SDDM took only $20$min to process approximately $3$ million documents, which is much faster than the other approaches.
Regarding perplexity scores, SDDM generally outperforms DM, which suggests that modeling time is beneficial. For the EJC dataset, SDM outperforms SDDM. Modeling groups might negatively affect perplexity because the majority of the EJC journals (groups) have very few articles (i.e. less than $100$ -- a setup challenging for many topic modeling algorithms). On the Wiki corpus each category (group) has sufficient amount of training documents and time-group partitioning considered by SDDM achieves the best perplexity score.
\begin{wrapfigure}[12]{r}{0.25\linewidth}
\renewcommand{\figurename}{Fig.}
\vspace{-.75in}
\centering
\begin{subfigure}[b]{0.25\textwidth}
\centering
\includegraphics[scale=0.15]{imgs/ejc_sdm_perplexity_gamma_1.pdf}
\vskip -0.08in
\caption{EJC perplexity}
\label{fig:ejc_sdm_perplexity}
\end{subfigure} \\
\begin{subfigure}[b]{0.25\textwidth}
\centering
\includegraphics[scale=0.15]{imgs/ejc_sdm_topic_gamma_1.pdf}
\vskip -0.08in
\caption{\# of topics in EJC}
\label{fig:ejc_sdm_topic}
\end{subfigure}
\vskip -0.05in
\caption{SDM parameters}
\label{fig:sdm_sensitivity}
\vskip -0.15in
\end{wrapfigure}
\subsection{Parameter choices}
The rate of topic dynamics of the SDM and SDDM is effectively controlled by $\tau_0$, where smaller values imply higher dynamics rate. Parameter $\tau_1$ controls variance of local topics around corresponding global topics in all of our models. This variance dictates how likely a local topic to be matched to an existing global topic. When this variance is small, the model will tend to identify local topics as new global topics more often. Lastly, $\gamma_0$ affects the probability of new topic discovery, which scales with time and number of groups. In the preceding experiments we set $\tau_0=2,\tau_1=1,\gamma_0=1$ for SDM; $\tau_1=2,\gamma_0=1$ for DM; $\tau_0=4,\tau_1=2,\gamma_0=2$ for SDDM. In Figure \ref{fig:sdm_sensitivity} we show heatpmaps for perplexity and number of learned topics, fixing $\gamma_0=1$ and varying $\tau_0$ and $\tau_1$. We see that for large $\tau_1$, SDM identifies more topics to fit the smaller variability constraint imposed by the parameter.
\section{Discussion and Conclusion}
\label{sec:discuss}
Our work suggests the naturalness of incorporating sophisticated Bayesian nonparametric techniques
in the inference of rich latent geometric structures of interest.
We demonstrated the feasibility of \emph{approximate} nonparametric learning at scale, by utilizing suitable geometric representations and devising fast algorithms for obtaining reasonable point estimates for such representations. Further directions include incorporating more meaningful geometric
features into the models (e.g., via more elaborated base measure modeling for the Beta Process) and developing efficient algorithms for full Bayesian inference.
For instance, the latent geometric structure of the problem is solely encoded in the base measure.
We want to explore choices of base measures for other geometric structures such as collections of k-means centroids, principal components, etc. Once an appropriate base measure is constructed, our Beta process based models can be utilized
to enable a new class of Bayesian nonparametric models amenable to scalable inference and suitable for analysis of large datasets. In our concurrent work we have utilized model construction similar to one from Section \ref{subsec:distributed} to perform Federated Learning of neural networks trained on heterogeneous data \citep{yurochkin2019pfnm} and proposed a general framework for model fusion \citep{yurochkin2019spahm}.
\subsubsection*{Acknowledgments}
This research is supported in part by grants NSF CAREER DMS-1351362, NSF CNS-1409303, a research
gift from Adobe Research and a Margaret and Herman Sokol Faculty Award to XN.
|
\section{Introduction}
Increasing interest for developing smart energy management system within scientific community has thrived for improvement of short term energy prediction (STEP) algorithms with high accuracy without any computational overload. Model predictive controller based energy management systems are capable of dealing with the uncertainties in energy demand to some extent by taking receding horizon approach. However, receding horizon control requires prediction of future renewable generation and demand (thermal and electrical) beforehand. The short-term prediction helps the energy management system to schedule the energy sources in more cost efficient way avoiding under or over energy generation, plan maintenance work without compromising consumers' comfort \cite{landa2015}. The performance of the algorithms are affected due to several factors. The renewable generation is often subjected to fluctuations induced due to meteorological factors such as irradiation, wind-speed, dust cover etc, which are both intrinsic and extrinsic to the operation of the PV panels. Similarly, both thermal and electrical demand are time-varying parameters that depend on numerous factors which includes type of the day (working day or weekend), month of a year, climatic conditions and so on. These factors are inherently non-linear and time-varying.
Fuzzy inference systems (FIS) have been evolved successfully for solving different real-time problems due to continuous input-output mapping and interpretation abilities \cite{wu2011continuity,rong2007adaptive,chatterjee2008augmented,uuguz2012adaptive}. In particular, since their advent, neural-fuzzy approaches have become the foremost tool as they inherently assimilate both the learning capability of a neural network and the ability of a FIS to capture and model underlying non-linear characteristics of real-life data with promising accuracies \cite{jang1993anfis}. Kasabov proposed one of the first adaptive neuro-fuzzy inference systems (NFIS) \cite{kasabov2001evolving}, in which rules and parameters are updated by the guidance of a hybrid online supervised/unsupervised learning scheme in response to new ensuing data. Dynamic evolving neuro-fuzzy inference system (DENFIS)\cite{kasabov2002denfis} uses a clustering method to evolve the rules and update parameters. This network chooses \textit{m}-most significant rules for prediction through the offline clustering technique, which makes DENFIS not suitable for online circumstances. Dynamic fuzzy neural network (D-FNN) \cite{wu2000dynamic,wu2001fast} dynamically adjusts the width of the RBF unit of the TSK-based extended RBF neural network by a hierarchical online self-organized learning depending on the total training data. The absence of total training data limits its usage in offline learning only. A self-constructing neuro-fuzzy inference network (SONFIN) \cite{juang1998online} proposes an input data alignment scheme for clustering and measures a projection-based correlation for evolving rules.
This paper focuses on three of the adaptive FIS that, in true sense, implements sequential learning strategies for rule and parameter, namely meta-cognitive fuzzu inference system (McFIS) \cite{Subramanian2012}, sequential adaptive fuzzy inference system (SAFIS) \cite{rong2006sequential} and evolving Takagi-Sugeno model (ETS) \cite{Angelov2004}. This approaches are well-established in several machine learning problems such as classification, system identification problems \cite{Subramanian2012,angelov2008evolving,rong2006sequential,Subramanian2013}. This investigation analyses and compares the performance of these approaches in terms of their prediction error and architectural complexity for load forecasting in urban buildings and renewable energy generation, where there is a deficiency of storing a large amount of historical data. Also, these dataset shows an ample amount of uncertainties depending on the season, type of the day etc. Apart from this, the behaviour of these approaches is also examined while predicting renewable energy generation with external input to verify the improvement in performance of these FIS.
The rest of the paper is organized as follows. Section II describes the STEP problem definition followed by the working principle of ETS, SAFIS and McFIS. Section III presents the dataset that are used for simulation and case study with parameter settings of these particular algorithms. Section IV concludes the study with a course of potential future implementation of these algorithms in real-time systems.
\section{Energy Forecasting Problem and Adaptive Neural-Fuzzy Approaches}
This section starts with energy prediction problem definition followed by a brief yet comprehensive descriptions of the algorithms that are mainly focused for energy demand and renewable generation prediction problems. The algorithms considered here have the capability of online learning and prediction. In case of online learning, the training data are collected and used for parameter update sequentially. The rule base and the parameters are upgraded or modified depending on the strength of the information possessed by the data sample.
\subsection{Energy Forecasting Problem}
STEP addresses the problems of one-hour-ahead to several-day-ahead energy prediction. NFIS learns from a set of training samples given by $\left\{ \left( \mathbf{u^1}, \mathbf{v^1}\right), \dots, \left( \mathbf{u}^k, \mathbf{v}^k\right), \dots \right\}$ where $\mathbf{u}^k = \left[p^k(t),\dots,~p^k(t-\nu+1), r^k(t),\dots,r^k(t-\mu+1) \right]^T \in$ $\mathbb{R}^\nu \times \mathbb{R}^\mu$ is the input vector which consists of the past $\nu$ energy demand time-series sample-points and $\mu$ previous input points to the dynamical system. $\mathbf{v}^k = \left[p^k(t+1),~p^k(t+2),\dots,~p^k(t+\gamma) \right]^T$ $\in \mathbb{R}^m$ is the vector of future responses. $\gamma$ is known as prediction horizon. Forecasting problem can be defined as functional mapping between input and output of a dynamical system $\mathbf{\Phi} : \mathbf{u}^k \in \mathbb{R}^n \times \mathbb{R}^p \to \mathbf{v}^k \in \mathbb{R}^m$ based on past response of the system. The predicted output of the system is given by
\begin{equation}
\hat{\mathbf{v}}^k = \hat{\mathbf{\Phi}} \left[\mathbf{u}^k, \mathbf{\lambda}\right].
\label{forecast_eqn}
\end{equation}
where, $\mathbf{\lambda}$ is the parameters of the FIS network. The objective is to approximate the function $\mathbf{\Phi[\cdot]}$ such that predicted response $\hat{\mathbf{v}}^k$ is as close as possible to system's actual response $\mathbf{v}^k$.
\subsection{Evolving Takagi-Sugeno Model (ETS)}
ETS proposed by Angelov \textit{et al.} \cite{Angelov2004} uses an on-line clustering technique to gradually evolve Takagi-Sugeno (TS) fuzzy model. It verifies the information content of the data sequentially to update or modify the fuzzy rules. The information content of each data is extracted using a information potential measurement and the spatial proximity of the data samples to the already existing fuzzy rules in the fuzzy sub-space. The algorithmic flow for on-line learning and prediction of ETS is listed as follows.
\begin{enumerate}
\item \textbf{Step 1:} During first iteration, the first data sample is considered as a the focus of the first cluster or rules. This first data sample forms the antecedent part of the first rule using a user-define membership function.
\item \textbf{Step 2:} As the next data sample is considered, the potential of the data sample in the fuzzy rule space is measured recursively using a Cauchy type function by calculating a projection of the distance between the current sample and the previous samples.
\item \textbf{Step 3:} The potential of the cluster centres are updated considering the data samples information in a recursive way using the information from the previous sample points.
\item \textbf{Step 4:} The potential of the new data sample in the fuzzy rule space is compared with the potential of the already existing rule centres. The decision of adding a new rule is made if the potential of the data sample is higher than the potential of the existing cluster centres. Alternatively, the cluster centres' potentials are updated in the next iteration as described in step 3.
\item \textbf{Step 5:} In the penultimate step, The parameters of the consequent part of the rule base are either updated globally using recursive least square (RLS) method or updated locally using weighted RLS technique.
\item \textbf{Step 6:} The final step is to predict the output of the data sample. The iteration is loop again continuous starting from step 2 as the new data sample is collected.
\end{enumerate}
In ETS, as the TS model is evolved in each time step, and the model parameters are modified depending on the gradual change in the cluster centres, the rule base is expected to grow.
\subsection{Sequential Adaptive Fuzzy Inference System (SAFIS)}
SAFIS \cite{rong2006sequential} uses the idea of the \textit{influence of a fuzzy rule} to upgrade the rule base. In statistical sense, the influence of the fuzzy rule is defined as the contribution of that particular fuzzy rule in predicting the overall output. It uses the distance information of the current sample from the existing rules to update the parameters. Only the parameters related to the nearest rules are updated using extended Kalman filter if the measured distance is below a certain threshold. Alternatively, the new rule is added. It also incorporates rule pruning technique considering that the influence of the particular rule is under pre-defined threshold. The usage of the current sample only for updating parameters and upgrading rule base are the reason of achievement of fast computation.
\subsection{Meta-cognitive Neuro-Fuzzy Inference System (McFIS)}
McFIS as proposed by Subramanian \textit{et al.} \cite{Subramanian2012} has been developed based on simple meta-cognition model of Nelson and Narens \cite{nelson1990metamemory}. The working principle of McFIS differs from other adaptive NFIS as discussed follows. The detail parameter update and rule growing and pruning along with architectural desriptions are given in \cite{Subramanian2012}.
\begin{enumerate}
\item The meta-cognitive unit acts as a self-regulatory learning component which controls the learning mechanism of the cognitive component by assessing the current knowledge and identifying the new knowledge based on the state of the cognitive component.
\item Depending on the prediction error knowledge, meta-cognition regulates the learning ability of the main NFIS network with \textit{how-to-learn}, \textit{when-to-learn} and \textit{what-to-learn}, for each samples observed sequentially by the network. Thus, McFIS has the ability to escape over-training.
\item In order to address the aforementioned learning strategies, McFIS takes three simple actions: {\em{(i)}} remove samples with similar information (\textit{sample deletion}); {\em{(ii)}} grow or prune rules and parameter update depending on the information content of the current sample (\textit{sample learning}); {\em{(iii)}} use less informative samples at a later stage of the learning process to tune the parameters (\textit{sample reserve}). This helps McFIS to avoid over-training but to keep generalization ability.
\end{enumerate}
\begin{table}[h]
\centering
\caption{1-h-ahead energy demand prediction error comparison}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{NFIS} & \textbf{RMSE} & \textbf{NDEI} & \textbf{RULES} \\ \hline
eTS & 0.1558 & 0.5948 & 12 \\ \hline
SAFIS & 0.2291 & 0.8660 & 17 \\ \hline
McFIS & \textbf{0.1555} & \textbf{0.5878} & \textbf{5} \\ \hline
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{5-h-ahead energy demand prediction error comparison}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{NFIS} & \textbf{RMSE} & \textbf{NDEI} & \textbf{RULES} \\ \hline
eTS & 0.2094 & 0.7439 & \textbf{6} \\ \hline
SAFIS & 0.2947 & 1.0310 & 22 \\ \hline
McFIS & \textbf{0.1657} & \textbf{0.5798} & 28 \\ \hline
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{5-min-ahead energy demand prediction error comparison}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{NFIS} & \textbf{RMSE} & \textbf{NDEI} & \textbf{RULES} \\ \hline
eTS & 0.1019 & 0.5021 & 16 \\ \hline
SAFIS & 0.1174 & 0.5793 & 162 \\ \hline
McFIS & \textbf{0.0964} & \textbf{0.4755} & \textbf{15} \\ \hline
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{1-h-ahead renewable energy prediction error comparison}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{NFIS} & \textbf{RMSE} & \textbf{NDEI} & \textbf{RULES} \\ \hline
eTS & 0.1334 & 0.6519 & 19 \\ \hline
SAFIS & 0.1496 & 0.7315 & 167 \\ \hline
McFIS & \textbf{0.1328} & \textbf{0.6491} & \textbf{5} \\ \hline
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{5-min-ahead renewable energy prediction error comparison including temperature variation}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{NFIS} & \textbf{RMSE} & \textbf{NDEI} & \textbf{RULES} \\ \hline
eTS & 0.0909 & 0.5323 & 25 \\ \hline
SAFIS & 0.0976 & 0.5724 & 163 \\ \hline
McFIS & \textbf{0.0867} & \textbf{0.5085} & \textbf{5} \\ \hline
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{1-h-ahead renewable energy prediction error comparison including temperature variation}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{NFIS} & \textbf{RMSE} & \textbf{NDEI} & \textbf{RULES} \\ \hline
eTS & 0.1365 & 0.7923 & 23 \\ \hline
SAFIS & 0.1239 & 0.7510 & 161 \\ \hline
McFIS & \textbf{0.1221} & \textbf{0.7114} & \textbf{7} \\ \hline
\end{tabular}
\end{table}
\section{Experimental Analysis}
\subsection{Description of the Dataset}
The energy demand data is measured in one of the academic buildings at Nanyang technological university campus, Singapore. The data is collected during the month of May, 2015 with a sampling interval of 1 h over a period of 7 days. Four previous samples are used as the input vector for learning and prediction in all of the algorithms i.e. $\nu=4$. $85$\% of the data points are used for sequential learning of these algorithms. For 1-h and 5-h ahead prediction of energy demand, $\gamma$ is considered to be 1 and 5 respectively.
The renewable energy generation data has been taken during the month of January, 2016 with sampling interval of 5 min. 70\% of the complete dataset is used for training. As similar to the energy data, $\nu=4$ is considered. For 5-min and 1-h ahead prediction of renewable energy, $\gamma$ is set at 1 and 12 respectively.
\begin{figure*}
\centering
\begin{subfigure}{.4\linewidth}
\centering
\includegraphics[scale=0.45]{E1.pdf}
\caption{}
\label{subfig:conv1}
\end{subfigure}
\hspace{10pt}
\begin{subfigure}{.4\linewidth}
\centering
\includegraphics[scale=0.45]{PV1.pdf}
\caption{}
\label{subfig:conv2}
\end{subfigure}
\begin{subfigure}{.4\linewidth}
\centering
\includegraphics[scale=0.45]{E5.pdf}
\caption{}
\label{subfig:conv4}
\end{subfigure}
\begin{subfigure}{.4\linewidth}
\centering
\includegraphics[scale=0.45]{PV1h.pdf}
\centering
\caption{}
\label{subfig:conv5}
\end{subfigure}
\caption{Actual vs predicted output data using McFIS. (a) 1-h-ahead energy demand prediction, (b) 5-min-ahead renewable energy prediction, (c) 5-h-ahead energy demand prediction, (d) 1-h-ahead renewable energy prediction.}
\label{fig:conv1}
\end{figure*}
\subsection{Prediction Performance Analysis}
The prediction errors of these approaches are presented in Table I-IV. The original dataset is normalized for further analysis and prediction. It is to be mentioned that, this analysis only considers time-series prediction, in which any external input to the dynamical system i.e. $u$ is absent. Two error measures termed as root mean squared error (RMSE) and non-destructive error index (NDEI) are used for analysing performance. Besides, the number of fuzzy rules used is also tabulated to reflect the architectural complexity of these inference systems during prediction.
In case of 1-h-ahead energy demand forecasting as shown in Table I, the prediction accuracies of both McFIS and ETS are comparable. However, McFIS achieves this accuracy using less fuzzy rules, which reduces network architecture complexity. Although McFIS uses a large number of fuzzy rules for 5-h-ahead energy demand forecasting, it reduces the prediction error of ETS and SAFIS by approximately 20\% and 40\% respectively as shown in Table II. In case of renewable energy prediction, McFIS performs significantly well in terms of both RMSE and fuzzy rules over other two approaches as depicted in Table III and IV.
The predicted value of the McFIS is depicted in Figure 1. Figure 1(a) and Figure 1(c) illustrates that McFIS is able to predict the sharp changes in time series dynamics. Similarly it is able to detect and predict the quick changing dynamics of the renewable energy prediction, as shown in Figure 1(b) and 1(d), although it is prominent in Figure 1(d) that, during night hours when there is no output from photo voltaic panel due to absence of solar irradiation, McFIS predicts a non-zero output.
\begin{figure*}
\centering
\begin{subfigure}{.3\linewidth}
\centering
\includegraphics[scale=0.33]{T1.pdf}
\caption{}
\label{subfig:conv1}
\end{subfigure}
\begin{subfigure}{.3\linewidth}
\centering
\includegraphics[scale=0.33]{P1T.pdf}
\caption{}
\label{subfig:conv2}
\end{subfigure}
\begin{subfigure}{.3\linewidth}
\centering
\includegraphics[scale=0.33]{P1hT.pdf}
\caption{}
\label{subfig:conv1}
\end{subfigure}
\caption{Renewable energy prediction with temperature as input signal using McFIS. (a) Temperature variation, (b) 5-min-ahead renewable energy prediction, (c) 1-h-ahead renewable energy prediction}
\label{fig:conv1}
\end{figure*}
\subsection{Renewable Energy Prediction with Temperature Variation}
This investigation also considers temperature as an input feature to the dynamical system to predict the renewable energy generation. Only current instant temperature is considered, i.e. $\mu=1$. The temperature variation as measured at a sampling period of 5 min during January 2016 over a period of 10 days is shown in Figure 2. Table V and VI tabulates the prediction error and number of fuzzy rules used. One can observe that although eTS and SAFIS fails to achieve a significant improvement, McFIS certainly improves the prediction error for 1-h-ahead forecasting and architectural complexity in both 5-min and 1-h-ahead prediction problems. Also it can be observed from Figure 2(c) that incorporating temperature as input variable helps to increase the prediction accuracy of McFIS during no-energy-generation hours.
\subsection{Rank-based Statistical Comparison}
In order to find the statistical significance of the outcome of McFIS algorithm, Friedman's rank test is realized. For ease of understanding, we denote the problems as follows.
\begin{enumerate}
\item[] \textit{F$_1$}: 1-h-ahead energy demand prediction,
\item[] \textit{F$_2$}: 5-h-ahead energy demand prediction,
\item[] \textit{F$_3$}: 5-min-ahead renewable energy prediction,
\item[] \textit{F$_4$}: 1-h-ahead renewable energy prediction,
\item[] \textit{F$_5$}: 5-min-ahead renewable energy prediction with temperature as input,
\item[] \textit{F$_6$}: 1-h-ahead renewable energy prediction with temperature as input.
\end{enumerate}
\begin{table}[]
\centering
\caption{Average ranking of the algorithms}
\begin{tabular}{l|c|c|c}
\hline
\textbf{Probelms} & \textbf{ETS} & \textbf{SAFIS} & \textbf{McFIS} \\ \hline
\textbf{F1} & 2 & 3 & 1 \\ \hline
\textbf{F2} & 2 & 3 & 1 \\ \hline
\textbf{F3} & 2 & 3 & 1 \\ \hline
\textbf{F4} & 2 & 3 & 1 \\ \hline
\textbf{F5} & 2 & 3 & 1 \\ \hline
\textbf{F6} & 3 & 2 & 1 \\ \hline
\textbf{Avg Rank} & 2.17 & 2.83 & \textbf{1} \\ \hline
\end{tabular}
\end{table}
\begin{table}[]
\centering
\caption{Average rank difference and statistics}
\begin{tabular}{c|c}
\hline
Algorithm & \begin{tabular}[c]{@{}l@{}}Difference between average \\ ranking w.r.t McFIS\end{tabular} \\ \hline
ETS & 1.17 \\ \hline
SAFIS & 1.83 \\ \hline
Critical difference (Bonferroni-Dunn) & \begin{tabular}[c]{@{}l@{}}1.3 ($\alpha$ = 0.05) , \\ 1.13 ($\alpha$ = 0.01)\end{tabular} \\ \hline
\end{tabular}
\end{table}
Table VII shows the rank, average rank of the these algorithms considering RMSE as the performance metric. Friedman test initially assumes that all the algorithms are performing similar. Depending on the average ranking, the test computes a $Q$ value. In this study, the Q value is found as 10.3068 which is greater than the Friedman statistical value atconfidence level of 95\% and 99\% which are 7 and 9 respectively for a problem with 3 treatments and 6 blocks problem \cite{demvsar2006}. As a result, the null hypothesis can be rejected and it can be concluded that the average ranking is statistically significant. In our case, as the number of hypothesis and number of dataset are low, a pairwise post-hoc Bonferroni-Dunn test will verify the statistical significance of the predictive algorithms under investigation which is recorded in Table VIII. This particular test states that the performance of a particular algorithm has statistical impact if the difference between the average rank is greater than a critical difference with some confidence \cite{demvsar2006}. Table VIII signifies that the performance of McFIS is better than that of SAFIS and eTS with a 99\% confidence level.
\section{Conclusion}
This investigation studies the performance of some of the well-established adaptive FIS in prediction short term energy demand and renewable generation. From the analysis it is found that McFIS has the ability to predict data with the promising accuracy by keeping the number of fuzzy rules as low as possible. It has also been observed that presence of external relevant parameters e.g. temperature variation in renewable generation not only improves computational requirement in McFIS but also shows a promising direction towards improving the prediction error. The low computational requirement as well as high prediction accuracy makes McFIS more suitable forecasting tool for STEP in urban buildings.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec.introduction}
The lattice Boltzmann (LB) method has developed into an attractive numerical
method over the past three decades for simulating complex fluid flows
\cite{Aidun2010, Gross2010, Huang2016.amr} and solving various partial
differential equations \cite{Ginzburg2005, Dellar2011, Chai2018}.
Historically, the LB method originates from the lattice gas automata (LGA)
to eliminate the statistical noise \cite{McNamara1988}, and thus it inherits
some distinguishing features from LGA, such as the simple algorithm (local
collision and linear streaming) and the easy incorporation of microscopic
interactions \cite{Shan1993, Ladd1994}. Afterward, it is found that the
classical LB model for hydrodynamic flows can be derived from the
Boltzmann-BGK equation via systematic discretization \cite{He1997a,
He1997b}, and then various LB models for multiphase flows
\cite{He1998.nonideal, Luo1998} and thermo-hydrodynamic flows (i.e., thermal
fluid flows) \cite{He1998.thermal, Guo2007} have been established from the
kinetic models in an {\it a priori} manner.
\par
Since most hydrodynamic flows involve some forms of thermal effects,
thermo-hydrodynamic flows are extensively encountered in nature and
engineering, and the LB method for simulating thermo-hydrodynamic flows has
attracted continuous attention since the early 1990s \cite{He1998.thermal,
Guo2007, Shan1997, Alexander1993, McNamara1997, Zheng2008, Lallemand2003,
Mezrhab2004, Mezrhab2010, Huang2014.ibm, Contrino2014}. However, it remains
open-ended though the LB method has achieved great success in simulating
isothermal fluid flows \cite{Sbragaglia2009, Succi2015}. Generally, the
existing LB models for thermo-hydrodynamic flows can be categorized into
three major groups: the multispeed approach \cite{McNamara1997, Zheng2008},
the double-distribution-function (DDF) approach \cite{He1998.thermal,
Guo2007, Shan1997}, and the hybrid approach \cite{Lallemand2003,
Mezrhab2004}. The multispeed approach uses a single distribution function
(DF) to describe the mass, momentum, and energy conservation laws, and thus
it requires more discrete velocities than the standard lattice (i.e., it
requires the multispeed lattice). By definition, the DDF approach consists
of double DFs, with one DF for the mass and momentum conservation laws and
the other DF for the energy conservation law. In the hybrid approach, the
mass and momentum conservation laws are described by one DF, while the
energy conservation law is described by a macroscopic governing equation
that is solved via the conventional computational fluid dynamics methods.
Severe numerical instability \cite{Lallemand2003} and complexity of boundary
condition treatment \cite{Lee2018} are usually encountered in the multispeed
approach. As to the hybrid approach, it acts as a compromised solution that
deviates from the mesoscopic LB method \cite{Lallemand2003}, and the viscous
dissipation is usually ignored in this approach \cite{Feng2018, Safari2018}.
On the contrary, the DDF approach, free of the above drawbacks, is most
widely studied and adopted in real applications.
\par
Most of the existing DDF LB models for thermo-hydrodynamic flows are
inherently a decoupling model, which means that the recovered equation of
state (EOS) is a decoupling EOS $p_\text{\tiny EOS}^{} = \rho R_g T_0$ ($R_g$ is the gas
constant and $T_0$ is the reference temperature), where the pressure is not
directly related to the temperature. Consequently, these LB models are
restricted to the thermo-hydrodynamic flows under the Boussinesq
approximation (i.e., the decoupling thermo-hydrodynamic flows). Based on the
DDF kinetic model constructed by Guo et al.\ \cite{Guo2007}, and by applying
the discretization of velocity space presented by Shan et al.\
\cite{Shan2006} that can lead to the temperature-independent discrete
velocities, Hung and Yang \cite{Hung2011} proposed a DDF LB model aimed at
recovering the ideal-gas EOS $p_\text{\tiny EOS}^{} = \rho R_g T$. However, the
deviation in the third-order moment of the equilibrium distribution function
(EDF) for density due to the constraint of standard lattice, as previously
identified by Prasianakis and Karlin \cite{Prasianakis2007}, is not
considered in Hung and Yang's model, and meanwhile, an error also exists in
their derived EDF for total energy. In 2012, by introducing the correction
term for the third-order moment of the EDF for density and deriving the
correct EDF for total energy, Li et al.\ \cite{Li2012} developed a DDF LB
model for simulating coupled thermo-hydrodynamic flows. The ideal-gas EOS
can be recovered by Li et al.'s model, and the simulation of natural
convection with a large temperature difference is reported \cite{Li2012}.
Following the similar way, Feng et al.\ \cite{Feng2015} proposed
three-dimensional DDF LB models. A correction term for the second-order
moment of the EDF for total energy is further introduced by Feng et al.\
\cite{Feng2015} to enhance the numerical stability of the LB equation for
total energy DF. Recently, the cascaded collision scheme is employed in the
LB equation for density DF to enhance the numerical stability by Fei and Luo
\cite{Fei2018}, while the single-relaxation-time (SRT) collision scheme is
still used in the LB equation for total energy DF.
\par
It is worth pointing out that the ideal-gas EOS is recovered by the above
DDF LB models \cite{Hung2011, Li2012, Feng2015, Fei2018}, which indicates
that these models are only applicable to the coupled thermo-hydrodynamic
flows of ideal gases. Moreover, in these models, the LB equation for total
energy DF is complicated due to the consideration of the viscous dissipation
and compression work, and thus it is difficult to employ the
multiple-relaxation-time (MRT) or cascaded collision schemes in this LB
equation to enhance the numerical stability although the MRT and cascaded
collision schemes have been employed in the LB equation for density DF
\cite{Li2012, Fei2018}. Most recently, we developed an LB model with
self-tuning EOS for multiphase flows \cite{Huang2018.eos}. Since the
recovered EOS can be self-tuned via a built-in variable, this model serves
as a good and distinct starting point for developing a novel LB model for
coupled thermo-hydrodynamic flows, which is the main objective of the
present work. To be specific, a novel MRT LB equation for solving the energy
conservation equation, with considering the viscous dissipation and
compression work, is developed. Furthermore, boundary condition treatment
for simulating coupled thermo-hydrodynamic flows is also proposed on the
basis of the judicious decomposition of DF into three parts rather than two.
The remainder of the present paper is organized as follows. In Section
\ref{sec.lbm}, a novel LB model for coupled thermo-hydrodynamic flows is
developed. In Section \ref{sec.bc}, boundary condition treatment is
proposed. Numerical validations of the present LB model are carried out in
Section \ref{sec.validation}, and a brief conclusion is drawn in Section
\ref{sec.conclusion}.
\par
\section{Lattice Boltzmann model} \label{sec.lbm}
The present LB model for coupled thermo-hydrodynamic flows is developed on
the basis of the recent LB model with self-tuning EOS for multiphase flows.
Double DFs are involved: one is the density DF used to solve the velocity
field (i.e., the mass-momentum conservation equations), and the other is the
total energy DF used to solve the temperature field (i.e., the energy
conservation equation). The full coupling of thermo-hydrodynamic effects is
achieved via the self-tuning EOS recovered by the LB equation for density DF
and the viscous dissipation and compression work considered in the LB
equation for total energy DF. Both the LB equations for density and total
energy DFs are based on the standard lattice. For the sake of simplicity and
clarity, the two-dimensional model will be developed here, and its extension
to three-dimensional model is straightforward. The standard two-dimensional
nine-velocity (D2Q9) lattice is given as \cite{Qian1992}
\begin{equation}
\mathbf{e}_i =
\begin{cases}
c \big( 0 ,\, 0 \big) ^\text{T}, & i=0, \\
c \big( \cos[(i-1)\pi/2] ,\, \sin[(i-1)\pi/2] \big) ^\text{T},
& i=1,2,3,4, \\
\sqrt{2}c \big( \cos[(2i-1)\pi/4] ,\, \sin[(2i-1)\pi/4] \big) ^\text{T},
& i=5,6,7,8, \\
\end{cases}
\end{equation}
where the lattice speed $c = \delta_x / \delta_t$ with $\delta_x$ and
$\delta_t$ being the lattice spacing and time step, respectively.
\par
\subsection{LB equation for density DF}
The recently developed LB equation for density DF that can recover a
self-tuning EOS is briefly introduced here for self-completeness. The MRT LB
equation for density DF $f_i (\mathbf{x}, t)$ can be expressed as
\cite{Huang2018.eos}
\begin{subequations}\label{eq.LBE.u}
\begin{equation}\label{eq.LBE.u.s}
f_i (\mathbf{x} + \mathbf{e}_i \delta_t, t + \delta_t)
= \bar{f} _i (\mathbf{x}, t),
\end{equation}
\begin{equation}\label{eq.LBE.u.c}
\bar{ \mathbf{m} } (\mathbf{x}, t) = \mathbf{m} + \delta_t \mathbf{F}_m -
\mathbf{S} \left( \mathbf{m} - \mathbf{m}^\text{eq} + \dfrac{\delta_t}{2}
\mathbf{F}_m \right) - \mathbf{R} \left( \mathbf{I} -
\dfrac{\mathbf{S}}{2}
\right) \left( \mathbf{m} - \mathbf{m}^\text{eq} + \dfrac{\delta_t}{2}
\mathbf{F}_m \right) - \delta_x \mathbf{T} \cdot \nabla \rho -
\dfrac{\delta_x}{c^2} \mathbf{X} \cdot \nabla p_\text{\tiny LBE}^{},
\end{equation}
\end{subequations}
where Eq.\ (\ref{eq.LBE.u.s}) is the streaming process executed in velocity
space and Eq.\ (\ref{eq.LBE.u.c}) is the collision process executed in
moment space at position $\mathbf{x}$ and time $t$. The moment of density DF
in Eq.\ (\ref{eq.LBE.u.c}) is given as $\mathbf{m} = \mathbf{M} (f_i)^\text{T}$.
Here, $\mathbf{M}$ is the dimensionless transformation matrix
\cite{Lallemand2000}
\begin{equation}\label{eq.M}
\mathbf{M} =
\begin{bmatrix}
1& 1& 1& 1& 1& 1& 1& 1& 1 \\
-4&-1&-1&-1&-1& 2& 2& 2& 2 \\
4&-2&-2&-2&-2& 1& 1& 1& 1 \\
0& 1& 0&-1& 0& 1&-1&-1& 1 \\
0&-2& 0& 2& 0& 1&-1&-1& 1 \\
0& 0& 1& 0&-1& 1& 1&-1&-1 \\
0& 0&-2& 0& 2& 1& 1&-1&-1 \\
0& 1&-1& 1&-1& 0& 0& 0& 0 \\
0& 0& 0& 0& 0& 1&-1& 1&-1 \\
\end{bmatrix},
\end{equation}
and $(f_i)^\text{T}$ denotes the vector $(f_0, f_1, \cdots, f_8)^\text{T}$. The
post-collision density DF in Eq.\ (\ref{eq.LBE.u.s}) is obtained via the
inverse transformation $( \bar{f}_i )^\text{T} = \mathbf{M}^{-1} \bar{\mathbf{m}}$, and the
post-collision moment $\bar{\mathbf{m}}$ is computed by Eq.\ (\ref{eq.LBE.u.c}). The
last three terms on the right-hand side (RHS) of Eq.\ (\ref{eq.LBE.u.c}) are
the correction terms aimed at eliminating the additional cubic terms of
velocity in the recovered momentum conservation equation
\cite{Huang2018.cubic}, where $p_\text{\tiny LBE}^{}$ denotes the recovered EOS by the
LB equation. The macroscopic density $\rho$ and velocity $\mathbf{u}$ are
defined as
\begin{equation}\label{eq.rho.u}
\rho = \sum\nolimits_i f_i, \quad
\rho \mathbf{u} = \sum\nolimits_i \mathbf{e}_i f_i + \dfrac{\delta_t}{2}
\mathbf{F},
\end{equation}
where $\mathbf{F}$ is the force term. In the recent LB model for multiphase
flows \cite{Huang2018.eos}, $\mathbf{F}$ is the total force due to the
long-range molecular interaction, while in the present LB model for coupled
thermo-hydrodynamic flows, $\mathbf{F}$ is simply an external force, such as
the gravity force.
\par
In Eq.\ (\ref{eq.LBE.u.c}), the equilibrium moment function for density DF
$\mathbf{m} ^\text{eq}$ is given as \cite{Huang2018.eos}
\begin{equation}
\begin{split}
\mathbf{m}^\text{eq} = \big[ \rho ,\,
2\alpha_1^{} \rho + 2\beta_1^{} \eta + 3 \rho |\hat{\mathbf{u}}|^2 ,\,
\alpha_2^{} \rho + \beta_2^{} \eta - 3 \rho |\hat{\mathbf{u}}|^2 + 9 \rho \hat{u}_x^2
\hat{u}_y^2 &,\,
\\
\rho \hat{u}_x ,\,
-\rho \hat{u}_x + 3 \rho \hat{u}_x \hat{u}_y^2 ,\,
\rho \hat{u}_y ,\,
-\rho \hat{u}_y + 3 \rho \hat{u}_y \hat{u}_x^2 &,\,
\rho (\hat{u}_x^2 - \hat{u}_y^2) ,\,
\rho \hat{u}_x \hat{u}_y \big] ^\text{T} ,
\end{split}
\end{equation}
where $\hat{\mathbf{u}} = \mathbf{u} /c$ and $\eta$ is the built-in variable aimed at
achieving a self-tuning EOS. The coefficients $\alpha_1^{}$ and $\beta_1^{}$
are set to $-1$ and $1$, respectively, while the coefficients $\alpha_2^{}$
and $\beta_2^{}$ are determined by Eq.\ (\ref{eq.LBE.u.coefficients}). The
discrete force term in moment space $\mathbf{F}_m$ is given as
\begin{equation}
\begin{split}
\mathbf{F}_m = \big[ 0 ,\,
6 \hat{\mathbf{F}} \cdot \hat{\mathbf{u}} ,\,
-6 \hat{\mathbf{F}} \cdot \hat{\mathbf{u}} + 9[\hat{\mathbf{F}} \hat{\mathbf{u}} \hat{\mathbf{u}} \hat{\mathbf{u}}]_{xxyy} ,\,
\hat{F}_x ,\,
-\hat{F}_x + 3[\hat{\mathbf{F}} \hat{\mathbf{u}} \hat{\mathbf{u}}]_{xyy} &,\,
\\
\hat{F}_y ,\,
-\hat{F}_y + 3[\hat{\mathbf{F}} \hat{\mathbf{u}} \hat{\mathbf{u}}]_{xxy} &,\,
2 (\hat{F}_x \hat{u}_x - \hat{F}_y \hat{u}_y) ,\,
\hat{F}_x \hat{u}_y + \hat{F}_y \hat{u}_x \big] ^\text{T} ,
\end{split}
\end{equation}
where $\hat{\mathbf{F}} = \mathbf{F} /c$, and the square bracket and its subscript
denote permutation and tensor index, respectively. For example, $[\hat{\mathbf{F}} \hat{\mathbf{u}}
\hat{\mathbf{u}} \hat{\mathbf{u}}]_{xxyy} = 2 \hat{F}_x \hat{u}_x \hat{u}_y^2 + 2 \hat{F}_y \hat{u}_y \hat{u}_x^2$. To
correctly recover the Newtonian viscous stress tensor, the collision matrix
in moment space $\mathbf{S}$ is modified as follows \cite{Huang2018.eos}
\begin{equation}
\mathbf{S} =
\begin{bmatrix}
s_0^{}& 0& 0& 0& 0& 0& 0& 0& 0\\
0& s_e^{}& k s_\varepsilon^{} \omega_e^{}&
0& h \hat{u}_x s_q^{} \omega_e^{}&
0& h \hat{u}_y s_q^{} \omega_e^{}& 0& 0\\
0& 0& s_\varepsilon^{}& 0& 0& 0& 0& 0& 0\\
0& 0& 0& s_j^{}& 0& 0& 0& 0& 0\\
0& 0& 0& 0& s_q^{}& 0& 0& 0& 0\\
0& 0& 0& 0& 0& s_j^{}& 0& 0& 0\\
0& 0& 0& 0& 0& 0& s_q^{}& 0& 0\\
0& 0& 0& 0& 2 b \hat{u}_x s_q^{} \omega_p^{}&
0& -2 b \hat{u}_y s_q^{} \omega_p^{}& s_p^{}& 0\\
0& 0& 0& 0& b \hat{u}_y s_q^{} \omega_p^{}&
0& b \hat{u}_x s_q^{} \omega_p^{}& 0& s_p^{}\\
\end{bmatrix},
\end{equation}
where $\omega_{e, p}^{} = s_{e, p}^{} /2 -1$, and $k$, $h$, and $b$ are the
coefficients. Through the Chapman-Enskog analysis, the coefficients in
$\mathbf{m}^\text{eq}$ and $\mathbf{S}$ should satisfy the following relations
\begin{equation}\label{eq.LBE.u.coefficients}
\alpha_2^{} = - \dfrac{2\alpha_1^{} + \varpi +1}{1-\varpi} ,\quad
\beta_2^{} = - \dfrac{2 \beta_1^{}}{1-\varpi} ,\quad
k = 1-\varpi ,\quad
h = \dfrac{6 \varpi (1-\varpi)}{1-3\varpi} ,\quad
b = \dfrac{1-\varpi}{1-3\varpi} ,
\end{equation}
where $\varpi$ is related to the bulk viscosity.
\par
In Eq.\ (\ref{eq.LBE.u.c}), the last three terms, together with the
high-order terms of velocity in $\mathbf{m}^\text{eq}$ and $\mathbf{F}_m$, are
introduced to eliminate the additional cubic terms of velocity
\cite{Huang2018.cubic}, which are not considered in the previous DDF LB
models for coupled thermo-hydrodynamic flows. The correction matrix
$\mathbf{R}$ is a $9 \times 9$ matrix and it is set as \cite{Huang2018.eos}
\begin{subequations}
\begin{equation}
\mathbf{R} =
\begin{bmatrix}
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& R_{11}& 0& 0& 0& 0& 0& R_{17}& R_{18}\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& R_{71}& 0& 0& 0& 0& 0& R_{77}& R_{78}\\
0& R_{81}& 0& 0& 0& 0& 0& R_{87}& R_{88}\\
\end{bmatrix},
\end{equation}
where the nonzero elements can be determined via the Chapman-Enskog analysis
as follows
\begin{equation}
\setlength{\arraycolsep}{0.6em}
\begin{array}{lll}
R_{11} = -\tfrac{ (9-15k-2h) s_e^{} }{4\varpi} (\hat{u}_x^2 + \hat{u}_y^2), &
R_{17} = -\tfrac{ 3(9-3k+2h) s_p^{} (2-s_e^{}) }{4(2-s_p^{})}
(\hat{u}_x^2 - \hat{u}_y^2), &
R_{18} = \tfrac{ 12(3k+h) s_p^{} (2-s_e^{}) }{2-s_p^{}} \hat{u}_x \hat{u}_y, \\
R_{71} = -\tfrac{ (3-4b) s_e^{} (2-s_p^{}) }{4\varpi (2-s_e^{})}
(\hat{u}_x^2 - \hat{u}_y^2), &
R_{77} = -\tfrac{ 3(3+4b) s_p^{} }{4} (\hat{u}_x^2 + \hat{u}_y^2), &
R_{78} = 0, \\
R_{81} = \tfrac{ b s_e^{} (2-s_p^{}) }{\varpi (2-s_e^{})} \hat{u}_x \hat{u}_y, &
R_{87} = 0, &
R_{88} = 6 b s_p^{} (\hat{u}_x^2 + \hat{u}_y^2). \\
\end{array}
\end{equation}
\end{subequations}
The correction matrix $\mathbf{T}$ is set as \cite{Huang2018.eos}
\begin{subequations}
\begin{equation}
\mathbf{T} = \big(
\mathbf{0} ,\;
\mathbf{T}_1 ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{T}_7 ,\;
\mathbf{T}_8 \big) ^\text{T},
\end{equation}
whose element is a vector implying that the dimensions of $\mathbf{T}$ are
$9 \times 2$. The nonzero elements in $\mathbf{T}$ can also be determined
via the Chapman-Enskog analysis as follows
\begin{equation}
\mathbf{T}_1 = \tfrac{ 3(2-s_e^{}) }{2}
\begin{bmatrix}
(1-k) \hat{u}_x^3 - (2k+h) \hat{u}_x \hat{u}_y^2 \\
(1-k) \hat{u}_y^3 - (2k+h) \hat{u}_x^2 \hat{u}_y
\end{bmatrix} ,\quad
\mathbf{T}_7 = \tfrac{ 2-s_p^{} }{2}
\begin{bmatrix}
\hat{u}_x^3 + 2b \hat{u}_x \hat{u}_y^2 \\
-\hat{u}_y^3 - 2b \hat{u}_x^2 \hat{u}_y
\end{bmatrix} , \quad
\mathbf{T}_8 = - \tfrac{ b(2-s_p^{}) }{2}
\begin{bmatrix}
\hat{u}_y^3 + 2 \hat{u}_x^2 \hat{u}_y \\
\hat{u}_x^3 + 2 \hat{u}_x \hat{u}_y^2
\end{bmatrix} .
\end{equation}
\end{subequations}
Similarly to $\mathbf{T}$, the correction matrix $\mathbf{X}$ is a $9 \times
2$ matrix and it is set as \cite{Huang2018.eos}
\begin{subequations}
\begin{equation}
\mathbf{X} = \big(
\mathbf{0} ,\;
\mathbf{X}_1 ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{0} ,\;
\mathbf{X}_7 ,\;
\mathbf{X}_8 \big) ^\text{T},
\end{equation}
where the nonzero elements are given as
\begin{equation}
\mathbf{X}_1 = \tfrac{ 9 (2k+h) (2-s_e^{}) }{2}
\begin{bmatrix}
\hat{u}_x \hat{u}_y^2 \\
\hat{u}_x^2 \hat{u}_y
\end{bmatrix} ,\quad
\mathbf{X}_7 = -3 b (2-s_p^{})
\begin{bmatrix}
\hat{u}_x \hat{u}_y^2 \\
-\hat{u}_x^2 \hat{u}_y
\end{bmatrix} ,\quad
\mathbf{X}_8 = \tfrac{ 3 b (2-s_p^{})}{2}
\begin{bmatrix}
\hat{u}_y^3 + 2 \hat{u}_x^2 \hat{u}_y \\
\hat{u}_x^3 + 2 \hat{u}_x \hat{u}_y^2
\end{bmatrix} .
\end{equation}
\end{subequations}
Here, we would like to point out that for the coupled thermo-hydrodynamic
flows under the low Mach number condition, these correction terms for the
additional cubic terms of velocity can be simply ignored. However, they are
kept in the present work for the sake of theoretical completeness and
computational accuracy.
\par
Through the Chapman-Enskog analysis, the following mass-momentum
conservation equations can be recovered \cite{Huang2018.eos}
\begin{equation}
\begin{cases}
\partial_t \rho + \nabla \cdot (\rho \mathbf{u}) = 0, \\
\partial_t (\rho \mathbf{u}) + \nabla \cdot (\rho \mathbf{uu}) =
-\nabla p_\text{\tiny LBE}^{} + \mathbf{F} + \nabla \cdot \mathbf{\Pi}, \\
\end{cases}
\end{equation}
where $p_\text{\tiny LBE}^{}$ and $\mathbf{\Pi}$ are the recovered EOS and viscous
stress tensor
\begin{equation}\label{eq.NSE}
p_\text{\tiny LBE}^{} = c_s^2 [ (2+\alpha_1^{}) \rho + \beta_1^{} \eta ] ,\quad
\mathbf{\Pi} = \rho \nu [\nabla \mathbf{u} +
\mathbf{u} \nabla - (\nabla \cdot \mathbf{u}) \mathbf{I}] + \rho
\varsigma
(\nabla \cdot \mathbf{u}) \mathbf{I} ,
\end{equation}
where the lattice sound speed $c_s = c / \sqrt{3}$, the kinematic viscosity
$\nu = c_s^2 \delta_t \big( s_p^{-1} - 0.5 \big)$, and the bulk viscosity
$\varsigma = \varpi c_s^2 \delta_t \big( s_e^{-1} - 0.5 \big)$. As seen in
Eq.\ (\ref{eq.NSE}), the recovered EOS $p_\text{\tiny LBE}^{}$ can be arbitrarily tuned
via the built-in variable $\eta$.
\par
\subsection{LB equation for total energy DF}
Since the EOS recovered by the above LB equation for solving the velocity
field can be self-tuned, we are now well equipped to simulate coupled
thermo-hydrodynamic flows. The remaining task is to develop an LB equation
for solving the temperature field, in which the viscous dissipation and
compression work are consistently considered.
\par
\subsubsection{Energy conservation equation}
The collision term of an LB equation conserves macroscopic quantity, and the
recovered macroscopic conservation equation for this quantity usually has a
conservative form (see Eq.\ (\ref{eq.NSE}) as an example). On the basis of
this principle, the total energy conservation equation, in which the viscous
dissipation and compression work are expressed as $\nabla \cdot (\mathbf{u}
\cdot \mathbf{\Pi}) - \nabla \cdot (p_\text{\tiny EOS}^{} \mathbf{u})$, is a better and
more natural starting point for directly developing an LB equation at the
discrete level than the internal energy conservation equation, in which the
viscous dissipation and compression work are expressed as $\mathbf{\Pi} :
\nabla \mathbf{u} - p_\text{\tiny EOS}^{} \nabla \cdot \mathbf{u}$. Here, $p_\text{\tiny EOS}^{}$ is
the pressure determined by the adopted EOS. To facilitate the development of
an LB equation, the total energy conservation equation is reformulated as
\begin{equation}\label{eq.ECE}
\partial_t (\rho E) + \nabla \cdot (\rho H \mathbf{u}) =
\nabla \cdot (\lambda \nabla T + \mathbf{u} \cdot \mathbf{\Pi}) +
\mathbf{u} \cdot \mathbf{F} + q,
\end{equation}
where $E$ is the total energy, $H = E + p_\text{\tiny EOS}^{} / \rho$ is the total
enthalpy, $T$ is the temperature that can be determined by the internal
energy $\epsilon$ ($\epsilon = E - | \mathbf{u} |^2/2$) and density $\rho$,
$\lambda$ is the heat conductivity, and $q$ is the source term. In Eq.\
(\ref{eq.ECE}), the viscous dissipation $\nabla \cdot (\mathbf{u} \cdot
\mathbf{\Pi})$ combines with the conduction term $\nabla \cdot (\lambda
\nabla T)$ to constitute the term $\nabla \cdot (\lambda \nabla T +
\mathbf{u} \cdot \mathbf{\Pi})$, and the compression work $- \nabla \cdot
(p_\text{\tiny EOS}^{} \mathbf{u})$ combines with the convection term $\nabla \cdot
(\rho E \mathbf{u})$ to constitute the term $\nabla \cdot (\rho H
\mathbf{u})$. Meanwhile, we can also combine the work done by force
$\mathbf{u} \cdot \mathbf{F}$ and the source term $q$ to constitute an
equivalent source term $q_e = \mathbf{u} \cdot \mathbf{F} + q$. Thus, Eq.\
(\ref{eq.ECE}) can be viewed as a general convection-diffusion equation with
source term. Here, we would like to point out that the above reformulation
is consistent with the Chapman-Enskog analysis, which means that the two
terms combined together are of the same order.
\par
\subsubsection{Viscous stress tensor}
To consider the viscous dissipation in the LB equation for total energy DF,
we first recall the recovery of viscous stress tensor by the above LB
equation for density DF. On the basis of the Chapman-Enskog analysis, the
viscous stress tensor $\mathbf{\Pi}$ is of order $\varepsilon^1$ and can be
expressed as \cite{Huang2018.eos}
\begin{equation}\label{eq.Pi}
\mathbf{\Pi} = \varepsilon^1 \mathbf{\Pi}^{(1)} = - \varepsilon^1 c^2
\begin{bmatrix}
\tfrac12 \tilde{G}_7^{(1)} & \tilde{G}_8^{(1)} \\
\tilde{G}_8^{(1)} & -\tfrac12 \tilde{G}_7^{(1)} \\
\end{bmatrix}
- \varepsilon^1 c^2
\begin{bmatrix}
\tfrac16 \tilde{G}_1^{(1)} & 0 \\
0 & \tfrac16 \tilde{G}_1^{(1)} \\
\end{bmatrix},
\end{equation}
where $\varepsilon$ is the small expansion parameter in the Chapman-Enskog
analysis and $\tilde{\mathbf{G}}^{(1)}$ is
\begin{equation}\label{eq.tG}
\tilde{\mathbf{G}}^{(1)} = \dfrac{ \mathbf{m}^{(1)} + \bar{\mathbf{m}}^{(1)} }{2} ,
\end{equation}
where $\mathbf{m} ^{(1)}$ and $\bar{\mathbf{m}}^{(1)}$ are the $\varepsilon^1
\text{-order}$ terms of $\mathbf{m}$ and $\bar{\mathbf{m}}$ in their Chapman-Enskog
expansions $\mathbf{m} = \sum\nolimits _{n=0}^{+\infty} \varepsilon^n
\mathbf{m} ^{(n)}$ and $\bar{\mathbf{m}} = \sum\nolimits _{n=0}^{+\infty} \varepsilon^n
\bar{\mathbf{m}}^{(n)}$, respectively. Here, it is worth pointing out that the
post-collision moment $\bar{\mathbf{m}}^{(1)}$ is kept in Eq.\ (\ref{eq.tG}) rather than
being substituted by Eq.\ (\ref{eq.LBE.u.c}). As a consequence, the
post-collision moment $\bar{\mathbf{m}}$, which is computed in the collision process of
density DF, can be directly utilized to consider the viscous dissipation in
the LB equation for total energy DF (see \ref{app.Implementation}).
Moreover, from the Chapman-Enskog analysis of the LB equation for density
DF, we can easily know that the $\varepsilon^0 \text{-order}$ terms of
$\mathbf{m}$ and $\bar{\mathbf{m}}$ satisfy
\begin{equation}\label{eq.m.order1}
\mathbf{m}^{(0)} = \bar{\mathbf{m}}^{(0)} = \mathbf{m} ^\text{eq}.
\end{equation}
\par
\subsubsection{LB equation}
For the energy conservation equation given by Eq.\ (\ref{eq.ECE}), the total
energy DF $g_i (\mathbf{x}, t)$ is introduced here, and the MRT LB equation
for $g_i (\mathbf{x}, t)$ is devised as
\begin{subequations}\label{eq.LBE.T}
\begin{equation}\label{eq.LBE.T.s}
g_i (\mathbf{x} + \mathbf{e}_i \delta_t, t + \delta_t) =
\bar{g} _i (\mathbf{x}, t),
\end{equation}
\begin{equation}\label{eq.LBE.T.c}
\bar{\mathbf{n}} (\mathbf{x}, t) = \mathbf{n} + \delta_t \mathbf{Q}_m -
\mathbf{L} \left( \mathbf{n} - \mathbf{n}^\text{eq} +
\dfrac{\delta_t}{2} \mathbf{Q}_m \right) + c^2 \mathbf{Y} \left(
\dfrac{\mathbf{m} + \bar{\mathbf{m}}}{2} - \mathbf{m}^\text{eq} \right),
\end{equation}
\end{subequations}
where Eqs.\ (\ref{eq.LBE.T.s}) and (\ref{eq.LBE.T.c}) represent the
streaming process in velocity space and the collision process in moment
space, respectively. The moment of total energy DF in Eq.\
(\ref{eq.LBE.T.c}) is given as $\mathbf{n} = \mathbf{M} (g_i) ^\text{T}$, the
post-collision total energy DF in Eq.\ (\ref{eq.LBE.T.s}) is obtained via
$(\bar{g}_i) ^\text{T} = \mathbf{M} ^{-1} \bar{\mathbf{n}}$, and the
post-collision moment $\bar{\mathbf{n}}$ is computed by Eq.\
(\ref{eq.LBE.T.c}). Here, the dimensionless transformation matrix
$\mathbf{M}$ is also given by Eq.\ (\ref{eq.M}). On the RHS of Eq.\
(\ref{eq.LBE.T.c}), the last density-DF-related term is introduced to
consider the viscous dissipation, in which $\mathbf{Y}$ is a $9 \times 9$
matrix that will be discussed and determined later. By definition, the
macroscopic total energy $\rho E$ is given as
\begin{equation}\label{eq.E}
\rho E = \sum\nolimits_i g_i + \dfrac{\delta_t}{2} q_e,
\end{equation}
where $q_e = \mathbf{u} \cdot \mathbf{F} + q$ is the equivalent source term.
Then, the total enthalpy $H$ and the temperature $T$ can be determined via
the thermodynamic relations $H = E + p_\text{\tiny EOS}^{} / \rho$ and $T = T (\epsilon,
\rho)$ (a function of internal energy $\epsilon = E - |\mathbf{u}| ^2/2$ and
density $\rho$), respectively. In the present work, a simple relation $T =
\epsilon / C_v$, though it strictly holds only for the ideal gases, is
adopted for the sake of simplicity, and more general or empirical relations
can be adopted as required by specific applications. Here, $C_v$ is the
specific heat at constant volume.
\par
To recover the targeted energy conservation equation, as well as inspired by
the ideas of our previous works on solid-liquid phase change
\cite{Huang2015, Huang2016.amr}, the equilibrium moment function for total
energy DF $\mathbf{n} ^\text{eq}$ is devised as
\begin{equation}
\begin{split}
\mathbf{n} ^\text{eq} = \big[
\rho E ,\,
-4 (\rho E - \rho_0^{} C_{p,0} T) + \gamma_1^{} \rho_0^{} C_{p,0} T ,\,
4 (\rho E - \rho_0^{} C_{p,0} T) + \gamma_2^{} \rho_0^{} C_{p,0} T &,\, \\
\rho H \hat{u}_x ,\,
-\rho H \hat{u}_x ,\,
\rho H \hat{u}_y ,\,
-\rho H \hat{u}_y &,\,
0 ,\,
0
\big] ^\text{T},
\end{split}
\end{equation}
where $\rho_0^{}$ and $C_{p,0}$ are the reference density and the reference
specific heat at constant pressure, respectively, and $\gamma_1^{}$ and
$\gamma_2^{}$ are the coefficients related to the heat conductivity.
Similarly to $\mathbf{n} ^\text{eq}$, the discrete source term in moment space
$\mathbf{Q}_m$ is devised as
\begin{equation}
\mathbf{Q}_m = \big(
q_e ,\,
\gamma_1^{} q_e ,\,
\gamma_2^{} q_e ,\,
q_e \hat{u}_x ,\,
-q_e \hat{u}_x ,\,
q_e \hat{u}_y ,\,
-q_e \hat{u}_y ,\,
0 ,\,
0
\big) ^\text{T}.
\end{equation}
To avoid the deviation term caused by the convection term recovered at the
order of $\varepsilon^1$ in the diffusion term recovered at the order of
$\varepsilon^2$, the collision matrix in moment space $\mathbf{L}$ is
modified as follows \cite{Huang2014.cde}
\begin{equation}
\mathbf{L} =
\begin{bmatrix}
\sigma_0^{}& 0& 0& 0& 0& 0& 0& 0& 0\\
0& \sigma_e^{}& 0& 0& 0& 0& 0& 0& 0\\
0& 0& \sigma_\varepsilon^{}& 0& 0& 0& 0& 0& 0\\
0& 0& 0& \sigma_j^{}& \sigma_q^{} \omega_j^{}& 0& 0& 0& 0\\
0& 0& 0& 0& \sigma_q^{}& 0& 0& 0& 0\\
0& 0& 0& 0& 0& \sigma_j^{}& \sigma_q^{} \omega_j^{}& 0& 0\\
0& 0& 0& 0& 0& 0& \sigma_q^{}& 0& 0\\
0& 0& 0& 0& 0& 0& 0& \sigma_p^{}& 0\\
0& 0& 0& 0& 0& 0& 0& 0& \sigma_p^{}\\
\end{bmatrix},
\end{equation}
where $\omega_j^{} = \sigma_j^{} /2 -1$.
\par
Since the viscous stress tensor $\mathbf{\Pi}$ is only related to
$\tilde{G}_1^{(1)}$, $\tilde{G}_7^{(1)}$, and $\tilde{G}_8^{(1)}$ (see Eq.\ (\ref{eq.Pi})),
the matrix $\mathbf{Y}$ in the density-DF-related term, which is introduced
in Eq.\ (\ref{eq.LBE.T.c}) to consider the viscous dissipation, is set as
follows
\begin{equation}
\mathbf{Y} =
\begin{bmatrix}
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& Y_{31}& 0& 0& 0& 0& 0& Y_{37}& Y_{38}\\
0& Y_{41}& 0& 0& 0& 0& 0& Y_{47}& Y_{48}\\
0& Y_{51}& 0& 0& 0& 0& 0& Y_{57}& Y_{58}\\
0& Y_{61}& 0& 0& 0& 0& 0& Y_{67}& Y_{68}\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
0& 0& 0& 0& 0& 0& 0& 0& 0\\
\end{bmatrix},
\end{equation}
where $Y_{3 \alpha} + Y_{4 \alpha} = 0$ and $Y_{5 \alpha} + Y_{6 \alpha} =
0$ for $\alpha = 1$, $7$, and $8$. Through the Chapman-Enskog analysis (see
\ref{app.CE}), the nonzero elements in $\mathbf{Y}$ can be determined as
follows
\begin{equation}\label{eq.Y.nonzero}
\begin{array}{lll}
Y_{31} = \hat{u}_x /3, & Y_{37} = \hat{u}_x, & Y_{38} = 2 \hat{u}_y, \\
Y_{51} = \hat{u}_y /3, & Y_{57} =-\hat{u}_y, & Y_{58} = 2 \hat{u}_x. \\
\end{array}
\end{equation}
Then, the following macroscopic conservation equation can be recovered
\begin{equation}\label{eq.ECE.LBE}
\partial_t (\rho E) + \nabla \cdot (\rho H \mathbf{u}) = \nabla \cdot
\left[
\left( \dfrac{2}{3} + \dfrac{\gamma_1^{}}{2} + \dfrac{\gamma_2^{}}{3}
\right)
\rho_0^{} C_{p,0} c^2 \delta_t \left( \dfrac{1}{\sigma_j^{}} - \dfrac12
\right)
\nabla T + \mathbf{u} \cdot \mathbf{\Pi} \right] + q_e.
\end{equation}
Compared with Eq.\ (\ref{eq.ECE}), the heat conductivity is given as
$\lambda = (2/3 + \gamma_1^{}/2 + \gamma_2^{}/3) \rho_0^{} C_{p,0} c^2 \delta_t
\big(\sigma_j^{-1} - 0.5\big)$. It can be seen from Eq.\ (\ref{eq.ECE.LBE})
that the viscous dissipation and compression work are correctly considered.
\par
Before proceeding further, some discussion on the present LB model for
coupled thermo-hydrodynamic flows is in order. First, the MRT collision
scheme is employed in both the LB equations for density and total energy
DFs, and the collision matrix in moment space is modified to be a
nondiagonal matrix rather than being set as the conventional diagonal
matrix. Second, the Prandtl number $P\mspace{-1.5mu}r = C_p \mu / \lambda$ can be
arbitrarily adjusted. Here, $C_p$ is the specific heat at constant pressure,
and $\mu = \rho \nu$ is the dynamic viscosity. Third, the specific heat
ratio $\gamma = C_p / C_v$ can also be arbitrarily adjusted. Note that $C_p
- C_v$ depends on the adopted EOS, and $C_p - C_v = R_g$ holds only for the
ideal-gas EOS. Lastly, and most importantly, an arbitrary EOS (including the
nonideal-gas EOS) $p_\text{\tiny EOS}^{}$ can be prescribed, and the built-in variable
$\eta$ is inversely calculated via $\eta = [ c_s^{-2} p_\text{\tiny LBE}^{} - (2 +
\alpha_1^{}) \rho ] \big/ \beta_1^{}$ with $p_\text{\tiny LBE}^{} = p_\text{\tiny EOS}^{}$.
\par
\section{Boundary condition treatment} \label{sec.bc}
In real applications, the boundary conditions are usually given in terms of
the macroscopic variables, and thus additional treatment is required to
obtain the mesoscopic DFs at the boundary node. In this section, we propose
the boundary condition treatment for simulating coupled thermo-hydrodynamic
flows.
\par
\subsection{Macroscopic variables}
For the velocity field, the nonslip velocity boundary condition is
considered and the velocity on the boundary is directly specified. Due to
the full coupling of thermo-hydrodynamic effects, the density may
significantly vary near the boundary and also has a direct effect on the
heat transfer process. Thus, it is important to ensure the mass conservation
at the boundary node for simulating coupled thermo-hydrodynamic flows. In
the present boundary condition treatment, the boundary node $\mathbf{x}_b$
is exactly placed on the wall boundary, as shown in Fig.\ \ref{fig.BC}. The
post-collision density DF $\bar{f}_i (\mathbf{x}_b, t)$ hitting the wall
(i.e., streaming out of the computational domain) reverses its direction as
follows
\begin{equation}\label{eq.BB.f}
f_{\bar{i}, \text{temp}} (\mathbf{x}_b, t + \delta_t) = \bar{f}_i
(\mathbf{x}_b, t),
\end{equation}
where $\bar{i}$ means $\mathbf{e}_{\bar{i}} = - \mathbf{e}_i$, and the
subscript ``temp'' implies that the density DF $f_{\bar{i}, \text{temp}}
(\mathbf{x}_b, t + \delta_t)$ is temporary. After this ``bounce-back''
process, all the unknown density DFs at $\mathbf{x}_b$ and $t + \delta_t$
due to the absence of adjacent nodes are now obtained. Then, the density
$\rho (\mathbf{x}_b, t+\delta_t)$ can be computed via definition as usual
(i.e., $\rho = \sum\nolimits_i f_i$). Note that the velocity $\mathbf{u}
(\mathbf{x}_b, t + \delta_t)$ is directly specified. Obviously, the local
conservation of mass can be strictly satisfied at the boundary node.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig01.pdf}
\caption{Schematic of boundary condition treatment with $\mathbf{x}_b$
denoting the boundary node placed on the wall, $\mathbf{x}_f$ and
$\mathbf{x}_{f\mspace{-4.0mu}f}$ denoting the nearest and next-nearest fluid nodes in the
normal direction, and the solid and dashed arrow lines denoting the known
and unknown distribution functions after the streaming process.}
\label{fig.BC}
\end{figure}
As for the temperature field, the Dirichlet boundary condition with
specified temperature and the Neumann boundary condition with zero heat flux
(i.e., the adiabatic boundary condition) are considered. For the Dirichlet
boundary condition, since the temperature $T (\mathbf{x}_b, t + \delta_t)$
is directly specified, all the involved macroscopic variables, such as the
total energy $E (\mathbf{x}_b, t + \delta_t)$, the pressure $p_\text{\tiny EOS}^{}
(\mathbf{x}_b, t + \delta_t)$, and the total enthalpy $H (\mathbf{x}_b, t +
\delta_t)$, can be determined via the corresponding thermodynamic relations.
For the Neumann boundary condition with zero heat flux, the post-collision
total energy DF $\bar{g}_i (\mathbf{x}_b, t)$ hitting the wall reverses its
direction as follows
\begin{equation}\label{eq.BB.g}
g_{\bar{i}, \text{temp}} (\mathbf{x}_b, t + \delta_t) = \bar{g}_i
(\mathbf{x}_b, t),
\end{equation}
and thus all the unknown total enthalpy DFs at $\mathbf{x}_b$ and $t +
\delta_t$ due to the absence of adjacent nodes are temporarily obtained.
Then, the total energy $E (\mathbf{x}_b, t + \delta_t)$ can be computed via
definition as usual (i.e., $\rho E = \sum\nolimits_i g_i + \delta_t q_e
/2$), and all the involved macroscopic variables, such as the temperature $T
(\mathbf{x}_b, t + \delta_t)$, the pressure $p_\text{\tiny EOS}^{} (\mathbf{x}_b, t +
\delta_t)$, and the total enthalpy $H (\mathbf{x}_b, t + \delta_t)$, can be
determined via the corresponding thermodynamic relations.
\par
\subsection{Density and total energy DFs}
At the boundary node, the unknown density and total energy DFs obtained via
Eqs.\ (\ref{eq.BB.f}) and (\ref{eq.BB.g}) are only used to compute the
macroscopic density and total energy. In the present boundary condition
treatment, all the known and unknown DFs at the boundary node will be
updated to make sure that the defining equations of density, velocity, and
total energy (i.e., Eqs.\ (\ref{eq.rho.u}) and (\ref{eq.E})) exactly hold at
the boundary node. For this purpose, we decompose the moment of DF (the same
as the DF) into its equilibrium, force (source), and nonequilibrium parts,
i.e.,
\begin{subequations}\label{eq.decompose}
\begin{equation}
\mathbf{m} = \mathbf{m}^\text{eq} - \dfrac{\delta_t}{2} \mathbf{F}_m +
\mathbf{m} ^\text{neq},
\end{equation}
\begin{equation}
\mathbf{n} = \mathbf{n}^\text{eq} - \dfrac{\delta_t}{2} \mathbf{Q}_m +
\mathbf{n} ^\text{neq}.
\end{equation}
\end{subequations}
Note that the present nonequilibrium part $\mathbf{m}^\text{neq}$
($\mathbf{n}^\text{neq}$) in Eq.\ (\ref{eq.decompose}) is different from the
previous nonequilibrium part defined as $\mathbf{m} - \mathbf{m}^\text{eq}$
($\mathbf{n} - \mathbf{n}^\text{eq}$) \cite{Kruger2017} when the force (source)
term exists. Since the equilibrium parts ($\mathbf{m}^\text{eq}$ and
$\mathbf{n}^\text{eq}$) and the force (source) parts ($- \delta_t \mathbf{F}_m /2$
and $-\delta_t \mathbf{Q}_m /2$) are determined by the macroscopic
variables, $\mathbf{m}^\text{eq} (\mathbf{x}_b, t+\delta_t) - \delta_t
\mathbf{F}_m (\mathbf{x}_b, t+\delta_t) /2$ and $\mathbf{n}^\text{eq}
(\mathbf{x}_b, t+\delta_t) - \delta_t \mathbf{Q}_m (\mathbf{x}_b,
t+\delta_t) /2$ can be directly computed. As to the nonequilibrium parts
($\mathbf{m} ^\text{neq}$ and $\mathbf{n} ^\text{neq}$) at $\mathbf{x}_b$
and $t + \delta_t$, extrapolations are employed following the idea of the
nonequilibrium-extrapolation approach \cite{Guo2002a, Guo2002b}. However,
instead of simply extrapolating $\mathbf{m} ^\text{neq}$ and $\mathbf{n}
^\text{neq}$, we introduce the following terms
\begin{subequations}\label{eq.neq.redefinition}
\begin{equation}
\tilde{\mathbf{m}} ^\text{neq} = \left( \mathbf{I} -
\dfrac{\mathbf{S}}{2}
\right) \mathbf{m} ^\text{neq},
\end{equation}
\begin{equation}
\tilde{\mathbf{n}} ^\text{neq} = \left( \mathbf{I} -
\dfrac{\mathbf{L}}{2}
\right) \mathbf{n} ^\text{neq} + \dfrac{c^2 \mathbf{Y}}{2} \left(
\mathbf{I}
- \dfrac{\mathbf{S}}{2} \right) \mathbf{m} ^\text{neq},
\end{equation}
\end{subequations}
where $\mathbf{I}$ is the $9 \times 9$ identity matrix; and then the first-
and second-order nonequilibrium extrapolations are given as
\begin{subequations}\label{eq.neq.extrapolation}
\begin{equation}
\begin{array}{ll}
\text{first-order:} & \tilde{\mathbf{m}} ^\text{neq} (\mathbf{x}_b, t +
\delta_t) = \tilde{\mathbf{m}} ^\text{neq} (\mathbf{x}_f, t +
\delta_t), \\
\text{second-order:} & \tilde{\mathbf{m}} ^\text{neq} (\mathbf{x}_b, t +
\delta_t) = 2 \tilde{\mathbf{m}} ^\text{neq} (\mathbf{x}_f, t +
\delta_t) - \tilde{\mathbf{m}} ^\text{neq} (\mathbf{x}_{f\mspace{-4.0mu}f}, t +
\delta_t);
\end{array}
\end{equation}
\begin{equation}
\begin{array}{ll}
\text{first-order:} & \tilde{\mathbf{n}} ^\text{neq} (\mathbf{x}_b, t +
\delta_t) = \tilde{\mathbf{n}} ^\text{neq} (\mathbf{x}_f, t +
\delta_t), \\
\text{second-order:} & \tilde{\mathbf{n}} ^\text{neq} (\mathbf{x}_b, t +
\delta_t) = 2 \tilde{\mathbf{n}} ^\text{neq} (\mathbf{x}_f, t +
\delta_t) - \tilde{\mathbf{n}} ^\text{neq} (\mathbf{x}_{f\mspace{-4.0mu}f}, t +
\delta_t);
\end{array}
\end{equation}
\end{subequations}
where $\mathbf{x}_f$ and $\mathbf{x}_{f\mspace{-4.0mu}f}$ denote the nearest and
next-nearest fluid nodes in the normal direction, as shown in Fig.\
\ref{fig.BC}. Based on our numerical tests, the first-order extrapolation
has better stability but lower accuracy than the second-order extrapolation.
Note that although the present collision matrices $\mathbf{S}$ and
$\mathbf{L}$ are nondiagonal, $\mathbf{I} - \mathbf{S}/2$ and $\mathbf{I} -
\mathbf{L}/2$ in Eq.\ (\ref{eq.neq.redefinition}) are still invertible, and
their inverse matrices are given in \ref{app.Inverse}. Therefore, Eq.\
(\ref{eq.neq.extrapolation}) is compatible and can be easily implemented due
to the special forms of $( \mathbf{I} - \mathbf{S}/2 )^{-1}$ and $(
\mathbf{I} - \mathbf{L}/2 )^{-1}$. Moreover, different from the previous
nonequilibrium-extrapolation approach \cite{Guo2002a, Guo2002b, Guo2007},
the present boundary condition treatment is applicable to the situation when
the dynamic viscosity and thermal conductivity significantly vary with
temperature and hence with space because the collision matrices are
considered in the present extrapolations of nonequilibrium parts.
\par
\section{Validations and discussions} \label{sec.validation}
In this section, simulations of thermal Poiseuille and Couette flows are
first carried out to validate the present LB model with self-tuning EOS for
coupled thermo-hydrodynamic flows. Three different EOSs, including the
decoupling EOS, the ideal-gas EOS, and the Carnahan-Starling EOS for
rigid-sphere fluids \cite{Carnahan1969}, are adopted, which are given in
order as follows
\begin{subequations}\label{eq.EOSs}
\begin{equation}\label{eq.EOSs.1}
p_\text{\tiny EOS}^{} = \rho R_g T_0,
\end{equation}
\begin{equation}\label{eq.EOSs.2}
p_\text{\tiny EOS}^{} = \rho R_g T,
\end{equation}
\begin{equation}\label{eq.EOSs.3}
p_\text{\tiny EOS}^{} = Z \rho R_g T \quad \text{with} \quad Z= \dfrac{1 + b\rho/4 +
(b\rho/4)^2 - (b\rho/4)^3}{ (1-b\rho/4)^3 },
\end{equation}
\end{subequations}
where $Z$ is the compressibility factor with the coefficient $b$ set to
$\sqrt{2} \pi / (3 \rho_0^{})$ here. Then, the present LB model is applied to
the simulation of natural convection in a square cavity with a large
temperature difference. The ideal-gas EOS is adopted and the Rayleigh number
varies from $10^3$ up to $10^8$. In the following simulations, $\varpi =
1/6$, $\gamma_1^{} = -2$, and $\gamma_2^{} = 2$ are chosen. The relaxation
parameters in $\mathbf{S}$ satisfy $s_0^{} = s_j^{} =1$, $\big( s_p^{-1} -
0.5 \big) \big( s_q^{-1} - 0.5 \big) = 1/12$, and $s_\varepsilon^{} =
s_e^{}$ \cite{Huang2016.3rd}, and the relaxation parameters in $\mathbf{L}$
satisfy $\sigma_0^{} =1$, $\big( \sigma_j^{-1} - 0.5 \big) \big(
\sigma_e^{-1} - 0.5 \big) = 1/4$, $\sigma_\varepsilon^{} = \sigma_p^{} =
\sigma_e^{}$, and $\sigma_q^{} = \sigma_j^{}$ \cite{Huang2016.amr}.
Meanwhile, the ratio of bulk to kinematic viscosity $\varsigma / \nu$ is
fixed at $1$ unless otherwise stated.
\par
\subsection{Thermal Poiseuille flow}
The thermal Poiseuille flow, driven by a constant force $\mathbf{F} = (F_x,
\, 0) ^\text{T}$ between two parallel walls, is first simulated. Both the lower
and upper walls are at rest, and the temperature of the lower and upper
walls are kept at $T_c$ and $T_h$ ($T_c < T_h$), respectively. The Prandtl
number $P\mspace{-1.5mu}r = C_p \mu / \lambda$, the specific heat at constant pressure
$C_p$, and the dynamic viscosity $\mu$ are assumed to be constant. Thus, the
analytical solutions for velocity and temperature are given as
\cite{Guo2007}
\begin{subequations}\label{eq.Poiseuille.u.T}
\begin{equation}
\dfrac{u_x}{U_0} = 4 \dfrac{y}{W} \left( 1- \dfrac{y}{W} \right), \quad
\dfrac{u_y}{U_0} =0,
\end{equation}
\begin{equation}
\dfrac{T - T_c}{T_h - T_c} = \dfrac{P\mspace{-1.5mu}r E\mspace{-1.5mu}c}{3} \left[ 1 - \left( 1 -
2\dfrac{y}{W} \right)^4 \right] + \dfrac{y}{W},
\end{equation}
\end{subequations}
where $W$ is the channel width, $U_0 = F_x W^2 / (8 \mu)$ is the maximum
velocity, and $E\mspace{-1.5mu}c = U_0^2 \big/ [C_p (T_h - T_c)]$ is the Eckert number. As
seen in Eq.\ (\ref{eq.Poiseuille.u.T}), the analytical solutions for
velocity and temperature are fully determined by $P\mspace{-1.5mu}r$ and $E\mspace{-1.5mu}c$. However,
the analytical solution for density further depends on both the initial
state and the adopted EOS. In the simulations, the density, velocity, and
temperature are initialized as $\rho_0^{}$, $\mathbf{0}$, and $T_0$ ($T_0 =
T_c$), respectively, and the initial pressure is determined by the adopted
EOS. Thus, for the decoupling EOS (i.e., Eq.\ (\ref{eq.EOSs.1})), the
analytical solution for density can be easily obtained as
\begin{subequations}\label{eq.rho.analytical}
\begin{equation}\label{eq.rho.analytical.1}
\dfrac{\rho}{\rho_0^{}} = 1;
\end{equation}
for the ideal-gas EOS (i.e., Eq.\ (\ref{eq.EOSs.2})), the analytical
solution for density is given as
\begin{equation}\label{eq.rho.analytical.2}
\dfrac{\rho}{\rho_0^{}} = A \dfrac{T_0}{T},
\end{equation}
where the coefficient $A^{-1} = \int\nolimits_0^W T_0 / T \text{d}y$; as for
the Carnahan-Starling EOS (i.e., Eq.\ (\ref{eq.EOSs.3})), the analytical
solution for density satisfies
\begin{equation}\label{eq.rho.analytical.3}
\int\limits_0^W \rho \text{d}y = \rho_0^{} W \quad \text{and} \quad
p_\infty^{} = Z \rho R_g T,
\end{equation}
\end{subequations}
where $p_\infty^{}$ is the final pressure in the channel. Although an
explicit expression for $\rho / \rho_0^{}$ cannot be derived from Eq.\
(\ref{eq.rho.analytical.3}), $\rho / \rho_0^{}$ can be easily obtained with high
precision using numerical integration.
\par
In the simulations, the lattice sound speed is set as
\begin{equation}\label{eq.cs}
c_s = \left. \sqrt{\partial_\rho p_\text{\tiny EOS}^{}} \right|
\lower 1.23ex \hbox{$\scriptstyle \rho = \rho_0^{}, T = T_0$},
\end{equation}
and the specific heats at constant pressure and volume are fixed at
\begin{equation}\label{eq.Cp}
C_p = 3.5 \dfrac{c_s^2}{T_0}, \quad C_v = C_p - \dfrac{p_\text{\tiny EOS}^{}
(\rho_0^{}, T_0)}{\rho_0^{} T_0}.
\end{equation}
With this configuration, the specific heat ratio $\gamma$ is $1.4$ for the
decoupling and ideal-gas EOSs and $1.101465$ for the Carnahan-Starling EOS.
The simulations are carried out on a $64 \times 64$ grid with lattice
spacing $\delta_x = 1/64$ and periodic boundary in $x \text{-direction}$.
The lower and upper walls are treated by the present boundary condition
treatment with second-order extrapolation. The basic parameters are set as
$R_g = 1$, $T_0 = 1$, and $\rho_0^{} = 1$, and the dimensionless relaxation time
for density DF, defined as $\tau = s_p^{-1}$, is fixed at $0.8$ for the
kinematic viscosity $\mu / \rho_0^{}$.
\par
Fig.\ \ref{fig.Poiseuille.u.T} shows the velocity $u_x / U_0$ and
temperature $(T - T_c) / (T_h - T_c)$ distributions across the channel and
compares the numerical results with the analytical solutions given by Eq.\
(\ref{eq.Poiseuille.u.T}). Two sets of $P\mspace{-1.5mu}r$ and $E\mspace{-1.5mu}c$ are considered here:
for the first set, $P\mspace{-1.5mu}r$ is fixed at $0.71$ and $E\mspace{-1.5mu}c$ varies from $0.1$ to
$100$; while for the second set, $E\mspace{-1.5mu}c$ is fixed at $10$ and $P\mspace{-1.5mu}r$ varies
from $0.1$ to $4$. As an important computational parameter, the lattice Mach
number, defined as $M\mspace{-1.5mu}a_\text{lattice} = U_0 / c_s$, is fixed at $0.2$ in
the simulations. Good agreement between the numerical results and the
analytical solutions can be observed in Fig.\ \ref{fig.Poiseuille.u.T},
which demonstrates that the effects of the viscous dissipation and
compression work are successfully captured by the present LB model. From
Fig.\ \ref{fig.Poiseuille.u.T}, we can also see that the distributions of
$u_x / U_0$ and $(T - T_c) / (T_h - T_c)$ obtained with different EOSs are
almost identical, which agrees with the aforementioned discussion. Note that
the simulation with $E\mspace{-1.5mu}c = 10$ and $P\mspace{-1.5mu}r = 0.1$ loses numerical stability
for the ideal-gas EOS. To further validate the present LB model with
self-tuning EOS, comparisons of the density $\rho / \rho_0^{}$ distributions are
carried out in Fig.\ \ref{fig.Poiseuille.rho}. Good agreement is observed
between the numerical results obtained with different EOSs and the
corresponding analytical solutions given by Eq.\ (\ref{eq.rho.analytical}),
which demonstrates that various EOSs (including the nonideal-gas EOS) can be
handled by the present LB model. For the decoupling EOS, $\rho / \rho_0^{}$
keeps constant across the channel; as for the ideal-gas and
Carnahan-Starling EOSs, $\rho / \rho_0^{}$ varies across the channel due to the
full coupling of thermo-hydrodynamic effects. In the Carnahan-Starling EOS,
the molecular volume is considered, which implies that the rigid-sphere
fluid is less compressible than the corresponding ideal gas. Therefore, the
variation in density across the channel obtained with the Carnahan-Starling
EOS is smaller than that obtained with the ideal-gas EOS, as clearly shown
in Fig.\ \ref{fig.Poiseuille.rho}.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig02.pdf}
\caption{Comparisons of velocity $u_x / U_0$ (left) and temperature
$(T - T_c) / (T_h - T_c)$ (right) distributions between the numerical
results and the analytical solutions for thermal Poiseuille flow with
(a) $P\mspace{-1.5mu}r = 0.71$ and $E\mspace{-1.5mu}c$ varying from $0.1$ to $100$, and
(b) $E\mspace{-1.5mu}c = 10$ and $P\mspace{-1.5mu}r$ varying from $0.1$ to $4$.}
\label{fig.Poiseuille.u.T}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig03.pdf}
\caption{Comparisons of density $\rho / \rho_0^{}$ distributions between
the numerical results obtained with different EOSs and the corresponding
analytical solutions for thermal Poiseuille flow with
(a) $P\mspace{-1.5mu}r = 0.71$ and $E\mspace{-1.5mu}c = 20$, and (b) $E\mspace{-1.5mu}c = 10$ and $P\mspace{-1.5mu}r=1$.}
\label{fig.Poiseuille.rho}
\end{figure}
Considering that the lattice Mach number plays an important role in the LB
method, we further investigate the accuracy of the present simulation with
respect to $M\mspace{-1.5mu}a_\text{lattice}$. It is worth pointing out that the lattice
Mach number is not only a computational parameter but also closely related
to the Eckert number (or the real Mach number) due to the lattice sound
speed given by Eq.\ (\ref{eq.cs}). In the following simulations, the Prandtl
number is fixed at $0.71$, and the Eckert number is set as $100 M\mspace{-1.5mu}a
_\text{lattice} ^2$ with $M\mspace{-1.5mu}a _\text{lattice}$ varying from $0.01$ to
$0.32$. Thus, the temperature difference $T_h - T_c$ remains unchanged for
different $M\mspace{-1.5mu}a _\text{lattice}$. The relative errors of velocity,
temperature, and density are calculated here, which are defined as
\begin{equation}\label{eq.Err}
E_\phi = \sqrt{ \dfrac{ \sum [f(\phi)_\text{numerical} -
f(\phi)_\text{analytical} ]^2}{ \sum f(\phi)_\text{analytical} ^2} },
\end{equation}
where $f(\phi)$ denotes the velocity $u_x / U_0$, temperature $(T - T_c) /
(T_h - T_c)$, and density $\rho / \rho_0^{}$ when $\phi = u$, $T$, and $\rho$,
respectively, the subscripts ``numerical'' and ``analytical'' denote the
numerical result and analytical solution of $f(\phi)$, respectively, and the
summation is over the computational domain. The relative errors $E_u$,
$E_T$, and $E_\rho$ versus $M\mspace{-1.5mu}a _\text{lattice}$ are shown in Fig.\
\ref{fig.Poiseuille.Ma}. As seen, the accuracy with respect to $M\mspace{-1.5mu}a
_\text{lattice}$ for velocity $u_x / U_0$ is fourth order when $M\mspace{-1.5mu}a
_\text{lattice}$ is relatively large and gradually decreases to second order
as $M\mspace{-1.5mu}a _\text{lattice}$ decreases, while the accuracy for temperature $(T -
T_c) / (T_h - T_c)$ and density $\rho / \rho_0^{}$ keep second order. Here, the
fourth-order accuracy for $u_x / U_0$ when $M\mspace{-1.5mu}a _\text{lattice}$ is
relatively large is due to the elimination of the additional cubic terms of
velocity, and the second-order accuracy may be caused by the boundary
condition treatment. Nevertheless, from Fig.\ \ref{fig.Poiseuille.Ma} we can
clearly see that satisfying results with different EOSs can be obtained by
the present LB model and boundary condition treatment under the low Mach
number condition.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig04.pdf}
\caption{Relative errors of (a) velocity $u_x / U_0$, (b) temperature
$(T - T_c) / (T_h - T_c)$, and (c) density $\rho / \rho_0^{}$ versus lattice
Mach number $M\mspace{-1.5mu}a_\text{lattice}$ when $P\mspace{-1.5mu}r = 0.71$ and
$E\mspace{-1.5mu}c = 100 M\mspace{-1.5mu}a _\text{lattice} ^2$. The symbols ``$\circ$'',
``$\vartriangle$'', and ``$\times$'' denote the results obtained with
decoupling, ideal-gas, and Carnahan-Starling EOSs, respectively,
and the solid, dashed, and dotted lines denote second-, third-, and
fourth-order accuracy with respect to $M\mspace{-1.5mu}a _\text{lattice}$, respectively.}
\label{fig.Poiseuille.Ma}
\end{figure}
\subsection{Thermal Couette flow}
The thermophysical properties (dynamic viscosity and thermal conductivity)
are assumed to be constant for the above thermal Poiseuille flow. To
validate that the present LB model is capable of handling the coupled
thermo-hydrodynamic flows with variable thermophysical properties, the
thermal Couette flow between two parallel walls is simulated in this
section. The lower wall is at rest and keeps adiabatic, and the upper wall
moves along $x$-direction with a constant velocity $U_0$ and keeps at a
constant temperature $T_0$. The Prandtl number $P\mspace{-1.5mu}r = C_p \mu / \lambda$
and the specific heat at constant pressure $C_p$ are assumed to be constant,
and thus $\lambda \propto \mu$. Considering a linear dependence of $\mu$ on
$T$ that is $\mu / \mu_0 = T / T_0$, the analytical solutions for velocity
and temperature are given as \cite{Liepmann1957}
\begin{subequations}\label{eq.Couette.u.T}
\begin{equation}
\dfrac{u_x}{U_0} + \dfrac{P\mspace{-1.5mu}r M\mspace{-1.5mu}a_e^2}{2} \left( \dfrac{u_x}{U_0} -
\dfrac13 \dfrac{u_x^3}{U_0^3} \right) = \left( 1 +
\dfrac{P\mspace{-1.5mu}r M\mspace{-1.5mu}a_e^2}{3} \right) \dfrac{y}{W}, \quad
\dfrac{u_y}{U_0} = 0,
\end{equation}
\begin{equation}
\dfrac{T}{T_0} = 1 + \dfrac{P\mspace{-1.5mu}r M\mspace{-1.5mu}a_e^2}{2} \left( 1 -
\dfrac{u_x^2}{U_0^2} \right),
\end{equation}
\end{subequations}
where $W$ is the channel width, and $M\mspace{-1.5mu}a_e = U_0 / \sqrt{C_p T_0}$ is an
equivalent Mach number different from but closely related to the lattice and
real Mach numbers. As it can be seen from Eq.\ (\ref{eq.Couette.u.T}), the
analytical solutions for $u_x / U_0$ and $T / T_0$ are fully determined by
$P\mspace{-1.5mu}r$ and $M\mspace{-1.5mu}a_e$. Similarly to the thermal Poiseuille flow, the analytical
solution for density $\rho / \rho_0^{}$ here is not only related to the initial
state, but it also depends on the adopted EOS. In the simulations, the
density, velocity, temperature, and pressure are initialized as $\rho_0^{}$,
$\mathbf{0}$, $T_0$, and $p_\text{\tiny EOS}^{} (\rho_0^{}, T_0)$, respectively. Thus, the
analytical solution for density $\rho / \rho_0^{}$ is also given by Eq.\
(\ref{eq.rho.analytical}), where the coefficient $A$ can be explicitly
written as $A = 1 + P\mspace{-1.5mu}r M\mspace{-1.5mu}a_e^2 \big/3$.
\par
In the simulations, all the simulation parameters are chosen the same as
those for the thermal Poiseuille flow, except that $\tau$ is fixed at $0.8$
for $\mu_0 / \rho_0^{}$. Since $\mu$ varies with $T$, $\tau$ also varies with
$T$ even for the decoupling EOS. Fig.\ \ref{fig.Couette.u.T} gives the
velocity $u_x / U_0$ and temperature $T / T_0$ distributions across the
channel for $P\mspace{-1.5mu}r = 0.71$ and $M\mspace{-1.5mu}a_e$ varying from $0.01$ to $0.15$, and for
$M\mspace{-1.5mu}a_e = 0.10$ and $P\mspace{-1.5mu}r$ varying from $0.1$ to $20$. Here, the simulation
with $M\mspace{-1.5mu}a_e = 0.10$ and $P\mspace{-1.5mu}r = 0.1$ also loses numerical stability for the
ideal-gas EOS. As seen in Fig.\ \ref{fig.Couette.u.T}, the numerical results
are in good agreement with the analytical solutions. Thus, the coupled
thermo-hydrodynamic flows with variable thermophysical properties can be
successfully handled by the present LB model. Fig.\ \ref{fig.Couette.u.T}
also verifies that the results ($u_x /U_0$ and $T / T_0$) obtained with
different EOSs are indistinguishable as long as the simulations are
numerically stable. Fig.\ \ref{fig.Couette.rho} compares the distributions
of density $\rho / \rho_0^{}$ obtained with different EOSs for $P\mspace{-1.5mu}r = 0.71$ and
$M\mspace{-1.5mu}a_e = 0.15$, and for $M\mspace{-1.5mu}a_e = 0.10$ and $P\mspace{-1.5mu}r = 20$. Good agreement
between the numerical results and the corresponding analytical solutions can
be observed, which reaffirms the applicability and accuracy of the present
LB model with self-tuning EOS.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig05.pdf}
\caption{Comparisons of velocity $u_x / U_0$ (left) and temperature
$T / T_0$ (right) distributions between the numerical results and the
analytical solutions for thermal Couette flow with (a) $P\mspace{-1.5mu}r = 0.71$ and
$M\mspace{-1.5mu}a_e$ varying from $0.01$ to $0.15$, and (b) $M\mspace{-1.5mu}a_e = 0.10$ and $P\mspace{-1.5mu}r$
varying from $0.1$ to $20$.}
\label{fig.Couette.u.T}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig06.pdf}
\caption{Comparisons of density $\rho / \rho_0^{}$ distributions between the
numerical results obtained with different EOSs and the corresponding
analytical solutions for thermal Couette flow with (a) $P\mspace{-1.5mu}r = 0.71$ and
$M\mspace{-1.5mu}a_e = 0.15$, and (b) $M\mspace{-1.5mu}a_e = 0.10$ and $P\mspace{-1.5mu}r = 20$.}
\label{fig.Couette.rho}
\end{figure}
The accuracy of the present simulation with respect to the lattice Mach
number $M\mspace{-1.5mu}a _\text{lattice} = U_0 / c_s$ is also investigated here.
Considering the specific heat at constant pressure given by Eq.\
(\ref{eq.Cp}), we have $M\mspace{-1.5mu}a_e = M\mspace{-1.5mu}a _\text{lattice} / \sqrt{3.5}$. In the
following simulations, the Prandtl number $P\mspace{-1.5mu}r$ is fixed at $0.71$, and the
lattice Mach number $M\mspace{-1.5mu}a _\text{lattice}$ varies from $0.01$ to $0.32$.
Fig.\ \ref{fig.Couette.Ma} shows the variations of the relative errors
$E_u$, $E_T$, and $E_\rho$ with $M\mspace{-1.5mu}a _\text{lattice}$. Here, the relative
error $E_\phi$ ($\phi = u$, $T$, and $\rho$) is also computed via Eq.\
(\ref{eq.Err}), in which $f(\phi)$ denotes $u_x / U_0$, $T / T_0$, and $\rho
/ \rho_0^{}$ when $\phi = u$, $T$, and $\rho$, respectively. It can be seen from
Fig.\ \ref{fig.Couette.Ma}(a) that the accuracy with respect to $M\mspace{-1.5mu}a
_\text{lattice}$ for velocity $u_x / U_0$ is fourth order and decreases to
second order when $M\mspace{-1.5mu}a _\text{lattice}$ and also $E_u$ are very small. A
similar trend can also be observed in Fig.\ \ref{fig.Couette.Ma}(b) for the
accuracy for temperature $T / T_0$. As to the accuracy for density $\rho /
\rho_0^{}$, it is fourth order and decreases rapidly when $E_\rho$ is rather
small for the decoupling EOS, while it is second order for the ideal-gas and
Carnahan-Starling EOSs. Here, the observed high-order accuracy with respect
to $M\mspace{-1.5mu}a _\text{lattice}$ can be explained by the elimination of the
additional cubic terms of velocity in the recovered momentum conservation
equation, and the deterioration of accuracy when $M\mspace{-1.5mu}a _\text{lattice}$ and
also $E_\phi$ ($\phi = u$, $T$, and $\rho$) are very small is probably
caused by the boundary condition treatment.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig07.pdf}
\caption{Relative errors of (a) velocity $u_x / U_0$, (b) temperature
$T / T_0$, and (c) density $\rho / \rho_0^{}$ versus lattice Mach number
$M\mspace{-1.5mu}a _\text{lattice}$ when $P\mspace{-1.5mu}r = 0.71$ and $M\mspace{-1.5mu}a_e = M\mspace{-1.5mu}a _\text{lattice}
/ \sqrt{3.5}$. The symbols ``$\circ$'', ``$\vartriangle$'', and
``$\times$'' denote the results obtained with decoupling, ideal-gas, and
Carnahan-Starling EOSs, respectively, and the solid, dashed, and dotted
lines denote second-, third-, and fourth-order accuracy with respect
to $M\mspace{-1.5mu}a _\text{lattice}$, respectively.}
\label{fig.Couette.Ma}
\end{figure}
\subsection{Natural convection in a square cavity}
To further validate the present LB model for coupled thermo-hydrodynamic
flows, the natural convection in a square cavity with a large temperature
difference is simulated in this section. All the four walls of the cavity
are at rest, among which the left (heating) and right (cooling) walls keep
at the temperature $T_h$ and $T_c$ ($T_h > T_c$), respectively, and the
horizontal walls keep adiabatic. The temperature difference between the
heating and cooling walls is quantified by a dimensionless parameter
$\varepsilon = (T_h - T_c) / (2T_0)$, where the reference temperature $T_0 =
(T_h + T_c) /2$. The ideal-gas EOS $p_\text{\tiny EOS}^{} = \rho R_g T$ is adopted here,
and thus $C_p - C_v = R_g$. The specific heat ratio $\gamma = C_p / C_v$ and
the Prandtl number $P\mspace{-1.5mu}r = C_p \mu / \lambda$ are assumed to be constant.
The dependence of dynamic viscosity on temperature is described by
Sutherland's law as follows \cite{Vierendeels2003}
\begin{equation}
\dfrac{\mu}{\mu^\ast} = \left( \dfrac{T}{T^\ast} \right) ^{3/2}
\dfrac{T^\ast + S}{T + S},
\end{equation}
where $T ^\ast = 273 \text{K}$, $S = 110.5 \text{K}$, and $\mu ^\ast$ is the
dynamic viscosity at $T ^\ast$. As a key dimensionless parameter associated
with natural convection, the Rayleigh number is defined as
\begin{equation}
R\mspace{-1.5mu}a = P\mspace{-1.5mu}r \dfrac{ |\mathbf{g}| \rho_0^2 (T_h - T_c) L^3}{T_0 \mu_0^2} ,
\end{equation}
where $\mathbf{g}$ is the gravity acceleration, $L$ is the side length of
the square cavity, and $\mu_0$ is the reference dynamic viscosity at $T_0$.
Initially, the ideal gas in the cavity stays still with temperature $T_0$
and density $\rho_0^{}$, and then the temperature of the left and right walls
are abruptly changed to $T_h$ and $T_c$, respectively. In the simulations,
the lattice sound speed is set as $c_s = \sqrt{R_g T_0}$, and the basic
parameters are chosen as $|\mathbf{g}| = 9.81 \text{m}/\text{s}^2$, $R_g =
287 \text{J} / (\text{kg} \cdot \text{K})$, $T_0 = 600 \text{K}$, and $\rho_0^{}
= p_0^{} / (R_g T_0)$ with $p_0^{} = 101325 \text{Pa}$. The Rayleigh number
$R\mspace{-1.5mu}a$ varies from $10^3$ up to $10^8$, while the remaining dimensionless
parameters are fixed at $\varepsilon = 0.6$, $\gamma = 1.4$, and $P\mspace{-1.5mu}r =
0.71$. The grid sizes $N_x \times N_y$ and the viscosity ratio $\varsigma /
\nu$ adopted for different $R\mspace{-1.5mu}a$ are listed in Table \ref{table.NxNy}, where
$\varsigma / \nu$ is set to $2$ and $4$ for $R\mspace{-1.5mu}a = 10^7$ and $10^8$,
respectively, to enhance the numerical stability and it is simply set to $1$
for $R\mspace{-1.5mu}a \leq 10^6$. As to the velocity and temperature boundary conditions
on all the four walls, they are realized by the present boundary condition
treatment with first-order extrapolation.
\par
\begin{table}[htbp]
\centering
\caption{Grid sizes $N_x \times N_y$ and viscosity ratio $\varsigma / \nu$
for different Rayleigh numbers $R\mspace{-1.5mu}a$.}
\label{table.NxNy}
\begin{tabular}{ccc ccc c}
\hline
$R\mspace{-1.5mu}a$ & $10^3$ & $10^4$ & $10^5$ & $10^6$ & $10^7$ & $10^8$ \\
\hline
$N_x \times N_y$ & $128 \times 128$ & $192 \times 192$ &
$256 \times 256$ & $512 \times 512$ & $1024 \times 1024$ &
$4096 \times 4096$ \\
$\varsigma / \nu$ & $1$ & $1$ & $1$ & $1$ & $2$ & $4$ \\
\hline
\end{tabular}
\end{table}
Fig.\ \ref{fig.Convection.fields} shows the streamlines, isotherms, and
density field for the natural convection when $R\mspace{-1.5mu}a$ varies from $10^3$ to
$10^8$. It can be seen from Fig.\ \ref{fig.Convection.fields} that a single
vortex with its center closer to the cooling wall appears in the cavity for
$R\mspace{-1.5mu}a = 10^3$. As $R\mspace{-1.5mu}a$ increases, the vortex is stretched by the natural
convection and breaks up into two vortices when $R\mspace{-1.5mu}a = 10^5$. As $R\mspace{-1.5mu}a$
further increases, the two vortices move closer to the heating and cooling
walls, respectively, and some small vortices are induced around the center
and in the lower-right and upper-left corners of the cavity when $R\mspace{-1.5mu}a =
10^7$. Meanwhile, a counter-rotating vortex also appears in the lower-right
corner and very close to the lower wall for $R\mspace{-1.5mu}a = 10^7$. When $R\mspace{-1.5mu}a$ reaches
$10^8$, the natural convection becomes unsteady, and many small vortices,
including some counter-rotating ones, are induced by the strong convection.
As to the heat transfer characteristics, it can be seen from Fig.\
\ref{fig.Convection.fields} that the isotherms are almost parallel to the
vertical walls when $R\mspace{-1.5mu}a = 10^3$, implying that the heat transfer is
dominated by conduction. As $R\mspace{-1.5mu}a$ increases, the isotherms around the cavity
center progressively incline and become parallel to the horizontal walls,
implying that the dominant mechanism for heat transfer changes from
conduction to convection. When $R\mspace{-1.5mu}a$ reaches $10^8$, the isotherms spread
along the heating and cooling walls in a very thin layer and become
horizontal almost in the entire cavity. All these observed streamline
patterns and isotherm characteristics are in good agreement with the
previous numerical results \cite{Li2012, Fei2018, Feng2018, Safari2018,
Vierendeels2003}, which are all obtained by the LB method except for the
benchmark solutions reported in Ref.\ \cite{Vierendeels2003}. Note that the
maximum $R\mspace{-1.5mu}a$ reported in Refs.\ \cite{Li2012} and \cite{Fei2018} are $10^5$
and $10^6$, respectively, and the maximum $R\mspace{-1.5mu}a$ reported in Refs.\
\cite{Feng2018, Safari2018, Vierendeels2003} are $10^7$. On the basis of the
present simulations, it is interesting to find that the natural convection
in a square cavity with a large temperature difference ($\varepsilon = 0.6$)
becomes unsteady when $R\mspace{-1.5mu}a = 10^8$, while the corresponding natural
convection with a small temperature difference (i.e., the Boussinesq
approximation is valid) keeps steady when $R\mspace{-1.5mu}a = 10^8$ and becomes unsteady
when $R\mspace{-1.5mu}a > 1.9 \times 10^8$ \cite{Paolucci1989, Mayne2001, Huang2014.mb}.
From Fig.\ \ref{fig.Convection.fields}, we can also see that the density
significantly varies over space, particularly in the vicinity of the cooling
wall, with its minimum and maximum values smaller and larger than $0.400
\text{kg} / \text{m}^3$ and $1.300 \text{kg} / \text{m}^3$, respectively.
Obviously, the Boussinesq approximation cannot be adopted here. In addition,
the density contours are similar to the isotherms to some extent, which
conforms to the low Mach number condition \cite{Paillere2000}. In fact, the
maximum Mach number is rather small for the natural convection simulated
here \cite{Vierendeels2003}.
\par
\begin{figure}[htbp]
\centering
\subfigure{\includegraphics[scale=1,draft=\figdraft]{Fig08a.pdf}}
\subfigure{\includegraphics[scale=1,draft=\figdraft]{Fig08b.pdf}}
\subfigure{\includegraphics[scale=1,draft=\figdraft]{Fig08c.pdf}}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure{\includegraphics[scale=1,draft=\figdraft]{Fig08d.pdf}}
\subfigure{\includegraphics[scale=1,draft=\figdraft]{Fig08e.pdf}}
\subfigure{\includegraphics[scale=1,draft=\figdraft]{Fig08f.pdf}}
\caption{Streamlines (left), isotherms (middle), and density field (right)
for natural convection in a square cavity with a large temperature
difference when $R\mspace{-1.5mu}a = 10^3$, $10^4$, $10^5$, $10^6$, $10^7$, and $10^8$.}
\label{fig.Convection.fields}
\end{figure}
To further validate the present results, the profiles of the horizontal
velocity along the vertical midplane and the vertical velocity along the
horizontal midplane are plotted in Fig.\ \ref{fig.Convection.profile} and
compared with the benchmark solutions obtained by Vierendeels et al.\
\cite{Vierendeels2003} using the finite difference (FD) method. Here, the
velocity and coordinate are normalized by the reference velocity
$U_\text{ref} = \sqrt{R\mspace{-1.5mu}a} \mspace{1.5mu} \mu_0 \big/ (\rho_0^{} L)$ and side
length $L$, respectively, i.e., $\mathbf{u} ^\ast = \mathbf{u} /
U_\text{ref}$ and $\mathbf{x} ^\ast = \mathbf{x} /L$. Note that the
convection becomes unsteady when $R\mspace{-1.5mu}a = 10^8$, and thus Fig.\
\ref{fig.Convection.profile}(f) shows the instantaneous profiles at some
time point. Excellent agreement between the present results and the
benchmark solutions can be observed. From Fig.\
\ref{fig.Convection.profile}, we can also see that the velocity profiles are
asymmetric with respect to the cavity center, which is caused by the
invalidation of the Boussinesq approximation. For quantitative comparison,
the average Nusselt number along the heating wall, the average pressure in
the cavity, and the maximum horizontal (vertical) velocity and its position
along the vertical (horizontal) midplane are computed and listed in Table
\ref{Table.Nuave}. Here, the average Nusselt number and pressure are defined
as
\begin{subequations}
\begin{equation}
N\mspace{-1.5mu}u_\text{ave} = \dfrac{1}{ \lambda_0 (T_h - T_c) } \int\limits_0^L
J_x (0, y) \text{d}y,
\end{equation}
\begin{equation}
p_\text{ave}^\ast = \dfrac{1}{L^2} \int\limits_0^L \int\limits_0^L
\dfrac{p_\text{\tiny EOS}^{} (x, y)}{p_0^{}} \text{d}x\text{d}y ,
\end{equation}
\end{subequations}
where $J_x (x,y)$ is the local heat flux in $x \text{-direction}$,
$\lambda_0^{}$ is the thermal conductivity at $T_0$, and the pressure is
normalized by $p_0^{}$. As seen in Table \ref{Table.Nuave}, the present
results agree well with the previous numerical results, which further
demonstrates the applicability and accuracy of the present LB model for
coupled thermo-hydrodynamic flows.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=1,draft=\figdraft]{Fig09.pdf}
\caption{Profiles of the horizontal velocity ($u_x^\ast$) along vertical
midplane ($y^\ast \text{-mid}$) and the vertical velocity ($u_y^\ast$) along
horizontal midplane ($x^\ast \text{-mid}$) for natural convection in a
square cavity with a large temperature difference when $R\mspace{-1.5mu}a = 10^3$, $10^4$,
$10^5$, $10^6$, $10^7$, and $10^8$. The solid lines are the present results
and the symbols are the benchmark solutions in
Ref.\ \cite{Vierendeels2003}.}
\label{fig.Convection.profile}
\end{figure}
\begin{table}[htbp]
\centering
\caption{Comparisons of the average Nusselt number along the heating wall
($N\mspace{-1.5mu}u_\text{ave}$), the average pressure in the cavity
($p_\text{ave}^\ast$), the maximum horizontal velocity
($|u_x^\ast|_\text{max}$) and its position ($y_\text{max}^\ast$)
along the vertical midplane, and the maximum vertical velocity
($|u_y^\ast|_\text{max}$) and its position ($x_\text{max}^\ast$)
along the horizontal midplane between the present and previous results.}
\label{Table.Nuave}
\newlength{\hj}\setlength{\hj}{1.5ex}
\setlength{\tabcolsep}{13.28pt} %
\begin{tabular}{ccc ccc cc}
\hline
$R\mspace{-1.5mu}a$ & Method & $\quad N\mspace{-1.5mu}u_\text{ave}$ & $\quad p_\text{ave}^\ast$
& $|u_x^\ast| _\text{max}$ & $y_\text{max}^\ast$
& $|u_y^\ast| _\text{max}$ & $x_\text{max}^\ast$ \\
\hline
$10^3$ & Present & $1.1063$
& $0.93443$ & $0.1653$ & $0.1641$ & $0.1911$ & $0.8984$ \\
~ & FD method \cite{Vierendeels2003} & $1.1077$
& $0.93805$ & $0.1649$ & $0.1618$ & $0.1926$ & $0.9036$ \\
~ & LB method \cite{Fei2018} & $1.106$
& -- & $0.1639$ & $0.1624$ & $0.1925$ & $0.9063$ \\
~ & LB method \cite{Li2012} & $1.111$
& -- & $0.1660$ & $0.1600$ & $0.1973$ & $0.9100$
\vspace{\hj} \\
$10^4$ & Present & $2.2123$
& $0.91144$ & $0.2360$ & $0.7813$ & $0.2857$ & $0.9271$ \\
~ & FD method \cite{Vierendeels2003} & $2.218$
& $0.91463$ & $0.2363$ & $0.7821$ & $0.2863$ & $0.9270$ \\
~ & LB method \cite{Fei2018} & $2.224$
& -- & $0.2372$ & $0.7813$ & $0.2859$ & $0.9312$ \\
~ & LB method \cite{Li2012} & $2.217$
& -- & $0.2364$ & $0.7800$ & $0.2874$ & $0.9267$
\vspace{\hj} \\
$10^5$ & Present & $4.4836$
& $0.91719$ & $0.1950$ & $0.8359$ & $0.3130$ & $0.0977$ \\
~ & FD method \cite{Vierendeels2003} & $4.480$
& $0.92196$ & $0.1946$ & $0.8364$ & $0.3166$ & $0.0948$ \\
~ & LB method \cite{Fei2018} & $4.512$
& -- & $0.1951$ & $0.8344$ & $0.3176$ & $0.0938$ \\
~ & LB method \cite{Li2012} & $4.454$
& -- & $0.1959$ & $0.8360$ & $0.3165$ & $0.0960$
\vspace{\hj} \\
$10^6$ & Present & $8.7406$
& $0.91820$ & $0.1193$ & $0.8516$ & $0.3141$ & $0.0547$ \\
~ & FD method \cite{Vierendeels2003} & $8.687$
& $0.92449$ & $0.1193$ & $0.8541$ & $0.3203$ & $0.0537$ \\
~ & LB method \cite{Fei2018} & $8.691$
& -- & $0.1202$ & $0.8551$ & $0.3159$ & $0.0540$
\vspace{\hj} \\
$10^7$ & Present & $16.4373$
& $0.91425$ & $0.0745$ & $0.8262$ & $0.3124$ & $0.0313$ \\
~ & FD method \cite{Vierendeels2003} & $16.240$
& $0.92263$ & $0.0749$ & $0.8260$ & $0.3229$ & $0.0305$
\vspace{\hj} \\
$10^8$ & Present & $29.9435$
& $0.91609$ & $0.0586$ & $0.9004$ & $0.3169$ & $0.0176$ \\
\hline
\end{tabular}
\end{table}
\section{Conclusions} \label{sec.conclusion}
A novel LB model for coupled thermo-hydrodynamic flows is developed in the
framework of the DDF approach. The velocity field is solved by the recently
developed LB equation for density DF, by which the recovered EOS can be
self-tuned via a built-in variable, implying that various EOSs can be
adopted in real applications. With the energy conservation equation properly
reformulated, a novel LB equation for total energy DF is directly developed
at the discrete level to solve the temperature field. The viscous
dissipation is recovered along with the conduction term by introducing a
density-DF-related term into this LB equation, while the compression work is
recovered along with the convection term by devising the equilibrium moment
function for total energy DF. The work done by force is absorbed into the
source term and then correctly incorporated into the LB equation via the
discrete source term. Moreover, by modifying the collision matrix, the
targeted energy conservation equation can be recovered without deviation
term. The development of the present LB model, with double MRT collision
schemes employed, is based on the standard lattice, and both the Prandtl
number and specific heat ratio can be arbitrarily adjusted. On the basis of
judiciously decomposing DF into its equilibrium, force (source), and
nonequilibrium parts, boundary condition treatment is further proposed for
simulating coupled thermo-hydrodynamic flows, which can ensure the local
conservation of mass, momentum, and energy at the boundary node. The
applicability and accuracy of the present LB model with self-tuning EOS are
first validated by simulating thermal Poiseuille and Couette flows with the
decoupling, ideal-gas, and Carnahan-Starling EOSs. Then, the present LB
model is successfully applied to the simulation of natural convection in a
square cavity with a large temperature difference for the Rayleigh number
ranging from $10^3$ up to $10^8$, and the obtained results agree very well
with the previous benchmark solutions.
\section*{Acknowledgements}
R.H.\ acknowledges the support by the Alexander von Humboldt Foundation,
Germany. This work was also supported by the National Natural Science
Foundation of China through Grant No.\ 51536005.
|
\section{Introduction}
In this note, we describe a new approach to understanding concentration properties of high energy eigenfunctions. Although the methods in~\cite{GT,Gdefect,CGT,CG17,CG18} (on which this note is based) apply to the quasimodes of a wide variety of pseudodifferential operators, we focus on the case of the Laplacian on a Riemannian manifold $(M,g)$ of dimension $n$ and consider only eigenfunctions i.e. solutions to
\begin{equation}
\label{e:laplace}
(-h^2\Delta_g-1)u_h=0
\end{equation}
for concreteness. Consider a submanifold $H\subset M$. We are interested in averages of the form
$$
\int_H u_hd\sigma_H
$$
where $\sigma_H$ denotes the volume measure induced on $H$ from $M$. We note that, when $H=\{x\}$ is a point in $M$, this average is given precisely by $u_h(x)$. Thus, using our methods we are able to obtain control on $L^\infty$ norms. We do not give the details of many proofs in this note, instead referring to the relevant papers. We review some of the previously existing results, state the new theorems, and describe the ideas central to the proofs.
Since the middle of the twentieth century~\cite{Ava,Lev,Ho68} many authors have been interested in the growth of eigenfunctions for self-adjoint elliptic operators. In particular, they prove that a solution to~\eqref{e:laplace} satisfies,
\begin{equation}
\label{e:Linf}
\|u_h\|_{L^\infty(M)}\leq Ch^{\frac{1-n}{2}}\|u_h\|_{L^2(M)}.
\end{equation}
If one considers the case of $(M,g)=(S^2,g_{\text{round}})$, the sphere with the round metric, it is not hard to construct the family of zonal harmonics, $Z_{h}$, with the property that
$$
ch^{-\frac{1}{2}}\|Z_{h}\|_{L^2(S^2)}\leq \|Z_{h}\|_{L^\infty(S^2)}\leq Ch^{-\frac{1}{2}}\|Z_{h}\|_{L^2(S^2)},\qquad (-h^2\Delta_{S^2}-1)Z_h=0,
$$
and hence that the estimate~\eqref{e:Linf} cannot be improved on a general manifold. Because of this, it is natural to try to understand situations in which~\eqref{e:laplace} is sharp.
It is also interesting to think of the question of $L^\infty$ norms as averages over points and to generalize that question to averages over submanifolds $H\subset M$. While it is a more recent line of inquiry than that of $L^\infty$ bounds, it dates at least to the early 1980's~\cite{Hej,Good}. The analog of~\eqref{e:laplace} was proved in~\cite{Zel}, where the author shows that if $H$ has codimension $k$, then
\begin{equation}
\label{e:average}
\Big|\int_H u_hd\sigma_H\Big|\leq Ch^{\frac{1-k}{2}}\|u_h\|_{L^2(M)}.
\end{equation}
Again, for every $1\leq k \leq n$, there are examples on the sphere of dimension $n$ which saturate the estimate~\eqref{e:average} and it is natural to characterize situations in which~\eqref{e:average} can be saturated.
\subsection{A review of previous $L^\infty$ results}
Before we can state the results on a general manifold $M$, we need some concepts from geometry. Let $T^*\!M$ denote the cotangent bundle to $M$, $H\subset M$ a submanifold with conormal bundle $N^*\!H$, and $S\!N^*\!H$, the unit conormal bundle to $H$,
$$
S\!N^*\!H:=\big\{(x,\xi)\in N^*\!H\mid |\xi|_{g(x)}=1\big\},
$$
where $|\cdot |_{g}$ denotes the metric induced on $T^*\!M$ by $g$. Note that $S\!N^*\!\{x\}=S^*_xM$ where $S^*_xM:=T^*_xM\cap S^*\!M$. Next, let $G^t:S^*\!M\to S^*\!M$ denote the geodesic flow.
We define the \emph{first return time} $T_H:S\!N^*\!H\to [0,\infty]$ by
$$
T_H(x,\xi):=\inf\{t>0\mid G^t(x,\xi)\in S\!N^*\!H\}.
$$
We then define the \emph{loop set of $H$}, $\mc{L}_H:=T_H^{-1}([0,\infty))$. Finally, we let $\sigma_{_{\!\!S\!N^*\!H}}$ be the volume induced on $S\!N^*\!H$ by the Liouville measure on $T^*\!M$.
\begin{theorem}[\cite{SoggeZelditch}]
\label{t:loops}
Let $(M,g)$ be a Riemannian manifold of dimension $n$. Suppose that $x\in M$ and $\sigma_{_{\!S^*_xM}}(\mc{L}_x)=0$. Then for $u_h$ solving~\eqref{e:laplace}
$$
|u_h(x)|=o\big(h^{\frac{1-n}{2}}\|u_h\|_{L^2(M)}\big).
$$
\end{theorem}
We define the first return map $\eta_H:\mc{L}_H\to S\!N^*\!H$ by
$$
\eta_H(x,\xi):=G^{T_H(x,\xi)}(x,\xi).
$$
and let
$$
\mc{L}_H^{\pm n}:=\bigcap_{k=0}^n\eta_H^{\pm k}(S\!N^*\!H),\qquad \mc{L}_H^\infty:=\bigcap_n\mc{L}_H^n.
$$
Next, define the \emph{recurrent set of $H$}
$$
\mc{R}_H:=\Big\{(x,\xi) \in \mc{L}_H^\infty \mid (x,\xi) \in \Big[\bigcap_{n>0}\overline{\bigcup_{k\geq n}\eta_H^k(x,\xi)}\,\Big]\bigcap \Big[ \bigcap_{n>0}\overline{\bigcup_{k\geq n}\eta_H^{-k}(x,\xi)}\,\Big]\,\Big\}.
$$
In~\cite{SoggeTothZelditch}, the authors show that $\mc{L}_x$ can be replaced by $\mc{R}_x$ in Theorem~\ref{t:loops}. Finally, in~\cite{SZ16I, SZ16II}, the authors obtain still more restrictive assumptions in the case that $(M,g)$ is real analytic. In fact, in the case of a real analytic surface, they are able to verify the conjecture that one can replace $\mc{R}_x$ by the set of directions $\xi$ so that the geodesic through $(x,\xi)$ is a smoothly closed loop.
If one wants to go beyond $o(1)$ improvements of~\eqref{e:Linf}, very few results are available. In~\cite{Berard77} (combined with~\cite{Bo16}), the author shows using the Hadamard parametrix that if $(M,g)$ is a manifold without conjugate points, then
\begin{equation}
\label{e:LinfLog}
\|u_h\|_{L^\infty(M)}\leq C\frac{h^{\frac{1-n}{2}}}{\sqrt{\log h^{-1}}}\|u_h\|_{L^2(M)}.
\end{equation}
The only polynomial improvements that the author is aware of appear in~\cite{I-s} where the authors study Hecke--Maas forms on certain arithmetic surfaces.
\subsection{A review of previous results on averages}
The study of when~\eqref{e:average} is saturated is much more recent and, until the methods of this note were introduced, the only improvements on~\eqref{e:average} available under no additional assumptions on $(M,g)$ are:
\begin{theorem}[\cite{Wym3}]
\label{t:loopAverage}
Suppose that $\sigma_{_{\!\!S\!N^*\!H}}(\mc{L}_H)=0$. Then
$$
\Big|\int_H u_h d\sigma_H\Big|= o\big(h^{\frac{1-k}{2}}\|u_h\|_{L^2(M)}\big)\
$$
\end{theorem}
The article~\cite{CS} provides $o(1)$ improvements on~\eqref{e:average} on surfaces of negative curvature when $H$ is a geodesic.
On the other hand, on manifolds with non-positive curvature the Hadamard parametrix is available and as a result logarithmic improvements of the form
\begin{equation}
\label{e:averageLog}
\Big|\int_Hu_hd\sigma_H\Big|\leq C\frac{h^{\frac{1-k}{2}}}{\sqrt{\log h^{-1}}}\|u_h\|_{L^2(M)}
\end{equation}
hold under a variety of assumptions on the pair $(M,H)$~\cite{SXZ,Wym2,Wym18}. However, none of these results give general dynamical conditions guaranteeing such improvements.
\subsection{Results of the microlocal techniques}
The question raised in all previous attempts to understand when~\eqref{e:Linf} and~\eqref{e:average} can be saturated can be thought of as `In which geometries can saturation occur?'. The question raised in~\cite{GT, CGT,Gdefect,CG17,CG18} is instead `How does an eigenfunction that saturates~\eqref{e:Linf} or~\eqref{e:average} behave?' It then turns out that a sufficiently good understanding of the answer to the latter question yields answers to the former. In fact, by describing the behavior of eigenfunctions saturating~\eqref{e:Linf}, we will be able to extend \emph{all} existing results. Moreover, our analysis of the eigenfunctions saturating these bounds demonstrates that the phenomena governing averages is identical to that governing $L^\infty$ bounds.
\subsubsection{Defect measures}
We begin by describing some of the results of~\cite{CG17} where we rely on defect measures to describe the behavior of $u_h$. Recall that a defect measure is a positive Radon measure, $\mu$, on $T^*\!M$ associated to a sequence of functions $\{u_h\}_{0<h<h_0}$ so that for any $a\in C_c^\infty(T^*M)$,
$$
\langle Op_h(a)u_h,u_h\rangle_{L^2(M)}\to \int a(x,\xi)d\mu
$$
where $Op_h(a)$ denotes the quantization of the symbol $a$ (see e.g. \cite[Appendix E]{ZwScat} for a description of quantization procedures).
See, for example,~\cite[Chapter 5]{EZB} for a treatment of these measures. We recall that every $L^2$ bounded sequence of functions $\{u_h\}$ has a subsequence with a defect measure, $\mu$ and, moreover if $u_h$ solves~\eqref{e:laplace}, then $\text{\ensuremath{\supp}} \mu\subset S^*\!M$ and $\mu$ is $G^t$ invariant.
Suppose that $\mu$ is a finite radon measure invariant under the geodesic flow. Then we define for any Borel $A\subset S\!N^*\!H$,
$$
\mu_H(A):=\lim_{\delta\to 0}\frac{1}{2\delta}\mu\Big(\bigcup_{|t|\leq \delta}G^t(A)\Big).
$$
We write $\omega\perp \nu$ when the measures $\omega$ and $\nu$ are mutually singular.
We then have the following consequence of~\cite[Theorem 6]{CG17} (see also~\cite[Theorem 2]{Gdefect} for the case $k=n$).
\begin{theorem}
\label{t:micro1}
Let $(M,g)$ be a smooth compact Riemannian manifold and $H\subset M$ a closed embedded submanifold of codimension $k$. Suppose that $u_h$ solves~\eqref{e:laplace} and has defect measure $\mu$. Let $f\in L^1(S\!N^*\!H;\sigma_{_{\!\!S\!N^*\!H}})$ so that
$$
\mu_H=fd\sigma_{_{\!\!S\!N^*\!H}}+\omega,\qquad \omega \perp \sigma_{_{\!\!S\!N^*\!H}}.
$$
Then there is $C_{n,k}>0$ depending only on $(n,k)$ so that for $A\subset H$ with smooth boundary,
$$
\Big|\int_A u_hd\sigma_H\Big|\leq C_{n,k}h^{\frac{1-k}{2}}\int_{\pi_H^{-1}(A)} \sqrt{f}d\sigma_{_{\!\!S\!N^*\!H}} +o(h^{\frac{1-k}{2}})
$$
where $\pi_H:S\!N^*\!H\to H$ is the natural projection.
\end{theorem}
Note that Theorem~\ref{t:micro1} can be interpreted as saying that \emph{every} eigenfunction which maximizes either~\eqref{e:Linf} or~\eqref{e:average} must have a component which behaves $o(1)$ microlocally the same as the canonical example on $S^n$. In particular, in order that $u_h$ maximize the $L^\infty$ bounds, there must be a point where $u_h$ behaves like the zonal harmonic, $Z_h$ (See e.g.~\cite[Section 4]{GT} for a description of the defect measures of zonal harmonics.)
As an easy consequence of Theorem~\ref{t:micro1} together with the Poincar\'e recurrence theorem we are able to replace $\mc{L}_H$ in Theorem~\ref{t:loopAverage} by $\mc{R}_H$.
\begin{corollary}[{{\cite[Theorem 2]{CG17}}}]
\label{c:norecur}
Suppose that $A\subset H$ has smooth boundary, $\sigma_{_{\!\!S\!N^*\!H}}(\pi_H^{-1}(A)\cap \mc{R}_H)=0$, and $u_h$ solves~\eqref{e:laplace} then
$$
\Big|\int_A u_hd\sigma_H\Big|=o\big(h^{\frac{1-k}{2}}\|u_h\|_{L^2(M)}\big).
$$
\end{corollary}
Using geometric arguments to show that $\sigma_{_{\!\!S\!N^*\!H}}(\mc{R}_H)=0$ in a variety of settings, we are then able to recover all existing $o(1)$ improvements over~\eqref{e:average} in~\cite[Theorem 4]{CG17}.
\begin{theorem}[{{\cite[Theorem 4]{CG17},\cite[Theorem 3]{CG18}}}]\label{t:app1}
In all of the following situations, $\sigma_{_{\!\!S\!N^*\!H}}(\mc{R}_H)=0$.
\begin{enumerate}[label=\textbf{\Alph*.},ref=\Alph*]
\item \label{a1} $(M,g)$ has no conjugate points and $H$ has codimension $k>\frac{n +1}{2}$. \smallskip
\item \label{a2}$(M,g)$ has no conjugate points and $H$ is a geodesic sphere.\smallskip
\item \label{a4}$(M,g)$ is a surface with Anosov geodesic flow and $H$ is any submanifold. \smallskip
\item \label{a3} $(M,g)$ has constant negative curvature and $H$ is any submanifold.\smallskip
\item \label{a6}$(M,g)$ has Anosov flow, non-positive curvature, and $k>1$. \smallskip
\item \label{a5}$(M,g)$ has Anosov geodesic flow and {non-positive curvature}, and $H$ is totally geodesic. \smallskip
\item \label{a7} $(M,g)$ has {Anosov geodesic flow} and $H$ is a subset $M$ that lifts to a horosphere.
\end{enumerate}
\end{theorem}
\smallskip
\subsubsection{Towards quantitative estimates}
In order to pass to the quantitative estimates from~\cite{CG18}, we will first describe some easy consequences of Theorem~\ref{t:micro1}. We say that $A\subset T^*\!M$ is \emph{$[t,T]$ non-self looping} if either
\begin{equation}
\label{e:nsl}
\begin{gathered}
G^s(A)\cap A=\emptyset,\qquad s\in[t,T],\qquad
\text{or}\qquad
G^{-s}(A)\cap A=\emptyset,\qquad s\in[t,T]\\
\end{gathered}
\end{equation}
We have the following Corollary of Theorem~\ref{t:micro1}.
\begin{corollary}
\label{c:nonLoop}
Suppose that there is a an $h$-independent covering $\big\{B,\{G_\ell\}_\ell\big\}$ of $S\!N^*\!H$ and $\{t_\ell\}_{\ell}$, $\{T_\ell\}_\ell\subset (1,\infty)$ with $t_\ell<T_\ell$ independent of $h$ so that
$$
S\!N^*\!H= B\,\cup \,\bigcup_{\ell} G_\ell.
$$
and $G_\ell$ is $[t_\ell,T_\ell]$ non-self looping an that $u_h$ solves~\eqref{e:laplace}. Then, there is $C>0$ so that for all $u_h$ solving~\eqref{e:laplace}
\begin{equation}
\label{e:prelimEst}
\Big|\int _Hu_hd\sigma_H\Big|\leq Ch^{\frac{1-k}{2}}\left(\sigma_{_{\!\!S\!N^*\!H}}(B)^{\frac{1}{2}}+\sum_\ell \frac{\sigma_{_{\!\!S\!N^*\!H}}(G_\ell)^{\frac{1}{2}}t_\ell^\frac{1}{2}}{T_\ell^{\frac{1}{2}}}+o(1)\right)\|u_h\|_{L^2(M)}
\end{equation}
\end{corollary}
In fact, Corollary~\ref{c:norecur} can be deduced from Corollary~\ref{c:nonLoop}. To see this, let $\{U_i\}_{i=1}^\infty$ be a basis for the topology of $S\!N^*\!H$. Then let $T>0$ and set
$$
E^{\pm,T}_i:=\{x\in U_i\mid G^t(x)\notin U_i, \pm t>T\},\qquad E_i^T=\bigcup_{\pm}E^{\pm,T}_i,\qquad E^\infty_i=\bigcup_{T>0} E_i^T.
$$
Let
$$
B_{N}=\Big[\bigcap_i \big(S\!N^*\!H\setminus E^T_i\big)\Big] \cup \Big(\bigcup_{j=N}^\infty E^T_j\setminus \bigcup_{k=1}^{N-1}E_k^T\Big),\qquad \qquad G_i=E^T_i.
$$
Then, since $G_i$ is $[T,S]$ non-self looping for any $S>T$, we apply~\eqref{e:prelimEst} to obtain
$$
\limsup_{h\to 0}\frac{h^{\frac{k-1}{2}}}{\|u_h\|_{L^2(M)}}\Big|\int _Hu_hd\sigma_H\Big|\leq C\Big(\sigma_{_{\!\!S\!N^*\!H}}(B_N)^{\frac{1}{2}}+\sum_{i=1}^{N-1}\frac{\sigma_{_{\!\!S\!N^*\!H}}(E^T_i)^{1/2}T^{\frac{1}{2}}}{S^{\frac{1}{2}}}\Big)
$$
Sending $S\to \infty$ gives
$$
\limsup_{h\to 0}\frac{h^{\frac{k-1}{2}}}{\|u_h\|_{L^2(M)}}\Big|\int _Hu_hd\sigma_H\Big|\leq C\sigma_{_{\!\!S\!N^*\!H}}(B_N)^{\frac{1}{2}}
$$
Sending $N\to \infty$ then gives
$$
\limsup_{h\to 0}\frac{h^{\frac{k-1}{2}}}{\|u_h\|_{L^2(M)}}\Big|\int _Hu_hd\sigma_H\Big|\leq C\sigma_{_{\!\!S\!N^*\!H}}\Big(\bigcap_i \big(S\!N^*\!H\setminus E^T_i\big)\Big)^{\frac{1}{2}}
$$
Finally, sending $T\to \infty$ gives
$$
\limsup_{h\to 0}\frac{h^{\frac{k-1}{2}}}{\|u_h\|_{L^2(M)}}\Big|\int _Hu_hd\sigma_H\Big|\leq C\sigma_{_{\!\!S\!N^*\!H}}\Big(\bigcap_i \big(S\!N^*\!H\setminus E^\infty_i\big)\Big)^{\frac{1}{2}}
$$
Now, suppose $x$ is not recurrent. Then, there exists $i, T$ so that $x\in U_i$, $G^t(x)\notin U_i$ either for $t>T$ or $-t>T$. In particular, $x\in E_i^T\subset E_i^\infty$. Therefore, if $x$ is not recurrent, then $x\in \cup_iE_i^\infty$. In particular,
$$
S\!N^*\!H\setminus \mc{R}_H\,\subset\, \bigcup_i E_i^\infty,\qquad \text{ so }\qquad \bigcap_i \big(S\!N^*\!H\setminus E_i^\infty\big)\subset \mc{R}_H.
$$
Therefore, if $\mc{R}_H$ has measure 0, then $\bigcap_i\big(S\!N^*\!H \setminus E_i^\infty\big) $ has measure 0 and we have obtained Corollary~\ref{c:norecur}.
The fact that $\rho\in \mc{R}_H$ does not contain \emph{any} quantitative information about how long it takes for the geodesic through $\rho$ to return to a given neighborhood of $\rho$. Because of this, one should not expect to have a quantitative version of Corollary~\ref{c:norecur}. However, Corollary~\ref{c:nonLoop} \emph{is} quantitative and one might hope that it holds even with $B$, $G_\ell$, and $[t_\ell,T_\ell]$ $h$-dependent. This is almost true, although we will require some additional structure of the sets $B$ and $G_\ell$ (see Theorem~\ref{t:nonLoop2}).
\subsubsection{Quantitative Estimates}
In order to state our quantitative estimates, we will need to define a few additional objects. We will use the metric induced by the Sasaki metric on $T^*\!M$ (see e.g. \cite{Eberlein73} for a description of the Sasaki metric) for convenience, but our results do not depend on the choice of metric on $T^*\!M$. First, fix $\mc{H}\subset T^*\!M$ a smooth hypersurface transverse to the geodesic flow so that $S\!N^*\!H\subset \mc{H}$.
Define $\psi:\mathbb{R}\times \mc{H}_{\Sigma}\to T^*\!M$ by $\psi(t,q)=\varphi_t(q)$. Next, let
$$
\tau_{_{\!\text{inj}H}}:=\sup\{\tau\leq 1: \psi|_{(-\tau,\tau)\times\mc{H}_{\Sigma}}\text{ is injective}\}.
$$
Given $A\subset T^*\!M$, define
$$
\Lambda_A^\tau:=\bigcup_{|t|\leq \tau}G^t(A).
$$
Then, for $r>0$ and $A\subset \mc{H}$, define
$$
\Lambda_A^\tau(r):=\Lambda_{A_r}^{\tau+r},\qquad A_r:=\{ \rho\in \mc{H}\mid d(\rho,A)<r\}.
$$
Finally, let $K_H>0$ be a bound for the the sectional curvatures of $H$ and for the second fundamental form of $H$.
\begin{theorem}
\label{t:micro2}
Let $H\subset M$ be a closed embedded submanifold of codimension $k$. There exist $C_{n,k}>0$ depending only on $n,k$, $\tau_0>0$ depending on $(M,g,\tau_{_{\!\text{inj}H}})$, and $R_0=R_0(n,k,K_H)$ so that the following holds.
Let $0<\tau<\tau_0$, $0\leq \delta<\frac{1}{2}$, $N>0$, and $R_0>R(h)\geq 5h^\delta$.
Then, there exists a family $\{\gamma_j\}_{j=1}^{N_h}$ of geodesics through $S\!N^*\!H$, and a partition of unity $\{\chi_j\}_{j=1}^{N_h}$ for $\Lambda^\tau_{_{\!\Sigma_{H\!,p}}}(h^\delta)$ with
$\chi_j\in S_\delta\cap C^\infty_c(T^*M;[0,1])$,
\[
\text{\ensuremath{\supp}} \chi_j\subset \Lambda_{\rho_j}^\tau(R(h)), \qquad \qquad \rho_j:=\gamma_j \cap S\!N^*\!H,
\]
so that for all $w\in C_c^\infty(H)$, $N>0$ there is $C_N>0$ and $h_0>0$ with the property that for any $0<h<h_0$ and all $u_h$ solving~\eqref{e:laplace}
\begin{align*}
h^{\frac{k-1}{2}}\Big|\int_Hwu_hd\sigma_H\Big|
&\leq C_{n,k} {R(h)^{\frac{n-1}{2}}\sum_{j}\frac{\|Op_h(\chi_j)u_h\|_{L^2(M)}}{\tau^{\frac{1}{2}}}}+C_Nh^N\|u_h\|_{L^2(M)}.
\end{align*}
\end{theorem}
Theorem~\ref{t:micro2} is a much finer analog of Theorem~\ref{t:micro1} and in particular can be interpreted as saying that \emph{every} eigenfunction which maximizes either~\eqref{e:Linf} or~\eqref{e:average} must have a component which behaves the same as the canonical example on $S^n$ microlocally on $h^\delta$ scales. In particular, in order that $u_h$ maximize the $L^\infty$ bounds, there must be a point where $u_h$ behaves like a zonal harmonic at scale $h^\delta$.
While at first it may seem difficult to use Theorem~\ref{t:micro2} in concrete situations, combining Theorem~\ref{t:micro2} with Egorov's theorem up to the Ehrenfest time (see e.g. \cite{DyGu14}) we obtain a purely dynamical estimate which is readily applicable.
We define the \emph{maximal expansion rate }
$$
\Lambda_{\max}:=\limsup_{|t|\to \infty}\frac{1}{|t|}{\log} \sup_{S^*\!M}\|dG^t(x,\xi)\|.
$$
Then the Ehrenfest time at frequency $h$ is
$$
T_e(h):=\frac{\log h^{-1}}{2\Lambda_{\max}}.
$$
Note that $\Lambda_{\max}\in[0,\infty)$ and if $\Lambda_{\max}=0$, we may replace it by an arbitrarily small positive constant. We have the following quantitative version of Corollary~\ref{c:nonLoop}.
\begin{theorem}[{{\cite[Theorem 5]{CG18}}}]
\label{t:nonLoop2}
Suppose that $H\subset M$ is a closed embedded submanifold of codimension $k$. Let $0<\delta<\frac{1}{2}$, $N>0$. There exist positive constants $h_0=h_0(M,g,K_H)$, $\tau_0=\tau_0(M,g,\tau_{_{\!\text{inj}H}})$, $R_0=R_0(n,k,K_H)$ and $C_{n,k}$ depending only on $n$ and $k$, and for each $0<\tau<\tau_0$ there exists and $C_N=C_N(\tau,\delta,M,g)>0$, so that the following holds.
Let $R_0>R(h)\geq 5h^\delta$,{ $\alpha< 1-2{\limsup_{h\to 0}\frac{\log R(h)}{\log h}}$,} and suppose $\{\Lambda_{_{\rho_j}}^\tau(R(h))\}_{j=1}^{N_h}$ is a cover of $\Lambda^\tau_{_{\!\Sigma_{H\!,p}}}(h^\delta)$ that is the union of $C_{n,k}$ subsets of disjoint tubes (the existence is guaranteed by \cite[Lemma 2.2]{CG18}). In addition, suppose there exist $\mc{B}\subset \{1,\dots N_h\}$ and a finite collection $\{\mc{G}_\ell\}_{\ell \in L} \subset \{1,\dots N_h\}$ with
$$
\{1,\dots N(R(h))\}\;\subset\; \mc{B} \cup \bigcup_{\ell \in L}\mc{G}_\ell,
$$
and so that for every $\ell \in L$ there exist $t_\ell(h)>0$ and ${T_\ell(h)}\leq 2 \alpha T_e(h)$ so that
$$
\bigcup_{j\in \mc{G}_\ell}\Lambda_{_{\rho_j}}^\tau(R(h))\;\;\text{ is }\;[t_\ell(h),T_{\ell}(h)]\text{ non-self looping}.
$$
Then, for all $w\in C_c^\infty(H)$, $N>0$ there exists $C_N>0$, $h_0>0$ so that for all for $u_h$ solving~\eqref{e:laplace} and $0<h<h_0$,
\begin{multline}
\label{e:estTube}
h^{\frac{k-1}{2}}\Big|\int_H w u_hd\sigma_H\Big|\leq \frac{C_{n,k}\|w\|_{L^\infty}R(h)^{\frac{n-1}{2}}}{\tau^{\frac{1}{2}}}
\!\left[|\mc{B}|^{\frac{1}{2}}+\sum_{\ell \in L }\frac{|\mc{G}_\ell|^{\frac{1}{2}}t_\ell^{\frac{1}{2}}(h)}{T^{\frac{1}{2}}_\ell(h)}+C_Nh^N\right]\!\!\|u_h\|_{L^2(M)}.
\end{multline}
\end{theorem}
Note that the term
$$
R(h)^{n-1}|\mc{G}_\ell|\propto \sigma_{_{\!\!S\!N^*\!H}}\Big(\bigcup_{j\in \mc{G}_\ell}\Lambda_{_{\rho_j}}^\tau(R(h))\cap S\!N^*\!H\Big).
$$
and in particular when $T_{\ell},\,R(h)$ are $h$ independent~\eqref{e:estTube} implies~\eqref{e:prelimEst}. Since~\eqref{e:prelimEst} implies Corollary~\ref{c:norecur}, Theorem~\ref{t:nonLoop2} should be thought of as a quantitative version of the non-recurrent condition. With this intuition in mind, we are able to construct effective covers by tubes in many geometric situations.
\begin{theorem}[{{\cite[Theorem 3]{CG18}}}]\label{T:applications}
Let $(M,g)$ be a smooth, compact Riemannian manifold of dimension $n$. Let $H\subset M$ be a closed embedded submanifold of codimension $k$. Suppose one of the following assumptions holds{:}
\begin{enumerate}[label=\textbf{\Alph*.},ref=\Alph*]
\item \label{a1} $(M,g)$ has no conjugate points and $H$ has codimension $k>\frac{n +1}{2}$. \smallskip
\item \label{a2}$(M,g)$ has no conjugate points and $H$ is a geodesic sphere.\smallskip
\item \label{a3}$(M,g)$ is a surface with Anosov geodesic flow. \smallskip
\item \label{a6}{$(M,g)$ has Anosov geodesic flow, non-positive curvature and $k>1$.} \smallskip
\item \label{a4}$(M,g)$ has Anosov geodesic flow and {non-positive curvature}, and $H$ is totally geodesic. \smallskip
\item \label{a5} $(M,g)$ has {Anosov geodesic flow} and $H$ is a subset of $M$ that lifts to a horosphere in the universal cover.
\end{enumerate}
Then there exists $C>0$ so that for all $w\in C_c^\infty(H)$ there is $h_0>0$ so that for $0<h<h_0$ and $u_h$ solving~\eqref{e:laplace}
\begin{equation}
\label{e:subEst}
\Big|\int_Hwu_hd\sigma_H\Big|\leq Ch^{\frac{1-k}{2}}\frac{\|u_h\|_{{_{\!L^2(M)}}}}{\sqrt{\log h^{-1}}}.
\end{equation}
\end{theorem}
Finally, there is some uniformity in the estimates from Theorem~\ref{t:micro2} and we can obtain $L^\infty$ estimates. To state these estimates we need to recall a few notions from Riemannian geometry. A Jacobi field along a geodesic $\gamma(t)$ is a vector field along $\gamma(t)$ satisfying
$$
D_t^2J+R(J,\dot{\gamma})\dot \gamma=0
$$
where $D_t$ denotes the covariant derivative along $\gamma$ and $R(\cdot,\cdot)(\cdot)$ denotes the Riemann curvature tensor (see e.g. \cite[Chapter 10]{LeeBook}). We say that $J$ is perpendicular of $\langle J,\dot\gamma\rangle_g=0$ and $\langle D_tJ,\dot\gamma\rangle_g=0$.
For a geodesic $\gamma$, we say that $\gamma$ has a conjugate point of multiplicity $m$ at $t_0$ if there are perpendicular Jacobi fields $\{J_i\}_{i=1}^m $ so that $J_i(0)=0$, $\{D_tJ_i(0)\}_{i=1}^m$ are linearly independent, and $J_i(t_0)=0$. Note that the maximum multiplicity of of a conjugate point is $n-1$ where $n$ is the dimension of the manifold $M$. Moreover, it is not hard to see that there exists $\delta>0$ so that for any geodesic $\gamma$ and any $t_0\in {\mathbb R}$, there are at most $n-1$ conjugate points counted with multiplicity in $[t_0-\delta,t_0+\delta]$.
Define
$$
{\Lambda}_x^{m,r,t}:=\big\{\gamma\in \Lambda: \gamma(0)=x,\,\exists\text{ at least }m\text{ conjugate points to } x \text{ in }\gamma([t-r,t+r])\big\},
$$
where we count conjugate points with multiplicity. Next, for a set $V \subset M$ write
$$
\mc{C}_{_{\!V}}^{m,r,t}:=\bigcup_{x\in V}\{\gamma(t): \gamma\in \Lambda_x^{m,r,t}\}.
$$
Note that the set $\mc{C}_{_{\!x}}^{n-1,0,t}$ is the set of points that are \emph{maximally conjugate to $x$} at time $t$. In particular, for $y\in \mc{C}_{_{\!x }}^{n-1,0,t}$ there is a geodesic $\gamma$ with $\gamma(0)=x$, $\gamma(t)=y$ and so that \emph{all} of the perpendicular Jacobi fields vanish at $t$. One case where this happens is on the sphere where $x$ and $y$ are antipodal points. While the condition
$x\notin \mc{C}_x^{n-1,0,t}$ for $t\geq t_0$ is enough to guarantee $o(1)$ improvements in $L^\infty$ bounds, a notion of \emph{uniform maximal self conjugacy} is necessary to have quantitative improvements.
\begin{theorem}[{{\cite[Theorem 1]{CG18}}}]
\label{t:noConj}
Let $U\subset M$ and suppose that there is $T>0, a>0$ so that for all $x\in U$,
$$
d\Big(x, \mc{C}_x^{n-1,r_a(t),t}\Big)\geq r_a(t),\qquad t\geq T
$$
where $r_a(t)=a^{-1}e^{-a t}$. Then
$$
\|u_h\|_{L^\infty(U)}\leq C\frac{h^{\frac{1-n}{2}}}{\sqrt{\log h^{-1}}}\|u_h\|_{L^2(M)}.
$$
\end{theorem}
It is not hard to see that Theorem~\ref{t:nonLoop2} implies even stronger estimates where we only assume certain volume control on the directions along which $x$ is maximally self-conjugate.
We note at this point that Theorems~\ref{T:applications} and~\ref{t:noConj} subsume \emph{all} previous conditions known to give logarithmic improvements and~\ref{t:app1} subsumes \emph{all} previous conditions known to give $o(1)$ improvements.
\bigskip
\noindent {\sc Acknowledgements.}
Thanks to Yaiza Canzani for comments on an early version of this note. The author is grateful to the National Science Foundation for support under the Mathematical Sciences Postdoctoral Research Fellowship DMS-1502661.
\section{The overall ideas of the proofs}
\subsection{The microlocal estimate}
The first important observation in the proof of Theorem~\ref{t:micro2} is that the most localized that an eigenfunction can be is to an $h^{\frac{1}{2}}$ tube around a single length $\sim1$ piece of geodesic. This is the case, for example, for the highest weight spherical harmonics on $S^2$ given by the restriction of $j^{\frac{n-1}{4}}(x_1+ix_2)^j$ to the sphere, $x_1^2+x_2^2+x_3^2=1$. It is then natural to think of building an eigenfunction out of pieces localized to such tubes. Locally, these pieces are of the form
$$
u_h(x)=h^{-\frac{1-n}{4}} e^{\frac{i}{h}x_1}e^{-\frac{|x|^2}{2h}}a(x)
$$
with the geodesic given by $\gamma=\{(x_1,0,1,0)\mid |x_1|<1\}$. Here, we have taken $(x_1,x',\xi_1,\xi')$ as coordinates on $T^*\!M$. We will refer to this type of object as a \emph{gaussian beam}.
The first step is then to understand how the average over $H$ of an eigenfunction localized to such a tube behaves. This is a two step process. First, if the tube is passing over the hypersurface in a direction which is not normal to the hypersurface, then the contribution is $O(h^\infty)$. Such a restriction is shown in Figure~\ref{f:nonnormal}. Since oscillation remains after restriction, the contribution from such a tube is $O(h^\infty)$. Once we have this in place, we need to study tubes passing normally over $H$ as in Figure~\ref{f:normal}.
\begin{figure}
\begin{tikzpicture}
\begin{scope}[shift= {(0,2.3)}]
\draw [ultra thick] (-4,-1)--(-4,1.5);
\draw[thick, domain =-5.1:-2.9, samples =100 ]plot (\ensuremath{\times}, {2.4*e^( -50*(\ensuremath{\times}+4)*(\ensuremath{\times}+4))-1}) node[right]{};
\draw[->](-4,-1.2)-- (-4.3,-1.2);
\draw[->] (-4,-1.2)--(-3.7,-1.2);
\node at (-4,-1.5){\scriptsize{$h^{\frac{1}{2}}$}};
\draw[->] (-5,.5)--(-5,1.4);
\draw[->] (-5,.5)--(-5,-1);
\node at (-5.3,.2){\scriptsize{$h^{-\frac{1}{4}}$}};
\draw [ultra thick,red] (-5.5,-1)--(-2.5,-1);
\draw node[circle,fill,color=blue,scale=.5] at (-4,-1){};
\node at (-4,1.9){\scriptsize{Profile across a gaussian beam}};
\node at (-4,-2.3){\scriptsize{Profile along a gaussian beam}};
\end{scope}
\draw[ domain=-5.5:-2.5,samples=500] plot (\ensuremath{\times},{sin( (15 *\ensuremath{\times}) r )-1.5}) node[right] {};
\draw[->](-4,-2.7)-- (-2.5,-2.7);
\draw[->] (-4,-2.7)--(-5.5,-2.7);
\node at (-4,-2.9){\scriptsize{$Ch$}};
\draw[->] (-6,-1.5)--(-6,-.5);
\draw[->] (-6,-1.5)--(-6,-2.5);
\node at (-6.5,-1.3){\scriptsize{$h^{-\frac{1}{4}}$}};
\draw[->,blue] (-5.5,-.3)--(-2.5,-.3);
\begin{scope}[shift={(0,2)}]
\draw[ultra thick, red] (0,1.5)-- (0,-1.5);
\node at (0,1.8){\footnotesize{$H$}};
\draw[->,blue] (45:-1.5)--(45:1.5);
\node at (3.3,1.2){$\|u_h\|_{L^2}=1$} ;
\node at (3.3,.2){$\int_{H}u_hd\sigma_{H}=O(h^\infty)$} ;
\end{scope}
\begin{scope}[shift={(6,0)}]
\draw[ domain=-5.5:-2.5,samples=500] plot (\ensuremath{\times},{e^( -10*(\ensuremath{\times}+4)*(\ensuremath{\times}+4))*sin((40*\ensuremath{\times}) r)-1.5}) node[right] {};
\draw[->](-4,-2.7)-- (-2.5,-2.7);
\draw[->] (-4,-2.7)--(-5.5,-2.7);
\node at (-4,-2.9){\scriptsize{$Ch^{\frac{1}{2}}$}};
\draw[->] (-6,-1.5)--(-6,-.5);
\draw[->] (-6,-1.5)--(-6,-2.5);
\node at (-6.5,-1.3){\scriptsize{$h^{-\frac{1}{4}}$}};
\draw[->,white] (-5.5,-.3)--(-2.5,-.3)node[above,midway,black]{\scriptsize{Profile after restriction to $H$}};
\end{scope}
\end{tikzpicture}
\caption{\label{f:nonnormal} Diagram when a gaussian beam passes over $H$ non-normally}
\end{figure}
\begin{figure}
\begin{tikzpicture}
\begin{scope}[shift= {(0,2.3)}]
\draw [ultra thick] (-4,-1)--(-4,1.5);
\draw[thick, domain =-5.1:-2.9, samples =100 ]plot (\ensuremath{\times}, {2.4*e^( -50*(\ensuremath{\times}+4)*(\ensuremath{\times}+4))-1}) node[right]{};
\draw[->](-4,-1.2)-- (-4.3,-1.2);
\draw[->] (-4,-1.2)--(-3.7,-1.2);
\node at (-4,-1.5){\scriptsize{$h^{\frac{1}{2}}$}};
\draw[->] (-5,.5)--(-5,1.4);
\draw[->] (-5,.5)--(-5,-1);
\node at (-5.3,.2){\scriptsize{$h^{-\frac{1}{4}}$}};
\draw [ultra thick,red] (-5.5,-1)--(-2.5,-1);
\draw node[circle,fill,color=blue,scale=.5] at (-4,-1){};
\node at (-4,1.9){\scriptsize{Profile across a gaussian beam}};
\node at (-4,-2.3){\scriptsize{Profile along a gaussian beam}};
\end{scope}
\draw[ domain=-5.5:-2.5,samples=500] plot (\ensuremath{\times},{sin( (15 *\ensuremath{\times}) r )-1.5}) node[right] {};
\draw[->](-4,-2.7)-- (-2.5,-2.7);
\draw[->] (-4,-2.7)--(-5.5,-2.7);
\node at (-4,-2.9){\scriptsize{$Ch$}};
\draw[->] (-6,-1.5)--(-6,-.5);
\draw[->] (-6,-1.5)--(-6,-2.5);
\node at (-6.5,-1.3){\scriptsize{$h^{-\frac{1}{4}}$}};
\draw[->,blue] (-5.5,-.3)--(-2.5,-.3);
\begin{scope}[shift={(0,2)}]
\draw[ultra thick, red] (0,1.5)-- (0,-1.5);
\node at (0,1.7){\footnotesize{{$H$}}};
\foreach \t in {0}
{ \draw[->,blue] (-1.5, \t) -- (1.5,\t); }
\node at (3.3,1.2){$\|u_h\|_{L^2}=1$} ;
\node at (3.3,.2){$\int_{H}u_hd\sigma_{H}\sim ch^{\frac{1}{4}}$} ;
\end{scope}
\begin{scope}[shift={(6,0)}]
\draw[ domain=-5.5:-2.5,samples=500] plot (\ensuremath{\times},{e^( -10*(\ensuremath{\times}+4)*(\ensuremath{\times}+4))-1.5}) node[right] {};
\draw[->](-4,-2.7)-- (-2.5,-2.7);
\draw[->] (-4,-2.7)--(-5.5,-2.7);
\node at (-4,-2.9){\scriptsize{$Ch^{\frac{1}{2}}$}};
\draw[->] (-6,-1.5)--(-6,-.5);
\draw[->] (-6,-1.5)--(-6,-2.5);
\node at (-6.5,-1.3){\scriptsize{$h^{-\frac{1}{4}}$}};
\draw[->,white] (-5.5,-.3)--(-2.5,-.3)node[above,midway,black]{\scriptsize{Profile after restriction to $H$}};
\end{scope}
\end{tikzpicture}
\caption{\label{f:normal} Diagram when a gaussian beam passes over $H$ normally}
\end{figure}
When we decompose eigenfunctions using tubes, we will use tubes of size $R(h)\geq h^\delta$ for some $0\leq\delta<\frac{1}{2}$ so that the symbolic calculus works well. It remains to understand how eigenfunctions localized to such tubes behave when restricted to submanifolds. The key observation is that localization to a small tube implies better control on oscillation and that this control gives improved $L^\infty$ estimates. In particular, imagine that we are working on ${\mathbb R}^n$ with coordinates $(x_1,x',\xi_1,\xi')$ on $T^*{\mathbb R}^n$ and
$$
\gamma=\{(x_1,0,1,0)\mid |x_1|<1\}\subset T^*{\mathbb R}^n.
$$
Then, assume that a function $u_h$ has frequencies only in $|\xi'|\leq R(h)$, i.e. with
$$
\mc{F}_h(u_h)(\xi):=\int e^{-\frac{i}{h}\langle x,\xi\rangle} u_h(x)dx,
$$
satisfying $\text{\ensuremath{\supp}} \mc{F}_h(u)\subset \{|\xi'|\leq R(h)\}$ modulo $O(h^\infty)$. Then
$$
\|(hD_{x'})^mu_h\|_{L^2}\leq CR(h)^m\|u_h\|_{L^2}.
$$
In particular, $u$ is oscillating at frequency $R(h)h^{-1}$ in the $x'$ variables.
Now, suppose that $H$ is given by $\{x_1=\bar{x}=0\}$ where $\bar{x}\in {\mathbb R}^{k-1}$ and $x'=(\bar{x},x'')$. By the standard Sobolev embedding $H^s({\mathbb R}^{m})\to L^\infty({\mathbb R}^{k-1})$ for $s>\frac{k-1}{2}$, such oscillation then implies that
$$
\|u_h(x_1,\cdot)\|_{L^\infty_{\bar{x}}L^2_{x''}}\leq CR(h)^{\frac{k-1}{2}}h^{\frac{1-k}{2}}\|u_h(x_1,\cdot)\|_{L^2_{x'}}.
$$
Then, if we assume in addition $(-h^2\Delta-1)u=0$, a standard energy estimate (see e.g. \cite[Chapter 7]{EZB}) implies that for $u$ localized close enough ($h$ independently) to $\gamma$,
$$
\|u_h(x_1,\cdot)\|_{L^2_{x'}}\leq C\|u_h\|_{L^2}.
$$
In particular,
$$
\|u_h\|_{L^\infty_{\bar{x}}L^2_{x''}}\leq CR(h)^{\frac{k-1}{2}}h^{\frac{1-k}{2}}\|u_h\|_{L^2}.
$$
Finally, if $u$ is also supported on $|x'|\leq R(h)$ modulo $O(h^\infty)$, then
\begin{equation}
\label{e:singleTube}
\Big|\int u_h(0,x'')dx''\Big|\leq CR(h)^{\frac{k}{2}}\|u_h(0,\cdot)\|_{L^2_{x''}}\leq CR(h)^{\frac{n-1}{2}}h^{\frac{1-k}{2}}\|u_h\|_{L^2}.
\end{equation}
In order to make this argument on a general manifold, we construct microlocal cutoffs, $\chi$, to $R(h)$ sized tubes around geodesics (see Figure~\ref{f:tube}) which essentially commute with the Laplacian near $H$. We are then able to use the calculus of pseudodifferential operators to obtain the estimate~\eqref{e:singleTube}.
\begin{figure}
\begin{tikzpicture}
\draw[thick, dashed, ->, scale=1.3] (-{sqrt(3)},-.5)-- ({sqrt(3)},.5)node[right] {$\gamma$};
\draw[thick] (-{sqrt(3)},-.5) ellipse (.2 and .375);
\draw[thick] ({sqrt(3)},.5) ellipse (.2 and .375);
\draw[thick] (-{sqrt(3)},-.5+.375)-- ({sqrt(3)},.5+.375);
\draw[thick] (-{sqrt(3)},-.5-.375)-- ({sqrt(3)},.5-.375);
\draw[gray, shift={(-.5,-.75)}] (-30:0)-- (-30:1.5)--++(90:2.5)--++(150:1.5) node[above]{$S\!N^*\!H$}--cycle;
\draw (0,0) node[below]{$\rho$};
\fill (0,0) circle (.05);
\end{tikzpicture}
\caption{\label{f:tube} A single tube}
\end{figure}
In order to complete the proof of Theorem~\ref{t:micro2}, we then cover $S\!N^*\!H$ by tubes as in Figure~\ref{f:cover}. In the case of $k=n$, combining the estimates is just a matter of applying the triangle inequality. However, when $k<n$, we must once again use that, modulo $O(h^\infty)$, the cutoffs are supported \emph{in physical space} at a distance $R(h)$ from a geodesic. Covering $H$ by balls of radius $R(h)$ and applying the triangle inequality in each ball then gives the required estimate.
\begin{figure}
\begin{tikzpicture}
\begin{scope}[scale=2]
\foreach \t in {0,15,..., 359}
{
\fill[light-gray,rotate=\t] (-.06,0) rectangle (.06,1.5);}
\foreach \t in {0,15,..., 359}
{
\draw[->] (\t:-1.5) -- (\t:1.5); }
\node at (.05,1.6){\footnotesize{$R(h)$}} ;
\node at (.12,1.1){$1$} ;
\node at (1.7,0){\footnotesize{{$\mathcal{T}_j$}}};
\fill[red](0,0)circle (.1);
\draw[ultra thick,dashed,->] (7.5:-1.7)node[left]{\red{$H$}}--(7.5:-.11) ;
\node at (0,1.9){normal slice};
\end{scope}
\begin{scope}[scale=2,shift={(3,0)}]
\foreach \t in {0,.24,..., 2}
{
\fill[fill=light-gray,shift={(\t,0)}] (-.06,0) rectangle (.06,1.5);
\fill[fill=light-blue,shift={(\t+.12,0)}](-.06,0) rectangle (.06,1.5);
\draw[->,shift={(\t,0)}] (0,0) -- (0,1.5);
\draw[->,shift={(\t+.12,0)}] (0,0) -- (0,1.5);}
\foreach \t in {0,.24,..., 2}
{
\fill[light-gray,shift={({\t},0)}] (-.06,0) rectangle (.06,-1.5);
\fill[fill=light-blue,shift={(\t+.12,0)}](-.06,0) rectangle (.06,-1.5);
\draw[->,shift={({\t+.12},0)}] (0,0) -- (0,-1.5);
\draw[->,shift={({\t},0)}] (0,0) -- (0,-1.5);}
\draw[ultra thick,red] (-.2,0)--(2.2,0) node[right]{$H$};
\node at (1,1.9){tangent slice};
\end{scope}
\end{tikzpicture}
\caption{\label{f:cover} The cover of $S\!N^*\!H$ by tubes. (left) The projection onto a plane normal to $H$. (right) The projection of the tubes onto a plane tangent to $H$. Note that each pair of tubes (pointing up and down) on the right corresponds to a whole sphere of tubes in $S\!N^*\!H$ as pictured on the left. The tubes alternate in color only to make it easier to distinguish adjacent tubes.}
\end{figure}
\subsection{From the microlocal estimate to a Theorem~\ref{t:nonLoop2}}
Passing from Theorem~\ref{t:micro2} to Theorem~\ref{t:nonLoop2} is an application of Egorov's theorem to long times. In particular, observe that if $\chi\in C_c^\infty(T^*\!M;[0,1])\cap S_\delta$ is $[t_0,T_0]$ non-self looping, then
\begin{equation}
\label{e:propSymb}
\Big|\frac{1}{T}\int_0^T\chi^2\circ\varphi_tdt\Big|\leq \frac{t_0}{T}.
\end{equation}
Here, we say $\chi\in S_\delta$ if
$$
\big|\partial_x^\alpha\partial_{\xi}^\beta \chi(x,\xi)\big|\leq C_{\alpha\beta}h^{-\delta(|\alpha|+|\beta|)}.
$$
Together with Egorov's theorem to the Ehrenfest time (see e.g. \cite[Proposition 3.8]{DyGu14}),~\eqref{e:propSymb} implies that
\begin{align*}
\|Op_h(\chi)u_h\|_{L^2}^2&=\left\langle \frac{1}{T}\int_0^{T}e^{ith\Delta_g}Op_h(\chi)^*Op_h(\chi)e^{-ith\Delta_g}dtu_h,u_h\right\rangle _{L^2(M)}\\
&\leq \frac{t_0}{T}\big(1+O(h^\epsilon)\big)\|u_h\|_{L^2(M)}^2.
\end{align*}
In particular, if $\bigcup_{j\in \mc{G}_\ell}\Lambda_{\rho_j}^\tau(R(h))$ is $[t_\ell,T_\ell]$ non-self looping, then since there are at most $C$ $\chi_j$ with overlapping supports,
\begin{equation}
\label{e:nonLoopingTubesEst}
\sum_{j\in \mc{G}_\ell}\|Op_h(\chi_j) u_h\|^2_{L^2}\leq C\Big\|\sum_{j\in \mc{G}_\ell}Op_h(\chi_j)u_h\Big\|_{L^2}^2\leq \frac{t_\ell}{T_\ell}\big(1+O(h^\epsilon)\big)\|u_h\|_{L^2(M)}^2.
\end{equation}
An application of Cauchy--Schwarz together with Theorem~\ref{t:micro2} and~\eqref{e:nonLoopingTubesEst} then gives Theorem~\ref{t:nonLoop2}.
\subsection{Construction of effective covers}
There are two mechanisms used to construct the effective covers for Theorem~\ref{T:applications}; contraction and rotation.
\subsubsection{Contraction}
\begin{figure}[h]
\begin{tikzpicture}
\begin{scope}[scale=4]
\draw[color=blue,thick, domain =0:2, samples =100 ]plot (\ensuremath{\times}, {.5*e^( -1*(\ensuremath{\times})}) node[right]{};
\draw[color=blue,thick, domain =0:2, samples =100 ]plot (\ensuremath{\times}, -{.5*e^( -1*(\ensuremath{\times})}) node[right]{};
\foreach \y in{0,0.5,...,2}{
\draw[red,thick] (\y, -.7)-- (\y, .7)node[above]{\tiny{$H$}} ;
\draw[ultra thick] (\y, -{.5*e^( -1*(\y)})-- (\y, {.5*e^( -1*(\y)}) node[midway,left]{\tiny{$G^t(B)$}};
}
\draw[thick,->] (0,-.75)--(2,-.75)node[right]{\tiny{$t\to\infty$}};
\end{scope}
\end{tikzpicture}
\caption{\label{f:tubeContract} Contraction mechanism for constructing effective covers.}
\end{figure}
In the contraction mechanism, pictured in Figure~\ref{f:tubeContract}, we use the fact that a subset of $S\!N^*\!H$ contracts under the flow either forward or backward in time. This is the case, for instance, when $H$ is contained in a stable or unstable horosphere. Under this condition, we start with a macroscopic set $A_0$ and flow it forward in time. We remove all of the pieces of
$$
B_0:=\bigcup_{t_0}^TG^t(A_0)
$$
intersecting $A_0$. Since $A_0$ is contracting, we may choose $t_0$ large enough so that $\sigma_{_{\!\!S\!N^*\!H}}(B_0)\leq \epsilon\sigma_{_{\!\!S\!N^*\!H}}(A_0)$. We then let $G_0=A_0\setminus B_0$ and $A_1=B_0$. By construction $G_0$ is $[t_0,T]$ non-self looping. We can then repeat the process replacing $T$ by $\frac{1}{2}T$ to obtain $G_1$ that is $[t_0,\frac{1}{2}T]$ non-self looping. Inductively repeating this process, we construct an effective non-self looping cover of $S\!N^*\!H$.
\subsubsection{Rotation}
\begin{figure}[h]
\begin{tikzpicture}
\foreach \t in{0}{
\begin{scope}[shift= {(2.4*\t-1.5,-4)},scale=1.5]
\draw[thick,->](\t,-1.7)node[below]{\tiny{${t=\t}$}}--(\t,1.7);
\draw[gray,->](\t,0)--(\t+.5,.5)node[right]{\color{gray}{\tiny{$H_{|\xi|_g^2}$}}};
\draw[thick,dashed,blue](\t-.3,-.9)--(\t+.4,1.2)node[right]{\tiny{\color{blue}$T_\rho S\!N^*\!H$}};
\draw[thick,->](\t-.85,.5)--(\t+.85,-.5);
\draw[thick,red] ({\t-.75*.2/(\t+1)},{-.75*.6*(\t+1)}) --({\t+.75*.2/(\t+1)}, {.75*.6*(\t+1)})node[left]{\tiny{\color{red}$G^t(B_\rho)$}};
\end{scope}
}
\foreach \t in{1,...,2}{
\begin{scope}[shift= {(2.4*\t-1.5,-4)},scale=1.5]
\draw[thick,->](\t,-1.7)node[below]{\tiny{${t=\t}$}}--(\t,1.7);
\draw[gray,->](\t,0)--(\t+.5,.5)node[right]{\color{gray}{\tiny{$H_{|\xi|_g^2}$}}};
\draw[thick,dashed,blue](\t-.3,-.9)--(\t+.4,1.2)node[right]{\tiny{\color{blue}$T_{_{\!G^{\t}(\rho)}} S\!N^*\!H$}};
\draw[thick,->](\t-.85,.5)--(\t+.85,-.5);
\draw[thick,red] ({\t-.75*.2/(\t+1)},{-.75*.6*(\t+1)}) --({\t+.75*.2/(\t+1)}, {.75*.6*(\t+1)})node[left]{\tiny{\color{red}$G^t(B_\rho)$}};
\end{scope}
}
\end{tikzpicture}
\caption{ \label{f:tubeRotate} Rotation mechanism for constructing effective covers.}
\end{figure}
In the rotation mechanism, pictured in Figure~\ref{f:tubeRotate}, a ball of small radius $B_\rho\subsetS\!N^*\!H$ rotates to become transverse to the plane of $T_{G^t(\rho)}S\!N^*\!H$ (when $G^t(\rho)\in S\!N^*\!H$) as $t\to \pm \infty$. In this situations, we can use the implicit function theorem to show that the intersection of $\bigcup_{t_0}^T\varphi_t(B_\rho)$ with $S\!N^*\!H$ is a finite union of lower dimensional subsets. Covering these lower dimensional subsets by tubes with small volume, we are able to construct an effective cover.
\subsubsection{Effective covers with no uniformly maximal self-conjugate points}
\begin{figure}[h]
\includegraphics[width=\textwidth]{conjugacy1.pdf}
\caption{\label{f:noConj} The construction of a non-self looping collection of tubes under a non-uniformly maximal self conjugacy assumption.}
\end{figure}
We now sketch the construction of the cover by tubes that is used to prove Theorem~\ref{t:noConj}. The crucial fact is that if $J$ is a Jacobi field along $\gamma$ with $J(0)=0$ and $\Gamma:(-\epsilon,\epsilon)\times {\mathbb R}\to M$ is a map so that
$$
\begin{gathered}
\Gamma(s,0)=\gamma(0),\qquad\qquad \gamma_s:t\mapsto\Gamma(s,t)\text{ is a geodesic},\qquad\qquad\partial_sD_t\Gamma(0,0)=D_tJ(0),
\end{gathered}
$$
then $\partial_s\gamma(0,t)=J(t)$. Said another way, if $J(t_0)$ is non-zero, then for $s\neq 0$ small $\gamma_s(t_0)\neq \gamma(0)$.
Translating this from the $SM\subset TM$ to $S^*\!M\subset T^*\!M$, this implies that there is a vector $V=(D_tJ(0))^\sharp\in T_{{\dot\gamma}^\sharp}S^*_xM$ so that
$$
d\pi dG^{t_0} V\neq 0
$$
where $\pi:T^*\!M\to M$ denotes the projection. Using this together with the implicit function theorem, we find a submanifold $B\subset S^*_xM$ of dimension$<n-1$ and a neighborhood $W$ of ${\dot\gamma}^\sharp$ so that for $t$ near $t_0$ and $\rho\in W\setminus B$, $G^t(\rho)\notin S^*_xM$. We can then cover $B$ by $\sim R(h)^{2-n}$ tubes.
Since $x$ is not maximally self-conjugate for $t>s_0$, we can repeat this argument near each point $\rho\in S^*_xM$ and then for approximately $T$ values of $t_0$, we produce a large collection of tubes, $\mc{G}$ whose union is $[s_0,T]$ non-self looping and $\sim T R(h)^{2-n}$ possibly looping tubes $\mc{B}$.
In order to make this construction work, we must control the size of the neighborhood $W$ near each $\rho$. It is precisely in this quantification where the uniformity in the non-maximally self conjugacy is used.
\section[#1]{{\Large{#1}}}}
\def\smallsection#1{\smallskip\noindent\textbf{#1}.}
\def\bigsection#1{\smallskip\noindent{\textbf{#1}.}\smallskip}
\newtheorem{theorem}{Theorem}
\newtheorem{prop}{Proposition}
\numberwithin{prop}{section}
\newtheorem{corol}{Corollary}
\numberwithin{corol}{section}
\newtheorem{lemma}{Lemma}
\numberwithin{lemma}{section}
\newtheorem{conjecture}{Conjecture}
\numberwithin{conjecture}{section}
\newenvironment{remarks}[1][]{\begin{remark}\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]\end{trivlist}\begin{itemize}}{\end{itemize}\end{remark}}
{\theoremstyle{definition}
\newtheorem{defin}{Definition}
\numberwithin{defin}{section}
}
\numberwithin{figure}{section}
\renewcommand{\Re}{\mathop{\rm Re}\nolimits}
\renewcommand{\Im}{\mathop{\rm Im}\nolimits}
\newcommand{\quotient}[2]{{\left.\raisebox{.2em}{$#1$}\middle/\raisebox{-.2em}{$#2$}\right.}}
\newcommand{C^\infty}{C^\infty}
\newcommand{C^\infty_{\rm{c}}}{C^\infty_{\rm{c}}}
\newcommand{\mathbb R}{\mathbb R}
\newcommand{\operatorname{Spec}}{\operatorname{Spec}}
\newcommand{\begin{flushleft}}{\begin{flushleft}}
\newcommand{\end{flushleft}}{\end{flushleft}}
\newcommand{\begin{flushright}}{\begin{flushright}}
\newcommand{\end{flushright}}{\end{flushright}}
\newcommand{\begin{center}}{\begin{center}}
\newcommand{\end{center}}{\end{center}}
\newcommand{\nq}{\vspace{1cm}
\newcommand{\nsq}{\vspace{5mm}
\newcommand{\bigcup}{\bigcup}
\newcommand{\bigcap}{\bigcap}
\newcommand{\mcal}[1]{\mathcal{#1}}
\newcommand{\unionLim}[3]{\bigcup _{#1=#2}^{#3}}
\newcommand{\isectLim}[3]{\bigcap _{#1=#2}^{#3}}
\newcommand{\recip}[1]{\frac{1}{#1}}
\newcommand{\Rightarrow}{\Rightarrow}
\newcommand{\Leftrightarrow}{\Leftrightarrow}
\newcommand{\mathbb{C}}{\mathbb{C}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{\begin{enumerate}}{\begin{enumerate}}
\newcommand{\end{enumerate}}{\end{enumerate}}
\newcommand{\sumLim}[3]{\sum_{#1=#2}^{#3}}
\newcommand{\liml}[2]{\lim_{#1 \to #2}}
\newcommand{\backepsilon}{\backepsilon}
\newcommand{\left( \begin{array}{cc}0&-i\\i&0\\ \end{array}\right)}{\left( \begin{array}{cc}0&-i\\i&0\\ \end{array}\right)}
\newcommand{\text{\textopenbullet}}{\text{\textopenbullet}}
\newcommand{\times}{\times}
\newcommand{\epsilon}{\epsilon}
\newcommand{\textbf{Proof:}}{\textbf{Proof:}}
\newcommand{\textbf{Lemma:}}{\textbf{Lemma:}}
\newcommand{\textbf{Corollary:}}{\textbf{Corollary:}}
\newcommand{\intL}[2]{\int_{#1}^{#2}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{Poincar\'{e} }{Poincar\'{e} }
\newcommand{\nonumber\\}{\nonumber\\}
\newcommand{\langle}{\langle}
\newcommand{\rangle}{\rangle}
\newcommand{\mathcal{D}_{\loc}}{\mathcal{D}_{\loc}}
\newcommand{{\int\negthickspace\negthickspace\negthickspace\negthinspace -}}{{\int\negthickspace\negthickspace\negthickspace\negthinspace -}}
\newcommand{Calder\'{o}n }{Calder\'{o}n }
\newcommand{-\!\!\!\!\!\!\int}{-\!\!\!\!\!\!\int}
\newcommand{\Psi_{\textup{hom}}}{\Psi_{\textup{hom}}}
\newcommand{\mc}[1]{\mathcal{#1}}
\newcommand{\partial\Omega}{\partial\Omega}
\theoremstyle{remark}
\newtheorem{remark}{Remark}
\newcommand{\Delta_{\partial\Omega,\delta'}}{\Delta_{\partial\Omega,\delta'}}
\newcommand{\Deltad}[1]{\Delta_{#1,\delta}}
\newcommand{\operatorname{sech}}{\operatorname{sech}}
\newcommand{\operatorname{arcsec}}{\operatorname{arcsec}}
\newcommand{\mc{F}_h}{\mc{F}_h}
\newcommand{\partial_{\nu}\mc{D}\ell}{\partial_{\nu}\mc{D}\ell}
\newcommand{N}{N}
\newcommand{\mc{S}\ell}{\mc{S}\ell}
\newcommand{\mc{D}\ell}{\mc{D}\ell}
\renewcommand{\S}{\mc{S}\ell}
\newcommand{C_c^\infty}{C_c^\infty}
\renewcommand{\O}[1]{\mathpzc{O}_{#1}}
\renewcommand{\o}[1]{\mathpzc{o}_{#1}}
\newcommand{\opht}[1]{\operatorname{Op_{h,#1}}}
\newcommand{\wt}[1]{\widetilde{#1}}
\newcommand{N_2}{N_2}
\newcommand{\Ph}[2]{\Psi^{#1}_{#2}}
\newcommand{D_{\Omega}}{D_{\Omega}}
\newcommand{$$}{$$} |
\section*{Introduction}
In this paper we investigate the Hochschild homology of a class of product algebras called the
\emph{smash biproducts}~\cite{CaenIonMiliZhu00} (or also referred as the \emph{twisted tensor
products}~\cite{CapSchichlVanzura95}) that include Hopf-cross products and Ore extensions.
Then we effectively calculate Hochschild homologies of quantum tori, multiparametric (quantum) affine spaces~\cite{GuccGucc97,Wamb93}, quantum complete
intersections~\cite{BerErd08}, quantum Weyl algebras~\cite{Richard:QuantumWeyl}, deformed
completed Weyl algebras~\cite{DuroMeljSamsSkod07,MeljSkod07}, and the quantum matrix algebra
$M_q(2)$~\cite{ArtinSchelterTate:QuantumDeformationsOfGLN,Kassel-book}.
Our strategy relies on splitting the Hochschild complex into a (twisted) product of two
Hochschild complexes induced by the decomposition of the underlying algebra. Then we combine
the homologies of the individual pieces via a suitable spectral sequence. In doing so, we do
not rely on any ad-hoc resolutions of the underlying algebras as bimodules over themselves. The
key observation is that if one of the component algebras has Hochschild homological dimension
less than 2, then the spectral sequence degenerates on the $E^2$-page yielding the result
immediately. We refer the reader to Section~\ref{sect:HomologyOfSmashBiproducts} for technical
details.
Here is a plan of the paper: In Section~\ref{sect:smash-biproducts} we recall basic definitions and tools needed to work with smash biproduct algebras, and then in Section~\ref{sect:HomologyOfSmashBiproducts} we construct our homological machinery. Section~\ref{sect:calculation} contains our calculations. In Section~\ref{subsect-quantum-tori} we calculate the Hochschild homologies of the quantum plane $k[x]\# k[y]$, the quantum cylinder $k[x]\# k[y,y^{-1}]$, and the quantum torus $k[x,x^{-1}]\# k[y,y^{-1}]$.
Next, in Section \ref{subsect-multip-aff-sp} we calculate the Hochschild homology of the multiparametric affine space $S(X_\nu,\Lambda)$~\cite{GuccGucc97}, as well as a particular case of quantum Weyl algebras~\cite{Richard:QuantumWeyl}. The Hochschild homology of the quantum complete intersection algebra $C_{a,b}:= k\langle x,y\rangle/\langle x^a, xy-qyx,y^b\rangle $ for every $a,b\geq 2$~\cite{BerErd08} is computed in Section \ref{subsect-quant-comp-int}. In Section \ref{subsect-quant-Weyl}, we deal with the quantum Weyl algebra $S(X_\mu,Y_\nu,\Lambda)$~ \cite{Richard:QuantumWeyl} in full generality. Subsection \ref{sect:WeylAlgebras} deals with the $(\G{g}, D)$-deformed Weyl algebra $A^{\rm pol}_{\G{g}, D} := S(V) \rtimes U(\G{g})$, and its completion $A_{\G{g}, D} := \widehat{S(V)} \rtimes U(\G{g})$~\cite{DuroMeljSamsSkod07,MeljSkod07}. Although deformed Weyl algebras are Hopf-cross products, neither component is necessarily of low Hochschild homological dimension. Nevertheless, our method plays a critical role in identifying the (continuous) Hochschild homology of $A_{\G{g}, D}$ with its dense subalgebra $A^{\rm pol}_{\G{g}, D}$. In particular, we obtain the results
\cite[Thm. 2.1]{FeigFeldShoi05} and \cite[Thm. 4]{FeigTsyg83} for the for the completed Weyl algebra $A_{2n}$, since the Hochschild homology of $A^{\rm pol}_{2n}$ is already known. The last subsection, Subsection \ref{subsect-quant-matrix}, is devoted to the Hochschild homology of the quantum matrix algebra $M_q(2)$. The fact that the algebra $M_q(2)$ is a short tower of Ore extensions~\cite{Kassel-book} makes it amenable for the tools we develop here. In fact, all quantum matrix algebras $M_q(n)$ are towers of Ore extensions~\cite{ArtinSchelterTate:QuantumDeformationsOfGLN,TorrLena00,Kassel-book}. However, the growth of the lengths of these towers renders the arguments we use in the present paper ineffective. As such, the homology of the algebras $M_q(n)$ for $n>2$, and the quantum linear groups~\cite{ParshallWang:QuantumLinearGroups} demand a different approach, which we postpone to a subsequent paper.
\subsection*{Notation and Conventions}
We fix a ground field $k$ of characteristic 0. All algebras are assumed to be unital and
associative, but not necessarily finite dimensional. We use $k[S]$ and $k\{S\}$ to denote
respectively the free commutative and free noncommutative polynomial algebras generated by a
set $S$. We also use $\left<S\right>$ to denote either the $k$-vector space spanned, or the
two-sided ideal, generated by the set $S$ depending on the context. If $G$ is a group,
$k[G]$ denotes the group algebra of $G$ over $k$.
\section{Smash biproducts, Hopf-cross products, and Ore extensions}\label{sect:smash-biproducts}
\subsection{The smash biproduct of algebras}~
Let us now recall from \cite{CaenIonMiliZhu00} the smash biproduct of
algebras, which are also called the twisted tensor
products~\cite{CapSchichlVanzura95}. Let $A$ and $B$ be two
algebras, and let $R\colon B\otimes A \longrightarrow A\otimes B$ be a linear map written
as $R(b\otimes a) =: {}^Ra\otimes b^R$ making the diagrams
\begin{equation}\label{DistributiveLaw}
\xymatrix{
& B\ar[dr]^{1\otimes B}\ar[dl]_{B\otimes 1}\\
B\otimes A \ar[rr]^{ R } & & A\otimes B\\
& A \ar[ul]^{1\otimes A} \ar[ur]_{A\otimes 1}
}
\qquad
\xymatrix{
B\otimes B\otimes A \ar[r]^{B\otimes R }\ar[d]_{\mu_B\otimes A}
& B\otimes A\otimes B \ar[r]^{ R \otimes B}
& A\otimes B\otimes B \ar[d]^{A\otimes \mu_B}\\
B\otimes A \ar[rr]^R & & A\otimes B\\
B\otimes A\otimes A \ar[r]_{ R \otimes B} \ar[u]^{B\otimes\mu_A}
& A\otimes B\otimes A \ar[r]_{A\otimes R }
& A\otimes A\otimes B \ar[u]_{\mu_A\otimes B}
}
\end{equation}
commutative. We shall call any such map a \emph{distributive law}. One can then show that $A\otimes B$ is a unital associative algebra with the multiplication
\[
(a\otimes b)(a'\otimes b') := a({}^Ra') \otimes (b^R)b',
\]
for any $a,a'\in A$ and $b,b' \in B$, and with the unit $1\otimes 1 \in A \otimes B$, if and only if the diagrams in \eqref{DistributiveLaw} commute, \cite[Thm. 2.5]{CaenIonMiliZhu00}. We will use $A\#_R B$ to denote this algebra.
Let us note that the natural inclusions
\[
i_A:A \longrightarrow A\#_R B, \quad a \mapsto a\otimes 1, \qquad i_B:B \longrightarrow A\#_R B, \quad b \mapsto 1\otimes b
\]
are algebra maps, and that the map
\[
\Psi:=\mu_{A\#_R B} \circ (i_A \otimes i_B) : A\otimes B \longrightarrow A\#_R B
\]
is an isomorphism of vector spaces. Then, the map $R$ can be recovered as
\[
R = \Psi^{-1} \circ \mu_{A\#_R B} \circ (i_B \otimes i_A).
\]
Let us next review a few (more concrete) examples of this construction.
\subsection{Ore extensions}~\label{subsect:OreExtensions}
Let $A$ be an algebra, and $k[X]$ be the polynomial algebra in one indeterminate. Let
$\alpha:A\to A$ be an algebra homomorphism, and $\delta:A\to A$ be a derivation. Then the
invertible distributive law defined as
\[ R\colon k[X]\otimes A \longrightarrow A\otimes k[X], \qquad R(X \otimes a) := \alpha(a) \otimes X + \delta(a) \otimes 1 \]
yields the Ore extension associated to the datum $(A,\alpha,\delta)$, {\it i.e.\/}\
$A\#_R k[X] \cong A[X,\alpha,\delta]$, \cite[Ex. 2.11]{CaenIonMiliZhu00}.
We refer the reader to \cite{FaddReshTakh89, Kassel-book} for an account of the Ore extensions in the
quantum group theory.
\subsection{Hopf-cross products}\label{HopfSmashProduct}~
We note from \cite[Ex. 4.2]{CaenIonMiliZhu00} that the smash biproduct construction covers
Hopf-cross products, i.e. algebras of the form $A \otimes H$ where $H $ is a Hopf algebra and
$ A$ is a (left) $H$-module algebra. The latter means that there is an action
$\triangleright \colon H \otimes A\to A$ which satisfies
\begin{equation}\label{ModuleAlgebra}
h\triangleright (ab) = (h_{(1)}\triangleright a)(h_{(2)}\triangleright b)
\end{equation}
for any $h\in H $, and any $a,b\in A$. Then, there is an algebra structure on $A \rtimes H := A \otimes H$ given by
\[
(a\otimes h)(b\otimes g) = a(h\ps{1}\triangleright b) \otimes h\ps{2}g,
\]
for any $a,b \in A$, and any $h,g \in H$, with unit $1 \otimes 1 \in A \rtimes H$. Now letting
\begin{equation}
\label{HopfDistributive}
R\colon H \otimes A\to A\otimes H, \qquad R(h\otimes a) = h_{(1)}\triangleright a\otimes h_{(2)}
\end{equation}
for any $a\in A$ and any $h\in H$ it follows from \eqref{ModuleAlgebra} that
$A \rtimes H = A\#_R H$.
\subsection{Algebras with automorphisms}~\label{GroupAction}
Let $A$ be any unital associative algebra, and let $G$ be a group acting on $A$ via
automorphisms. Consider the distributive law $R\colon k[G]\otimes A\to A\otimes k[G]$ given
by
\begin{equation}
R(g\otimes a) = (g\triangleright a)\otimes g
\end{equation}
for every $a\in A$ and $g\in G$. Let us denote the smash biproduct algebra $A\#_R k[G]$
simply by $A\# G$. This is a special case of the smash product given by the distributive
law~\eqref{HopfDistributive} for the Hopf algebra $H=k[G]$.
\subsection{Algebras with derivations}~\label{AlgebrasWithDerivation}
Let $\G{g}$ be a Lie algebra, and $A$ a unital associative algebra on which $\G{g}$ acts by derivations; that is,
\[
X\triangleright (ab) = (X\triangleright a)b + a(X\triangleright b)
\]
for any $a,b\in A$, and any $X\in\G{g}$. Then,
\[
R\colon U(\G{g})\otimes A\to A\otimes U(\G{g}), \qquad R(X\otimes a) = (X\triangleright a)\otimes 1 + a\otimes X
\]
determines an invertible distributive law, and hence the smashed biproduct algebra $A \#_R U(\G{g})$ which we simply denote by $A \# \G{g}$.
Let us note that the above conditions endows $A$ with a $U(\G{g})$-module algebra structure, that is, the algebra $A \# \G{g}$ is a special case of the Hopf crossed product $A\rtimes H$ with $H=U(\G{g})$, and the distributive law~\eqref{HopfDistributive}.
\section{Homology of smash biproducts}\label{sect:HomologyOfSmashBiproducts}
\subsection{The bar complex}~
Given an associative algebra $A$, together with a right $A$-module $V$ and
a left $A$-module $W$, one can form the two sided bar complex
\begin{equation}\label{bar-complex}
{\rm CB}_*(V,A,W) := \bigoplus_{n\geq 0} V\otimes A^{\otimes n}\otimes W
\end{equation}
together with the differentials $d\colon {\rm CB}_n(V,A,W)\longrightarrow {\rm CB}_{n-1}(V,A,W)$, $n\geq 1$, defined as
\begin{align}\label{bar-diff}
d(v\otimes a_1\otimes\cdots\otimes a_n\otimes w)
&= v\cdot a_1\otimes a_2\otimes\cdots\otimes a_n\otimes w \nonumber\\
& + \sum_{j=1}^{n-1}(-1)^j v\otimes\cdots\otimes a_j a_{j+1}\otimes\cdots\otimes w \\
& + (-1)^n v\otimes a_1\otimes\cdots \otimes a_{n-1}\otimes a_n\cdot w.\nonumber
\end{align}
\subsection{Hochschild homology}~
Now, let $A$ be a $k$-algebra and $V$ be an $A$-bimodule, or equivalently, a right module over the enveloping algebra $A^e:= A \otimes A^{\rm op}$. Then the Hochschild homology of $A$, with coefficients in $V$, is defined to be
\[
H_\ast(A,V) := {\rm Tor}^{A^e}_\ast(A,V).
\]
Equivalently, see for instance \cite[Prop. 1.1.13]{Loday-book}, the Hochschild homology of $A$, with coefficients in the $A$-bimodule $V$, is the homology of the complex
\[
{\rm CH}_\ast(A,V) := \bigoplus_{n\geq 0}{\rm CH}_n(A,V), \qquad {\rm CH}_n(A,V) := V \otimes A^{\otimes\,n}
\]
with respect to the differential $b\colon {\rm CH}_n(A,V)\longrightarrow {\rm CH}_{n-1}(A,V)$, defined for $n\geq 1$ as
\begin{align}\label{Hochschild-diff}
b(v\otimes a_1\otimes\cdots\otimes a_n)
& = v\cdot a_1\otimes a_2\otimes\cdots\otimes a_n \nonumber\\
& + \sum_{j=1}^{n-1}(-1)^j v\otimes a_1 \otimes\cdots\otimes a_j a_{j+1}\otimes\cdots\otimes a_n \\
& + (-1)^n a_n\cdot v\otimes a_1\otimes\cdots \otimes a_{n-1}.\nonumber
\end{align}
\subsection{Hochschild homology with invertible distributive laws}~\label{Subsect-Hochschild-smash-E1-E2}
Let $R\colon B\otimes A\to A\otimes B$ be an invertible distributive law, and $V$ an
$A\#_R B$-bimodule, that is, the diagram
\begin{equation}
\xymatrix{
B\otimes A\otimes V \ar@/^1pc/[dd]\ar[r] & B\otimes V \ar[rd] &
& V\otimes A \ar[dl] & \ar[l] V\otimes B\otimes A \ar@/^1pc/[dd] \\
& & V\\
A\otimes B\otimes V \ar[r]\ar@/^1pc/[uu] & A\otimes V \ar[ur] &
& V \otimes B \ar[ul] & \ar[l] V\otimes A\otimes B \ar@/^1pc/[uu]
}
\end{equation}
is commutative. We now introduce a bisimplicial complex computing the Hochschild homology, with coefficients, of the algebra $A\#_R B$. We shall need an iterative use of the distributive law (and its inverse), as such, we make use of the notation
\[
R(b_j\otimes a_i) = {}^{R_{ji}}a_i \otimes b_j^{R_{ji}}, \qquad R^{-1}(a_i\otimes b_j) = {}^{L_{ji}}b_j \otimes a_i^{L_{ji}}.
\]
\begin{proposition}\label{prop:diagonal}
Given a smash biproduct algebra $A\#_R B$ with an invertible distributive law $R:B\otimes A \to A \otimes B$, and an $A\#_R B$-bimodule $V$, the isomorphism
\begin{align*}
v \otimes (a_1, b_1) & \ot\cdots\ot (a_n,b_n) \mapsto \\
& a_1^{L_{11} \ldots L_{n1}} \ot\cdots\ot a_{n-1}^{L_{(n-1)(n-1)} L_{(n-1)n}} \otimes a_n^{L_{nn}}\otimes v \otimes {}^{L_{11}} b_1\otimes {}^{L_{21} L_{22}}b_2 \ot\cdots\ot {}^{L_{n1} \ldots L_{nn}}b_n
\end{align*}
identifies the Hochschild homology complex of the algebra $A\#_R B$, with coefficients in $V$, with the diagonal of the bisimplicial complex given by $\C{B}_{p,q}(A,V,B):=A^{\otimes\,p} \otimes V \otimes B^{\otimes\,q}$ with the structure maps
\begin{align*}
\partial^h_i & (a_1\ot\cdots\ot a_q\otimes v\otimes b_1\ot\cdots\ot b_p)\\
:= & \begin{cases}
a_1\ot\cdots\ot a_q\otimes v\triangleleft b_1\ot\cdots\ot b_p & \text{ if } i=0,\\
a_1\ot\cdots\ot a_q\otimes v \otimes b_1\ot\cdots\ot b_ib_{i+1}\ot\cdots\ot b_p & \text{ if } 1\leq i\leq p-1,\\
{}^{R_{p1} }a_1\ot\cdots\ot {}^{R_{pq} }a_q\otimes b_p^{R_{p1} \ldots R_{pq}}\triangleright v \otimes b_1\ot\cdots\ot b_{p-1} & \text{ if } i=p,
\end{cases}\\\\
\partial^v_j & (a_1\ot\cdots\ot a_q\otimes v\otimes b_1\ot\cdots\ot b_p)\\
:= & \begin{cases}
a_2\ot\cdots\ot a_q\otimes v\triangleleft {}^{R_{11} \ldots R_{p1}} a_1\otimes b_1^{R_{11}}\ot\cdots\ot b_p^{R_{p1}} & \text{ if } j=0,\\
a_1\ot\cdots\ot a_ja_{j+1} \otimes a_q\ot\cdots\ot v\otimes b_1\ot\cdots\ot b_p & \text{ if } 1\leq j\leq q-1, \\
a_1\ot\cdots\ot a_{q-1}\otimes a_q\triangleright v\otimes b_1\ot\cdots\ot b_p & \text{ if } j=q.
\end{cases}
\end{align*}
\end{proposition}
\begin{proof}
We observe that
\begin{align*}
\partial^h_p\partial^v_0 & (a_1\ot\cdots\ot a_q\otimes v\otimes b_1\ot\cdots\ot b_p) \\
= & \partial^h_p\left(a_2\ot\cdots\ot a_q\otimes v\triangleleft {}^{R_{11} \ldots R_{p1}} a_1\otimes b_1^{R_{11}}\ot\cdots\ot b_p^{R_{p1}}\right) \\
= & {}^{R_{p2}}a_2\ot\cdots\ot {}^{R_{pq}}a_q\otimes b_p^{R_{p1}\ldots R_{pq}}\triangleright\left(v\triangleleft {}^{R_{11} \ldots R_{p1}} a_1\right)\otimes b_1^{R_{11}}\ot\cdots\ot b_{p-1}^{R_{(p-1)1}} \\
= & {}^{R_{p2}}a_2\ot\cdots\ot {}^{R_{pq}}a_q\otimes \left(b_p^{R_{p1}\ldots R_{pq}}\triangleright v\right)\triangleleft {}^{R_{11} \ldots R_{p1}}a_1\otimes b_1^{R_{11}}\ot\cdots\ot b_{p-1}^{R_{(p-1)1}} \\
= & \partial^v_0\left({}^{R_{p1} }a_1\ot\cdots\ot {}^{R_{pq} }a_q\otimes b_p^{R_{p1} \ldots R_{pq}}\triangleright v \otimes b_1\ot\cdots\ot b_{p-1}\right) \\
= & \partial^v_0\partial^h_p(a_1\ot\cdots\ot a_q\otimes v\otimes b_1\ot\cdots\ot b_p).
\end{align*}
For $1 \leq i \leq p-1$ and $1 \leq j \leq q-1$, the commutativity $\partial^h_i\partial^v_j = \partial^v_j\partial^h_i$ follows from the commutativity of the second diagram in \eqref{DistributiveLaw}.
\end{proof}
Considering the spectral sequence associated to the filtration by the rows (or the columns) of the bicomplex given in Proposition \ref{prop:diagonal}, we obtain the following result.
\begin{theorem}\label{thm:HochschildOfSmashProducts}
Let $R\colon B\otimes A\to A\otimes B$ be an invertible distributive law, and let $V$ be an $A\#_R B$-bimodule. Then there are two spectral sequences whose $E^1$-terms are given by
\begin{equation}\label{eqn-E1}
E^1_{p,q} = H_q(A,{\rm CH}_p(B,V)),\qquad \qquad {}'E^1_{p,q} = H_p(B,{\rm CH}_q(A,V))
\end{equation}
which converge to the Hochschild homology of the smash biproduct $A\#_R B$, with coefficients in $V$.
\end{theorem}
\begin{corollary}\label{SpecialCase}
Let $\G{g}$ be a Lie algebra action on an algebra $A$ by derivations, and let $G$ be a
group action on an algebra $B$ by automorphisms. Let also $V$ be an $A\# \G{g}$-bimodule, and $W$ a $B\# G$-bimodule. Then there are two spectral sequences such that
\begin{equation}
\label{eq:LieAction}
H_{p+q}(A\# \G{g},V) \Leftarrow {}'E^1_{p,q} = H_p(\G{g},{\rm CH}_q(A,V)^{ad}),
\end{equation}
and
\begin{equation}
\label{eq:GroupAction}
H_{p+q}(B\# G,W) \Leftarrow {}'E^1_{p,q} = H_p(G,{\rm CH}_q(B,W)^{ad}).
\end{equation}
\end{corollary}
\subsection{Homology of smash biproducts by amenable algebras}~
We will call a unital associative algebra $B$ as \emph{amenable} if it has Hochschild
homological dimension 0, in other words $B$ is a $B^e$-flat
module~\cite{Johnson:AmenableBanachAlgebras}. In the discrete (algebraic) case, the typical
examples are group algebras $k[G]$ of finite groups where the characteristic of the field $k$ does not
divide the order of the group $G$. In the measurable case, as in the case dealt originally
in~\cite{Johnson:AmenableBanachAlgebras}, the typical examples are of the form
$L^1(G)$ where $G$ is an amenable group. In particular, all compact groups and locally
compact abelian groups are amenable,~\cite[Chap.3]{Pier:AmenableGroups}.
In the sequel, we make frequent use of the notation
\begin{equation}\label{eqn-H_0}
V_B:= \frac{V}{[B,V]} = \frac{V}{\langle bv-vb\mid v\in V, b\in B \rangle} \cong H_0(B,V)
\end{equation}
and
\begin{equation}
V^B := \{v\in V|\ bv=vb, \text{ for all } b\in B\} \cong H^0(B,V)
\end{equation}
for every $B$-bimodule $V$.
\begin{theorem}\label{thm:amenable}
Let $A$ and $B$ be two unital associative algebras, where $B$ is amenable, and let $R\colon B\otimes A\to A\otimes B$ be an invertible distributive law. Then,
\[
H_n(A\#_RB,V)\cong H_n(A,V)_B
\]
for any $n\geq 0$, and for any $A\#_R B$-bimodule $V$.
\end{theorem}
\begin{proof}
By Theorem \ref{thm:HochschildOfSmashProducts} we have
\[
H_\ast(A\#_RB,V) \Leftarrow {}'E^1_{p,q} = H_p(B,{\rm CH}_q(A,V)).
\]
Since $B$ is amenable, we have
\[
{}'E^1_{p,q} \cong H_p(B,{\rm CH}_q(A,V)) \cong \begin{cases}
{\rm CH}_q(A,V)_B & \text{ if } p =0,\\
0 & \text{ otherwise}.
\end{cases}
\]
Furthermore, since $B$ is flat as a left $B^e$-module, the functor $(\ \cdot\ )\otimes_{B^e}B$ is exact. As such,
\[
H_\ast(A\#_RB,V) \Leftarrow {}'E^2_{p,q} = \begin{cases}
H_q(A,V)_B & \text{ if } p =0,\\
0 & \text{ otherwise}.
\end{cases}\]
The result then follows from the spectral sequence consisting of only one column.
\end{proof}
In particular, for $V = A$ and $B = k[G]$, where $G$ is a finite group, we have the following.
\begin{corollary}
Let $G$ be a finite group acting on a unital associative algebra $A$ by automorphisms. Then
\[
H_\ast(A\# G,A) \cong HH_\ast(A)_{kG} \cong HH_\ast(A)^G,
\]
where $HH_\ast(A)^G$ denotes the space of $G$-invariants (under the diagonal action).
\end{corollary}
\subsection{Homology of smash biproducts by smooth algebras}~
A unital associative $k$-algebra $B$ is called \emph{smooth} if $B$ has Hochschild
cohomological dimension 1, {\it i.e.\/}\ the kernel of the multiplication map $\mu_B:B\otimes B\to B$
is $B^e$-projective; \cite[Lemma 2.3]{Schelter:SmoothAlgebras}. Such algebras are also
referred as \emph{quasi-free}; \cite{CuntzQuillen:NonsingularityI}. Among the most basic examples of smooth algebras, there are $k[x]$ and $k[x,x^{-1}]$, see \cite[Ex. 3.4.3]{Loday-book}. Similarly, we call an
algebra \emph{$m$-smooth} if $\Omega_{B|k}$ has Hochschild homological dimension $m$. As such, an
ordinary smooth algebra is $0$-smooth. In \cite{vdBer98}, a smooth algebra is defined to an
algebra $B$ whose projective $B^e$-resolution is finite, and contains only finitely generated
$B^e$-modules. The quintessential example is $B=S(V)$; the polynomial algebra with
$m+1=\dim_k(V)$-indeterminates which is $m$-smooth.
\begin{lemma}\label{lem:aux}
Let $B$ be a smooth algebra, and $V$ a $B$-bimodule. Then,
\[
H_1(B,V) \cong V^B \cong H^0(B,V).
\]
\end{lemma}
\begin{proof}
Since $B$ is smooth
\[
0 \to \ker(\mu_B)\to B\otimes B \to B \to 0
\]
is a $B^e$-projective resolution of $B$. Then we immediately see that
\[
H_1(B,V)\cong \ker\left(V\otimes_{B^e}\ker(\mu_B)\to V\right).
\]
The claim then follows from $\ker(\mu_B)$ being generated, as a $B^e$-module, by the elements of the form $1\otimes x - x\otimes 1\in B^e$.
\end{proof}
\begin{theorem}\label{SmoothExtensions}
Let $A$ and $B$ be two algebras with $B$ being smooth, and let $R\colon B\otimes A\to A\otimes B$ be an invertible distributive law. Then, for any $A\#_R B$-bimodule $V$,
\[
H_n(A\#_R B,V) \cong H_n({\rm CH}_*(A,V)_B)\oplus H_{n-1}({\rm CH}_*(A,V)^B)
\]
for all $n\geq 0$.
\end{theorem}
\begin{proof}
In view of the smoothness of $B$, Theorem~\ref{thm:HochschildOfSmashProducts} implies that the 1st page $'E^1_{p,q}$ of the spectral sequence of Theorem \ref{thm:HochschildOfSmashProducts} consists of two columns at $p=0$ and $p=1$. Then by \eqref{eqn-H_0}
\[
'E^2_{0,q} = H_q(H_0(B,{\rm CH}_\ast(A,V))) \cong H_q({\rm CH}_\ast(A,V)_B) ,
\]
and by Lemma~\ref{lem:aux}
\[
'E^2_{1,q} = H_q(H_1(B,{\rm CH}_\ast(A,V)))\cong H_q({\rm CH}_\ast(A,V)^B)
\]
as we wanted to show.
\end{proof}
\section{Computations}\label{sect:calculation}
\subsection{Galois extensions}\label{Galois-ext}~
Let $K/k$ be a finite Galois extension, and let $G$ be the Galois group of this extension. Accordingly, we have the $K$-algebra $K\#G$, and from Theorem~\ref{thm:amenable} we obtain
\[
H_n(K\# G,K) = H_n(K)^G.
\]
The homology of $K$, on the other hand, is given by
\[
HH_n(K) = \begin{cases}
K & \text{ if } n= 0, \\
0 & \text{ otherwise},
\end{cases}
\]
when regarded as a $K$-algebra. Finally $G$ being the Galois group of the extension $K/k$, we have $K^G=k$. Hence,
\[
H_n(K\# G,K) =
\begin{cases}
k & \text{ if } n=0,\\
0 & \text{ otherwise.}
\end{cases}
\]
\subsection{Group rings with faithful characters}\label{subsect-group-ring}~
Let $G$ be a group and let $\sigma\colon G\to k^\times$ be a character. Let us define
\begin{equation}\label{R-distr-law}
R\colon k[x,x^{-1}]\otimes k[G]\to k[G]\otimes k[x,x^{-1}], \qquad R(x^n\otimes g) := \sigma(g)^n g\otimes x^n
\end{equation}
for any $n\geq 0$, and any $g\in G$. Then \eqref{R-distr-law} obeys the diagram \eqref{DistributiveLaw}, and we have the smash biproduct algebra $k[G]\# k[x,x^{-1}]$. Now, since $k[x,x^{-1}]$ is smooth, we have by Theorem \ref{SmoothExtensions}
\begin{align*}
H_n(k[G]\# k[x,x^{-1}])
\cong & H_n\Big({\rm CH}_\ast(k[G],k[G]\# k[x,x^{-1}])_{k[x,x^{-1}]}\Big)\\
& \oplus H_{n-1}\Big({\rm CH}_\ast(k[G],k[G]\# k[x,x^{-1}])^{k[x,x^{-1}]}\Big).
\end{align*}
On the other hand,
\[ \Big[x, g_1\otimes\cdots\otimes g_n\otimes (g_0\otimes x^{\ell-1})\Big] =
(\sigma(g_0g_1\cdots g_n) - 1)g_1\otimes\cdots\otimes g_n\otimes (g_0\otimes x^{\ell} )\]
Now, assume $\sigma$ is faithful, {\it i.e.\/}\ $\sigma(g) = 1$ if and only if $g=e$. In
view of the faithfulness of the character we get
\[ {\rm CH}_\ast(k[G],k[G]\# k[x,x^{-1}])_{k[x,x^{-1}]} = 0, \]
whereas
\[
{\rm CH}_\ast(k[G],k[G]\# k[x,x^{-1}])^{k[x,x^{-1}]} \cong {\rm CH}_\ast^{(e)}(k[G])\otimes k[x,x^{-1}],
\]
where
\begin{align*}
{\rm CH}_n^{(e)} (k[G])
:= & \left<g_1 \ot\cdots\ot g_n \otimes g_0 \in {\rm CH}_n(k[G])\mid g_1\cdots g_ng_0 = e\right>.
\end{align*}
In order to compute the homology of this subcomplex we shall need the following lemma which
goes back to Eilenberg and MacLane~\cite{EilenbergMacLane:MacLaneHomology}.
(See~\cite{SiegelWitherspoon:Hochschild} for a history of Hochschild (co)homology of group
algebras.)
\begin{lemma}\label{technical2}
Let $G$ be an arbitrary group and consider ${\rm CH}_*(k[G])$. Now, consider the subcomplex
${\rm CH}^{(e)}_*(k[G])$ generated by homogeneous tensors of the form
$g_0\otimes\cdots\otimes g_n$ such that the product $g_0\cdots g_n$ is the unit element.
Then
\[ H_\ast({\rm CH}_\ast^{(e)}(k[G]) \cong H_\ast(k[G],k). \]
\end{lemma}
\begin{proof}
The morphism of complexes ${\rm CH}_n(k[G],k)\to {\rm CH}_n^{(e)}(k[G],k[G])$, given by
\[ g_1\otimes\cdots\otimes g_n \mapsto g_1\otimes\cdots\otimes g_n \otimes g_n^{-1}\cdots g_1^{-1}, \]
is an isomorphism.
\end{proof}
Accordingly, we have
\[
H_\ast\Big({\rm CH}_\ast(k[G],k[G]\# k[x,x^{-1}])^{k[x,x^{-1}]} \Big) \cong H_\ast(k[G],k)\otimes k[x,x^{-1}],
\]
and hence
\[
HH_n(k[G]\# k[x,x^{-1}]) \cong H_{n-1}(G,k)\otimes k[x,x^{-1}].
\]
Let $G$ be a finitely generated abelian group of the form $G = G^f\times G^t$ where $G^t$
is the maximal torsion subgroup and $G^f$ is the maximal torsion-free subgroup of $G$.
Since $H_n(G^t,k)=0$ for $n\geq 1$, we see that the higher homology is determined $G^{f}$.
Let $a$ be the free rank of $G$. Then
\[ HH_n(k[G]\# k[x,x^{-1}]) \cong k^{\binom{a}{n-1}}\otimes k[x,x^{-1}]. \]
\subsection{Smash products of Lorentz polynomials}\label{subsect-quantum-tori}~
Below, we shall consider the Hochschild homology of the smash products of polynomial algebra $k[x]$ and the Laurent polynomial algebra $k[x,x^{-1}]$ in various combinations, but always with the particular distributive law given by
\[
R(y^j\otimes x^i) = q^{ij} x^i\otimes y^j
\]
for some fixed $q\in k^\times$ which is not a root of unity.
\subsubsection{The quantum plane}~
Let us begin with $A=k[x]$ and $B=k[y]$. In this case, for any $x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i y^j \in {\rm CH}_m(A,A\#_R B)$, we have
\begin{align*
\left[y^j,x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i \right]
= & \left(q^{(i_1+\ldots +i_m + i)j}-1\right)x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i y^{j},
\end{align*}
that is,
\[ {\rm CH}_m(A,A\#_R B)_B \cong {\rm CH}_m(k,B/k) \oplus {\rm CH}_m(A) \]
whereas,
\[
{\rm CH}_m(A,A\#_R B)^B \cong{\rm CH}_m(k,B).
\]
Accordingly, we see that
\begin{align*}
HH_n(A\#_R B)
\cong & H_n({\rm CH}_*(A,A\#_R B)_B) \oplus H_{n-1}({\rm CH}_*(A,A\#_R B)^B) \\
\cong & \begin{cases}
k[x]\oplus yk[y] & \text{ if } n=0,\\
k[x]\oplus k[y] & \text{ if } n=1,\\
0 & \text{ if } n\geq 2.
\end{cases}
\end{align*}
\subsubsection{The quantum cylinder}~
We now consider $A=k[x]$ and $B=k[y,y^{-1}]$. In this case, for any $x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i y^j \in {\rm CH}_m(A,A\#_R B)$ so that $i_1+\ldots +i_m + i \neq 0$, we may choose $j_1,j_2 \in \B{Z}$ such that $j_1+j_2 = j$ and that $j_1 \neq 0$ to get
\begin{align*
\left[y^{j_1},x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^iy^{j_2} \right]
= & \left(q^{(i_1+\ldots +i_m + i)j_1}-1\right)x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i y^{j} ,
\end{align*}
that is,
\[
{\rm CH}_m(A,A\#_R B)_B \cong {\rm CH}_m(k,B) \cong {\rm CH}_m(A,A\#_R B)^B.
\]
Accordingly, we see that
\[
HH_n(A\#_R B) \cong
\begin{cases}
k[y,y^{-1}]& \text{ if } n=0,1,\\
0 & \text{ if } n\geq 2.
\end{cases}
\]
\subsubsection{The quantum torus}~
We continue with $A=k[x,x^{-1}]$ and $B=k[y,y^{-1}]$. In this case, for any $x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i y^j \in {\rm CH}_m(A,A\#_R B)$ so that $i_1+\ldots +i_m + i \neq 0$, we may choose $j_1,j_2 \in \B{Z}$ so that $j_1+j_2 = j$ and that $j_1 \neq 0$ to get
\begin{align*
\left[y^{j_1},x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^iy^{j_2} \right]
= & \left(q^{(i_1+\ldots +i_m + i)j_1}-1\right)x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^i y^{j} ,
\end{align*}
that is,
\[
{\rm CH}_m(A,A\#_R B)_B = {\rm CH}_m(A,A\#_R B)^B
\cong \left< x^{i_1} \ot\cdots\ot x^{i_m} \otimes x^iy^j \mid i_1 + \ldots + i_m + i = 0\right>,
\]
Just like we did in Lemma we next identify the above complexes with the complex computing the homology of the group algebra $k[\B{Z}]$ with coefficients in $k[y,y^{-1}]$ via
\begin{align*}
{\rm CH}_n(A,A\#_R B)_B
\cong & {\rm CH}_n(A,A\#_R B)^B \to {\rm CH}_n(k[\B{Z}],k[y,y^{-1}]), \\
& x^{\ell_1} \ot\cdots\ot x^{\ell_n} \otimes \left(x^{-\ell_1 - \ldots -\ell_n} \otimes y^j\right) \mapsto \ell_1 \ot\cdots\ot \ell_n \otimes y^j,
\end{align*}
where the right $\B{Z}$-module structure on $k[y,y^{-1}]$ is given by $y^j \triangleleft i := q^{ij}$, and the left $\B{Z}$-module structure is trivial. As such,
\begin{align*}
& HH_n(A\#_R B) \cong H_n\Big({\rm CH}_\ast(A,A\#_R B)_B\Big) \oplus H_{n-1}\Big({\rm CH}_\ast(A,A\#_R B)^B\Big) \cong \\
&\hspace{3cm} H_n(k[\B{Z}],k[y,y^{-1}]) \oplus H_{n-1}(k[\B{Z}],k[y,y^{-1}]).
\end{align*}
It is then evident, in view of \cite[Ex.III.1.1]{Brown:CohomologyOfGroups}, that
\[
HH_n(A\#_R B) \cong
\begin{cases}
k & \text{ if } n=0,\\
k \oplus k & \text{ if } n=1,\\
0 & \text{ if } n\geq 2.
\end{cases}
\]
\subsection{The multiparametric affine space}\label{subsect-multip-aff-sp}~
The multiparametric affine space $S(X_\nu,\Lambda)$ defined in \cite[Subsect. 3.1]{GuccGucc97}
is the algebra generated by $X_\nu:=\{x_1, \ldots, x_\nu\}$ subject to the relations
\begin{equation}\label{QuantumPlane1}
x_jx_i = q_{i,j}x_ix_j,
\end{equation}
where for $1 \leq i,j \leq \nu$, and the structure constants $\Lambda:=(q_{i,j})$ form a set
of nonzero elements in $k$ satisfying
\begin{equation}\label{QuantumPlane2}
q_{i,i} = 1, \qquad q_{i,j}q_{j,i} = 1,
\end{equation}
for all $i < j$. For the sake of simplicity, we are going to drop $\Lambda$ from the notation.
Now, we can write $S(X_\nu)$ as an iterated sequence of smash biproducts
\[ S(X_\nu) = \underbrace{k[x_1]\# \cdots \# k[x_n]}_{\text{$\nu$-times}} = S(X_{\nu-1})\#
k[x_\nu] \] In view of Theorem \ref{SmoothExtensions}, we have
\begin{align*}
HH_n(S(X_\nu))
\cong & H_n\Big({\rm CH}_\ast\big(S(X_{\nu-1}),S(X_\nu)\big)_{k[x_\nu]}\Big)
\oplus H_{n-1}\Big({\rm CH}_\ast\big(S(X_{\nu-1}),S(X_\nu)\big)^{k[x_\nu]}\Big).
\end{align*}
\subsubsection{The free case}~
Now, let us assume for $1 \leq i,j \leq \nu$ that $\Lambda$ generates a free abelian group of rank $\nu(\nu-1)/2$ in $k^\times$. Then,
\[
{\rm CH}_\ast\big(S(X_{\nu-1}),S(X_\nu)\big)_{k[x_\nu]} \cong {\rm CH}_\ast\big(S(X_{\nu-1}),S(X_{\nu-1})\big),
\]
and
\[
{\rm CH}_\ast\big(S(X_{\nu-1}),S(X_\nu)\big)^{k[x_\nu]} \cong {\rm CH}_\ast\big(k,k[x_\nu]\big).
\]
It is then immediate to see that
\[
HH_n(S(X_\nu)) \cong
\begin{cases}
k\oplus \bigoplus_{i=1}^\nu x_ik[x_i] & \text{ if } n=0, \\
\bigoplus_{i=1}^\nu\,k[x_i] \otimes x_i & \text{ if } n=1, \\
0 & \text{ if } n\geq 2,
\end{cases}
\]
which is the case $r=0$ in \cite[Thm. 4.4.1]{Richard:QuantumWeyl}.
\subsubsection{The non-free case}~
Hochschild complex admits the $\B{N}^\nu$-grading on $S(X_\nu,\Lambda)$ where we have
\[
HH_\ast(S(X_\nu)) = \bigoplus_{(m_1,\ldots,m_\nu) \in \B{N}^\nu}\,HH_\ast^{(m_1,\ldots,m_\nu)}(S(X_\nu)),
\]
The same grading applies to both ${\rm CH}_\ast\big(S(X_{\nu-1}),S(X_\nu)\big)_{k[x_\nu]}$ and
${\rm CH}_\ast\big(S(X_{\nu-1}),S(X_\nu)\big)^{k[x_\nu]}$. Then it suffices to consider only the
components with multi-degree $(m_1,\ldots,m_\nu) \in \B{N}^\nu$.
Now, consider the cone $C$ in the lattice $ \B{N}^\nu$ of elements
$(m_1,\ldots,m_\nu) \in \B{N}^\nu$ with the property that
\begin{equation}\label{eqn-prod}
\prod_{i=1}^\nu \, q_{j,i}^{m_i} = 1
\end{equation}
for every $1\leq j\leq \nu$. This translates into
\begin{equation}\label{eqn-commutation}
x_j \cdot \left(x_1^{m_1}\ldots x_\nu^{m_\nu}\right)
= \left(x_1^{m_1}\ldots x_\nu^{m_\nu}\right)\cdot x_j
\end{equation}
for $1\leq j\leq \nu$. Hence, it follows at once from \eqref{eqn-prod} and
\eqref{eqn-commutation} that
\[
\dim_k HH_n^{(m_1,\ldots,m_\nu)}(S(X_\nu)) =
\begin{cases}
\binom{h(m_1,\ldots,m_\nu)}{n} & \text{ if } (m_1,\ldots,m_\nu)\in C\\
0 & \text{ otherwise,}
\end{cases}
\]
where we define
\[ h(m_1,\ldots,m_\nu) = \# \{ m_i\mid m_i > 0 \}, \] i.e. it counts the number of $m_i>0$.
Our result agrees with \cite[Thm. 3.1.1]{GuccGucc97}, and also with \cite[Thm. 6.1]{Wamb93}.
\subsection{Quantum complete intersections}\label{subsect-quant-comp-int}~
Given two integers $a,b\geq 2$, and $q \in k$ which is not a root of unity; let $C_{a,b}$ be
the quotient of the algebra $k\{x,y\}$ generated by two noncommuting indeterminates divided by
the two sided ideal given by the relations
\begin{equation}
x^a,\quad yx-qxy, \quad y^b
\end{equation}
as given in \cite{BerErd08}. We note that the choice of
$a=b=2$ yields the quantum exterior algebra \cite{Bergh:QuantumExteriorAlgebras}. A linear
basis of this algebra is given by monomials of the form $x^iy^j$ where $0\leq i < a$ and
$0\leq j < b$.
Now, consider the truncated polynomial algebras $T_a := k[x]/\langle x^a\rangle$, and
$T_b := k[y]/\langle y^b\rangle$ determined by $a,b\geq 2$. Then
\[
R:T_b\otimes T_a \to T_a \otimes T_b; \qquad R(y^j\otimes x^i) := q^{ij}x^i\otimes y^j,
\]
an invertible distributive law. Thus we have $C_{a,b} \cong T_a \#_R T_b$.
The Hochschild homology of a truncated polynomial algebra $T_a:= k[x]/\left<x^a\right>$ has
the periodic resolution $(P_*,\partial_*)$ where $P_n = T_a\otimes T_a$ and
\begin{equation*}
\partial_n =
\begin{cases}
1\otimes x- x\otimes 1 & \text{ if $n$ is odd,}\\
\displaystyle\sum_{i=0}^{a-1} x^i\otimes x^{a-1-i} & \text{ if $n$ is even.}
\end{cases}
\end{equation*}
for $n\geq 1$ \cite[Sect. 5.9]{Kassel:CohomologyOfAssociativeAlgebras}, see
also~\cite[E.4.1.8]{Loday-book}. Tensoring (over $T_a \otimes T_a$) with a $T_a$-bimodule $U$,
then $H_n(T_a,U)$ appears as the homology of the complex
\begin{equation}\label{eqn-truncated-poly-alg-comp}
\xymatrix{
\cdots \ar[r] & U \ar[r]^{\partial_4} & U \ar[r]^{\partial_3} &U \ar[r]^{\partial_2} & U \ar[r]^{\partial_1} & U \ar[r] & 0,
}
\end{equation}
where for $n \geq 1$,
\begin{equation}\label{eqn-truncated-poly-alg-diff}
\partial_n(u) =
\begin{cases}
xu-ux & \text{ if $n$ is odd,} \\
\displaystyle\sum_{i=0}^{a-1}x^iu x^{a-1-i} & \text{ if $n$ is even.}
\end{cases}
\end{equation}
More precisely, for any $m\geq 1$ and
$x^{i_1}\ot\cdots\ot x^{i_m}\otimes x^iy^j \in {\rm CH}_m(T_a,T_a\#_R T_b)$ we have
\begin{align*}
\partial_n\left(x^{i_1}\ot\cdots\ot x^{i_m}\otimes x^iy^j\right) =
\begin{cases}
\left(q^{i_1+\cdots+i_m + i} - 1\right)\,x^{i_1}\ot\cdots\ot x^{i_m}\otimes x^iy^{j+1} & \text{ if $n$ is off,}\\
\sum_{s=0}^{b-1}\, q^{(i_1+\cdots+i_m + i)s} \,x^{i_1}\ot\cdots\ot x^{i_m}\otimes x^iy^{b-1+j} & \text{ if $n$ is even.}
\end{cases}
\end{align*}
We now see that $HH_q(T_a) \cong k^{a-1}$ for $q>0$, as in \cite[Prop. 5.4.15]{Loday-book}
Accordingly, by Theorem~\ref{thm:HochschildOfSmashProducts} we get
\[
H_p(T_b,{\rm CH}_\ast(T_a,C_{a,b})) =\begin{cases}
{\rm CH}_\ast(k,T_b/k)\oplus {\rm CH}_\ast(T_a), & \text{ if } p=0, \\
{\rm CH}_\ast(k,T_b/\langle y^{b-1}\rangle) & \text{ if $p$ is odd,} \\
{\rm CH}_\ast(k,T_b/k) & \text{ if $p>0$ is even.}
\end{cases}
\]
Then the 1st page of the spectral sequence~\eqref{eqn-E1} appears as
\[ HH_\ast(C_{a,b}) \Leftarrow {}'E^1_{p,q} \cong
\begin{cases}
{\rm CH}_q(k,T_b/k)\oplus {\rm CH}_q(T_a) & \text{ if } p=0,\\
{\rm CH}_q(k,T_b/\langle y^{b-1}\rangle) & \text{ if $p$ is odd,} \\
{\rm CH}_q(k,T_b/k)& \text{ if $p>0$ is even.}
\end{cases}
\]
Finally, since $H_q(k,T_b/\langle y^{b-1}\rangle) = 0 = H_q(k,T_b/k)$ for $q > 0$, the
vertical homology yields
\[
{}'E^2_{p,q} \cong
\begin{cases}
T_b/k \oplus k^a & \text{ if } p=q=0\\
k^{a-1} & \text{ if $p=0$ and $q>0$}, \\
T_b/\langle y^{b-1}\rangle & \text{ if $p$ is odd and $q=0$,}\\
T_b/k & \text{ if $p>0$ is even and $q=0$},\\
0 & \text{ otherwise.}
\end{cases}
\]
In other words,
\[
\dim_k HH_n(C_{a,b})
= \begin{cases}
a+b-1 & \text{ if } n =0, \\
a+b-2 & \text{ if } n \geq 1,
\end{cases}
\]
as in \cite[Thm. 3.1]{BerErd08}.
\subsection{The deformed (completed) Weyl algebras}\label{sect:WeylAlgebras}~
Let $\G{g}$ be a finite dimensional Lie algebra, and $V$ be a vector space. Let also $S(V)$ denote the symmetric algebra on $V$, and
$\widehat{S(V)}$ be the completed symmetric algebra over $V$, {\it i.e.\/}\ the algebra of
formal power series in $\dim(V)$-many variables. We note that the algebra $\widehat{S(V)}$ is isomorphic to the dual algebra
$S(V)^\vee$ of $S(V)$ viewed as the coalgebra of polynomials over $\dim(V)$-many commuting
variables.
Let us now recall the $(\G{g}, D)$-deformed Weyl algebras from
\cite{DuroMeljSamsSkod07,MeljSkod07}. Given $ D:\G{g} \to {\rm Hom}(V,\widehat{S(V)})$, extending it to a morphism $ D:\G{g} \to {\rm Der}(\widehat{S(V)})$ of Lie algebras~\cite[Subsect. 1.2]{MeljSkod07}, one arrives at the Hopf-cross product,
$A_{\G{g}, D} := \widehat{S(V)} \rtimes U(\G{g})$, called the $(\G{g}, D)$-deformed (completed) Weyl algebra.
We are going to identify the (continuous) Hochschild homology of the algebra
$A_{\G{g}, D}$ with the homology of its dense subalgebra $A^{\rm pol}_{\G{g}, D} := S(V) \rtimes U(\G{g})$.
\begin{proposition}\label{Completion}
If $ D(g)(v)\in \Bbbk\left<1\right>\oplus V$, for any $g\in \G{g}$ and
any $v\in V$, then there is an isomorphism of the form
$HH_\ast(A^{\rm pol}_{\G{g}, D}) \cong HH^{cont}_\ast (A_{\G{g}, D})$, where the right hand side refers to the continuous Hochschild homology.
\end{proposition}
\begin{proof}
It follows at once from the hypothesis that, the distributive law between $\widehat{S(V)}$ and $U(\G{g})$ restricts to a
degree preserving distributive law between $S(V)$ and $U(\G{g})$. As such, we have a subcomplex
\[
{\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m)}\subseteq {\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)
\]
of terms whose total degree in $S(V)$ is less than or equal to $m\in\B{N}$. This collection yields a projective system of complexes together with the natural epimorphisms
\begin{equation}\label{eq:projective-system}
{\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m+1)} \to {\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m)}
\end{equation}
The system satisfies the Mittag-Leffler condition~\cite{Dimi04,Emma96} by definition. Then we conclude from
\cite[Prop. 2 and Thm. 5]{Emma96} that
\begin{equation}\label{inv-lim}
HH^{cont}_\ast(A_{\G{g}, D})
\cong \lim_{\underset{m}{\longleftarrow}} HH_\ast \left(S(V)\rtimes U(\G{g})\right)_{(m)}
\end{equation}
where the left hand side is the continuous Hochschild homology of $A_{\G{g}, D}$.
On the other hand, since the distributive law is degree preserving, the collection
${\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m)}$ forms also an injective system of complexes via the natural embeddings
\begin{equation}\label{eq:injective-system}
{\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m)} \to {\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m+1)},
\end{equation}
which results in
\[
{\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)
= \lim_{\underset{m}{\longrightarrow}} {\rm CH}_\ast \left( S(V)\rtimes U(\G{g})\right)_{(m)} .
\]
Since the homology $HH_\ast \left( S(V)\rtimes U(\G{g})\right)$ is bounded and finite dimensional at every degree, by Corollary~\ref{SpecialCase}, we see that there is an index $N \in \B{N}$ such that for every $m\geq N$ the natural injections in \eqref{eq:injective-system}, and therefore the natural
projections in \eqref{eq:projective-system}, induce quasi-isomorphisms. In other words,
\[
HH_\ast \Big( S(V)\rtimes U(\G{g})\Big) \cong HH^{cont}_\ast \Big( \widehat{S(V)}\rtimes U(\G{g})\Big)
\]
as we wanted to show.
\end{proof}
In particular, for the Weyl algebra $A^{\rm pol}_{2n}$, and its completion $A_{2n}$, in view of the fact that the Hochschild homology of $A^{\rm pol}_{2n}$ is known, see for
example~\cite[Sect. 5.10]{Kassel:CohomologyOfAssociativeAlgebras}, we get
\cite[Thm. 2.1]{FeigFeldShoi05} and \cite[Thm. 4]{FeigTsyg83} in
\begin{equation}
HH^{cont}_m(A_{2n}) \cong HH_m(A^{\rm pol}_{2n}) =
\begin{cases}
k & \text{ if } m= 2n\\
0 & \text{ otherwise}
\end{cases}
\end{equation}
for any $n\geq 1$ and any $m\geq 0$, as an immediate corollary to Proposition~\ref{Completion}.
\subsection{Quantum Weyl algebras}\label{subsect-quant-Weyl}~
In this subsection we are going to discuss the Hochschild homology of quantum Weyl algebras
$S(X_\mu,Y_\nu,\Lambda)$ generated by $X_\mu:=\{x_1,\ldots,x_\mu\}$ and
$Y_\nu:=\{y_1,\ldots,y_\nu\}$, subject to the relations
\begin{align}\label{comm-rel}
x_ix_j - q_{i,j}x_jx_i, \quad
y_iy_j - q_{i,j}y_jy_i, \quad
y_jx_i - q_{i,j}x_iy_j, \quad
x_iy_i - y_ix_i + 1
\end{align}
for all $i\neq j$ where $1\leq i\leq \mu$, $1\leq j\leq \nu$, and $\Lambda:=(q_{i,j})$ of
\eqref{QuantumPlane2}. Since there is a an isomorphism of algebras of the form
$S(X_\mu,Y_\nu,\Lambda) \to S(X_\nu,Y_\mu,\Lambda^t)$ given by
\begin{align*} & x_i\mapsto - y_i, \qquad y_j\mapsto x_j,
\end{align*}
we may assume, without loss of generality, that $\nu\geq\mu$. We also assume that $\Lambda$
generates a free abelian group of rank $\nu(\nu-1)/2$ in $k^\times$ as
in~\cite{Richard:QuantumWeyl}. For simplicity, from now on we are going to drop the parameter
matrix $\Lambda$ from the notation.
\subsubsection{The case $\mu=\nu$}~
Let $\nu=\mu$, and let $A_2(x_i,y_i)$ be the copy of the ordinary Weyl algebra in
$S(X_\mu,Y_\mu)$ generated by $x_i$ and $y_i$ for $1\leq i,j\leq \mu$. We recall also the $A_2(x_1,y_1)^e$-resolution
\[ 0 \to A_2(x_1,y_1)^e \xra{\partial_2} A_2(x_1,y_1)^e\oplus A_2(x_1,y_1)^e\xra{\partial_1} A_2(x_1,y_1)^e\to 0, \]
where
\[ \partial_2(a\otimes b) = (ax_1\otimes b-a\otimes x_1 b)e_1 - (ay_1\otimes b - a\otimes y_1 b)e_2, \]
and
\[ \partial_1((a_1\otimes b_1)e_1 + (a_2\otimes b_2)e_2)
= a_1y_1\otimes b_1 - a_1\otimes y_1 b_1 + a_2x_1\otimes b_2 - a_2\otimes x_1 b_2,
\]
for $A_2(x_1,y_1)$, \cite[Sect. 5.10]{Kassel:CohomologyOfAssociativeAlgebras}. Since $S(X_\mu,Y_\mu) \cong S(X_{\mu-1},Y_{\mu-1}) \# A_2(x_1,y_1)$, in view of
Theorem~\ref{thm:HochschildOfSmashProducts}, we have
\[ HH_*(S(X_\mu,Y_\mu)) \Leftarrow {}'E^1_{p,q} =
H_p\Big(A_2(x_1,y_1),{\rm CH}_q\big(S(X_{\mu-1},Y_{\mu-1}),S(X_\mu,Y_\mu)\big)\Big).
\]
Now we note that
\begin{align*}
& \left[y_1 , x_2^{a_2}y_2^{b_2}\ot\cdots\ot x_\mu^{a_\mu}y_\mu^{b_\mu} \otimes
x_1^ay_1^b \right] \\
& = (q_{1,2}^{B_2-A_2}\cdots q_{1,\mu}^{B_\mu-A_\mu}-1)
x_2^{a_2}y_2^{b_2}\ot\cdots\ot x_\mu^{a_\mu}y_\mu^{b_\mu} \otimes x_1^a y_1^{b+1}\\
& \quad + aq_{1,2}^{B_2-A_2}\cdots q_{1,\mu}^{B_\mu-A_\mu} x_2^{a_2}y_2^{b_2}\ot\cdots\ot x_\mu^{a_\mu}y_\mu^{b_\mu} \otimes x_1^{a-1}y_1^b,
\end{align*}
and that
\begin{align*}
& \left[x_1, x_2^{a_2}y_2^{b_2}\ot\cdots\ot x_\mu^{a_\mu}y_\mu^{b_\mu}
\otimes x_1^ay_1^b\right]\\
& = (q_{1,2}^{A_2-B_2}\cdots q_{1,\mu}^{A_\mu-B_\mu}-1)
x_2^{a_2}y_2^{b_2}\ot\cdots\ot x_\mu^{a_\mu}y_\mu^{b_\mu} \otimes x_1^{a+1} y_1^b\\
&\quad + bq_{1,2}^{A_2-B_2}\cdots q_{1,\mu}^{A_\mu-B_\mu}x_2^{a_2}y_2^{b_2}\ot\cdots\ot x_\mu^{a_\mu}y_\mu^{b_\mu} \otimes x_1^ay_1^{b-1},
\end{align*}
where $A_i$ and $B_j$, $2\leq i,j \leq \mu$, denote the total degree of $x_i$ and $y_j$ respectively. Hence, upon tensoring the above resolution with the complex ${\rm CH}_*(S(X_{\mu-1},Y_{\mu-1}), S(X_\mu,Y_\mu))$ we see that
\[ {}'E^1_{p,q} \cong
\begin{cases}
{\rm CH}_q(S(X_{\mu-1},Y_{\mu-1}), S(X_\mu,Y_\mu))^{A_2(x_1,y_1)} & \text{ if } p=2,\\
0 & \text{ otherwise.}
\end{cases}
\]
The invariant subcomplex consists of the tensors whose total degree of $x_i$ terms is
equal to the total degree of $y_i$ terms for each $2 \leq i \leq \mu$, with no $x_1$ or $y_1$ terms. Moreover, this subcomplex is a direct summand (as differential graded modules) of the
Hochschild complex ${\rm CH}_q(S(X_\mu,Y_\mu))$, as such, we see that
\begin{align*}
{}'E^2_{2,q} \cong H_q\big({\rm CH}_\ast(S(X_{\mu-1},Y_{\mu-1}), S(X_\mu,Y_\mu))^{A_2(x_1,y_1)}\big) =
HH_q\left(S(X_{\mu-1},Y_{\mu-1})\right)^{A_2(x_1,y_1)} .
\end{align*}
In other words,
\[
HH_n(S(X_\mu,Y_\mu)) = HH_{n-2}\left(S(X_{\mu-1},Y_{\mu-1})\right)^{A_2(x_1,y_1)}.
\]
Reducing recursively then, we obtain
\begin{equation}
HH_n(S(X_\mu,Y_\mu)) \cong \begin{cases}
k & \text{ if } n = 2\mu, \\
0 & \text{ otherwise}.
\end{cases}
\end{equation}
\subsubsection{The case $\nu>\mu$}~
Considering
\[ S(X_\mu,Y_\nu) \cong S(X_\mu,Y_{\nu-1}) \#_R k[y_\nu], \]
Theorem \ref{SmoothExtensions} yields
\begin{align*}
HH_n\big(S(X_\mu, & Y_\nu)\big)\\
\cong & H_n\Bigg({\rm CH}_\ast\bigg(S(X_\mu, Y_{\nu-1}), S(X_\mu,Y_\nu)\bigg)_{k[y_\nu]}\Bigg) \oplus H_{n-1}\Bigg({\rm CH}_\ast\bigg(S(X_\mu, Y_{\nu-1}), S(X_\mu,Y_\nu)\bigg)^{k[y_\nu]}\Bigg).
\end{align*}
It is evident from the commutation relations \eqref{comm-rel} that
\begin{align*}
{\rm CH}_\ast \bigg(S(X_\mu, & Y_{\nu-1}), S(X_\mu,Y_\nu)\bigg)_{k[y_\nu]}\\
\cong & {\rm CH}_\ast\bigg(S(X_\mu, Y_{\nu-1}), S(X_\mu,Y_{\nu-1})\bigg)
\oplus {\rm CH}_\ast\bigg(k, k[y_\nu]/k\bigg).
\end{align*}
and that the $k[y_\nu]$-invariant subcomplex ${\rm CH}_\ast\big(S(X_\mu, Y_{\nu-1}), S(X_\mu,Y_\nu)\big)^{k[y_\nu]}$
may be written as a direct sum of ${\rm CH}_\ast\big(k, k[y_\nu]\big)$
and a complex with the same total degree of $x_i$'s and $y_i$'s, for each $1 \leq i \leq \mu$. However, since these complexes are accompanied with an extra $\otimes y_{\nu}$, the reduction performed in the previous subsection reveals that the homology (on the next page) of the latter is trivial. As a result,
\begin{align*}
HH_n(S(X_\mu,Y_\nu))
\cong & HH_n(S(X_\mu,Y_{\nu-1})) \oplus H_n\Big({\rm CH}_\ast\big(k, k[y_\nu]/k\big)\Big)\\
& \oplus H_{n-1}\Big({\rm CH}_\ast\big(k, k[y_\nu]\otimes y_\nu\big)\Big).
\end{align*}
Hence, inductively, we see that
\[
HH_n(S(X_\mu,Y_\nu))
\cong \begin{cases}
\bigoplus_{j=\mu+1}^\nu\,y_jk[y_j]
& \text{ if } n=0, \\
\bigoplus_{j=\mu+1}^\nu\,k[y_j] \otimes y_j
& \text{ if } n=1, \\
k & \text{ if } n= 2\mu, \\
0 & \text{ otherwise},
\end{cases}
\]
as in \cite[Thm. 4.4.1]{Richard:QuantumWeyl}.
\subsection{The algebra of quantum matrices}\label{subsect-quant-matrix}~
Given $q \neq 1$, the algebra $M_q(2)$ of quantum matrices is defined in \cite[Def. IV.3.2]{Kassel-book} as the quotient algebra generated by $\{a,b,c,d\}$ subject to the relations
\begin{align*}
ba & = qab, & db & = q bd, \\
ca & = qac, & dc & = qcd, \\
bc & = cb, & ad - da & = (q^{-1} - q)bc.
\end{align*}
We note also that the center of $M_q(2)$ is generated by the quantum determinant $D_q := ad - q^{-1}bc \in M_q(2)$, see \cite[Thm. 1.6]{NoumYamaMima93}).
More importantly, $M_q(2)$ is a tower of Ore extensions $A_1\subseteq A_2\subseteq A_3\subseteq M_q(2)$ given by
\[ A_1 = k[a], \quad A_2 \cong A_1[b,\alpha_1,0], \qquad A_3 \cong A_2[c,\alpha_2,0], \qquad M_q(2) \cong A_3[d,\alpha_3,\delta],
\]
see, for instance, \cite[Sect. IV.4]{Kassel-book}. The structure morphisms $\alpha_i\colon A_i\to A_i$ are given as
\begin{align*}
\alpha_1(a) & := qa, \\
\alpha_2(a) & := qa, & \alpha_2(b) &:= b, \\
\alpha_3(a) & := a, & \alpha_3(b) &:= qb, & \alpha_3(c) &:= qc.
\end{align*}
for $i=1,2,3$, and $\delta\colon A_3\to A_3$ is given as
\begin{align*}
\delta(b^jc^k) & := 0 &
\delta(a^ib^jc^k) & := (q - q^{-1})\frac{1-q^{2i}}{1-q^2}a^{i-1}b^{j+1}c^{k+1}.
\end{align*}
Let us also recall from \cite[Lemma 1]{GeLiuSun92} the relations
\begin{equation}\label{eqn-ad-da-commutation}
d^na = ad^n - q\left(1-q^{-2n}\right)bcd^{n-1}, \qquad da^n = a^nd + q\left(q^{-2n}-1\right)a^{n-1}bc.
\end{equation}
It follows from Theorem~\ref{thm:HochschildOfSmashProducts} that
\[
HH_\ast(M_q(2)) \Leftarrow E^1_{i,j} := H_j(A_3,{\rm CH}_i(k[d],M_q(2))),
\]
where, setting $U_i:= {\rm CH}_i(k[d],M_q(2))$, the same results yields also
\[
H_\ast(A_3, U_i) \Leftarrow {}'E^1_{m,n} := H_m(k[c],{\rm CH}_n(A_2,U_i))
= \begin{cases}
{\rm CH}_n(A_2,U_i)_{k[c]} & \text{ if } \,\, m = 0, \\
{\rm CH}_n(A_2,U_i)^{k[c]} \otimes c & \text{ if } \,\, m = 1, \\
0 & \text{ if } \,\, m \geq 2.
\end{cases}
\]
On the next step we note that
\[
\Big[k[c],M_q(2)\Big] = \Big\langle a^rb^sc^td^\ell \mid r\neq \ell,\ t>0 \Big\rangle,
\]
and hence,
\[
\frac{M_q(2)}{\Big[k[c],M_q(2)\Big]} := \Big\langle a^rb^sd^\ell, a^\ell b^s c^td^\ell \mid r,s,\ell\geq 0, \quad t>0 \Big\rangle.
\]
Following the ideas of Proposition \ref{prop:diagonal} we may identify the ($A_2$, or vertical) homology of the complex
\begin{align*}
{\rm CH}_n & (A_2,U_i)_{k[c]}\\
\cong & \Big\langle a^{r_1}b^{s_1}\ot\cdots\ot a^{r_n}b^{s_n} \otimes a^{r_{n+1}}b^{s_{n+1}} d^{\ell_1} \otimes d^{\ell_2} \ot\cdots\ot d^{\ell_{i+1}} \mid r_u,s_v,\ell_w\geq 0 \Big\rangle \\
& \oplus \Big\langle a^{r_1}b^{s_1}\ot\cdots\ot a^{r_n}b^{s_n} \otimes a^{r_{n+1}}b^{s_{n+1}}c^td^{\ell_1} \otimes d^{\ell_2} \ot\cdots\ot d^{\ell_{i+1}} \mid \\
& \qquad r_u,s_v,\ell_w\geq 0, t > 0, \sum_u r_u = \sum_w \ell_w \Big\rangle
\end{align*}
with the total homology of the bicomplex
\begin{align*}
{\rm CH}_{\alpha,\beta} & (A_2,U_i)_{k[c]} \\
\cong & \Big\langle a^{r_1}\ot\cdots\ot a^{r_\beta}\otimes a^{r_{\beta+1}} b^s d^{\ell_1} \otimes d^{\ell_2} \ot\cdots\ot d^{\ell_{i+1}} \otimes b^{s_1}\ot\cdots\ot b^{s_\alpha} \mid r_u,s,\ell_w\geq 0 \Big\rangle \\
& \oplus \Big\langle a^{r_1}\ot\cdots\ot a^{r_\beta}\otimes a^{r_{\beta+1}} b^sc^td^{\ell_1} \otimes d^{\ell_2} \ot\cdots\ot d^{\ell_{i+1}} \otimes b^{s_1}\ot\cdots\ot b^{s_\alpha} \mid \\
& \qquad r_u,s,t,\ell_w\geq 0, t>0, \sum_u r_u = \sum_w \ell_w\Big\rangle.
\end{align*}
The homology of this bicomplex, in turn, is approximated by the spectral sequence
\begin{align}\label{eqn-homology-E1-coinv}
H_\ast & \left({\rm CH}_\ast(A_2,U_i)_{k[c]} \right) \nonumber
= H_\ast\left({\rm CH}_{\ast,\ast}(A_2,U_i)_{k[c]} \right)\\
\Leftarrow {}'E^1_{\alpha,\beta} &
\cong \begin{cases}
\Big\langle a^{r_1}\ot\cdots\ot a^{r_\beta}\otimes a^rb^sc^td^\ell \otimes d^{\ell_1} \ot\cdots\ot d^{\ell_i} \mid \\
\qquad r,r_1,\ldots r_\beta,s,t,\ell,\ell_1,\ldots,\ell_i\geq 0,\quad r_1 + \ldots + r_\beta + r = \ell_1 + \ldots + \ell_i +\ell\Big\rangle\\
\Big\langle a^{r_1}\ot\cdots\ot a^{r_\beta}\otimes a^rd^\ell \otimes d^{\ell_1} \ot\cdots\ot d^{\ell_i} \mid r,r_1,\ldots r_\beta,s_1,\ldots,s_\alpha,\ell,\ell_1,\ldots,\ell_i\geq 0\big\}\Big\rangle & \text{ if } \alpha= 0, \\
\Big\langle a^{r_1}\ot\cdots\ot a^{r_\beta}\otimes a^rb^sc^td^\ell \otimes d^{\ell_1} \ot\cdots\ot d^{\ell_i} \mid \\
\qquad r,r_1,\ldots r_\beta,s,t,\ell,\ell_1,\ldots,\ell_i\geq 0,\quad r_1 + \ldots + r_\beta + r = \ell_1 + \ldots + \ell_i +\ell \Big\rangle \otimes b
& \text{ if } \alpha= 1, \\
0 & \text{ if } \alpha\geq 2.
\end{cases}
\end{align}
Noticing, in view of \ref{eqn-ad-da-commutation}, that
\[
\Big[a, a^{\ell-1}b^sc^td^\ell\Big] = (1-q^{s+t})a^\ell b^sc^td^\ell - q(1-q^{-2\ell}) a^{\ell-1}b^{s+1}c^{t+1}d^{\ell-1},
\]
in the vertical homology we arrive at
\begin{align}\label{eqn-homology-E2-coinv}
H_\ast & \left({\rm CH}_\ast(A_2,U_i)_{k[c]} \right) \nonumber\\
\Leftarrow & {}'E^2_{\alpha,\beta}
\cong \begin{cases}
{\rm CH}_i(k,k[b,c]) \oplus {\rm CH}_i(k[d],k[a,d]) & \text{ if } \alpha= 0, \beta=0, \\
a \otimes {\rm CH}_i(k,k[a]) & \text{ if } \alpha= 0,\ \beta=1\\
{\rm CH}_i(k,k[b,c])\otimes b & \text{ if } \alpha= 1,\ \beta=0, \\
0 & \text{ otherwise.
\end{cases}
\end{align}
Now, on the other hand,
\[
M_q(2)^{k[c]} \cong \Big\langle \big\{ a^rb^s c^td^r \mid r,s,t\geq 0\big\}\Big\rangle,
\]
as such,
\begin{align*}
{\rm CH}_n(A_2,U_i)^{k[c]}
\cong & \Big\langle a^{r_1}b^{s_1}\ot\cdots\ot a^{r_n}b^{s_n} \otimes a^rb^sc^td^\ell \otimes d^{\ell_1} \ot\cdots\ot d^{\ell_i} \mid \\
& \qquad r,r_1,\ldots r_n,s,s_1,\ldots,s_n,t,\ell,\ell_1,\ldots,\ell_i\geq 0, \quad r_1 + \ldots + r_n + r = \ell_1 + \ldots + \ell_i +\ell\Big\rangle.
\end{align*}
Similarly above, we may identify the ($A_2$, or vertical) homology of this complex with the total homology of the bicomplex
\begin{align*}
{\rm CH}_{\mu,\nu}(A_2,U_i)^{k[c]}
\cong & \Big\langle a^{r_1}\ot\cdots\ot a^{r_\nu}\otimes a^rb^sc^td^\ell \otimes d^{\ell_1} \ot\cdots\ot d^{\ell_i} \otimes b^{s_1}\ot\cdots\ot b^{s_\mu} \mid \\
& \qquad r,r_1,\ldots r_\nu,s,s_1,\ldots,s_\mu,t,\ell,\ell_1,\ldots,\ell_i\geq 0, \quad r_1 + \ldots + r_\nu + r = \ell_1 + \ldots + \ell_i +\ell\Big\rangle.
\end{align*}
Now, following the same line of thought in \eqref{eqn-homology-E1-coinv} and \eqref{eqn-homology-E2-coinv}, we obtain
\begin{align*}
H_\ast \left({\rm CH}_\ast(A_2,U_i)^{k[c]} \otimes c\right) \Leftarrow & {}'E^2_{\alpha,\beta} \\
\cong &
\begin{cases}
{\rm CH}_i(k,k[b,c])\otimes c & \text{ if } \alpha= 0, \ \beta=0, \\
{\rm CH}_i(k,k[b,c]) \otimes b\otimes c & \text{ if } \alpha= 1,\ \beta=0, \\
0 & \text{ otherwise. }
\end{cases}
\end{align*}
We thus conclude,
\begin{align*}
H_\ast(A_3, U_i)
\Leftarrow & {}'E^2_{m,n}
\cong \begin{cases}
{\rm CH}_i(k,k[b,c]) \oplus {\rm CH}_i(k[d],k[a,d])
& \text{ if } m=0\text{ and } n= 0, \\
a \otimes {\rm CH}_i(k,k[a]) \oplus {\rm CH}_i(k,k[b,c]) \otimes b
& \text{ if } m=0\text{ and } n= 1, \\
{\rm CH}_i(k,k[b,c]) \otimes c
& \text{ if } m=1\text{ and } n= 0, \\
{\rm CH}_i(k,k[b,c]) \otimes b \otimes c
& \text{ if } m=1\text{ and } n= 1, \\
0 & \text{ otherwise.}
\end{cases}
\end{align*}
As such,
\begin{align*}
HH_\ast(M_q(2))
\Leftarrow & E^1_{i,j} \\
:= & H_j(A_3,{\rm CH}_i(k[d],M_q(2))) \\
\cong & \begin{cases}
{\rm CH}_i(k,k[b,c]) \oplus {\rm CH}_i(k[d],k[a,d])
& \text{ if } j = 0, \\
a \otimes {\rm CH}_i(k,k[a]) \oplus {\rm CH}_i(k,k[b,c]) \otimes b \oplus {\rm CH}_i(k,k[b,c]) \otimes c
& \text{ if } j = 1, \\
{\rm CH}_i(k,k[b,c]) \otimes b \otimes c
& \text{ if } j = 2, \\
0 & \text{ if } j \geq 3,
\end{cases}
\end{align*}
and
\begin{align*}
HH_\ast(M_q(2)) \Leftarrow & E^2_{i,j} \\
\cong & \begin{cases}
k[b,c] \oplus k[a,d]
& \text{ if } j = 0 \text{ and } i =0, \\
k[d] \otimes d
& \text{ if } j = 0 \text{ and } i =1, \\
a \otimes k[a] \oplus k[b,c] \otimes b \oplus k[b,c] \otimes c
& \text{ if } j = 1 \text{ and } i =0, \\
k[b,c]\otimes b \otimes c
& \text{ if } j = 2 \text{ and } i =0, \\
0 & \text{ if } j \geq 3 \text{ or } i\geq 1.
\end{cases}
\end{align*}
In other words, we obtain
\[
HH_n(M_q(2)) \cong
\begin{cases}
k[b,c] \oplus k[a,d]
& \text{ if } n =0, \\
k[a] \oplus k[b,c] \oplus k[b,c] \oplus k[d]
& \text{ if } n =1, \\
k[b,c]
& \text{ if } n =2, \\
0 & \text{ if } n \geq 3.
\end{cases}
\]
\begin{remark}
We note that the result above does not follow \cite[Coroll. 2.5]{GuccGucc97} since the
extension $M_q(2)=A_3[d,\alpha_3,\delta]$ does not satisfy the hypothesis therein.
\end{remark}
\bibliographystyle{plain}
|
\section{Introduction}
Ultra-cold molecules can be used to test fundamental symmetries, investigate the behavior of strongly-interacting quantum systems, process quantum information, and study collisions and chemistry at low temperature. Direct laser cooling is one way to produce molecules at microkelvin temperatures. Laser cooling has been used to cool several diatomic molecules~\cite{Shuman2010,Hummon2013,Zhelyazkova2014,Lim2018}, and even a polyatomic molecule~\cite{Kozyryev2017}, and magneto-optical traps (MOTs) of SrF~\cite{Barry2014, McCarron2015, Norrgard2016, Steinecker2016}, CaF~\cite{Truppe2017, Williams2017, Anderegg2017} and YO~\cite{Collopy2018} have been demonstrated. Some of the properties of these laser cooled and magneto-optically trapped molecules are similar to their atomic counterparts, while other properties are strikingly different and not fully understood. For example, the spring constants and damping coefficients are both roughly 100 times smaller in molecular MOTs compared to typical alkali atomic MOTs, and the temperature of the molecular MOTs are up to 100 times higher than the Doppler limit, whereas atomic MOTs normally have temperatures close to the Doppler limit.
The main differences between the properties of laser cooled molecules and most laser cooled atoms stems from the different way in which a closed optical cycling transition is achieved. For atoms, it is usual to drive a transition from a ground state of angular momentum $F$ to an excited state of angular momentum $F'=F+1$, referred to as a type-I transition. In this case, there are no dark states, so the atom can scatter the laser light indefinitely. The transition used for laser cooling of molecules is between the lowest rotational level of an electronically excited state, and the first rotationally-excited level of the ground electronic state. This choice avoids decays to multiple rotational states~\cite{Stuhl2008}, but introduces type-II transitions that have $F'\leq F$. In this case, there are dark states amongst the ground state sub-levels. These dark states have a great effect on the scattering rate and associated position and velocity-dependent forces, which in turn influence the properties of a molasses or MOT.
In previous work \cite{Devlin2016}, we studied these position and velocity-dependent forces in 3D molasses and MOTs operating on type-I and type-II transitions, for idealized systems with a single hyperfine ground state and a single hyperfine excited state. In this paper, we present a general method for modeling laser cooling and magneto-optical trapping of molecules, calculate the forces on a real molecule, work out the effect of these forces on the properties of a molasses and a MOT, and compare our findings to experimental results. Some findings are also compared to previous work that used a rate-model approach~\cite{Tarbutt2015,Tarbutt2015b}. We begin by presenting a general model which can be applied to any laser-cooled atom or molecule. Then, we focus our attention on CaF molecules cooled using the {A~$^2\Pi_{1/2}$~--~X~$^2\Sigma^+$~P(1)} transition, since MOTs and molasses of these molecules have recently been studied in depth~~\cite{Truppe2017b, Williams2017, Anderegg2017}. We examine the predictions of the model and compare them to the results of experiments, first for a molasses, and then for the computationally more complex case of the MOT.
\section{Method}
\label{OBEs}
\subsection{Generalised optical Bloch equations}
We start with a set of optical Bloch equations (OBEs) which describe the time evolution of the internal state of the molecule in the presence of a set of near-resonant laser fields with angular frequencies $\omega_{k}$, and a static magnetic field. To derive the OBEs, we use an identical method to that described in our earlier work \cite{Devlin2016}, generalized to the case of multiple lower and upper hyperfine components. We do not re-derive these equations here, but we do present them in a general form useful to any laser cooling experiment where multiple transitions participate.
The basis states of the system are labeled $\ket{e/g,F_a,M_a}_{\rm s}$ where $e$ denotes an electronically excited state and $g$ a ground electronic state, $F_a$ is the angular momentum, $M_{a}$ is the $z$ projection of angular momentum and $a$ is an index labeling the state. Where there is no ambiguity, we use the shorthand $\ket{e/g,F_a,M_a}_{\rm s}\equiv \ket{e/g,a}_{\rm s}$. The state $\ket{e/g,a}_{\rm s}$ has energy $\hbar\omega_{e/g,a}$, and it is convenient to define the time-dependent Heisenberg picture operators $\ket{e/g,a}\bra{e/g,b}=e^{-i\omega_{e/g,a}t}\ket{e/g,a}\mathbin{\subrangle{{\rm s}}}e^{i\omega_{e/g,b}t}\mathbin{\sublangle{{\rm s}}\!\!}\bra{e/g,b}$. We find that the expectation values $\langle |e/g,a\rangle \langle e/g,b|\rangle$
evolve according to the following generalized optical Bloch equations:
\begin{align}
\frac{\text{d} \big\langle |g,a\rangle \langle e,b|\big\rangle }{\text{d} \tau }=& \sum_{\substack{k,q,c}}\frac{i G_k f_{k,q}}{2 \sqrt{2}}e^{i(\overline{\omega}_{e,b}-\overline{\omega}_{g,c}-\overline{\omega}_k)\tau}\mathbin{\sublangle{{\rm s}}\!\!\left\langle e,b\left|\bar{d}_q\right|g,c\right\rangle\subrangle{{\rm s}}}\langle |g,a\rangle \langle g,c|\rangle \nonumber\\
&-\sum_{\substack{k,q,c'}}\frac{i G_k f_{k,q}}{2 \sqrt{2}}e^{i (\overline{\omega}_{e,c'}-\overline{\omega}_{g,a}-\overline{\omega}_k)\tau} \mathbin{\sublangle{{\rm s}}\!\!\left\langle e,c'\left|\bar{d}_q\right|g,a\right\rangle\subrangle{{\rm s}}}\langle |e,c'\rangle \langle e,b| \rangle\nonumber\\
&+\sum_{\substack{q,n}}i (-1)^q \beta _q \,\mathbin{\sublangle{{\rm s}}\!\!\left\langle e,F_b,M_b\left|\bar{\mu}_{-q}\right|e,F_b,n\right\rangle\subrangle{{\rm s}}}\langle |g,F_a,M_a\rangle \langle e,F_b,n|\rangle \nonumber\\
&-\sum_{\substack{q,m}}i (-1)^q \beta _q \,\mathbin{\sublangle{{\rm s}}\!\!\left\langle g,F_a,m\left|\bar{\mu }_{-q}\right|g,F_a,M_a\right\rangle\subrangle{{\rm s}}} \langle
|g,F_a,m\rangle \langle e,F_b,M_b|\rangle \nonumber\\
&-\frac{1}{2} \langle |g,a\rangle \langle e,b|\rangle\,
\label{eq:eg}
\end{align}
\begin{align}
\frac{\text{d} \big\langle |e,a\rangle \langle e,b|\big\rangle }{\text{d} \tau }=& \sum_{k,q,c}\frac{i G_k}{2 \sqrt{2}}\bigg(f_{k,q} e^{i (\overline{\omega} _{e,b}-\overline{\omega} _{g,c}-\overline{\omega} _k)\tau} \mathbin{\sublangle{{\rm s}}\!\!\left\langle e,b\left|\bar{d}_q\right|g,c\right\rangle\subrangle{{\rm s}}} \langle |e,a\rangle \langle g,c|\rangle \nonumber\\
&-\left(f_{k,q}\right)^* e^{-i (\overline{\omega}_{e,a}-\overline{\omega} _{g,c}-\overline{\omega} _k)\tau} \mathbin{\sublangle{{\rm s}}\!\! \left\langle g,c\left|\bar{d}_q\right|e,a\right\rangle\subrangle{{\rm s}}} \langle |g,c\rangle \langle e,b| \rangle\bigg)\nonumber\\
&+\sum_{q,n} i (-1)^q \beta _q \, \mathbin{\sublangle{{\rm s}}\!\! \left\langle e,F_b,M_{b}\left|\bar{\mu }_{-q}\right|e,F_{b},n\right\rangle\subrangle{{\rm s}}} \langle |e,F_a,M_a\rangle \langle e,F_b,n|\rangle \nonumber\\
&-\sum_{q,m} i (-1)^q \beta _q \,\mathbin{\sublangle{{\rm s}}\!\! \left\langle e,F_{a},m\left|\bar{\mu }_{-q}\right|e,F_a,M_{a}\right\rangle\subrangle{{\rm s}}} \langle |e,F_a,m\rangle \langle e,F_b,M_b|\rangle \nonumber\\
&-\langle |e,a\rangle \langle e,b|\rangle
\,,
\label{eq:ee}
\end{align}
\begin{align}
\frac{\text{d} \big\langle |g,a\rangle \langle g,b|\big\rangle }{\text{d} \tau }=&\sum_{k,q,c'}\frac{-i G_k}{2 \sqrt{2}}\bigg(f_{k,q} e^{i (\overline{\omega} _{e,c'}-\overline{\omega}_{g,a}-\overline{\omega} _k)\tau} \mathbin{\subrangle{{\rm s}}\!\! \left\langle e,c'\left|\bar{d}_q\right|g,a\right\rangle\subrangle{{\rm s}}} \langle |e,c'\rangle\langle g,b| \rangle\nonumber\\
&-\left(f_{k,q}\right)^* e^{-i (\overline{\omega} _{e,c'}-\overline{\omega} _{g,b}-\overline{\omega} _k)\tau} \mathbin{\sublangle{{\rm s}}\!\!\left\langle g,b\left|\bar{d}_q\right|e,c'\right\rangle\subrangle{{\rm s}}} \langle |g,a\rangle \langle e,c'|\rangle \bigg)\nonumber\\
&+ \sum_{q,n}i (-1)^q \beta _q \,\mathbin{\sublangle{{\rm s}}\!\!\left\langle g,F_b,M_{b}\left|\overline{\mu} _{-q}\right|g,F_b,n\right\rangle\subrangle{{\rm s}}} \langle |g,F_a,M_a\rangle \langle g,F_b,n|\rangle\nonumber\\
&-\sum_{q,m} i (-1)^q \beta _q \,\mathbin{\sublangle{{\rm s}}\!\!\left\langle g,F_a,m\left|\overline{\mu} _{-q}\right|g,F_a,M_{a}\right\rangle\subrangle{{\rm s}}} \langle |g,F_a,m\rangle \langle g,F_b,M_b|\rangle \nonumber\\
&+\sum_{q,c',c''} \mathbin{\sublangle{{\rm s}}\!\!\left\langle e,c'\left|\bar{d}_q\right|g,a\right\rangle\subrangle{{\rm s}}} \,\,\,\mathbin{\sublangle{{\rm s}}\!\!\left\langle g,b\left|\bar{d}_q\right|e,c''\right\rangle\subrangle{{\rm s}}} \nonumber\\
&\times e^{i(\overline{\omega} _{e,c'}-\overline{\omega} _{e,c''}+\overline{\omega} _{g,b}-\overline{\omega} _{g,a})\tau}\langle |e,c'\rangle\langle e,c''| \rangle \, .
\label{eq:gg}
\end{align}
In these equations, the first two terms represent interactions with the lasers, the next two terms capture the effect of an applied magnetic field, and the remaining terms are caused by spontaneous emission. The summations are over the laser frequencies $k$, the polarizations $q=-1,0,1$, all ground states $|g,c\rangle$, all excited state $|e,c'\rangle$ or $|e,c''\rangle$ and all sublevels $-F_a\leq m\leq F_a$ and $-F_b\leq n\leq F_b$. Spontaneous emission has been introduced via the radiation reaction approximation \cite{Ungar1989}, in which the total electric field interacting with the molecular dipole $\boldsymbol{\hat{d}}$ is written as the sum of the applied electric fields from the lasers and a reaction field $\boldsymbol{\hat{E}}_\textrm{RR}=\tfrac{1}{6\pi\varepsilon_0c^3}\tfrac{\textrm{d}^3}{\textrm{d}t^3}\boldsymbol{\hat{d}}\approx\tfrac{i k^3}{6\pi\varepsilon_0}\boldsymbol{\hat{d}}$, where in the last step we have assumed that the laser frequencies are so close that $\omega_k$ can be replaced by a single $\omega$. We have made the rotating wave approximation and have assumed that the magnetic field is small enough that all Zeeman shifts are linear. For the specific case of CaF which we consider later, this is a good approximation for fields below 10~G, which is the relevant range for both the molasses and the MOT. To put the equations into natural units, we have used a dimensionless time $\tau = \Gamma t$, and dimensionless angular frequencies $\overline{\omega}_{i} = \omega_{i}/\Gamma$, where $\Gamma$ is the natural decay rate of the excited state. The classical electric field with frequency component $\omega_{k}$ is written as $\boldsymbol{E}_k(\boldsymbol{x},t) = {\cal E}_k \sum_q f_{k,q}(\boldsymbol{x})\epsilon_q^* \cos(\omega_k t)$ and has intensity $I_{k}=\frac{1}{2}c\epsilon_0{\cal E}_k^2$. Here, $\boldsymbol{x}(t)$ is the position of the molecule at time $t$ and the $\epsilon_q$ are the usual spherical basis vectors. The dimensionless parameter $G_{k}$ is defined by $G_{k}=\sqrt{I_k/I_\textrm{sat}}$, where $I_\textrm{sat}=\pi h c \Gamma/(3\lambda^3)$ is the usual expression for the saturation intensity. The applied magnetic field is written as $\boldsymbol{B} = \frac{\hbar\Gamma}{\mu_\textrm{B}} \sum_q \beta_q \epsilon_q^*$ which defines the dimensionless quantity $\beta$. The dimensionless matrix elements of the electric dipole moment operator are
\begin{align*}
\mathbin{\sublangle{{\rm s}}\!\! \left\langle g,a\left|\bar{d}_q\right|e,b\right\rangle\subrangle{{\rm s}}}=&\frac{(-1)^{F_a-M_{a}}\sqrt{2F_b+1}\,\,\mathbin{\sublangle{{\rm s}}\langle g,F_a\|d\|e,F_b\rangle\subrangle{{\rm s}}}}{\sqrt{\sum_c\left|\,\,\mathbin{\sublangle{{\rm s}}\langle e,F_b\|d\|g,F_c\rangle\subrangle{{\rm s}}}\right|^2}}\begin{pmatrix}
F_a & 1 & F_b \\
-M_{a} & q & M_{b}
\end{pmatrix}\,,
\end{align*}
where the sum over $c$ includes all ground states.
The dimensionless matrix elements of the magnetic moment operator are
\begin{align*}
\mathbin{\sublangle{{\rm s}}\!\! \left\langle e/g,F_a,m\left|\bar{\mu}_q\right|e/g,F_a,n\right\rangle\subrangle{{\rm s}}}=&-g_{F_a}(-1)^{F_a-m}\sqrt{F_a(F_a+1)(2F_a+1)}
\begin{pmatrix}
F_a & 1 & F_a\\
-m & q & n
\end{pmatrix}\, ,
\end{align*}
where $g_{F_a}$ is the magnetic $g$-factor of the state $|e/g,a\rangle$.
We use these equations to calculate how the internal state of a moving molecule evolves over time. The equations may contain far-off-resonant couplings, particularly if there are large energy splittings amongst the ground or excited states. These terms have almost no effect on the results, but can greatly slow down the simulations, so we remove terms oscillating at frequencies greater than a certain threshold, typically $10\Gamma$. In general, for a real multi-level molecular system in which each transition can be excited by several laser frequencies, the molecular operators do not come to a steady state, but continue to vary as a function of time even if the molecule is stationary. If the Hamiltonian is periodic in time, then once any transients relating to the initial conditions die away, the expectation values of the molecular operators also become periodic in time, with the same periodicity as the Hamiltonian \cite{Yudin2016}. In this work, we use rounded values for all frequencies and speeds which sets the periodicity of the Hamiltonian, and hence ensures the molecular operators will eventually reach a periodic quasi-steady state.
Starting with the population evenly distributed over all ground states, we solve the initial value problem for the ordinary differential equations (\ref{eq:eg}-\ref{eq:gg}) using an explicit Runge-Kutta numerical method, implemented using the \emph{Mathematica} software package, propagating the molecular operators for a large number of time steps ($\tau \approx 1000$). We check how closely the system has reached the quasi-steady state after this initial propagation period by comparing the values of the internal state variables and the velocity-dependent force (defined below) averaged over the next two time periods of the Hamiltonian. A large difference in the calculated values indicates that a longer time period should be used.
An alternative method of finding the periodic steady state is to calculate the eigenvectors of the propagation matrix \cite{Yudin2016}. This may be faster, though so far in our investigations both solution methods take a similar amount of time.
The quasi-steady state expectation values of the molecular operators are now used to calculate several relevant properties. One is the expectation value of the force operator, $\boldsymbol{\hat{f}}=\textrm{d}\boldsymbol{\hat{P}}/\textrm{d}t=-\nabla\hat{H}$, which may be written as
\begin{equation}
\boldsymbol{F}\! =\!\langle \boldsymbol{\hat{f}}\rangle\! =\!\!\sum_{k,q,c,c'}\!\!\frac{-h\Gamma G_k}{2\sqrt{2} \lambda}e^{-i (\overline{\omega} _{k}+\overline{\omega} _{g,c}-\overline{\omega}_{e,c'})\tau} \mathbin{\sublangle{{\rm s}}\!\!\left\langle e,c'\left|\bar{d}_q\right|g,c\right\rangle\subrangle{{\rm s}}}\big\langle |e,c'\rangle \langle g,c|\big\rangle\frac{\partial f_{k,q}(\boldsymbol{x})}{\partial\boldsymbol{x}}+\textrm{c.c.}
\end{equation}
We calculate the force, averaged over one period of the Hamiltonian's oscillation, as the molecule is dragged at constant velocity through the light field. Repeating this for various velocities gives the velocity-dependent force curve, and the gradient of this curve around the equilibrium velocity gives the damping coefficient. We note that this method of determining the velocity-dependent force has been used in several studies on laser cooling of atoms, e.g. \cite{Ungar1989,Molmer1991} and is known to give good agreement with the force obtained from Monte Carlo simulations when the speed is high enough~\cite{Ungar1989}. That comparison suggests that, for the molecular system considered in this paper, the method will be accurate for all speeds above 0.1~m/s.
Another important quantity is the momentum diffusion tensor, whose components are
\begin{align}
D_{ij}=\frac{1}{2}&\frac{\textrm{d}}{\textrm{d}t}\left(\langle \hat{P}_i\hat{P}_j\rangle-\langle \hat{P}_i\rangle\langle \hat{P}_j\rangle\right)=\textrm{Re}\int_{-\infty}^t\left[\langle \hat{f}_i(\tau)\hat{f}_j(t)\rangle-\langle \hat{f}_i(\tau)\rangle \langle \hat{f}_j(t)\rangle\right]\textrm{d}\tau.\label{eq:fullDiff}
\end{align}
Here $\hat{P}_i$ and $\hat{f}_i$ are the Cartesian components of the momentum and force operator respectively. The momentum diffusion tensor is diagonal if we choose the quantization axis as one of the coordinate axes, and if the laser beams also propagate along a coordinate axis \cite{Ungar1989}. If, furthermore, the laser configuration is identical along all three coordinate axes, then the three diagonal coefficients are all equal, $D_{xx} = D_{yy} = D_{zz} = D$. The diffusion constant includes the effects of (i) the random momentum kicks due to spontaneous emission, (ii) the random momentum kicks due to fluctuations in the difference between the number of photons absorbed from each of the laser beams, and (iii) the fluctuations in the dipole force that arise as the molecule hops between its quantum states in the presence of the intensity and polarization gradients of the light~\cite{Gordon1980,Dalibard1989}. Evaluating Eq.~(\ref{eq:fullDiff}) for a molecule in motion through a complicated light field is a substantial challenge, since the term $\langle \hat{f}_i(\tau)\hat{f}_j(t)\rangle$ involves the expectation value of products of molecular operators at different times of the form $\langle \left(|g\rangle \langle e|(\tau)\right)\left(|e\rangle \langle g|(t)\right)\rangle$, which are distinct from the expectation values of the operators yielded by solving the OBEs. M\o lmer \cite{Mo/lmer1991} provides a method for calculating $D$ for a stationary atom with a single lower and upper hyperfine state interacting with a single laser beam. However, this method is not straightforward to apply to our case since, in the presence of several laser beams of different frequencies, there is no longer a stationary solution to the OBEs. A later paper \cite{Agarwal1993} provides a method for calculating the first order velocity dependence of the diffusion tensor for a simpler system, which could potentially be adapted to our case. Instead, we choose to approximate the diffusion tensor by a simple expression that ignores stimulated emission and only considers the random nature of spontaneous emission and the corresponding absorption events. In this approximation, the diffusion constant $D_{\rm s}$ is related to the total excited state population $N_e=\sum_{c'}\big\langle |e,c'\rangle \langle e,c'|\big\rangle$ according to
\begin{align}
D_{\rm{s}}=\frac{1}{3}\hbar^2k^2\Gamma N_e.
\label{eq:diff}
\end{align}
This approximation neglects the extra diffusion caused by the light standing wave pattern, so can only provide a lower limit to the diffusion constant. As with the force, we average the diffusion constant over a period of the Hamiltonian in subsequent calculations of the temperatures and damping constants.
From the force and diffusion constant, we can predict how the fraction of molecules in the phase space element $\textrm{d}\boldsymbol{x}\, \textrm{d}\boldsymbol{v}\,\textrm{d}t$, denoted by $W(\boldsymbol{x},\boldsymbol{v},t)\textrm{d}\boldsymbol{x}\, \textrm{d}\boldsymbol{v}\, \textrm{d}t$, evolves over time. This evolution is governed by a Fokker-Planck-Kramers equation \cite{Marksteiner1996,Molmer1994}, which for a molecule of mass $m$, moving in 3D is
\begin{equation}
\frac{\partial W(\boldsymbol{x},\boldsymbol{v},t)}{\partial t}+\sum_{i}v_i\frac{\partial W(\boldsymbol{x},\boldsymbol{v},t)}{\partial x_i}=\sum_{i}\frac{\partial }{\partial v_i}\left(\frac{-F_i(\boldsymbol{x},\boldsymbol{v})}{m}W(\boldsymbol{x},\boldsymbol{v},t)+\frac{D_{ii}(\boldsymbol{x},\boldsymbol{v})}{m^2}\frac{\partial W(\boldsymbol{x},\boldsymbol{v},t)}{\partial v_i}\right).
\label{eq:fpe}
\end{equation}
Using the methods discussed above, we find $F(\boldsymbol{x},\boldsymbol{v})$ and $D_{s}(\boldsymbol{x},\boldsymbol{v})$.
To calculate their effective values for a millimeter-sized cloud of moving molecules, we average them over a cube of size $\lambda$, and also average over all directions of motion. These average values are labeled $\tilde{F}_i(\boldsymbol{x},v)$ and $\tilde{D}_{s}(\boldsymbol{x},v)$, and they are the ones we use in Eq.~(\ref{eq:fpe}). Due to the symmetric arrangement of the laser beams, the force transverse to the direction of motion averages to zero, leaving only a force in the direction of motion. Thus, the average force can be written as $\tilde{F}(\boldsymbol{x},v)\hat{\boldsymbol{v}}$, where $v$ is the speed and $\hat{\boldsymbol{v}}$ is a unit vector in the direction of $\boldsymbol{v}$. In the special case where the intensity and magnetic field are uniform over the size scale of the molecular distribution, the position dependence vanishes from Eq.~(\ref{eq:fpe}), which we find reduces to the following equation for the probability density, $W(v,t)$:
\begin{equation}
\frac{\partial}{\partial t}v^2W(v,t)=\frac{\partial }{\partial v}\left(\frac{-\tilde{F}(v)}{m}v^2W(v,t)+\frac{v^2\tilde{D}_{s}(v)}{m^2}\frac{\partial W(v,t)}{\partial v}\right).
\label{eq:fpe3}
\end{equation}
We are often interested in the steady-state solution of this equation, which is
\begin{align}
W(v) & = W_0 \exp\left[m \int_0^{v}\frac{\tilde{F}(u)}{\tilde{D}_s(u)}\textrm{d}u\right].
\label{eq:temp}
\end{align}
The fraction of molecules with speeds between $v$ and $v+dv$ is $W(v)\,4\pi v^2 dv$, and $W_0$ is chosen so that ${\int W(v)\, 4\pi v^2 dv=1}$. If the force is a linear drag force $\tilde{F}(v)=-\alpha v$, and the diffusion constant does not depend on speed, then this integral leads to a Maxwell-Boltzmann speed distribution with temperature $T=\tilde{D}_s/k_B\alpha$. More generally, for any other $W(v)$, we calculate $\overline{v^2}$, the variance in the speed, and so get the equivalent temperature of a Maxwell-Boltzmann speed distribution with that variance,
\begin{align}
T=\frac{m}{3k_B}\int_0^{u}v^2W(v)\,4\pi v^{2}\,\textrm{d}v \, ,
\label{eq:temperatureDefinition}
\end{align}
where $u$ is the upper speed to which the functions $\tilde{F}$ and $\tilde{D}_s$ have been found, chosen to be sufficiently high so as not to affect the distribution $W$.
We can now apply these general equations to the laser cooling of calcium monofluoride. In order to do this, it is helpful to summarize some salient experimental details so we can assess what features of the problem need to be considered.
\section{Application to laser cooling and magneto-optical trapping of CaF}
\begin{figure}[tb]
\centering
\includegraphics[scale=1]{Fig1.pdf}
\caption{(a) Main levels used in the laser cooling of CaF. (b) The hyperfine levels in the $A^2\Pi_{1/2}(v=0)$ and $X^2\Sigma^+(v=0)$ states, and the frequencies of the main cooling laser used to address them. Also indicated is the polarization of each frequency component, as used in the experiments described in \cite{Truppe2017,Williams2017} and in our simulations. Here, $\sigma^{\pm}$ means that the restoring beams of the MOT drive $\Delta M=\pm 1$ transitions in a coordinate system whose $z$-axis is along the magnetic field. The numbers to the left of the lines are the hyperfine intervals in MHz. (c) The $\eta$ parameter as a function of the intensity and detuning of the cooling laser. $\eta$ is the the total population in $X^2\Sigma^+(v=0)$ and $A^2\Pi_{1/2}(v=0)$ in the case where decays to $X^2\Sigma^+(v=1)$ are forbidden, divided by the total population when decays to this level are allowed.}
\label{fig:CaFenergylevel}
\end{figure}
Our aim is to develop a comprehensive understanding of laser cooling and magneto-optical trapping of molecules. We focus here on the experimental results obtained for dc MOTs and optical molasses of CaF~\cite{Truppe2017, Williams2017, Anderegg2017}, though we expect our methods, and many of our conclusions, to be equally applicable to other molecules and to radio-frequency MOTs~\cite{Norrgard2016}. Figure \ref{fig:CaFenergylevel}(a) shows the relevant energy levels and optical transitions used in the experiments. We use the notation $I_{ij}$ to refer to the total intensity of the lasers addressing the $v=i\rightarrow v'=j$ transition. The main cooling laser drives the transition $\textrm{A} ^2\Pi_{1/2}(v=0, J=1/2)\leftarrow\textrm{X} ^2\Sigma^{+}(v=0,N=1)$ which has a natural linewidth of $\Gamma=2\pi\times 8.3$ MHz, and a saturation intensity of $I_\textrm{sat}=4.9$~mW/cm$^2$. Population that leaks into $\textrm{X} ^2\Sigma^{+}(v=1,N=1)$ is returned to the cooling cycle using a second laser which we refer to as the repump laser. Additional lasers (not shown in the figure) are used to recover population that leaks to higher-lying vibrational states, but they play such a minor role that we can safely neglect them in the simulations. The cooling laser has four main frequency components to address the four hyperfine ground states shown in Fig.~\ref{fig:CaFenergylevel}(b). The frequency components are derived from a single laser beam; one portion is passed through an electro-optic modulator (EOM) at 74.5 MHz, making three sidebands which address the $F=1^-$, $F=0$ and $F=2$ states, while another portion is passed through an acousto-optic modulator at 48 MHz to address the $F=1^+$ state \cite{Williams2017}. These components are overlapped and directed onto the molecules in three orthogonal pairs of counter-propagating circularly polarized beams in the standard MOT configuration. In the simulations, their intensities are all equal, and their frequencies are $\omega_0+\Delta_{00}+\omega_k$, where $\Delta_{00}$ is a common detuning of the $v=0 \rightarrow v'=0$ laser, $\omega_0$ is the frequency of the $\textrm{X}^2\Sigma^+(v=0, N=1, F=2) \rightarrow \textrm{A}^2\Pi_{1/2}(v=0,J=1/2, F=1)$ transition, and {$\omega_k=2\pi\times\{-2.90,24.15,72.29,146.00\}$ MHz}. With this definition of $\Delta_{00}$, the simulations predict that the maximum scattering rate at high intensity occurs when $\Delta_{00}=0$, and that there is Doppler cooling when $\Delta_{00}<0$. In the experiment, the EOM produces additional, unwanted, sidebands, but these are detuned from any transition by more than 70~MHz so we neglect them. The repump laser has the same set of four frequency components, but here the common detuning is fixed at zero.
The parameters used for modeling the MOT and the molasses are summarized in Tab.~\ref{tab:Parameters}. When modeling the MOT, we use laser beams that have Gaussian intensity distributions with $1/e^2$ radii of 8.1~mm, as in the experiment. The magnetic field is $\mathbf{B}=A(\boldsymbol{x}+\boldsymbol{y}-2\boldsymbol{z})$, with $A=15.3$~G/cm. When modeling the molasses, we assume a homogeneous magnetic field randomly oriented with respect to the coordinate axes, and typically use beams with a uniform intensity equal to the true peak intensity. This is valid because the cloud of molecules in the molasses is much smaller than the beam size. When we calculate the capture velocity of the molasses, we consider the full beam profile.
The simulation process is as follows. We first round all angular frequencies and detunings to $\omega_{{\rm min}}$ and round velocities to $\omega_{{\rm min}} \lambda/(2\pi)$, choosing $\omega_{{\rm min}}=10^{-2}\Gamma$ when examining the behavior at high velocity, and $\omega_{{\rm min}}=10^{-3}\Gamma$ if we want more resolution at low velocities. After this procedure, the dimensionless laser frequencies $\omega_k/\Gamma$ are $\bar{\omega}_k=\{-0.36,2.91,8.71,17.59\}$ or $\bar{\omega}_k=\{-0.354,2.909,8.708,17.588\}$. This makes the equations periodic, with period $\mathcal{T}=2\pi/\omega_{{\rm min}}$. Then, for a particular choice of laser intensity, detuning, and applied magnetic field, we solve the OBEs for a molecule moving at constant velocity, $\boldsymbol{v}$, until the quasi-steady-state is reached. From these results we calculate the force and excited-state population averaged over the period $\mathcal{T}$. We do this for random selections of different initial positions, directions of travel, and laser phases, and then average together the results to obtain the mean force and population at this speed $v=|\boldsymbol{v}|$. By repeating this procedure for a range of $v$, we obtain the velocity-dependence of the force and the excited-state population. We use the bootstrap method \cite{BradEandTibshirani1993} to estimate the uncertainty on the mean curve derived this way. Using rounded equations allows the sub-wavelength position-dependent fluctuations in the force and excited state population to be averaged over completely.
\begin{table}[tb]
\caption{Typical experimental parameters used in the simulations. The columns list the intensity of cooling laser, intensity of repump laser, detuning of cooling laser, $1/e^2$ radius of intensity distribution, radial magnetic field gradient. The intensity is the peak intensity due to all four frequency components and all six beams.}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
& $I_{00}$ (mW/cm$^{2}$) & $I_{10}$ (mW/cm$^{2}$) & $\Delta_{00}$ ($\Gamma$) & $w$ (mm) & $A$ (G/cm)\\
\hline
MOT & 2.9--466 & 591 & -0.75 & 8.1 & 15.3 \\
Molasses & 2.9--466 & 591 & 2.50 & 8.1 & 0\\
\hline
\end{tabular}
\label{tab:Parameters}
\end{center}
\end{table}
The system described above consists of 28 molecular states, and thus 405 unique coupled equations. This system is small enough that a quasi-steady state solution to the OBEs, and the associated force and diffusion constant, can be computed in around 140~s on a single processor. However, the process of averaging over different trajectories and laser phases described above typically requires hundreds of individual steady-state solutions of the OBEs, which makes it desirable to speed up the calculation. One way to do that is to neglect decay to $X^2\Sigma^+(v=1)$. This approximation has to be treated with caution, because unlike the other states neglected, there is often significant population in this state. Since the repump light is on resonance, we do not expect the $A^2\Pi_{1/2}(v=0)\leftarrow X^2\Sigma^+(v=1)$ transition to contribute directly to the position-dependent or velocity-dependent forces, but it does contribute indirectly by altering the populations of the various states. In particular, neglecting $X^2\Sigma^+(v=1)$ will lead to an overestimate of the populations in $X^2\Sigma^+(v=0)$ and $A^2\Pi_{1/2}(v=0)$, which in turn leads to an overestimate of the force and the diffusion constant. We have investigated this by solving the rate equations for the system \cite{Tarbutt2015b} with and without the $X^2\Sigma^+(v=1)$ levels. For this investigation, we fix $I_{10}$ and vary $I_{00}$ and $\Delta_{00}$ over wide ranges. In all cases, we find that including the $X^2\Sigma^+(v=1)$ levels reduces the population in $A^2\Pi_{1/2}(v=0)$ and $X^2\Sigma^+(v=0)$ by a common factor $\eta(I_{00},\Delta_{00})$. In light of these observations, we suggest that fairly accurate results can be found by solving the OBEs without including the $X^2\Sigma^+(v=1)$ levels, and then dividing both the force and the excited state population by the correction factor $\eta$. The value of $\eta$ is plotted as a function of $I_{00}$ and $\Delta_{00}$ in Fig.~\ref{fig:CaFenergylevel}(c). To validate this approach, we solve the OBEs for a few specific parameters both with and without the $v=1$ levels. We find that when $\Delta_{00}=2.5\Gamma$ (as used in the molasses), the $\eta$-scaled force and excited state population curves agree very well with the full simulations at both high and low values of $I_{00}$. The agreement is also good when $\Delta_{00}=-0.75\Gamma$ (as used in the MOT) and the intensity is low, but less good at higher intensity. For the highest $I_{00}$ used, the $\eta$-scaled excited-state population is accurate, but the $\eta$-scaled force is 60\% lower than given by the full simulations. This underestimate should be kept in mind when considering simulations of the MOT at the highest intensities.
\section{Forces and excited-state populations in the CaF MOT and molasses}
The essential properties of a CaF MOT and molasses can be understood from the velocity-dependence of the acceleration and the excited-state probability, particularly at low velocity. Figure \ref{fig:basicPhysics}(a) shows the acceleration in the direction of the velocity, $a_\textbf{v}(v)$, as a function of the speed of a CaF molecule, for both red-detuned and blue-detuned light. The magnetic field is set to zero, and the other parameters are given in the caption. For high speeds the acceleration is negative for red-detuned light and positive for blue-detuned light, corresponding to normal Doppler cooling or heating. At lower speeds, polarization-gradient forces dominate over Doppler forces and the acceleration changes sign. We note that, despite the complexity of the CaF system, the force curve is very similar to those found for type-II systems with just one ground state and one excited state~\cite{Devlin2016}, and we conclude that the mechanisms at work are the same as for those simpler systems~\cite{Devlin2016, Weidemuller1994, Shahriar1993}. For a stationary molecule excited on a type-II transition between integer valued\footnote{If $F$ is half-integer, the transition from $F$ to $F'=F$ is only dark in circularly polarized light \cite{Berkeland2002}.} hyperfine levels, there is one dark state when $F=F'$, and two dark states when $F=F'+1$. A moving molecule will tend to be pumped into a dark state near the intensity anti-nodes, where the pumping rate is highest, and will tend to make a non-adiabatic transition back to a bright state close to the nodes, where the splitting between bright and dark states, arising from the ac Stark shift, is smallest. For blue-detuned light, the ac Stark shift is positive, so the bright states have higher energy at the anti-nodes than at the nodes. Thus, the molecule will continually lose energy to the light field, leading to a cooling force. For red-detuned light, the sign of the ac Stark shift is reversed, so the molecule continually receives energy from the light field. In Fig.~\ref{fig:basicPhysics}(a), the speed where the force crosses zero is around 5 m/s. This is the rms speed of a 60 mK Maxwell-Boltzmann distribution, giving an approximate temperature scale where the Doppler and polarization-gradient forces are balanced.
Figure \ref{fig:basicPhysics}(b) shows how the acceleration curve changes as the intensity is reduced. Here, the detuning is negative and close to that typically used for the MOT. Both the range and magnitude of the sub-Doppler force is reduced as the intensity is lowered. This explains the experimental observation that lowering the intensity lowers the temperature.
\begin{figure}[tb]
\begin{center}
{\includegraphics{Fig2.pdf}}
\caption{Acceleration and excited-state population as a function of speed. For each curve, the shaded band indicates the 67\% confidence interval determined from the distribution of multiple simulations. (a) Acceleration curves for $I_{00}=456$ mW/cm$^2$. Blue curve (positive acceleration at higher speeds): $\eta=1.29$, $\Delta_{00}=2.61\Gamma$; red curve (negative acceleration at higher speed): $\eta=1.32$, $\Delta_{00}=-2.39\Gamma$. (b) Acceleration curves for $\Delta_{00}=-0.64\Gamma$ and for a series of intensities, $I_{00}$; dark red: 456~mW/cm$^2$, $\eta=1.65$; red: 229~mW/cm$^2$, $\eta=1.48$; orange: 112~mW/cm$^2$, $\eta=1.32$; yellow: 46~mW/cm$^2$, $\eta=1.16$. (c) Steady-state excited-state fraction when $I_{00}=456$ mW/cm$^2$, $\eta=1.45$, and $\Delta_{00}=2.61\Gamma$. (d) Excited-state fraction when $I_{00}=228$ mW/cm$^2$, $\eta=1.48$, $\Delta_{00}=-0.64\Gamma$. In (c,d), the solid curve is calculated from the optical Bloch equations, while the dashed line is the prediction of a rate equation model~\cite{Tarbutt2015}.}
\label{fig:basicPhysics}
\end{center}
\end{figure}
Figures \ref{fig:basicPhysics}(c,d) show the excited state fraction as a function of speed for a positive detuning close to the one used in the molasses (part c), and for a negative detuning close to the one used for the MOT (part d). Here, we compare the population found by solving the OBEs (solid lines) to the predictions of a rate model (dashed lines), described fully in Ref.~\cite{Tarbutt2015}. At all velocities, the OBEs predict a lower excited state fraction, and hence a lower scattering rate, than predicted by the rate model, indicating that transient dark states play an important role in reducing the scattering rate. Close to zero velocity, the excited state fraction dips even further, as the molecule optically pumps into the dark states. The excited state fraction does not quite drop to zero at zero velocity, as would be expected for an isolated type-II transition driven by a single laser frequency. This behavior can be understood with the help of the level structure shown in Fig.~\ref{fig:CaFenergylevel}(b). Population cannot build up in $F=0$ since regardless of the laser polarization this can always be excited to $F'=1$. The presence of the nearly degenerate pair of excited hyperfine states $F'=0$ and $F'=1$ means that any molecules in $F=1^-$ or $F=1^+$ can always interact with the elliptical light field formed by the superposition of three pairs of $\sigma^+\sigma^-$ beams whose phases are not controlled. Population might be expected to build up in $F=2$, since this can only be excited to $F'=1$ and so it appears to be an isolated type-II transition. However when multiple lasers with different frequencies and polarizations drive the same type-II transition, the states that are dark to one laser beam are, in general, bright to the other. Except in certain special cases, for instance if one of the two fields driving the $F\rightarrow F-1$ transition is circularly polarized and the other linear, there is no guaranteed time-independent orientation of the dipole which is simultaneously orthogonal to all of the applied frequency components of light. The molecule therefore never decouples completely from the light field, even at zero velocity.
It is also worth noting that as well as dark states formed between Zeeman sub-levels of a particular hyperfine state, there can also be Raman dark states which are superpositions of two or more Zeeman sub-levels from different hyperfine states. The laser sidebands are phase-coherent, so we can expect these Raman dark states to be stable. To assess their importance, we artificially destabilize these dark states by adding a term to the right hand side of Eq.~\eqref{eq:gg} of the form $-\gamma(1-\delta(\omega_a-\omega_b))\langle |g,a\rangle \langle g,b|\rangle$, where $\delta(x)$ is the Kronecker delta function and we set $\gamma=10$, so that coherences between hyperfine levels are rapidly damped away. We find that this increases the excited state fraction by about 20\%, reduces the range of the polarization-gradient force by about 50\%, and reduces the magnitude of this force by 33\%, without changing its slope at low velocities. This shows that optical pumping into Raman dark states, and non-adiabatic transitions out of these states, is an important part of the polarization-gradient cooling mechanism in this multi-level system.
\section{Simulations of the molasses}
\subsection{Damping constant and capture velocity}
\begin{figure}[tb]
\begin{center}
\includegraphics{Fig3.pdf}
\caption{Predicting the temperature of a CaF molasses. For all curves, parameters are $I_{00}=456$~mW/cm$^2$, $\eta=1.29$, and $\Delta_{00}=2.61\Gamma$. (a) Excited state population as a function of speed. (b) Thick, negative blue curve: Acceleration parallel to velocity versus speed. Other curves: speed distribution, $W(v,t)$, for $t=\{0,0.1,0.2,1\}$~ms. The initial distribution has a temperature of $1.4$~mK, and the distributions get narrower with time. (c) Lower, blue curve: simulated temperature as a function of molasses cooling time. Green diamonds: experimental data. Upper green line: fit of experimental data to $T=T_{{\rm final}} + (T_{{\rm init}}-T_{{\rm final}})e^{-t/\tau_{{\rm cool}}}$. (d) Displacement versus time for molecules of various initial speeds, showing which are effectively cooled to rest. Curves are labeled by the initial speed. Inset: fraction of molecules captured in the molasses as a function of starting temperature.}
\label{fig:molassescooling}
\end{center}
\end{figure}
We now compare the results of our simulations to experimental data. We start by considering cooling in the molasses, since this is easier to simulate than the MOT. We simulate the case where $\Delta_{00}=2.61\Gamma$ and $I_{00}=456$~mW/cm$^2$, and assume that the molasses is loaded from a thermal distribution with a temperature of 1.4~mK, which is common in the experiments. Figures \ref{fig:molassescooling}(a,b) show the excited-state population ($N_e$) and the acceleration ($a_v$), as functions of the speed. For a linear damping, $a_v=-\alpha v/m$, we would expect the temperature of an initially hot sample to decay exponentially with a $1/e$ time constant of $t_d=m/2\alpha$. Taking the gradient of $a_v$ near zero velocity, we find {$\alpha=10.2(0.5)\times10^3$ s$^{-1}$}, implying a characteristic damping time of {$t_d=49(2)$ $\mu$s}. However, the $a_v$ curve is only linear for speeds below about 0.5~m/s, while the initial velocity distribution extends to significantly higher speeds. To get a better estimate of the damping time we solve Eq.~\eqref{eq:fpe}, taking the diffusion constant given by Eq.~\eqref{eq:diff}, and the $a_v$ and $N_e$ curves shown in Fig.~\ref{fig:molassescooling}(a,b). The resulting velocity distributions are shown for four different times in Fig.~\ref{fig:molassescooling}(b). From distributions such as these, we obtain the predicted temperature as a function of time, which is shown by the blue line in Fig.~\ref{fig:molassescooling}(c). The figure also shows recent experimental data obtained at an intensity of $I_{00}=456$~mW/cm$^2$, and under conditions where the magnetic field is carefully zeroed and the laser detuning is switched rapidly from the MOT to the molasses phase. The predicted damping time is 101(1) $\mu$s, fairly close to the measured time of 160(30) $\mu$s. The predicted final temperature is 30 $\mu$K, about 5 times lower than the measured value of 144 $\mu$K. We attribute this discrepancy to the incomplete treatment of the diffusion constant, which omits the fluctuations in the dipole force. For a type-I transition in a one-dimensional lin$\perp$lin configuration, the diffusion related to the fluctuations of the dipole force is approximately $\Delta^2/\Gamma^2$ times greater than $D_s$ \cite{Dalibard1989}. There are intensity gradients in the 3D molasses, so we can expect the same mechanism to be present. If the type-II transition exhibits similar scaling, this would lead to a seven-fold increase in the temperature, bringing it closer to the experimentally observed result. A more thorough treatment of the diffusion constant for this multilevel system in 3D is desirable. We also note that at 30 $\mu K$, the thermal de Broglie wavelength of 40~nm is approaching the wavelength of the light, so the validity of the classical treatment of the molecular position and momentum begins to break down. A full quantum treatment of the position and momentum of the molecule is called for to analyze fully these lowest temperature cases.
Next, we consider the capture velocity of the molasses. We calculate $a_v(v)$ for many different intensities, and then use the known laser beam profile to generate the map $a_v(v,r)$, where $r$ is the displacement from the center of the molasses. We then consider a CaF molecule traveling outwards from the center of the molasses, and plot its displacement as a function of time. Figure \ref{fig:molassescooling}(d) shows a series of these curves for several different starting velocities. Here, the parameters are $\Delta_{00}=2.61\Gamma$ and $I_{00}=456$ mW/cm$^2$. The maximum speed at which the molasses is able to bring the molecule to rest is $v_c=3.1$~m/s. To estimate the fraction of molecules that can be cooled, we integrate a Maxwell-Boltzmann speed distribution up to $v_c$ for a range of temperatures. The results are plotted in the inset to Fig.~\ref{fig:molassescooling}(d). For initial temperatures below 5~mK all the molecules are cooled by the molasses. At higher temperatures, a fraction of them escape from the molasses before they can be cooled.
\subsection{Dependence of temperature on intensity, detuning and magnetic field}
To investigate the effects of a background magnetic field, we apply the same procedure as described above, with a randomly oriented uniform field applied. We first solve the OBEs to calculate $a_v(v)$ and $D_s(v)$ for various magnetic field strengths. Figure \ref{fig:molassessensitivity}(a) shows the linear slope of the acceleration close to zero velocity, $\alpha= -\left(\frac{d a_v}{d v}\right)_{v=0}$, and $N_e(0)=\frac{3D_s(0)}{\hbar^2k^2\Gamma}$ as a function of the absolute value of the magnetic field. We see that the damping decreases linearly with magnetic field strength, whereas the diffusion constant increases linearly over the range considered. If we simply use these linear gradients, along with $k_{\rm B} T =D_{s}/\alpha$, we would expect the temperature $T$ in $\mu K$ as a function of the magnetic field $B$ in mT to be $T(B)=21 + 130B + 200B^2+310B^3$, where terms of order $B^4$ have a negligible contribution over the range of $B$ considered here. To investigate the effects of the full velocity dependence of $a_v(v)$ and $D_s(v)$, we find the steady-state temperature using Eq.~\eqref{eq:fpe3}. The results are shown in Fig.~\ref{fig:molassessensitivity}(b), and show a very different dependence to the one expected from the linear approximation made above. The temperature fits well to a purely quadratic dependence on $B$, with a curvature of 1070(50) $\mu$K/(mT)$^2$. The experiment also found a quadratic dependence, but with the larger coefficient of 5740(30) $\mu$K/(mT)$^2$. The discrepancy between predicted and measured curvatures can again be explained by a systematic underestimation of the temperature because of the missing part of the diffusion constant, again by a similar factor of 5.6.
\begin{figure}[tb]
\begin{center}
{\includegraphics{Fig4.pdf}}
\caption{ (a) Excited state population, $N_e$, at zero velocity, and the damping constant, $\alpha$, as a function of magnetic field. (b) Molasses temperature as a function of magnetic field. Parameters are $I_{00}=456$ mW/cm$^2$, $\eta$=1.29 $\Delta_{00}=2.61\Gamma$. Red line: quadratic fit to blue points. (c) Molasses temperature as a function of cooling laser intensity, $I_{00}$, for $\Delta_{00}=2.61\Gamma$. Values of $\eta$ for points running left to right are $\{1.01,1.02,1.05,1.10,1.17,1.29\}$. A thermal distribution with a temperature of 1~mK is loaded into the molasses. Blue circles: steady-state temperatures; Orange crosses: temperatures after 5~ms of cooling; Green diamonds: data from reference \cite{Truppe2017} with $\Delta_{00}=2.5\Gamma$. (d) Molasses temperature as a function of laser detuning, $\Delta_{00}$, with $I_{00}=467$ mW/cm$^2$. Values of $\eta$ for points running left to right are $\{ 1.82,1.80,1.77,1.74,1.71,1.47,1.25,1.12,1.07\}$. A thermal distribution with a temperature of 1.4~mK is loaded into the molasses. Blue circles: steady-state temperatures; Orange crosses: temperatures after 5~ms of molasses cooling.}
\label{fig:molassessensitivity}
\end{center}
\end{figure}
Figure \ref{fig:molassessensitivity}(c) shows how the molasses temperature depends on the laser intensity, $I_{00}$. Here, the blue circles give the predicted steady-state temperature, the orange crosses give the temperature after 5~ms of molasses cooling, and the green diamonds are the experimental data points. At high intensity, we see once again a five fold underestimation of the temperature. As the intensity is decreased, both in the simulation and in the experiment, the temperature is reduced. This is because the low velocity part of $a_v$ is independent of intensity, but lowering the intensity reduces the excited state fraction and hence the diffusion constant. In both the experiment and the simulation, the optimum intensity is around 100~mW/cm$^2$. At lower intensities than this, the temperature rises, and we see a difference between the predicted steady-state temperature and its value after only 5~ms of cooling. This rise in temperature at low intensity occurs because the velocity range of the sub-Doppler force is so low that it can only cool the slowest molecules, so the value of $a_v$ averaged over the velocities of the molecules is reduced. At very low intensities, $I_{00} \lesssim 20$~mW/cm$^2$, we see that the temperature of the molasses after 5 ms is close to 1~mK, which is the temperature at which the molasses was loaded, implying that the molasses no longer cools the distribution at all. In fact, for these intensities the entire molecular distribution is heated to velocities greater than 5 m/s and we could not find a steady-state temperature (which is why there are no blue circles plotted for these low intensities).
Figure \ref{fig:molassessensitivity}(d), shows how the temperature depends on detuning. The temperature decreases rapidly as the detuning is increased from zero, reaches a minimum around $\Delta_{00}=3\Gamma$, exactly as in the experiments, and then gradually increases with further increases in detuning. As can be seen from Fig.~\ref{fig:CaFenergylevel}(b), some of the laser frequency components become resonant with some hyperfine levels as the detuning is scanned over this range. Surprisingly, we do not see any structure in the plot of temperature versus detuning that reflects the hyperfine structure of the molecule. Indeed, at the detuning where the temperature is minimized, the lowest frequency component of the laser is near-resonant with the upper $F=1$ level, but this does not appear to raise the temperature.
\section{Simulations of the MOT}
\subsection{Method for modelling the MOT}
\begin{figure}[tb]
\begin{center}
\begin{tabular}{cc}
\putindeepbox[1pt]{\includegraphics{Fig5.pdf}}
\end{tabular}
\caption{Top: $a_v(v)$ for a molecule at the center of the MOT. Middle: $N_e(v)$ for a molecule at the center of the MOT. Bottom: Blue line is $v^2W(v)$ calculated using $a_v(v)$ and $N_e(v)$, orange dashed line is a Maxwell-Boltzmann speed distribution of temperature T=14.5~mK. Parameters are $I_{00}=234$~mW/cm$^2$, $\Delta=-0.64 \Gamma$ and $\eta=1.49.$}
\label{fig:motPropertiesSimple}
\end{center}
\end{figure}
The MOT is more complicated to model than the molasses, because the position-dependence of the magnetic field and laser intensity modify the force and excited state population. This means that both $a_v$ and $D_s$ depend on the axial and radial displacement from the MOT center, as well as the speed. Before discussing how we deal with the additional complication, let us first focus on the behavior of a molecule at the very center of the MOT. The top two panes in Fig.~\ref{fig:motPropertiesSimple} plot $a_v$ and $N_e$ at the center of a MOT with $I_{00}=234$ mW/cm$^2$ and $\Delta_{00} =-0.64 \Gamma$. The lower pane shows the steady-state molecular speed distribution found from these $a_v(v)$ and $N_e(v)$ using Eq.~\eqref{eq:temp}. The change in sign of the damping force at a speed of around 2.5 m/s, where Doppler cooling turns into Sisyphus heating, leads to a peak in the speed distribution at this point. The temperature of this distribution, calculated using Eq.~\eqref{eq:temp} and Eq.~\eqref{eq:temperatureDefinition}, is 14.5~mK, far higher than the Doppler-limited temperature of 570 $\mu$K for this detuning and intensity. The distribution looks nothing like a Maxwell-Boltzmann distribution at this temperature, which is shown, for comparison, by the orange dashed line. Nevertheless, and remarkably, we find that when a collection of molecules with speeds drawn from this distribution expands freely, their rms width $\sigma$ increases as a function of time according to $\sigma^2(t)=\sigma_0^2+k_BTt^2/m$, with the same temperature $T$ as found above.
Next we consider what happens away from the center of the MOT. Figure \ref{fig:radialField} shows contour plots of $a_v(x,v)$ and $N_e(x,v)$ as a function of speed $v$ and displacement $x$ along the $x-$axis. In (a), cooling is indicated in blue-yellow, and heating in red-orange. In (b), darker regions indicate low excited-state probability, and lighter regions indicate higher probability. Focusing on the heating at low velocity, we see that as the distance from the center of the MOT increases, the gradient of the force curve at zero velocity decreases, the maximum value of the force also decreases, and the range over which the polarization-gradient heating acts is reduced. All are due to the increasing magnetic field which reduces the effectiveness of the polarization-gradient force. Around 5~mm from the center the heating force has almost vanished, and the force curve is dominated by Doppler cooling. At even larger distances, the decreasing laser intensity reduces the Doppler cooling force. Figure \ref{fig:radialField} (b) shows that as the displacement from the MOT center increases towards $\sim 3$~mm, the excited-state population increases for all speeds, but increases most strongly when the speed is low. Since the temperature in the MOT is primarily determined by the zero crossing point of $a_v (v)$, these changes in $N_e$ do no have much effect on the temperature. They affect the total scattering rate of the MOT, tending to make it brighter than the equivalent molasses. Once the displacement grows larger than around 7.5~mm, the decreasing laser intensity strongly reduces the excited state fraction. Experimentally, it is found that for the choice of parameters used here, the radial rms width of the cloud is about 2.0~mm. Therefore, for accurate predictions of the MOT properties, we must account for the position dependence of $a_v$ and $N_e$.
\begin{figure}[tb]
\begin{center}
\includegraphics{Fig6.pdf}
\caption{(a) $a_v$ and (b) $N_e$ as a function of displacement along the $x$ axis and speed $v$. Parameters are: $I_{00}=468$ mW/cm$^2$, $
\Delta=-0.64\Gamma$. $\eta$ ranges from $1.65$ in the center to $1.09$ at $x=10$~mm.}
\label{fig:radialField}
\end{center}
\end{figure}
Using our calculation of $a_v(x,v)$ and $N_e(x,v)$, it is possible to calculate the complete phase-space distribution by numerically solving Eq.~(\ref{eq:fpe}). Because this is difficult, we adopt a simpler method where we assume that the density distribution is already known from experiment, and that momentum and position are uncorrelated. In this case, the phase-space distribution is $W(\boldsymbol{x},\boldsymbol{v},t)=U(\boldsymbol{x})P(\boldsymbol{v},t)$. Integrating Eq.~(\ref{eq:fpe}) over the spatial coordinates, and using the fact that $U(\boldsymbol{x})\rightarrow 0$ as $\{x,y,z\} \rightarrow \pm \infty$, we obtain
\begin{equation}
\frac{\partial}{\partial t}v^2P(v,t)=\frac{\partial }{\partial v}\left(\frac{-\mathcal{F}(v)}{m}v^2P(v,t)+\frac{v^2\mathcal{D}_{S}(v)}{m^2}\frac{\partial P(v,t)}{\partial v}\right),
\label{eq:fpeMOT}
\end{equation}
where
\begin{align}
\mathcal{F}(v)&=\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty} U(\boldsymbol{x})\tilde{F}(\boldsymbol{x},v)\textrm{d}x \textrm{d}y\textrm{d}z \,,\\
\mathcal{D}_{s}(v)&=\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty} U(\boldsymbol{x})\tilde{D}_{s}(\boldsymbol{x},v)\textrm{d}x \textrm{d}y\textrm{d}z\, ,
\end{align}
are the position-weighted averages of the force and diffusion constant. Note that, similar to Eq.~(\ref{eq:diff}), $\mathcal{D}_{s}$ is related to the position-weighted excited-state population, $\mathcal{N}_{e}$, as $\mathcal{D}_{s}(v) = \hbar^2 k^2 \Gamma \mathcal{N}_{e}(v)/3$.
The force is the sum of a trapping term, which is antisymmetric under $\boldsymbol{x}\rightarrow\boldsymbol{-x}$, and a damping term which is symmetric. Assuming $U(\boldsymbol{x})$ is symmetric, we see that the trapping term makes no contribution to the integral and we only need to use the damping term. To perform the integrals, we find $\tilde{F}(\boldsymbol{x},\boldsymbol{v})$ and $\tilde{D}_s(\boldsymbol{x},\boldsymbol{v})$ by repeatedly solving the OBEs using the methods discussed previously, where the intensity and magnetic field take their values at position $\boldsymbol{x}$ in the MOT. We take $U(\boldsymbol{x})$ to be the density distribution measured in the experiment,
\begin{align}
U(\boldsymbol{x})&=\frac{1}{\sigma_x^2\sigma_z(2\pi)^{3/2}}e^{-\frac{x^2+y^2}{2\sigma_x^2}} e^{-\frac{z^2}{2\sigma_z^2}}\, ,
\end{align}
where $\sigma_x$ and $\sigma_z$ are the measured rms widths in the radial and axial directions. We then calculate the weighting integrals to arrive at $\mathcal{F}$ and $\mathcal{D}_{s}$. As we shall see, this approach reproduces many of the the observed properties of the MOT.
\subsection{MOT properties}
Measurements of the MOT temperature as a function of $I_{00}$~\cite{Williams2017} are plotted as green diamonds in Fig.~\ref{fig:motProperties}(a), while the calculated steady-state temperatures found by solving Eq.~(\ref{eq:fpeMOT}) are shown as blue circles in this figure. We see fairly good agreement for the points between 10 and 500~mW/cm$^2$. Below 10~mW/cm$^2$, the predicted steady-state temperature decreases with decreasing intensity, whereas the measured points trend upwards. Here, at least part of the discrepancy might be caused by the long thermalization time of the MOT at these low intensities. In the experiments, the intensity is first held at 468~mW/cm$^2$ to load the MOT, then ramped down to its new value over 20~ms, then held at this value for 5~ms, before measuring the temperature. We replicate this procedure numerically by first calculating the velocity distribution for a MOT with $I_{00}=468$~mW/cm$^2$, then calculating how this distribution is modified when $I_{00}$ is varied as in the experiment. The results of these simulations are shown as orange crosses in the figure. At high intensities, this procedure gives the same temperature as in the steady-state (the orange crosses lie on top of the blue circles). At lower intensities, we see that the 20~ms cooling ramp and/or 5~ms hold time are not slow enough for the temperature to reach the steady-state. As a result, the temperature rises at low $I_{00}$. The agreement between simulation and experiment is reasonably good over the whole range of intensities (spanning 3 orders of magnitude). Notably, our model gives far better predictions of the temperature than the Doppler-limited temperature predicted by a rate equation model, which is plotted as the black dashed line in the figure.
\begin{figure}[htb]
\begin{center}
\includegraphics{Fig7.pdf}
\caption{A comparison between experimentally measured properties of the CaF MOT \cite{Williams2017} (green diamonds) and the numerical simulations (other points) as a function of $I_{00}$. The detuning is $\Delta=-0.64\Gamma$ in the simulations and $\Delta=-0.75\Gamma$ for the experimental data. (a) Temperature. Blue circles show the predicted steady-state temperature. Orange crosses show the temperature following the intensity ramp performed in the experiment (see main text). The black dashed line is the predicted Doppler cooling limit (see Eq.(14) of Ref.~\cite{Williams2017}). (b) Photon scattering rate. (c) Damping rate for radial oscillations. Red triangles: damping rate inferred from the gradient of the acceleration curve at zero velocity. Blue circles: damping rate inferred from the time constant for the temperature to approach equilibrium. (d) Frequency of radial oscillations.}
\label{fig:motProperties}
\end{center}
\end{figure}
Once Eq.~(\ref{eq:fpeMOT}) has been solved in the steady state to give $P(v)=\lim_{t\to\infty} P(v,t)$ at each intensity, it can be used to calculate the scattering rate, which is
\begin{align}
R_{\rm sc} = \Gamma \int_{0}^{\infty} \mathcal{N}_e(v) P(v)\,4\pi v^{2}dv\,.
\end{align}
This scattering rate is plotted in Fig.~\ref{fig:motProperties} (b), and is seen to agree excellently with the measured scattering rate. By comparison, the rate equation model, which cannot capture optical pumping into transient dark states, overestimates the scattering rate by a factor of 2~\cite{Williams2017}.
Experimentally, it is found that following a radial push, the cloud executes damped harmonic motion, with the position, $r$, of the center of the cloud, following the equation
\begin{align*}
r''=-\omega^2r-b r'\, .
\end{align*}
Here, $\omega$ is the angular oscillation frequency, and $b$ is the damping rate. The green diamonds in Fig.~\ref{fig:motProperties} (c) shows the measured damping rate as a function of intensity. We can determine the damping rate from the simulations in two different ways. Firstly, we can calculate the slope of $\mathcal{F}(v)/m$ around the zero crossing velocity $v_0$, $\alpha'= -\left(\frac{d a_v}{d v}\right)_{v=v_0}$, which should be equal to $b$. The results obtained this way are plotted as red triangles in Fig.~\ref{fig:motProperties} (c). Alternatively, we can perturb the steady state distribution $P(v)$ by changing it to $P(v-v_{\rm push})$, where $v_{\rm push}=2.3$~m/s is the speed given to the molecules in the measurements of the damped oscillations. We then find $\tau_{\rm d}$, the $1/e$ time constant with which the temperature is damped, which should be related to the velocity damping constant by $b=1/(2\tau_{\rm d})$. These points are plotted as blue circles in this figure. Both methods yield similar results, and both have the same qualitative shape as the experimentally measured distribution. At low intensities the agreement between the simulations and the experiment is good. However, the simulations overestimate the damping constant by a factor of 3--5 at higher intensities. While not perfect, the agreement here is again better than the rate equation model, which overestimates the damping constant by a factor of 2-3 at low intensities and 5-10 at higher intensities.
Figure \ref{fig:motProperties} (d) shows the trap oscillation frequency $\omega/(2\pi)$ as a function of $I_{00}$.
The trapping force is much weaker than the damping forces or local dipole forces, so special care must be taken to resolve its contribution. The procedure is as follows. We apply a homogeneous magnetic field along $(\boldsymbol{\hat{x}}+\boldsymbol{\hat{y}})/\sqrt{2}$ for a particular $I_{00}$. We fix the speed at $v=3$ m/s and we solve the OBEs repeatedly for a random choice of directions and for a range of positive and negative magnetic fields. We calculate the component of the force along $(\boldsymbol{\hat{x}}+\boldsymbol{\hat{y}})/\sqrt{2}$ and average together the results from many simulations. For each choice of direction, we ensure that the exact opposite direction is included in the set of simulations. This ensures that the dominant velocity-dependent part of the force (ideally) averages to zero. Because, in the MOT, the magnetic field is proportional to the displacement, the slope of the acceleration at zero magnetic field is proportional to $\omega^2$. The uncertainty in determining $\omega$ is large, because of the noise from the residual damping force. Nevertheless, the oscillation frequencies agree reasonably well with the experimentally measured results. In the CaF MOT, the confining force is expected to arise primarily from a dual-frequency effect between the two laser components closest in frequency to the $F=2$ hyperfine component~\cite{Tarbutt2015b}. This dual-frequency effect should only occur if the two frequency components have opposite polarizations. We have simulated the MOT trapping force for the case where all frequency components have identical polarization, and find that the trapping force vanishes. This result verifies that the dual-frequency mechanism is indeed the mechanism responsible for trapping the molecules.
The capture velocity of the MOT can be found by calculating the maximum velocity a molecule can have as it enters the MOT region (taken to be at a radius of 10~mm) if it is to be captured. Using the force map $\tilde{F}(\boldsymbol{x},v)$ for the radial plane, at an intensity of $I_{00}=468$ mW/cm$^2$, we calculate a capture velocity of 14~m/s. This is close to the measured value of {$11.2^{+1.2}_{-2.0}$ m/s}~\cite{Williams2017}.
\section{Summary and conclusions}
We have presented a general method for modeling laser cooling and magneto-optical trapping of atoms and molecules, and used the method to understand recent results from experiments with CaF molecules. Our method uses generalized optical Bloch equations to calculate the three-dimensional steady-state force and momentum diffusion constant, taking into account all relevant levels of the molecule and all frequency components of the light. Then, we use the Fokker-Planck-Kramers equation to determine the evolution of the velocity distribution, and the steady-state distribution. Our simulation results show broad agreement with experimental results across a wide range of parameters, and help to improve our understanding of laser cooling and magneto-optical trapping of molecules.
In our previous work~\cite{Devlin2016}, we considered model systems with just one hyperfine ground state and one hyperfine excited state. We found that, for type-II systems driven by red-detuned light, there is Doppler cooling at high speed, but Sisyphus heating at low speed. For blue-detuned light the forces are reversed. In the present work, we have modeled all the levels of CaF relevant to laser cooling, and find that the velocity-dependent force curves are very similar to those found for the simpler systems. Notably, the force curve crosses zero at two speeds: zero, and a specific speed where the Doppler and Sisyphus forces cancel. By solving the Fokker-Planck-Kramers equation, we find the velocity distributions resulting from these unusual force curves. When the light is blue-detuned, and the initial velocity is small enough, the velocity is damped towards zero and the distribution is approximately thermal. In this regime, the temperature can be determined from the damping constant at low velocity, and the momentum diffusion constant. When the light is red-detuned, the velocity distribution is far from being a thermal distribution, and peaks near the special velocity where the force crosses zero, as we would intuitively expect. Despite its non-thermal nature, we find that the ballistic expansion of a cloud with such a velocity distribution is similar to that of a Maxwell-Boltzmann distribution, allowing a temperature to be assigned. This temperature is predicted to be about 10~mK at the highest intensities used in the experiments, and is predicted to decrease as the laser intensity is lowered, mainly because the zero crossing of the force curve shifts to lower speed at lower intensity. All these predictions agree well with experimental results. Thus, we quantitatively confirm the hypothesis that the balance between Doppler cooling and Sisyphus heating is responsible for the high temperatures observed in molecular MOTs, and also in type-II atomic MOTs. In this regime, momentum diffusion does not play a strong role in determining the temperature.
The excited-state population calculated from the optical Bloch equations is smaller than predicted by a rate model~\cite{Tarbutt2015,Tarbutt2015b} at all speeds, and drops dramatically at low speeds as the molecules are optically pumped into transient dark states. We find that dark states formed between Zeeman sub-levels of a particular hyperfine level, and Raman dark states which are superpositions of Zeeman sub-levels from different hyperfine states, all play a role in reducing the scattering rate. Similarly, non-adiabatic transitions out of these dark states are important to the Sisyphus cooling / heating mechanism in this system. Our calculated scattering rate in the MOT agrees very well with the measured values across a wide range of intensities. We note that the excited-state population does not drop to zero at zero velocity, showing that there are no time-independent dark states. This is because different frequency components of the light have different polarizations. When a type-II transition is driven by two components of light, there is a time-independent dark state if they have the same polarization (but different detunings), or if they have the same detuning (but different polarizations). If the different frequency components have different polarizations and different detunings, we find there is no time-independent dark state, except in some special cases.
Previous work using a rate model approach concluded that the confining force in a static MOT of CaF is due mainly to a dual-frequency mechanism that arises when two frequency components of opposite polarisation address the same transition with different detunings~\cite{Tarbutt2015b}. This mechanism, which is also analyzed in Ref.~\cite{Cournol2016}, provides both Doppler cooling and strong confinement in cases where little or no confining force is present with one frequency component alone, or with two components of the same polarization. Our simulations using the OBEs support the conclusion that this dual-frequency mechanism is responsible for the trapping force in the CaF MOT.
Simulations of molecules loaded from a MOT into a molasses predict that the temperature drops on a timescale of about 100~$\mu$s, similar to what is measured. Heating in the molasses is due to the randomness of photon absorption and spontaneous emission events, and due to fluctuations of the dipole force. Our model does not include the last of these, which is particularly difficult to calculate. Across a wide range of parameters, the model consistently predicts steady-state temperatures 3-6 times lower than measured, indicating that the dipole force fluctuations contribute significantly to heating of the molasses. A method for treating this heating mechanism for a multi-level system in 3D would be valuable. At low velocities, the acceleration curve is independent of intensity, but the scattering rate decreases with decreasing intensity. As a result, the temperature of the molasses decreases with decreasing intensity. This is seen in the simulations and the experiments. The cooling time gets longer at low intensities however, and at very low intensities there seems to be no cooling at all. Applying a magnetic field to the molasses increases the scattering rate at zero velocity, and decreases the damping constant, so the temperature increases with magnetic field. The model predicts a quadratic dependence of the temperature on magnetic field, which is also the dependence seen experimentally.
The methods presented in this work can also be used to study laser cooling and trapping of other diatomic or polyatomic species, or for investigating unusual magneto-optical trapping arrangements. To this end, we have presented the equations in a general form so that they can be used by others. We have already used our model to study a blue-detuned MOT of $^{87}$Rb~\cite{Jarvis2018}, and to investigate laser cooling of SrF and YbF molecules~\cite{Lim2018}, where we find the same qualitative behavior as described here for CaF. Future applications include the study of MOTs for molecules with very different energy level structures, the investigation of $\Lambda$-enhanced gray molasses cooling~\cite{Cheuk2018} and other novel cooling schemes, and the study of laser cooling within optical dipole traps~\cite{Anderegg2018}.
Underlying data may be accessed from Zenodo\footnote{\lowercase{10.5281/zenodo.1473592}} and used under the Creative Commons CCZero license.
\acknowledgements
This research was supported by STFC grant no. ST/N000242 and by EPSRC under grants EP/P01058X/1 and EP/M027716/1. We are grateful to the CaF team at Imperial College London for helpful discussions and for providing experimental data.
|
\section*{Abstract}
Stability of running on rough terrain depends on the propagation of perturbations due to the ground.
We consider stability within the sagittal plane and model the dynamics of running as a two-dimensional body with an alternating aerial and stance phase.
Stance is modeled as a passive, impulsive collision followed by an active, impulsive push-off that compensates for collisional losses.
Such a runner has infinitely many strategies to maintain periodic gaits on flat ground.
However, these strategies differ in how perturbations due to terrain unevenness are propagated.
Instabilities manifest as tumbling (orientational instability) or failing to maintain a steady speed (translational instability).
We find that open-loop strategies that avoid sensory feedback are sufficient to maintain stability on step-like terrains with piecewise flat surfaces that randomly vary in height.
However, these open-loop runners lose orientational stability on rough terrains whose slope and height vary randomly.
Only by avoiding tangential collisions is orientational stability recovered.
Tangential collisions may be avoided through leg-retraction to match foot and ground speed at touch down.
By analyzing the propagation of perturbations, we derive a single dimensionless parameter that governs stability and guides the design and control of both biological and robotic runners.
\clearpage
\section{Introduction}
Legged terrestrial animals run stably on rough terrains, despite potential difficulties such as sensory latencies and the highly dynamic nature of running.
Our current understanding of how running animals negotiate rough terrains is based on studies where the animal experiences obstacles in the form of a single step up or down \citep{Daley2006,Birn2012}, or a random sequence of up and down steps \citep{Grimmer2008,Voloshina2015}.
However, natural terrains exhibit not only variations in height but also in slope, and it is unclear how our understanding of running on step-like terrains translates to such natural terrains.
Mathematical studies of running over rough terrains reflect the experiments and focus on stability when running on step-like terrains that are piecewise flat \citep{Daley2010,Blum2014,Karssen2015}.
Furthermore, models of runners with massless legs, such as the spring-legged-inverted-pendulum (SLIP) \citep{Blickhan1989,McMahon1990,Blickhan1993}, cannot distinguish between different slopes of the terrain and only respond to variations in height.
Assuming a massless leg enforces the ground reaction force vector to always align with the leg \citep{Srinivasan2008} regardless of the terrain's slope beneath the foot.
More detailed models that mimic the anatomy of specific animals or robots avoid this limitation of SLIP models \citep{Karssen2015}, at the cost of generalizability.
Thus there is a need for generalizable models of running that incorporate dependence on both terrain slope and height, and yet remain sufficiently abstract to glean principles that may underlie stability on rough terrains.
Stability may be governed by many factors, including sensory feedback control \citep{Pearson1993,Pearson1995,Dickinson2000}, the inherently stabilizing mechanical response of the animal's body \citep{Holmes2006}, energy dissipation within the body \citep{Daley2006}, and feed-forward strategies such as swing-leg retraction \citep{Seyfarth2003}.
The slowest is often sensory feedback control that has latencies comparable to or greater than the stance duration.
For example, at endurance running speeds for humans, the stance lasts around 200~ms \citep{Cavagna1964} and only slightly longer than the shortest proprioceptive feedback delay of 70--100~ms or visual feedback delay of 150--200~ms \citep{vanBeers2002}.
To better understand the inherent stability or instability of the dynamics of running, we consider only passive mechanical and anticipatory strategies in this study without relying on active feedback control.
Studies of running birds and the role of open-loop stability of running find that increased energy dissipation during stance may help stability when faced with an unexpected drop in terrain height \citep{Daley2006}.
Consistent with the role of energy dissipation, experiments with humans find that metabolic power increases by 5\% to run on step-like terrains versus flat ground \citep{Voloshina2015}.
Walking over rough terrains leads to an increase of 28\% in metabolic power \citep{Voloshina2013}, higher in both relative and absolute terms.
The difference in energetics may indicate that the dynamics of running are inherently less unstable, but such an analysis on natural rough terrains has not been carried out.
Therefore, we incorporate energy dissipation in our examination of open-loop strategies to address the effect of dissipation on stability.
Not relying on feedback control within a single stance does not preclude active strategies that rely on anticipation or internal models, sometimes called feed-forward strategies.
Computational studies of walking have demonstrated the role of look-ahead strategies that use the height and slope of the oncoming terrain in planning the control \citep{Byl2009}.
Evidence for the importance of feed-forward strategies for running come from computational studies of SLIP-like running dynamics \citep{Seyfarth2003} that show how swing-leg retraction automatically modulates the landing angle in response to unexpected variations in the terrain height.
However, these studies on running have not yet considered the effect of slope variations in the terrain.
Thus in our study, we analyze anticipatory strategies that incorporate the slope of the oncoming terrain.
An extreme and simplified approximation of running is that of a point mass with an impulsive and instantaneous stance followed by projectile flight.
Such an approximation appears as a natural solution to the problem of minimizing measures of metabolic energy consumption when the desired forward speed exceeds critical levels, and subject to other constraints such as step length \citep{Srinivasan2006a}.
SLIP-like models are an unfolding of these point-mass instantaneous-stance models to have finite stance duration.
They have helped us understand the kinetics of stance \citep{Holmes2006} and the energetics of producing forces \citep{Srinivasan2006a} on flat terrains, and the role of swing-leg retraction on piecewise flat terrains \citep{Blum2014}.
However, these point-mass models possess no sense of body orientation during the aerial phase and are therefore immune to falling by tumbling.
In this study, we unfold the point-mass, instantaneous-stance model by using a finite moment of inertia for the runner, while still maintaining an impulsive stance.
A finite moment of inertia defines a body orientation and thus enables the examination of the effect of angular momentum fluctuations induced by stance.
Such a model maps the net effect of the ground forces over stance as a linear impulse applied at the contact point and an angular impulse applied at the center of mass.
These impulses lead to a change in the linear and angular momentum of the whole runner because of the passive and active forces during stance.
As we discuss later, the angular impulse captures the effect of a finite stance duration and configuration changes during stance.
Section~\ref{sec:model} develops a sagittal-plane model that incorporates a finite moment of inertia, inelastic 2D collisions, and an active push-off so that both terrain slope and height variations affect stability.
In section~\ref{sec:mc}, we use Monte Carlo simulations with random variation of ground height and slope to examine open-loop strategies, the effects of energy dissipation, and strategies that anticipate the slope of the terrain.
Section~\ref{sec:stability} derives the linearized dynamical equations and analyzes their stability.
Using the linearization, we find a single dimensionless parameter that governs stability in section~\ref{sec:paramdepend}, which in turn guides morphological design for stability.
We conclude in section~\ref{sec:discussion} with a discussion of the limitations and generality of our analyses, its relationship to experimental results, and generate testable predictions for future experiments.
\section{Mathematical model of sagittal plane running}
\label{sec:model}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{model_fail_cood.pdf}
\caption{Bouncing as a model of running.
{\bfseries a}, The outline of a human running at 3.5~m/s, created from motion capture data, shows stance and aerial phases over a single step. The stance leg and ipsilateral arm are in red, and the center of mass trajectory is shown as a blue, dashed curve.
{\bfseries b}, The runner pushes-off the ground by applying a linear impulse $\vec{J}_{\rm imp}$ at the contact point P, and the effect of additional torques about the center of mass arising from configuration changes during stance are captured by an angular impulse $J_{\phi}$ at the center of mass.
At the end of the aerial phase, the runner undergoes a passive collision with the ground at the new contact point P.
The momentum lost due to the collision in directions tangential and normal to the terrain surface is dictated by the parameters $\et$ and $\en$, respectively.
{\bfseries c}, The runner can fail in two ways: \emph{orientational failure} when orientation at touchdown exceeds the tip-over threshold, i.e.\ $\abs{\phi^-} > \phi_{\rm tol}$, or \emph{translational failure} when the forward velocity at take-off drops below a chosen threshold, e.g.\ $v_{\rm G,x}^+<0.01$.}
\label{fig:problemsetup}
\end{figure}
We model the runner in the sagittal plane as a rigid body (Fig.~\ref{fig:problemsetup}a) of mass $m$, radius of gyration $r_g$, i.e.\ moment of inertia $\I = m r_{g}^2$ about its center of mass, and radius $r_\ell$ (leg length). All quantities are in units such that $m = 1,r_\ell = 1$ and the acceleration due to gravity $g=1$.
See section~\ref{sec:notation} for notation used in this paper.
\subsection{Aerial and stance phases}
\label{sec:running dynamics}
A single step is comprised of an aerial and a stance phase.
The aerial phase is modeled as a drag-free projectile in uniform gravity.
Stance involves two successive parts: a passive collision with the ground followed by an active push-off.
The passive collision is two-dimensional and parameterized by two coefficients of restitution $\en$ along the normal to the ground and $\et$ along the tangent to the ground.
The active push-off applies a linear impulse $\vec{J}_{\rm imp}$ at the contact point P and a rotational impulse $J_{\phi}$ at the center of mass G.
The governing dynamical equations are,
\begin{subequations} \label{eqn:dynamics of running}
\begin{eqnarray}
\label{eqn:passive}
\text{passive collision:}\ &
&\vec{v}_{\rm P}^c = \begin{pmatrix}
\epsilon_{\rm t} & 0 \\
0 & -\epsilon_{\rm n} \end{pmatrix} \vec{v}_{\rm P}^-, \\
\label{eqn:ang-mom}
& &{H}_{\rm /P}^c - {H}_{\rm P}^- = 0,\\
\label{eqn:pushoff}
\text{push-off:}\ &
&\vec{v}_{\rm G}^+ = \vec{v}_{\rm G}^c + \vec{J}_{\rm imp}, \quad \I\omega^+ = \I\omega^c + J_{\rm imp,t} + J_{\phi}, \\
\label{eqn:flight}
\text{flight:}\ &
&\ddot{x}_{\rm G}(t) = 0,\quad \ddot{y}_{\rm G}(t) = -1,\quad \ddot{\phi}(t) = 0,\ \text{and}\\
\label{eqn:flight initial conditions}
\text{initial conditions:}\ &
&\begin{pmatrix}x_{\rm G}(0) \\ y_{\rm G}(0)\\ \phi(0) \end{pmatrix} = \begin{pmatrix} x_{\rm G}^+ \\ y_{\rm G}^+\\ \phi^+ \end{pmatrix},\; \begin{pmatrix} \dot{x}_{\rm G}(0) \\ \dot{y}_{\rm G}(0)\\ \dot{\phi}(0) \end{pmatrix} = \begin{pmatrix} v_{\rm G,x}^+ \\ v_{\rm G,y}^+\\ \omega^+ \end{pmatrix}.
\end{eqnarray}
\end{subequations}
Horizontal and vertical positions are denoted by $x$ and $y$, respectively, orientation by $\phi$, velocity by $\vec{v}$, angular velocity by $\omega$, moment of inertia by $\I$,
and angular momentum by ${H}$.
Superscript `$-$' denotes variables immediately preceding the collision, `$c$' after the passive collision and `$+$' after the active push-off.
Subscripts P and G refer to quantities associated with the foot and center of mass, respectively
The mechanical state of the runner is parameterized by the center of mass positions $(x_{\rm G}, y_{\rm G})$, body orientation $\phi$, and their respective velocities $(v_{\rm G,x}, v_{\rm G,y})$ and $\omega$.
Because the stance is assumed to be instantaneous, the velocities may change discontinuously but the position and orientation remain constant during stance.
The instantaneous stance assumption also implies that unmodeled finite forces such as gravity or air-drag do not contribute to the impulse on the runner.
However, the active rotational impulse ${J}_{\phi}$ applied at the center of mass G captures the effects of varying posture over stance and the changing center of pressure on the ground.
We examine this approximation and its implications in the discussion.
\subsection{Stance: passive collision}
\label{sec:eten}
The runner may control the passive collisional impulse with the ground by varying the parameters $\en$ and $\et$.
Because the collisional impulse passes through P, the angular momentum of the runner about the contact point ${H}_{\rm /P}$ does not change (equation~\eqref{eqn:ang-mom}) and governs the change in the angular velocity $\omega$ due to the collision.
The passive normal collision can vary from perfectly inelastic to perfectly elastic, and is parameterized by the normal coefficient of restitution $0\le\en\le 1$.
The fraction $\epsilon_{\rm n}^2$ models elastic energy stored and recovered during stance.
When $\en = 0$, the runner is completely dissipative and $\en = 1$ implies perfectly energy conserving.
The tangential coefficient of restitution $\et$, parameterizes the tangential impulse when the foot undergoes a collision with the ground.
Modulation of $\et$ is the feature that distinguishes open-loop versus anticipatory strategies in our model.
Consider the example where the runner modulates $\et$ by varying the tangential foot speed at touch-down.
The open-loop runner would vary the foot speed by assuming that the terrain is flat and that its own mechanical state matches that of a perfectly periodic and steady speed runner.
On a rough terrain the mechanical state and the terrain slope vary from step-to-step.
Thus the intended tangential collision $\etr$ and the actual tangential collision $\et$ may not be equal for the open-loop runner.
The anticipatory runner would use information of the terrain's slope in the oncoming step and its own mechanical state to make sure that the intended and actual foot speed at touchdown match.
Thus the actual and the intended (controlled) tangential collision $\etr$ are equal for the anticipatory runner.
The relationship between $\etr$ and $\et$ are therefore,
\begin{equation}
\et =
\begin{cases}
\etr \left(\frac{v_{\rm x0}}{v_{\rm P,t}^-}\right) &\text{: open-loop}, \\
\etr &\text{: anticipatory},\\
\end{cases}
\label{eqn:omega-etr}
\end{equation}
where $v_{\rm P,t}^-$ is the tangential velocity of P just before landing and $v_{\rm x0}$ is the steady forward velocity of the center of mass on flat ground.
The two policies are identical on flat ground and when the body has no angular velocity prior to landing.
A numerical examination of the relationship between $\etr$ and $\et$ on rough terrain, is presented in supplement~\ref{ext-sec:et-vs-etc}.
\subsection{Stance: active push-off}
\label{sec:pushoff}
Stance ends with the application of an active, linear push-off impulse $\vec{J}_{\rm imp}$ at the contact point P and an active angular push-off impulse $J_\phi$ at the center of mass G.
We constrain these impulses so that in the absence of external perturbations or other disturbances the runner is perfectly periodic and remains upright ($\phi(t)=0$) on flat ground.
Importantly, once the impulses are chosen for flat ground, they are not allowed to vary step-to-step on any other terrain to reflect the absence of active feedback control.
Together, these conditions imply that that the active push-off impulses $\vec{J}_{\rm imp}$ and $J_\phi$ depend only on $\en$, $\et$, $v_{\rm x0}$ and $v_{\rm y0}$, and no other parameters, according to
\begin{subequations}
\label{eqn:jimp}
\begin{eqnarray}
\label{eqn:jimp1}\vec{j}_{\rm imp} &=& \begin{pmatrix} v_{\rm x0} \\ v_{\rm y0} \end{pmatrix} - \begin{pmatrix} (\et + \frac{1-\et}{1 + I_{\rm /G}})v_{\rm x0} \\ -\en v_{\rm y0} \end{pmatrix},\\
\label{eqn:jimp2}\quad j_\phi &=& 0
\end{eqnarray}
\end{subequations}
Thus the center of mass of a periodic runner on flat ground has a constant forward speed $v_{\rm x0}$ and vertical speed $v_{\rm y0}$ at every step.
On rough terrains, there are two options for defining the application of the invariant linear impulse on every step.
First, the impulse vector $\vec{J}_{\rm imp}$ may be held constant in every step with respect to gravity (x-y frame in Fig.~\ref{fig:problemsetup}a), which we call the \emph{lab-fixed} push-off policy.
Second, the impulse vector may be held constant in every step with respect to the normal direction to the terrain at the point of contact (t-n frame in Fig.~\ref{fig:problemsetup}b), which we call a \emph{terrain-fixed} push-off policy.
The terrain-fixed policy may be considered a better approximation of what animals do, because the normal to the terrain and the leg orientation are often coupled, whereas leg orientation at contact and gravity may vary from step to step.
Implicit in preferring the terrain-fixed policy is the assumption that joint torques to apply forces are planned in an ego-centric (body-fixed) frame of reference.
For the disc-like model of a runner that we use, the terrain-fixed and body-fixed policies are identical.
Detailed expressions for the velocities in the stance phase, as well as expressions for $\vec{J}_{\rm imp}$ under both push-off policies are given in supplement~\ref{ext-sec:modeldetails}.
We present an complete analysis of the lab-fixed push-off policy in supplement~\ref{ext-sec:labpushoff} and focus on the terrain-fixed policy in the main paper.
\section{Monte Carlo simulations}
\label{sec:mc}
A sagittal plane runner can only fail by two modes, when the body orientation exceeds a chosen threshold (\emph{orientational failure}), or by failing to move forward any longer (\emph{translational failure}).
We choose the orientational threshold $\phi_{\rm tol}$ as the angle of tilt to passively topple a human who is standing with their feet apart in a pose resembling double-stance in walking.
We perform Monte Carlo~ simulations on step-like and undulating rough terrains to estimate the statistics of failure for both open-loop and anticipatory runners.
Stability is quantified by the mean steps to failure, like previous studies of rough terrain walking \citep{Byl2009}.
The terrain is modeled as a piecewise linear interpolation of an underlying random grid.
The grid points are separated by a distance $\lambda$ and the heights $h$ of the grid points are chosen from a uniform random distribution (table~\ref{table:parameters}, section~\ref{sec:terrain model}).
A linear interpolation between the grid points yields a terrain with random variations in both slope and height.
Corners at grid point implies an indeterminate slope, and we therefore define an effective slope at the grid points by interpolating the slope before and after the point (details in section~\ref{sec:landing}).
Parameter values that represent a human-like runner (table~\ref{table:parameters}) are used for all Monte Carlo~ simulations, unless indicated otherwise.
\begin{table}
\centering
\begin{tabular}{*9c}
\toprule
\multicolumn{5}{c}{Runner} & \multicolumn{2}{c}{Terrain} & \multicolumn{2}{c}{Monte Carlo} \\
\addlinespace
$I_{\rm /G}$ & $\epsilon_{\rm n}$ & $\phi_{\rm tol}$ & $v_{\rm x0}$ & $v_{\rm y0}$ & $h$ & $\lambda$ & $M$ & MAX \\
\cmidrule(lr){1-5}\cmidrule(lr){6-7}\cmidrule(lr){8-9}
0.17 & 0.63 & $\pi/6$ & 0.96 & 0.26 & $\thicksim\mathcal{U}(-0.03, 0.03)$ & 0.1 & $10^5$ & $10^3$ \\
\bottomrule
\end{tabular}
\caption{Parameter values for a human-like runner.
Units are chosen such that $m=1$, $g = 1$, and $r_\ell = 1$.
Parameters describing the runner are discussed in section~\ref{sec:model}.
The heights $h$ at grid points defining the terrain are chosen from a uniform distribution over the range $[-0.03,0.03]$ (section~\ref{sec:terrain model}).
The ensemble size used in the Monte Carlo~ simulations is $M$, and MAX is the number of steps to which the runner is simulated.
All runners failed before reaching MAX.
We elaborate on the choice of these values in section~\ref{sec:simulation}.
}\label{table:parameters}
\end{table}
\subsection{Open-loop runners on rough terrains}
\label{sec:openloop}
Open-loop runners always fail through an orientational instability on rough terrains regardless of the energy dissipated per step (Fig.~\ref{fig:Large perturbations}a,d).
The open-loop runners with human-like inertia and size take $9.6 \pm 4.1$ steps (mean $\pm$ standard deviation) before tumbling while only 1\% of the runners fall within 3 steps (Fig.~\ref{fig:Large perturbations}a).
Decreasing $\en$ from 1 to 0 increases the mean steps to failure by just 2 steps (Fig.~\ref{fig:Large perturbations}d).
Thus, dissipating more energy per step in the normal collision has minimal influence on stability.
The tangential collision, parameterized by $\etr$, has little or no influence because the contour lines of the mean steps to failure are nearly parallel to the $\etr$ axis (Fig.~\ref{fig:Large perturbations}d).
Therefore, energy dissipation or modulating the tangential collision are both ineffective stabilization strategies for purely open-loop runners.
\subsection{Open-loop runners on step-like terrains}
\label{sec:step-like terrains}
The purely open-loop runner remains stable on step-like terrains that are piecewise flat and possess only height variations (Fig.~\ref{fig:Large perturbations}b).
This is because forward and vertical dynamics are decoupled on piecewise flat terrains, and hence the open-loop runner does not fall as long as the step height is smaller than the apex height of the aerial phase.
This result suggests a foot placement strategy for running on any rough terrain, namely to aim to land on flat patches of the ground so that stability is maintained with little reliance on feedback control.
However, such a strategy would require visual surveying of the terrain up ahead and planning the location of foot falls.
Increased footfall probability on flat regions of the terrain, along with low probability of footfalls on highly sloped regions of the terrain may be evidence for such a foot placement strategy in experiments.
\subsection{Effect of terrain geometry}
\label{sec:height distribution}
The exact step-to-step variation in the terrain's height and slope depend on the distribution function used to generate the random terrain (Fig.~\ref{fig:Large perturbations}e, inset).
However, we find that the distribution underlying the rough terrain has little effect on the distribution of the steps to failure (Fig.~\ref{fig:Large perturbations}e) when assessed using three different functions to generate rough terrains: von~Mises, uniform and beta.
Runners took $9.6\pm4.1$ (mean $\pm$ std.\ dev.) steps before tumbling on terrains characterized by von~Mises and uniform distributions, and $10.8\pm4.7$ steps on the terrains characterized by the beta distribution.
All steps to failure distributions are unimodal, but skewed.
A Markov model for the step-to-step dynamics (supplement~\ref{ext-sec:levy}) lends insight into the nearly invariant shape of the steps-to-failure distribution.
The insensitivity may arise from the terrain roughness being uncorrelated from step-to-step (terrain's correlation length $\lambda \ll 1$), and thus the net effect of the perturbations resembles a Gaussian noise process that is propagated by the dynamics of running.
\begin{figure}
\centering
\includegraphics{finite_perturbations.pdf}
\caption{The effect of the tangential collision, energy dissipation and terrain geometry on running stability for a human-like runner, found using Monte Carlo~ simulations.
{\bfseries a}, Open-loop runners with $\etr = 0$ (orange circles) lose orientational stability on the rough terrain while anticipatory runners with $\etr = 1$ (blue squares) maintain orientation.
{\bfseries b}, On the step-like terrain, open-loop runners (purple star) maintain forward speed and orientation as the probability of failure, orientational or translational, is zero.
Open-loop and anticipatory runners are identical on step-like terrains
{\bfseries c}, Anticipatory runners slow down on the rough terrain, eventually completely losing forward speed.
Whereas human-like open-loop runners also lose forward speed, they lose orientational stability before completely losing forward momentum.
{\bfseries d}, A contour plot of mean steps taken by open-loop runners as a function of $\etr$ and $\en$ finds that contours are approximately parallel to the $\etr$ axis, while the steps taken increases with decreasing $\en$.
{\bfseries e}, Steps to failure distributions for human-like open-loop runners on rough terrain with height distributions for the grid points drawn from von~Mises (yellow diamond, mean = 0, $\kappa = 6$), Beta (orange square, $\alpha = 1.9,\, \beta = 2.3$) and uniform distributions (blue circle).
The inset shows the probability density functions for the three distributions used to generate the terrain: von~Mises (yellow), Beta (orange) and uniform (blue).
The distributions were scaled and shifted such that mean height $=0$, and range = $0.060r_\ell$ (table 1, section~\ref{sec:simulation}).}
\label{fig:Large perturbations}
\end{figure}
\subsection{Anticipatory runners on rough terrains: tangential collisions}
\label{sec:anticipate}
\begin{figure}
\centering
\includegraphics{anticipatory.pdf}
\caption{Effect of tangential collisions and energy dissipation on running stability for anticipatory runners.
In each panel, the contour plot of mean steps taken over the entire range of independent parameters is shown together with a zoom-in of contours that are bunched together.
{\bfseries a}, The contour plot of mean steps taken by anticipatory runners as a function of $\etr$ and $\en$ shows that contours are bunched close together around $\etr \simeq 1$, with the maximum steps taken at $\en = 0, \etr = 1$ (top plot) and minimum at $\en = 1, \etr = 0$ (bottom plot).
{\bfseries b}, Effect of noise in $\etr$. Contour plot of mean steps taken as a function of $\etr$ and $\Delta\epsilon_{\rm t}$.
The optimal $\etr$ (red circles) is shown for each value of $\Delta\epsilon_{\rm t}$ simulated.
}
\label{fig:anticipatory}
\end{figure}
Runners that use anticipatory strategies to control the tangential passive collision maintain orientational stability if they entirely avoid tangential collisions using $\etr=1$.
But these runners eventually fail by completely losing forward momentum (Fig.~\ref{fig:Large perturbations}c).
Recall that because of the active push-off, the loss of forward momentum is not simply a break of symmetry by the passive tangential collision.
Through a more careful analysis, we find that the mean slope encountered by the runners is positive and not zero, i.e.\ the terrain preferentially impedes the forward momentum (supplement~Fig.~\ref{ext-fig:terrainsmooth}d).
For human-like parameters, anticipatory runners take $75 \pm 40.9$ (mean $\pm$ std.\ dev) steps before completely losing forward momentum and only 1\% of the runners stop moving forward within 15 steps.
In contrast, over 80\% of the open-loop runners fail within 15 steps.
For the anticipatory runner, permitting tangential collisions $\etr < 1$ induces orientational failures and the mean steps to failure decreases.
For example, with $\en=0$ the mean steps to failure when $\etr=1$ is 85 and decreases to 20 when $\etr=0.95$ (Fig.~\ref{fig:anticipatory}a).
A 5\% decrease in $\etr$ caused an over three-fold decrease in the mean steps to failure.
Importantly, the dominant mode of failure switches from translational failures to orientational failures (supplement~Fig.~\ref{ext-fig:ant-runners}a).
At $\etr = 0$ and independent of $\en$, the anticipatory and open-loop strategies are identical.
Thus the anticipatory runner substantially improves stability by avoiding tangential collisions.
Increasing energy dissipation in the normal collision increases the number of steps taken by the anticipatory runner.
For example, at $\etr = 1$, where runners only fail by losing forward speed, increasing energy dissipation by changing from $\epsilon_{\rm n} = 1$ to $\epsilon_{\rm n} = 0$, increases the mean number of steps taken before failure by two-fold, from 40 to 85 (Fig.~\ref{fig:anticipatory}a).
Away from $\etr = 1$, energy dissipation has a smaller effect on stability.
When $\etr \approx 0$, the anticipatory runners resemble the open-loop runners and the mean steps to failure increases by only 2 steps despite $\en$ decreasing from 1 to 0 (Fig.~\ref{fig:anticipatory}a).
Thus, for the anticipatory runner using $\etr \approx 1$, increasing energy dissipation in the direction normal to the terrain is an effective means to improve stability, unlike for the open-loop runner.
\subsection{Noise in anticipatory strategies}
\label{sec:noisy retract}
The sensitivity of the steps to failure with respect to tangential collisions prompts an examination of the effect of stochasticity in how a runner may control the tangential collision.
After all, no runner can exactly control the tangential collision from step-to-step.
For example, errors in sensing the terrain profile as well as motor noise may prevent accurate implementation of a desired $\etr$.
We model such sources of noise in controlling the tangential collision as
\begin{subequations}
\begin{eqnarray}
\label{eqn:etnoise}
\epsilon_{\rm t, noisy} &=& \etr + \Delta\et \eta, \\
\text{where }\eta &\thicksim& \mathcal{U}[-1,1], \, \Delta\et \in \mathbb{R}.
\end{eqnarray}
\end{subequations}
The uniformly distributed zero-mean random variable $\eta$ models random step-to-step noise in $\etr$ and $\Delta\epsilon_{\rm t}$ is the noise intensity.
We find that incurring tangential collisions ($\etr<1$) is optimal when there is non-zero noise ($\Delta\epsilon_{\rm t}>0$).
This is unlike the noiseless anticipatory runner whose optimum is $\etr = 1$.
However, noise in controlling tangential collisions does affect stability and the mean steps to failure are severely reduced (Fig.~\ref{fig:anticipatory}b).
For example, compared to a noiseless human-like runner, the mean steps to failure drops nine-fold for a runner with noise intensity $\Delta\epsilon_{\rm t} = 0.1$, and the optimum $\etr$ decreases by 1\% to $\etr = 0.99$ (Fig.~\ref{fig:anticipatory}b).
Additional noise in the tangential collision of open-loop runners reduces the number of steps taken, but does not alter the dependence of steps taken on $\etr$ (supplement~\ref{ext-sec:open-loop noise}).
Therefore, for anticipatory runners, noise in controlling the tangential collision implies that incurring a slight tangential collision is optimal but at the cost of stability.
\subsection{Predictions for $\epsilon_{\rm t}$ in experiments}
\label{sec:eth}
A main finding of our analyses is the importance of minimizing tangential collisions with the ground when running on rough terrains.
But measuring $\et$ on rough terrains is challenging because it needs a well-defined point of contact under the foot, precise knowledge of the terrain's slope in 3D at that point, and measurement of the reaction force along that tangent.
To facilitate comparisons with experimental data, we consider an easier to measure correlate of $\et$ via the parameter $\hat{\epsilon}_{\rm t}$ that is defined as
\begin{equation}
\hat{\epsilon}_{\rm t} = 1 - \frac{\Delta v_x}{v_x},
\end{equation}
where $\Delta v_x/v_x$ is the fraction of the forward momentum of the runner lost due to the passive collision.
On perfectly flat terrain, $\et=\hat{\epsilon}_{\rm t}$.
In the Monte Carlo~ simulations, $\hat{\epsilon}_{\rm t}$ is characterized by a distribution that evolves with increasing steps (Fig.~\ref{fig:ethat}a, supplement~Fig.~\ref{ext-fig:stepwise-eth}a).
The dependence of $\hat{\epsilon}_{\rm t}$ on steps taken arises because the runner is slowing down, and thus $v_x$ and consequently $\Delta v_x$ change from step-to-step.
But, the mean of $\hat{\epsilon}_{\rm t}$ appears to converge to a constant after just 3 steps for all values of $\etr$ (supplement~Fig.~\ref{ext-fig:stepwise-eth}b).
Importantly, mean $\etht$ increases linearly with $\etr$ (Fig.~\ref{fig:ethat}b) and is this a reliable correlate of the true tangential collision.
However, $\etht$ has a reduced range; mean $\etht = 0.81$ at $\etr = 0$, and mean $\etht = 0.97$ at $\etr = 1$.
The standard deviation of the distributions converges to a value between 0.05 and 0.1 by approximately 10 steps for most values of $\etr$ except when $\etr \to 1$ (Fig.~\ref{fig:ethat}a, supplement~Fig.~\ref{ext-fig:stepwise-eth}b).
For comparison, reported values of $\hat{\epsilon}_{\rm t} $ from experiments with human runners on flat and two rough terrains are $0.94\pm0.01$ (mean $\pm$ standard deviation) identically \citep{Dhawale2018ASB}.
These experimental data are consistent with the prediction that optimal anticipatory runners should maintain $\etr=1$.
\begin{figure}
\centering
\includegraphics{ethat.pdf}
\caption{Estimated tangential coefficient of restitution $\hat{\epsilon}_{\rm t}$ for anticipatory runners using Monte Carlo~ simulations with an ensemble size of $10^6$. {\bfseries a}, Probability density function of $\hat{\epsilon}_{\rm t}$ for human-like anticipatory runners with $\etr = 1$ on rough terrain after 3 steps and after 20 steps. While the standard deviation almost doubles between the two distributions shown here (supplement~Fig.~\ref{ext-fig:stepwise-eth}c), the mean of the distribution converges by 3 steps (supplement~Fig.~\ref{ext-fig:stepwise-eth}b). {\bfseries b}, Mean $\hat{\epsilon}_{\rm t}$, converges by 3 steps for all values of $\etr$ (supplement~Fig.~\ref{ext-fig:stepwise-eth}b), and is always less than 1, ranging from 0.81 at $\etr = 0$ to 0.97 at $\etr = 1$.}
\label{fig:ethat}
\end{figure}
\subsection{Modulation of $\et$}
The tangential collisional impulse depends on the speed of the foot at collision and also on how that collisional impulse is transmitted to the center of mass.
For example, the foot collision may not affect the center of mass very much if the intervening joints between the foot and the body are compliant.
The transmission of collisions is treated in terms of sprung and unsprung masses in models of automobiles and in running biomechanics \citep{McGeer1990R}.
If collisional impulses at the foot are faithfully transmitted to the center of mass, the retraction rate $\omega_{\rm ret}$ is related to $\et$ as,
\begin{subequations}\label{eqn:et-omega_ret-reln}
\begin{eqnarray}
\et &=& \frac{\omega_{\rm ret}}{v_{\rm p,t}^-}, \\
\text{where}\ \omega_{\rm ret} &=& \begin{cases} \etr v_{\rm p,t}^-\ \text{: anticipatory},\\
\etr v_{\rm x0} \ \text{: open-loop} \end{cases}.
\end{eqnarray}
\end{subequations}
Using leg retraction to control $\et$ implies that the optimal retraction rate zeros the tangential foot speed at landing.
Equivalently, the foot may be allowed to collide with the ground and yet achieve $\et\approx 1$ by maintaining low stiffness in the leg's joints.
\section{Linear stability analysis}
\label{sec:stability}
For periodic dynamic systems linear stability is defined as the response to small perturbations in the neighborhood of a periodic orbit \citep{Full2002,Holmes2006,Bruijn2013} and analyzed using Floquet theory \citep{Guckenheimer1983,Holmes2006}.
Floquet analysis for the stability of a periodic orbit defines a transverse cross-section to the orbit and a discrete return map from initial conditions on the cross-section back to the same cross-section after a complete period.
The eigenvalues of the return map, called Floquet multipliers, are independent of the chosen cross-section and govern the stability of the periodic solution to small perturbations \citep{Guckenheimer1983}.
Here we consider the anticipatory runner and discuss the open-loop runner in supplement~\ref{ext-sec:jordan} because the unstable modes of both variants are the same.
The mechanical state of the runner is represented by $\vec{\zeta} = (x, y, \phi, v_x, v_y, \omega)\tr$, where $(x, y)$ and $\phi$ denote the center of mass position and orientation, and $(v_x, v_y)$ and $\omega$ are the respective velocities, all measured in a Newtonian reference frame that translates forward at a constant speed $v_{\rm x0}$.
A steady runner is periodic in this translating Newtonian frame of reference.
We define a transverse cross-section (Poincar\'{e} section) at the apex of the aerial phase ($v_y = 0$) following the approach of \citet{Full2002} and \citet{Seyfarth2003}.
The equations~\eqref{eqn:dynamics of running} yield the step-to-step return map $\vec{f}_{\rm an}$ and its linearization $\vec{\rm T}_{\rm an} $ in terms of a the mechanical state $\vec{\psi}$ in a translating frame according to
\begin{subequations} \label{eqn:floquet analysis}
\begin{eqnarray}
\vec{\psi} &=& (x, y, \phi, v_x, \omega)\tr, \label{eqn:poincare state}\\
\vec{\psi}_{n+1} &=& \vec{f}_{\rm an}\left(\vec{\psi}_n\right), \label{eqn:return map}\\
\Delta\vec{\psi}_{n+1} &=& \vec{\rm T}_{\rm an}\Delta\vec{\psi}_n, \label{eqn:linearized return map}\\
\text{where}\ \Delta\vec{\psi} &=& \vec{\psi}-\vec{\psi}^\ast,\ \vec{\rm T}_{\rm an} = \frac{\partial \vec{f}_{\rm an}}{\partial \vec{\psi}}\bigg|_{\vec{\psi}^*}.\label{eqn:linearized return map definition}
\end{eqnarray}
\end{subequations}
The Poincar\'{e} map given by equation~\eqref{eqn:return map} has a fixed point at $\vec{\psi}^*=\vec{0}$ when the terrain is flat and corresponds to an exactly periodic runner on flat ground.
\begin{figure}
\centering
\includegraphics{returnmap.pdf}
\caption{Illustration of the trajectory of the runner in state space in a reference frame that is translating along with the runner with velocity $v_{\rm x0}$.
The runner appears periodic in this reference frame and the runner's mechanical state follows a periodic orbit.
The return map $f_\bullet$ ($\bullet$ is `ol' or `an' for open-loop or anticipatory, respectively) is defined from the apex of the aerial phase ($v_y = 0$) to apex of the following aerial phase.
$\vec{\psi^*}$ is the fixed point of the return map and $\vec{\psi}_n$ is a small perturbation away from the fixed point $\vec{\psi}^*$ at step $n$.
In the next step, $\vec{\psi}_n$ maps to $\vec{\psi}_{n+1}$ at the apex of the following aerial phase under action of the return map $\vec{f}_\bullet$.}
\label{fig:returnmap}
\end{figure}
The linearized return map $\vec{\rm T}_{\rm an}$ has three eigenvalues equal to one and the others are all less than one.
The eigenvalues with magnitude less than one correspond to stable modes so that perturbations along their respective eigenvectors will always decay.
The remaining three eigenvalue are all $\lambda = 1$ with algebraic multiplicity equal to 3 and geometric multiplicity equal to 2.
This implies that there are only two independent eigenvectors corresponding to the three unity eigenvalues and the matrix $\vec{T}_{\rm an}$ is therefore non-diagonalizable.
For non-diagonalizable systems, the Jordan decomposition is used to analyze stability in terms of generalized eigenvectors (supplement~\ref{ext-sec:jordan}), and implies that the modes (eigenvectors) associated with these eigenvalues cannot be decoupled and analyzed independently.
The two eigenvectors $\vec{\nu}_1$, $\vec{\nu}_2$ and the third generalized eigenvector $\vec{\nu}_3$ corresponding to the repeat eigenvalue $\lambda = 1$ span a subspace in which the dynamics of the return map don't simply decay back to the origin.
For a diagonalizable system, any perturbation within this subspace would neither decay nor grow.
However, the non-diagonalizable nature of $\vec{T}_{\rm an}$ leads to the outcome that a perturbation $\Delta\vec{\psi}_0$ within this subspace grows with increasing steps.
The eigenvectors $\vec{\nu}_1$, $\vec{\nu}_2$, $\vec{\nu}_3$, the initial perturbation $\Delta\vec{\psi}_0 $, and its growth after $n$ steps to $\Delta\vec{\psi}_n$ are given by,
\begin{subequations}\label{eqn:generalized instability}
\begin{eqnarray}
\vec{\nu}_1 &=& \begin{pmatrix} 0 & 0 & 1 & 0 & 0 \end{pmatrix}\tr,\
\vec{\nu}_2 = \begin{pmatrix} 1 & 0 & 0 & 0 & 0 \end{pmatrix}\tr,\
\vec{\nu}_3 = \begin{pmatrix} 0 & 0 & 0 & \frac{-1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \end{pmatrix}\tr,\label{eqn:eigenvectors}\\
\Delta\vec{\psi}_0 &=& \sum_{k=1}^3 \alpha_k \nu_k,\ \text{and}\\
\Delta\vec{\psi}_n &=& n \alpha_3 \sqrt{2} \en v_{\rm y0} (\vec{\nu}_1 - \vec{\nu}_2) + \Delta\vec{\psi}_0,\ \text{respectively}\label{eqn:subspace perturbation}.
\end{eqnarray}
\end{subequations}
As $n$ grows larger, the asymptotic approximation (denoted by $\approx$) is given by
\begin{equation}
\Delta\vec{\psi}_n \approx n\, \alpha_3 \sqrt{2}\en v_{\rm y0} \begin{pmatrix} -1 \\ 0 \\ 1 \\ 0 \\ 0 \end{pmatrix}\ \text{where}\ n\gg 1. \label{eqn:lingrowth}
\end{equation}
Only a perturbation of magnitude $\alpha_3$ along $\vec{\nu}_3$ affects stability and leads to a nearly linear growth within the subspace spanned by the eigenvectors $\vec{\nu}_1, \vec{\nu}_2$.
Perturbations along $\vec{\nu}_1$ or $\vec{\nu}_2$ neither grow nor decay because these represent invariance with respect to rotations and translations of the reference frame, respectively.
A perturbation along the generalized eigenvector $\vec{\nu}_{3}$ may be geometrically viewed as one that conserves the velocity of the contact point on flat terrain but changes the angular momentum of the runner about its center of mass.
Therefore, any perturbation to the angular momentum will affect both orientation and forward speed.
For the special case of the anticipatory runner that completely avoids tangential collisions, the linearized return map $\vec{T}_{\rm an}$ with $\etr = 1$ has eigenvalue $\lambda = 1$ of algebraic multiplicity 4 and geometric multiplicity 2, and one eigenvalue with $|\lambda| < 1$.
The eigenvectors $\vec{\nu}_1, \vec{\nu}_2$ and the generalized eigenvectors $\vec{\nu}_3,\vec{\nu}_4$ associated with $\lambda = 1$ form a basis for a subspace within which an initial perturbation $\vec{\psi}_0$ grows linearly with the number of steps $n$ in a subspace spanned by eigenvectors $\vec{\nu}_1,\vec{\nu}_2$, i.e.\
\begin{subequations}
\begin{eqnarray}
\vec{\nu}_1 &=& \begin{pmatrix} 0 \\ 0 \\ 1 \\ 0 \\ 0 \end{pmatrix},\
\vec{\nu}_2 = \begin{pmatrix} 1 \\ 0 \\ 0 \\ 0 \\ 0 \end{pmatrix},\
\vec{\nu}_3 = \begin{pmatrix} 0 \\ 0 \\ 0 \\ 0 \\ 1 \end{pmatrix},
\vec{\nu}_4 = \begin{pmatrix} 0 \\ 0 \\ 0 \\ 1 \\ 0 \end{pmatrix}, \label{eqn:eigenvectors etc=1}\\
\Delta\vec{\psi}_0 &=& \sum_{k=1}^4 \alpha_k \nu_k,\
\Delta\vec{\psi}_n = n\, (\alpha_3 a_1 \vec{\nu_1} + \alpha_4 a_2 \vec{\nu_2}) + \Delta\vec{\psi}_0,\\
\Delta\vec{\psi}_n &\approx& n\ 2\en v_{\rm y0} \begin{pmatrix} \alpha_4 \\ 0 \\ \alpha_3 \\ 0 \\ 0 \end{pmatrix} \text{for}\ n \gg 1.
\label{eqn:etc=1 lingrowth}
\end{eqnarray}
\end{subequations}
A perturbation to angular velocity $\omega$ causes a linear growth in orientation $\phi$, and a perturbation to the linear velocity $v_x$ causes a linear growth in position $x$.
However, an anticipatory runner with $\etr = 1$ avoids angular velocity perturbations due to the terrain altogether, i.e.\ $\alpha_3 = 0$.
Therefore, only forward speed is affected due to the remaining unstable mode $\vec{\nu}_4$.
Although there are no unstable eigenvalues with magnitude greater than one, we find that the dynamics of running lead to an unstable growth with increasing steps.
The growth due to non-diagonalizability of the return map is linearly proportional to the number of steps, and not geometric as is the case for simple unstable eigenvalues.
Importantly, the primary effect of the instability is to affect the forward speed and orientation, consistent with the numerical simulations that use finite perturbations and nonlinear dynamics.
Also in agreement with the simulations, the only instability is translational when $\etr=1$.
\section{Scaling analysis of the orientational failure mode}
\label{sec:paramdepend}
The mean steps to failure depends on many parameters, but none of the parameters separately predict the failure statistics (supplement~Fig.~\ref{ext-fig:general}).
As most runners undergo orientational failures, we investigated whether the amount of body rotation accumulated over a single step due to a terrain slope perturbation would predict failure statistics.
If a runner with a periodic trajectory on flat ground encounters a sloped terrain of angle $\theta$, the orientation $\phi_\bullet$ at the next landing will no longer be vertical.
This orientation $\phi_\bullet$ accumulated over one step depends on the take-off vertical velocity $v_{y,\bullet}^+$ via the aerial phase time $2v_{y,\bullet}^+$, and take-off angular velocity $\omega_\bullet^+$, as $\phi_\bullet = 2v_{y,\bullet}^+\omega_\bullet^+$.
The subscript `$\bullet$' is a placeholder for `ol' or `an' as the orientation change depends on whether the runner is purely open-loop (ol) or employs anticipatory (an) control.
We hypothesize that the mean steps to failure $N_\bullet$ is a function of the orientational threshold $\phi_{\rm tol}$ and the orientation change over a single step $\phi_\bullet$ alone, i.e.\ $N_\bullet = s_\bullet(\phi_{\rm tol},\phi_\bullet)$.
Substituting the form of $s_\bullet(\phi_{\rm tol},\phi_\bullet)$ derived in supplement~\ref{ext-sec:levy}, we show that the mean steps to failure $N_\bullet$ is predicted to scale according to,
\begin{equation}
\label{eqn:paramdep}
N_\bullet \sim \frac{\phi_{\rm tol}}{\phi_\bullet},
\end{equation}
where the expression for $\phi_\bullet$ is given in supplement~equation~\eqref{ext-eqn:phi-bullet-compl}
The mean steps to failure in simulations performed with many different parameter values (supplement~\ref{ext-sec:bodyrot}) are well-approximated by a single function of a dimensionless parameter $\phi_{\rm tol}/\phi_\bullet$ (Fig.~\ref{fig:Smallperturbation}).
The collapse of the simulation data highlights that the spin accumulated in one step due to a single perturbation (equation~\eqref{eqn:paramdep}) captures the fundamental principle underlying orientational failures.
Importantly, this dimensionless parameter collapses the simulation data better than any individual parameter (supplement~Fig.~\ref{ext-fig:general}).
Thus, the single parameter $\phi_{\rm tol}/\phi_\bullet$ quantifies stability of runners of different sizes and mass distributions.
\begin{figure}[!thb]
\centering
\includegraphics{generalization.pdf}
\caption{Generalizing results from section~\ref{sec:openloop} and section~\ref{sec:anticipate} to a wider range of physical and terrain parameters.
Mean steps to failure from the Monte Carlo simulations is plotted against {\bfseries a}, $\phi_{\rm tol}/\phi_{\rm an}$ and {\bfseries b}, $\phi_{\rm tol}/\phi_{\rm ol}$ for different values of $\phi_{\rm tol}$.
The mean steps to failure depend mostly on a single dimensionless parameter $\phi_{\rm tol}/\phi_{\bullet}$.
All simulation parameters were varied independently in these simulations.
But, for clarity, only variations in $\phi_{\rm tol}$ are identified with different marker types.}
\label{fig:Smallperturbation}
\end{figure}
The dimensionless parameter $\phi_{\rm tol}/\phi_\bullet$ also captures the parametric dependence of mean steps to failure on $\etr$ and $\en$ as seen from comparing the contour plots of $\phi_{\rm tol}/\phi_\bullet$ shown in Fig.~\ref{fig:asymptotic analysis contours} against that of the direct simulations in Fig.~\ref{fig:Large perturbations}d and Fig.~\ref{fig:anticipatory}a.
The dependence of mean steps to failure on $\etr$ and $\en$ for small slopes of the terrain is understood using a series expansion of $\phi_\bullet$ in terms of $\theta$ as given by,
\begin{subequations}\label{eqn:series exp}
\begin{eqnarray}
\label{eqn:phi-ol}
\phi_{\rm ol} &=& \bigg(\frac{2v_{\rm y0}^2}{1+\I}\bigg)\theta + \bigg(\frac{3-\I}{(1+\I)^2}+\frac{4\I}{(1+\I)^2}\etr + \frac{2}{1+\I}\en \bigg) v_{\rm x0}v_{\rm y0}\, \theta^2 + O(\theta^3),\\
\label{eqn:phi-an}
\nonumber \phi_{\rm an} &=& \bigg(\frac{2v_{\rm y0}^2}{1+\I}(1-\etr)\bigg)\theta + \bigg(\frac{3-\I}{(1+\I)^2} + \frac{5\I-3}{(1+\I)^2}\etr -\frac{4 \I}{(1+\I)^2}\etr^2 + \\
& & \frac{2 (1-\etr)}{1+\I}\en \bigg)v_{\rm x0}v_{\rm y0}\,\theta^2 + O(\theta^3).
\end{eqnarray}
\end{subequations}
For the open-loop strategy, neither of the collision parameters, $\en$ or $\etr$, appear in the linear (leading order) term.
When using an anticipatory strategy, the tangential collision parameter $\etr$ appears to leading order.
The normal collision parameter $\en$ affects the second order dependence on $\theta$ for both strategies.
These show why it is impossible to avoid orientational failures for the open-loop strategy, but may be avoided when using the anticipatory strategy by choosing $\etr=1$ and $\en=0$.
\begin{figure}[!thb]
\centering
\includegraphics{one-step-contours.pdf}
\caption{Contour plots of {\bfseries a}, $\phi_{\rm tol}/\phi_{\rm ol}$ and {\bfseries b}, $\phi_{\rm tol}/\phi_{\rm an}$ as a function of $\en$ and $\etr$ reveal that a single parameter captures the dependence of mean steps to failure of both open-loop runners (Fig.~\ref{fig:Large perturbations}d) and anticipatory runners (Fig.~\ref{fig:anticipatory}a) on the collision parameters $\en$ and $\etr$.
Recall that the only controllable parameters for the runners in these simulations are $\en$ and $\etr$.
The complete expression for $\phi_\bullet$, shown in supplement~equation~\eqref{ext-eqn:phi-bullet-compl}, was used to generate these plots with parameter values drawn from table 1, and $\phi_{\rm tol} = 1$.
For the anticipatory runner, we restricted the maximum value of $\etr$ to 0.99. For higher values of $\etr$ orientational failures are rare and thus not accounted for by $\phi_{\rm tol}/\phi_\bullet$.
}
\label{fig:asymptotic analysis contours}
\end{figure}
For the open-loop runner, $\phi_{\rm tol}/\phi_{\rm ol}$ is independent of $\en$ and $\etr$ to first order in $\theta$ (equation~\eqref{eqn:phi-ol}).
Hence the contours in Fig.~\ref{fig:asymptotic analysis contours}a (which resemble the contours in Fig.~\ref{fig:Large perturbations}d from the Monte Carlo~ simulations) show a weak dependence on $\en$ and $\etr$ that arises from the $\theta^2$ term in equation~\eqref{eqn:phi-ol}.
The parameter $\phi_{\rm ol}$ is smallest when $\en = \etr = 0$, and largest when $\en = \etr = 1$.
For a human-like runner, $\I\ll 1$ (table 1), and thus the $\theta^2$ term in equation~\eqref{eqn:phi-ol} can be reduced to $(3+2\en)v_{\rm x0}v_{\rm y0}\theta^2$, with no dependence on $\etr$ at the asymptotic limit of $\I\ll 1$.
The asymptotic analysis of $\phi_{\rm ol}$ therefore explains why the contours of mean steps to failure in the Monte Carlo~ simulations are nearly parallel to the $\etr$ axis and increase only slightly when $\en$ is decreased (Fig.~\ref{fig:Large perturbations}d).
For the anticipatory runner, the first order term in the expansion depends on $\etr$ (equation~\ref{eqn:phi-an}), unlike the case for the open-loop runner (equation~\eqref{eqn:phi-ol}).
Nearly perfect anticipation corresponds to $\etr \to 1$.
At this limit $\phi_{\rm an} \to 0$ and thus $N=\phi_{\rm tol}/\phi_{\rm an} \to \infty$, explaining why the contours of mean steps to failure in the Monte Carlo~ simulations are tightly bunched together in the neighborhood of $\etr = 1$ (Fig.~\ref{fig:anticipatory}a) and nearly parallel to the $\en$ axis.
Like for the open-loop runner, $\phi_{\rm an}$ also shows a dependence on $\en$ only in the $\theta^2$ term of the power series expansion in equation~\ref{eqn:phi-an}.
As $\en$ decreases so does $\phi_{\rm an}$, and thus $\phi_{\rm tol}/\phi_{\rm an}$ increases, capturing the trend observed in the Monte Carlo~ simulations where decreasing $\en$ increases steps taken for the anticipatory runner (Fig.~\ref{fig:anticipatory}a).
For the anticipatory runner, unlike the open-loop runner, the $\en$ dependence is coupled to $\etr$, and thus the sensitivity of the parameter $\phi_{\rm tol}/\phi_\bullet$ to changes in $\en$ depend on the value of $\etr$.
The limit of $\etr = 0$, where $\phi_{\rm an} = \phi_{\rm ol}$ has already been discussed above, for the open-loop runner.
To analyze the case where $\etr \to 1$, we approximate $\phi_{\rm an}$ in the limit where $\I << 1$ (e.g.\ human-like runners) as
\begin{equation}
\phi_{\rm an} \approx 2v_{\rm y0}^2(1-\etr)\theta + (1-\etr)(2\en+3)v_{\rm x0}v_{\rm y0}\theta^2.
\label{eqn:phi-an-approx}
\end{equation}
To understand the dependence of the mean steps to failure $N$ on $\en$ and $\etr$, we consider the limit of small angles of the terrain slope $\theta \ll 1$.
Using equation~\eqref{eqn:phi-an-approx}, and for small $\theta$ we find that mean steps to failure $N = \phi_{\rm tol}/\phi_{\rm an}$ and its sensitivity to changes in $\en$ are given by
\begin{subequations}
\begin{eqnarray}
\label{eqn:N approx}
N &=& \left(\frac{1}{1-\etr}\right)\frac{\phi_{\rm tol}}{v_{\rm y0}\theta}(2v_{\rm y0} - 3v_{\rm x0}\theta - 2v_{\rm x0}\theta\en),\\
\label{eqn:N approx sensitivity en}
\frac{\partial N}{\partial \en} &=& -\left(\frac{1}{1-\etr}\right) \frac{2 \phi_{\rm tol} v_{\rm x0}}{v_{\rm y0}}.
\end{eqnarray}
\end{subequations}
Therefore, $N$ is more sensitive to changes in $\en$ when $\etr \to 1$.
This resembles Fig.~\ref{fig:anticipatory}a where the mean steps to failure from the Monte Carlo~ simulations increases significantly as $\en$ is reduced when $\etr \to 1$, as opposed to when $\etr \to 0$ where there is much lesser sensitivity of the mean steps to failure with respect to changes in $\en$.
Improving running stability by increasing mean steps to failure helps provide more time for feedback driven corrections in real-world runners.
The analysis of mean-steps to failure in the simplified runners without any feedback ability suggests that increasing $\phi_{\rm tol}$ and decreasing $\phi_\bullet$ are both effective strategies to negotiate rough terrains.
Therefore, besides altering $\etr$ and $\en$ in order to increase $\phi_{\rm tol}/\phi_\bullet$ as already discussed, increasing $\I$ and reducing $v_{y0}$ also improves stability.
\section{Discussion}
\label{sec:discussion}
We show that purely open-loop strategies with no feedback control cannot stabilize sagittal-plane dynamics during running.
Such open-loop runners fail primarily by losing orientational stability and tumbling.
Using an anticipatory strategy to eliminate tangential collisions with the ground eliminates orientational instabilities but leads to a steady slowing down of the runner.
However, on a step-like piecewise flat terrain the open strategy is sufficient to stabilize the runner without losing forward speed, and so is the anticipatory strategy.
If an anticipatory strategy is implemented noisily, i.e.\ the tangential collisions are low but not entirely eliminated, the runner suffers orientational instabilities.
However, both the orientational and the translational instabilities are weak when using an anticipatory strategy and the growth of the instability is only linearly proportional to the number of steps taken and not a higher power.
The exact number of steps to failure depend on many parameters including the inertial, geometry, collision parameters and the thresholds in orientation and speed for failure.
These large set of parameters may be combined into a single dimensionless parameter that captures the failure statistics, when can also guide the morphological design of stable runners.
\subsection*{Impulsive stance assumption}
An impulsive stance phase implies that the stance impulse is defined, but not the detailed time history of forces.
Thus the model may be used study the dynamics and stability over multiple steps, but it cannot be used to find the actuation patterns that would achieve the desired impulse.
Such simplified models may used to specify the desired collisional and push-off impulses as constraints that should be be met.
More detailed models could then be used to calculate the stance force profiles as a constrained search or optimization problem.
The model also ignores the impulse due to the finite forces of gravity, because stance is treated as instantaneous.
Relaxing the assumption of instantaneous stance implies that body-weight affects the body's angular momentum about the contact point according to,
\begin{equation}
\label{eqn:ang mom finite stance}
{H}_{\rm /P}^+-{H}_{\rm /P}^- = \int\limits_{0}^{T_{\rm stance}} {M}_{\rm /P}(t)\, dt,
\end{equation}
where $T_{\rm stance}$ is the stance duration and ${M}_{\rm /P}(t)$ is the time-varying moment of the body weight about the contact point P.
The torque due to gravitational forces about the contact point is proportional to body weight and the time-varying horizontal distance from the center of mass to the contact point.
The gravitational contribution is zero for a symmetric stance, and highest for the most asymmetric stance.
Assuming a constant forward speed during stance and a 20$^\circ$ touchdown angle, the maximum change in angular momentum about the contact point, i.e.\ the integral in equation~\eqref{eqn:ang mom finite stance}, is $\lvert\Delta{H}_{\rm /P}\rvert \lessapprox 0.15$ in the same dimensionless units as before.
For a typical human runner, the resultant orientation change in a single step is $\Delta\phi \lessapprox 0.01$, negligibly small compared to the influence of the terrain.
Thus, ignoring torques induced by gravity has minimal impact on our conclusions.
\begin{figure}[!thb]
\centering
\includegraphics{finite-stance-cartoon.pdf}
\caption{Equivalence between a runner using a freedom finite stance with internal degrees of freedom versus an infinitesimal stance with an applied torque impulse.
The contact of the general runner is represented as an impulse due to the collision at touchdown (left, pink vector) plus a net impulse due to the push-off (right, pink vector).
In this example, by varying duration of stance, the runner can selectively vary the lever arm of the push-off impulse and thus the angular impulse about its center of mass without altering the linear impulse.
The net effect of a finite stance and change in configuration is therefore captured by a linear impulse at the point of contact (pink vector at ground) and an angular impulse at the center of mass ($J_{\phi}$).
}
\label{fig:finite infinitesimal equivalence}
\end{figure}
A finite stance duration and the associated change in configuration allows a runner to control the body's sagittal plane angular momentum, independently from the forward and upward linear momenta.
This may be understood in terms of breaking the symmetry of the stance phase or applying a large forward impulse and yet having the net ground reaction impulse pass through the center of mass (no contribution to angular momentum).
To not lose such control when considering an impulse stance, we permit the application of an arbitrary angular impulse $J_{\phi}$ at the center of mass during push-off (equation~\eqref{eqn:pushoff}).
Thus having a finite stance duration and change of body configuration over stance is equivalent to $J_{\rm phi}$ (Fig.~\ref{fig:finite infinitesimal equivalence}).
The angular impulse also provides a means to accommodate torques due to a finite base of support and a moving center of pressure during stance.
However, recall that the constraint that the applied active push-off impulses should lead to perfectly periodic gait on flat terrain implies that $J_{\phi}=0$.
In our model, both open-loop and anticipatory runners slow down on rough terrain.
Regaining forward speed needs a feedback controller, and then the additional control authority offered by $J_{\phi}$ would be necessary to vary forward speed without affecting the body's angular momentum or vertical momentum.
\subsection*{Point contact assumption}
Another limitation arises from considering a point-like contact that cannot capture effects associated with the spatial extent of the foot.
These effects include the spatial filtering of terrain roughness and the application of a net torque about the initial contact point.
The inclusion of $J_{\phi}$ in the model captures the application of torques, but there is no explicit means of incorporating the ability of the foot to act as a spatial filter \citep{Sharbafi2017}.
Therefore, careful consideration should be given to the spatial frequency (wave number) of the roughness of the terrain when using a model with a point contact.
\subsection*{Timescale for feedback corrections}
Open-loop runners with human-like parameters have a 99\% chance of taking at least 3 steps without failing by exceeding the orientation threshold, while anticipatory runners ($\etr = 1$) can take upto 15 steps with the same probability of completely losing forward momentum.
This implies that the open-loop runner employing the slowest sensory modality (visual feedback delay $\approx200$~ms \citep{vanBeers2002}) has 7 feedback cycles to correct for instabilities at endurance running speeds of 3m/s (step period $\approx500$~ms \citep{Cavagna1964}), with only an approximately 1\% chance of an orientational failure.
Thus, while sensory feedback is required to run on rough terrains, timescales associated with sensory feedback delays do not limit the runner's ability to maintain stability because of the nature of the instability.
Furthermore, employing an appropriate anticipatory strategy ($\etr = 1$) eliminates the orientational instability entirely, thereby further extending the timescale over which feedback is necessary.
\subsection*{Leg retraction}
Analyses of running models with leg mass suggest that optimal retraction rate is defined by stability demands, although these studies were limited to step-like terrains \citep{Karssen2015}.
Experiments with runners on flat ground which measure the angle of the foot's velocity vector with respect to the ground suggest that foot velocity is perhaps not modulated in the manner we hypothesize \citep{Blum2010}.
In the study by \citet{Blum2010}, the mean angle made by the subjects' foot velocity vector with the ground was 165$^\circ$, whereas our prediction based on zero tangential speed of the foot at touchdown would imply that the angle should be 90$^\circ$.
Given these differences, we propose that the low values of $\hat{\epsilon}_{\rm t}$ for human runners observed by \citet{Dhawale2018ASB} may result from joint stiffness modulation in the leg rather than precise control of the foot speed through leg retraction.
Modulating foot and leg stiffness allows the runner to minimize the tangential collision and yet employ leg retraction strategies that accomplish other goals such as hypothesized by \citet{Seyfarth2003} and \citet{Birn2014}.
\subsection*{Energy dissipation}
Besides leg retraction, energy dissipation may also aid in stability based on studies of walking \citep{Kuo1999,Donelan2001} and running \citep{Daley2006,Arellano2011,Arellano2012}.
Our model shows that while increasing energy dissipation in the direction normal to the terrain does increase the number of steps taken for open-loop and anticipatory runners, dissipating energy in the tangential collision is detrimental to stability.
However, whether energy dissipation helps or hinders depends on the details of what is meant by ``open-loop''.
For example if the runner uses a \emph{lab-fixed} push-off policy instead of the \emph{terrain-fixed} push-off described in the main text, dissipating energy in the normal direction is also detrimental to orientational stability (supplement~\ref{ext-sec:labpushoff}).
Thus the hypothesized trade-off between energy consumption and stability is not universally true in our models.
Our results are consistent with experiments on running birds encountering sudden terrain drops, as the birds do not always dissipate energy on the perturbation step \citep{Daley2006}.
Our results might provide a means to understand why the increase in energy consumption for humans running on step-like terrains is only 5\% \citep{Voloshina2015}.
We find that open-loop strategies are sufficient to maintain stability on step-like terrains and additional energy expenditure provides little added benefit.
\subsection*{Implications of scaling analysis to body plan of animals}
The single parameter $\phi_{\rm tol}/\phi_\bullet$ that predicts mean steps to failure (Fig.~\ref{fig:Smallperturbation}) generalizes our results beyond runners with human-like parameters and can be used as a criteria to assess a runner's stability.
This is because runners with a larger $\phi_{\rm tol}/\phi_\bullet$ should be able to maintain orientation for a greater number of steps in the absence of sensory feedback control.
We have discussed how energy storage in the direction normal to the terrain ($\en$) and tangential collision modulation ($\etr$) affects $\phi_{\rm tol}/\phi_\bullet$ in section~\ref{sec:paramdepend}, and now turn to the implications the parameter has on how body morphology and mass distribution affect stability.
The parameter $\phi_{\rm tol}$ is the maximum angle of tilt the runner can accumulate before it falls.
By employing larger (base/height) ratios, i.e.\ adopting a landscape rather than a portrait orientation when viewed in the sagittal plane, animals can increase $\phi_{\rm tol}$ and thereby increase $\phi_{\rm tol}/\phi_{\rm \bullet}$.
Quadrupeds such as cats and dogs, and other adept runners such as cockroaches possess such an aspect ratio.
Another way to increase $\phi_{\rm tol}$ is by increasing the range of motion of the leg with respect to the body.
Even if the body begins to tilt, the ability to place the foot in front of the runner initiates stance and hence allows the runner to correct for body orientation.
In our simulations, the choice of $\phi_{\rm tol}$ value is based on this consideration of the leg angle for humans.
Ostriches are another example of an animal with a portrait orientation but who are adept runners, perhaps in part due to the large of range of motion of their legs.
Penguins, who are not known to be adept runners, occupy the opposite end of $\phi_{\rm tol}$ scale due to possessing a portrait orientation and low range of motion of their legs compared to other bipeds such as humans, turkeys, and ostriches.
Because $\phi_{\rm tol}/\phi_\bullet \propto \I/v_{\rm y0}^2$ (equations~\eqref{eqn:series exp}), lowering take-off angles for a given forward speed would be beneficial to stability.
However, very low take-off angles increase the risk of tripping on rough terrains.
Altering body mass distribution to increase the radius of gyration $r_g$ relative to leg length $r_\ell$ also reduces $\phi_\bullet$ and thereby increases stability.
This can be achieved by increasing distal masses in appendages like arms and legs.
However, increasing distal masses in the leg increases the metabolic cost of running \citep{Myers1985} via increased energetic cost associated with swinging the leg \citep{Marsh2004,Doke2005} and may also lead to potentially injurious collisions.
Alternatively, light legs with a bulky, extended torso or large head, like in horses and bison, also increases $r_g$.
Lastly, tails in animals like cats and lizards, while used for active stabilization and in complex maneuvers like righting reflexes \citep{Jusufi2011,Libby2012}, are yet another means to increase $r_g$, thereby reducing $\phi_\bullet$.
Anticipatory runners further benefit from setting $\et \approx 1$ like observed in humans \citep{Dhawale2018ASB}, thereby drastically reducing $\phi_\bullet$ as discussed in section~\ref{sec:paramdepend}.
Thus, the dimensionless parameter $\phi_{\rm tol}/\phi_\bullet$ is qualitatively consistent with the body morphology of adept and poor animal runners and we propose that it can be used as a design criteria for running robots.
\section{Methods}
\label{sec:methods}
\subsection{Simulation methods}
\label{sec:simulation}
All simulations were performed using custom-written C programs. Parameter values given in table 1 are used for simulations of the human-like runner.
These values are chosen for the purpose of illustration, however our qualitative results are not sensitive to these values, and the scaling analysis in section~\ref{sec:paramdepend} addresses the generalization of these numerical results to runners and terrains with varying parameter values.
\subsection{Parameter values for a human-like runner}
The rationale for chosing the human-like parameter values is as follows.
The moment of inertia value we use is derived from estimates made by \citep{Erdmann1999}, who find that moment of inertia about the center of mass in the sagittal plane is $\approx 13$~kg.m$^2$ for a 75~kg human.
The value for $\epsilon_{\rm n} = 0.63$ corresponds to $\approx40$\% elastic energy stored over one gait cycle, similar to estimates by \citep{Cavagna1964,Cavagna1977b,Alexander1987}.
The orientation bound $\phi_{\rm tol}$ is equal to $\pi/6$ as it approximately half the angle between the legs during double stance in walking.
Forward speed at take-off $v_{\rm x0} = 0.96$ corresponds to 3~m/s for a leg length of 1~m and vertical speed at take-off $v_{\rm y0} = 0.26$ corresponds to $\approx 0.8$~m/s \citep{Dhawale2018ASB}.
Distributions had nearly converged by an ensemble size of $10^4$, hence we simulate for $10^5$ instances (supplement~\ref{ext-sec:convergence}).
\subsection{Terrain model}
\label{sec:terrain model}
The terrain is modelled as piecewse linear.
This is achieved by first defining a one-dimensional grid with fixed grid spacing $\lambda$.
Interpolating heights between the grid points $k$, located at $x_k$ to intermediate points $(x_t,y_t)$ in the $k$th terrain patch, yields a piecewise linear, continuous terrain profile, where terrain slope $m_k$ is discontinous at the grid points,
\begin{subequations}\label{eqn:terrain-model}
\begin{eqnarray}
\text{patch}\ k:\, y_t &=& m_k x_t + c_k,\\
\text{where}\ x_t &\in& [x_k, x_{k+1}], \\
\text{continuity condition:}\ m_k x_{k+1} + c_k &=& m_{k+1} x_{k+1} + c_{k+1},
\end{eqnarray}
\end{subequations}
where $m_k$ and $c_k$ are constants within a patch.
Terrain heights at all grid points are distributed according to $\thicksim\mathcal{U}(-0.03, 0.03)$ (table 1).
Our choice of the uniform distribution $\mathcal{U}$ is to improve simulation speed, even though beta distributions described in Fig.~\ref{fig:Large perturbations}e most closely matched artificial terrain used in experiments \citep{Dhawale2015DW,Dhawale2018ASB}.
The range of heights $h \in [-0.03, 0.03]$ and grid spacing $\lambda = 0.1$ was chosen to match the artificially constructed rough terrains \citep{Dhawale2015DW,Dhawale2018ASB}.
Step-like terrains with no slope distributions were simulated by picking a height from the probability distribution prior to landing.
If the chosen landing height was above the apex height of any portion of the runner, we chose another landing height from the distribution.
The probability of this resampling occuring is $\sim 10^{-4}$.
\subsection{Calculating ground contact point}
\label{sec:landing}
The aerial phase ends when the runner collides with the ground.
The landing position is determined by solving for the unknown intersection point $x_t$ of the runner's aerial phase trajectory $(x_G,y_G)$ with the condition for tangential contact between runner and ground,
\begin{subequations}\label{eqn:terrain contact}
\begin{eqnarray}
\label{eqn:ballflight}
\text{parabolic flight:}\ y_{\rm G} &=& b_0 + b_1 x_{\rm G} + b_2 x_{\rm G}^2, \\
\text{touchdown:}\ y_{\rm G} &=& y_t + \frac{1}{\sqrt{1+m_k^2}},\; x_{\rm G} = x_t - \frac{m_k}{\sqrt{1+m_k^2}},
\label{eqn:contact}
\end{eqnarray}
\end{subequations}
where $b_0,b_1,b_2$ are constants that define the aerial phase trajectory.
Equations~\eqref{eqn:terrain-model}-\eqref{eqn:terrain contact} solved simultaneously yield a quadratic equation in $x_t$,
\begin{subequations}\label{eqn:landing}
\begin{eqnarray}
\label{eqn:quadratic}
A x_t^2 + B x_t + C &=& 0, \\
\text{where } A &=& b_2, \\
B &=& b1 - 2 b_2 \frac{m_k}{\sqrt{1+m_k^2}} - m_k, \\
C &=& b_0 - \frac{b_1 m_k - 1}{\sqrt{1+m_k^2}} + b_2 \frac{m_k^2}{1+m_k^2} - c_k.
\end{eqnarray}
\end{subequations}
The larger of the two roots of this quadratic is the true landing point $x_{\rm P}$, if the roots are real and the larger of the two roots is greater than $x_k$.
The other real root is always less than $x_k$ and corresponds to the intersection of the aerial phase trajectory with the terrain patch closer to the take-off point.
On flat terrain, the smaller root is the location of the take-off point.
Having solved for $x_{\rm P}$, the position of the center of mass at landing is determined using equation~\eqref{eqn:contact}.
However, if the runner lands on a grid point, the position of the center of mass appears to be indeterminate as the grid point $x_k$ is associated with two slopes, $m_k$ and $m_{k+1}$.
In fact, we detect a corner collision if the larger root of equation~\eqref{eqn:quadratic} is less than $x_k$, or if the roots are complex.
Thus, we now know the position of contact point P, $x_{\rm P} = x_k$, but cannot determine $(x_G,y_G)$ at contact using equation~\eqref{eqn:contact}, since $x_k$ is associated with slopes $m_k$ and $m_{k+1}$.
We determine a unique slope at the point $x_k$ by accounting for the aerial phase trajectory.
Substituting $x_t = x_k$ in equation~\eqref{eqn:quadratic}, we write equations~\eqref{eqn:landing} as a quartic polynomial in unknown $m_k$.
We numerically find all the roots using the Jenkins-Traub algorithm \citep{Jenkins1970} and pick the real root that corresponds to first contact between the ground and runner, i.e.\ when the parabolic trajectory describing the aerial phase is above the ground.
\subsection{Notation}
\label{sec:notation}
Scalars are denoted by italic symbols (e.g.\ $m$ for mass of the runner, $I$ for the moment of inertia), vectors by bold, italic symbols ($\vec{J}_{\rm imp}$ for push-off impulse, $\vec{v}$ for velocity), points or landmarks in capitalized non-italic symbols (such as center of mass G in Fig.~\ref{fig:problemsetup}) and capitalized, bold, non-italic symbols for matrices (such as return map matrix $\vec{\rm T}_{\rm an}$).
Vectors associated with a point, such as velocity of center of mass G are written as $\vec{v}_{\rm G}$, with the uppercase alphabet in the subscript specifying the point in the plane.
Angular momentum vectors or moment of inertia variables are subscripted with `/A' representing angular momentum or moment of inertia computed about point A, such as $I_{\rm /G}$ representing moment of inertia about center of mass G.
The component of velocity $\vec{v}_{\rm A}$ (velocity of point A) in the $\hat{x}$ direction is denoted with a subscript `A,t', e.g.\ tangential velocity of the contact point P is written as $v_{\rm P,t}$.
The symbols $v_{x0}, v_{y0}$ denote the initial forward and vertical velocities of the runner at take-off.
Variables just before collision with the terrain are denoted by the superscript `-', after passive collision but before push-off by the superscript `c', and just after push-off by the superscript `+'. For example, angular velocity before collision is $\omega^-$, after passive collision is $\omega^c$ and just after push-off is $\omega^+$.
\section{Authors' Contribution}
MV conceived of the model.
ND and MV ran the simulations.
All authors were involved in the analysis of the model; SM, MV and ND performed the linear stability analysis, MV and ND performed the one-step analysis to capture mean statistics, MV, SM and ND did the analysis of the steps to failure distributions.
ND and MV wrote the manuscript, and all authors edited it.
\section{Acknowledgments and funding}
This work was funded by the Human Frontiers Science Program and the Wellcome/DBT India Alliance.
\section{Competing interests}
We have no competing interests.
\singlespacing
|
\section{Introduction}
\label{sec:intro}
Modern technologies have generated a large number of datasets that possess a matrix structure for classification purpose. For example, in neuropsychiatric disease studies, it is often of interest to evaluate the prediction accuracy of prognostic biomarkers by relating two-dimensional imaging predictors, e.g., electroencephalography (EEG) and magnetoencephalography, to clinical outcomes such as diagnostic status \citep{Mu2011}. In this paper, we focus on extending one of the most commonly used classification methods, Fisher linear discriminant analysis (LDA) to matrix-valued predictors. Progress has been made in recent years on developing sparse LDA using $\ell_1$-regularization \citep{tibshirani1996}, including \citet{Shao2011, Fan2012, mai2012}. However, all these methods only deal with vector-valued covariates; and it remains challenging to accommodate the matrix structure. Naively transforming the matrix data into a high-dimensional vector will result in unsatisfactory results for several reasons. First, vectorization destroys the structural information within the matrix such as shapes and spatial correlations. Second, turning a $p \times q$ matrix into a $pq \times 1$ vector generates unmanageably high dimensionality. E.g., estimating the population precision matrix for LDA can be troublesome if $pq \gg n$. Third, $\ell_1$-regularization does not necessarily work well because the underlying two-dimensional signals are usually approximately low-rank rather than $\ell_0$-sparse.
Recently, there are some development of regression methods for matrix data. \citet{chen2013reduced} invented an adaptive nuclear norm penalization approach for low-rank matrix approximation. \citet{zhou2014} proposed a class of regularized matrix regression methods based on spectral regularization. \citet*{wang2017} developed a generalized scalar-on-image regression model via total variation. \citet{kong2018l2rm} proposed a low-rank linear regression model with high-dimensional matrix response and high dimensional scalar covariates, while \citet{hu2019nonparametric} developed a nonparametric matrix response regression model.
In this paper, we propose a new matrix LDA approach by building on the equivalence between the classical LDA and the ordinary least squares. We formulate the binary classification as a nuclear norm penalized least squares problem, which efficiently exploits the low rank structure of the two-dimensional discriminant direction matrix. The involved optimization is amenable to the accelerated proximal gradient method. Although our problem is formulated as a penalized regression problem, a fundamental difference is that the covariates ${\bf X}_i$ and the residuals $\epsilon_i$ are no longer independent in our case. This requires extra effort for developing the risk bound and rank consistency result. The risk bound is explicit in terms of the rank of the image, image size, sample size, and the eigenvalues of the covariance matrix for the image covariates. This result also implies estimation consistency provided the $p \times q$ image satisfies $\max(p,q) = o(n/\log^3 n)$. Under stronger conditions, we show that the rank of the coefficient matrix can be consistently estimated as well. The proof is based on exploiting the spectral norm of random matrices with mixture-of-Gaussian components and extending the results in \citet{Bach2008} to allow diverging matrix dimensions. Finally, we prove that our method enjoys classification error consistency.
It is worth noting that the 2D image classification problem has been studied by \citet{zhong2015matrix}, where they proposed a penalized matrix discriminant analysis method (PMDA) that projects the matrix coefficient into row space and column space separately. Those two projections are then estimated iteratively and integrated together for classification. Compared with PMDA, we make the following contributions. First, the rank of the PMDA is set as one because of the separability assumption, while we allow the rank of the direction matrix to take general positive integer values and the rank can then be selected by a data driven procedure. Our rank assumption is more flexible in practice and hence often leads to a lower mis-classification error in the numerical studies. Second, our method adopts a direct estimation approach by solving a nuclear norm penalized regression problem, which is computationally much faster compared with PMDA, where the estimation involves an iterative procedure for calculating the inverse of covariance matrices during each iteration. Third, our method can handle the high-dimensional data when image dimensions $p $ and $ q $ are much larger than the sample size, which is the case for many applications; while PMDA cannot handle the case when $ p+q>n$. Finally, we have provided theoretical guarantee for our estimator when $ p $ and $ q $ diverge with $ n $. In particular, we have developed an non-asymptotic error bound for the estimated LDA direction, as well as results on rank consistency and classification error consistency. These results are stronger compared with the root-$n$ consistency of the LDA direction in \citet{zhong2015matrix}, where both $ p $ and $ q $ are assumed to be fixed.
\section{Method}
We first define some useful notations. Let $\text{vec}(\cdot)$ be a vectorization operator, which stacks the entries of a matrix into a column vector. The inner product between two matrices of same size is defined as $\langle {\bf M, N} \rangle=\text{tr}({\bf M}^{{ \mathrm{\scriptscriptstyle T} }} {\bf N})=\langle \text{vec}({\bf M}), \text{vec}({\bf N}) \rangle$.
Consider a binary classification problem, where $\bf X$ is a two-dimensional image covariate with dimension $p\times q$ and $G = 1, 2$ denotes the class labels. The LDA assumes that $\text{vec} ({\bf X}) \mid G = g\sim N({\bf \mu}_g, {\bf \Sigma})$, $\text{pr}(G=1) = \pi_1$, and
$\text{pr}(G=2) = \pi_2$. Suppose we have $n$ subjects with $n_1$ subjects belonging to class 1 and $n_2=n-n_1$ subjects to class 2. It is well known that LDA is connected to the linear regression with the class labels as responses \citep{duda2012pattern, mika2002}. When $ pq<n $, the classical LDA is equivalent to solving
\begin{align} \label{OLS}
(\hat\beta_0^{\text{ols}}, \hat{\bf B}^{\text{ols}} ) = \argmin_{\beta_0,B}\sum_{i=1}^n\Big (y_i - \beta_0 - \langle {\bf{X}_i, B}\rangle \Big)^2,
\end{align}
where ${\mathbf X}_i$ is the image covariate from subject $i$, $\mathbf B$ is the coefficient matrix for the image covariate and it represents the direction of the linear discriminant classifier, $\beta_0$ is the intercept, and the response $y_i = -n/n_1$ if subject $i$ is in class 1, and $y_i = n/n_2$ if subject $i$ is in class 2. Although this connection gives the exact LDA direction when $ pq < n $, it has two potential drawbacks. First, when $ pq>n $, the equivalence between Fisher LDA and (\ref{OLS}) is lost because of the non-uniqueness of solution. Second, the formulation (\ref{OLS}) does not incorporate the 2D image structure when estimating the direction because $ \langle {\bf X}_i, {\bf B} \rangle=\langle \text{vec}({\bf X}_i), \text{vec}({\bf B}) \rangle$. These motivate us to consider a penalized version of (\ref{OLS}) as follows
\begin{align} \label{POLS1}
(\hat\beta_0, \hat {\bf B}) = \argmin_{\beta_0, {\bf B}}\frac{1}{2n} \sum_{i=1}^n \Big(y_i - \beta_0 - \langle{\bf X}_i,{\bf B} \rangle \Big)^2 + \omega_n\|{\bf B}\|_\ast,
\end{align}
where the nuclear norm $ \|{\bf B}\|_{\ast}=\sum_{j}\sigma_j({\bf B}) $ and $ \sigma_j({\bf B}) $s are the singular values of the matrix $ {\bf B} $. The nuclear norm $ \|{\bf B}\|_\ast $ plays an important role because it imposes a low rank structure in the estimated direction $ \hat {\bf B} $. An alternative choice is to add a Lasso type penalty, i.e. $ \omega_n\|{\bf B}\|_{1,1}=\omega_n\sum_{j=1}^p \sum_{k=1}^q |b_{jk}|$, where $ b_{jk} $ is the $jk$-th element of $ {\bf B} $. However, the Lasso type penalty can only identify at most $ n $ nonzero components, and for most cases in imaging studies, the signal is usually not that sparse. More importantly, the Lasso type of penalty ignores the matrix structure because it is equivalent to vectorizing the array and applying sparse LDA. Once $ \hat {\bf B} $ from (\ref{POLS1}) is obtained, a naive classification rule will assign the $i$-th subject to class 2 if $ \langle {\bf X}_i, \hat {\bf B} \rangle + \hat\beta_0 > 0 $. However, it can be shown that the intercept $ \hat\beta_0 $ obtained from (\ref{POLS1}) is not optimal. Instead, we use the optimal intercept $\tilde\beta_0$ that minimizes the training error after obtaining $\hat {\bf B}$. \citet{mai2012} showed that the intercept of LDA actually has a closed form. Their derivations can be easily applied to our case. In particular, if $(\hat\mu_2 - \hat\mu_1)^{{ \mathrm{\scriptscriptstyle T} }}\text{vec}(\hat {\bf B}) >0$, then
\begin{eqnarray}\label{optimalintercept}
\tilde\beta_0=-(\hat\mu_1 +\hat\mu_2)^{{ \mathrm{\scriptscriptstyle T} }}\text{vec}(\hat {\bf B})/2 + \text{vec}(\hat {\bf B})^{{ \mathrm{\scriptscriptstyle T} }}\hat{\bf \Sigma} \text{vec}(\hat {\bf B})\{(\hat\mu_2 - \hat\mu_1)^{{ \mathrm{\scriptscriptstyle T} }}\text{vec}(\hat {\bf B})\}^{-1}\log(n_2/n_1),
\end{eqnarray}
where $\hat\mu_g$ is the sample mean for subjects in class $g$ and $\hat\Sigma$ is the estimated covariance matrix. If $(\hat\mu_2 - \hat\mu_1)^{{ \mathrm{\scriptscriptstyle T} }}\text{vec}(\hat {\bf B}) <0$, we can plug $ -\hat {\bf B} $ into (\ref{optimalintercept}) to obtain the optimal intercept $ \tilde\beta_0 $. The optimal classification rule is to assign the $i$-th subject to class 2 if $ \langle {\bf X}_i, \hat {\bf B} \rangle + \tilde\beta_0 > 0$.
For any fixed $\omega_n$, the optimization problem in \eqref{POLS1} can be solved using the accelerated proximal gradient method \citep{nesterov1983, beck2009}. \citet*{zhou2014} studied the algorithm for the nuclear norm regularized matrix regression. As we know, nuclear norm is not differentiable. Fortunately, its subderivative $\partial \|.\|_\ast$ exists. Therefore \eqref{POLS1} has local minima $(\hat\beta_0, \hat {\bf B})$ if and only if $0 \in -\frac{1}{n}\sum_{i=1}^n {\bf X}_i\epsilon_i + \omega_n\partial \|\hat {\bf B}\|_\ast.$ Thanks to the convexity of nuclear norm, the local minima is global as well. Based on these facts, singular value thresholding method for nuclear norm regularization was deployed for building blocks of the Nesterov's method. Compared with classical gradient decent method with convergence of $O(t^{-1})$, where $t$ denotes the number of iteration, Nesterov's accelerated gradient decent method achieves convergence rate of $O(t^{-2})$. It differs from traditional algorithms by utilizing the estimators from previous two iterations to generate the next estimator. For computational algorithm, we use the {\tt matrix\_sparsereg} function in the Matlab TensorReg Toolbox (\url{https://hua-zhou.github.io/TensorReg/}) for solving nuclear norm penalized matrix regression. It implements an optimal Nesterov acceleration of the proximal gradient algorithm. Actually one contribution of our paper is to link matrix LDA to regularized matrix regression so that the computational machinery developed for the latter can be applied to matrix LDA problems. For tuning of the $ \omega_n $, we adopt the \textsc{bic} derived by \citet*{zhou2014} under the nuclear norm regularized matrix regression framework.
\section{Theory}
In this section we discuss the theoretical properties of the proposed regularization estimator. Denote the residuals $\epsilon_i=y_i - \beta_0 - \langle{\bf X}_i,{\bf B} \rangle$ and the true coefficient matrix by ${\bf B}_0$. {By the equivalence between LDA direction and least squares, we know $\text{vec}({\bf B}_0)$ can be written as $c{\bf \Sigma}^{-1}(\mu_2 - \mu_1)$ for some positive constant $c$.} Consider the singular value decomposition ${\bf B}_0 = {\bf U}_0\text{Diag}(S_0){\bf V}_0^{{ \mathrm{\scriptscriptstyle T} }}$ with ${\bf U}_0 \in \mathrm{R}^{p \times r}$ and ${\bf V}_0 \in \mathrm{R}^{q \times r}$. Let ${\bf U}_{0\perp} \in \mathrm{R}^{p \times (p-r)} $ and ${\bf V}_{0\perp} \in \mathrm{R}^{q \times (q-r)}$ be (arbitrary) orthogonal complements of ${\bf U}_0$ and ${\bf V}_0$, respectively. We make the following assumptions.
\begin{itemize}
\item [(A1)] We assume that the second-order moment of the covariate ${\bf X}$, $\text{E}(\text{vec}({\bf X})\text{vec}({\bf X})^{{ \mathrm{\scriptscriptstyle T} }})$, denoted by ${\bf \Sigma}_{xx}$, satisfies $\lambda_l \leq \lambda_{\min}({\bf \Sigma}_{xx}) \leq \lambda_{\max}({\bf \Sigma}_{xx}) \leq \lambda_u$, where $\lambda_{\min}({\bf \Sigma}_{xx})$ and $\lambda_{\max}({\bf \Sigma}_{xx})$ are the smallest and largest eigenvalues of ${\bf \Sigma}_{xx}$, respectively, and $\lambda_l, \lambda_u$ are some positive constants.
\item [(A2)] Let $ r = \text{rank}({\bf B}_0)$ be the unknown rank of the true coefficient matrix ${\bf B}_0$. Define ${\bf \Lambda} \in \mathrm{R}^{(p-r)\times(q-r)}$ as
\begin{align*}
\text{vec}({\bf \Lambda}) = \{({\bf V}_{0\perp}\otimes {\bf U}_{0\perp})^{{ \mathrm{\scriptscriptstyle T} }}{\bf \Sigma}^{-1}({\bf V}_{0\perp}\otimes {\bf U}_{0\perp})\}^{-1}
\{({\bf V}_{0\perp}\otimes {\bf U}_{0\perp})^{{ \mathrm{\scriptscriptstyle T} }}{\bf \Sigma}^{-1}({\bf V}_0\otimes {\bf U}_0)\text{vec}({\bf I})\}.
\end{align*}
We assume its spectral norm $\|{\bf \Lambda}\|_2 < 1$.
\item [(A3)] Assume the quantities $\omega_n$, $\{\min(p,q)\}^{1/2}n^{-1/2}\omega_n^{-1}$, $ \min(p,q)n^{-1/2}$, $\omega_np^{1/2}q^{1/2}\min(p,q)$ tend to $0$ as $n \rightarrow \infty$.
\item [(A4)] There exists a positive constant $C_{\mu}$ such that $\| {\bf \mu}_2 - \mu_1 \|_2 \leq C_{\mu} (\sqrt{p} + \sqrt{q})$.
\end{itemize}
Condition (A1) requires bounded eigenvalues for the covariance matrix of the vectored covariate, which is standard in the literature. Condition (A2) is similar with the strict consistency condition in \citet{Bach2008}. It is needed to establish rank consistency. This condition extends the classical strong irrepresentable condition in \citet{Zhao2006}, which is commonly used for proving model selection consistency of Lasso. The major difference between our Assumption (A2) and the similar assumption in \citet{Bach2008} is that the number of parameters is fixed in \citet{Bach2008} while in our case the number is diverging with $n$. Therefore we will need to assume that the regularization parameter $\omega_n$ decays slower than the one in \citet{Bach2008}. Condition (A3) puts more requirement on the order of $p,q$, and $w_n$ in order to obtain consistent rank estimation in addition to consistent coefficient estimation. This is expected since rank estimation consistency is usually not implied by parameter estimation consistency. Condition (A4) can be viewed as a sparsity assumption on $B_0$. Recall the solution (the slope) to classical LDA problem with vector covariates depends on the term $\mu_2 - \mu_1$. This assumption essentially implies that there are at most $O(\max(p,q))$ number of $O(1)$ elements in the true coefficient matrix ${\bf B}_0$ given the rank of ${\bf B}_0$ is fixed.
Next, we briefly review two important concepts, namely decomposable regularizer and strong convex loss function, proposed by \citet{Nega2012} and highlight their connection to the risk bound property for our estimator.
\begin{definition}
A regularizer $R(\cdot)$ is decomposable with respect to a given pair of subspaces $(M, N)$ where $M \subseteq N^\perp$ if
\[
R(u + v) = R(u) + R(v) \text{\quad for all } u\in M, v\in N.
\]
\end{definition}
In our setting, $R(\cdot)$ is the nuclear norm. Considering a matrix ${\bf B} \in \mathcal R^{p\times q}$ to be estimated, we observe that nuclear norm is decomposable given a pair of subspaces:
\[
M({\bf U}, {\bf V}) := \{{\bf B} \in \mathcal R^{p\times q} \mid \text{row}({\bf B}) \subseteq {\bf V}, \text{ col}({\bf B}) \subseteq {\bf U}\},
\]
\[
N({\bf U}, {\bf V}) := \{{\bf B} \in \mathcal R^{p\times q} \mid \text{row}({\bf B}) \subseteq {\bf V}^{\perp}, \text{ col}({\bf B}) \subseteq {\bf U}^{\perp}\},
\]
where ${\bf U}, {\bf V}$ represent $B$'s left and right singular { vectors}. For any pair of matrices ${\bf B}_1 \in M$ and ${\bf B}_2 \in N$, the inner product of ${\bf B}_1, {\bf B}_2$ is $0$ due to their mutually orthogonal rows and columns. Hence we conclude $R({\bf B}_1 + {\bf B}_2) = R({\bf B}_1) + R({\bf B}_2)$. Since we assume the true parameter has a low rank structure, we expect the regularized estimator to have a large value of projection on $M({\bf U}, {\bf V})$ and a relatively small valued projection on $N({\bf U}, {\bf V})$.
When the loss function $L(\hat\beta_0, \hat {\bf B}_{\omega_n} )$ defined as $ \frac{1}{2n} \sum_{i=1}^n \Big(y_i - \hat\beta_0 - \langle {\bf X}_i, \hat {\bf B}_{\omega_n} \rangle \Big)^2$ is close to $L(\beta_0, {\bf B}_0)$, it is insufficient to claim $\hat {\bf B}_{\omega_n} - {\bf B}_0$ is small if the loss function $L$ is relatively flat. This is why the strong convexity condition is required.
\begin{definition}
For a given loss function $L$ and norm $\|.\|$, we say $L$ is strong convex with curvature $k_L$ and tolerance function $\tau_L$ if
\[
\delta L({\bf \Delta}, {\bf B}_0) \geq k_L\|{\bf \Delta}\|^2 - \tau_L^2({\bf B}_0), \text{\quad for any }\delta \in \mathcal C(M,N;{\bf B}_0),
\]
where $\mathcal C(M,N;{\bf B}_0) := \{{\bf \Delta} \in \mathcal R^{p\times q} \mid R({\bf \Delta}_{N})\leq 3R({\bf \Delta}_{N^\perp}) + 4R({\bf B}_{0N})\}$.
\end{definition}
Now we are ready to state the main result on the risk bound for our estimate. The proof is provided in the Appendix B.
\begin{theorem}\label{tm1}
Suppose that (A1) and (A4) hold. Let $\hat{{\bf B}}$ be the solution to \eqref{POLS1}. If $$\omega_n \geq \frac{12 (\log n)^{3/2} (C_{\mu} + \lambda_u^{1/2}) (\sqrt{p} + \sqrt{q} + \sqrt{\log n}) }{\sqrt{n}}, $$ then with probability of at least $1 - C n^{-1}$ for some constant $C>0$,
\begin{align*}
\|\hat {\bf B} -{\bf B}_0\|_F^2 + |\hat \beta - \beta_0^*|^2 \leq 9\frac{\omega_n^2}{\lambda_l}r,
\end{align*}
where $\beta_0^* = \beta_0 - \pi_2^{-1} \{ c - 1+(\pi_2 - \pi_2^2) ({\bf D}^{{ \mathrm{\scriptscriptstyle T} }} {\bf \Sigma}^{-1}{\bf D})\}$ and $c$ is some positive constant.
\end{theorem}
Theorem \ref{tm1} gives a non-asymptotic risk bound for the proposed estimators. In other words, the results hold for any positive $\omega_n$ satisfying the conditions there. However, in order to ensure the consistency of the proposed estimators, we will need the risk bound to go to $0$, which requires $\omega_n \rightarrow 0$ and $\max(p,q) = o\left(n/(r \log^3 n)\right)$. If the rank of ${\bf B}_0$ is fixed, then both $p$ and $q$ can diverge with $n$ at the order of $o(n/\log^3 n)$ and their product $p q > n$. This result is compatible with Theorem 1 in \citet{Rask2015}. Note that the estimated intercept $\hat \beta$ converges to $\beta_0^*$, which deviates from the truth $\beta_0$. This is expected because the solution to OLS is only equivalent with LDA's solution in terms of the slope ${\bf B}$, not on $\beta_0$. More precisely, for OLS, by taking the derivative of squared loss function with respect to $\beta_0$ and set it to $0$, we essentially require $E (\epsilon) = 0$. However, this does not hold in our case. Instead we need to shift the residual $\epsilon$ by $d$ to balance off the bias in the cross-product term E$(\epsilon {\bf X})$. The proof of the theorem uses Gaussian comparison inequality which allows us to deal with $\text{vec}({\bf X})$ following a general Gaussian distribution instead of standard Gaussian distribution given that the largest singular value of ${\bf \Sigma}_{xx}$ is bounded. Based on this connection, we further utilize concentration property of spectral norm of Gaussian random matrices.
Next we show that $\hat{\bf B}$ is rank-consistent under stronger conditions.
\begin{theorem}\label{thm2}
Suppose that (A1)--(A4) hold. Then the estimate $\hat {\bf B}$ is rank-consistent, that is, $P(\mbox{rank}(\hat {\bf B})=\mbox{rank}({{\bf B}}_0))\rightarrow 1$ as $ n\rightarrow \infty$.
\end{theorem}
Similar to Lasso, estimation consistency does not guarantee correct rank estimation for matrix regularization. In fact, the assumptions here are stronger than those in Theorem \ref{tm1}. For example, Theorem \ref{tm1} allows $p + q = o(n/\log^3n)$ while Theorem \ref{thm2} requires $\max(p,q) = o\left( n^{1/3} \log^{-3/2} n\right )$ if $\min(p,q) = O(1)$. The proof is based on the arguments in \citet{Bach2008} with modifications to allow diverging $p$ and $q$.
\begin{remark}
Although nuclear norm penalized least squares is used to estimate the classification direction, there is a fundamental difference between our theorems and the theoretical results derived for nuclear norm penalized least squares regression \citep{Bach2008, Nega2012}. The previous work assumes that the data obey a linear regression model with covariates-independent additive noise, which is not true in our case. In particular, the covariates ${\bf X}_i$ and the residuals $\epsilon_i$ are no longer independent in our problem, which brings additional challenges in developing theoretical results.
\end{remark}
Next we state a classification error consistency result. To be consistent with the notation in the classification literature, for subject $i$, we use $Y_i \in \{-1,1\}$ to denote its true label, $\hat{f}_n({\bf X_i})$ as the classified label for which $\hat{f}_n$ is the classification rule obtained by solving \eqref{POLS1}, and $l(Y_i ,f({\bf X_i})) = I\{Y_i \neq \text{sign}(f({\bf X_i}))\} $ as the 0-1 loss function. Define the risk of $\hat{f}_n$ as $R(\hat{f}_n) = E_{\bf X} l(Y ,\hat{f}_n({\bf X}))$ and the Bayes risk as $R^* = \inf_f R(f)$. In addition, we assume that the true label $Y_i$ given ${\bf X_i}$ is determined by the linear classification rule with coefficients $\beta_0^*$ and ${\bf B}_0$. Then the following theorem shows that the proposed classifier achieves the Bayes optimal risk under certain conditions. The proof, given in the Appendix B, is based on the general results in \citet{Zhang2004}, where the author studied the optimal Bayes error rate using a classifier obtained by minimizing a convex upper bound of the classification error function.
\begin{theorem}\label{tm3}
Assume the same conditions for Theorem \ref{tm1} hold and $\omega_n \rightarrow 0$. Then $R(\hat{f}_n) \rightarrow R^* $ as $n \rightarrow \infty$.
\end{theorem}
\section{Numerical results}
\subsection{Simulation}
We conduct simulation studies to evaluate the numerical performance of our proposed method. We compare its performance with that of a few alternatives: ``Lasso LDA'', which adopts a naive Lasso penalty in LDA without taking into account matrix structure, the regularized matrix logistic regression \citep{zhou2014} using nuclear norm and Lasso penalties, denoted by ``Logistic Nuclear'' and ``Logistic Lasso'', and the penalized matrix discriminant analysis (PMDA) approach proposed by \citet{zhong2015matrix}. We generate $n \in \{100,200,500\}$ samples from two classes with weights $(\pi_1,\pi_2) \in \{(0.5,0.5),(0.75,0.25)\}$. For each class, we generate predictors from a bivariate normal distribution with means $\mu_g$, $g = 1, 2$, and covariance ${\bf \Sigma}$. We set $\mu_1 = 0$ and $\mu_2 = {\bf \Sigma} \text{vec}({\bf B}_0)$. The covariance matrix ${\bf \Sigma}$ has a 2D autoregressive structure: $ {\rm cov}({\bf x}_{i_1,j_1}, {\bf x}_{i_2,j_2})=0.5^{|i_1-i_2| + |j_1-j_2|}$ for $ 1\leq i_1\leq p $ and $ 1\leq j_1\leq q $. The true signal ${\bf B}_0$ is generated based on a 64-by-64 image. We consider three settings: a cross, a triangle and a butterfly. These pictures are shown in Figure \ref{fig2}(a) respectively. In particular, the white color denotes value 0 and black denotes 0$.$05. We apply each fitted model to an independent test data set of size $1000$ and summarize the misclassification rates based on 1000 Monte Carlo replications. The results are contained in Table \ref{tb1}.
\begin{table}
\centering
\def~{\hphantom{0}}
\caption{Simulation results: misclassification rates ($\%$) and associated standard errors obtained from our method, Lasso LDA, Logistic Nuclear (L-Nuclear), Logistic Lasso (L-Lasso) and penalized matrix discriminant analysis (PMDA) based on 1000 Monte Carlo replications. }{%
\begin{tabular}{lccccccc} \hline
Shape &n&$(\pi_1, \pi_2)$& Ours & Lasso LDA& L-Nuclear &L-Lasso& PMDA\\[5pt]
Cross & 100&(0$.$5,0$.$5)&3$.$65(0$.$02) &17$.$81(0$.$07) &3$.$70(0$.$02)&19$.$51(0$.$07)&*\\
& 100&(0$.$75,0$.$25) & 3$.$32(0$.$02) &14$.$89(0$.$05) &6$.$62(0$.$04) &18$.$84(0$.$04)&* \\
& 200 &(0$.$5,0$.$5)&3$.$22(0$.$02) &11$.$69(0$.$05) &3$.$26(0$.$02) &13$.$39(0$.$05)&26$.$93(0$.$05) \\
& 200 &(0$.$75,0$.$25)&2$.$87(0$.$02) &9$.$89(0$.$04) &4$.$14(0$.$03) &16$.$27(0$.$04)&19$.$58(0$.$08) \\
& 500 &(0$.$5,0$.$5)&3$.$09(0$.$02) & 6$.$97(0$.$03) &3$.$11(0$.$02) &8$.$19(0$.$04)&25$.$17(0$.$04) \\
& 500 &(0$.$75,0$.$25)&2$.$62(0$.$02) &5$.$81(0$.$03) &3$.$59(0$.$02) &14$.$91(0$.$03)&12$.$05(0$.$04) \\[ 10pt]
Triangle & 100&(0$.$5,0$.$5) &3$.$12(0$.$02) &15$.$73(0$.$06) &3$.$11(0$.$02) &17$.$70(0$.$07)&* \\
& 100&(0$.$75,0$.$25) & 2$.$66(0$.$02)&13$.$72(0$.$05) &6$.$10(0$.$04) &17$.$19(0$.$04)&* \\
& 200 &(0$.$5,0$.$5)&2$.$85(0$.$02) &9$.$90(0$.$04) &2$.$81(0$.$02) &11$.$81(0$.$04)& 30$.$17(0$.$08)\\
& 200 &(0$.$75,0$.$25)&2$.$43(0$.$02) &8$.$72(0$.$03) &3$.$62(0$.$02) &13$.$40(0$.$04)&24$.$63(0$.$10) \\
& 500 &(0$.$5,0$.$5)&2$.$67(0$.$02) & 5$.$67(0$.$03)&2$.$73(0$.$02) &6$.$96(0$.$03)&25$.$92(0$.$04)\\
& 500 &(0$.$75,0$.$25)&2$.$29(0$.$01) &4$.$89(0$.$02) &2$.$74(0$.$02) &9$.$97(0$.$03) &14$.$69(0$.$05)\\ [ 10pt]
Butterfly & 100&(0$.$5,0$.$5) &3$.$86(0$.$02) &17$.$10(0$.$06) &4$.$16(0$.$02) &18$.$82(0$.$07)&* \\
& 100&(0$.$75,0$.$25) & 3$.$47(0$.$02)&14$.$79(0$.$05) &7$.$14(0$.$04) &17$.$78(0$.$04)&* \\
& 200 &(0$.$5,0$.$5)&3$.$67(0$.$02) &11$.$00(0$.$04) &3$.$78(0$.$02)&12$.$66(0$.$05)&29$.$79(0$.$07)\\
& 200 &(0$.$75,0$.$25)&3$.$26(0$.$02) &9$.$80(0$.$04) &4$.$50(0$.$02) &13$.$93(0$.$04)&23$.$83(0$.$09) \\
& 500 &(0$.$5,0$.$5)&3$.$56(0$.$02) & 6$.$50(0$.$03) &3$.$52(0$.$02) &7$.$70(0$.$03)&25$.$77(0$.$04) \\
& 500&(0$.$75,0$.$25) & 3$.$02(0$.$02)&5$.$74(0$.$03) &3$.$51(0$.$02) &10$.$49(0$.$03)&14$.$66(0$.$05) \\ \hline
\end{tabular}}
\label{tb1}
\end{table}
The results show that our method performs much better than ``Lasso LDA" and ``Logistic Lasso" under all scenarios. This is expected because these two methods ignore the matrix structure. For ``Logistic Nuclear'', it has similar misclassification rates with our method for balanced data, but does not perform as good as ours for unbalanced data. We have also plotted the estimates using nuclear norm and $\ell_1$-norm from one randomly selected Monte Carlo replicate in Figure \ref{fig2}(b)(c). It can be seen that the proposed nuclear norm regularization is much better than $\ell_1$-regularization in recovering the matrix signal in different shapes. By comparing the recovery of different shapes in Column (b) in Figure \ref{fig2}, we find that our method works better for cross than for triangle and butterfly. This is expected since triangle and butterfly do not have the low rank structure.
We also compare the performance of our method with that of PMDA proposed by \citet{zhong2015matrix}. In Table \ref{tb1}, it can be seen that our proposed method has a lower mis-classification rate under all scenarios. This is because we allow flexible values of the rank for the linear discriminant direction $ {\bf B} $, while in \citet{zhong2015matrix}, their assumption is equivalent to assuming $ {\bf B} $ is of rank 1. In particular, using their notation, for binary case, their direction $ {\bf B}=\boldsymbol\beta_1\boldsymbol\xi^{{ \mathrm{\scriptscriptstyle T} }}$, where $\boldsymbol\beta_1 \in \mathbb{R}^{p}$ and $\boldsymbol\xi\in \mathbb{R}^{q}$. Since the true ranks of $ {\bf B} $ in our simulation studies are all of rank greater than 1, it is not surprising that our method outperforms PMDA. Moreover, PMDA does not apply to the case where $n < p + q$, i.e., the sample size is far smaller than the summation of image dimensions. Therefore, their method does not apply to one of our simulation settings $(n,p,q)=(100,64,64)$ and we mark their results using $*$ in Table \ref{tb1}. We also compare the computation time between PMDA and our method. In simulation, when $n=200$ and true signal is a cross, given a fixed regularization parameter, the system running time of PMDA is around 1.5 minutes whereas the system running time of our method is no more than 13 seconds. Here system running time is measured on a Macbook Pro laptop with a 2.9 GHz Intel Core i5. This is because PMDA essentially solves least square problems with $L_1$ penalty in each iteration when setting $\omega_1 =0$ in Algorithm 2 in \citet{zhong2015matrix}. Our method is based on the Nesterov optimal gradient method which avoids computing inverse of covariance matrix and hence has a faster convergence rate.
\begin{figure}[h]
\begin{tabular}{ccc}
\includegraphics[width=0.3\linewidth]{f1}\label{(a)}&
\includegraphics[width=0.3\linewidth]{f2}\label{(b)}&
\includegraphics[width=0.3\linewidth]{f3}\label{(c)}\\
\includegraphics[width=0.3\linewidth]{sanjiao1}\label{(a)}&
\includegraphics[width=0.3\linewidth]{sanjiao2}\label{(b)}&
\includegraphics[width=0.3\linewidth]{sanjiao3}\label{(c)}\\
\includegraphics[width=0.3\linewidth]{hudie1}\label{(a)}&
\includegraphics[width=0.3\linewidth]{hudie2}\label{(b)}&
\includegraphics[width=0.3\linewidth]{hudie3}\label{(c)}\\
(a)&(b)&(c)\\
\end{tabular}
\caption{The figures for cross image: (a) original signal; (b) our nuclear regularization estimate; (c) $\ell_1$-regularized estimate.}\label{fig2}
\end{figure}
\subsection{Real data application}
We apply our method to an EEG dataset, which is available at \url{https://archive.ics.uci.edu/ml/datasets/EEG+Database}. The data was collected by the Neurodynamics Laboratory to study the EEG correlates of genetic predisposition to alcoholism. It contained measurements from 64 electrodes placed on each subject's scalps sampled at 256 Hz (3$.$9-msec epoch) for 1 second. Each subject was exposed to three stimuli: a single stimulus, two matched stimuli, two unmatched stimuli. Among the 122 subjects in the study, 77 were alcoholic individuals and 45 were controls. More details about the study can be found in \citet{zhang1995}. In statistics literature, EEG data has been analyzed using different models, for example, \citet{Gao20191} considered an unsupervised approach for clustering EEG data, \citet{Gao20192} and \citet{Gao2018} considered an evolutionary state-space model and graphical model for better understanding brain connectivity, respectively. However, these methods are not directly applicable for classification purpose here.
In our data analysis, for each subject, we use the average of all 120 runs for each subject under single-stimulus condition and use that as the covariate ${\bf x}_i$, which is a $256\times64$ matrix. The classification label is \emph{alcoholic} or not.
We randomly divide the data set into training set of 81 subjects and test set of 41 subjects for 100 times, and each time fit the model on the training set and apply it on the test set to obtain the average mis-classification rate and its standard error. The results for different methods are summarized in Table \ref{tb2}. It can be seen that the proposed method has a significant lower mis-classification rate compared with other methods, which agrees with the simulation findings for the unbalanced data. PMDA does not work here since $p+q >n $ ($(n,p,q)=(122, 256, 64)$). We also check the fitted signal matrix and it agrees well with the one obtained by \citet{zhou2014}.
In terms of computational efficiency, we measured the computation time among Lasso LDA, Logistic Nuclear, Logistic Lasso and our method based on one evaluation of the data, that is, partitioning the data into training and test sets, fitting the model on the training set and applying it on the test set. The running time for Lasso LDA, Logistic Nuclear, Logistic Lasso and our method is 0.67s, 1.79s, 1.27s and 1.87s, respectively. The system running time is measured in Matlab R2015b on a Macbook Pro laptop with a 2.9 GHz Intel Core i5.
\begin{table}
\centering
\def~{\hphantom{0}}
\caption{EEG data analysis: misclassification rates ($\%$) and associated standard errors.}{%
\begin{tabular}{lrrrr}
Our method & Lasso LDA & Logistic Nuclear& Logistic Lasso & PMDA\\
22$.$20(0$.$53)&24$.$12(0$.$70)&24$.$44(0$.$80)&26$.$24(0$.$91)&*
\end{tabular}}
\label{tb2}
\end{table}
\begin{comment}
\begin{table}
\centering
\def~{\hphantom{0}}
\caption{Electroencephalography data analysis: misclassification rates ($\%$).}{%
\begin{tabular}{lcccc}
Method& leave-one-out &5-fold &10-fold &20-fold \\
Our method&0$.$205&0$.$172&0$.$189&0$.$180\\
Lasso LDA& 0$.$238& 0$.$221&0$.$250 &0$.$262\\
Logistic Nuclear& 0$.$230&0$.$214&0$.$222&0$.$181\\
Logistic Lasso&0$.$246&0$.$287&0$.$271&0$.$264\\
\end{tabular}}
\label{tb2}
\end{table}
\end{comment}
\section{Discussion}
In the literature, total variation (TV) regularization has also been commonly used for modeling image data in addition to the proposed nuclear norm regularization. Their focuses are slightly different -- the former is on structured sparse pattern and the later is on low-rank pattern. The main reason that we choose to focus on the nuclear norm regularization in this paper is because we have found that low rankness is a more reasonable assumption than sparseness assumption in our real data application. In particular, the mis-classification errors of our method is lower than the sparse method (LASSO) in our real data analysis. The TV regularization is an interesting direction to explore as it requires new computational algorithms and theories; and thus we leave this for the future research.
In this paper, we only consider the situation where all the image measurements are taking at the same scale, that is, the dimension of the image covariates $p$ and $q$ are equal for every study subject. We believe this is the case for most applications. For the special cases when image dimensions vary across subjects, our method may still be applicable by first resizing the image to the same scale. It will be of future interest to develop flexible statistical methods to handle image data that can be of different sizes in general.
\section*{Acknowledgment}
The authors would like to thank the Editor, Associate Editor and two reviewers for their constructive comments, which have substantially improved the paper. Shen's research is partially supported by Simons Foundation Award 512620 and NSF DMS-1509023. Zhou's research is partially supported by NIH grants R01HG006139, R01GM53275 and NSF DMS-1310319. Kong's research
is partially supported by the Natural Science and Engineering Research Council of Canada.
|
\section{Introduction}
\label{sec:int}
Giant extrasolar planets that orbit their host stars at distances shorter than $\approx$ 1 AU but farther away than the hot-Jupiter pile-up at $\approx$ 0.1 AU, are termed ``warm" giants. They have been efficiently discovered by radial velocity (RV) surveys \citep[e.g.,][]{hebrard16,jenkins17}, and have a wide distribution for their eccentricities, with a median of $\approx0.25$. The origin for these eccentricities is a topic of active research because
the migration of planets through interactions with the protoplanetary disc predicts circular orbits \citep{dunhill:2013}, while planet-planet scattering after disc dispersal at typical warm giant orbital distances should generate usually planet collisions rather than high eccentricity excitations \citep{petrovich:2014}.
Transiting giants are key for constraining theories of orbital evolution of exoplanets. Besides providing the true mass of the planet, follow-up observations can be carried out to constrain the sky-projected spin-orbit angle (obliquity) of the system, which is a tracer of the migration history of the planet \citep[e.g.,][]{zhou:2015, esposito:2017, mancini:2018}. While the obliquity for hot giant ($P < 10$ d) systems can be affected by strong tidal interactions \citep{triaud:2013,dawson:2014}, the periastra of warm giants are large enough that significant changes in the spin of the outer layers of the star are avoided, and thus the primordial obliquity produced by the migration mechanism should be conserved.
Unfortunately, the number of known transiting warm giants around nearby stars is still very low. In addition to the scaling of the transit probability as $a^{-1}$, the photometric detection of planets with $P > 10$ days requires
a high duty cycle, which puts strong limitations on the ability of ground-based wide-angle photometric surveys \citep[e.g.,][]{bakos:2004,pollacco:2006,bakos:2013} to discover warm giants. From the total of $\approx 250$ transiting giant planets detected from the ground, only 5 have orbital periods longer than 10 d \citep{kovacs:2010,hatp17,wasp117,brahm:2016:hs17,wasp130}. On the other hand, the \textit{Kepler} and CoRoT space missions found dozens of warm giants \citep[e.g. ][]{corot9,corot10,dawson:2012,borsato:2014}, but orbiting mostly faint stars, for which detailed follow-up observations are very challenging.
Due to their relatively low equilibrium temperatures ($\ensuremath{T_{\rm eq}} < 1000$ K), transiting warm giants are important objects for characterizing the internal
structure of extrasolar giant planets since their atmospheres are not subject to the yet unknown mechanisms that inflate the radii of typical hot Jupiters \citep[for a review see][]{fortney:2010}.
For warm giants, standard models of planetary structure can be used to infer their internal composition from mass and radii measurements \citep[e.g.,][]{thorngren:2016}.
In this work we present the discovery of an eccentric warm giant planet orbiting a bright star, having physical parameters similar to those of Saturn. This discovery was made in the context of the K2CL collaboration, which has discovered a number of planetary systems using K2 data \citep{brahm:2016:k2,espinoza:2017:k2,jones:2017,giles:2018,soto:2018,k2-232,k2-261}.
\section{Observations} \label{sec:obs}
\subsection{K2}
Observations of campaign 15 (field centered at RA=15:34:28 and DEC=-20:04:44) of the K2 mission \citep{howell:2014} took place between August 23 and November 20 of 2017. The data of K2 campaign 15 was released on March 2018.
We followed the steps described in previous K2CL discoveries to process the light curves and identify transiting planet candidates. Briefly, the K2 light curves for
Campaign 15 were detrended using our implementation of the EVEREST algorithm \citep{luger:2016}, and a Box-Least-Squares \citep[BLS;][]{BLS} algorithm was used to find candidate
box-shaped signals. The candidates that showed power above the noise
level were then visually inspected to reject evident eclipsing binary systems
and/or variable stars. We identified 23 candidates in this field. Among those candidates, K2-287\ (EPIC 249451861) stood out as a high priority candidate for follow-up due to its relative long period, deep flat-bottomed transits, and bright host star ($V=11.4$ mag).
The detrended light curves of the six transits observed for K2-287\ by K2 are displayed in Figure~ \ref{fig:lc}.
\begin{figure*}
\plotone{CL001-15_phot.pdf}
\caption{De-trended K2 photometry of K2-287. Black points are individual 30-min cadence K2 data The transits of K2-287b\ are clearly seen. \label{fig:lc}}
\end{figure*}
\subsection{Spectroscopy}
We obtained 52 R=48000 spectra between March and July of 2018 using the FEROS spectrograph \citep{kaufer:99} mounted on the 2.2 MPG telescope in La Silla observatory. Each spectrum achieved a signal-to-noise ratio of $\approx90$ per spectral resolution element. The instrumental drift was determined via comparison with a simultaneous fiber illuminated with a ThAr+Ne lamp. We obtained additionally 25 R=115000 spectra between March and August of 2018 using the HARPS spectrograph \citep{mayor:2003}. Typical signal-to-noise ratio for these spectra ranged between 30 and 50 per spectral resolution element.
Both FEROS and HARPS data were processed with the \texttt{CERES}\ suite of echelle pipelines \citep{brahm:2017:ceres}, which produce radial velocities and bisector spans in addition to reduced spectra.
Radial velocities and bisector spans are presented in Table~\ref{tab:rvs} with their corresponding uncertainties, and the radial velocities are displayed as a function
of time in Figure~\ref{fig:rvstime}. No large amplitude variations were identified which could be associated with eclipsing binary scenarios for the K2-287\ system and no additional stellar components were evident in the spectra. The radial velocities present a time
correlated variation in phase with the photometric ephemeris, with an amplitude consistent with the one expected to be produced by a giant planet. We find no correlation between the radial velocities and the bisector spans (95\% confidence intervals for the Pearson coefficient are $[-0.19,0.21]$, see Figure~\ref{fig:bis}).
\begin{figure*}
\plotone{cl001-15_rvs-time.pdf}
\caption{Radial velocity (RV) curve for K2-287\ obtained with FEROS (red) and HARPS (black). The black line corresponds to the Keplerian model with the posterior parameters found in Section \ref{sec:glob}.\label{fig:rvstime}}
\end{figure*}
\begin{figure}
\plotone{rv-bs.pdf}
\caption{Radial velocity (RV) versus bisector span (BIS) scatter plot using data from our spectroscopic observations of K2-287. We find that the data is consistent with no correlation. \label{fig:bis}}
\end{figure}
\subsection{Ground-based photometry}
\label{ssec:ground}
On July 14 of 2018 we observed the primary transit of K2-287\ with the Chilean-Hungarian Automated Telescope (CHAT), installed at Las Campanas Observatory, Chile.
CHAT is a newly commissioned 0.7m telescope, built by members of the HATSouth \citep{bakos:2013} team, and dedicated to the follow-up of transiting exoplanets.
A more detailed account of the CHAT facility will be published at a future date (Jord\'an et al 2018, in prep\footnote{\url{https://www.exoplanetscience2.org/sites/default/files/submission-attachments/poster_aj.pdf}}).
Observations were obtained in the Sloan i' band and the adopted exposure time was of 53 s per image, resulting in a peak pixel
flux for K2-287\ of $\approx$ 45000 ADU during the whole sequence. The observations covered
a fraction of the bottom part of the transit and the egress (see Figure~\ref{fig:pht}). The same event was also monitored by one telescope of the Las Cumbres Observatory 1m network \citep{brown:2013:lcogt} at Cerro Tololo Inter-American Observatory, Chile. Observations were obtained with the Sinistro camera with 2mm of defocus in the Sloan i band. The adopted exposure time for the 88 observations taken was 60 s, and reduced images were obtained with the standard Las Cumbres Observatory pipeline (BANZAI pipeline).
The light curves for CHAT and the Las Cumbres 1m telescope were produced from the reduced images using a dedicated pipeline (Espinoza et al 2018, in prep).
The light curves were detrended by describing the systematic trends as a Gaussian Process with an exponential squared kernel depending on time, airmass and centroid position and whose parameters are estimated simultaneously with those of the transit. A photometric jitter term is also included; this parameter is passed on as a fixed parameter in the final global analysis that determines the planetary parameters (\S~\ref{sec:glob}). In more detail, the magnitude time series is modeled as
\begin{equation}
m_i = Z + x_1c_{1,i} + x_2 c_{i,2} + \delta_i + \epsilon_i
\end{equation}
\noindent where $Z$ is a zeropoint, $c_1$ and $c_2$ are comparison light curves, $x_1$ and $x_2$ are parameters weighting the light curves, $\delta$ is the transit model and $\epsilon$ is a Gaussian Process to model the noise. The subscript $i$ denotes evaluation at the time $t=t_i$ of the time series. For the Gaussian process, we assume a kernel given by
\begin{equation}
k_{ij} = A \exp\left[- \sum_m \alpha_m(x_{m,i} - x_{m,j})^2\right] + \sigma^2\delta_{ij}.
\end{equation}
The variables $x_{m}$ are normalized time ($m=0$), flux centroid in $x$ ($m = 1$) and flux centroid in $y$ ($m=2$); $\delta_{ij}$ is the Kronecker delta. The normalization is carried out by setting the mean to 0 and the variance to 1. The priors on the kernel hyper parameters were taken to be the same as the ones defined in \citet{gibson:2014}, the priors for the photometric jitter term $\sigma$ and $A$ were taken to be uniform in the logarithm between $0.01$ and $100$, with $\sigma$ and $A$ expressed in mmag.
In Figure~\ref{fig:chat-lcogt} we show the CHAT and LCOGT light curves with the weighted comparison stars subtracted along with the Gaussian process posterior model for the systematics.
\begin{figure*}
\plottwo{phot_CHAT.pdf}{phot_LCOGT.pdf}
\caption{Ground-based light curves for the July 14 2018 transit of K2-287b\ obtained with CHAT (left panel) and a LCOGT 1m telescope at CTIO (right panel). The red lines represent the posterior Gaussian process models for remaining systematics after subtracting the transit and weighted comparison stars and obtained as described in \S\ref{ssec:ground} \label{fig:chat-lcogt}}
\end{figure*}
\subsection{GAIA DR2}
Observations of K2-287\ by GAIA were reported in DR2 \citep{gaia, gaia:dr2}. From GAIA DR2, K2-287\ has a parallax of $6.29 \pm 0.05$ mas,
an effective temperature of $\ensuremath{T_{\rm eff}} = 4994 \pm 80$ K and a radius
of $\ensuremath{{\rm R}_{\star}} = 1.18 \pm 0.04 \,\, \ensuremath{{\rm R}_{\odot}}$. We used the observed
parallax for K2-287\ measured by GAIA for estimating a more precise value of \ensuremath{{\rm R}_{\star}}\ by combining it with the atmospheric parameters obtained from the spectra as described in \S~\ref{sec:ana}. We corrected the GAIA DR2 parallax for the systematic offset of -82 $\mu$as reported in \citet{stassun:2018}.
Two additional sources to K2-287\ are identified by GAIA inside the adopted K2 aperture ($\approx 12\arcsec$). However, both stars are too faint ($\Delta G > 7.8$ mag) to produce any significant effect on the planetary and stellar parameters found in \S~\ref{sec:ana}. The radial velocity variations in-phase with the transit signal, which are caused by K2-287, confirm that the transit is not caused by a blended stellar eclipsing binary on one of the companions.
\section{Analysis} \label{sec:ana}
\subsection{Stellar parameters}
As in previous K2CL discoveries we estimated the atmospheric parameters of the host star by comparing the co-added high resolution spectrum to a grid of synthetic models through the \texttt{ZASPE}\ code \citep{brahm:2016:zaspe}. In particular, for K2-287\ we used the co-added FEROS spectra, because they provide the higher signal-to-noise ratio spectra, and because the synthetic grid of models used by \texttt{ZASPE}\ was empirically calibrated using FEROS spectra of standard stars. Briefly, \texttt{ZASPE}\ performs an iterative search of the optimal model through $\chi^2$ minimization on the spectral zones that are most sensitive to changes in the atmospheric parameters. The models with specific values of atmospheric parameters are generated via tri-linear interpolation of a precomputed grid generated using the ATLAS9 models \citep{atlas9}. The interpolated model is then degraded to match the spectrograph resolution by convolving it with a Gaussian kernel that includes the instrumental resolution of the observed spectrum and an assumed macroturbulence value given by the relation presented in \citet{valenti:2005}. The spectrum is also convolved with a rotational kernel that depends on \ensuremath{v \sin{i}}, which is considered as a free parameter. The uncertainties in the estimated parameters are obtained from Monte Carlo simulations that consider that the principal source of error comes from the systematic mismatch between the optimal model and the data, which in turn arises from poorly constrained parameters of the atomic transitions and possible deviations from solar abundances.
We obtained the following stellar atmospheric parameters for K2-287: \ensuremath{T_{\rm eff}}=5695 $\pm$ 58 K, \ensuremath{\log{g}}=4.4 $\pm$ 0.15 dex, \ensuremath{{\rm [Fe/H]}}=0.20 $\pm$ 0.04 dex, and \ensuremath{v \sin{i}}=3.2 $\pm$ 0.2 km s$^{-1}$. The \ensuremath{T_{\rm eff}}\ value obtained with \texttt{ZASPE}\ is significantly different to that reported by GAIA DR2, but is consistent that of the K2 input catalog \citep{huber:2016}.
The stellar radius is computed from the GAIA parallax measurement, the available photometry, and the atmospheric parameters. As in \citet{k2-261}, we used a \texttt{BT-Settl-CIFIST} spectral energy distribution model \citep{baraffe:2015} with the atmospheric parameters derived with \texttt{ZASPE}\ to generate a set of synthetic magnitudes at the distance computed from the GAIA parallax. These magnitudes are compared to those presented in table \ref{tab:stprops} for a given value of \ensuremath{{\rm R}_{\star}}. We also consider an extinction coefficient A$_V$ in our modeling which affects the synthetic magnitudes by using the prescription of \citet{cardelli:89}. We explore the parameter space for \ensuremath{{\rm R}_{\star}}\ and A$_V$ using the \texttt{emcee} package \citet{emcee:2013}, using uniform priors in both parameters. We found that K2-287\ has a radius of $\ensuremath{{\rm R}_{\star}}=1.07 \pm 0.01$ \ensuremath{{\rm R}_{\odot}}\ and has a reddening of A$_V=0.56 \pm 0.03$ mag, which is consistent with what is reported by GAIA DR2.
Finally, the stellar mass and evolutionary stage for K2-287\ are obtained by comparing the estimation of \ensuremath{{\rm R}_{\star}}\ and the spectroscopic \ensuremath{T_{\rm eff}}\ with the predictions of the Yonsei-Yale evolutionary models \citep{yi:2001}. We use the interpolator provided with the isochrones to generate a model
with specific values of \ensuremath{{\rm M}_{\star}}, age, and \ensuremath{{\rm [Fe/H]}}, where \ensuremath{{\rm [Fe/H]}}\ is fixed to the value found in the spectroscopic analysis. We explore the parameter space for \ensuremath{{\rm M}_{\star}}\ and stellar age using the
\texttt{emcee} package, using uniform priors in both parameters. We find that the mass and age of K2-287\ are $\ensuremath{{\rm M}_{\star}} = 1.036 \pm 0.033$ $\ensuremath{{\rm M}_{\odot}}$ and 5.6 $\pm$ 1.6 Gyr (see Figure \ref{fig:iso}), similar to those of the Sun. The stellar parameters we adopted for K2-287\ are summarized in Table~\ref{tab:stprops}.
\begin{figure}
\plotone{CL001_15_txt_iso.pdf}
\caption{Yonsei-Yale isochrones for the metallicity of K2-287\ in the \ensuremath{T_{\rm eff}}--\ensuremath{{\rm R}_{\star}}\ plane. From left to right the isochrones correspond to 1, 3, 5, 7 and 9 Gyr. The position of K2-287\ is at the center of the blue shaded region, which marks the 3$\sigma$ confidence region for \ensuremath{T_{\rm eff}}\ and \ensuremath{{\rm R}_{\star}}.\label{fig:iso}}
\end{figure}
\begin{deluxetable*}{lrc}[b!]
\tablecaption{Stellar properties of K2-287\ \label{tab:stprops}}
\tablecolumns{3}
\tablewidth{0pt}
\tablehead{
\colhead{Parameter} &
\colhead{Value} &
\colhead{Reference} \\
}
\startdata
Names \dotfill & K2-287 & EPIC \\
& 2MASS J15321784-2221297 & 2MASS \\
& TYC 6196-185-1 & TYCHO \\
& WISE J153217.84-222129.9 & WISE \\
RA \dotfill (J2000) & 15h32m17.84s & EPIC\\
DEC \dotfill (J2000) & -22d21m29.74s & EPIC\\
pm$^{\rm RA}$ \hfill (mas yr$^{-1}$) & -4.59 $\pm$ 0.11& GAIA\\
pm$^{\rm DEC}$ \dotfill (mas yr$^{-1}$) & -17.899 $\pm$ 0.074 & GAIA\\
$\pi$ \dotfill (mas)& 6.288 $\pm$ 0.051 & GAIA \\
\hline
K$_p$ \dotfill (mag) & 11.058 & EPIC\\
B \dotfill (mag) & 12.009 $\pm$ 0.169 & APASS\\
g' \dotfill (mag) & 11.727 $\pm$ 0.010 & APASS\\
V \dotfill (mag) &11.410 $\pm$ 0.129 & APASS\\
r' \dotfill (mag) & 11.029 $\pm$ 0.010 & APASS\\
i' \dotfill (mag) & 10.772 $\pm$ 0.020 & APASS\\
J \dotfill (mag) & 9.677 $\pm$ 0.023 & 2MASS\\
H \dotfill (mag) & 9.283 $\pm$ 0.025 & 2MASS\\
K$_s$ \dotfill (mag) & 9.188 $\pm$ 0.021 & 2MASS\\
WISE1 \dotfill (mag) & 9.114 $\pm$ 0.022 & WISE\\
WISE2 \dotfill (mag) & 9.148 $\pm$ 0.019 & WISE\\
WISE3 \dotfill (mag) & 9.089 $\pm$ 0.034 & WISE\\
\hline
\ensuremath{T_{\rm eff}} \dotfill (K) & 5695 $\pm$ 58& \texttt{zaspe}\\
\ensuremath{\log{g}} \dotfill (dex) & 4.398 $\pm$ 0.015 & \texttt{zaspe}\\
\ensuremath{{\rm [Fe/H]}} \dotfill (dex) & +0.20 $\pm$ 0.04 & \texttt{zaspe}\\
\ensuremath{v \sin{i}} \dotfill (km s$^{-1}$) & 3.2 $\pm$ 0.2 & \texttt{zaspe}\\
\ensuremath{{\rm M}_{\star}} \dotfill (\ensuremath{{\rm M}_{\odot}}) & 1.056 $\pm$ 0.022 & YY + GAIA\\
\ensuremath{{\rm R}_{\star}} \dotfill (\ensuremath{{\rm R}_{\odot}}) & 1.07 $\pm$ 0.01 & GAIA + this work\\
Age \dotfill (Gyr) & 4.5 $\pm$ 1 & YY + GAIA\\
$\rho_\star$ \dotfill (g cm$^{-3}$) & 1.217 $\pm$ 0.045& YY + GAIA\\
\enddata
\end{deluxetable*}
\subsection{Global modeling}
\label{sec:glob}
\begin{figure*}
\plotone{CL001-15_transit.pdf}
\caption{The top panels show from left to right: the phase folded Kepler $K2$ photometry ($K_p$ band), the CHAT follow up photometry ($i$ band), and the LCO follow-up photometry ($i$ band) for K2-287. For the three cases, the model generated with the derived parameters of \texttt{EXONAILER} is plotted with a blue line. The bottom panels show the corresponding residuals. \label{fig:pht}}
\end{figure*}
\begin{figure*}
\plotone{CL001-15_rv.pdf}
\caption{The top panel presents the radial velocities for K2-287\ (filled circles) obtained with FEROS and HARPS as a function of the orbital phase. The RV model with the derived orbital parameters for K2-287b\ corresponds to the blue solid line. The bottom panel shows the residuals obtained
for these radial velocity measurements.
\label{fig:phr}}
\end{figure*}
In order to determine the orbital and transit parameters of the K2-287b\
system we performed a joint analysis of the detrended K2 photometry, the
follow-up photometry, and the radial velocities.
As in previous planet discoveries of the K2CL collaboration, we
used the \texttt{exonailer} code which is described in
detail in \citet{espinoza:2016:exo}. Briefly, we model the transit light
curves using the \texttt{batman} package \citep{kreidberg:2015} by taking
into account the effect on the transit shape produced by the long integration time of the
long-cadence K2 data \citep{ kipping:2010}.
To avoid systematic biases in the determination of the transit parameters
we considered the limb-darkening coefficients as additional free parameters in the transit modeling \citep{EJ:2015}, with the complexity of limb-darkening law chosen following the criteria presented in \citet{espinoza:2016:lds}. In our case, we select the quadratic limb-darkening law, whose coefficients were fit using the
uninformative sampling technique of \citet{Kipping:LDs}.
We also include a
photometric jitter parameter for the K2 data, which allow us to have an estimation of the level of stellar noise in the light curve. The radial velocities are modeled with
the \texttt{radvel} package \citep{fulton:2018}, where we considered systemic velocity and jitter factors for the data of each spectrograph.
We use the stellar density estimated in our stellar modeling as an extra ``data point" in our global fit as described in \citet{k2-232}. Briefly, there is a term in the likelihood of the form
\begin{eqnarray*}
p(\vec{y}_{\rho_*}|\theta ) = \frac{1}{\sqrt{2\pi \sigma^2_{\rho_*}}} \exp -\frac{(\rho_* - \rho_*^m)^2}
{2\sigma_{\rho_*}^2} ,
\end{eqnarray*}
where
\begin{eqnarray*}
\rho^m_* = \frac{3\pi}{GP^2}\left(\frac{a}{R_*}\right)^3
\end{eqnarray*}
by Newton's version of Kepler's law, and
$\rho_*$ and $\sigma_{\rho_*}$ are the mean stellar density and its standard-deviation,
respectively, derived from our stellar analysis. In essence, because the period $P$ is
tightly constrained by the observed periodic transits, this extra term puts a strong
constraint on $a/R_*$, which in turn helps to extract information about the eccentricity $e$ and argument of periastron $\omega$ from the duration of the transit. Resulting planet parameters are set out in Table~\ref{tab:plprops}, the best-fit orbit solution in Figures~\ref{fig:rvstime} and \ref{fig:phr} and the best-fit light curves in Figure~\ref{fig:pht}.
\begin{deluxetable*}{lrc}[b!]
\tablecaption{Planetary properties of the K2-287\ system. For the priors, $N(\mu,\sigma)$ stands for a normal distribution with mean $\mu$ and standard deviation $\sigma$, $U(a,b)$ stands for a uniform distribution between $a$ and $b$, and $J(a,b)$ stands for a Jeffrey's prior defined between $a$ and $b$.\label{tab:plprops}}
\tablecolumns{3}
\tablenum{2}
\tablewidth{0pt}
\tablehead{
\colhead{Parameter} &
\colhead{Prior} &
\colhead{Value} \\
}
\startdata
P (days) & $N(14.893,0.01)$ & 14.893291 $\pm$ 0.000025\\
T$_0$ (BJD)& $N(2458001.722,0.01)$& 2458001.72138 $\pm$ 0.00016 \\
$a$/R$_\star$ & $U(1,300)$ & 23.87$_{-0.31}^{+0.30}$ \\
\ensuremath{{\rm R_P}}/\ensuremath{{\rm R}_{\star}} & $U(0.001,0.5)$ & 0.08014$_{-0.00098}^{+0.00086}$ \\
$\sigma_w^{\rm K2}$ (ppm) & $J(10,50000)$ & 47.7$^{+0.54}_{0.54}$\\
q$_1^{\rm K2}$ & $U(0,1)$ & 0.32$^{+0.06}_{-0.05}$ \\
q$_2^{\rm K2}$ & $U(0,1)$& 0.57$^{+0.13}_{-0.11}$ \\
q$_1^{\rm CHAT}$ &$U(0,1)$ & 0.83$^{+0.12}_{-0.17}$ \\
q$_2^{\rm CHAT}$ &$U(0,1)$ & 0.15$^{+0.16}_{-0.11}$ \\
q$_1^{\rm LCO}$ & $U(0,1)$& 0.62$^{+0.20}_{-0.19}$ \\
q$_2^{\rm LCO}$ & $U(0,1)$& 0.08$^{+0.11}_{-0.06}$ \\
K (m s$^{-1}$) & $N(0,100)$& 28.8$^{+2.3}_{-2.2}$ \\
$e$ & $U(0,1)$ & 0.478$^{+0.025}_{-0.026}$ \\
$i$ (deg) & $U(0,90)$ & 88.13$^{+0.1}_{-0.08}$\\
$\omega$ (deg) & $U(0,360)$ & 10.1$^{+4.6}_{-4.2}$ \\
$\gamma_{\rm FEROS}$ (m s$^{-1}$)& $N(32963.2,0.1)$& 32930.41$^{+0.10}_{-0.10}$ \\
$\gamma_{\rm HARPS}$ (m s$^{-1}$)& $N(32930.4,0.1)$ & 32963.19$^{+0.10}_{-0.10}$ \\
$\sigma_{\rm FEROS}$ (m s$^{-1}$)& $J(0.1,100)$ & 16.0$^{+2.1}_{-1.8}$ \\
$\sigma_{\rm HARPS}$ (m s$^{-1}$) & $J(0.1,100)$ & 4.8$^{+1.8}_{-1.6}$ \\
\hline
\ensuremath{{\rm M_P}}\ (\ensuremath{{\rm M_{J}}})& & 0.315 $\pm$ 0.027\\
\ensuremath{{\rm R_P}}\ (\ensuremath{{\rm R_J}})& & 0.847 $\pm$ 0.013\\
$a$ (AU) & & 0.1206$_{-0.0008}^{+0.0008}$\\
\ensuremath{T_{\rm eq}}\tablenotemark{a} (K) & & 804$_{-7}^{+8}$\\
\enddata
\tablenotetext{a}{Time-averaged equilibrium temperature computed according to equation~16 of \citet{mendez:2017}}
\end{deluxetable*}
\section{Discussion} \label{sec:dis}
By combining data from the Kepler K2 mission and ground based photometry and spectroscopy,
we have confirmed the planetary nature of a $P=14.9$ d candidate around the $V=11.4$ mag G-type
star K2-287. We found that the physical parameters of K2-287b\ (\ensuremath{{\rm M_P}} = \ensuremath{0.317 \pm 0.026 }\ \ensuremath{{\rm M_{J}}}, \ensuremath{{\rm R_P}} = \ensuremath{0.833 \pm 0.013 }\ \ensuremath{{\rm R_J}}) are consistent to those of Saturn. The non-inflated structure of K2-287b\ is
expected given its relatively low time-averaged equilibrium temperature of \ensuremath{T_{\rm eq}} = 808 $\pm$ 8 K.
In Figure \ref{fig:mr} the mass and radius of K2-287b\ are compared to those for the full population of transiting planets with parameters measured to a precision of 20\% or better.
Two other transiting planets, orbiting fainter stars, that share similar structural properties to K2-287b\ are HAT-P-38b \citep{sato:2012} and HATS-20b \citep{bhatti:2016}, which have equilibrium temperatures that are higher but relatively close to the $\ensuremath{T_{\rm eq}} \approx 1000$ K limit below which the inflation mechanism of hot Jupiters does not play a significant role \citep{kovacs:2010,demory:2011}.
By using the simple planet structural models of \citet{fortney:2007} we find that the observed properties of K2-287b\ are consistent with having a solid core of $M_c = 31 \pm 4 M_{\oplus}$. However, models that consider the presence of solid material in the envelope of the planet are required to
obtain a more reliable estimate for the heavy element content of K2-287b\ \citep[e.g.,][]{thorngren:2016}.
\begin{figure*}
\plotone{m-r.pdf}
\caption{Mass-Radius diagram for the full population of transiting planets with both parameters measured to at least 20\% precision. The points are color-coded by equilibrium temperature. K2-287b\ is the object in the plot that has error bars and is indicated by the arrow. The dashed
gray lines correspond to iso-density curves of 0.3, 3, and 30 g cm$^{-3}$, while the solid
line represents the prediction of the \citet{fortney:2007} structural model with a central core
mass of 10 M$_{\oplus}$.
Due to its relatively low equilibrium temperature, K2-287b\ lies in a sparsely populated region of the parameter space of moderately compact giant planets. \label{fig:mr}}
\end{figure*}
\begin{figure*}
\plotone{p-j.pdf}
\caption{Population of well characterized giant planets having \ensuremath{{\rm R_P}} $>$ 0.4 \ensuremath{{\rm R_J}}\ in the orbital period -- J magnitude plane. K2-287b\ is inside a black square. The size of the points represent the eccentricity of the orbit, while the color indicates the discovery method/mission (blue: ground-based photometry, yellow: RV planets, orange: CoRoT, red: \textit{Kepler}, green: \textit{Kepler} K2). The \textit{Kepler} K2 mission has been the most effective source for discovering transiting bright (J $<$ 11) warm (P $>$ 10 d) giant planets. \label{fig:pj}}
\end{figure*}
The numerous radial velocity measurements obtained for the K2-287\ system allow us to constrain the eccentricity of the planet to be $e=0.478 \pm 0.025$. Even though K2-287b\ is among the most eccentric extrasolar planets to have a period shorter than 50 days, its periastron distance is not
small enough to cause a significant migration by tidal interactions throughout the main sequence lifetime of the host star. Specifically, by using the equations of \citet{jackson:2009}, we find that in the absence of external sources of gravitational interaction, K2-287b\ should have possessed an eccentricity of $e\approx0.65$ and a semi-major
axis of $a\approx0.15$ AU when the system was 0.1 Gyr old. Under the same assumptions, we expect that K2-287b\ would be engulfed by its host star at an age of $\approx$12 Gyr before being able to reach full circularization at a distance of $a\approx0.1$ AU.
These orbital properties for K2-287b\ and those of the majority of eccentric warm giants are not easy to explain. If K2-287b\ was formed \textit{in situ} \citep{huang:2016} at 0.15 AU or migrated to this position via interactions with the protoplanetary disc \citep{lin:1997}, its eccentricity could have been excited by the influence of another massive object in the system after disc dispersal. However, planet-planet scattering \citep{ford:2008} at these close-in orbits generally produces planet collisions rather than eccentricity excitation \citep{petrovich:2014}.
An alternative proposition for the existence of these eccentric systems is that they
are being subject to secular gravitational interactions produced by another distant
planet or star in the system \citep{rasio:1996}, with the planet experiencing long term
cyclic variations in its eccentricity and spin orbit angle. In this scenario, the planet
migrates by tidal interactions only during the high eccentricity stages, but it is usually
found with moderate eccentricities. Further observations on the K2-287\ system could help
support this mechanism as the responsible for its relatively high eccentricity, particularly given that \citet{petrovich:2016} concludes that high-eccentricity migration excited by an outer planetary companion can account for most of the warm giants with $e>0.4$. Specifically, long term radial velocity monitoring and the search for transit timing variations could be used to detect the relatively close companions to migrating warm Jupiters proposed by \citet{dong:2014}. Future astrometric searches of companions with GAIA could also be used to find companions and infer the predicted mutual inclination between both orbits, which are predicted to be high \citet{anderson:2017}.
Finally, it is worth noting that an important fraction of the transiting warm giants amenable for detailed characterization ($J<11$ mag) have been discovered in the last couple of years thanks to the K2 mission (see Figure~\ref{fig:pj}). The combination of relatively long observing campaigns per field, and the increased number of fields monitored, have allowed the discovery and dynamical characterization of several warm giant planets with data from the K2 mission \citep[see Figure~\ref{fig:pj}, ][]{k2-24,k2-99,barragan:2017,shporer:2017,k2-232,k2-234,k2-261,k2-261b}.
While not particularly designed to discover warm giants, the TESS mission \citep{tess} is expected to discover $\approx$ 120 additional warm giants with $\ensuremath{{\rm R_P}} > 4R_\oplus$ and an incident flux $F < 150 F_\oplus$, where $F_\oplus$ is the incident flux at Earth, around $J\lesssim 11$ mag stars \citep{barclay:2018}. With such population at hand, it will be possible to compare the distributions of eccentricities and obliquities to predictions from different migration mechanisms \citep[e.g. ][]{petrovich:2016} in order to establish a clearer picture about how eccentric warm giant planets originate.
\acknowledgements
A.J.\ acknowledges support from FONDECYT project 1171208, CONICYT project BASAL AFB-170002, and by the Ministry for the Economy, Development, and Tourism's Programa Iniciativa Cient\'{i}fica Milenio through grant IC\,120009, awarded to the Millennium Institute of Astrophysics (MAS). R.B.\ acknowledges support from FONDECYT Post-doctoral Fellowship Project 3180246, and from the Millennium Institute of Astrophysics (MAS).
M.R.D.\ acknowledges support by CONICYT-PFCHA/Doctorado Nacional 21140646, Chile. A.Z.\ acknowledges support by CONICYT-PFCHA/Doctorado
Nacional 21170536, Chile.
J.S.J.\ acknowledges support by FONDECYT project 1161218 and
partial support by CONICYT project BASAL AFB-170002.
This paper includes data collected by the K2 mission. Funding for the K2 mission is provided by the NASA Science Mission directorate.
This work has made use of data from
the European Space Agency (ESA) mission Gaia (\url{https:
//www.cosmos.esa.int/gaia}), processed by the Gaia Data
Processing and Analysis Consortium (DPAC, \url{ https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding for
the DPAC has been provided by national institutions, in particular
the institutions participating in the Gaia Multilateral
Agreement.
Based on observations collected at the European
Organisation for Astronomical Research in the Southern
Hemisphere under ESO programmes 0101.C-0497, 0101.C-0407, 0101.C-0510.
\vspace{5mm}
\facilities{CHAT~0.7m, LCOGT~1m, MPG~2.2m, ESO~3.6m, \textit{Kepler}, GAIA, APASS, 2MASS, WISE}
\software{EXO-NAILER \citep{espinoza:2016:exo},
CERES \citep{brahm:2017:ceres,jordan:2014},
ZASPE \citep{brahm:2016:zaspe,brahm:2015},
radvel \citep{fulton:2018}
}
|
\section{Introduction}
Causality principle lies the foundations of all the Science \cite{pearl}. That a cause precedes its effect is strictly linked, in Special Relativity, to the impossibility of superluminal communications (no-signaling condition). In Quantum Mechanics (QM) this principle is ensured by the kinematic structure of the theory and by the linearity of the dynamics, a circumstance known as `peaceful coexistence between QM and Special Relativity'. As a consequence, entanglement between two space-like separated regions cannot be exploited for superluminal communication. This consideration led some authors to include the no-signaling condition in the basic set of axioms of QM \cite{Simon}. Recently a further generalization has been introduced as well, known as the principle of information causality: it states that there cannot be more information available than was transmitted \cite{paw}.\\
Notwithstanding, the
issue of a possible modification of non-relativistic QM in a nonlinear
sense has a long-standing history, starting from Wigner's suggestion \cit
{Wigner}. The motivation for this search is twofold. On one side, there was
a need to understand whether linearity is a foundational aspect of QM or
should be considered just an approximation, though a very good one. On the
other side, the measurement problem, and the seemingly subjective border
separating quantum and classical domain, led people to explore nonlinear
modifications of the Schr\"{o}dinger equation. This subject, however, should
not be confused with the issue of the nonlinear versions of the Schr\"{o}dinger equation considered on completely different grounds. In fact there
exists a vast literature on what is called `nonlinear Schr\"{o}dinger
equation', as the famous Gross-Pitaevskii equation describing Bose-Einstein
condensates, which has to be considered as a mean field limit in the
framework of the standard theory and, at the best, only a test-case in
nonlinear dynamics.
Despite some serious efforts to introduce nonlinear corrections \cit
{WeinbergAnnals} assuming that the observables include all the usual linear
Hermitian operators, stringent limitations on the allowed nonlinearities
became soon evident on the basis of the no-signaling condition
\cite{Gisin, Polchinski}. Nowadays this latter is widely recognized to be a very powerful condition. In fact, for example, it can be used alone to find the
maximum fidelity in the copy of quantum states \cite{BuzekHillery}.
Interestingly, as Polchinski showed \cite{Polchinski}, while one can
construct a theory free from causality problems by means of the stringent
requirement that the Hamiltonian of each subsystem depends (in the formalism
of Ref.\cite{WeinbergAnnals}) only on its density matrix, unusual (Everett)
communications take place among the different branches of the wave function.
That is, the state of a system at any given instant of time and in a
specific world (or mind!) depends also on what happened or \textit{is}
happening in some other worlds (minds). Incidentally, this circumstance
could be used, in principle, to devise an experiment to test the existence
of parallel worlds, independently of a detailed knowledge of human brain
perception's physiology (required, for example, in the conceptual experiment
discussed by Deutsch \cite{Deutsch}).
We note that these foundational problems of QM are not isolated, but appear
in close connection with the still elusive theory of quantum gravity. In
particular, the way in which gravitational fields are produced by quantum
matter is still controversial. Indeed it is not even clear if gravity has to
be quantized at all, in which case the existence of a quantum superposition
of space-times is implied (see \cite{Penrose}
for some related conceptual problems); or gravity is intrinsically
classical, and should be treated consequently. A natural candidate for this
latter hypothesis is the semi-classical gravity, which (at least in its
simplest form) prescribes to take as the source of the field (appearing on
the right hand side of the Einstein field equation) the expectation of the
quantum energy-momentum tensor. Its Newtonian limit, the famous Newton-Sch
\"{o}dinger equation, turns out to be a non-linear quantum mechanical
equation \cite{Bahrami, Giulini}. While showing some interesting features,
such as the existence of self-localized stationary solutions \cite{Moroz},
this equation is plagued by the causality problems mentioned above \cit
{Bahrami}.
To give a further chance to semiclassical gravity, various attempts have
been made, inspired and connected with the phenomenological collapse models,
to add \textit{ad hoc} stochastic terms to the Newton-Schr\"{o}dinger term
in such a way that non-classical (delocalized) states become unstable,
rapidly collapsing to well-localized states \cite{Bahrami,Pearle&Squires,Nimm}.
In the present work we take a somehow different viewpoint, by looking at the
Newton-Schr\"{o}dinger equation as a mean field approximation of more
fundamental quantum mechanical equations, technically in a quite similar way
in which Gross-Pitaewski nonlinear equation is derived from standard QM. In
particular, a single-particle \textit{N-S} equation is regarded as the mean
field approximation of an equation of $N$ identical copies of the particle,
gravitationally interacting among them, as $N$ tends to infinity. Indeed, it
was showed that the limit of $N\rightarrow \infty $ gives back the
Newton-Schr\"{o}dinger equation \cite{DeFil+DeFilMaim}. More specifically,
within this model, physical observations are referred to only one of these
particles, while the remaining ones are considered to belong to an \textit
hidden} system. The $N$ particles interact uniquely via gravitational
interaction, while the global state of the system is constrained to be
symmetric with respect to particles state permutations. (Incidentally, it is
interesting in this respect the observation by Adler pointing to an
interpretational problem with particles self-interaction within the Hartee
approximation of the Newton-Schr\"{o}dinger equation \cite{Adler}). The evolved
physical state is obtained by tracing out the unobservable degrees of
freedom of the $N-1$ particles (see Appendix \ref{general} for a
self-contained second-quantization general formulation of the model).
The model, known as Nonunitary Newtonian Gravity (NNG, from now on), has
been studied in some detail, in particular the limit $N=2$, showing the
interesting property of (entropic) dynamical self-localization for masses
above the (sharp) threshold of $10^{11}$ proton masses ($m_{p}$ from now on), with
precise signatures susceptible to future experimental tests \cit
{DeFMaimPRD}, \cite{DeFilMaimRob}, \cite{Nostro}.
Since generally pure states evolve into mixed states even for isolated
systems \cite{DeFMaimPRD},\cite{NostroTwo}, the fundamental description of
physical reality cannot be associated to the wave function. Instead, density
matrix has to be considered as a fundamental description of the Nature (see,
concerning this last point, the latest approach of S. Weinberg to the
foundations of QM \cite{WeinStateVec}).
While it can be inferred directly by the well-posedness of the model that it
is free from causality problems, in the following we will show this fact
explicitly, unfolding its physical basis. In particular we will elucidate
how the basic mechanism operates and guarantees the no-signaling condition.
It turns out that this mechanism is strictly related to the dynamical
mechanism of state reduction naturally embedded within the model and
completely suppressed in the limit $N\rightarrow \infty $ (see Appendix \re
{reduction} for an explicit demonstration). This is the same effect
operating in the stochastic versions of Newton-Schr\"{o}dinger, but here it
emerges naturally and is not an \textit{ad hoc} prescription, while state
reduction appears to be a necessary built-in consequence of the no-signaling
condition.
The plan of the present work is as follows. In Section 2, we present in the
simplest possible setting the Newton-Schr\"{o}dinger limit, helping to
elucidate its physical rationale. Then, in Section 3 we analyze a specific EPR situation,
showing how in the simplest $N=2$ case the oddities of causality violations are effectively cured. In Section 4, we show how a communication among
Everett worlds is instead possible, though strongly limited by the model
dynamics itself. Some concluding remarks, in particular on a variant of the
Everett Many Worlds interpretation suggested by our results, end the paper.
\section{Newton-Schr\"{o}dinger limit}
Let's begin by seeing how a superposition of states looks like within the
model. Consider in the ordinary QM setting a superposition of two states of
a body corresponding to two different locations of its CM. Its
representation in the theory goes as follows. Given $p,\,q\,>0$, with $p+q=1
, the CM state of the system and of its hidden counterparts, \textit{i.e.}
the \textit{metastate}, is
\begin{equation}
\Vert \Psi ^{(N)}\rangle \rangle =\otimes _{i=1}^{N}(\sqrt{p}|x^{i}\rangle
\sqrt{q}|y^{i}\rangle )=\sum_{k=0}^{N}\sqrt{\left(
\begin{array}{c}
N \\
\end{array
\right) }p^{\frac{k}{2}}q^{\frac{N-k}{2}}|x\rangle ^{\otimes k}|y\rangle
^{\otimes (N-k)}, \label{metastat}
\end{equation
where
\begin{equation*}
|x\rangle ^{\otimes k}|y\rangle ^{\otimes (N-k)}\equiv \frac{1}{\sqrt{N_{P}}
\sum_{P}P\Vert x^{1}\dots x^{k}y^{k+1}\dots y^{N}\rangle \rangle .
\end{equation*
(A proof of equation (\ref{metastat}) can be found in Appendix A).
Now, starting from the normalization
\begin{equation*}
\langle \langle \Psi ^{(N)}\Vert \Psi ^{(N)}\rangle \rangle
=\sum_{k=0}^{N}\left(
\begin{array}{c}
N \\
\end{array
\right) p^{k}q^{N-k},
\end{equation*
for $N$ sufficiently large the binomial term can be approximated by a
gaussian
\begin{equation*}
B_{p}(k,N)\sim \aleph (Np,N(1-p)),
\end{equation*
and, putting $\alpha =\frac{k}{N}$, we get
\begin{equation}
\langle \langle \Psi ^{(N)}\Vert \Psi ^{(N)}\rangle \rangle =\int_{0}^{1
\frac{1}{\sqrt{2\pi p(1-p)/N}}e^{-\frac{(\alpha -p)^{2}}{2p(1-p)/N}}d\alpha
\xrightarrow[N\rightarrow \infty]{}\int_{0}^{1}\delta (\alpha -p)\,d\alpha .
\end{equation
Then, in the limit $N\rightarrow \infty $ the only surviving contribution of
the original sum is that with $k=Np$, \emph{i.e.} the huge superposition of
states reduces to the single (central) term:
\begin{equation*}
\Vert \Psi ^{(N)}\rangle \rangle \xrightarrow[N\rightarrow \infty]{}\Vert
\Psi ^{\infty }\rangle \rangle \propto \lim_{N\rightarrow \infty }|x\rangle
^{\bigotimes Np}|y\rangle ^{\bigotimes N(1-p)}.
\end{equation*
In other words, the superposition reduces to the state in which a fraction
\textit{p} of meta-matter is displaced to \textit{x} position and the
remaining part to position \textit{y}.
\section{Analysis of an ERP situation}
Following the simple argument given in Ref.\cite{Pearle&Squires}, let's
consider a sphere of matter of ordinary density of radius $R$, and let the
state $|Z\rangle $ denote the sphere with center on the z-axis at $z=Z,$ and
suppose the state of the sphere is, within the ordinary QM setting, $(|+Z\rangle
+|-Z\rangle )/\sqrt{2}.$ A probe mass moving along the $x-$ axis will,
according to the non-relativistic quantum theory of gravity, become
entangled with the state of the sphere, resulting in the state vector
(|+Z\rangle |up\rangle +|-Z\rangle |down\rangle )/\sqrt{2}$, where
|up\rangle $ ($|down\rangle $) means that the probe is deflected in the
positive (negative) $z-$direction. According to semi-classical gravity, the
probe mass should be undeflected. This was experimentally tested under the
hypothesis that QM continues to hold in the macroscopic domain \cit
{PageGeilker}, with the (not unexpected) result that the mass is deflected.
Indeed, semiclassical gravity implies precisely a modification of QM in this
domain, so the hypothesis is self-contradictory.
Instead, a serious theoretical objection to semiclassical gravity is, as
said above, that it allows superluminal communication and then causality violation. To see this, consider
the entangled state $(|+Z\rangle |0\rangle +|-Z\rangle |1\rangle )/\sqrt{2}$,
where the state $|0\rangle $ and $|1\rangle $ denote orthogonal states of a
two state system (qbit) which is at a large distance from the sphere, but
close to a \textquotedblleft sender\textquotedblright. A probe mass is then
used as before.
If the sender chooses not to measure the system, the ``receiver'', who is
close to the sphere and uses the probe mass as described above, finds it
undeflected. If, on the other hand, the sender chooses to measure the
system, thereby finding it to be in the state $\vert 0\rangle $ or $\vert
1\rangle $, the sphere will immediately be in the state $\vert +Z\rangle $
or $\vert -Z\rangle $ respectively. Then the receiver will be able to see
this because the probe mass will now be deflected up or down.
Let's translate now this conceptual experiment into our model of replicas.
The entangled state of the \textit{sphere+q-bit} is analogous to Eq. (\re
{metastat}),
\begin{eqnarray}\label{metastatq}
\|\Psi^{(N)}\rangle\rangle &=&\bigotimes _{i=1}^N\frac{1}{\sqrt{2}}(\rvert+Z^{i}\rangle\vert 0^{i}\rangle+\rvert-Z^{i}\rangle\vert 1^{i}\rangle) =\notag \\
& = &\sum_{k=0}^{N}\sqrt{\left(\begin{array}{c}N \\k\end{array}\right)}\biggl(\frac{1}{2}\biggr)^{N/2}|+Z;0\rangle^{\otimes k} |-Z;1\rangle^{\otimes (N-k)},
\end{eqnarray}
where the tensor product terms are defined in the same way as in Eq. (\re
{metastat}).
Consider now the probe particle (supposed to have a mass much more smaller
than the lump) shut just over the superposition. Including the probe in the
system's description, we have that an initially unentangled (global)
state of the probe and of the composite \textit{sphere+q-bit} system evolves
towards an entangled one,
\begin{eqnarray}\label{product}
\Vert \varphi_I^{(N)}\rangle\rangle\bigotimes \Vert \Psi^{(N)}\rangle\rangle \xrightarrow[time\mbox{ }evolution]{} \sum_{k=0}^{N} \sqrt{\left(\begin{array}{c}N \\k\end{array}\right)}\biggl(\frac{1}{2}\biggr)^{N/2}\Vert \varphi_k^{(N)}\rangle\rangle\vert +Z;0\rangle^{\bigotimes k}\vert -Z;1\rangle^{\bigotimes (N-k)},
\end{eqnarray}
where $\Vert \varphi _{I}^{(N)}\rangle \rangle =\bigotimes_{m=1}^{N}|\varphi
_{I}^{m}\rangle $ is the product of identical copies of the probe's state
(it is assumed that probe's mass is so small that gravity-induced internal
entanglement among copies is irrelevant; see the example of the particle in
Earth's gravitational field in Ref.\cite{DeFMaimPRD}). The meaning of the
index $k$ is clarified in Fig. 1.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.5]{grafico.eps}\\
\caption{\small{Seen from above, deflection branches of the probe particle passing over the massive sphere in superposition of Eq. (\ref{product}).}} \label{fig:deflection}
\end{figure}
For simplicity let's consider the simplest case with $N=2$, for which we
have only three branches in the superposition (downward, central and upward
trajectories). Generalization to a generic $N$ is straightforward.
First of all, observe that in order to detect a whatever deflection in the
probe particle's trajectory, the size of the wave packets describing the
particle's states in the superposition should be smaller than the deflection
itself. Moreover, their spreading along the path have to be taken into
account before the position measurement of the particle's position along the
\textit{z}-axis. Denoting with $v_{x}$ the velocity of the particle, $m$ its
mass, and $M$ the mass of the sphere, the time during which the sphere's
gravitational attraction is effective is of the order of $T\sim R/v_{x}.$
Then it should be
\begin{equation*}
\frac{GMT}{R^{2}}\gtrsim \dfrac{\hbar }{mZ}.
\end{equation*}
\begin{figure}[tbph]
\centering
\includegraphics[scale=1]{Interdiction_Zone.eps}
\caption{{\protect\small {White area denotes the `interdiction' zone in the
diagram (\textit{M, time}), for which superluminal communications would be
possible. $T$ and $\tau_g$, representing the upper and lower lines respectively,
are defined within the main text.}}}
\label{fig:interdition}
\end{figure}
On the other hand, a peculiar dynamical signature of the models (studied in
Ref.\cite{DeFMaimPRD}; see also the calculation below in the limiting case
N=2$) is that, for a lump of mass \textit{M} above a threshold of $10^{11} m_{p}$, a superposition of the CM wave packets separated by a
distance much smaller than the body's size undergoes a rapid state reduction
after a characteristic time $\tau _{g}\sim \hbar G^{-1}M^{-5/3}\rho ^{-1/3}$
(see Appendix \ref{reduction} for a self-contained derivation), leading from
the initial superposition to an ensemble of localized states (see footnote~\ref{nota} on page~\pageref{nota}). The additional condition for the detection of a significant deflection of
the probe is that the measurement time should be smaller than this reduction
time, which together with the constraint above gives the set of conditions
that must be satisfied simultaneously:
\begin{equation}
\left\{
\begin{array}{ll}
T & \gtrsim \biggl(\dfrac{\hbar }{m}\biggr)^{1/3}\biggl[\frac{1}{2}\biggl
\dfrac{4}{3}\pi \rho \biggr)^{\frac{2}{3}}GM^{\frac{1}{3}}\biggr]^{-\frac{2}
3}} \\
T & \ll \tau _{g}=\hbar G^{-1}M^{-5/3}\rho ^{-1/3}.
\end{array
\right. \label{conditions}
\end{equation
As said previously, the generalization to a generic \textit{N} is
straightforward. In other words, our conclusion can be stated by saying that
gravity-induced state reduction is so rapid to forbid a sufficiently long
measurement, which would otherwise permit a deflection discrimination with
respect to the spreading of the wave packet. These conditions have been
depicted in Fig. \ref{fig:interdition}, from which it is clear that
they cannot be satisfied together.
In other words the `peaceful coexistence' between (deterministic) QM and
special relativity indeed \textit{imply} linearity (see for example \cit
{Simon}), unless a certain amount of predictability loss is present.
\section{Everett phone}
As mentioned above, one of the drawbacks of the introduction of
nonlinearities in a theory with a properly restricted observational algebra
is the appearance of bizarre communications among Everett branches of the
wave function. This circumstance appears also within the nonlinear model
described here. At variance with the other nonlinear modifications of the
theory, in this case predictability loss strongly limits such communication
possibilities as far as macroscopic (mass) cat states are involved, as we
are going to show. Let's start by looking at the interaction among branches;
then we turn to the possibility to use these interactions to construct, in
principle, an Everett phone.
The existence of effective interactions among Everett branches can be
immediately inferred by considering the following physical situation.
Assume an initial superposition of localized states of an isolated lump of
matter (whose mass we will assume just above the localization mass
threshold). Branches or worlds independence means that, following the
evolution of each of the states independently from each other and, then,
forming the final superposition of the evolved states at some later time, is
equivalent to considering the time evolution of the global state from the
beginning. This is clearly not true within the present model, since the
state reduction dynamics rely precisely on the existence of all the other
states in the superposition, while the exact reduction time depends also on
the spatial distribution of the localized states. (For example, a long
cigar-shaped matter distribution would be reduced more slowly in comparison
with a ball-like distribution occupying the same volume.)
Said in other words, the evolution of a superposition of localized states
does not coincide with the superposition of the evolved localized states.
This is a statement of the existence of interactions among the branches.
Incidentally, we stress that Everett branches should not be confused with
the copy/copies of the physical system, which represent just a useful and
easy way to formulate the model.
Let's now illustrate, following Polchinski's proof of principle, how an
Everett phone could be constructed on the basis\ of the theory. Consider a
two-level system $S$ initially in the state $|\Phi _{0}\rangle $ and a
system $M$, initially in the state $|M_{0}\rangle $, that we can imagine
composed in general of \textit{environment} + \textit{recording device} +
(eventually) \textit{an observer's mind}.\newline
Measuring the q-bit state in the $\{|1\rangle ,|0\rangle \}$ basis, we
get
\begin{equation}
|\Phi _{0}\rangle \otimes |M_{0}\rangle \longrightarrow \frac{1}{\sqrt{2}
\biggl(|1\rangle |M_{1}\rangle +|0\rangle |M_{2}\rangle \biggr).
\label{statoin}
\end{equation
Calling $\sigma _{3}$ the third Pauli matrix, if we obtain $1$, it is $\rho
^{+}=\frac{1+\sigma _{3}}{4}$; otherwise $\rho ^{-}=\frac{1-\sigma _{3}}{4}$.
In this second case, the observer (or the recording automatic system) can
follow one of two actions: (\textit{a}) nothing; (\textit{b}) rotates the
q-bit state into the $\widehat{\mathbf{i}}$ direction. We analyze these cases
in the NNG representation, starting from the second case.
\begin{itemize}
\item \textit{Case} (\textit{b}).
\end{itemize}
We write our (meta-)state $||\psi(t)\rangle\rangle^{(I)}$ as
\begin{eqnarray}
&&\frac{1}{2}\biggl(|1M_{1}\rangle +|\mathbf{i}M_{2}\rangle \biggr)\otimes
\biggl(|\widetilde{1M_{1}}\rangle +|\widetilde{\mathbf{i}M_{2}}\rangle
\biggr) \notag \label{metastatobt} \\
&=&\frac{1}{2}\biggl(|1\rangle |\widetilde{1}\rangle |M_{1}\rangle
\widetilde{M_{1}}\rangle +|1M_{1}\rangle |\widetilde{\mathbf{i}M_{2}}\rangle
+|\mathbf{i}\rangle |\widetilde{1}\rangle |M_{2}\rangle |\widetilde{M_{1}
\rangle +|\mathbf{i}\rangle |\widetilde{\mathbf{i}}\rangle |M_{2}\rangle
\widetilde{M_{2}}\rangle \biggr)\notag \\
&\xrightarrow[time\mbox{ }evolution]{} &\frac{1}{2}\biggl(|1M_{1};\widetilde{1M_{1}}\rangle
_{t}+|1M_{1};\widetilde{\mathbf{i}M_{2}}\rangle _{t}\biggr).
\end{eqnarray
(Remember that $|\mathbf{i}\rangle =\frac{1}{\sqrt{2}}\bigl(|1\rangle
+i|0\rangle \bigr)$.) Suppose that, at the beginning, state $|1\rangle $ was
measured; in (\ref{metastatobt}), that part of\ the wave function not living
in the branch of $M_{1}$ has to be disregarded. Writing the most general
time-evolved relevant states in the form
\begin{equation*}
|1M_{1};\widetilde{1M_{1}}\rangle _{t}=\sum_{(i,\mu )(\tilde{\imath},\tilde
\mu})}c_{(i,\mu )(\tilde{\imath},\tilde{\mu})}^{(1)}|i\mu ;\tilde{\imath
\tilde{\mu}\rangle _{0}
\end{equation*
and
\begin{equation*}
|1M_{1};\widetilde{\widehat{\mathbf{i}}M_{2}}\rangle _{t}=\sum_{(i,\mu )
\tilde{\imath},\tilde{\mu})}c_{(i,\mu )(\tilde{\imath},\tilde{\mu
)}^{(2)}|i\mu ;\tilde{\imath}\tilde{\mu}\rangle _{0}\ ,
\end{equation*
the general expression for the evolved q-bit state is ($\widetilde{S}$ is the
hidden q-bit)
\begin{eqnarray}
\rho (t) &=&\overset{}{\underset{M,\ \widetilde{S}}{Tr}}\biggl[\Vert \psi
(t)\rangle \rangle ^{(I)}\langle \langle \psi (t)\Vert \biggr]=\sum_{\alpha
,(\tilde{j},\tilde{\alpha})}\langle \alpha ,\tilde{j},\tilde{\alpha}\Vert
\psi (t)\rangle \rangle ^{(I)}{}\langle \langle \psi (t)\Vert \alpha ,\tilde
j},\tilde{\alpha}\rangle = \notag \label{traccia} \\
&=&\frac{1}{4}\sum_{i,i^{\prime }}\biggl[\sum_{\alpha ,(\tilde{j},\tilde
\alpha})}\biggl(c_{(i,\alpha )(\tilde{j},\tilde{\alpha})}^{(1)}c_{(i^{\prime
},\alpha )(\tilde{j},\tilde{\alpha})}^{(1)\ast }+c_{(i,\alpha )(\tilde{j}
\tilde{\alpha})}^{(2)}c_{(i^{\prime },\alpha )(\tilde{j},\tilde{\alpha
)}^{(2)\ast }+c_{(i,\alpha )(\tilde{j},\tilde{\alpha})}^{(1)}c_{(i^{\prime
},\alpha )(\tilde{j},\tilde{\alpha})}^{(2)\ast }+c_{(i,\alpha )(\tilde{j}
\tilde{\alpha})}^{(2)}c_{(i^{\prime },\alpha )(\tilde{j},\tilde{\alpha
)}^{(1)\ast }\biggr)\biggr]|i\rangle \langle i^{\prime }|= \notag \\
&\equiv &\dfrac{1}{4}\sum_{i,i^{\prime }}C_{i,i^{\prime }}|i\rangle \langle
i^{\prime }|\ ,
\end{eqnarray}
from which
\begin{equation}
\langle \sigma _{3}\rangle =Tr\biggl[\sigma _{3}\rho _{s}^{(t)}\biggr
=C_{1,1}-C_{2,2}. \label{spinmedio}
\end{equation}
\begin{itemize}
\item \textit{Case} (\textit{a}).
\end{itemize}
In this case our meta-state is
\begin{equation*}
\frac{1}{2}\biggl(|1M_{1}\rangle +|0M_{2}\rangle \biggr)\otimes \biggl(
\widetilde{1M_{1}}\rangle +|\widetilde{0M_{2}}\rangle \biggr).
\end{equation*
With a bit of algebra we can verify that the above expression remains\
formally unchanged, but with values of $c_{(i,\alpha )(\tilde{j},\tilde
\alpha})}^{(2)}$ generally changed.
To summarize the above scheme, the observer measuring $0$ at the beginning is
able to send a bit of information (say, `$a$' if he chooses action (\textit{
}), `$b$' if he chooses action (\textit{b})) to the observer who originally
measured $1$. In this way a procedure to communicate between two Everett
branches is set up.
In order to ensure the consistency of our argument, we can verify that in
the absence of gravitational interactions, case (\textit{a}) and case
\textit{b}) coincide. For this purpose it is convenient to use `mixed basis
states' (in which one factor of the product tensor is taken at the initial
time as before, while the other is considered at time $t$) to represent the
time evolved states, \textit{i.e.} writing
\begin{equation*}
|1\ M_{1};\widetilde{1\ M_{1}}\rangle _{t}=\sum_{(i,\mu )(\tilde{\imath}
\tilde{\mu})}c_{(i,\mu )(\tilde{\imath},\tilde{\mu})}^{(1)}|i\mu \rangle _{0
\underset{=e^{-\frac{i\tilde{H_{0}}t}{\hbar }}|\tilde{\imath},\tilde{\mu
\rangle _{0}}{\underbrace{|\tilde{\imath},\tilde{\mu}\rangle _{t}}}
\end{equation*
and
\begin{equation*}
|1\ M_{1};\widetilde{\mathbf{i}\ M_{2}}\rangle _{t}=\sum_{(i,\mu )(\tilde
\imath},\tilde{\mu})}c_{(i,\mu )(\tilde{\imath},\tilde{\mu})}^{(2)}|i\mu
\rangle _{0}|\tilde{\imath},\tilde{\mu}\rangle _{t}\ .
\end{equation*}
Remembering that ${}_{0}\langle \widetilde{i\mu }|\widetilde{1M_{2}}\rangle
_{0}={}_{t}\langle \widetilde{i\mu }|\widetilde{1M_{2}}\rangle _{t}=\delta _
\widetilde{i\mu }}\delta _{\widetilde{1M_{2}}}$, it is immediate to prove
the identity of $\langle \sigma_{3}\rangle $ for the two cases, given by
\begin{equation*}
\langle \sigma _{3}\rangle =2\sum_{\alpha }\biggl[\biggl\vert{}_{0}\langle
1,\alpha |1,M_{1}\rangle _{t}\biggr\vert^{2}-\biggl\vert{}_{0}\langle
2,\alpha |1,M_{1}\rangle _{t}\biggr\vert^{2}\biggr].
\end{equation*}
It's important to note that, apart from formal consistency, the above result
show that an Everett phone cannot work for truly microscopic systems.
As the last point concerning branch communications, it's easy to see that if
the superposition $M_{1}+M_{2}$ involves sufficiently spatially-separated
massive systems, then the expectation values $\langle \sigma _{3}\rangle ~
for case (\textit{a}) and (\textit{b}) become identical in a very short
time, meaning that after that time Everett Universes' communication
possibilities are strongly suppressed.
Suppose, in fact, that $M_{1}$ and $M_{2}$ are approximate CM's position
eigenstates with a mass above threshold $m\gtrsim 10^{12} m_{p}$ at
a distance $\sim \left( m_{p}/m\right) ^{1/2}cm$ from each other\footnote{
By `superposition of
two approximate position eigenstates' we have meant,
throughout the text, a superposition of two contiguous clusters of
localized states; as a matter of fact, a necessary condition for a complete
state reduction within the characteristic time $\tau _{g}$ is that the
superposition is composed by a large number of localized states, of width
\sim \left( m_{p}/m\right) ^{1/2}cm$ (see \cite{DeFilMaimRob} for an
explicit numerical simulation of this case). Otherwise, a superposition of
two really separated localized states would lead to a rapid oscillation of
coherences in the basis of positions. Indeed, for all practical purposes,
for sufficiently massive bodies, Nonunitary Gravity acts to reduce the
overall quantum state at the very beginning of the (unitary) process of
superposition formation.\label{nota}}. As before, the dynamics in this situation lead to
a rapid reduction of the state within the characteristic time $\tau _{g}$.
For the sake of simplicity, let's suppose that the spin dynamics is
slower than $\tau _{g}$; then, with respect to the case (\textit{b}), the
state $\frac{1}{\sqrt{2}}(|1M_{1}\rangle +|\widehat{\mathbf{i}}M_{2}\rangle )
$ evolves in a time $t\gtrsim \tau _{g}$ to give the almost diagonal state
\begin{equation*}
\rho (t)\simeq \frac{1}{2}\biggl(|1M_{1}\rangle \langle 1M_{1}|+|\widehat
\mathbf{i}}M_{2}\rangle \langle \widehat{\mathbf{i}}M_{2}|\biggr)\ ,
\end{equation*
from which, after tracing out system $M$, we see that there is a given
probability to obtain $1$ associated to $M_{1}$, and the complementary
probability to obtain $\widehat{\mathbf{i}}$ associated to (mind, recording
device, etc.) $M_{2}$ without any `transfer' from branch of $M_{2}$ to
branch $M_{1}$ after a time $t\gtrsim \tau _{g}$.
In conclusion, to set-up an efficient Everett communication, we need
macroscopic mass superpositions (because gravity is responsible for
nonlinearity, the basic ingredient for Everett communication!). At the same
time, the macroscopic nature of the required quantum superposition implies
its rapid decay, putting a severe trade-off between efficiency and practical
feasibility of the Everett phone.
\section{Conclusions}
We have explicitly shown how the well-known causality problems of the Newton-Schrödinger equation can be cured in a natural and satisfactory way at the price of introducing a certain amount of
predictability loss into the theory. The character of this extra-level of
indeterminism is such that it is completely irrelevant for microscopic
systems, while prompting large entropy production when macroscopic bodies
are involved. Being the ensuing theory, equivalent to NNG, a fully
consistent QM model, we have shown in an explicit physical setting how the
theory itself gets rid automatically of the superluminal communication
channels. This has been accomplished by analyzing an (ideal) EPR-type
experiment involving the superposition of two distinct CM position states of
a massive body. It has been shown that this circumstance arises from the
intrinsic mechanism of spontaneous state reduction of the model, which is
completely suppressed in the Newton-Schr\"{o}dinger limit. As a matter of
fact, the present analysis can be considered as an independent argument for
the introduction of NNG.
As to the experimental consequences of our calculations, one could not design an experimental setting to detect the average gravitational field of a whatever quantum superposition of localized states, even if semiclassical gravity is the correct description for all practical purposes. Turning it the other way around, one could not use semiclassical gravity to design an apparatus to send faster-than-light signals.
Besides, since within NNG density matrix plays a fundamental role, amounting
to the most complete characterization of a physical system's state, the
Everett Many World Interpretation appears to be the most natural conceptual
framework of that theory.
As in other approaches in which non-linearity was introduced at a
fundamental level in QM, the possibility of constructing an Everett phone
connecting different branches of the wave function emerges, though the
mechanism it is based on appears to be strongly inhibited. In fact the
theory gets rid of the huge number of branches continuously forming by
turning each time the macroscopic states superposition into ensembles of
localized states through gravitational self-interaction.
The severe restriction to branching proliferation would be responsible for
keeping the process confined within the microscopic-to-mesoscopic realm. To
further clarify this point, let's consider the famous Schr\"{o}dinger's cat
thought experiment in whatever of its countless versions. The main point
concerning this thought experiment is that a microscopic system's dynamics
(described by the Schr\"{o}dinger equation) is amplified up to the level of
a macroscopic object (like a gun or a bottle of poison), and from it to a
poor cat, turning the world split into two macroscopically well distinct
branches: one in which the cat stands up happy and alive, and one in which
the same cat is lying stretched out on the floor. Decoherence theory adds to
this picture the participation of the surrounding environment in this
branching, but the principle remains exactly the same. What happens when
Nonunitary Gravity is acting? That as soon as the gun's trigger (or the
poison vapor) begins to form superpositions of mesoscopically different mass
displacement states, the overall (cat+killing apparatus+microscopic
system+surrounding environment) pure state is rapidly converted into an
ensemble of classical-like states, meaning that each state has its own
probability of occurrence. It is important to stress that this would not
correspond to the subjective coarse graining, usually given by tracing out
the environmental degrees of freedom. It would, instead, be associated with
a fundamental coarse graining, the adjective `fundamental' meaning that one
cannot resort even in principle to a more complete description, regardless
of how precise and technologically sophisticated his measuring apparatuses
would be.
Notwithstanding, of course in the microscopic realm the branching of the
wave function continues undisturbed, while some remote echoes of other
parallel worlds could eventually take a role in the biological processes
involved in the brain's functioning.
As a final remark of moral order, a \textit{soft} version of the Everett
formulation like the one just described, if found to conform to physical
reality, would presumably sound less horrible, and even
psychologically/morally more acceptable than the quite disturbing notion of
an infinite number of replicas of ourselves doing who knows what, who knows
where.
\textit{Note added}. During the completion of this paper we became aware of
a work by S. De Filippo, which also points out the ability of NNG model in avoiding
superluminal communications \cite{NotaDeFilippo}.
\begin{appendices}
\numberwithin{equation}{section}
\section{Proof of Eq. (\protect\ref{metastat})}
Let's proceed by induction. For $N=1,$ Eq. \ref{metastat}) is trivially
verified. Assuming that it is verified for $N,$ it is sufficient to prove
that it is verified for $N+1$, \textit{i.e.} we have to prove the
equivalence of
\begin{equation}
\biggl(\sum_{k=0}^{N}\sqrt{\binom{N}{k}}p^{\frac{k}{2}}q^{\frac{N-k}{2
}\vert x\rangle ^{\otimes k} \vert y\rangle ^{\otimes (N-k)}\biggr)\times
\sqrt{p}\vert x^{N+1}\rangle +\sqrt{q}\vert y^{N+1}\rangle)
\end{equation}
with
\begin{equation}
\sum_{k=0}^{N+1}\sqrt{\binom{N+1}{k}}p^{\frac{k}{2}}q^{\frac{N-k+1}{2}}\frac{
1}{\sqrt{N p^{\prime \prime }}}\sum_{P^{\prime \prime }}P^{\prime \prime
}\Vert x^1,\dots ,x^{k}, y^{k+1},\dots ,y^{N+1}\rangle\rangle .
\end{equation}
Both expressions are decomposable in a unique way as sums of terms each with
a fixed number $k$ of $x$. So we can show the equivalence of the two
expressions term by term:
\begin{eqnarray*}
&&\sqrt{\binom{N}{k}}p^{\frac{k}{2}}q^{\frac{N-k+1}{2}}\frac{1}{\sqrt{N p}}
\sum_{P}P\Vert x^{1},\dots ,x^{k}, y^{k+1},\dots ,y^{N}\rangle\rangle \vert
y^{N+1}\rangle \\
& + & \sqrt{\binom{N}{k-1}}p^{\frac{k}{2}}q^{\frac{N-k+1}{2}}\frac{1}{\sqrt{
N p^{\prime }}}\sum_{P^{\prime }}P^{\prime} \Vert x^1,\dots ,x^{k-1},
y^{k},\dots ,y^{N}\rangle\rangle \vert x^{N+1}\rangle \\
& = & \sqrt{N p^{\prime \prime }} p^{\frac{k}{2}}q^{\frac{N-k+1}{2}}\frac{1}{
\sqrt{N p^{\prime \prime }}}\sum_{P^{\prime \prime }}P^{\prime \prime
1},\dots ,x^{k}, y^{k+1},\dots ,y^{N+1}\rangle\rangle \vert y^{N+1}\rangle \\
& \equiv & \sqrt{\binom{N+1}{k}}p^{\frac{k}{2}}q^{\frac{N-k+1}{2}}\vert
x\rangle^{\otimes k} \vert y\rangle^{\otimes (N+1-k)}.
\end{eqnarray*}
The total number of permutations involved in the above transformations is
equal to
\begin{equation*}
N_{p}+N_{p^{\prime }}\equiv \binom{N}{k}+\binom{N}{k-1}=\binom{N+1}{k}\equiv
N p^{\prime \prime }.
\end{equation*}
\section{General formulation of the model}
\label{general}
For a more general formulation of the model, it is convenient to switch to
second quantization. Let $H[\psi ^{\dagger },\psi ]$ denote the second
quantized non relativistic Hamiltonian of a finite number of particle
species, like electrons, nuclei, ions, atoms and/or molecules, according to
the energy scale. For notational simplicity, $\psi ^{\dagger },\,\psi $
denote the whole set $\psi _{j}^{\dagger }(x),\,\psi _{j}(x)$ of
creation-annihilation operators, \textit{i.e.} one couple per particle
species and spin component. This Hamiltonian includes the usual
electromagnetic interaction accounted for in atomic and molecular physics.
To incorporate gravitational interactions including self-interactions, we
introduce a color quantum number $\alpha =1,2,\dots ,N$, in such a way that
each couple $\psi _{j}^{\dagger }(x),\,\psi _{j}(x)$ is replaced by $N$
couples $\psi _{j,\alpha }^{\dagger }(x),\,\psi _{j,\alpha }(x)$ of
creation-annihilation operators. The overall Hamiltonian, including
gravitational interactions and acting on the tensor product $\otimes
_{\alpha }F_{\alpha }$ of the Fock space of the $\psi _{\alpha }$ operators,
is then given by
\begin{equation}
H_{G}=\sum_{\alpha =1}^{N}H[\psi ^{\dagger },\psi ]-\frac{G}{N-1
\sum_{j,k}m_{j}m_{k}\sum_{\alpha <\beta }\int dx\,dy\frac{\psi _{j,\alpha
}^{\dagger }(x),\,\psi _{j,\alpha }(x)\psi _{k,\beta }^{\dagger }(y),\,\psi
_{k,\beta }(y)}{|x-y|},
\end{equation
where here and henceforth Greek indices denote color indices, $\psi _{\alpha
}\equiv (\psi _{1,\alpha },\psi _{2,\alpha },\dots ,\psi _{N,\alpha })$ and
m_{i}$ denotes the mass of the $i-$th particle species, while $G$ is the
gravitational constant. While the $\psi _{\alpha }$ operators obey the same
statistics as the original operators $\psi $, we take advantage of the
arbitrariness pertaining to distinct operators and, for simplicity, we chose
them commuting with one another: $\alpha \neq \beta \Rightarrow \lbrack \psi
_{\alpha },\psi _{\beta }]_{-}=[\psi _{\alpha },\psi _{\beta }^{\dagger
}]_{-}=0.$ The metaparticle state space $S$ is identified with the subspace
of $\otimes _{\alpha }F_{\alpha }$ including the metastate obtained from the
vacuum $\Vert 0\rangle \rangle =\otimes _{\alpha }|0\rangle _{\alpha }$
applying operators built in terms of the product $\prod_{\alpha =1}^{N}\psi
_{j,\alpha }^{\dagger }(x_{\alpha })$ and symmetrical with respect to
arbitrary permutations of the color indices, which, as a consequence, for
each particle species, have the same number of metaparticles of each color.
This is a consistent definition since the time evolution generated by the
overall Hamiltonian is a group of (unitary) endomorphism of $S$. If we
prepare a pure $n-$particle state, represented in the original setting,
excluding gravitational interactions, by
\begin{equation*}
|g\rangle \doteq \int d^{n}x\,g(x_{1},x_{2},\dots ,x_{n})\psi
_{j_{1}}^{\dagger }(x_{1}),\psi _{j_{2}}^{\dagger }(x_{2})\dots ,\psi
_{j_{n}}^{\dagger }(x_{n})|0\rangle ,
\end{equation*
its representative in $S$ is given by the metastate
\begin{equation*}
\Vert g^{\otimes N}\rangle \rangle \doteq \prod_{\alpha }\biggl[\int
d^{n}x\,g(x_{1},x_{2},\dots ,x_{n})\psi _{j_{1},\alpha }^{\dagger
}(x_{1}),\psi _{j_{2},\alpha }^{\dagger }(x_{2})\dots ,\psi _{j_{n},\alpha
}^{\dagger }(x_{n})\biggr]\Vert 0\rangle \rangle .
\end{equation*}
As for the physical algebra, it is identified with the operator algebra of
say the $\alpha =1$ metaworld. In view of this, expectation values can be
evaluated by previously tracing out the unobservable operators, namely with
\alpha >1$, and then taking the average of an operator belonging to the
physical algebra. It should be made clear that we are not prescribing an
\textit{ad hoc} restriction of the observable algebra. Once the constraint
restricting $\otimes _{\alpha }F_{\alpha }$ to $S$ is taken into account,
in order to get an effective gravitational interaction between particles of
one and the same color, the resulting state space does not contain states
that can distinguish between operators of different color. The only way to
accommodate a faithful representation of the physical algebra within the
metastate space is to restrict the algebra to that of $\psi _{1}$ operators.
Note that the resulting constrained\ theory is, by construction, a fully
consistent QM theory.
\subsection{State reduction}
\label{reduction}
The evolution operator in the interaction representation mapping an initial
physical state $\rho (0)=|\Phi (0)\rangle _{1}\left\langle \Phi
(0)\right\vert $ into the evolved physical state $\rho (t)$ can be written,
according to \cite{DeFil+DeFilMaim, NotaDeFilippo}, as
\begin{eqnarray}
M(t) &=&\int \mathcal{D}[\varphi ]\prod_{\alpha }\mathcal{D}[\varphi
_{\alpha }]\mathcal{D}[\varphi ^{\prime }]\prod_{\alpha }D[\varphi ^{\prime
{_{\alpha }}]exp{\biggl[\frac{ic^{2}}{2\hbar }\int dt\,dx\bigl[\varphi
\nabla ^{2}\varphi -\sum_{\alpha }\varphi _{\alpha }\nabla ^{2}\varphi
_{\alpha }-\varphi ^{\prime 2}\varphi ^{\prime }+\sum_{\alpha }\varphi
_{\alpha }^{\prime }\nabla ^{2}\varphi _{\alpha }^{\prime }\bigr]}\biggr]
\notag \label{MdiT} \\
&\times &\biggl[\bigotimes_{\alpha =2}^{N}\;{}_{\alpha }\langle \Phi (0)
\biggr]T^{-1}\exp {\biggl[i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}
\sum_{\alpha =2}^{N}\int dt\,dx[\varphi ^{\prime }(x,t)+\varphi _{\alpha
}^{\prime }(x,t)]\psi _{\alpha }^{\dagger }(x,t)\psi _{\alpha }(x,t)\biggr]}
\notag \\
&\times &T\exp {\biggl[-i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}
\sum_{\alpha =2}^{N}\int dt\,dx[\varphi (x,t)+\varphi _{\alpha }(x,t)]\psi
_{\alpha }^{\dagger }(x,t)\psi _{\alpha }(x,t)\biggr]}\bigotimes_{\alpha
=2}^{N}|\Phi (0)\rangle _{\alpha } \\
&\times &T\exp {\biggl[-i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}}\int
dt\,dx[\varphi (x,t)+\varphi _{1}(x,t)]\psi _{1}^{\dagger }(x,t)\psi
_{1}(x,t)\biggr]}|\Phi (0)\rangle _{1} \notag \\
&\times &{}_{1}\langle \Phi (0)|T^{-1}\exp {\biggl[i\frac{2mc}{\hbar }\sqrt
\frac{\pi G}{N-1}}\int dt\,dx[\varphi ^{\prime }(x,t)+\varphi _{1}^{\prime
}(x,t)]\psi _{1}^{\dagger }(x,t)\psi _{1}(x,t)\biggr]}. \notag
\end{eqnarray
Let's consider localized states $|z\rangle _{1}$, which are approximate
eigenstates of the density operator $\psi _{1}^{\dagger }(x,t)\psi
_{1}(x,t)|z\rangle _{1}\approx n(x-z)|z\rangle _{1}$ that are quasi
stationary, apart from a slow spreading proportional to $N$ (associated to
the center of metamass $Nm$ spreading). Taking as initial state a
superposition of a large number $\mathcal{N}$ of localized states,
\begin{equation*}
|\Phi (0)\rangle _{1}=\frac{1}{\sqrt{\mathcal{N}}}\sum_{j=1}^{\mathcal{N
}|z_{j}\rangle _{1},
\end{equation*
where $\mathcal{N}$ is the number of states in superposition.
We want to evaluate explicitly the matrix element
\begin{eqnarray*}
&&{}_{1}\langle z_{h}|M(t)|z_{k}\rangle _{1} \\
&=&\int \mathcal{D}[\varphi ]\prod_{\alpha }\mathcal{D}[\varphi _{\alpha }
\mathcal{D}[\varphi ^{\prime }]\prod_{\alpha }D[\varphi ^{\prime }{_{\alpha
}]exp{\biggl[\frac{ic^{2}}{2\hbar }\int dt\,dx\bigl[\varphi \nabla
^{2}\varphi -\sum_{\alpha }\varphi _{\alpha }\nabla ^{2}\varphi _{\alpha
}-\varphi ^{\prime 2}\varphi ^{\prime }+\sum_{\alpha }\varphi _{\alpha
}^{\prime }\nabla ^{2}\varphi _{\alpha }^{\prime }\bigr]}\biggr] \\
&\times &\biggl[\bigotimes_{\alpha =2}^{N}\frac{1}{\sqrt{\mathcal{N}}
\sum_{j=1}^{\mathcal{N}}\,{}_{\alpha }\langle z_{j}|\biggr]T^{-1}\exp
\biggl[i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}}\sum_{\alpha =2}^{N}\int
dt\,dx[\varphi ^{\prime }(x,t)+\varphi _{\alpha }^{\prime }(x,t)]\psi
_{\alpha }^{\dagger }\psi _{\alpha }(x,t)\biggr]} \\
&\times &T\exp {\biggl[-i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}
\sum_{\alpha =2}^{N}\int dt\,dx[\varphi (x,t)+\varphi _{\alpha }(x,t)]\psi
_{\alpha }^{\dagger }(x,t)\psi _{\alpha }(x,t)\biggr]}\biggl
\bigotimes_{\alpha =2}^{N}\frac{1}{\sqrt{\mathcal{N}}}\sum_{j=1}^{\mathcal{N
}|z_{j}\rangle _{\alpha }\biggr] \\
&\times &{}_{1}\langle z_{h}|T\exp {\biggl[-i\frac{2mc}{\hbar }\sqrt{\frac
\pi G}{N-1}}\int dt\,dx[\varphi (x,t)+\varphi _{1}(x,t)]\psi _{1}^{\dagger
}(x,t)\psi _{1}(x,t)\biggr]}|\Phi (0)\rangle _{1} \\
&\times &{}_{1}\langle \Phi (0)|T^{-1}\exp {\biggl[i\frac{2mc}{\hbar }\sqrt
\frac{\pi G}{N-1}}\int dt\,dx[\varphi ^{\prime }(x,t)+\varphi _{1}^{\prime
}(x,t)]\psi _{1}^{\dagger }(x,t)\psi _{1}(x,t)\biggr]|z_{k}\rangle _{1}}
\end{eqnarray*}
\begin{eqnarray*}
&=&\int \mathcal{D}[\varphi ]\prod_{\alpha }\mathcal{D}[\varphi _{\alpha }
\mathcal{D}[\varphi ^{\prime }]\prod_{\alpha }D[\varphi ^{\prime }{_{\alpha
}]exp{\biggl[\frac{ic^{2}}{2\hbar }\int dt\,dx\bigl[\varphi \nabla
^{2}\varphi -\sum_{\alpha }\varphi _{\alpha }\nabla ^{2}\varphi _{\alpha
}-\varphi ^{\prime 2}\varphi ^{\prime }+\sum_{\alpha }\varphi _{\alpha
}^{\prime }\nabla ^{2}\varphi _{\alpha }^{\prime }\bigr]}\biggr] \\
&\times &\biggl[\bigotimes_{\alpha =2}^{N}\frac{1}{\sqrt{\mathcal{N}}
\sum_{j=1}^{\mathcal{N}}\,{}_{\alpha }\langle z_{j}|\biggr]\exp {\biggl[
\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}}\sum_{\alpha =2}^{N}\int
dt\,dx[\varphi ^{\prime }(x,t)+\varphi _{\alpha }^{\prime }(x,t)]n(x-z_{j}
\biggr]} \\
&\times &\exp {\biggl[-i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}
\sum_{\alpha =2}^{N}\int dt\,dx[\varphi (x,t)+\varphi _{\alpha
}(x,t)]n(x-z_{j})\biggr]}\biggl[\bigotimes_{\alpha =2}^{N}\frac{1}{\sqrt
\mathcal{N}}}\sum_{j=1}^{\mathcal{N}}|z_{j}\rangle _{\alpha }\biggr]\times \\
&\times &\exp {\biggl[-i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}}\int
dt\,dx[\varphi (x,t)+\varphi _{1}(x,t)]n(x-z_{h})\biggr]} \\
&\times &\exp {\biggl[i\frac{2mc}{\hbar }\sqrt{\frac{\pi G}{N-1}}\int
dt\,dx[\varphi ^{\prime }(x,t)+\varphi _{1}^{\prime }(x,t)]n(x-z_{k})\biggr]
\times {}_{1}\langle z_{h}|\Phi (0)\rangle _{1}\langle \Phi (0)|z_{k}\rangle
_{1}
\end{eqnarray*
\begin{eqnarray}
&=&\frac{1}{\mathcal{N}^{N}}\sum_{j_{2},\dots ,j_{a}\dots ,j_{N}}\exp
\biggl[-\frac{im^{2}}{2\hbar }\frac{G}{N-1}t\int dx\,dy\frac{[\sum_{\alpha
=2}^{N}n(x-z_{j_{\alpha }})+n(x-z_{h})][\sum_{\alpha ^{\prime
}=2}^{N}n(y-z_{j_{\alpha ^{\prime }}})+n(y-z_{h})]}{|x-y|}\biggr]} \notag \\
&\times &\exp {\biggl[\frac{im^{2}}{2\hbar }\frac{G}{N-1}t\int dx\,dy\frac
[\sum_{\alpha =2}^{N}n(x-z_{j_{\alpha }})+n(x-z_{k})][\sum_{\alpha ^{\prime
}=2}^{N}n(y-z_{j_{\alpha ^{\prime }}})+n(y-z_{k})]}{|x-y|}\biggr]} \notag \\
&\times &\exp {\biggl[-\frac{im^{2}}{2\hbar }\frac{G}{N-1}t\int dx\,dy\biggl
\frac{n(x-z_{h})n(y-z_{h})+\sum_{\alpha =2}^{N}n(x-z_{j_{\alpha
}})n(y-z_{j_{\alpha }})}{|x-y|}\biggr]\biggr]} \notag \\
&\times &\exp {\biggl[-\frac{im^{2}}{2\hbar }\frac{G}{N-1}t\int dx\,dy\biggl
\frac{n(x-z_{k})n(y-z_{k})+\sum_{\alpha =2}^{N}n(x-z_{j_{\alpha
}})n(y-z_{j_{\alpha }})}{|x-y|}\biggr]\biggr]} \notag \\
&=&\frac{1}{\mathcal{N}^{N}}\sum_{j_{2},\dots ,j_{a}\dots ,j_{N}}\exp
\biggl[\frac{im^{2}}{2\hbar }\frac{G}{N-1}t\int dx\,dy\frac{[\sum_{\alpha
=2}^{N}n(x-z_{j_{\alpha }})n(y-z_{h})-\sum_{\alpha =2}^{N}n(x-z_{j_{\alpha
}})n(y-z_{k})]}{|x-y|}\biggr]} \label{espo} \\
&=&\frac{1}{\mathcal{N}^{N}}\biggl[\sum_{j=1}^{\mathcal{N}}e^{i\frac
\mathcal{A}_{j}}{N-1}}\biggr]^{N-1}, \notag
\end{eqnarray}
where
\begin{equation*}
\mathcal{A}_{j}={\frac{m^{2}G\ t}{2}\int dx\,dy\frac
[n(x-z_{j})n(y-z_{h})-n(x-z_{j})n(y-z_{k})]}{|x-y|}.}
\end{equation*}
Now, using the property of the multinomial distributions for $Multi(n
\mathbf{p})\underset{n\rightarrow \infty }{\approx }\mathfrak{N}(n\mathbf{p
,n\Sigma ),$ where $\mathfrak{N}$ is the normal multivariate distribution
and $\Sigma =P-\mathbf{p}\mathbf{p}^{T},$ with $P$ being the diagonal matrix
whose diagonal is formed by the probability vector $\mathbf{p},$ we can rewrite
\ref{espo}) as
\begin{equation*}
\frac{1}{\mathcal{N}^{N}}\sum_{\substack{ k1,\dots ,k_{\mathcal{N}} \\
\mbox{with }k_{1}+\dots +k_{\mathcal{N}}=N-1}}\binom{N-1}{k_{1}k_{2}\dots k_
\mathcal{N}}}\biggl[\frac{e^{i\frac{\mathcal{A}_{1}}{N-1}}}{\mathcal{N}
\biggr]^{k_{1}}\biggl[\frac{e^{i\frac{\mathcal{A}_{2}}{N-1}}}{\mathcal{N}
\biggr]^{k_{2}}\dots \biggl[\frac{e^{i\frac{\mathcal{A}_{\mathcal{N}}}{N-1}
}{\mathcal{N}}\biggr]^{k_{\mathcal{N}}},
\end{equation*
which, introducing the rescaled variables $\chi _{i}\equiv \frac{k_{i}}{N-1}$
and passing from sum to integral, transforms into
\begin{eqnarray}
&&\frac{e^{i\mathcal{A}_{\mathcal{N}}}}{\mathcal{N}}\int_{\substack{ \chi
_{i}>0,\forall i=1,\dots ,\mathcal{N}-1 \\ \mbox{with}\sum_{i=1}^{\mathcal{N
-1}\chi _{i}\leq 1}}d\chi _{1}\dots d\chi _{\mathcal{N}}\prod_{j=1}^
\mathcal{N}-1}\frac{1}{\sqrt{2\pi }\sqrt{\frac{1}{\mathcal{N}(\mathcal{N}-1)
}}\exp {\biggl[-\frac{(\chi _{j}-\frac{1}{\mathcal{N}})^{2}}{\frac{2}
\mathcal{N}(N-1)}}+i(\mathcal{A}_{j}-\mathcal{A_{\mathcal{N}}})\chi _{j
\biggr]} \notag \\
&\xrightarrow[N\rightarrow \infty]{}&\frac{e^{i\mathcal{A}_{\mathcal{N}}}}
\mathcal{N}}\int_{\substack{ \chi _{i}>0,\forall i=1,\dots ,\mathcal{N}-1 \\
\mbox{with}\sum_{i=1}^{\mathcal{N}-1}\chi _{i}\leq 1}}d\chi _{1}\dots d\chi
_{\mathcal{N}}\delta \biggl(\chi _{1}-\frac{1}{\mathcal{N}}\biggr)\times
\dots \times \delta \biggl(\chi _{\mathcal{N}-1}-\frac{1}{\mathcal{N}}\biggr
e^{i(\mathcal{A}_{j}-\mathcal{A}_{\mathcal{N}})\chi _{j}}= \notag \\
&=&\frac{1}{\mathcal{N}}e^{i/\mathcal{N}\sum_{j=1}^{\mathcal{N}}\mathcal{A
_{j}}. \label{lunga}
\end{eqnarray
Then
\begin{equation*}
\left\vert _{1}\langle z_{h}|M(t)|z_{k}\rangle _{1}\right\vert \underset
N\rightarrow \infty }{\longrightarrow }\frac{1}{\mathcal{N}},
\end{equation*}
while, in the same limit, the spreading time tends to infinity, from which we
conclude that the mechanism of random phase cancelation, leading for finite
N$ to a rapid decoherence of the superposition of localized states on one
side, and the spreading of the wave function on the other side, are
completely suppressed in the limit $N\rightarrow \infty $.
It is worth noting that, for the simplest case of $N=2$ treated in the main
text, we get easily the expression for the characteristic state reduction
time. Starting from the intermediate expression \ref{espo} specialized
for $N=2$,
\begin{equation*}
{}\frac{1}{\mathcal{N}^{2}}\sum_{j=1}^{\mathcal{N}}\exp {\biggl[\frac{
im^{2}G\ t}{2\hbar }\int dx\,dy\frac{
[n(x-z_{j})n(y-z_{h})-n(x-z_{j})n(y-z_{k})]}{|x-y|}\biggr],}
\end{equation*}
we see that phase cancelation occurs when exponentials reach values of order
$1$, which happens in a characteristic time $\tau _{g}=\hbar
G^{-1}M^{-5/3}\rho ^{-1/3}\ $ \cite{DeFMaimPRD}. For example, for a very fine
grain of sand, of mass $10^{-6}gr.$, we have $\tau _{g}\sim 10^{-10}\sec .
, consistently short with respect to the mass-independent spreading time of
10^{3}\sec .$
\end{appendices}
|
\subsection{Case study: Boundary attack}
The boundary attack uses an iterative update rule to gradually move the adversarial image closer to the original image, maintaining that the image remains adversarial at each step. Starting from random noise, the algorithm samples a noise matrix $\eta \sim N(0,1)^{d \times d}$ at each iteration and adds it to the current iterate $\mathbf{z}$ after appropriate scaling. This point is then projected onto the sphere of center $\mathbf{x}$, the target image, and radius $\|\mathbf{z}\|_2$ so that the next iterate never moves away from $\mathbf{x}$. Finally, we contract towards $\mathbf{x}$ by $\epsilon$, and the new iterate $\tilde{\mathbf{z}}$ is accepted only if it remains adversarial. This guarantees that terminating the algorithm at any point still results in an adversarial image, but the perturbation magnitude reduces with each contraction step.
\paragraph{Modification.} To construct low frequency perturbation using the boundary attack, we constrain the noise matrix $\eta$ to be sampled from $\idct_r(N(0,1)^{d \times d})$ instead. Figure~\ref{fig:boundary} illustrates the modified attack. Sampling low frequency noise instead of Gaussian noise is particularly beneficial to the boundary attack: After adding the noise matrix $\eta$, if the iterate is not adversarial, the algorithm must re-sample a noise matrix and perform another model query. By restricting to the low frequency subspace, which has a larger density of adversarial directions, this step succeeds more often, speeding up convergence towards the target image. We term this variant of the boundary attack as \emph{low frequency boundary attack} (LF-BA) and the original boundary attack as RGB-BA.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.49\textwidth]{fig/queries_ba.pdf}
\includegraphics[width=0.49\textwidth]{fig/queries_nes.pdf}
\caption{Distribution of the number of queries required for a successful attack (defined as achieving a perturbation MSE of 0.001 or lower for RGB-BA/LF-BA). See text for details.}
\label{fig:queries_comparison}
\vspace{-1em}
\end{figure*}
\paragraph{Hyperparameters.} The boundary attack has two hyperparameters: noise step size $\delta$ and contraction step size $\epsilon$. Both step sizes are adjusted based on the success rate of the past few candidates, \emph{i.e.}, if $\tilde{\mathbf{z}}$ is accepted often, we can contract towards the target $\mathbf{x}$ more aggressively by increasing $\epsilon$ and vice versa, and similarly for $\delta$. For the low frequency variant, we find that fixing $\delta$ to a large value is beneficial for speeding up convergence, while also reducing the number of model queries by half. For all experiments, we fix $\delta = 0.2$ and initialize $\epsilon = 0.01$.
Selecting the frequency ratio $r$ is more crucial. Different images may admit adversarial perturbations at different frequency ranges, and thus we would like the algorithm to automatically discover the right frequency on a per-image basis. We use Hyperband \citep{li2016hyperband}, a bandit-type algorithm for selecting hyperparameters, to optimize the frequency ratio $r$. We initialize Hyperband with multiple runs of the attack for every frequency ratio $r \in \{\frac{1}{4},\frac{1}{8},\frac{1}{16},\frac{1}{32}\}$. Repeatedly after $T'$ iterations, the least successful half of the parallel runs is terminated until one final frequency remains. This setting is continued until the total number of model queries reaches $T$.
\subsection{Case study: NES attack}
Natural evolution strategies (NES) \citep{wierstra2014natural} is a black-box optimization technique that has been recently proposed for its use in black-box attacks \citep{ilyas2018blackbox}. The attacker constructs the adversarial image $\mathbf{z}$ by minimizing a continuous-valued adversarial loss $\ell$ returned by black-box query to the model. However, instead of minimizing $\ell$ directly, the NES attack minimizes the loss at all points near $\mathbf{z}$. More precisely, we specify a search distribution $\mathcal{D}$ and minimize:
\begin{equation}
\min_{\mathbf{z}} \mathbb{E}_{\eta \sim \mathcal{D}}[\ell(\mathbf{z} + \eta)] \hspace{4pt} \text{subject to} \hspace{4pt} d(\mathbf{x}, \mathbf{z}) \leq \rho,
\label{eq:nes}
\end{equation}
where $\rho$ is some perceptibility threshold. When the search distribution $\mathcal{D}$ is chosen to be an isotropic Gaussian, i.e. $\mathcal{D} = N(0, \sigma^2)^{d \times d}$, the gradient of the objective function in \autoref{eq:nes} becomes
\begin{equation*}
\nabla_{\mathbf{z}} \mathbb{E}_{\eta \sim \mathcal{D}}[\ell(\mathbf{z} + \eta)] = \frac{1}{\sigma^2} \mathbb{E}_{\eta \sim \mathcal{D}}[\ell(\mathbf{z} + \eta) \cdot \eta].
\end{equation*}
Thus, \autoref{eq:nes} can be minimized with stochastic gradient descent by sampling a batch of noise vectors $\eta_1,\ldots,\eta_m \sim N(0, \sigma^2)^{d \times d}$ and computing the (mini-batch) stochastic gradient
\begin{equation}
\nabla_{\mathbf{z}} \mathbb{E}_{\eta \sim \mathcal{D}}[\ell(\mathbf{z} + \eta)] \approx \frac{1}{m \sigma^2} \sum_{i=1}^m \ell(\mathbf{z} + \eta_i) \cdot \eta_i.
\label{eq:nes_update}
\end{equation}
One way to interpret this update rule is that the procedure pushes $\mathbf{z}$ away from regions of low adversarial density --- directions $\eta_i$ for which $\ell(\mathbf{z} + \eta_i)$ is high.
The perceptibility constraint can be enforced by projecting to the feasible region at every step. For this attack, the max-norm $\| \cdot \|_\infty$ is used as the perceptibility metric, hence the projection step reduces to clipping of each dimension in the adversarial perturbation to the range $[-\rho, \rho]$.
\paragraph{Modification.} The low frequency distribution defined in \autoref{sec:dct} can be readily incorporated into the NES attack. We replace the Gaussian search distribution with its low frequency version, i.e. we sample a batch of noise vectors $\eta_1,\ldots,\eta_m \sim \idct_r(N(0, \sigma^2)^{d \times d})$ instead. The stochastic gradient remains identical to \autoref{eq:nes_update}. Note that since each $\eta_i$ is low-frequency, this process results in a low frequency adversarial perturbation. We term the original NES attack using search distribution in pixel space as RGB-NES and the low frequency variant as \emph{low frequency NES} (LF-NES).
\paragraph{Hyperparameters.} The NES attack has two hyperparameters: $\rho$, which controls the perceptibility of adversarial perturbation, and $\sigma$, which controls the width of the search distribution. We set $\rho = 0.03$ to match the average $L_2$-norm of perturbations generated by RGB-BA/LF-BA, and set $\sigma = 0.001$ as suggested by the authors. Intriguingly, the frequency ratio $r$ is not very sensitive for LF-NES. Setting a single value of $r$ for all images is sufficiently effective, and we choose the same value of $r = 1/2$ in all of our experiments for simplicity.
\subsection{Defenses against black-box attacks}
\section{INTRODUCTION}
\input{introduction.tex}
\section{BACKGROUND}
\label{sec:background}
\input{background.tex}
\section{LOW FREQUENCY IMAGE SUBSPACE}
\label{sec:method}
\input{method.tex}
\section{APPLICATION TO BLACK-BOX ATTACKS}
\label{sec:application}
\input{application.tex}
\section{EMPIRICAL EVALUATION}
\label{sec:experiment}
\input{experiment.tex}
\section{DISCUSSION AND FUTURE WORK}
\input{conclusion.tex}
\newpage
\subsection{Universality of low frequency subspace}
Results in the previous section support our hypothesis that restricting the search space to LF-DCT substantially increases the sample success rate of random adversarial directions. However, the dimensionality reduction does impose a restriction on the possible solutions of attack algorithms. To examine the effects of this limitation, we apply our low-frequency restriction to white-box attacks by projecting the gradient onto the LF-DCT space.
\paragraph{Low frequency gradient descent.}
Let $\ell_{y}$ denote the adversarial loss, e.g. \autoref{eq:margin_loss}. For a given $r \in (0, 1]$ and $v \in \mathbb{R}^{rd \times rd}$, define $V \in \mathbb{R}^{d \times d}$ by
\begin{equation*}
V_{i,j} =
\begin{cases}
v_{i,j} & \text{if } 1 \leq i,j \leq rd \\
0 & \text{otherwise,}
\end{cases}
\end{equation*}
The wave coefficient matrix $V$ contains $v$ as its submatrix and only includes frequencies lower than $rd$. The low frequency perturbation domain can then be parametrized as $\Delta = \idct(V)$. To optimize with gradient descent, let $\bar{\Delta}$ and $\bar{V}$ be vectorizations of $\Delta$ and $V$, i.e., $\bar{\Delta}_{i_1 * d + i_2} = \Delta_{i_1, i_2}$ and similarly for $\bar{V}$. From \autoref{eq:idct}, it is easy to see that each coordinate of $\bar{\Delta}$ is a linear function of $\bar{V}$, hence $\idct$ is a linear transformation, whose adjoint is precisely the linear transformation defined by $\dct$. For any vector $\mathbf{z}$, its right-product with the Jacobian of $\idct$ is given by $J_{\idct} \cdot \mathbf{z} = \dct(\mathbf{z})$. Thus we may apply the chain rule to compute
\begin{equation*}
\frac{\partial \ell}{\partial V} = \dct \left( \frac{\partial \ell}{\partial \Delta} \right), \frac{\partial \ell}{\partial v} = \left[\frac{\partial \ell}{\partial V}\right]_{1:rd,1:rd},
\end{equation*}
which is equivalent to applying DCT to the gradient and dropping the high frequency coefficients.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig/whitebox_attack_samples.pdf}
\vspace{-3ex}
\caption{A sample image perturbed by the Carlini-Wagner attack using the full image space and low frequency space with different $r$. The adversarial perturbation (second row) has clearly different pattern across different frequency ranges.}
\label{fig:whitebox_sample}
\end{figure}
\begin{table}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccc}
\hline
~& $d'$ & MSE & Success Rate (\%) \\
\hline
RGB ($r = 1$) & 150528 & $2.78 \times 10^{-5}$ & 100.0 \\
LF-DCT ($r = 1/8$) & 2352 & $6.94 \times 10^{-5}$ & 100.0 \\
LF-DCT ($r = 1/16$) & 588 & $1.61 \times 10^{-4}$ & 95.5 \\
LF-DCT ($r = 1/32$) & 147 & $1.56 \times 10^{-4}$ & 56.0 \\
\hline
\end{tabular}
}
\caption{Average MSE and accuracy after Carlini-Wagner attack with different frequency ratios $r$. $d' = 3 \times rd \times rd$ is the effective adversarial space dimensionality. At $r=1/8$, optimizing in the frequency space of dimensionality 2352 is as effective as optimizing in the full image space.}
\label{table:whitebox_mse}
\vspace{-1em}
\end{table}
\paragraph{Adversarial optimality in low frequency subspace.}
Table \ref{table:whitebox_mse} shows average perturbation MSE and model accuracy after the Carlini-Wagner attack \citep{carlini2017towards} in low frequency space. The original attack in pixel space corresponds to $r = 1$. The images have three color channels and the effective subspace dimensionality is $d' = 3 \times rd \times rd$. For $r = 1/8$, the attack can achieve perfect (100\%) success rate, while the resulting MSE is only roughly 3 times larger --- despite that the search space dimensionality is only $1/64$ of the full image space.
This result further supports that the density of adversarial examples is much higher in the low frequency domain, and that searching exclusively in this restricted subspace consistently yields near-optimal adversarial perturbations.
As expected, choosing a very small frequency ratio eventually impacts success rate, as the subspace dimensionality is too low to admit adversarial perturbations.
\autoref{fig:whitebox_sample} shows the resulting adversarial images and perturbations corresponding to frequency ratios $r$. All perturbations are imperceptible but when isolated (bottom row) reveal increasingly smooth patterns as r decreases.
\paragraph{Advantages of low frequency perturbation.} While the remainder of this paper focuses on the benefits of low frequency adversarial perturbation in the black-box setting, we highlight that there are advantages in the white-box setting as well. \citet{sharma2019low} showed that low frequency gradient-based attacks enjoy greater efficiency and can transfer significantly better to defended models. In particular, their attack is able to completely circumvent all of the top-placing defense entries at the NeurIPS 2017 competition. Furthermore, they observe that the benefit of low frequency perturbation is not merely due to dimensionality reduction --- perturbing exclusively the high frequency components does not give the same benefit.
|
\section{Introduction}
\label{introduction}
With the penetration of internet among masses, the content being posted on social media channels has uptaken. Specifically, in the Indian subcontinent, number of Internet users has crossed 500 mi\footnote{\url{https://bit.ly/2MCXz2Q
}}, and is rising rapidly due to inexpensive data\footnote{\url{https://bit.ly/2NUjOGh
}}. With this rise, comes the problem of hate speech, offensive and abusive posts on social media. Although there are many previous works which deal with Hindi and English hate speech (the top two languages in India), but very few on the code-switched version (Hinglish) of the two \cite{mathur2018detecting}. This is partially due to the following reasons: (i) Hinglish consists of no-fixed grammar and vocabulary. It derives a part of its semantics from Devnagari and another part from the Roman script. (ii) Hinglish speech and written text consists of a concoction of words spoken in Hindi as well as English, but written in the Roman script. This makes the spellings variable and dependent on the writer of the text.
Hence code-switched languages present tough challenges in terms of parsing and getting the meaning out of the text. For instance, the sentence, \emph{``Modiji foreign yatra par hai''}, is in the Hinglish language. Somewhat correct translation of this would be, \emph{``Mr. Modi is on a foriegn tour''}. However, even this translation has some flaws due to no direct translation available for the word \emph{ji}, which is used to show respect. Verbatim translation would lead to \emph{``Mr. Modi foreign tour on is''}. Moreover, the word \emph{yatra} here, can have phonetic variations, which would result in multiple spellings of the word as \emph{yatra, yaatra, yaatraa}, etc. Also, the problem of hate speech has been rising in India, and according to the policies of the government and the various social networks, one is not allowed to misuse his right to speech to abuse some other community or religion. Due to the various difficulties associated with the Hinglish language, it is challenging to automatically detect and ban such kind of speech.
Thus, with this in mind, we build a transfer learning based model for the code-switched language Hinglish, which outperforms the baseline model of \cite{mathur2018detecting}. We also release the embeddings and the model trained.
\begin{table}[]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Hinglish} & \textbf{English} & \textbf{Hinglish} & \textbf{English} \\ \hline
acha & good & gunda & thug \\
s**la & blo*dy & ra*di & h*oker \\
\hline
\end{tabular}
\caption{Examples of word-pairs in Hinglish-English dictionary}
\label{tab:examples_hinglish_english}
\end{table}
\begin{table*}[]
\centering
\begin{tabular}{|c|c|c|}
\hline
\textbf{Category} & \textbf{Tweet} & \textbf{Translation} \\ \hline
Benign & sache sapooto aap ka balidan hamesha yaad rahega & True sons, your sacrifice would be remembered.\\
Hate Inducing & Bik gya Porkistan & Porkistan (Derogatory term for Pakistan) has been sold\\
Abusive & Kis m*darch*d ki he giri hui harkt & Which m*therf*cker has done this \\
\hline
\end{tabular}
\caption{Examples of tweets in the dataset and their translations}
\label{tab:examples_tweets}
\end{table*}
\section{Methodology}
Our methodology primarily consists of these steps: Pre-processing of the dataset, training of word embeddings, training of the classifier model and then using that on HEOT dataset.
\subsection{Pre-Processing}
In this work, we use the datasets released by \cite{davidson2017automated} and HEOT dataset provided by \cite{mathur2018detecting}. The datasets obtained pass through these steps of processing: (i) Removal of punctuatios, stopwords, URLs, numbers, emoticons, \emph{etc.} This was then followed by transliteration using the Xlit-Crowd conversion dictionary \footnote{\url{https://github.com/chsasank/indic-transliteration}} and translation of each word to English using Hindi to English dictionary\footnote{\url{http://www.cfilt.iitb.ac.in/~hdict/webinterface_user/}}. To deal with the spelling variations, we manually added some common variations of popular Hinglish words. Final dictionary comprised of 7200 word pairs. Additionally, to deal with profane words, which are not present in Xlit-Crowd, we had to make a profanity dictionary (with 209 profane words) as well. Table \ref{tab:examples_hinglish_english} gives some examples from the dictionary.
\subsection{Training Word Embeddings}
We tried Glove \cite{pennington2014glove} and Twitter word2vec \cite{godin2015multimedia} code for training embeddings for the processed tweets. The embeddings were trained on both the datasets provided by \cite{davidson2017automated} and HEOT. These embeddings help to learn distributed representations of tweets. After experimentation, we kept the size of embeddings fixed to 100.
\begin{table}[]
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Model} & \textbf{Accuracy}\\ \hline
Davidson et al. & 0.57\\
Our Model with embeddings trained on Glove & \textbf{0.87} \\
Our Model with embeddings trained on Word2Vec & 0.82\\
Our Model with pre-trained Word2Vec embeddings & 0.59 \\
Mathur et al & 0.83 \\
\hline
\end{tabular}
\caption{Comparison of accuracy scores on HEOT dataset}
\label{tab:results_heot}
\end{table}
\begin{table}[]
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Model} & \textbf{Accuracy}\\ \hline
Davidson et al. & \textbf{0.90}\\
Our Model with embeddings trained on Glove & 0.89 \\
Our Model with embeddings trained on Word2Vec & 0.86 \\
Mathur et al. & 0.75 \\
\hline
\end{tabular}
\caption{Comparison of accuracy scores on \cite{davidson2017automated} dataset}
\label{tab:results_davidson}
\end{table}
\subsection{Classifier Model}
Both the HEOT and \cite{davidson2017automated} datasets contain tweets which are annotated in three categories: offensive, abusive and none (or benign). Some examples from the dataset are shown in Table \ref{tab:examples_tweets}. We use a LSTM based classifier model for training our model to classify these tweets into these three categories. An overview of the model is given in the Figure \ref{ml_model}. The model consists of one layer of LSTM followed by three dense layers. The LSTM layer uses a dropout value of 0.2. Categorical crossentropy loss was used for the last layer due to the presence of multiple classes. We use Adam optimizer along with L2 regularisation to prevent overfitting. As indicated by the Figure \ref{ml_model}, the model was initially trained on the dataset provided by \cite{davidson2017automated}, and then re-trained on the HEOT dataset so as to benefit from the transfer of learned features in the last stage. The model hyperparameters were experimentally selected by trying out a large number of combinations through grid search.
\section{Results}
Table \ref{tab:results_heot} shows the performance of our model (after getting trained on \cite{davidson2017automated}) with two types of embeddings in comparison to the models by \cite{mathur2018detecting} and \cite{davidson2017automated} on the HEOT dataset averaged over three runs. We also compare results on pre-trained embeddings. As shown in the table, our model when given Glove embeddings performs better than all other models. For comparison purposes, in Table \ref{tab:results_davidson} we have also evaluated our results on the dataset by \cite{davidson2017automated}.
\begin{figure}
\centering
\includegraphics[scale = 0.28]{lstm_model.jpg}
\caption{LSTM based model for tweet classification}
\label{ml_model}
\end{figure}
\section{Conclusion}
In this paper, we presented a pipeline which given Hinglish text can classify it into three categories: offensive, abusive and benign. This LSTM based model performs better than the other systems present. We also release the code, the dictionary made and the embeddings trained in the process. We believe this model would be useful in hate speech detection tasks for code-switched languages.
\fontsize{9.0pt}{10.0pt}
\selectfont
\bibliographystyle{aaai}
|
\section{INTRODUCTION}
In recent years, an increasing interest has been paid to metasurfaces (MSs); the planar arrays of well-patterned and optically thin scatterers distinguished by the induced electric and magnetic surface currents of their subwavelength engineered resonators, which are capable of abruptly changing the local state of light \cite{yu2011light,sun2012high,ding2017gradient,meinzer2014plasmonic,zhao2014recent,ra2015thin,chen2016review,epstein2016huygens}.
A reasonable dissipation loss owing to a relatively short interaction length, the ease of fabrication via planar lithographic processing due to the smaller physical footprint, the amenability for implementation of tunable devices thanks to their ultrathin flat configurations, and the compatibility with the widespread CMOS technology have made MSs more appealing than bulky metamaterial counterparts for tailoring the spatial, spectral, and temporal responses of light wavefronts in flat optics \cite{kuznetsov2016optically,arbabi2016multiwavelength,su2018advances,glybovski2016metasurfaces,arbabi2017planar,genevet2017recent,baranov2017all,forouzmand2018composite}.
To date, a bulk of research has been conducted on static MSs in which the geometry and compositions of their constituents are elaborately designed for a specific functionality that remain unchanged during the operation. However, to efficaciously harness MSs unlimited capabilities for the real-world applications such as adaptive optics, imaging, and microscopy, developing reconfigurable paradigms are becoming indispensable. For this reason, several approaches have been proposed in which external stimuli including chemical reactions, heat, elastic forces, magnetic fields, optical pumping, and electrostatic fields are applied to actively adjust the optical properties of MSs \cite{ferrera2017dynamic,makarov2017light}.
Among the aforementioned mechanisms, dynamic electro-optical metadevices incorporating active functional materials such as transition metal dichalcogenides (TMDCs), liquid crystal (LC), graphene, phase-transition materials (PTMs), highly doped semiconductors, and transparent conducting oxides (TCOs) show superiority in terms of the tuning range, response time, energy consumption, robustness, and pixel-by-pixel reconfiguration.
Nevertheless, fabrication complexities in the realization of defectless uniform TMDCs \cite{sun2016optical}, the surface anchoring in optically thick LC-based MSs \cite{franklin2015polarization,sautter2015active,chen2016electrically,su2017liquid,xie2017liquid,komar2018dynamic}, the highly inherent loss of graphene in the NIR and VIS regimes \cite{yao2013wide,smirnova2015tunable,abdollahramezani2015beam,dabidian2016experimental,huang2017dynamical,sherrott2017experimental}, the restricted tuning range and volatility of VO$_{2}$ \cite{wang2015switchable,liu2016hybrid,zhu2017dynamically,dong2018lithography}, and the low accessible index contrast of doped semiconductor-based MSs \cite{iyer2015reconfigurable,iyer2016electrically} introduce serious impediments in the realization of the maximum potential tuning.
Despite the advantages of TCOs, specifically indium tin oxide (ITO) as the most widely used one, including ultrafast modulation speed ($\sim$fs \cite{taghinejad2018hot}) and tunable electro-optical properties through the pre-/post-depositional processes \cite{howes2018dynamic,taghinejad2018ultrafast,huang2016gate,park2016dynamic,kafaie2018dual,forouzmand2018tunable}, the large unity-order index variation is achieved at the expense of dramatic dissipative losses within the exceptional epsilon-near-zero region restricted to the infrared spectrum. Furthermore, their ultra-thin ($\sim$1nm) inhomogeneous active layer strictly degrades the reflection efficiency, limiting their functionalities to the NIR regime. Overall, it is worth providing a clear comparison between the recently reported reconfigurable metadevices with different control mechanisms (Table \ref{table.comparisons}).
Since the light-matter interaction length and time are limited by the subwavelength pixel size of conventional MSs, and the quality factor of constituent resonators are moderate, inevitably, a drastic modulation of the complex refractive index is necessary to ensure considerable spectral shift of geometrical resonances which guarantees the full control over the impinging wavefronts at NIR and VIS spectrum ranges. Accordingly, the development of alternative functional materials which not only afford notable refractive index change upon excitation with external stimulus but also provide compact, inexpensive, and scalable platforms for high-speed and power-efficient architectures are highly desirable.
Chalcogenide-based phase-change materials (PCMs) as the spotlight of rewritable optical data storage disks and electronic flash memories offer a striking portfolio of properties. Amongst them, semiconducting germanium antimony telluride (Ge$_{2}$Sb$_{2}$Te$_{5}$), shortly GST, has lately garnered widespread interests in the emerging metaphotonics applications. Such an interest is due to the subwavelength scalability (down to the nm size), ultrafast switching speed (picosecond or less), high switching robustness (potentially up to 10$^{15}$ cycles), power efficient non-volatility, high thermal stability, and adaptability with the CMOS technology. More importantly, the refractive index of GST can be selectively controlled within the intermediate phases between two extreme states, namely amorphous and crystalline, rendering remarkedly different optical and electrical characteristics.
So far, most studies have investigated phase-change metadevices based on tailoring the amplitude response of incident light \cite{wuttig2017phase}, restricting the possible functionalities to the free space optical switching \cite{gholipour2013all,chen2013hybrid,rude2016ultrafast,alaee2016phase,ahmadivand2017optical,pogrebnyakov2018reconfigurable,qu2018polarization}, perfect absorption \cite{chen2015tunable,mkhitaryan2017tunable,tian2016visible}, imaging \cite{hosseini2014optoelectronic,rios2016color}, and thermal emission \cite{du2017control,qu2017dynamic}. Very recently, a growing attention has been drawn to the phase front manipulation with applications in beam steering and beam shaping \cite{chen2015engineering,colburn2017metasurfaces,yin2017beam,chu2016active,de2018nonvolatile}. The major drawbacks of conventional PCM-based metadevices are two fold. First, each designed structure is only capable of tailoring a specific wavefront dynamic (i.e., amplitude, phase, or polarization), minimizing the beam reconfigurability potential for real-world applications. Second, the post-fabrication modulation relies on the focused optical pulses or bulky thermal heater, as common time-controlled external stimuli, thus hinders the feasibility of device integration.
Herein, we leverage the multilevel non-volatile partial crystallization of GST and the underlying physics governing the operation of a gate-tunable meta-reflectarray (MRA) architecture to actively control the amplitude, phase, and polarization of an incident wavefront within a uniquely optimized configuration in the technologically relevant wavelength region of 1.55 $\mathbf{\mu}$m.
Indeed, the strong localized field enhancement provided by the excited plasmonic nanoantennas along with the large accessible refractive index of surrounding media in the view of reversible and reconfigurable GST material enable us to readily tailor the desired resonance response.
Then, by employing the concept of gradient MSs consisting of an array of well-engineered, geometrically similar plasmonic/GST nanoantennas, we will show the fully active modulation of the optical state of impinging light over subwavelength scales both spectrally and spatially. To validate the performance of our proposed approach in entirely controlling the wavefront, several reconfigurable, ultra-thin, and miniaturized nanophotonic devices with manifold functionalities will be demonstrated. Finally, we perform transient thermoelectric simulations to monitor the effect of temporal temperature change on the ultrafast conversion mechanism of GST upon applying the gate-voltage to the metadevice. Such a comprehensive study gives an extensive insight for the implementation of reconfigurable components favorable for a diverse applications in the next generation adaptive photonics and optoelectronics metasystems such as light detection and ranging (LiDAR), holograms, and modulators.
\section{STRUCTURE DESIGN, FUNCTIONAL MATERIAL, AND UNDERLYING PHYSICS}
\begin{figure} [b]
\centering
\includegraphics[trim=0cm 3.5cm 11.3cm 0cm,width=8.3cm,clip]{fig0.pdf}\\
\caption
{Three-dimensional illustration of the proposed electrically reconfigurable MRA. (a) The MRA structure consists of patterned gold ribbons on the GST nanostripes, which are separated from each other by SiO$_{2}$ spacers, deposited on a gold back-reflector. The whole structure is supposed to be resided on a silicon substrate with 500$~\mathrm{\mu}$m thickness. The gate voltage can be applied between the nanoantenna and the bottom gold to partially convert GST from amorphous to the crystalline states. (b) Cross-sectional view of the meta-atom with the periodicity of p = 550 nm. The width and thickness of each nanoantenna are w=340 nm and t=30 nm, respectively, and the thickness of GST nanostripes is h=180 nm.
}
\label{fig.3D perspective}
\end{figure}
Figure \ref{fig.3D perspective} depicts the structure of the proposed electrically tunable MS as well as a cross section of its unit cell, namely a meta-atom. The structure is comprised of a 1D array of periodic gold/GST nanoantennas, which are separated by SiO$_{2}$ spacers, deposited on an optically thick gold back-reflector. Here, the pivotal structural parameters including the GST nanostripes height (d) and width (w) as well as the lattice constant (p) of the array are well optimized to obtain arbitrary reflection characteristics via dynamic control over the optical properties of the structure granted by selectively adjusting the crystalline phase of GST nanostripes. To do so, nanoantennas can be connected individually or as a cluster (associated with the desired functionalities) to separate in-plane electrodes. Then, electrical gating across the top nanoribbon and the bottom back-reflector is realized to trigger the thermal annealing and melt-quenching processes necessary for implementation of multilevel crystallized states (see Methods).
The ellipsometrically measured complex permittivity of GST in amorphous and crystalline states are shown in Figure \ref{fig.GST dielectric function} \cite{shportko2008resonant}. While the large dielectric function of the crystalline state stems from highly polarizable delocalized resonant bonding, the lower dielectric function of the amorphous phase is mostly ascribed to covalent bonds \cite{wong2010phase,wuttig2017phase}. Transition between the two abovementioned extreme phases is practically accomplished by an external stimulus, with specified power for a predefined time duration, which generates localized Joule heating in the GST nanostripe (see Methods). Accordingly, in the intermediate states, GST can be viewed as an arbitrary spread of amorphous and crystalline elements \cite{raoux2009phase,loke2012breaking}. Among several effective-medium theories proposed for the description of the dielectric function of heterogeneous media, we utilize the well-known Lorentz-Lorenz relation \cite{chu2016active} to approximate the effective permittivity of partially crystallized GST as follows:
\begin{equation}
\frac{\epsilon_{eff}(\lambda)-1}{\epsilon_{eff}(\lambda)+2}={L_{c}}\times\frac{\epsilon_{c}(\lambda)-1}{\epsilon_{c}(\lambda)+2}+({L_{c}}-1)\times\frac{\epsilon_{a}(\lambda)-1}{\epsilon_{a}(\lambda)+2}
\label{Lorentz-Lorenz}
\end{equation}
where for a specific wavelength, $\epsilon_{c}(\lambda)$ and $\epsilon_{a}(\lambda)$ are the permittivity of crystalline and amorphous GST, respectively, and ${L_{c}}$, ranging from 0 (amorphous) to 1 (crystalline), is the crystallization fraction of the GST.
To analyze the proposed hybrid phase-change MS, numerical simulations were conducted by finite-element method (FEM) and finite-integral technique (FIT) with realistic material parameters (see Methods). Figure \ref{fig.reflection and profiles} depicts the scattering properties of the proposed meta-atom with h=180 nm , w=340 nm, and p=550 nm assuming the GST nanostripe crystallization level (${L_{c}}$) varies from 0 (as-deposited) to 100$\%$ (fully crystalline), when the structure is excited by a normal propagative transverse magnetic (TM) polarized light.
\begin{figure}[t]
\centering
\begin{tabular}{@{}cccc@{}}
\includegraphics[trim=0cm 6cm 0cm 0cm,width=8.3cm,clip]{fig22.pdf}\\
\includegraphics[trim=0cm 11.5cm 13cm 0cm,width=8.3cm,clip]{fig23.pdf}\\
\end{tabular}
\caption{Reflection response of the proposed tunable MS. (a) Reflection amplitude spectra of meta-atom for several crystallization fraction. Upon partial to ultimate conversion of GST nanostripe, the SR-SPP ($\lambda_{1}$) and PR-SPP ($\lambda_{2}$) resonance wavelengths can be gradually tuned resulting in different reflectivity. (b) Corresponding reflection phase spectra of the meta-atom. The wavelength region in which the fundamental modes of the structure communicate with each other is highlighted with the gray color. Magnetic field intensity profiles of the (c) SR-SPP, and (d) PR-SPP modes excited within the cross-sectional view of the meta-atom with $L_{c}^{60}$. }
\label{fig.reflection and profiles}
\end{figure}
As shown in the reflection amplitude and phase spectra of the MS in Figures \ref{fig.reflection and profiles}(a) and (b), the MS structure has two fundamental resonant modes, where the interaction of both resonance modes with the incident wave is considerably affected by changing the level of crystallization in the GST nanostripes. Figure \ref{fig.reflection and profiles}(c) and (d) show the mode profile of the two resonant modes of the structures at the GST intermediate crystallization factor $L_{c}^{60}$. As it can be seen from the mode profiles, the long wavelength mode corresponds to the local resonance of the lateral plasmonic Fabry-Pero$\mathrm{\acute{t}}$ (F-P) mode of the gold/GST nanoantenna, namely the short-range surface plasmon polariton (SR-SPP) working in the over-coupling regime. Also, the short wavelength mode is the resonance of the 1D grating structure at propagating wavevector (k=0), called propagative surface plasmon polariton (PR-SPP) operating in the under-coupling condition.
Upon illumination with a transverse magnetic (TM) polarized light, i.e., SR-SPP, a tightly-confined slow propagative electromagnetic wave coupled to the collective electron oscillations of metal, will be excited at the nanoribbon terminations and begins propagating back and forth between the two end-faces \cite{barnard2008spectral}. When SR-SPP experiences the truncations, it partially reflects and looses a fraction of its energy to the free-space scattered field. A nanoribbon with an appropriately chosen width can be treated as an F-P resonator generating a constructive interference between forward and backward traveling SR-SPP modes leading to a highly confined, higher mode-index resonant field adjacent to the nanoribbon. The resonance condition is given by \cite{schuller2010plasmonics}:
\begin{equation}
\omega_{\textrm{res},m}=\frac{m\pi-\phi(\epsilon_{\textrm{eff}})}{2\pi}\frac{\lambda_{0}}{n_{\textrm{SR-SPP}}(\epsilon_{\textrm{eff}})}
\label{resonance condition}
\end{equation}
where ${n_{\textrm{SR-SPP}}(\epsilon_{\textrm{eff}})}$ is the propagative SR-SPPs mode index provided by a metal film with the same thickness as the nanoribbon, $m$ is a positive integer indexing the resonance order, and $\phi(\epsilon_{\textrm{eff}})$ is a phase change due to the reflection at end-faces. It should be underlined that both $\phi(\epsilon_{\textrm{eff}})$ and ${n_{\textrm{SR-SPP}}(\epsilon_{\textrm{eff}})}$ are functions of metal and ambient dielectric permittivities. Although eq. \ref{resonance condition} provides a physical insight for explaining the resonance condition, anticipating the exact position of resonant scattering is dependent on $\phi(\epsilon_{\textrm{eff}})$ which is complicated to be evaluated, especially when the surrounding material is dispersive and lossy. It is worth mentioning that decreasing the gap size pushes the SR-SPP to confine more tightly to the metal surfaces, leading to the domination of the absorption. Meanwhile, due to the large refractive indices of multistate GST, by decreasing the gap size only a small portion of the incident light couples to gap surface plasmons. As a result, the electromagnetic field enhancement is so weak that by modifying the crystallization fraction, a negligible change in the reflection spectra is expected. To overcome this issue, the design of GST nanostripe with an optimized height is indispensable; Figure \ref{fig.reflection and profiles}(c) demonstrates the coupling between third-order (m=3) SR-SPPs of the gold nanoribbon and the enhanced SR-SPPs-like mode excited on the surface of back-reflector which results in the enhanced light-matter interaction in the GST gap nanostripe. Here, the absorption mechanism can be explained by the Ohmic losses in the nanoribbon and the back-mirror as well as the energy absorbed in the GST nanostripe itself upon multiple reflections of the standing wave between subsequent dielectric spacers.
The origin of the second absorption dip, i.e., $\lambda_{2}$, can be predicted according to the grating equation:
\begin{equation}
{\frac{\omega}{c}\textrm{sin}(\theta)\pm{n\frac{2\pi}{p}}=\pm{\beta}}
\label{grating condition}
\end{equation}
where $\theta$ is the incidence angle, n is the grating diffraction order, and $\beta$ is the wavenumber of the propagating mode along the interface of the gold substrate. A simple inspection of the symmetric mode profile in Figure \ref{fig.reflection and profiles}(d) confirms such a mode roots in the two-way propagation of the excited SPPs (PR-SPPs) which tunnel into the GST nanostipe. It is worth mentioning that the diffraction from the sharp edges of individual nanoribbons can provide the extra momentum required to generate PR-SPPs under the normal excitation \cite{pors2013efficient,doi:10.1021/acsphotonics.7b00959}. Moreover, Figure \ref{fig.reflection and profiles} (c) and (d) verifies that PR-SPPs do not have similar enhancement and localization degree of SR-SPPs primarily because nanoribbons only perturb the incoming light with slightly confined field in the GST nanostripe. In this case, the absorption phenomenon is mostly related to the Ohmic losses of the metallic back-reflector. In the Supplementary section, we justify the presence of the aforementioned modes by further investigating the evolvement of resonances upon modification of structures geometrical parameters and the angle of the excitation.
Considering the characteristics of two fundamental modes, we carry out full-wave simulations to investigate the evolution of the complex reflection coefficient representing the optical behavior of metastructures. For the sake of brevity, we assume the MRA as a closed optical system that limits the energy exchange through a channel. Basically, the total reflected complex field is governed by the non-resonant and resonant responses of the structure represented by \cite{park2016dynamic}
\begin{equation}
{r_{\textrm{tot},L_{c}}=r_{\textmd{non-res},L_{c}}\textmd{exp}(i\theta_{\textmd{non-res},L_{c}})+r_{\textmd{res},L_{c}}\textmd{exp}(i\theta_{\textmd{res},L_{c}})}
\label{non-resonant and resonant responses}
\end{equation}
in which the subscripts $L_{c}$ denotes the level of the crystallization, $\textmd{non-res}$ and $\textmd{res}$ stand for non-resonant and resonant, respectively. According to Figure \ref{fig.polar plot}, the non-resonant term attributed to the direct reflection from the metallic back-reflector is almost being intact for all levels of crystallization, i.e., $r_{\textmd{non-res},L_{c}}\textmd{exp}(i\theta_{\textmd{non-res},L_{c}})\approx-1$, which is reasonable due to the high impedance characteristic of the gold in NIR. In other words, the reflection phase remains constant, i.e., $180^{\circ}$ owing to the non-resonant effect, which can serve as a reference in the future study.
However, changing the crystalline state of the GST nanostripe significantly impacts the resonant response. Figure \ref{fig.polar plot} demonstrates that upon conversion of GST from amorphous to the crystalline state, the reflection phase evolves from 208$^{\circ}$ to -107$^{\circ}$ while the scattering cross section first decreases and then increases.
Comparing the electric field profiles with the reference profile reveals that the system is mostly in an over-coupling state, in which the resonant response dominates the non-resonant one, near amorphous state (bluish dots), then approaches the under-coupling regime (reddish dots), where the non-resonant response overcomes the resonant one, and finally comes back to the over-coupling zone (greenish dots) when approaches the crystalline state. Such fundamental regime exchange around the operational wavelength, i.e., 1550 nm, due to material phase change results in a broadband response (see Figure \ref{fig.reflection and profiles}(b)) which makes the proposed structure unique.
\begin{figure} [t]
\centering
\includegraphics[trim=12cm 4cm 0cm 0cm,width=8.3cm,clip]{fig3.pdf}\\
\caption
{Polar plot representation for the trajectory of the total complex reflection coefficient with increasing the crystallization level. The MRA moves form the over-coupling condition in its near amorphous phase (bluish dots) towards the under-coupling state in the intermediate phases (reddish dots), and it again enters the over-coupling regime in crystalline-like phases (greenish dots). The non-resonant reflection responses accommodated in the center grayish box confirms the negligible variation of both amplitude and phase for all crystallization levels. To compare the phase shift between different crystallization levels, the corresponding electric field profiles of resonant and non-resonant reflected light are illustrated. $L_{c}^{n}$ represents the crystallization fraction of $n\%$ for the GST nanostripe of each resonator.
}
\label{fig.polar plot}
\end{figure}
\begin{figure} [t]
\centering
\includegraphics[trim=12.5cm 0cm 0cm 0cm,width=8.3cm,clip]{fig4.pdf}\\
\caption
{Amplitude and phase evolution of the hybrid meta-atom upon conversion from the amorphous to crystalline state. (a) Normalized reflected $E_{x}$ field patterns from the MRA for the eleven crystallization levels shown in Figure \ref{fig.eleven crystallization levels}(b). Each strip is calculated by an independent full-wave numerical simulation under the illumination of a normally x-polarized plane wave with $\lambda=1550$ nm. The gray dashed line defining the wavefront verifies the achievable total phase shift of $315^{\circ}$. (b) Evolution of reflection amplitude (left axis) and phase shift (right axis) as a function of the crystallization level. Inset conceptual schematic depicts the gradual non-volatile phase transition of GST molecules from the amorphous to the fully crystalline.
}
\label{fig.eleven crystallization levels}
\end{figure}
\section{TOWARDS IMPLEMENTATION OF A HIGHLY RECONFIGURABLE METADEVICE}
Figures \ref{fig.eleven crystallization levels}(a) and (b) illustrate the proposed meta-atom, which enables the control of the reflected wave phase over a wide dynamic range of 315$^{\circ}$ while preserving the reflection amplitude above 15\%. By selectively modifying each meta-atom leveraging the tunable gate-voltage approach (see Methods), a broad range of patterns can be imprinted to the incident light by the MRA. Here, we are interested in the design and realization of three types of functionalities within a uniquely optimized structure in which: i) the phase front of light is processed, e.g., beam focusing, based on geometrical optics concepts and the generalized law of refraction, ii) the amplitude response of the wavefront is manipulated, e.g., switching, by using the impedance theory, and iii) polarization property of the beam is tailored, e.g., circular polarization conversion, by employing Jones calculations.
To encode the phase, amplitude, or polarization response associated with each specific functionality, we first discretize the related transverse distribution (defined by a unique equation) on the surface of MRA and then extract a sufficient sequence of samples. Indeed, each sample represents the reflection value of the corresponding meta-atom with an appropriate crystallization level chosen from Figure \ref{fig.eleven crystallization levels}(b). By judiciously arranging those meta-atoms in the lateral dimension, the structure mimics the behavior of the corresponding function. It is notable that all geometrical dimensions remain unchanged throughout the paper to highlight the effectiveness of our proposed architecture in manipulating the incident light dynamics in the desired way just by adjusting the external stimulus.
\subsection{ON THE ACTIVE MANIPULATION OF PHASE DISTRIBUTION}
Among a plethora of optical functionalities which can be realized via the phase-front engineering such as focusing, steering, and holography, we focus on the design and implementation of an efficient, reconfigurable, high numerical aperture, and diffraction-limited metalens. Following the equal optical path principle, to make a linearly polarized light converge at a specific focal length, a spatially varying phase shift needs to be imprinted to the impinging wavefront along the transverse direction of the metalens. The required quadratic phase profile follows \cite{cheng2014wave}
\begin{equation}
\phi(x)=\dfrac{2\pi}{\lambda}[\sqrt{f^{2}+x^{2}}-f]
\label{lens}
\end{equation}
where $f$ is the focal length and $\lambda$ is the wavelength in the background medium. In Figure \ref{fig.focusing}(a), the theoretical phase distribution for $f=20\lambda$ (solid blue line) is compared with the phase profile realized in the full-wave simulation (red circle-shaped marks). Although the discretization is limited to eleven steps (corresponding to the experimentally-demonstrated number of crystallization levels \cite{rios2015integrated}), the electric field intensity distribution in Figure \ref{fig.focusing}(b, top) illustrates that focusing occurs at a distance of $20.5\lambda$ in a very good agreement with the analytical result. Moreover, the normalized field intensity at the focal plane has the full-width at half-maximum (FWHM) of 995 nm and numerical aperture NA=0.65 justifying the diffraction-limited feature of the designed lens (see Figure \ref{fig.focusing}(c)). Although for such a small size lens ($\approx35\lambda$), the efficiency (i.e., the ratio of reflected power at the focal plane with lateral dimension of three times FWHM to the reflected power from a back-reflector) is 5.7\%. It should be emphasized that a larger aperture size results in a performance improvement due to the local periodicity effect retention. To reveal the superiority of our proposed reconfigurable metadevice, we investigate two more scenarios in which the focal point of metalens is controlled just by carefully assigning the GST crystallization profile. Electric field intensity profiles for metalenses with f=$10\lambda$ and f=$5\lambda$ are depicted in Figures \ref{fig.focusing}(b, middle) and (b, bottom), respectively. In these designs, simulated focal lengths are f=$10.3\lambda$ (with FWHM=740 nm and NA=0.86) and f=$5.2\lambda$ (with FWHM=615 nm and NA=0.96), respectively. For the sake of clarity, the realized phase profile, the required lateral crystallization distribution, and the calculated depth of focus for the investigated lenses are illustrated in Figure \ref{fig.focusing1}. All numerical simulations are in a good agreement with the analytical results while negligible discrepancies in the performance primarily originate from in-plane coupling between the elements, imprecise discretization of phase profiles, the abrupt phase shift in adjacent meta-atoms, and the finite size of the metalenses. In the Supplementary section, we investigated other optical functionalities, such as phased array antenna, airy beam generator and beam splitter, for the on demand tailoring of the phase front of light.
\begin{figure} [t]
\centering
\includegraphics[trim=7.5cm 0cm 0cm 0cm,width=8.3cm,clip]{fig5.pdf}\\
\caption
{Reconfigurable metalens with diverse focal lengths. (a) The desired (solid blue line) and realized (red circle-shaped marks) phase discontinuities to implement a metalens with the focal length of 20$\lambda$. (b) Calculated intensity profiles of the reflected beam by the tunable metalens in the transverse cross section for nominal focal lengths of (top) f=20$\lambda$, (middle) f=10$\lambda$, and (bottom) f=5$\lambda$. (c) Comparison of the Airy disk intensity (solid black line) with the square normalized amplitude of the electric field calculated at the distances of f=20$\lambda$, f=10$\lambda$, and f=5$\lambda$ (dashed lines). The lateral dimension of lenses is D=100p, and dashed horizontal lines indicate the focal plane. The impinging linearly-polarized light wavelength is $\lambda=1550$ nm.
}
\label{fig.focusing}
\end{figure}
\subsection{CONTROL OVER THE AMPLITUDE RESPONSE}
Further study on our proposed multifunctional meta-atom reveals that by uniformly controlling the crystallization level of GST along the MRA, the reflectance spectra can be independently control in a dynamic way, which enables several applications such as switching and perfect absorption. Essentially, such a freedom allows the implementation of ultra-fast and compact optical amplitude modulators. Delving further, the common drawback of most optical modulators is their large intrinsic insertion loss (IL), which is the device loss in the ON-state. On the other hand, for any modulator, the highest possible extinction ratio (ER), defined as the difference in loss between the ON-state and OFF-state of the modulator, is necessary to reduce the bit error rate \cite{yao2013wide,kruger2012design}. However, there is a tradeoff between the IL and the attainable ER meaning that a judiciously engineered architecture is always preferred to make a reasonable compromise between these parameters. In order to fulfill these criteria, we harness the gate-tunable meta-atom working in two distinct crystallization levels of $L_{c}^{0}$/$L_{c}^{60}$ corresponding to the ON/OFF states, respectively. Figure \ref{fig.modulator} illustrates that our proposed MRA notably increases the effective figure-of-merit (FOM) and modulation depth (MD) formulated as
\begin{align}
\textrm{FOM}&=\dfrac{\textrm{ER}}{\textrm{IL}}=\dfrac{-10\textrm{log}_{10}({|r_\textrm{min}(\lambda)|}/{|r_\textrm{max}(\lambda)|})}{-10\textrm{log}_{10}({|r_\textrm{max}(\lambda)|})}
, \nonumber \\
\textrm{MD}&=1-\dfrac{{|r_\textrm{min}(\lambda)|^{2}}}{{|r_\textrm{max}(\lambda)|^{2}}}
\label{equ1}
\end{align}
where $r_\textrm{min}(\lambda)$ and $r_\textrm{max}(\lambda)$ are the reflectivity of the modulator in the ON-state and OFF-state for each wavelength, respectively. Due to the improved IL and enhanced ER, a significant FOM and MD are obtained over a considerable wavelength range, verifying the high performance of our optimized modulator. It is noteworthy that high ER, i.e., ER$>$7 dB, is preferable for most applications such as data-rate interconnects and high-sensitivity sensing; however, ER$\sim$4 dB can be sufficient for specific applications such as short-distance data transmission.
\begin{figure} [t]
\centering
\includegraphics[trim=0cm 6.7cm 0cm 0cm,width=8.3cm,clip]{fig10.pdf}\\
\caption
{High performance behavior of the optical amplitude metamodulator. (a) The IL and ER diagrams for the proposed modulator in which ON- and OFF-states are corresponding to the $L_{c}^{60}$ and $L_{c}^{0}$, respectively. (b) The FOM and MD curves achieved from numerical simulations justify the unprecedented response of the modulator in the telecommunication band.
}
\label{fig.modulator}
\end{figure}
\subsection{TUNING THE POLARIZATION STATE OF INCIDENT BEAM}
Finally, we leveraged the relative phase difference between two orthogonal polarization states to realize an active optical waveplate converting an incoming arbitrarily polarized beam to the general form of elliptically polarized light. We consider the Jones matrix representation of the MS as
\begin{align}
J_\textrm{MS}=
\begin{bmatrix}
a~\textmd{exp}(i\phi_{x}) & 0 \\
0 & b~\textmd{exp}(i\phi_{y})
\end{bmatrix}
\label{equ9}
\end{align}
where $a$ and $b$ are the reflection amplitudes, and $\phi_{x}$ and $\phi_{y}$ are the reflection phases of the horizontally and vertically polarized light, i.e., perpendicular and parallel to the grating's direction. According to Figure \ref{fig.polar plot}, for all crystallization levels $b$ and $\phi_{y}$ can be considered 1 and $\pi$, respectively. As a result, reflected output fields can be calculated as
\begin{align}
\begin{bmatrix}
E_\textrm{H,out} \\
E_\textrm{V,out}
\end{bmatrix}
&=
\begin{bmatrix}
a~\textmd{exp}(i\phi_{x}) & 0 \\
0 & \textmd{exp}(i\pi)
\end{bmatrix}
\begin{bmatrix}
A~\textmd{exp}(i\phi_\textrm{in}) \\
\textmd{exp}(i\pi)
\end{bmatrix} \nonumber
\\
&=
\begin{bmatrix}
aA[\textmd{exp}(i(\phi_{x}+\phi_\textrm{in}))] \\
1
\end{bmatrix}
\label{equ10}
\end{align}
in which $A$ and $\phi_\textrm{in}$ are the amplitude and phase of incident wave, respectively. Based on the properly chosen state of input polarization and the status of the reflected x-polarized light, the MS generally behaves as a reflective elliptical polarizer. However, by satisfying particular conditions, several interesting scenarios are expected where: 1) $\phi_{x}+\phi_\textrm{in}=\pi/2$ and $A=a^{-1}$ which results in the left-handed circularly polarized light at the output, 2) $\phi_{x}+\phi_\textrm{in}=3\pi/2$ and $A=a^{-1}$ leading to right-handed circularly polarized light at the output, and 3) $\phi_{x}+\phi_\textrm{in}=\pi$ or $2\pi$ which gives rise to a linear-polarized light at the output. Figure \ref{fig.polarization} comprehensively represents the relative reflectivity ($|r_{x}|/|r_{y}|$) and the phase difference ($\phi_{x}-\phi_{y}$) between the two orthogonally polarized light as a function of the crystallization level over a wide spectral range. These cases justify the strength of our proposed MRA in engineering the polarization state of incident light for potential applications in photography, biosensing, and communication applications.
\begin{figure} [t]
\centering
\includegraphics[trim=0cm 0cm 9.5cm 0cm,width=8.3cm,clip]{fig11.pdf}\\
\caption
{Active MSs for the demonstration of the polarization conversion. The reflection amplitude ratio ($|r_{x}|/|r_{y}|$) and the relative phase ($\phi_{x}-\phi_{y}$) between the two orthogonally polarized light as a function of the crystallization level versus the wavelength. Different polarization states can be realized in the output by moving on the surface.
}
\label{fig.polarization}
\end{figure}
\section{CONCLUSIONS}
In summary, we have theoretically investigated the capability of an optimized hybrid plasmonic-GST architecture in manipulating the phase, amplitude, and polarization of incident beam both spectrally and spatially. We leveraged the two highly confined fundamental modes, namely SR-SPP and PR-SPP, and stark refractive index contrast granted by the successive stages of phase-change of GST to fully control the dynamics of light. Accordingly, we have demonstrated a uniquely reconfigurable structure to realize diverse optical functionalities, such as beam focusing, amplitude modulation, and polarization conversion, just by tunning the crystallization level of the constituent meta-atoms with the easy-to-handle electrical gate tunning. Due to the simultaneously significant modulation of amplitude and phase, such miniaturized, efficient, active, and high-speed MSs offer superior platforms for realization of practical metadevices and even metasystems necessary for the on-demand applications such as holograms, adaptive optics, and LiDAR.
\clearpage
\newpage
\newcommand{\beginsupplement}{%
\renewcommand{\thetable}{S\arabic{table}}%
\renewcommand{\thefigure}{S\arabic{figure}}%
\renewcommand{\thesection}{S\arabic{section}}%
\renewcommand{\thesubsection}{S\arabic{section}.\arabic{subsection}
}
\setcounter{figure}{0}
\setcounter{section}{0}
\setcounter{subsection}{0}
\setcounter{table}{0}
\section*{SUPPLEMENTARY INFORMATION}
\beginsupplement
\begin{figure} [H]
\centering
\includegraphics[trim=8.7cm 0cm 0cm 0cm,width=8.3cm,clip]{fig1.pdf}\\
\caption
{Dielectric function of GST. Real (blue on left axis) and imaginary (red on right axis) parts of the permittivity for amorphous GST (a-GST, solid lines) and crystalline GST (c-GST, dashed lines) states depending on the wavelength. Inset shows the refractive index of for a-GST and c-GST at the operational wavelength, i.e., $\lambda$=1550 nm.
}
\label{fig.GST dielectric function}
\end{figure}
\begin{figure} [b]
\centering
\includegraphics[trim=8.5cm 0cm 0cm 0cm,width=8.3cm,clip]{fig55.pdf}\\
\caption
{The desired (solid blue line) and realized (red circle-shaped marks) phase discontinuities to implement metalenses with the focal lengths of (a) 10$\lambda$ and (b) 5$\lambda$. (c) The normalized intensity distribution at y=0 and corresponding depth of focus (DoF) for the the investigated metalenses. (d) The required crystallization level distribution along the lateral direction for three defined cases. The lateral dimension of lenses is D=100p and dashed horizontal lines indicate the focal plane. The impinging linearly-polarized light wavelength is $\lambda=1550$ nm.
}
\label{fig.focusing1}
\end{figure}
\section{Parametric study of fundamental modes}
To get a better understanding on the relation of the dominant resonances with the structural parameters, a parametric study is conducted in which the reflectance of the proposed MRA is investigated as a function of nanoantenna width, lattice constant, intermediate layer height, and incidence angle versus the frequency. Figure \ref{fig.dependence}(a) illustrates upon variation of gold nanoribbon width around the optimized value, i.e. w=340 nm, a considerable shift happens for the SR-SPP mode, as anticipated by eq. \ref{resonance condition}, while the second resonance almost remains intact. On the other hand, according to Figure \ref{fig.dependence}(b), the PR-SPP mode is highly dependent on the structure periodicity, which was predicted by eq. \ref{grating condition}. It is more obvious that both modes behave similarly in case the height of intermediate layer slight changes around the resonance condition (see Figure \ref{fig.dependence}(c)). Actually, the thick layer not only hinders the interaction of SR-SPP mode on the gold nanoribbon and substrate in the GST gap feed but also disturbs the PR-SPP by weakening the light perturbation. The nature of the two modes is well clarified within Figure \ref{fig.dependence}(d) which illustrates by modifying the angle of incidence, PR-SPP is affected more than the SR-SPP mode.
\begin{figure} [t]
\centering
\includegraphics[trim=0cm 0cm 8cm 0cm,width=8.3cm,clip]{fig24.pdf}\\
\caption
{Dependence of the dominant SR-SPP and PR-SPP modes on the structural parameters and excitation incidence angle. 2D reflection amplitude color maps for different (a) gold nanoribbon widths, (b) MRA lattice constants, (c) heights of the GST nanostripe, and (d) angles of incident beam versus the frequency.
}
\label{fig.dependence}
\end{figure}
\section{Phased array antenna}
Phased array principle is a powerful tool to systematically design an array of antenna realizing a complicated radiation pattern in the far-field; that is the total far-field pattern of the array can be calculated by multiplying the individual element pattern and pattern of the designated array, so-called array factor (AF), in which the elements are replaced by isotropic radiators \cite{stutzman2012antenna}. Figure \ref{fig.phased array antenna} represents a uniformly arranged, linear array of N elements which are driven by phase shifters and attenuators in a general case. In our case, each element is indeed a super cell which includes $m$ ($m$ is integer) consecutive nanoantennae with period $P$ organized in the $y$ direction. We assume that the progressive phase-shift between two adjacent super cells is $\alpha$, and the reflection amplitude of all super cells are the same ($A_{n}$=$A_{0}$ for $n=1, 2, 3, ...$). Accordingly, AF for an equally spaced N-element array antenna is given by:
\begin{align}
\textrm{AF}&=A_{0}\sum_{n=0}^{N-1} e^{jn\psi}=A_{0}(1+e^{j1\psi}+e^{j2\psi}+...+e^{j(N-1)\psi}) \nonumber \\
&=A_{0}\frac{1-e^{jN\psi}}{1-e^{j\psi}}=A_{0}e^{j(N-1)\psi/2}\frac{\textrm{sin}(N\psi/2)}{\textrm{sin}(\psi/2)}
\label{phased array antenna1}
\end{align}
and the normalized AF is calculated as:
\begin{align}
\textrm{AF}_\textrm{normalized}=\frac{\textrm{sin}(N\psi/2)}{N\textrm{sin}(\psi/2)}
\label{phased array antenna2}
\end{align}
which shows AF is a $2\pi$-periodic function with $\psi={\beta}d\textrm{sin}(\theta)+\alpha$, where $\beta=2\pi/\lambda$, and $\alpha$ is the key parameter enables more potent control over scanning of the reflected beam. For a given $\psi$, such a formulation allows us calculate the maximum value of the pointing direction in terms of the polar angle. Considering the visible region is the upper hemisphere, i.e. $0<\theta<180$, thus the acceptable range for $\psi$ is: $\alpha-\beta d<\psi<\alpha+\beta d$. In our case, the progressive phase shift is chosen $\alpha=180$, determined by crystallization fraction of super cells, and each super cell contains $m=2,3,$ or $4$ similar meta-atoms which results in element spacing of $d=2P$, $d=3P$, or $d=4P$, respectively. Figure \ref{fig.phased array antenna}(b) and (c) clarify a simple procedure to relate the normalized array factor for 61 super cells as a function of $\psi$ in cartesian coordinates to the corresponding polar plot as a function of $\theta$. The intersection of vertical dashed lines and semi-circles with radius $\beta{d}$ (highlighted with dots) which are located at distance $\alpha$ from the origin, defines the reflection angle of the main beam, i.e., $\theta=\textrm{sin}^{-1}[\alpha/(\beta d)]$. According to Figure \ref{fig.eleven crystallization levels}, $L_{c}^{40}$ and $L_{c}^{80}$ are the two optimal values for crystallization levels providing two different dynamics with $180^{\circ}$ phase difference and almost the same amplitude. As a result, by choosing a cluster of reconfigured meta-atoms in terms of crystallization fraction in each super cell, and arranging these super cells in a 2D planar configuration, we can tailor the radiation pattern of the incident beam in the desired way. Figure \ref{fig.phased array antenna}(c) corroborates good agreement between the angle of diffraction achieved from full-wave simulations (represented by lobes) and the analytical results (represented by intersection dots), based on eq. \ref{phased array antenna2}, for three different configurations. From Figure \ref{fig.phased array antenna}(d), the total electric field distribution upon excitation with a TM-polarized light at $\lambda$=1550 nm clearly pictures the progressive phase-shift of $180^{\circ}$ in the transverse direction.
\begin{figure} [t]
\centering
\includegraphics[trim=0cm 1.5cm 8cm 0cm,width=8.3cm,clip]{fig8.pdf}\\
\caption
{Representation of optical phased array antenna. (a) Schematic of a general phased array antenna capable of controlling beam propagation in free space. (b) The array factor for 61 super cells as a function of $\psi$. The profile around zero is enlarged in the inset. (c) The corresponding polar patterns for three specific schemes which differ in terms of number of meta-atoms accommodated in a super cell. Dots represent the exact splitting angles achieved from analytical results while reflected beams from the designed MRA are depicted with lobes. (d) Simulated electric field intensity of the reflected wave (top) for three distinct cases investigated in panel (c) and the gate voltage arrangement required for each scenario (bottom).
}
\label{fig.phased array antenna}
\end{figure}
\section{Airy beam generator}
Fundamentally, for any incident light beam with spatially varying field distribution, a designer MS can be employed to tailor the beam field envelopes. Recently, Airy beams have garnered significant research interests thanks to their unique properties including self-healing, non-diffracting, and self-bending \cite{ding2017multiwavelength}. In fact, Airy wave packets establish a particular class of waves accelerating along parabolic trajectories. They suggest novel opportunities such as three dimensional super resolution imaging, optical interconnects, and optical micromanipulation applications \cite{song2017compact}.The phase and amplitude distributions for a 1D Airy beam generator can be described as \cite{minovich2011generation}
\begin{align}
\phi_\textrm{airy}(\xi,s)=&\textrm{Ai}\bigg[s-\bigg(\dfrac{\xi}{2}\bigg)^{2}+ia\xi\bigg]\text{exp}\bigg[as-\bigg(\dfrac{a\xi^{2}}{2}\bigg) \nonumber
\\&-i\bigg(\dfrac{\xi^{3}}{12}\bigg)+i\bigg(\dfrac{a^{2}\xi}{2}\bigg)+i\bigg(\dfrac{s\xi}{2}\bigg)\bigg]
\label{Airy beam}
\end{align}
where Ai($\cdot$) represents the Airy function, $s=x/x_{0}$ is a dimensionless transverse coordinate in which $x_{0}$ is the half width of the main lobe, $\xi=z/kx_{0}^2$ is a normalized propagation distance in which $k_{0}=2\pi/\lambda$ is the propagation constant in free space, and $a\ll1$ is a positive parameter to obtain truncated Airy beam distribution with finite power. Considering $a=0.05$ and $x_{0}=2.875~\mu$m, the initial field envelope of 1D Airy beam satisfies $\phi_\textrm{airy}(0,s)=\textrm{Ai}(s)\text{exp}(as)$ on the z=0 plane whose amplitude and phase profiles are illustrated in Figures \ref{fig.Airy beam}(a) and (b), respectively. The envelope of Airy beam depicts an exponential oscillating behavior associated with alternating positive maxima and negative minima; consequently, its phase distribution represents a periodic segments with $180^{\circ}$ phase difference. However, simultaneous realization of ideal phase and amplitude profiles is challenging since these two features are interrelated according to Figure \ref{fig.reflection and profiles}. Fortunately, a simplified yet practical approach is just assuming the field amplitude constant and only modulate the phase profile along the transverse direction. Such an accurate approximation stems from the fact that Airy beam's exceptional self-healing feature naturally retrieves the field profile along the propagation path even in the presence of a nonuniform obstacle (see inset of Figure \ref{fig.Airy beam}(c)). The simulated field profile of reflected Airy beam plotted in Figure \ref{fig.Airy beam}(c) shows clearly its special non-diffracting and self-bending features. Furthermore, its main lobe well follows the targeted trajectory obtained from the analytical formulation shown with white dashed curve. It is noteworthy that introducing an additional linear compensating phase across the MS enables us to control radiation pattern of an Airy beam to specific direction. Accordingly, based on generalized Snell's law, the total reflection phase shift can be derived as
\begin{align}
\varphi_\textrm{reflection}(x,\theta)=\textrm{arg}(\phi_\textrm{airy}(0,s))-k_{0}\textrm{sin}(\theta)x+2n\pi
\label{Snell's law Airy beam}
\end{align}
where $\theta$ is the reflection angle of Airy beam and $n$ is an integer. Figures \ref{fig.Airy beam}(d) and (e) illustrate propagation of the truncated Airy beam with three distinct phase profiles associated with different reflection angles, i.e., $\theta=-10^{\circ}$ and $10^{\circ}$ based on eq. \ref{Snell's law Airy beam}. Clearly, the simulation results verify the validity of our design process not only in generating an Airy beam but also in directing the reflected beam toward a specific angle.
\begin{figure} [t]
\centering
\includegraphics[trim=0cm 5cm 0cm 0cm,width=8.3cm,clip]{fig7.pdf}\\
\caption
{Generation of self-healing, non-diffracting and self-bending Airy beam using reconfigurable MSs. Ideal (a) absolute value and (b) phase of the amplitude function of Airy beam $\phi_\textrm{airy}(0,s)=\textrm{Ai}(s)\text{exp}(as)$ with $a=0.05$ and $x_{0}=2.875~\mu$m at the MS interface. (c) Simulated electric field intensity of the generated Airy beam (inset: self-healing feature of Airy beam in presence of a square obstacle at the path of main lobe). Electric field density for the steered Airy beam satisfying eq. \ref{Snell's law Airy beam} associated with (d) $\theta=-10^{\circ}$ and (e) $\theta=10^{\circ}$. White dashed curve represents the main lobe path of the Airy beam achieved from analytical formulation.
}
\label{fig.Airy beam}
\end{figure}
\section{Beam bender and splitter}
Anomalous reflection and spectrum splitting are amongst most useful functionalities necessary in optical spectroscopy. By introducing interfacial phase discontinuity on the MSs, which imprints an additional momentum to the impinging light, it would be possible to redirect the light to any arbitrary direction. Considering the y-symmetry in our configuration, the generalized Snell's law is simplified to \cite{yu2014flat}
\begin{equation}
\theta=\textrm{arcsin}[\dfrac{\lambda}{2\pi}\dfrac{\textrm{d}\phi}{\textrm{d}x}]
\label{bending}
\end{equation}
in which $\theta$ is the reflection angle and ${\textrm{d}\phi}/{\textrm{d}x}$ is the phase gradient along the x direction which is actually a linear phase function in our case. Given the anomalous reflection angle, through an elaborate arrangement of meta-atoms (as discussed before), a tunable beam bender with the capability of switching between different steering angles can be realized. Figure \ref{fig.bending and steering}(a) presents the simulated electric field distribution reflected from the MSs for two optional angles, i.e., $\theta=15^{\circ}$ and $\theta=45^{\circ}$, which shows a very good agreement with the desired arrow. Furthermore, according to eq. \ref{bending}, the direction of the reflected wave is determined by the sign of ${\textrm{d}\phi}/{\textrm{d}x}$. Considering two beam benders which their phase gradients are in opposite signs; a beam splitter can be formed just by neighboring them. To corroborate the performance of such simple technique, the near-field simulations are presented in Figure \ref{fig.bending and steering}(b) for a tunable beam splitter with two arbitrary angles, $\theta=48^{\circ}$ and $\theta=70^{\circ}$. It is notable that by addressing the issues already mentioned for the performance of beam focusing, small differences between the analytical and simulation results can be resolved here.
\begin{figure} [t]
\centering
\includegraphics[trim=26.5cm 10.8cm 0cm 0cm,width=6.3cm,clip]{fig6.pdf}\\
\caption
{Active beam steering and splitting with wide angular tunability. Simulated cross-sections of the electric intensity distributions (near-field snapshot in time) of the planar (a) beam bender with nominal steering angles of $\theta=45^{\circ}$ (top) and $\theta=15^{\circ}$ (bottom), and (b) beam splitter with nominal splitting angles of $2\theta=48^{\circ}$ (top) and $2\theta=70^{\circ}$ (bottom). The wavelength of incident field is $\lambda=1550$ nm and the lateral dimension of both structures is D=100p.
}
\label{fig.bending and steering}
\end{figure}
\clearpage
\newpage
\begin{table}
\centering
\scriptsize
\caption{Comparison of different types of reconfigurable metadevices}
\begin{adjustbox}{angle=90}
\centering
\begin{tabular}
{cp{20mm}p{15mm}p{19mm}p{32mm}p{12mm}p{13mm}p{15mm}p{13.7mm}p{20mm}}
\toprule
\rowcolor{black!27}
&\hspace{-5mm}approach
&material
&trigger
&modulation
&speed
&stimuli
&wavelength
&integration/
&functionality
\\
\rowcolor{black!27}
&
&
&
&
&
&
&
&robustness
&
\\
\midrule
& \hspace{-5mm}phase transition
& LC
& thermal \cite{sautter2015active}
& amplitude ($\sim$5-fold)
& NA
& NA
& 1.55 $\mu$m
& low/low
& modulator
\\
&
&
& electrical \cite{buchnev2013electro}
& amplitude ($\sim$5-fold)
& NA
& NA
& 1.55 $\mu$m
& low/low
& switch
\\
&
&
& electrical \cite{franklin2015polarization}
& amplitude
& NA
& NA
& VIS
& low/low
& structural color
\\
&
&
& electrical \cite{chen2016electrically}
& amplitude
& NA
& NA
& VIS
& low/low
& beam deflection
\\
&
& VO$_{2}$
& electrical \cite{liu2016hybrid}
& amplitude ($\sim$\%85)
& $\sim$1 s
& $\sim$1 $\mu$J
& 3-4 $\mu$m
& high/high
& imaging
\\
&
&
& electrical \cite{zhu2017dynamically}
& amplitude ($\sim$\%33)
& 1.27 ms
& $\sim$21 nJ
& 1.1 $\mu$m
& high/high
& modulator
\\
&
&
& thermal \cite{kats2012ultra}
& amplitude ($\sim$\%80)
& NA
& NA
& 11.6 $\mu$m
& high/high
& modulator
\\
&
&
& thermal \cite{dicken2009frequency}
& amplitude
& NA
& NA
& 1.5-5 $\mu$m
& high/high
& modulator
\\
&
&
& optical \cite{dong2018lithography}
& amplitude ($\sim$4-fold)
& 40 ms
& 11.5 mW
& 10.6 $\mu$m
& high/high
& hologram
\\
&
& GLS
& electrical \cite{samson2010metamaterial}
& amplitude (15\%)
& 10 ms
& NA
& 1.55 $\mu$m
& high/high
& switch
\\
&
& GST
& optical \cite{de2018nonvolatile}
& phase
& 200 ns
& 6 mW
& 1.55 $\mu$m
& high/high
& beam steering
\\
&
&
& optical \cite{wang2016optically}
& amplitude
& 85 fs
& 9 $\mu$J
& 730 nm
& high/high
& beam focusing
\\
&
&
& thermal \cite{qu2017dynamic}
& emission
& NA
& NA
& 6.5-9.5 $\mu$m
& high/high
& emitter
\\
&
&
& thermal \cite{yin2017beam}
& polarization
& NA
& NA
& 3.1 $\mu$m
& high/high
& beam focusing
\\
&
&
& electrical \cite{hosseini2014optoelectronic}
& amplitude
& 100 ns
& 100 $\mu$j
& 350-750 nm
& high/high
& hologram
\\
&
&
& optical \cite{chen2015engineering}
& phase
& NA
& 4 mW
& 1.55 $\mu$m
& high/high
& beam focusing
\\
&
&
& thermal \cite{tittl2015switchable}
& emission
& NA
& NA
& 3.4-3.9 $\mu$m
& high/high
& thermal imaging
\\
\rowcolor{gray!20}
& \hspace{-5mm}carrier doping
& graphene
& electrical \cite{yao2014electrically}
& phase
& NA
& 40 V
& 7.7 $\mu$m
& high/high
& polarizer
\\
\rowcolor{gray!20}
&
&
& electrical \cite{emani2013electrical}
& amplitude
& NA
& 60 V
& 1.8 $\mu$m
& high/high
& switch
\\
\rowcolor{gray!20}
&
&
& electrical \cite{lee2012switching}
& amplitude (47\%)
& 100 kHz
& 350 V
& 300 $\mu$m
& high/high
& switch
\\
\rowcolor{gray!20}
&
&
& electrical \cite{sherrott2017experimental}
& phase
& NA
& 170 V
& 8.7 $\mu$m
& high/high
& switch
\\
\rowcolor{gray!20}
&
& ITO
& electrical \cite{yi2013voltage}
& amplitude
& 500 kHz
& 20 V
& 4 $\mu$m
& high/high
& absorber
\\
\rowcolor{gray!20}
&
&
& electrical \cite{huang2016gate}
& phase ($\pi$)
& 500 kHz
& 2.5 V
& 1.55 $\mu$m
& high/high
& phased array antenna
\\
\rowcolor{gray!20}
&
&
& electrical \cite{park2016dynamic}
& phase ($\pi$)
& NA
& 80 V
& 6 $\mu$m
& high/high
& polarizer
\\
\rowcolor{gray!20}
&
& InSb
& thermal \cite{iyer2017ultrawide}
& phase ($3\pi/2$)
& NA
& NA
& 11.7 $\mu$m
& high/high
& phased array antenna
\\
\rowcolor{gray!20}
&
& Perovskite
& thermal \cite{iyer2017ultrawide}
& amplitude (100\%)
& 500 ps
& 35 $\mu$J
& 1.55 $\mu$m
& low/low
& modulator
\\
& \hspace{-5mm}elasticity
& PDMS
& mechanical \cite{kamali2016highly}
& phase (2$\pi$)
& NA
& NA
& 915 nm
& low/low
& beam focusing
\\
&
&
& mechanical \cite{ee2016tunable}
& phase (2$\pi$)
& NA
& NA
& 632.8 nm
& low/low
& beam focusing
\\
&
&
& mechanical \cite{pryce2010highly}
& frequency
& NA
& NA
& 3.37 $\mu$m
& low/low
& beam focusing
\\
\rowcolor{gray!20}
& \hspace{-5mm}MEMS
& Si/SiN$_{x}$
& electro-mechanical \cite{arbabi2018mems}
& phase
& 4 kHz
& 85 V
& 622-784 nm
& low/low
& beam focusing
\\
\rowcolor{gray!20}
&
& Au/SiO$_{2}$
& electro-mechanical \cite{roy2018dynamic}
& phase
& 1 kHz
& 60 V
& 4.6 $\mu$m
& low/low
& beam focusing
\\
\rowcolor{gray!20}
&
& Au/SiO$_{2}$
& electro-mechanical \cite{liu2013dynamic}
& amplitude (56\%)
& 30 kHz
& 16 V
& 6.3 $\mu$m
& low/low
& modulator
\\
& \hspace{-5mm}Lorentz force
&
& electrical \cite{valente2015magneto}
& amplitude (2.5\%)
& 5 $\mu$s
& NA
& 1 $\mu$m
& high/high
& switch
\\
\bottomrule
\end{tabular}
\end{adjustbox}
\label{table.comparisons}
\end{table}
\clearpage
\newpage
\newcommand{\beginmethods}{%
\renewcommand{\thetable}{M\arabic{table}}%
\renewcommand{\thefigure}{M\arabic{figure}}%
\renewcommand{\thesection}{M\arabic{section}}%
\renewcommand{\thesubsection}{M\arabic{section}.\arabic{subsection}
}
\setcounter{figure}{0}
\setcounter{section}{0}
\setcounter{subsection}{0}
\setcounter{table}{0}
\section*{METHODS}
\beginmethods
\section{Solving electromagnetic scattering of MRA using FEM and FIT methods}
The full-wave simulations were carried out in EM solvers COMSOL Multiphysics (based on the FEM) and validated in CST Microwave Studio (based on the FIT). The structure is supposed to be elongated infinitely in y direction, so periodic boundary condition is applied at x direction and perfectly match layer absorbing boundaries are applied in the z direction. The structure is normally illuminated with a
TM-polarized (H-field parallel to the long axis of stripes) plane wave propagating in z direction, and the reflected fields into the above free space are measured. Thickness of the gold nanoribbon, the underlying GST nanostripe, and the silica spacers are 30 nm, 180 nm, 180 nm, respectively. Moreover, as long as the thickness of gold substrate is considered greater than the skin depth, the transmittance will be negligible; therefore, we fix it at 100 nm for all the proposed designs. The period of MRA is chosen 550 nm which not only ensures the higher diffraction orders will be well suppressed but also meets resolution limit in fabrication process The dispersive behavior of gold can be approximated through the well-known Drude model with: $\epsilon_{gold}(\omega)=\epsilon_{0}[\epsilon_{\infty}-{\omega_{p}}^{2}(\omega^{2}+i\omega\Gamma)^{-1}]$ in which $\epsilon_{\infty}=1.53$, $\omega_{p}=2\pi\times2.069 $ PHz is the plasma frequency, $\Gamma=2\pi\times17.64$ THz is the collision frequency, $\omega$ is the angular frequency, and $\epsilon_{0}$ is the absolute permittivity. The permittivity of SiO$_{2}$ is considered as a constant value ($\epsilon_{\rm{SiO}_{2}}=2.25$) in NIR region. It is worth mentioning that the dielectric spacer prevents heat diffusion between adjacent unit cells during conversion process.
\section{Solving heat transfer model of MRA using FEM method}
To investigate the conversion mechanism of GST gap nanostripe in the proposed gate-voltage tunable MRA, COMSOL Multiphysics software package was used for Joule heating simulation effect and consequently transient thermal behavior study. To make a simple yet comprehensive model, we made certain assumptions providing useful qualitative insights for future improvement of the switching speed and power. A coupled multiphysics model including ``Electric Currents (ec)'' module, which is applied to the gold strip in case of lateral heating and gold strip/back-mirror in case of vertical heating, and ``Heat Transfer in Solid (ht)'' module, incorporating the entire meta-atom, was employed.
The governing transient heat transfer equation is described by \cite{chen2012nanosecond}
\begin{equation}
{C_{s}\rho\frac{\partial T}{\partial t}+\nabla .(-k \nabla .T)=Q_{s}
}
\label{equ11}
\end{equation}
where $C_{s}$ is the specific heat capacity, $\rho$ is density, $k$ is the thermal conductivity, $T$ is the time- and space-dependent temperature, and $Q_{s}$ is the heat source per volume. Material thermal properties used for simulation are summarized in Table \ref{table.heat transfer model}. To consider a real practical model, we assume the structure was attached to a $500~\mu$m silicon substrate.
In the ``Electric currents (ec)'' physics model all boundary conditions were set to be electrically insulating except the boundaries attached to the gates. While, a normal constant current density $J_{s}=I/A$ is applied to the input gate, such that $I$ is the total current applied to the metadevice and $A$ is the cross section of the gold nanoribbon, the output gate is electrically ground. In the ``Heat Transfer in Solid (ht)'' physics model, we assumed no net heat flux in the transverse plane meaning that thermal insulation boundary condition can be applied to all side-walls to prevent inward and outward heat transfer. Also, the top and bottom surfaces were given a heat flux condition with ambient temperature of $T_{0}=293.15$ K and the heat transfer coefficient of $5~$W/(m$^{2}$K).
To model the heat fluxes and temperature gradients at the interfaces, thermal boundary resistance (TBR) is applied at interior boundaries. While TBR for GST/SiO2 and GST/Au are considered $58~\textrm{m}^{2}\textrm{k}/\textrm{GW}$ and $5.1~\textrm{m}^{2}\textrm{k}/\textrm{GW}$, respectively, for other interior boundaries it is assumed $25~\textrm{m}^{2}\textrm{k}/\textrm{GW}$, which is typical for many systems.
To gain a complete understanding of the electrothermal process, we investigated the temperature dependency of the meta-atom on the applied current pulse (see Figure \ref{fig.electro-thermal}). One can observe that the temperature within GST nanostripe can be raised rapidly from the crystallization temperature, $T_\textrm{c}=160^{\circ}$ C, to the melting temperature, $T_\textrm{m}=600^{\circ}$ C, by increasing the pulse intensity, duration, and numbers. Such temperature variation can be translated to the multistate crystallization by correlating the simulated temperature data with the reflectance evolution achieved in the experiments; as a result, one can calculate the exact power to excite a specific state of GST nanostripe.
Such capability will be crucial for realization of practical reconfigurable nanophotonic devices. It is noteworthy that increasing the applied current intensity, which results in localized Joule heating, can substantially improved the switching behavior of the meta-atom and consequently the dynamic response of metadevice.
\begin{figure} [b]
\centering
\includegraphics[trim=0cm 2cm 8.5cm 0cm,width=8.3cm,clip]{fig12.pdf}\\
\caption
{Dependence of GST nanostripe temperature on the intensity, duration, and number of applied electrical current pulses. The point of interest is set at the geometrical center of the GST gap nanostripe.
}
\label{fig.electro-thermal}
\end{figure}
\clearpage
\newpage
\begin{table}
\centering
\footnotesize
\caption{Thermal properties of materials used in the heat transfer model \cite{cao2014broadband,wuttig2017phase}}
\begin{adjustbox}{angle=0}
\begin{tabular}{cp{22mm}p{16mm}p{20mm}p{21mm}}\toprule
\rowcolor{black!27}
&Special heat capacity
&Density
&Thermal conductivity
&Electrical conductivity
\\
\rowcolor{black!27}
&$Cs$ (J/(kg.K))
&$\rho$ (kg/m$^3$)
&$k$ (W/(m.k))
&$\sigma$ (S/m)
\\
\midrule
Gold
&129
&19300
&182
&45.6$\times 10^{6}$\\
\rowcolor{gray!20}
GST &
220
&6150
&Temperature dependent \cite{cao2014broadband}
&Phase dependent \cite{cao2014broadband}\\
Silica
&741
&2200
&1
&0\\
\rowcolor{gray!20}
Air
&1
&353/T[K]
&0.03
&0\\
Si
&700
&2328
&130
&1000\\
\bottomrule
\end{tabular}
\label{table.heat transfer model}
\end{adjustbox}
\end{table}
\clearpage
\newpage
\bibliographystyle{lpr}
|
\section{Introduction}
3D face reconstruction from monocular images is a long-standing goal in computer vision with applications in face
recognition, film industry, animation and other areas. Earlier efforts date back to late nineties and introduce
morphable face models~\cite{BaselMM}. Traditional methods address this task with optimization-based techniques and
analysis-through-synthesis methods~\cite{BaselFitting, Piotraschke2016Automated3F, Blanz2003, thies2016face, GarridoRigs}. More
recently, reg\-ression-based methods started to emerge \cite{zhu2015discriminative, ZFace, 3DDFA, MOFA}. In particular,
the task has seen an increasing interest from the CNN community over the past few years \cite{3DDFA, MOFA, VRN,
SynthLearn, tran2017regressing}. However, the applicability of neural networks remains difficult due to the lack of large-scale training data.
Possible solutions include the use of synthetic data \cite{ZFace, SynthLearn}, incorporation of unsupervised
training criteria \cite{MOFA}, or combination of both \cite{Richardson_2017_CVPR}. Another option is to produce
semi-synthetic data by applying an optimization-based algorithm with proven accuracy to a database of faces
\cite{3DDFA, VRN, tran2017regressing}.
Optimization-based methods for morphable model fitting vary in many respects. Some design choices include image
formation model, regularization and optimization strategy. Another source of variation is the kind of face attributes
being used. Traditional formulation employs face texture \cite{BaselMM}. It uses morphable model to generate a
synthetic face image and optimizes for parameters that would minimize the difference between the synthetic image
and the target. However, this formulation also relies on a sparse set of facial landmarks used for initialization.
Earlier methods used manually annotated landmarks \cite{Blanz2003}. The user was required to annotate a few facial
points by hand. Recent explosion of facial landmarking methods \cite{Kazemi, SDM, FAN, MDM} made this process automatic
and the set of landmarks became richer. This posed the question if morphable model fitting could be done based purely
on landmarks \cite{ABas}. It is especially desirable because algorithms based on landmarks are much faster and
suitable for real-time performance while texture-based algorithms are quite slow (on the order of 1 minute per image).
Unfortunately existing literature reports only few quantitative evaluations of optimization-based fitting algorithms.
Some works assume that landmark-based fitting provides satisfactory accuracy
\cite{HighFidelityNormalization, PerspectiveAware} while others demonstrate its limitations \cite{ABas, OcContours}.
Some use texture-based algorithms at the cost of higher computational demands, but the advantage in accuracy is not
quantified \cite{thies2016face, GarridoRigs, Saito2017PhotorealisticFT}.
The situation is further complicated by the lack of standard benchmarks with reliable ground truth and well-defined evaluation
procedures.
We implement a morphable model fitting algorithm and tune its parameters in two scenarios: relying solely on
landmarks and using landmarks in combination with the texture. We test this algorithm on images from BU4DFE dataset
\cite{BU4DFE} and demonstrate that incorporation of texture significantly improves the accuracy.
It is desirable to enjoy both the accuracy of texture-based reconstruction algorithms and the high processing speed
enabled by network-based methods. To this end, we use the fitting algorithm to process 300W database of faces
\cite{300W} and train a neural network to predict facial geometry on the resulting semi-synthetic dataset. It is
important to keep in mind that the applicability of the fitting algorithm is limited by the expressive power of the
morphable model. In particular, it doesn't handle large occlusions and extreme lighting conditions very well. To rule
the failures out, we visually inspect the processed dataset and delete failed examples. We compare our dataset with a
similarly produced 300W-3D \cite{3DDFA} and show that our dataset allows to learn more accurate models. We make our
dataset publicly available \footnote{\url{https://github.com/nchinaev/MobileFace}}.
An important consideration for CNN training is the loss function. Standard losses become problematic when predicting
parameters of morphable face models due to the different nature and scales of individual parameters.
To resolve this issue, the MSE loss needs to be reweighted and some ad-hoc weighting schemes have been used in the past~\cite{3DDFA}.
We present a loss function that accounts for individual contributions of morphable model parameters in a clear and
intuitive manner by constructing a 3D model and directly comparing it to the ground truth in the 3D space and in the 2D
projected space.
This work provides the following contributions:
(i) we evaluate variants of the fitting algorithm on a database of facial scans providing quantitative evidence of
texture-based algorithms superiority;
(ii) we train a MobileNet-based neural network that allows for fast facial shape reconstruction even on mobile devices;
(iii) we propose an intuitive loss function for CNN training;
(iv) we make our evaluation code and datasets publicly available.
\subsection{Related Work}
Algorithms for monocular 3d face shape reconstruction may be broadly classified into two following categories:
optimization-based and regression-based. Optimization-based approaches make assumptions about the nature of image
formation and express them in the form of energy functions. This is possible because faces represent a set of objects
that one can collect some strong priors about. One popular form of such prior is a morphable model. Another way to
model image formation is shape from shading technique \cite{Kemelmacher1, Kemelmacher2, roth2016adaptive}. This
class of algorithms has a drawback of high computational complexity. Regression-based methods learn from
data. The absence of large datasets for this task is a limitation that can be addressed in several ways outlined below.
\medskip\\
\noindent
\textbf{Learning From Synthetic Data.} Synthetic data may be produced by rendering facial scans \cite{ZFace} or by
rendering images from a morphable model \cite{SynthLearn}. Corresponding ground truth 3d models are readily available
in this case because they were used for rendering. These approaches have two limitations: first, the variability in
facial shapes is only limited to the subjects participating in acquisition, and second, the image formation is limited
by the exact illumination model used for rendering.
\medskip\\
\noindent
\textbf{Unsupervised Learning}. Tewari et al. \cite{MOFA} incorporate rendering process into their learning framework.
This rendering layer is implemented in a way that it can be back-propagated through. This allows to circumvent the
necessity of having ground truth 3d models for images and makes it possible to learn from datasets containing face
images alone. In the follow up work Tewari et al. \cite{MOFACooler} go further and learn corrections to the morphable
model. Richardson et al. \cite{Richardson_2017_CVPR} incorporate shape from shading into learning process to learn
finer details.
\medskip\\
\noindent
\textbf{Fitting + Learning.} Most closely related to our work are works of Zhu et al. \cite{3DDFA} and Tran et al.
\cite{tran2017regressing}. They both use fitting algorithms to generate datasets for neural network training. However,
accuracies of the respective fitting algorithms \cite{BaselFitting} and \cite{Piotraschke2016Automated3F} in the
context of evaluation on datasets of facial scans are not reported by their authors. This raises two questions: what is
the maximum accuracy attainable by learning from the results of these fitting methods and what are the gaps between the
fitting methods and the respective learned networks? We evaluate accuracies of our fitting methods and networks on
images from BU4DFE dataset in our work.
\section{MobileFace}
Our main objective is to create fast and compact face shape predictor suitable for real-time inference on mobile
devices. To achieve this goal we train a network to predict morphable model parameters (to be introduced in
Sec. \ref{sec:MM}). Those include parameters related to 3d shape $\vec{\alpha_{id}}$ and
$\vec{\alpha_{exp}}$, as well as those related to projection of the model from 3d space to the image plane: translation
$\vec{t}$, three angles $\phi$, $\gamma$, $\theta$ and projection $f$, $P_x$, $P_y$. Vector $\vec{p} \in
\mathbb{R}^{118}$ is a concatenation of all the morphable model parameters predicted by the network:
\begin{align}\label{eq:vectorP}
\vec{p} = \left(\begin{array}{ccccccccc} \vec{\alpha_{id}}^T & ~~\vec{\alpha_{exp}}^T & ~~\vec{t}^T & ~~\phi & ~~\gamma
& ~~\theta & ~~f & ~~Px & ~~Py \end{array} \right)^T
\end{align}
\subsection{Loss Functions}
We experiment with two losses in this work. The first MSE loss can be defined as
\begin{align}\label{eq:MSEloss}
\texttt{Loss}_{\texttt{MSE}} = \sum_{i}{||\vec{p^i} - \vec{p^i}_{gt}||_2^2}.
\end{align}
Such a loss, however, is likely to be sub-optimal as it treats parameters $\vec{p}$ of different nature and scales equally. They impact the $3d$
reconstruction accuracy and the projection accuracy differently. One way to overcome this is to use the outputs of the
network to construct 3d meshes $\vec{S}(\vec{p^i})$ and compare them with ground truth $\vec{S}_{gt}$ during training
\cite{dou2017end}. However, such a loss alone would only allow to learn parameters related to the 3d shape:
$\vec{\alpha_{id}}$ and $\vec{\alpha_{exp}}$. To allow the network to learn other parameters, we propose to augment
this loss by an additional term on model projections $P(\vec{p^i})$:
\begin{align}\label{eq:loss}
\texttt{Loss}_{\texttt{2d + 3d},~l_2} = \sum_{i}{\left(||\vec{S}(\vec{p^i}) -\vec{S_{gt}})||_2^2 +
||P(\vec{p^i}) - P_{gt}||_2^2\right)}
\end{align}
Subscript $l_2$ indicates that this loss uses $l_2$ norm for individual vertices. Likewise, we define
\begin{align}\label{eq:loss_l1}
\texttt{Loss}_{\texttt{2d + 3d},~l_1} = \sum_{i}{\left(||\vec{S}(\vec{p^i}) -\vec{S_{gt}})||_1 + ||P(\vec{p^i}) -
P_{gt}||_1\right)}
\end{align}
We provide details of $\vec{S}(\vec{p^i})$ construction and $P(\vec{p^i})$ projection in the next subsection.
\subsection{Morphable Model}\label{sec:MM}
\textbf{Geometry Model.} Facial geometries are represented as meshes. Morphable models allow to generate variability
in both face identity and expression. This is done by adding parametrized displacements to a template face model called
the mean shape. We use the mean shape and $80$ modes from Basel Face Model \cite{BaselMM} to generate identities and
$29$ modes obtained from Face Warehouse dataset \cite{FW} to generate expressions. The meshes are controlled by two
parameter vectors $\vec{\alpha}_{id} \in \mathbb{R}^{80}$ and $\vec{\alpha}_{exp} \in \mathbb{R}^{29}$:
\begin{align}\label{eq:MM}
\vec{S} = \vec{M} + A_{id} \cdot \vec{\alpha}_{id} + A_{exp} \cdot \vec{\alpha}_{exp}.
\end{align}
Vector $\vec{S} \in \mathbb{R}^{3\cdot N}$ stores the coordinates of $N$ mesh vertices. $\vec{M}$ is the mean shape.
Matrices $A_{id} \in \mathbb{R}^{3\cdot N \times 80}$, $A_{exp} \in \mathbb{R}^{3\cdot N
\times 29}$ are the modes of variation.
\\
\\
\textbf{Projection Model.} Projection model translates face mesh from the 3d space to a 2d plane. Rotation matrix $R$
and translation vector $\vec{t}$ apply a rigid transformation to the mesh. Projection matrix with three parameters $f$,
$Px$, $Py$ transforms mesh coordinates to the homogeneous space. For a vertex $\vec{v} = \left(x_m, y_m, z_m\right)^T$ the
transformation is defined as:
\begin{align}\label{eq:Rt}
\left(\begin{array}{c} x_t \\
y_t \\
z_t \end{array}\right) =
\Pi \cdot
\left(\begin{array}{cc} R & ~\vec{t} \end{array}\right) \cdot
\left(\begin{array}{c} x_m \\
y_m \\
z_m \\
1\end{array}\right),
~~~\Pi = \left(\begin{array}{ccc} f & ~0 & ~P_x \\ 0 & ~f & ~P_y \\ 0 & ~0 & ~1 \end{array}\right),
\end{align}
and the final projection of a vertex to the image plane is defined by $u$ and $v$ as:
\begin{align}\label{eq:uv}
u = \frac{x_t}{z_t},~~~v = \frac{y_t}{z_t}.
\end{align}
The projection is defined by $9$ parameters including three rotation angles, three translations and three
parameters of the projection matrix $\Pi$. We denote projected coordinates by:
\begin{align}\label{eq:P}
P(\Pi, R, \vec{t}, \vec{S}) = \left(\begin{array}{cccc} u_1 & u_2 & \ldots & u_N \\ v_1 & v_2 & \ldots & v_N \end{array}
\right)^T
\end{align}
\subsection{Data Preparation}
Our objective here is to produce a dataset of image-model pairs for neural network training. We use the
fitting algorithm detailed in Sec.
\ref{sec:opt} to process the 300W database of annotated face images \cite{300W}. Despite its accuracy reported in Sec.
\ref{sec:exp_fitting} this algorithm has two limitations. First, the expressive power of the morphable model is
inherently limited due to laboratory conditions in which the model was obtained and due to the lighting model being
used. Hence, the model can't generate occlusions and extreme lighting conditions. Second, the hyperparameters of the
algorithm have been tuned for a dataset taken under controlled conditions. Due to these limitations, the algorithm
inevitably fails on some of the in-the-wild photos. To overcome this shortcoming, we visually inspect the results and
delete failed photos. Note that we do not use any specific criteria and this deletion is guided by the visual appeal
of the models, hence it may be performed by an untrained individual. This leaves us with an even smaller amount of
images than has initially been in the 300W dataset, namely $2300$ images.
\renewcommand{\arraystretch}{0}
\begin{figure}
\begin{center}
\begin{tabular}{ccccc}
\subfloat{\includegraphics[width = 0.13\textwidth]{black_girl.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{asian_woman.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{bearded_gentleman.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{elijah.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{Indian.png}}\\
\subfloat{\includegraphics[width = 0.13\textwidth]{black_girl_original.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{asian_woman_original.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{bearded_original.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{elijah_original.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{indian_original.png}}\\
\end{tabular}
\caption{Example images and corresponding curated ground truth from our training set.}
\label{fig:exTrain}
\end{center}
\end{figure}
\renewcommand{\arraystretch}{1}
This necessitates data augmentation. We
randomly add blur and noise in both RGB and HSV spaces. Since some of the images with large occlusions have been
deleted during visual inspection, we compensate for this and randomly occlude images with black rectangles of varied
sizes~\cite{cutout}. Fig.~\ref{fig:exTrain} shows some examples of our training images.
\subsection{Network Architecture}
Architecture of our network is based on MobileNet \cite{MobileNet}. It consists of interleaving convolution and
depth-wise convolution \cite{DepthWise} layers followed by average pooling and one fully connected layer. Each
convolution layer is followed by a batch normalization step \cite{BatchNorm} and a ReLU activation. Input images are
resized to $96 \times 96$. The final fully-connected layer generates the outputs vector $\vec{p}$ eq. (\ref{eq:vectorP}). Main changes
compared to the original architecture in \cite{MobileNet} include the decreased input image size $96 \times 96 \times 3$, the first convolution
filter is resized to $3 \times 3 \times 3 \times 10$, the following filters are scaled accordingly, global average pooling is performed
over $2 \times 2$ region, and the shape of the FC layer is $320 \times 118$.
\section{Morphable Model Fitting}
We use morphable model fitting to generate 3d models of real-world faces to be used for neural network training. Our
implementation follows standard practices \cite{thies2016face, GarridoRigs}. Geometry and projection models have been
defined in (Sec. \ref{sec:MM}). Texture model and lighting allow to generate face images. Morphable model fitting aims
to revert the process of image formation by finding the combination of parameters that will result in a synthetic image
resembling the target image as closely as possible.
\subsection{Image Formation}
\textbf{Texture Model.} Face texture is modeled similarly to eq. (\ref{eq:MM}). Each vertex of the mesh is assigned
three RGB
values generated from a linear model controlled by a parameter vector $\vec{\beta}$:
\begin{align}\label{eq:MT}
\vec{T} = \vec{T}_0 + B \cdot \vec{\beta}.
\end{align}
We use texture mean and modes from BFM \cite{BaselMM}.
\\
\\
\textbf{Lighting Model.} We use the Spherical Harmonics basis \cite{invrend, envmap} for light computation. The
illumination of a vertex having albedo $\rho$ and normal $\vec{n}$ is computed as
\begin{align}\label{light}
I = \rho \cdot \left({\begin{array}{cc} \vec{n}^T & 1\end{array}}\right) \cdot M \cdot \left({\begin{array}{c} \vec{n}
\\ 1\end{array}}\right),
\end{align}
$M$ is as in \cite{envmap} having $9$ controllable parameters per channel. RGB intensities are
computed separately thus giving overall $9 \cdot 3 = 27$ lighting parameters, $\vec{l} \in \mathbb{R}^{27}$ is the
parameter vector. Albedo $\rho$ is dependent on $\vec{\beta}$ and computed as in eq. (\ref{eq:MT}).
\subsection{Energy Function}
Energy function expresses the discrepancy between the original attributes of an image and the ones generated from the
morphable model:
\begin{align}\label{texture_term}
\begin{split}
E = E_{\texttt{tex}} + c_{\texttt{lands}} \cdot E_{\texttt{lands}} + r_{beta, 2} \cdot E_{\texttt{reg,tex}}
+ r_{exp, 2} \cdot E_{\texttt{reg,exp}}.
\end{split}
\end{align}
We describe individual terms of this energy function below.
\\
\\
\textbf{Texture.} The texture term $E_{\texttt{tex}}$ measures the difference between the target image and the one rendered from the
model. We translate both rendered and target images to a standardized UV frame as in \cite{BaselFitting} to unify all
the image resolutions. Visibility mask
$\mathscr{M}$ cancels out the invisible pixels.
\begin{align}\label{eq:texture}
E_{\texttt{tex}} = \frac{||\mathscr{M} \cdot (I_{\texttt{target}} - I_{\texttt{rendered}})||}{|\mathscr{M}|}.
\end{align}
We produce $I_{\texttt{rendered}}$ by applying eq. (\ref{light}) and $I_{\texttt{target}}$ by sampling from the
target image at the positions of projected vertices $P$ eq. (\ref{eq:P}). Visibility mask $\mathscr{M}$ is computed
based on the orientations of vertex normals. We test three alternative norms in place of $||\cdot||$: $l_1$, $l_2$ and
$l_{2,1}$ norm \cite{thies2016face} that sums $l2$ norms computed for individual pixels.
\\
\\
\textbf{Landmarks.} We use the landmark detector of \cite{Kazemi}. Row indices $\mathscr{L} = \{k_i\}_{i = 1}^{68}$ for
matrix $P$ eq. (\ref{eq:P}) correspond to the $68$ landmarks. Detected landmarks are $L \in \mathbb{R}^{68\times2}$.
The landmark term is defined as:
\begin{align}\label{eq:lands}
E_{\texttt{lands}} = ||L - P_{\mathscr{L}, :}||_2^2.
\end{align}
One problem with this term is that indices $\mathscr{L}$ are view-dependent due to the landmark marching. We adopt a
solution similar to that of \cite{HighFidelityNormalization} and annotate parallel lines of vertices for the landmarks
on the border.
\\
\\
\textbf{Regularization.} We assume multivariate Gaussian priors on morphable model parameters as defined below and use
$\sigma_{id}$ and $\sigma_{tex}$ provided by~\cite{BaselMM}.
\begin{align}\label{reg}
E_{\texttt{reg,id}} = \sum_{i = 1}^{80}{\frac{\alpha_{id, i}^2}{\sigma_{id, i}^2}},
~~ E_{\texttt{reg,exp}} = \sum_{i = 1}^{29}{\frac{\alpha_{exp, i}^2}{\sigma_{exp, i}^2}},
~~ E_{\texttt{reg,tex}} = \sum_{i = 1}^{80}{\frac{\beta_{i}^2}{\sigma_{tex, i}^2}}
\end{align}
We regularize neither lighting nor projection parameters.
\subsection{Optimization}\label{sec:opt}
Optimization process is divided into two major steps: First, we minimize the landmark term:
\begin{align}\label{landmark_term}
E_1 = E_{\texttt{lands}} + r_{id, 1} \cdot E_{\texttt{reg,id}} + r_{exp, 1} \cdot E_{\texttt{reg,exp}}.
\end{align}
We then minimize the full energy function eq. (\ref{texture_term}). These two steps are also divided into sub-steps
minimizing the energy function with respect to specific parameters similarly to \cite{GarridoRigs}. We minimize the
energy function with respect to only one type of parameters at any moment.
We do not include identity regularization into eq. (\ref{texture_term}) because it did not improve accuracy in
our experiments.
\section{Experiments}
We carry out three sets of experiments. First, we study the effect of different settings for the fitting of the
morphable model used in this paper. Second, we experiment with different losses and datasets for neural network
training. Finally, we present a comparison of our method with other recent approaches.
Unfortunately current research in 3d face reconstruction is lacking standardized benchmarks and evaluation protocols.
As a result, evaluations presented in research papers vary in the type of error metrics and datasets
used (see Table~\ref{table:methodsTestsets}). This makes the results from many works difficult to compare. We hope
to contribute towards filling this gap by providing the standard evaluation code and a testing set of
images\footnote{\url{https://github.com/nchinaev/MobileFace}}.
\textbf{BU4DFE Selection.} Tulyakov et al. \cite{tulyakov} provide annotations for a total of $3000$ selected scans
from BU4DFE. We divide this selection into two equally sized subsets BU4DFE-test and BU4DFE-val. We report final
results on the former and experiment with hyperparameters on the latter. For the purpose of evaluation we use
annotations to initialize the ICP alignment.
\subsection{Implementation Details}
We trained networks for the total of $3\cdot 10^5$ iterations with the batches of size $128$. We added $l_2$ weight
decay with coefficient of $10^{-4}$ for regularization. We used Adam optimizer \cite{Adam} with learning rate of
$10^{-4}$ for iterations before $2\cdot10^5$-th and $10^{-5}$ after. Other settings for the optimizer are standard.
Coefficients for morphable model fitting are $r_{id, 1} = 0.001$, $r_{exp, 1} = 0.1$, $r_{beta, 2} = 0.001$,
$c_{\texttt{lands}} = 10$, $r_{exp, 2} = 10$.
\subsection{Accuracy Evaluation}
Accuracy of 3D reconstruction is estimated by comparing the resulting 3D model to the ground truth facial scan. To
compare the models, we first perform ICP alignment. Having reconstructed facial mesh $\vec{S}$ and the
ground truth scan $\vec{S_{gt}}$, we project vertices of $\vec{S}$ on $\vec{S_{gt}}$ and Procrustes-align $\vec{S}$ to
the projections. These two steps are iterated until convergence.
\\
\\
\textbf{Error Measure.} To account for variations in scan sizes, we use a normalization term
\begin{align}\label{eq:1}
C(\vec{S_{gt}}) = ||\vec{S_{gt}^0}||_2^2,
\end{align}
where $\vec{S_{gt}^0}$ is $\vec{S_{gt}}$ with the mean of each x, y, z coordinate subtracted. The dissimilarity measure
between $\vec{S}$ and $\vec{S_{gt}}$ is
\begin{align}\label{eq:errorMetric}
d(\vec{S}, \vec{S_{gt}}) = c_s \cdot\frac{||\vec{S} - \vec{S_{gt}}||_2^2}{C(\vec{S_{gt}})}
\end{align}
The scaling factor $c_s = 100$ is included for convenience.
\setlength{\tabcolsep}{4pt}
\begin{table}
\begin{center}
\caption{Methods and their corresponding testsets.}
\label{table:methodsTestsets}
\begin{tabular}{c|c}
\hline\noalign{\smallskip}
Work & Testset\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Jackson et al. \cite{VRN} & AFLW2000-3D, renders from BU4DFE and MICC\\
Tran et al. \cite{tran2017regressing} & MICC video frames \\
Tewari et al. \cite{MOFA} & synthetic data; Face Warehouse \\
Dou et al. \cite{dou2017end} & UHDB31, FRGC2, BU-3DFE \\
Roth et al. \cite{roth2016adaptive} & renders from BU4DFE \\
& \\
\end{tabular}
\end{center}
\end{table}
\setlength{\tabcolsep}{1.4pt}
\subsection{Morphable Model Fitting}\label{sec:exp_fitting}
We compare the accuracy of the fitting algorithm in two major settings: using only landmarks and using landmarks in
combination with texture. To put the numbers in a context, we establish two baselines. First baseline is attained by
computing the reconstruction error for the mean shape. This demonstrates the performance of a hypothetical
dummy algorithm that always outputs the mean shape for any input. Second baseline is computed by registering the
morphable model to the scans in 3d. It demonstrates the performance of a hypothetical best method that is only bounded
by the descriptive power of the morphable model.
Landmark-based fitting is done by optimizing eq. (\ref{landmark_term}) from sec. \ref{sec:opt}. Texture-based fitting
is done by optimizing both eq. (\ref{landmark_term}) and eq. (\ref{texture_term}). Fig. \ref{fig:landsVStex} shows
cumulative error distributions. It is clear from the graph that texture-based fitting significantly outperforms
landmark-based fitting which is only as accurate as the meanshape baseline. However, there is still a wide gap between
the performance of the texture-based fitting and the theoretical limit.
Figs. \ref{fig:texVariants1}, \ref{fig:texVariants2} show the performance of texture-based fitting algorithm with
different settings. The settings differ in the type of norm being used for texture term computation and the amount of
regularization. In particular,
Fig. \ref{fig:texVariants1} demonstrates that the choice of the norm plays an important role with $l_{2, 1}$ and $l_1$
norms outperforming $l_2$. Fig. \ref{fig:texVariants2} shows that the algorithm is quite sensitive to the
regularization, hence the regularization coefficients need to be carefully tuned.
\begin{figure}
\centering
\includegraphics[height=6.5cm]{landsVStex.eps}
\caption{Evaluation of fitting methods on BU4DFE-test. Areas under curve are computed for normalized mesh
distances ranging from $0$ to $1$. Shorter span of x-axis is used for visual clarity.}
\label{fig:landsVStex}
\end{figure}
\subsection{Neural Network}
We train the network on our dataset of image-model pairs. For the sake of comparison, we also train it on 300W-3D
\cite{3DDFA}. The training is performed in different settings: using different loss functions and using manually
cleaned version of the dataset versus non-cleaned. The tests are performed on BU4DFE-val. Figs.
\ref{fig:cedExperiments1}, \ref{fig:cedExperiments2} show cumulative error distributions. These experiments support
following claims:
\begin{itemize}
\item Learning from our dataset gives better results than learning from 300W-3D,
\item Our loss function improves results compared to baseline MSE loss function,
\item Manual deletion of failed photos by an untrained individual improves results.
\end{itemize}
\begin{figure}
\centering
\subfloat[Evaluations for different norms. $r_{exp, 2} = 0$]
{\includegraphics[width=0.48\textwidth]{texVariants1.eps}\label{fig:texVariants1}}
\hfill
\subfloat[Evaluations with different regularizations.
$l_{2, 1}$ norm is used in all cases.]
{\includegraphics[width=0.48\textwidth]{texVariants2.eps}\label{fig:texVariants2}}
\hfill
\caption{Evaluation of texture-based fitting algorithm on BU4DFE-val with different settings. }
\end{figure}
\begin{figure}
\centering
\subfloat[Comparison of networks trained on our cleaned dataset with different losses.]
{\includegraphics[width=0.48\textwidth]{cedExperiments1.eps}\label{fig:cedExperiments1}}
\hfill
\subfloat[Comparison of networks trained with $\texttt{Loss}_{\texttt{2d + 3d},~l_2}$ on different datasets.]
{\includegraphics[width=0.48\textwidth]{cedExperiments2.eps}\label{fig:cedExperiments2}}
\hfill
\caption{Evaluation on BU4DFE-val for networks trained with different losses on different datasets.}
\end{figure}
\subsection{Comparison with the State of the Art}
\begin{figure}
\centering
\includegraphics[height=6.5cm]{cedComparisons.eps}
\caption{Comparison of methods on BU4DFE-test.}
\label{fig:cedComparison}
\end{figure}
\noindent
\textbf{Quantitative Results.} Fig. \ref{fig:cedComparison} presents evaluations of our network and a few recent
methods on BU4DFE-test. Error metric is as in eq. (\ref{eq:errorMetric}). The work of Tran
et al. \cite{tran2017extreme} is based on \cite{tran2017regressing} and their code allows to produce
non-neutral models, therefore we present an evaluation of \cite{tran2017extreme} and not \cite{tran2017regressing}. We
do not present an evaluation of 3DDFA \cite{3DDFA} because Jackson et al. \cite{VRN} have already demonstrated that
3DDFA is inferior to their method. We do not include the work of Sela et al. \cite{Sela} into comparison because we
were not able to reproduce their results. For Jackson et al. \cite{VRN} we were able to reproduce their error on
AFLW2000-3D. We used MATLAB implementation of isosurface algorithm to transform their volumes into meshes. Tran et. al
\cite{tran2017extreme} do not present an evaluation on 3d scans, however we were able to roughly reproduce an error for
MICC dataset \cite{MICC} from their earlier work \cite{tran2017regressing}. We noticed that their method is sensitive
to the exact selection of frames from MICC videos. For an optimal selection of frames the error equals $1.43$ which is
less than $1.57$ reported in \cite{tran2017regressing}. In the worst case error equals $2.34$. Tewari et al.
\cite{MOFA} did not open-source their implementation of MOFA, but authors kindly provided their reconstructed models
for our testset.
It is seen from the graph that our network performs on a par with other recent methods being slightly ahead of the
second-best method. Additionally, the size of our model is orders of magnitude smaller, see Table
\ref{table:comparisons} for a comparison. We used Intel Core i5-4460 for CPU experiments, NVIDIA GeForce GTX 1080 for
GPU experiments (except for Tewari \cite{MOFA}, they used NVIDIA Titan X Pascal) and Samsung Galaxy S7 for ARM
experiments.
Table \ref{table:hassnerMicc} presents a comparison with Tran et al. \cite{tran2017regressing} on MICC dataset
\cite{MICC}. This is a dataset of $53$ subjects. For each of the subjects it provides three videos and a neutral facial
scan. It is therefore crucial that a method being evaluated on this dataset should output neutral models for any input.
Method of Tran et al. \cite{tran2017regressing} is specifically designed for this purpose. We adapt our method to this
scenario by setting $\vec{\alpha}_{exp}$ to zero. We randomly select $23$ frames per individual and form $23$
corresponding testsets. We compute errors over these testsets and average those. One important aspect affecting the
errors is the use of scaling during ICP alignment: Tran et al. \cite{tran2017regressing} did not allow models to scale
during the alignment. We present evaluations in both settings.
Table \ref{table:TewariFW} presents a comparison with Tewari et al. \cite{MOFA} and Garrido et al. \cite{GarridoRigs}
on a selection of $9$ subjects from Face Warehouse \cite{FW} dataset. Version of Tewari et al. network with surrogate
loss has been used for this and previous evaluations.
\setlength{\tabcolsep}{4pt}
\begin{table}
\begin{center}
\caption{Comparison with Tran et al. \cite{tran2017regressing} on MICC dataset \cite{MICC}. All numbers are in mm.}
\label{table:hassnerMicc}
\begin{tabular}{l|ll}
\hline\noalign{\smallskip}
Method & w.o. scale & w. scale \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Tran \cite{tran2017regressing} & 1.83 $\pm$ 0.04 & 1.46 $\pm$ 0.03 \\
Our & \textbf{1.70} $\pm$ 0.02 & \textbf{1.33} $\pm$ 0.02 \\
\end{tabular}
\end{center}
\end{table}
\setlength{\tabcolsep}{1.4pt}
\setlength{\tabcolsep}{4pt}
\begin{table}
\begin{center}
\caption{Comparisons on a selection from Face Warehouse dataset. All numbers are in mm.}
\label{table:TewariFW}
\begin{tabular}{l|lll}
\hline\noalign{\smallskip}
Method & Our & Tewari \cite{MOFA} & Garrido \cite{GarridoRigs}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Error & 1.8 $\pm$ 0.07& 1.7 & 1.4 \\
\hline
\end{tabular}
\end{center}
\end{table}
\setlength{\tabcolsep}{1.4pt}
\newline
\textbf{Qualitative Results.} Fig. \ref{fig:qualitative} shows a comparison with Jackson et al. \cite{VRN},
Tewari er al. \cite{MOFA} and Tran et al. \cite{tran2017extreme} on a few images from BU4DFE-test.
\section{Conclusions}
We have presented an evaluation of monocular morphable model fitting algorithms and a learning framework. It
is demonstrated that incorporation of texture term into the energy function significantly improves fitting accuracy.
Gains in the accuracy are quantified. We have trained a neural network using the outputs of the fitting algorithms
as training data. Our trained network is shown to perform on a par with existing approaches for the
task of monocular 3d face reconstruction while showing faster speed and smaller model size. Running time of our
network on a mobile devise is shown to be $3.6$ milliseconds enabling real-time applications. Our datasets and code for
evaluation are made publicly available.
\setlength{\tabcolsep}{4pt}
\begin{table}
\begin{center}
\caption{Model size and running time comparison.}
\label{table:comparisons}
\begin{tabular}{llllll}
\hline\noalign{\smallskip}
Method & Model size & GPU & CPU & ARM & AUC\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Jackson \cite{VRN} & 1.4 GB & - & 2.7 s & - & 0.849\\
Tewari \cite{MOFA} & 0.27 GB & 4 ms & - & - & 0.848\\
Tran \cite{tran2017extreme} & 0.35 GB & 40 ms & - & - & 0.810\\
Our & \textbf{1.5 MB} & \textbf{1 ms} & \textbf{1.8 ms} & 3.6 ms &
\textbf{0.854}\\
\hline
\end{tabular}
\end{center}
\end{table}
\setlength{\tabcolsep}{1.4pt}
\begin{figure}
\begin{center}
\begin{tabular}{ccccccc}
Original &
Fitting &
Our &
Jackson\cite{VRN} &
Tewari \cite{MOFA} &
Tran \cite{tran2017extreme} &
GT
\\
\subfloat{\includegraphics[width = 0.13\textwidth]{F015.jpg}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F015_fitting.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F015_our.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F015_VRN.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F015_MOFA.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F015_tran.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F015_gt.png}}
\\
\subfloat{\includegraphics[width = 0.13\textwidth]{F017.jpg}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F017_fitting.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F017_our.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F017_VRN.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F017_MOFA.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F017_tran.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F017_gt.png}}
\\
\subfloat{\includegraphics[width = 0.13\textwidth]{F025.jpg}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F025_fitting.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F025_our.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F025_VRN.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F025_MOFA.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F025_tran.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{F025_gt.png}}
\\
\subfloat{\includegraphics[width = 0.13\textwidth]{M002.jpg}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M002_fitting.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M002_our.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M002_VRN.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M002_MOFA.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M002_tran.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M002_gt.png}}
\\
\subfloat{\includegraphics[width = 0.13\textwidth]{M030.jpg}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M030_fitting.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M030_our.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M030_VRN.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M030_MOFA.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M030_tran.png}} &
\subfloat{\includegraphics[width = 0.13\textwidth]{M030_gt.png}}
\end{tabular}
\caption{Qualitative results. Images are from BU4DFE-test. Our implementation
of fitting was used for the second column.}
\label{fig:qualitative}
\end{center}
\end{figure}
\bibliographystyle{splncs}
|
\section{Introduction}
\label{sec:introduction}
Nonlinear optical processes enable complex manipulation of light and have been exploited extensively both in the classical and quantum regime for a wide variety of purposes, e.g. classical single- and multiple-channel frequency conversion \cite{Kumar1990, Chou1999}, optical parametric amplification \cite{Cerullo2003}, generation of squeezed states and entangled photons \cite{Pysher2008, Thyagarajan2009,Stefszky2017}, frequency conversion for single-photon detection \cite{Albota2004, Roussev2004, Vandevender2004} and to interface single photons with quantum memories \cite{Pelc2012a, Rutz2017, Maring2017}.
Realizing nonlinear processes in integrated waveguides is fundamental in bringing quantum protocols and devices closer to every-day life \cite{Orieux2016}.
Integrated nonlinear waveguides offer a few advantages over bulk nonlinear crystals, since they achieve a stronger nonlinear interaction by increasing the field confinement over longer lengths and can be interfaced more easily with fibre networks \cite{Montaut2017}.
Moreover, they can be integrated along with other linear and nonlinear elements to generate and manipulate different quantum states of light \cite{Krapick2013, Kruse2015, Sansoni2017,Lenzini2018arxiv}.
However, the nonlinear properties of integrated waveguides critically depend upon their quality and any fabrication imperfection can degrade the final performance.
In the classical regime, studies have already been performed to understand the relationship between fabrication imperfections and phasematching properties.
In particular, Lim et al. \cite{Lim1990} introduced the concept of \textit{noncritically phasematched} waveguides, i.e. waveguides that are specifically designed to minimise the impact of fabrication imperfections. They also derived fabrication conditions for noncritically phasematched thin-film and slab waveguides.
Experimentally, noncritical phasematching conditions for second harmonic generation have been investigated in annealed proton-exchanged lithium niobate (APE-LN) waveguides \cite{Bortz1994}.
In the field of quantum optics, different studies have addressed the influence of fabrication imperfections on the generation of photon pairs through parametric down conversion (PDC) in waveguides \cite{Pelc2010,Pelc2011a,Phillips2013} and photonics crystal fibres \cite{FrancisJones2016}.
The vast majority of these analyses has investigated the connection between fabrication imperfections and maximum conversion efficiency of the system.
For quantum applications, however, other properties become more critical depending on the intended task of the device, e.g. the phasematching bandwidth or its shape. Therefore, it is important to analyse the influence of fabrication imperfections in quantum devices systems bearing in mind their specific application.
In this paper, we study the impact of fabrication imperfections on the performance of waveguides in nonlinear crystals.
In particular, we analyse a variety of quantum processes realized in titanium in-diffused lithium niobate (Ti:LN) waveguides and show how their nonlinear performance is degraded by the presence of errors on the waveguide width.
In section \ref{subsec:mathdescription} we derive a qualitative expression that relates the length of a device to the maximum fabrication error tolerable. In section \ref{subsec:numericalanalysis} we apply this relation to estimate the effect of fabrication errors for a variety of different quantum optics processes of interest realized in titanium in-diffused lithium niobate (Ti:LN) waveguides. In section \ref{subsec:statanalysis} the effect of different types of fabrication imperfections is investigated by means of stochastic simulations and in section \ref{subsec:applications} we discuss how fabrication imperfections affect quantum state generation and manipulation. We focus our attention to the cases of squeezing generation, high-dimensional frequency bin encoding and efficient bandwidth compression of single photons.
Finally, in section \ref{sec:discussion} we discuss the results of the simulations and describe possible ways to overcome the fabrication limits that degrade the phasematching.
\section{Qualitative model for fabrication tolerances}
\label{sec:qualitativeModel}
\subsection{Mathematical model}
\label{subsec:mathdescription}
We begin deriving a simple model describing the effect of fabrication imperfections on the efficiency of a nonlinear process.
In particular, we assume that the fabrication imperfection is constant along the sample length.
Consider a general three-wave mixing (TWM) process in a waveguide
\begin{eqnarray*}
\omega_3 &= \omega_2 + \omega_1\\
\Delta\beta &=\beta_3 - \beta_2 - \beta_1,
\end{eqnarray*}
where $\omega_i$ and $\beta_i$ are the frequencies and the momenta of the fields involved, respectively, and $i=1,2,3$ denotes the three interacting fields.
If the momentum mismatch $\Delta\beta$ is constant along the waveguide, an exact solution for the phasematching spectrum $\phi (\Delta\beta)$ of the process is given by \cite{Boyd2008}
\begin{equation}
\phi(\Delta\beta) \propto \frac{1}{L}\int_0^L\rme^{\rmi\Delta\beta z}\rmd z \Rightarrow \phi(\Delta\beta)\propto\mathrm{sinc}\left(\frac{\Delta\beta L}{2}\right)\mathrm{e}^{\rmi\frac{\Delta\beta L}{2}},
\label{eq:ideal_pm}
\end{equation}
where $L$ is the crystal length. In the case of quasi-phasematching, $\Delta\beta$ has to include the effect of the grating vector $\beta_{QPM} = 2\pi/\Lambda$.
The propagation constants $\beta_i = 2\pi n_i/\lambda_i$ depend on the refractive index $n_i$ seen by the light field as it propagates in the crystal.
Fields propagating in a waveguide see an \textit{effective} refractive index dependent on the local refractive index distribution $n(x,y,z)$ \cite{Suhara2003}.
If $n(x,y,z)$ does not vary along the propagation axis $z$ of the waveguide, then $n^{eff}$ is constant and the waveguide is said to be \textit{homogeneous}. This assumption simplifies the treatment of phasematching in waveguide structures and in the rest of this section we will consider this scenario. The analysis of spatially-varying $n^{eff}(z)$ will be analysed in detail in section \ref{subsec:statanalysis}.
\begin{figure}[tbp]
\centering
\includegraphics[width=.5\textwidth]{figure1}
\caption{Illustration of the effect on the phasematching of a uniform variation $\delta f$ of the waveguide fabrication parameter $f$. The solid blue line shows the intensity of a desired nonlinear process for a waveguide with fabrication parameter $f_0$, plotted against the dimensionless parameter $\frac{\Delta\beta L}{2}$. A variation $\delta f$ of the fabrication parameter will shift the phasematching curve (dashdotted orange line). For shifts greater than the HWHM of the phasematching, the efficiency of the target process, represented by the dashed blue line, drops below 50\%. This condition is used to establish a simple criterion to indicate when the fabrication error $\delta f$ moves the process outside the chosen tolerance limits.}
\label{img:bw_description}
\end{figure}
Following the approach presented in \cite{Lim1990}, we consider a homogeneous waveguide designed for a specific TWM process.
For simplicity, we analyse the influence of a single fabrication parameter having a nominal value $f_0$.
Such a parameter can represent, for example, the waveguide width, depth, exchange temperature, etc.
Due to fabrication imperfections, the fabrication parameter $f_{prod}$ during the production can be off from the designed one by $\delta f = f_{prod}-f_0$. The fabrication error $\delta f$ will modify the $n^{eff}$ of the waveguide, which in turn has an impact on $\Delta\beta$. For this reason, a fabrication error $\delta f$ will shift the position of the phasematching curve $\phi(\Delta\beta)$ and reduce the overall efficiency of the process, as shown in Figure \ref{img:bw_description}.
To specify the fabrication tolerances for the waveguide production, we allow variations in $\Delta\beta$ such that the efficiency of the target process remains greater than 50\% of the ideal value:
\begin{equation}
\sinc{\frac{\Delta\beta L}{2}}^2\geq 0.5\Rightarrow \left|\frac{\Delta\beta L}{2}\right|\leq \Gamma,
\label{eq:bandwidth}
\end{equation}
where $\Gamma\approx 1.39$ is the half-width at half maximum of $\mathrm{sinc}(x)^2$.
Expanding $\Delta\beta$ in a Taylor series
\begin{equation*}
\Delta\beta = \Delta\beta(f_0) + \left.\partial_f\Delta\beta\right|_{f_0}\cdot\delta f + o(\delta f^2),
\end{equation*}
and noticing that, for the target process, $\Delta\beta(f_0)=0$, we can approximate (\ref{eq:bandwidth}) to the first order as
\begin{equation}
\left|\left.\partial_f\Delta\beta\right|_{f_0}\cdot\delta f\right|\frac{L}{2}\leq \Gamma.
\label{eq:sensitivity_waveguide}
\end{equation}
The parameter $\partial_f\Delta\beta$ can be referred to as the \textit{process sensitivity} to parameter $f$ because it relates the length of a waveguide to the maximum fabrication error allowable. In fact, assuming a maximum fabrication error of $\delta f_{max}$, from (\ref{eq:sensitivity_waveguide}) we can determine the maximum waveguide length $L_{max}$ to ensure that the process efficiency is greater than 50\%
\begin{equation}
L_{max} = \frac{2 \Gamma}{\left|\partial_f\Delta\beta\right|\cdot \delta f_{max} }.
\label{eq:sensitivity}
\end{equation}
It is therefore clear that any fabrication error poses an unavoidable constraint on the waveguide length.
However, one can see that if $|\partial_f\Delta\beta|$ approaches $0$, then $\delta f_{max}$ tends to infinity. Under this condition, the waveguide becomes first-order insensitive to the fabrication parameter variations. The condition $\partial_f\Delta\beta = 0$ is known as \textit{noncritical phasematching} and has been investigated in detail in previous works \cite{Lim1990,Bortz1994}.
Two main conclusions can be drawn from the analysis presented in this section. Firstly, for a desired process, there is an inverse proportionality between maximum waveguide length and fabrication errors. This means that the technological accuracy poses a well defined limit on the maximum length of the waveguides. The second conclusion is that technological imperfections can be mitigated if the \textit{process sensitivity} is minimized through careful waveguide design, thereby approaching noncritical phasematching \cite{Lim1990,Bortz1994}.
It is worth stressing that these conclusions are independent of the specific waveguide technology or waveguide geometry and therefore can be applied to all systems described by (\ref{eq:ideal_pm}).
While the model has been derived for a constant fabrication error, which will not usually be the case, its predictions still provide a qualitative description of device performance in the presence of fabrication imperfections, as we will show in the following sections.
\subsection{Numerical analysis of $|\partial_w\Delta\beta|$ for Ti:LN waveguides}
\label{subsec:numericalanalysis}
We now apply the previous theory to titanium in-diffused lithium niobate (Ti:LN) waveguides in order to study the technological limits of this platform.
Ti:LN waveguides have been widely used for classical and quantum applications \cite{Regener1988, Amin1997, Kanbara1999, Thyagarajan2009,Eckstein2011, Luo2015, Stefszky2017, Stefszky2018}. They exhibit extremely low propagation losses ($<$ 0.1 dB/cm), can guide both TE and TM polarization modes, possess high nonlinearity, allow on-chip manipulation of the light field via integrated beamsplitters and acousto- and electrooptical modulators and can be easily interfaced to fibre network via pigtailing \cite{Montaut2017}.
As illustrated in Figure \ref{img:wg_fabrication}, Ti:LN waveguides are produced by photolithographic patterning of titanium stripes with definite widths $w$ and thicknesses $\tau$ on top of a LiNbO${}_3$ substrate.
Subsequently, titanium is diffused inside the LN lattice by heating the sample in an oven.
The resulting waveguide is defined by the initial titanium stripe geometry, the exchange temperature and the exchange time.
Finally, periodic poling is performed by electric field poling after photolithographic patterning of the electrodes on the crystal faces.
Fabrication errors can occur at different steps, e.g. inhomogeneous illumination conditions can affect the patterning of the titanium stripes, or temperature gradients in the diffusion oven can lead to inhomogeneous diffusion of the titanium.
All these imperfections add up and can cause local deviations of the waveguide profile with respect to the ideal, homogeneous case.
Here we use (\ref{eq:sensitivity}) to study qualitatively the fabrication limits of Ti:LN waveguides.
To simplify the treatment, we choose to consider only one source of error, namely variation of the width $w$ of the Ti stripe (from now on we will refer to $w$ simply as the waveguide width).
In order to estimate $\partial_w\Delta\beta$, the effective refractive index of the guided modes as a function of the wavelength, polarization and waveguide width is needed. We employ a finite element solver written in Python implementing the model described in \cite{Strake1988} to calculate the Sellmeier equations of waveguides produced with different widths $w$.
The process sensitivity $\partial_w \Delta\beta$ as a function of $w$ for different processes is shown in Figure \ref{img:sensitivity_LN} for a number of processes of interest for quantum applications, namely:
type-0 PDC \cite{Stefszky2017}; type-II PDC \cite{Sansoni2017}; the quantum pulse gate (QPG) \cite{Eckstein2011}; the resonant PDC source described in \cite{Luo2015} and counter-propagating PDC generating photons at 1510nm and 1550nm.
\begin{figure}[tbp]
\begin{minipage}[t]{.45\textwidth}
\centering
\includegraphics[width=.6\textwidth]{figure2}
\caption{Standard Ti:LiNbO${}_3$ waveguide fabrication technique. From top to bottom: a) deposition of a titanium layer on top of the lithium niobate substrate and spin coating of a photoresist layer; b) photolithographic patterning of the photoresist; c) etching of the titanium to define the Ti stripe to be diffused; d) diffusion of the Ti stripe in the substrate to define the waveguide.}
\label{img:wg_fabrication}
\end{minipage}\hfill%
\begin{minipage}[t]{.45\textwidth}
\centering
\includegraphics[width=\textwidth]{figure3}
\caption{Calculated sensitivity $\partial_w\Delta\beta$ of different processes to variations of the Ti stripe width for Ti:LN waveguides. The processes analysed are: type-0 PDC (775nm$\rightarrow$1550nm, e$\rightarrow$ee), type-II PDC (775nm$\rightarrow$1550nm, o$\rightarrow$eo), quantum pulse gate \cite{Eckstein2011, Allgaier2017} (1550nm+860nm$\rightarrow$553nm, oe$\rightarrow$o), resonant PDC \cite{Luo2015} (532nm$\rightarrow$890nm+1320nm, o$\rightarrow$eo) and counter-propagating PDC (765nm$\rightarrow$1510nm+1550nm, e$\rightarrow$ee). }
\label{img:sensitivity_LN}
\end{minipage}
\end{figure}
\begin{figure}[btp]
\centering
\subfloat[Type-0 PDC 775nm$\rightarrow$1550nm]
{
\includegraphics[width=.4\textwidth]{figure4a}
\label{subpl:LNSHG0}} \quad
\subfloat[Quantum pulse gate]
{\includegraphics[width=.4\textwidth]{figure4b}
\label{subpl:LNQPG}}
\caption{The dependence of the maximum width error on the chosen waveguide length is shown for type-0 PDC (\ref{subpl:LNSHG0}) and for the quantum pulse gate (\ref{subpl:LNQPG}). This dependence is shown for different waveguide widths. The plots show that the maximum allowable width error decreases as $1/L$ and that wider waveguides are less sensitive to the width error. Note that the 13$\mu$m line in Figure \ref{subpl:LNSHG0} is not present because it is first-order immune to noise.}
\label{img:LNsensit}
\end{figure}
Recall that the waveguide is first-order immune to noise if the condition $\partial_w\Delta\beta=0$ is met.
Among the processes considered, only the resonant PDC process is non-critically phasematched in a regime where the waveguide is single-mode at telecom wavelengths. The type-0 PDC process is noncritically phasematched for $w$=13$\mu$m, but the waveguide is spatially multimode for this width. This is unfortunate as single-mode operation is often required.
Another important observation is that each process has a different sensitivity; even the ones involving similar wavelengths exhibit very different behaviour, e.g. type-0 and type-II PDC. Therefore, the process sensitivity has to be investigated independently for every process under consideration.
Using the calculated process sensitivities, we can estimate the maximum tolerable width error depending on the desired sample length using (\ref{eq:sensitivity}).
The results for type-0 PDC and the QPG are displayed in Figure \ref{img:LNsensit}. The model predicts that width errors of $|\delta w|\leq 0.2 \mu$m already limit the maximum waveguide lengths for these two processes to around $10$mm. The results for the other processes are reported in the supplementary material.
As we will show in the following section, this simple estimation still provides a good indication of the region where fabrication imperfection may start to play a role.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.5\textwidth]{figure5}
\caption{Example of the effect of different noise profiles on the phasematching intensity spectrum. In the main plot, the phasematchings of the waveguides without noise (black dotted line), with additive white gaussian noise (blue dashdotted line) and with $1/f$ noise spectrum (orange solid line) are shown. In the inset, the respective waveguide width profiles are reported. The device under consideration is a 20mm-long Ti:LiNbO${}_3$ waveguide for type-0 PDC 775nm $\rightarrow$ 1550nm, characterized via the reverse process, SHG.}
\label{img:noise_comparison}
\end{figure}
\section{Phasematching in inhomogeneous guiding structures.}
\label{sec:inhomogeneousWaveguides}
\subsection{Impact of different noise profiles on the phasematching.}
\label{subsec:statanalysis}
The analysis conducted in the previous sections considered only homogeneous waveguides.
However, in reality fabrication errors can occur randomly along the waveguide, thus leading to the production of \textit{inhomogeneous} waveguides, where the refractive index distribution varies along $z$.
For this reason, a spatially-varying fabrication parameter $f_{prod}(z)$ leads to a momentum mismatch $\Delta\beta(z)$ that varies along the waveguide.
In this case equation (\ref{eq:ideal_pm}) does not hold anymore and a more general expression has to be considered \cite{Helmfrid1992}
\begin{equation}
\phi \propto \frac{1}{L}\int_0^L \rme^{\rmi\int_0^z\Delta\beta(\xi) \rmd\xi} \rmd z.
\label{eq:general_pm}
\end{equation}
Integration of (\ref{eq:general_pm}) is possible usually only numerically and by assuming specific profiles for the momentum mismatch variation $\Delta\beta(z)$ along the waveguide.
Moreover, the phasematching spectrum will not result in the usual $\mathrm{sinc}^2$ shape \cite{Helmfrid1991,Pelc2010,Pelc2011a,Phillips2013}.
In the past, investigation of waveguides with variable dispersion profiles has been restricted to classical SHG systems assuming simple profiles for $\Delta\beta(z)$ \cite{Helmfrid1991}.
On the other hand, random fabrication errors may dramatically affect the desired quantum state produced in waveguide systems.
Therefore, in the remaining sections we study the effect of randomly variable dispersion relations in waveguides designed for quantum processes.
Here we study the phasematching properties of inhomogeneous Ti:LN waveguides as a function of the Ti stripe width $w$ and its maximum error $\delta w$.
Generating different profiles for $w(z)$ and calculating the relative momentum mismatch $\Delta\beta(z)$, we can integrate numerically (\ref{eq:general_pm}) to calculate the relative phasematching spectra.
The details of the simulations of this section are presented in the supplementary material, section 2.
For simplicity, we investigate fabrication errors $\delta w(z)$ characterized by two types of noise spectra, namely additive white gaussian (AWG) and $1/f$ noise. AWG noise describes uncorrelated noise fluctuations along the waveguide, while $1/f$ noise is characterised by spatial correlations and accounts for long-range drifts in the production parameters.
An example of how these noise spectra affect the width profile and the phasematching spectra is shown in Figure \ref{img:noise_comparison}.
To understand the main differences between the two types of noise, we study the performance of a 20mm-long, 7$\mu$m-wide waveguide designed for a type-0 second harmonic generation (SHG) pumped at 1550nm. We investigate the degradation of the conversion efficiency for values of the fabrication error $\delta w \in [0, 1.0]\mu$m for both types of noise.
The results of the simulations are presented in Figure \ref{img:comparison_awgn_pink}.
The two types of noise have very different impact on the maximum achievable conversion efficiency: AWG noise has a negligible influence, while $1/f$ noise can drastically decrease it.
Furthermore, the reduction of conversion efficiency is accompanied by an increase of the phasematching bandwidth, especially for errors $\delta w>0.25 \mu$m, whose broadened phasematching is shown in the insets of Figure \ref{img:comparison_awgn_pink}.
The same analysis has been performed for the other processes characterized in Figure \ref{img:sensitivity_LN} and the results are similar: AWG noise consistently has a negligible impact on the average maximum conversion efficiency, while $1/f$ noise rapidly degrades the performance of the device as $\delta w$ increases.
These results are well in agreement with previous studies on different systems. The presence of AWG noise on the poling grating of periodically-poled waveguides has been previously analysed in \cite{Pelc2010, Pelc2011a, Phillips2013} and showed only a minor influence on the maximum conversion efficiency. Moreover, a comparison between correlated and uncorrelated noise has been investigated in photonics crystal fibres, showing that imperfections with long-range correlations drastically effect the parametric gain of nonlinear processes \cite{Farahmand2004}.
In the rest of the paper, we will focus our attention exclusively on the effects of $1/f$ noise, since this is the main cause of phasematching distortions.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.5\textwidth]{figure6}
\caption{Maximum efficiency of a nonlinear process as a function of the waveguide width error for the ideal waveguide (black dotted line), a waveguide with AWG noise (blue dashdotted line) and $1/f$ noise (orange solid line). Shaded regions correspond to the range of results of 40 simulations for each datapoint.
The insets show the effect of $1/f$ noise on the phasematching intensity in comparison to an ideal waveguide.
The shaded area represent the range of simulated intensity spectra from 40 simulations.
A broadening of the average phasematching spectrum, more prominent side lobes and reduction of the efficiency are evident.
The device under consideration is a 20mm-long Ti:LiNbO${}_3$ waveguide for type-0 PDC 775nm $\rightarrow$ 1550nm, characterized via the reverse process, SHG.}
\label{img:comparison_awgn_pink}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.5\textwidth]{figure7}
\caption{Maximum efficiency of a nonlinear process as a function of the waveguide width error, in the presence of $1/f$ noise. The three different curves are calculated for nominal widths of 7$\mu$m (blue dotted line), 13$\mu$m (orange solid line) and 18$\mu$m (green dashed line). Shaded regions are the errorbars as retrieved from stochastic simulations. Note that the 13$\mu$m waveguide is noncritically phase matched and so is virtually immune to the presence of noise. The device under consideration is a 20mm-long Ti:LiNbO${}_3$ waveguide for type-0 PDC 775nm $\rightarrow$ 1550nm, characterized via the reverse process, SHG.}
\label{img:SHG0_stat_analysis}
\end{figure}
Having established a framework suitable for the study of waveguide inhomogeneities, we can now compare the approximate results derived in section \ref{sec:qualitativeModel} with numerical simulations.
An important result was the ability to predict design of noncritically phasematched waveguides. In particular, we calculated that a type-0 PDC process pumped at 775nm is noncritically phasematched for $w=13\mu$m.
This result is confirmed by evaluating the conversion efficiency of the reverse process, a type-0 SHG pumped at 1550nm, as a function of the waveguide width $w$ and the error $\delta w$ in presence of $1/f$ noise.
Indeed, Figure \ref{img:SHG0_stat_analysis} shows that a 13$\mu$m-wide waveguide is practically immune to $1/f$ noise when close to noncritical phasematching.
Moreover, from the calculations reported in Figure \ref{subpl:LNSHG0}, we expect that a 7$\mu$m-wide, 20mm-long waveguide will be sensitive to noise values $\delta w\geq 0.1 \mu$m. As shown in Figure \ref{img:SHG0_stat_analysis}, for $\delta w \geq 0.1\mu$m, the maximum efficiency rapidly degrades below 90\% of the ideal maximum. This confirms that the simplified model can provide reliable qualitative information about the waveguides' sensitivity to noise and thus the evaluation of the process sensitivity is can provide useful technological boundaries for the process quality of waveguide production.
\subsection{Applications}
\label{subsec:applications}
The theory presented so far is now applied to three different systems of interest in quantum optics. In fact, we will show that it is necessary to consider the impact of fabrication errors in these systems to correctly model and estimate their performance.
In section \ref{subsec:squeezing} we analyse the effect of waveguide inhomogeneities on the maximum squeezing attainable in a waveguided system; in section \ref{subsec:hdqkd} we estimate how noise reduces the maximum number of bins in a frequency-bin encoding (FBE) scheme; in section \ref{subsec:QPG} we study the effect of waveguide width noise on the bandwidth compression factor of a frequency conversion device.
\subsubsection{Impact of fabrication errors in squeezing generation.}
\label{subsec:squeezing}
We first consider a waveguide structure designed to produce continuous-wave (CW) single-mode squeezed states in a single-pass configuration.
These states are the foundation for continuous-variable (CV) quantum optics: they can be used as a basis for CV quantum computing \cite{Menicucci2006}, they have been used to generate complex quantum states such as EPR entanglement \cite{Bowen2002} and CV cluster states \cite{Yukawa2008}, and they have been used in sensing and metrology in order to improve the sensitivity of measurements, e.g. in gravitational-wave astronomy \cite{Aasi2013}.
We consider here a 7$\mu$m-wide Ti:LN waveguide pumped at 775nm that produces type-0 squeezing at 1550nm in a single-pass configuration.
It can be shown that both the losses of the fundamental field and the strength of the nonlinear process are critical to the amount of squeezing produced \cite{Serkland1995}.
We begin by neglecting the losses, thereby exclusively investigating the effect of waveguide width imperfections on the strength of the nonlinear process.
The strength of the nonlinear process can be found by performing second harmonic generation in such a sample, from which one can calculate the normalized conversion efficiency using
\begin{equation}
\eta_{norm} =\frac{P_{SH}}{P_{FF}^2 L^2}
\label{eq:normalized_conversion_efficiency}
\end{equation}
A common misconception is that, due to its definition, $\eta_{norm}$ is independent of length.
However, the qualitative model presented in Figure \ref{subpl:LNSHG0} shows that longer waveguides are more susceptible to fabrication imperfections, therefore we expect $\eta_{norm}$ to be dependent on waveguide length.
To calculate $\eta_{norm}$ in presence of fabrication errors, we numerically simulate the phasematching spectra of the system for different sample lengths $L\in$ [10, 60]mm and width error magnitude $\delta w\in$ [0, 0.5]$\mu$m.
For each parameter combination, we calculate the maximum conversion efficiency of 40 randomly generated systems to estimate the average normalized conversion efficiency.
From Figure \ref{subimg:norm_conv_eff_shg0}, it is evident that the normalized conversion efficiency is critically dependent on both $L$ and $\delta w$.
The simulations reveal that both $\eta_{norm}$ and the waveguide length of each sample are necessary to fairly compare the performance of different devices.
Furthermore it can be seen that the normalized conversion efficiency drops from 49 \%/Wcm$^2$ to 40\%/Wcm$^2$ for 10mm-long waveguides and below 15\%/Wcm$^2$ for 60mm-long waveguides.
Therefore, one will overestimate the benefits of longer samples if the impact of fabrication errors is not taken into account.
From the calculated normalized conversion efficiencies, one can estimate the amount of squeezing that can be produced in this device.
Follwing \cite{Serkland1995}, the maximum squeezing $S$ achievable in a single-pass CW waveguide can be given by
\begin{equation}
S=\left(\rme^{-2\sqrt{\eta_{norm}P_{in}}L} \rme^{-\alpha L}\right)+1-\rme^{-\alpha L},
\end{equation}
where $P_{in}$ is the input pump power of the squeezer and $\alpha$ is the loss for the squeezed field. We assume a negligible effect of the losses for the 775nm pump.
We consider $P_{in}$ = 500mW at 775nm and propagation losses $\alpha$ equal to 0.1 dB/cm, a safe estimate of the average losses measured in Ti:LN waveguides \cite{Stefszky2017}.
The squeezing $S$ produced as a function of $L$ and $\delta w$ is shown in Figure \ref{subimg:squeezing}.
It can be seen that, for a given $\delta w$, there exists a waveguide length that maximizes the squeezing produced.
Moreover, this optimal length increases as the magnitude of the width error increases.
This is due to a complex interplay between the nonlinear interaction strength and the losses; as the waveguide length increases, the positive effect of an increase in the interaction length is counteracted by an increase in the total losses and a simultaneous reduction of the normalized conversion efficiency.
The simulations show that the system under investigation (with 500mW of pump power) can produce around -9.5 dB of squeezing, choosing a waveguide with an optimized length of 40mm, if the error is below $\delta w\leq$ 0.1$\mu$m.
To reduce the impact of fabrication errors, one can consider the use of noncritically phasematched systems.
For the system under consideration, this can be done by choosing a 13$\mu$m-wide waveguide, as shown in Figures \ref{img:LNsensit} and \ref{img:SHG0_stat_analysis}.
In this case we expect a normalized conversion efficiency that is independent of the waveguide length and the fabrication imperfections.
Note that insensitivity to fabrication imperfections is equivalent to having no fabrication imperfection.
Therefore, the squeezing produced in a noncritically phasematched waveguide corresponds to the values at $\delta w =0\mu$m in Figure \ref{subimg:squeezing}, neglecting a minor deviation in $\eta_{norm}$ due to differences in the overlap of the interacting fields in the wider waveguide.
\begin{figure}[tbp]
\begin{minipage}[t]{.45\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figure8}
\caption{Normalised conversion efficiency as a function of the error on the waveguide width, for lossless samples of varying lengths. It can be seen that both the length of the sample and the magnitude of the width error have a strong impact on the normalized conversion efficiency, even in the absence of losses. The device under consideration is a 7$\mu$m-wide Ti:LiNbO${}_3$ waveguide for type-0 PDC 775nm $\rightarrow$ 1550nm, characterized via the reverse process, SHG.}
\label{subimg:norm_conv_eff_shg0}
\end{minipage}\hfill%
\begin{minipage}[t]{.45\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figure9}
\caption{Squeezing relative to shot noise exiting the waveguide for samples with different lengths and widths error. The process is pumped with 500mW of CW input at 775nm. Losses for the fundamental field are assumed to be 0.1 dB/cm. The device under consideration is a 7$\mu$m-wide Ti:LiNbO${}_3$ waveguide for type-0 PDC 775nm $\rightarrow$ 1550nm}
\label{subimg:squeezing}
\end{minipage}
\end{figure}
\subsubsection{Impact of fabrication errors on quantum information encoding.}
\label{subsec:hdqkd}
Frequency-bin encoding (FBE) is an attractive scheme for implementation of quantum information processing protocols because it offers an unbounded space for high-dimensional encoding compatible with standard fibre networks.
Furthermore, FBE can be implemented using PDC sources, which are a versatile and tunable platform that has been developed for many years.
Here, we study the limitations of Ti:LN waveguides as pulsed PDC sources for FBE and evaluate the impact of fabrication imperfections on such systems.
We consider a type-0 PDC source in a Ti:LN waveguide, pumped with at 775nm, generating pairs of frequency-bin entangled photons in the telecom C-band, between 1530nm and 1570nm.
The physical device is analoguous to the one presented in \cite{Olislager2010}.
Typically, PDC sources for FBE are pumped with CW light \cite{Olislager2010,Zhong2015}, however, pulsed systems provide advantages in terms of synchronization between the communicating parties.
Therefore, we consider a pulsed pump laser with a pump bandwidth broader than the phasematching.
FBE benefits from having a large number of encoding bins; however, it is also important to minimize cross-talk between them.
As a compromise between these two factors, we define the frequency-bin bandwidth $\Delta b$ as the full width at half maximum (FWHM) of the phasematching spectrum and each bin is separated by $\Delta b/2$, as illustrated in Figure \ref{img:bin_bandwidth}.
From this definition, given the available frequency band $\Delta\lambda$, one can calculate the number of available bins $n_{bins}$
\begin{equation}
n_{bins} = \frac{\Delta\lambda}{1.5\Delta b},
\end{equation}
where $\Delta\lambda=40nm$ is the bandwidth of the telecom C-band.
The number of available bins is then used as a figure of merit for the system.
The bin bandwidth $\Delta b$ is extracted from the phasematching spectrum by fitting it with a Gaussian and taking the FWHM of the fit.
This analysis is applied for varying sample lengths $L$ and width error magnitudes $\delta w$ and the results of the calculations are shown in Figure \ref{img:numb_bins}.
Solid lines represent 7$\mu$m-wide waveguides, while dashed ones represent 13$\mu$m-wide waveguides. Shaded areas represent the range of simulation results over 40 iterations for each data point.
The simulations show that it is possible to implement more than 70 bins in a 60mm-long, provided that fabrication errors are minimal.
However, for the 7$\mu$m-wide waveguides, the number of bins available in longer waveguides drops rapidly with increasing width error to a minimum of approximately 10.
In fact, for errors above 0.4$\mu$m, 10mm-long waveguides may outperform 60mm-long ones.
The reason is that longer samples theoretically have much narrower phasematching bandwidth and these are much more susceptible to fabrication errors, as highlighted in section \ref{subsec:mathdescription}.
For this reason shorter samples can outperform longer ones, in the presence of large fabrication errors.
In contrast, the 13$\mu$m-wide waveguides do not show a reduction in the number of bins as the noise increases, as can be seen from the dashed lines in Figure \ref{img:numb_bins}.
This is due to the fact that they are noncritically phasematched and therefore immune to fabrication errors.
\begin{figure}[tbp]
\begin{minipage}[t]{.45\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figure10}
\caption{Definition of maximum bin bandwidth. A simple criterion to ensure low cross-talk between the different frequency bins is to define the bin size equal to the FWHM of the phasematching bandwidth, and distance between bins equal to half the FWHM.}
\label{img:bin_bandwidth}
\end{minipage}\hfill%
\begin{minipage}[t]{.45\textwidth}
\centering
\includegraphics[width=1.\textwidth]{figure11}
\caption{Maximum number of bins available for a frequency-encoded HDQKD protocol. A $1/f$ noise spectrum with maximum error $\delta w$ has been considered.
The bin size is set to be equal to the phasematching FWHM and the distance between the channels is half the FWHM. Solid lines are for a 7$\mu$m-wide waveguide, dashed line for a 13$\mu$m-wide waveguide.}
\label{img:numb_bins}
\end{minipage}
\end{figure}
\subsubsection{Impact of fabrication errors on the performance of a bandwidth compressor.}
\label{subsec:QPG}
Interfacing components operating at different wavelengths is a critical challenge for quantum optical networks.
Reduction of transmission losses is paramount in most applications and therefore transmission in the telecom C-band is desired, where losses are minimal.
However, many quantum devices operate outside this frequency band and therefore efficient frequency conversion between these bands is required.
Furthermore, it is often necessary to efficiently match the bandwidth of different quantum devices.
Both bandwidth matching and frequency conversion can be efficiently achieved in the integrated quantum pulse gate (QPG), a device that implements type-II sum frequency generation (SFG) in a waveguide \cite{Allgaier2017}.
The integrated QPG in \cite{Allgaier2017} was implemented in a 7$\mu$m-wide, 27mm-long Ti:LN waveguide, designed to convert single photons from the telecom C-band to 550nm.
The measured bandwidth compression factor (BCF) was $\Delta\nu_{in} / \Delta\nu_{out} = 7.47\pm 0.01$, where $\Delta\nu_{in/out}$ is the spectral bandwidth of the input and output photons.
In this device, the compression factor is directly related to the phasematching bandwidth: the narrower the bandwidth, the higher the compression factor.
We have already shown in section \ref{subsec:hdqkd} that fabrication imperfections can increase the phasematching bandwidth of a given process.
Therefore, we expect that the compression factor will reduce in the presence of fabrication imperfections.
We consider a 7$\mu$m-wide waveguide with different lengths $L$ and varying magnitude $\delta w$ of $1/f$ noise on the waveguide width.
The input bandwidth is set to $\Delta\nu_{in} = 963\pm 11$ GHz, while the output bandwidth $\Delta \nu_{out}$ is defined as the FWHM of a Gaussian fit to the phasematching spectrum, following the method of \cite{Allgaier2017}.
Each datapoint has been simulated 40 times and the results are shown in Figure \ref{subimg:bw_compr_7um}.
The calculated BCFs are represented in solid lines, while the shaded regions represent the range of variation in the simulated data.
Simulations show that a 40mm-long sample provides a BCF of $\sim$64, in the absence of fabrication imperfections.
This corresponds to an output bandwidth of $\sim$15GHz or, equivalently, a 30ps-long pulse, under the approximation of Gaussian phasematching.
In Figure \ref{subimg:bw_compr_7um} it is also shown, with a dashed line, the BCF measured in \cite{Allgaier2017} for their 27mm-long waveguide.
It is immediately evident that the measured compression factor is well below the theoretically predicted value.
In fact, calculations show that a 27mm-long sample should provide a compression factor close to 45 in the absence of imperfections; however, the experiment measured a compression factor of only 7.45.
Such a reduction would only be expected in the presence of width error $\delta w\geq 0.4\mu$m, as illustrated in the Figure.
Allgaier \textit{et al.} \cite{Allgaier2017} also characterized the phasematching spectrum of their device and the measurement showed deviations from the expected sinc$^{2}$ profile, as shown in Figure \ref{img:sim_JSA}
These deviations indicates the presence of non-negligible fabrication imperfections, as it was shown in Figure \ref{img:comparison_awgn_pink} that the presence of $1/f$ noise leads to more prominent side lobes and an asymmetric phasematching profile.
Assuming a $1/f$-noise profile and $\delta w = 0.4\mu$m, simulations have been able to reproduce the asymmetry and the prominent side lobes present in the measured phasematching, as can be seen comparing the measured and simulated spectra in Figure \ref{subimg:noisy_qpg} and Figure \ref{subimg:meas} respectively.
The limited compression and the shape of the phasematching lead us to conclude that the presence of fabrication imperfections limits the performance of the device presented in \cite{Allgaier2017}.
A previously discussed method to overcome these limitations is to design the process to be noncritically phasematched.
Although Figure \ref{img:sensitivity_LN} shows that noncritical phasematching cannot be found for the QPG, it is possible to reduce the process sensitivity by increasing the waveguide width.
Figure \ref{subimg:bw_compr_13um} shows the calculated BCFs for a 13$\mu$m-wide waveguide as the length and the magnitude of the noise are varied.
The results show a greatly reduced sensitivity to noise, as large fabrication errors have a lower impact on the BCF.
These systems would reliably permit BCFs above 40, resulting in pulses with bandwidths below 25 GHz.
Interestingly, these bandwidths correspond to pulses longer than 20ps at 550nm, a regime that is often difficult to reach.
These results highlight the fact that proper system design can have a drastic impact on the final performance of such devices.
\begin{figure}[hbtp]
\centering
\subfloat[Bandwidth compression for a 7$\mu$m-wide waveguide.]{
\includegraphics[width=.4\textwidth]{figure12a}
\label{subimg:bw_compr_7um}}\quad
\subfloat[Bandwidth compression for a 13$\mu$m-wide waveguide.]{
\includegraphics[width=.4\textwidth]{figure12a}
\label{subimg:bw_compr_13um}}
\caption{Bandwidth compression factor resulting from the up-conversion of 963 GHz-broad telecom photons, using the quantum pulse gate \cite{Allgaier2017}. The compression factor has been calculated for different waveguide width errors $\delta w$ and lengths $L$, in presence of noise with $1/f$ noise spectrum. Moreover, two different nominal width have been investigated, namely $w=7$ $\mu$m in Figure \Ref{subimg:bw_compr_7um} and $13$ $\mu$m in Figure \Ref{subimg:bw_compr_13um}. The experimental value reported in \cite{Allgaier2017} is also shown as a dashed black line in both plots for comparison.}
\label{img:bwcompression}
\end{figure}
\begin{figure}[hbtp]
\centering
\subfloat[Ideal phasematching]{
\includegraphics[width=.3\textwidth]{figure13a}
\label{subimg:noiseless_qpg}}\quad
\subfloat[Phasematching measured in \cite{Allgaier2017}.]{
\includegraphics[width=.27\textwidth]{figure13b}\quad
\label{subimg:meas}}
\subfloat[Simulated phasematching for a noisy waveguide.]{
\includegraphics[width=.3\textwidth]{figure13c}
\label{subimg:noisy_qpg}}
\caption{Theoretical versus experimental performance of the bandwidth compression device presented in \cite{Allgaier2017}. In Figure \Ref{subimg:noiseless_qpg} the ideal phasematching is shown and can be compared with the measured phasematching shown in Figure \Ref{subimg:meas}. A phasematching similar to the measured one can be produced assuming a $1/f$ noise on the waveguide width and a maximum width error $\delta w$ = 0.4$\mu$m, as shown in Figure \Ref{subimg:noisy_qpg}.}
\label{img:sim_JSA}
\end{figure}
\section{Discussion}
\label{sec:discussion}
So far, we have shown that fabrication imperfections in integrated nonlinear systems limit the useful length and maximum efficiency of the devices.
A natural question is then how to overcome these limitations and optimize device performance.
Three general methods can be employed to improve the overall efficiency of the devices: reducing the magnitude of fabrication imperfections, designing the process to be insensitive to fabrication errors and reducing the impact of fabrication imperfections by using shorter waveguides.
Whilst it may be possible to reduce the magnitude of fabrication imperfections for a given production process, these devices are often realized using state-of-the-art technology and it may not be possible to make further improvements.
In any case, the unavoidable presence of technological errors during waveguide production will impose an ultimate limit to device performance.
It is therefore crucial to devise other methods to overcome these limits.
One solution is to design the process to reduce the sensitivity to fabrication errors, as discussed in \ref{subsec:mathdescription}.
Ideally, one would design waveguides in an noncritically phasematched regime \cite{Lim1990}; however, this is not always possible, as shown in Figure \ref{img:sensitivity_LN}.
Nevertheless, choosing the appropriate design will minimize the process sensitivity and reduce the impact of fabrication imperfections, as discussed at the end of section \ref{subsec:QPG}.
Another method to reduce the sensitivity to fabrication errors is to choose shorter waveguides; an unavoidable drawback is the reduction of the conversion efficiency.
This can be compensated for by employing multi-pass schemes, e.g. double-pass systems or cavity configurations.
This technique, well suited for CW systems \cite{Stefszky2017}, removes the need for long waveguides at the cost of increased device complexity.
\section{Conclusions}
\label{sec:conclusions}
In this paper we derived a framework to study the limits posed by fabrication imperfections on the nonlinear performance of waveguide devices.
A qualitative model was first developed to describe the effect of fabrication imperfections on the performance of nonlinear waveguides.
This model showed that long waveguides are more susceptible to fabrication errors occurring during the production.
A quantitative model was then introduced, which is able to account for inhomogeneities along the length of the waveguide.
Applying this model to Ti:LN waveguides revealed that fabrication imperfections with long range spatial correlations leads to a reduced conversion efficiency and a distortion of the phasematching spectrum.
Finally, we studied the impact of imperfections on three prominent quantum processes.
Despite each process having different figures of merit, e.g. conversion efficiency or phasematching bandwidth, the performance of each process was found to be limited by the presence of fabrication imperfections.
Therefore, it is crucial to take fabrication imperfections into account when designing quantum optics devices.
\section*{Acknowledgement}
The authors thank J. M. Donohue and C. Eigner for valuable discussion and helpful comments.
This research has received funding from the European Union (EU) Horizon 2020 Research and Innovation program under Grant Agreement No. 665148.
\clearpage
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.